Stay Ahead, Stay ONMINE

Image Captioning, Transformer Mode On

Introduction In my previous article, I discussed one of the earliest Deep Learning approaches for image captioning. If you’re interested in reading it, you can find the link to that article at the end of this one. Today, I would like to talk about Image Captioning again, but this time with the more advanced neural network architecture. The deep learning I am going to talk about is the one proposed in the paper titled “CPTR: Full Transformer Network for Image Captioning,” written by Liu et al. back in 2021 [1]. Specifically, here I will reproduce the model proposed in the paper and explain the underlying theory behind the architecture. However, keep in mind that I won’t actually demonstrate the training process since I only want to focus on the model architecture. The idea behind CPTR In fact, the main idea of the CPTR architecture is exactly the same as the earlier image captioning model, as both use the encoder-decoder structure. Previously, in the paper titled “Show and Tell: A Neural Image Caption Generator” [2], the models used are GoogLeNet (a.k.a. Inception V1) and LSTM for the two components, respectively. The illustration of the model proposed in the Show and Tell paper is shown in the following figure. Figure 1. The neural network architecture for image captioning proposed in the Show and Tell paper [2]. Despite having the same encoder-decoder structure, what makes CPTR different from the previous approach is the basis of the encoder and the decoder themselves. In CPTR, we combine the encoder part of the ViT (Vision Transformer) model with the decoder part of the original Transformer model. The use of transformer-based architecture for both components is essentially where the name CPTR comes from: CaPtion TransformeR. Note that the discussions in this article are going to be highly related to ViT and Transformer, so I highly recommend you read my previous article about these two topics if you’re not yet familiar with them. You can find the links at the end of this article. Figure 2 shows what the original ViT architecture looks like. Everything inside the green box is the encoder part of the architecture to be adopted as the CPTR encoder. Figure 2. The Vision Transformer (ViT) architecture [3]. Next, Figure 3 displays the original Transformer architecture. The components enclosed in the blue box are the layers that we are going to implement in the CPTR decoder. Figure 3. The original Transformer architecture [4]. If we combine the components inside the green and blue boxes above, we are going to obtain the architecture shown in Figure 4 below. This is exactly what the CPTR model we are going to implement looks like. The idea here is that the ViT Encoder (green) works by encoding the input image into a specific tensor representation which will then be used as the basis of the Transformer Decoder (blue) to generate the corresponding caption. Figure 4. The CPTR architecture [5]. That’s pretty much everything you need to know for now. I’ll explain more about the details as we go through the implementation. Module imports & parameter configuration As always, the first thing we need to do in the code is to import the required modules. In this case, we only import torch and torch.nn since we are about to implement the model from scratch. # Codeblock 1 import torch import torch.nn as nn Next, we are going to initialize some parameters in Codeblock 2. If you have read my previous article about image captioning with GoogLeNet and LSTM, you’ll notice that here, we got a lot more parameters to initialize. In this article, I want to reproduce the CPTR model as closely as possible to the original one, so the parameters mentioned in the paper will be used in this implementation. # Codeblock 2 BATCH_SIZE = 1 #(1) IMAGE_SIZE = 384 #(2) IN_CHANNELS = 3 #(3) SEQ_LENGTH = 30 #(4) VOCAB_SIZE = 10000 #(5) EMBED_DIM = 768 #(6) PATCH_SIZE = 16 #(7) NUM_PATCHES = (IMAGE_SIZE//PATCH_SIZE) ** 2 #(8) NUM_ENCODER_BLOCKS = 12 #(9) NUM_DECODER_BLOCKS = 4 #(10) NUM_HEADS = 12 #(11) HIDDEN_DIM = EMBED_DIM * 4 #(12) DROP_PROB = 0.1 #(13) The first parameter I want to explain is the BATCH_SIZE, which is written at the line marked with #(1). The number assigned to this variable is not quite important in our case since we are not actually going to train this model. This parameter is set to 1 because, by default, PyTorch treats input tensors as a batch of samples. Here I assume that we only have a single sample in a batch.  Next, remember that in the case of image captioning we are dealing with images and texts simultaneously. This essentially means that we need to set the parameters for the two. It is mentioned in the paper that the model accepts an RGB image of size 384×384 for the encoder input. Hence, we assign the values for IMAGE_SIZE and IN_CHANNELS variables based on this information (#(2) and #(3)). On the other hand, the paper does not mention the parameters for the captions. So, here I assume that the length of the caption is no more than 30 words (#(4)), with the vocabulary size estimated at 10000 unique words (#(5)). The remaining parameters are related to the model configuration. Here we set the EMBED_DIM variable to 768 (#(6)). In the encoder side, this number indicates the length of the feature vector that represents each 16×16 image patch (#(7)). The same concept also applies to the decoder side, but in that case the feature vector will represent a single word in the caption. Talking more specifically about the PATCH_SIZE parameter, we are going to use the value to compute the total number of patches in the input image. Since the image has the size of 384×384, there will be 576 patches in total (#(8)). When it comes to using an encoder-decoder architecture, it is possible to specify the number of encoder and decoder blocks to be used. Using more blocks typically allows the model to perform better in terms of the accuracy, yet in return, it will require more computational power. The authors of this paper decided to stack 12 encoder blocks (#(9)) and 4 decoder blocks (#(10)). Next, since CPTR is a transformer-based model, it is necessary to specify the number of attention heads within the attention blocks inside the encoders and the decoders, which in this case authors use 12 attention heads (#(11)). The value for the HIDDEN_DIM parameter is not mentioned anywhere in the paper. However, according to the ViT and the Transformer paper, this parameter is configured to be 4 times larger than EMBED_DIM (#(12)). The dropout rate is not mentioned in the paper either. Hence, I arbitrarily set DROP_PROB to 0.1 (#(13)). Encoder As the modules and parameters have been set up, now that we will get into the encoder part of the network. In this section we are going to implement and explain every single component inside the green box in Figure 4 one by one. Patch embedding Figure 5. Dividing the input image into patches and converting them into vectors [5]. You can see in Figure 5 above that the first step to be done is dividing the input image into patches. This is essentially done because instead of focusing on local patterns like CNNs, ViT captures global context by learning the relationships between these patches. We can model this process with the Patcher class shown in the Codeblock 3 below. For the sake of simplicity, here I also include the process inside the patch embedding block within the same class. # Codeblock 3 class Patcher(nn.Module): def __init__(self): super().__init__() #(1) self.unfold = nn.Unfold(kernel_size=PATCH_SIZE, stride=PATCH_SIZE) #(2) self.linear_projection = nn.Linear(in_features=IN_CHANNELS*PATCH_SIZE*PATCH_SIZE, out_features=EMBED_DIM) def forward(self, images): print(f’imagestt: {images.size()}’) images = self.unfold(images) #(3) print(f’after unfoldt: {images.size()}’) images = images.permute(0, 2, 1) #(4) print(f’after permutet: {images.size()}’) features = self.linear_projection(images) #(5) print(f’after lin projt: {features.size()}’) return features The patching itself is done using the nn.Unfold layer (#(1)). Here we need to set both the kernel_size and stride parameters to PATCH_SIZE (16) so that the resulting patches do not overlap with each other. This layer also automatically flattens these patches once it is applied to the input image. Meanwhile, the nn.Linear layer (#(2)) is employed to perform linear projection, i.e., the process done by the patch embedding block. By setting the out_features parameter to EMBED_DIM, this layer will map every single flattened patch into a feature vector of length 768. The entire process should make more sense once you read the forward() method. You can see at line #(3) in the same codeblock that the input image is directly processed by the unfold layer. Next, we need to process the resulting tensor with the permute() method (#(4)) to swap the first and the second axis before feeding it to the linear_projection layer (#(5)). Additionally, here I also print out the tensor dimension after each layer so that you can better understand the transformation made at each step. In order to check if our Patcher class works properly, we can just pass a dummy tensor through the network. Look at the Codeblock 4 below to see how I do it. # Codeblock 4 patcher = Patcher() images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE) features = patcher(images) # Codeblock 4 Output images : torch.Size([1, 3, 384, 384]) after unfold : torch.Size([1, 768, 576]) #(1) after permute : torch.Size([1, 576, 768]) #(2) after lin proj : torch.Size([1, 576, 768]) #(3) The tensor I passed above represents an RGB image of size 384×384. Here we can see that after the unfold operation is performed, the tensor dimension changed to 1×768×576 (#(1)), denoting the flattened 3×16×16 patch for each of the 576 patches. Unfortunately, this output shape does not match what we need. Remember that in ViT, we perceive image patches as a sequence, so we need to swap the 1st and 2nd axes because typically, the 1st dimension of a tensor represents the temporal axis, while the 2nd one represents the feature vector of each timestep. As the permute() operation is performed, our tensor is now having the dimension of 1×576×768 (#(2)). Lastly, we pass this tensor through the linear projection layer, which the resulting tensor shape remains the same since we set the EMBED_DIM parameter to the same size (768) (#(3)). Despite having the same dimension, the information contained in the final tensor should be richer thanks to the transformation applied by the trainable weights of the linear projection layer. Learnable positional embedding Figure 6. Injecting the learnable positional embeddings into the embedded image patches [5]. After the input image has successfully been converted into a sequence of patches, the next thing to do is to inject the so-called positional embedding tensor. This is essentially done because a transformer without positional embedding is permutation-invariant, meaning that it treats the input sequence as if their order does not matter. Interestingly, since an image is not a literal sequence, we should set the positional embedding to be learnable such that it will be able to somewhat reorder the patch sequence that it thinks works best in representing the spatial information. However, keep in mind that the term “reordering” here does not mean that we physically rearrange the sequence. Rather, it does so by adjusting the embedding weights. The implementation is pretty simple. All we need to do is just to initialize a tensor using nn.Parameter which the dimension is set to match with the output from the Patcher model, i.e., 576×768. Also, don’t forget to write requires_grad=True just to ensure that the tensor is trainable. Look at the Codeblock 5 below for the details. # Codeblock 5 class LearnableEmbedding(nn.Module): def __init__(self): super().__init__() self.learnable_embedding = nn.Parameter(torch.randn(size=(NUM_PATCHES, EMBED_DIM)), requires_grad=True) def forward(self): pos_embed = self.learnable_embedding print(f’learnable embeddingt: {pos_embed.size()}’) return pos_embed Now let’s run the following codeblock to see whether our LearnableEmbedding class works properly. You can see in the printed output that it successfully created the positional embedding tensor as expected. # Codeblock 6 learnable_embedding = LearnableEmbedding() pos_embed = learnable_embedding() # Codeblock 6 Output learnable embedding : torch.Size([576, 768]) The main encoder block Figure 7. The main encoder block [5]. The next thing we are going to do is to construct the main encoder block displayed in the Figure 7 above. Here you can see that this block consists of several sub-components, namely self-attention, layer norm, FFN (Feed-Forward Network), and another layer norm. The Codeblock 7a below shows how I initialize these layers inside the __init__() method of the EncoderBlock class. # Codeblock 7a class EncoderBlock(nn.Module): def __init__(self): super().__init__() #(1) self.self_attention = nn.MultiheadAttention(embed_dim=EMBED_DIM, num_heads=NUM_HEADS, batch_first=True, #(2) dropout=DROP_PROB) self.layer_norm_0 = nn.LayerNorm(EMBED_DIM) #(3) self.ffn = nn.Sequential( #(4) nn.Linear(in_features=EMBED_DIM, out_features=HIDDEN_DIM), nn.GELU(), nn.Dropout(p=DROP_PROB), nn.Linear(in_features=HIDDEN_DIM, out_features=EMBED_DIM), ) self.layer_norm_1 = nn.LayerNorm(EMBED_DIM) #(5) I’ve previously mentioned that the idea of ViT is to capture the relationships between patches within an image. This process is done by the multihead attention layer I initialize at line #(1) in the above codeblock. One thing to keep in mind here is that we need to set the batch_first parameter to True (#(2)). This is essentially done so that the attention layer will be compatible with our tensor shape, in which the batch dimension (batch_size) is at the 0th axis of the tensor. Next, the two layer normalization layers need to be initialized separately, as shown at line #(3) and #(5). Lastly, we initialize the FFN block at line #(4), which the layers stacked using nn.Sequential follows the structure defined in the following equation. Figure 8. The operations done inside the FFN block [1]. As the __init__() method is complete, we will now continue with the forward() method. Let’s take a look at the Codeblock 7b below. # Codeblock 7b def forward(self, features): #(1) residual = features #(2) print(f’features & residualt: {residual.size()}’) #(3) features, self_attn_weights = self.self_attention(query=features, key=features, value=features) print(f’after self attentiont: {features.size()}’) print(f”self attn weightst: {self_attn_weights.shape}”) features = self.layer_norm_0(features + residual) #(4) print(f’after normtt: {features.size()}’) residual = features print(f’nfeatures & residualt: {residual.size()}’) features = self.ffn(features) #(5) print(f’after ffntt: {features.size()}’) features = self.layer_norm_1(features + residual) print(f’after normtt: {features.size()}’) return features Here you can see that the input tensor is named features (#(1)). I name it this way because the input of the EncoderBlock is the image that has already been processed with Patcher and LearnableEmbedding, instead of a raw image. Before doing anything, notice in the encoder block that there is a branch separated from the main flow which then returns back to the normalization layer. This branch is commonly known as a residual connection. To implement this, we need to store the original input tensor to the residual variable as I demonstrate at line #(2). As the input tensor has been copied, now we are ready to process the original input with the multihead attention layer (#(3)). Since this is a self-attention (not a cross-attention), the query, key, and value inputs for this layer are all derived from the features tensor. Next, the layer normalization operation is then performed at line #(4), which the input for this layer already contains information from the attention block as well as the residual connection. The remaining steps are basically the same as what I just explained, except that here we replace the self-attention block with FFN (#(5)). In the following codeblock, I’ll test the EncoderBlock class by passing a dummy tensor of size 1×576×768, simulating an output tensor from the previous operations. # Codeblock 8 encoder_block = EncoderBlock() features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM) features = encoder_block(features) Below is what the tensor dimension looks like throughout the entire process inside the model. # Codeblock 8 Output features & residual : torch.Size([1, 576, 768]) #(1) after self attention : torch.Size([1, 576, 768]) self attn weights : torch.Size([1, 576, 576]) #(2) after norm : torch.Size([1, 576, 768]) features & residual : torch.Size([1, 576, 768]) after ffn : torch.Size([1, 576, 768]) #(3) after norm : torch.Size([1, 576, 768]) #(4) Here you can see that the final output tensor (#(4)) has the same size as the input (#(1)), allowing us to stack multiple encoder blocks without having to worry about messing up the tensor dimensions. Not only that, the size of the tensor also appears to be unchanged from the beginning all the way to the last layer. In fact, there are actually lots of transformations performed inside the attention block, but we just can’t see it since the entire process is done internally by the nn.MultiheadAttention layer. One of the tensors produced in the layer that we can observe is the attention weight (#(2)). This weight matrix, which has the size of 576×576, is responsible for storing information regarding the relationships between one patch and every other patch in the image. Furthermore, changes in tensor dimension actually also happened inside the FFN layer. The feature vector of each patch which has the initial length of 768 changed to 3072 and immediately shrunk back to 768 again (#(3)). However, this transformation is not printed since the process is wrapped with nn.Sequential back at line #(4) in Codeblock 7a. ViT encoder Figure 9. The entire ViT Encoder in the CPTR architecture [5]. As we have finished implementing all encoder components, now that we will assemble them to construct the actual ViT Encoder. We are going to do it in the Encoder class in Codeblock 9. # Codeblock 9 class Encoder(nn.Module): def __init__(self): super().__init__() self.patcher = Patcher() #(1) self.learnable_embedding = LearnableEmbedding() #(2) #(3) self.encoder_blocks = nn.ModuleList(EncoderBlock() for _ in range(NUM_ENCODER_BLOCKS)) def forward(self, images): #(4) print(f’imagesttt: {images.size()}’) features = self.patcher(images) #(5) print(f’after patchertt: {features.size()}’) features = features + self.learnable_embedding() #(6) print(f’after learn embedt: {features.size()}’) for i, encoder_block in enumerate(self.encoder_blocks): features = encoder_block(features) #(7) print(f”after encoder block #{i}t: {features.shape}”) return features Inside the __init__() method, what we need to do is to initialize all components we created earlier, i.e., Patcher (#(1)), LearnableEmbedding (#(2)), and EncoderBlock (#(3)). In this case, the EncoderBlock is initialized inside nn.ModuleList since we want to repeat it NUM_ENCODER_BLOCKS (12) times. To the forward() method, it initially works by accepting raw image as the input (#(4)). We then process it with the patcher layer (#(5)) to divide the image into small patches and transform them with the linear projection operation. The learnable positional embedding tensor is then injected into the resulting output by element-wise addition (#(6)). Lastly, we pass it into the 12 encoder blocks sequentially with a simple for loop (#(7)). Now, in Codeblock 10, I am going to pass a dummy image through the entire encoder. Note that since I want to focus on the flow of this Encoder class, I re-run the previous classes we created earlier with the print() functions commented out so that the outputs will look neat. # Codeblock 10 encoder = Encoder() images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE) features = encoder(images) And below is what the flow of the tensor looks like. Here, we can see that our dummy input image successfully passed through all layers in the network, including the encoder blocks that we repeat 12 times. The resulting output tensor is now context-aware, meaning that it already contains information about the relationships between patches within the image. Therefore, this tensor is now ready to be processed further with the decoder, which will later be discussed in the subsequent section. # Codeblock 10 Output images : torch.Size([1, 3, 384, 384]) after patcher : torch.Size([1, 576, 768]) after learn embed : torch.Size([1, 576, 768]) after encoder block #0 : torch.Size([1, 576, 768]) after encoder block #1 : torch.Size([1, 576, 768]) after encoder block #2 : torch.Size([1, 576, 768]) after encoder block #3 : torch.Size([1, 576, 768]) after encoder block #4 : torch.Size([1, 576, 768]) after encoder block #5 : torch.Size([1, 576, 768]) after encoder block #6 : torch.Size([1, 576, 768]) after encoder block #7 : torch.Size([1, 576, 768]) after encoder block #8 : torch.Size([1, 576, 768]) after encoder block #9 : torch.Size([1, 576, 768]) after encoder block #10 : torch.Size([1, 576, 768]) after encoder block #11 : torch.Size([1, 576, 768]) ViT encoder (alternative) I want to show you something before we talk about the decoder. If you think that our approach above is too complicated, it is actually possible for you to use nn.TransformerEncoderLayer from PyTorch so that you don’t need to implement the EncoderBlock class from scratch. To do so, I am going to reimplement the Encoder class, but this time I’ll name it EncoderTorch. # Codeblock 11 class EncoderTorch(nn.Module): def __init__(self): super().__init__() self.patcher = Patcher() self.learnable_embedding = LearnableEmbedding() #(1) encoder_block = nn.TransformerEncoderLayer(d_model=EMBED_DIM, nhead=NUM_HEADS, dim_feedforward=HIDDEN_DIM, dropout=DROP_PROB, batch_first=True) #(2) self.encoder_blocks = nn.TransformerEncoder(encoder_layer=encoder_block, num_layers=NUM_ENCODER_BLOCKS) def forward(self, images): print(f’imagesttt: {images.size()}’) features = self.patcher(images) print(f’after patchertt: {features.size()}’) features = features + self.learnable_embedding() print(f’after learn embedt: {features.size()}’) features = self.encoder_blocks(features) #(3) print(f’after encoder blockst: {features.size()}’) return features What we basically do in the above codeblock is that instead of using the EncoderBlock class, here we use nn.TransformerEncoderLayer (#(1)), which will automatically create a single encoder block based on the parameters we pass to it. To repeat it multiple times, we can just use nn.TransformerEncoder and pass a number to the num_layers parameter (#(2)). With this approach, we don’t necessarily need to write the forward pass in a loop like what we did earlier (#(3)). The testing code in the Codeblock 12 below is exactly the same as the one in Codeblock 10, except that here I use the EncoderTorch class. You can also see here that the output is basically the same as the previous one. # Codeblock 12 encoder_torch = EncoderTorch() images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE) features = encoder_torch(images) # Codeblock 12 Output images : torch.Size([1, 3, 384, 384]) after patcher : torch.Size([1, 576, 768]) after learn embed : torch.Size([1, 576, 768]) after encoder blocks : torch.Size([1, 576, 768]) Decoder As we have successfully created the encoder part of the CPTR architecture, now that we will talk about the decoder. In this section I am going to implement every single component inside the blue box in Figure 4. Based on the figure, we can see that the decoder accepts two inputs, i.e., the image caption ground truth (the lower part of the blue box) and the sequence of embedded patches produced by the encoder (the arrow coming from the green box). It is important to know that the architecture drawn in Figure 4 is intended to illustrate the training phase, where the entire caption ground truth is fed into the decoder. Later in the inference phase, we only provide a (Beginning of Sentence) token for the caption input. The decoder will then predict each word sequentially based on the given image and the previously generated words. This process is commonly known as an autoregressive mechanism. Sinusoidal positional embedding Figure 10. Where the sinusoidal positional embedding component is located in the decoder [5]. If you take a look at the CPTR model, you’ll see that the first step in the decoder is to convert each word into the corresponding feature vector representation using the word embedding block. However, since this step is very easy, we are going to implement it later. Now let’s assume that this word vectorization process is already done, so we can move to the positional embedding part. As I’ve mentioned earlier, since transformer is permutation-invariant by nature, we need to apply positional embedding to the input sequence. Different from the previous one, here we use the so-called sinusoidal positional embedding. We can think of it like a method to label each word vector by assigning numbers obtained from a sinusoidal wave. By doing so, we can expect our model to understand word orders thanks to the information given by the wave patterns. If you go back to Codeblock 6 Output, you’ll see that the positional embedding tensor in the encoder has the size of NUM_PATCHES × EMBED_DIM (576×768). What we basically want to do in the decoder is to create a tensor having the size of SEQ_LENGTH × EMBED_DIM (30×768), which the values are computed based on the equation shown in Figure 11. This tensor is then set to be non-trainable because a sequence of words must maintain a fixed order to preserve its meaning. Figure 11. The equation for creating sinusoidal positional encoding proposed in the Transformer paper [6]. Here I want to explain the following code quickly because I actually have discussed this more thoroughly in my previous article about Transformer. Generally speaking, what we basically do here is to create the sine and cosine wave using torch.sin() (#(1)) and torch.cos() (#(2)). The resulting two tensors are then merged using the code at line #(3) and #(4). # Codeblock 13 class SinusoidalEmbedding(nn.Module): def forward(self): pos = torch.arange(SEQ_LENGTH).reshape(SEQ_LENGTH, 1) print(f”postt: {pos.shape}”) i = torch.arange(0, EMBED_DIM, 2) denominator = torch.pow(10000, i/EMBED_DIM) print(f”denominatort: {denominator.shape}”) even_pos_embed = torch.sin(pos/denominator) #(1) odd_pos_embed = torch.cos(pos/denominator) #(2) print(f”even_pos_embedt: {even_pos_embed.shape}”) stacked = torch.stack([even_pos_embed, odd_pos_embed], dim=2) #(3) print(f”stackedtt: {stacked.shape}”) pos_embed = torch.flatten(stacked, start_dim=1, end_dim=2) #(4) print(f”pos_embedt: {pos_embed.shape}”) return pos_embed Now we can check if the SinusoidalEmbedding class above works properly by running the Codeblock 14 below. As expected earlier, here you can see that the resulting tensor has the size of 30×768. This dimension matches with the tensor obtained by the process done in the word embedding block, allowing them to be summed in an element-wise manner. # Codeblock 14 sinusoidal_embedding = SinusoidalEmbedding() pos_embed = sinusoidal_embedding() # Codeblock 14 Output pos : torch.Size([30, 1]) denominator : torch.Size([384]) even_pos_embed : torch.Size([30, 384]) stacked : torch.Size([30, 384, 2]) pos_embed : torch.Size([30, 768]) Look-ahead mask Figure 12. A look-ahead mask needs to be applied to the masked-self attention layer [5]. The next thing I am going to talk about in the decoder is the masked self-attention layer highlighted in the above figure. I am not going to code the attention mechanism from scratch. Rather, I’ll only implement the so-called look-ahead mask, which will be useful for the self-attention layer so that it doesn’t attend to the subsequent words in the caption during the training phase. The way to do it is pretty easy, what we need to do is just to create a triangular matrix which the size is set to match with the attention weight matrix, i.e., SEQ_LENGTH × SEQ_LENGTH (30×30). Look at the create_mask()function below for the details. # Codeblock 15 def create_mask(seq_length): mask = torch.tril(torch.ones((seq_length, seq_length))) #(1) mask[mask == 0] = -float(‘inf’) #(2) mask[mask == 1] = 0 #(3) return mask Even though creating a triangular matrix can simply be done with torch.tril() and torch.ones() (#(1)), but here we need to make a little modification by changing the 0 values to -inf (#(2)) and the 1s to 0 (#(3)). This is essentially done because the nn.MultiheadAttention layer applies the mask by element-wise addition. By assigning -inf to the subsequent words, the attention mechanism will completely ignore them. Again, the internal process inside an attention layer has also been discussed in detail in my previous article about transformer. Now I am going to run the function with seq_length=7 so that you can see what the mask actually looks like. Later in the complete flow, we need to set the seq_length parameter to SEQ_LENGTH (30) so that it matches with the actual caption length. # Codeblock 16 mask_example = create_mask(seq_length=7) mask_example # Codeblock 16 Output tensor([[0., -inf, -inf, -inf, -inf, -inf, -inf], [0., 0., -inf, -inf, -inf, -inf, -inf], [0., 0., 0., -inf, -inf, -inf, -inf], [0., 0., 0., 0., -inf, -inf, -inf], [0., 0., 0., 0., 0., -inf, -inf], [0., 0., 0., 0., 0., 0., -inf], [0., 0., 0., 0., 0., 0., 0.]]) The main decoder block Figure 13. The main decoder block [5]. We can see in the above figure that the structure of the decoder block is a bit longer than that of the encoder block. It seems like everything is nearly the same, except that the decoder part has a cross-attention mechanism and an additional layer normalization step placed after it. This cross-attention layer can actually be perceived as the bridge between the encoder and the decoder, as it is employed to capture the relationships between each word in the caption and every single patch in the input image. The two arrows coming from the encoder are the key and value inputs for the attention layer, whereas the query is derived from the previous layer in the decoder itself. Look at the Codeblock 17a and 17b below to see the implementation of the entire decoder block. # Codeblock 17a class DecoderBlock(nn.Module): def __init__(self): super().__init__() #(1) self.self_attention = nn.MultiheadAttention(embed_dim=EMBED_DIM, num_heads=NUM_HEADS, batch_first=True, dropout=DROP_PROB) #(2) self.layer_norm_0 = nn.LayerNorm(EMBED_DIM) #(3) self.cross_attention = nn.MultiheadAttention(embed_dim=EMBED_DIM, num_heads=NUM_HEADS, batch_first=True, dropout=DROP_PROB) #(4) self.layer_norm_1 = nn.LayerNorm(EMBED_DIM) #(5) self.ffn = nn.Sequential( nn.Linear(in_features=EMBED_DIM, out_features=HIDDEN_DIM), nn.GELU(), nn.Dropout(p=DROP_PROB), nn.Linear(in_features=HIDDEN_DIM, out_features=EMBED_DIM), ) #(6) self.layer_norm_2 = nn.LayerNorm(EMBED_DIM) In the __init__() method, we first initialize both self-attention (#(1)) and cross-attention (#(3)) layers with nn.MultiheadAttention. These two layers appear to be exactly the same now, but later you’ll see the difference in the forward() method. The three layer normalization operations are initialized separately as shown at line #(2), #(4) and #(6), since each of them will contain different normalization parameters. Lastly, the ffn layer (#(5)) is exactly the same as the one in the encoder, which basically follows the equation back in Figure 8. Talking about the forward() method below, it initially works by accepting three inputs: features, captions, and attn_mask, which each of them denotes the tensor coming from the encoder, the tensor from the decoder itself, and a look-ahead mask, respectively (#(1)). The remaining steps are somewhat similar to that of the EncoderBlock, except that here we repeat the multihead attention block twice. The first attention mechanism takes captions as the query, key, and value parameters (#(2)). This is essentially done because we want the layer to capture the context within the captions tensor itself — hence the name self-attention. Here we also need to pass the attn_mask parameter to this layer so that it cannot see the subsequent words during the training phase. The second attention mechanism is different (#(3)). Since we want to combine the information from the encoder and the decoder, we need to pass the captions tensor as the query, whereas the features tensor will be passed as the key and value — hence the name cross-attention. A look-ahead mask is not necessary in the cross-attention layer since later in the inference phase the model will be able to see the entire input image at once rather than looking at the patches one by one. As the tensor has been processed by the two attention layers, we will then pass it through the feed forward network (#(4)). Lastly, don’t forget to create the residual connections and apply the layer normalization steps after each sub-component. # Codeblock 17b def forward(self, features, captions, attn_mask): #(1) print(f”attn_masktt: {attn_mask.shape}”) residual = captions print(f”captions & residualt: {captions.shape}”) #(2) captions, self_attn_weights = self.self_attention(query=captions, key=captions, value=captions, attn_mask=attn_mask) print(f”after self attentiont: {captions.shape}”) print(f”self attn weightst: {self_attn_weights.shape}”) captions = self.layer_norm_0(captions + residual) print(f”after normtt: {captions.shape}”) print(f”nfeaturestt: {features.shape}”) residual = captions print(f”captions & residualt: {captions.shape}”) #(3) captions, cross_attn_weights = self.cross_attention(query=captions, key=features, value=features) print(f”after cross attentiont: {captions.shape}”) print(f”cross attn weightst: {cross_attn_weights.shape}”) captions = self.layer_norm_1(captions + residual) print(f”after normtt: {captions.shape}”) residual = captions print(f”ncaptions & residualt: {captions.shape}”) captions = self.ffn(captions) #(4) print(f”after ffntt: {captions.shape}”) captions = self.layer_norm_2(captions + residual) print(f”after normtt: {captions.shape}”) return captions As the DecoderBlock class is completed, we can now test it with the following code. # Codeblock 18 decoder_block = DecoderBlock() features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM) #(1) captions = torch.randn(BATCH_SIZE, SEQ_LENGTH, EMBED_DIM) #(2) look_ahead_mask = create_mask(seq_length=SEQ_LENGTH) #(3) captions = decoder_block(features, captions, look_ahead_mask) Here we assume that features is a tensor containing a sequence of patch embeddings produced by the encoder (#(1)), while captions is a sequence of embedded words (#(2)). The seq_length parameter of the look-ahead mask is set to SEQ_LENGTH (30) to match it to the number of words in the caption (#(3)). The tensor dimensions after each step are displayed in the following output. # Codeblock 18 Output attn_mask : torch.Size([30, 30]) captions & residual : torch.Size([1, 30, 768]) after self attention : torch.Size([1, 30, 768]) self attn weights : torch.Size([1, 30, 30]) #(1) after norm : torch.Size([1, 30, 768]) features : torch.Size([1, 576, 768]) captions & residual : torch.Size([1, 30, 768]) after cross attention : torch.Size([1, 30, 768]) cross attn weights : torch.Size([1, 30, 576]) #(2) after norm : torch.Size([1, 30, 768]) captions & residual : torch.Size([1, 30, 768]) after ffn : torch.Size([1, 30, 768]) after norm : torch.Size([1, 30, 768]) Here we can see that our DecoderBlock class works properly as it successfully processed the input tensors all the way to the last layer in the network. Here I want you to take a closer look at the attention weights at lines #(1) and #(2). Based on these two lines, we can confirm that our decoder implementation is correct since the attention weight produced by the self-attention layer has the size of 30×30 (#(1)), which basically means that this layer really captured the context within the input caption. Meanwhile, the attention weight matrix generated by the cross-attention layer has the size of 30×576 (#(2)), indicating that it successfully captured the relationships between the words and the patches. This essentially implies that after cross-attention operation is performed, the resulting captions tensor has been enriched with the information from the image. Transformer decoder Figure 14. The entire Transformer Decoder in the CPTR architecture [5]. Now that we have successfully created all components for the entire decoder, what I am going to do next is to put them together into a single class. Look at the Codeblock 19a and 19b below to see how I do that. # Codeblock 19a class Decoder(nn.Module): def __init__(self): super().__init__() #(1) self.embedding = nn.Embedding(num_embeddings=VOCAB_SIZE, embedding_dim=EMBED_DIM) #(2) self.sinusoidal_embedding = SinusoidalEmbedding() #(3) self.decoder_blocks = nn.ModuleList(DecoderBlock() for _ in range(NUM_DECODER_BLOCKS)) #(4) self.linear = nn.Linear(in_features=EMBED_DIM, out_features=VOCAB_SIZE) If you compare this Decoder class with the Encoder class from codeblock 9, you’ll notice that they are somewhat similar in terms of the structure. In the encoder, we convert image patches into vectors using Patcher, while in the decoder we convert every single word in the caption into a vector using the nn.Embedding layer (#(1)), which I haven’t explained earlier. Afterward, we initialize the positional embedding layer, where for the decoder we use the sinusoidal rather than the trainable one (#(2)). Next, we stack multiple decoder blocks using nn.ModuleList (#(3)). The linear layer written at line #(4), which doesn’t exist in the encoder, is necessary to be implemented here since it will be responsible to map each of the embedded words into a vector of length VOCAB_SIZE (10000). Later on, this vector will contain the logit of every word in the dictionary, and what we need to do afterward is just to take the index containing the highest value, i.e., the most likely word to be predicted. The flow of the tensors within the forward() method itself is also pretty similar to the one in the Encoder class. In the Codeblock 19b below we pass features, captions, and attn_mask as the input (#(1)). Keep in mind that in this case the captions tensor contains the raw word sequence, so we need to vectorize these words with the embedding layer beforehand (#(2)). Next, we inject the sinusoidal positional embedding tensor using the code at line #(3) before eventually passing it through the four decoder blocks sequentially (#(4)). Finally, we pass the resulting tensor through the last linear layer to obtain the prediction logits (#(5)). # Codeblock 19b def forward(self, features, captions, attn_mask): #(1) print(f”featurestt: {features.shape}”) print(f”captionstt: {captions.shape}”) captions = self.embedding(captions) #(2) print(f”after embeddingtt: {captions.shape}”) captions = captions + self.sinusoidal_embedding() #(3) print(f”after sin embedtt: {captions.shape}”) for i, decoder_block in enumerate(self.decoder_blocks): captions = decoder_block(features, captions, attn_mask) #(4) print(f”after decoder block #{i}t: {captions.shape}”) captions = self.linear(captions) #(5) print(f”after lineartt: {captions.shape}”) return captions At this point you might be wondering why we don’t implement the softmax activation function as drawn in the illustration. This is essentially because during the training phase, softmax is typically included within the loss function, whereas in the inference phase, the index of the largest value will remain the same regardless of whether softmax is applied. Now let’s run the following testing code to check whether there are errors in our implementation. Previously I mentioned that the captions input of the Decoder class is a raw word sequence. To simulate this, we can simply create a sequence of random integers ranging between 0 and VOCAB_SIZE (10000) with the length of SEQ_LENGTH (30) words (#(1)). # Codeblock 20 decoder = Decoder() features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM) captions = torch.randint(0, VOCAB_SIZE, (BATCH_SIZE, SEQ_LENGTH)) #(1) captions = decoder(features, captions, look_ahead_mask) And below is what the resulting output looks like. Here you can see in the last line that the linear layer produced a tensor of size 30×10000, indicating that our decoder model is now capable of predicting the logit scores for each word in the vocabulary across all 30 sequence positions. # Codeblock 20 Output features : torch.Size([1, 576, 768]) captions : torch.Size([1, 30]) after embedding : torch.Size([1, 30, 768]) after sin embed : torch.Size([1, 30, 768]) after decoder block #0 : torch.Size([1, 30, 768]) after decoder block #1 : torch.Size([1, 30, 768]) after decoder block #2 : torch.Size([1, 30, 768]) after decoder block #3 : torch.Size([1, 30, 768]) after linear : torch.Size([1, 30, 10000]) Transformer decoder (alternative) It is actually also possible to make the code simpler by replacing the DecoderBlock class with the nn.TransformerDecoderLayer, just like what we did in the ViT Encoder. Below is what the code looks like if we use this approach instead. # Codeblock 21 class DecoderTorch(nn.Module): def __init__(self): super().__init__() self.embedding = nn.Embedding(num_embeddings=VOCAB_SIZE, embedding_dim=EMBED_DIM) self.sinusoidal_embedding = SinusoidalEmbedding() #(1) decoder_block = nn.TransformerDecoderLayer(d_model=EMBED_DIM, nhead=NUM_HEADS, dim_feedforward=HIDDEN_DIM, dropout=DROP_PROB, batch_first=True) #(2) self.decoder_blocks = nn.TransformerDecoder(decoder_layer=decoder_block, num_layers=NUM_DECODER_BLOCKS) self.linear = nn.Linear(in_features=EMBED_DIM, out_features=VOCAB_SIZE) def forward(self, features, captions, tgt_mask): print(f”featurestt: {features.shape}”) print(f”captionstt: {captions.shape}”) captions = self.embedding(captions) print(f”after embeddingtt: {captions.shape}”) captions = captions + self.sinusoidal_embedding() print(f”after sin embedtt: {captions.shape}”) #(3) captions = self.decoder_blocks(tgt=captions, memory=features, tgt_mask=tgt_mask) print(f”after decoder blockst: {captions.shape}”) captions = self.linear(captions) print(f”after lineartt: {captions.shape}”) return captions The main difference you will see in the __init__() method is the use of nn.TransformerDecoderLayer and nn.TransformerDecoder at line #(1) and #(2), where the former is used to initialize a single decoder block, and the latter is for repeating the block multiple times. Next, the forward() method is mostly similar to the one in the Decoder class, except that the forward propagation on the decoder blocks is automatically repeated four times without needing to be put inside a loop (#(3)). One thing that you need to pay attention to in the decoder_blocks layer is that the tensor coming from the encoder (features) must be passed as the argument for the memory parameter. Meanwhile, the tensor from the decoder itself (captions) has to be passed as the input to the tgt parameter. The testing code for the DecoderTorch model below is basically the same as the one written in Codeblock 20. Here you can see that this model also generates the final output tensor of size 30×10000. # Codeblock 22 decoder_torch = DecoderTorch() features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM) captions = torch.randint(0, VOCAB_SIZE, (BATCH_SIZE, SEQ_LENGTH)) captions = decoder_torch(features, captions, look_ahead_mask) # Codeblock 22 Output features : torch.Size([1, 576, 768]) captions : torch.Size([1, 30]) after embedding : torch.Size([1, 30, 768]) after sin embed : torch.Size([1, 30, 768]) after decoder blocks : torch.Size([1, 30, 768]) after linear : torch.Size([1, 30, 10000]) The entire CPTR model Finally, it’s time to put the encoder and the decoder part we just created into a single class to actually construct the CPTR architecture. You can see in Codeblock 23 below that the implementation is very simple. All we need to do here is just to initialize the encoder (#(1)) and the decoder (#(2)) components, then pass the raw images and the corresponding caption ground truths as well as the look-ahead mask to the forward() method (#(3)). Additionally, it is also possible for you to replace the Encoder and the Decoder with EncoderTorch and DecoderTorch, respectively. # Codeblock 23 class EncoderDecoder(nn.Module): def __init__(self): super().__init__() self.encoder = Encoder() #EncoderTorch() #(1) self.decoder = Decoder() #DecoderTorch() #(2) def forward(self, images, captions, look_ahead_mask): #(3) print(f”imagesttt: {images.shape}”) print(f”captionstt: {captions.shape}”) features = self.encoder(images) print(f”after encodertt: {features.shape}”) captions = self.decoder(features, captions, look_ahead_mask) print(f”after decodertt: {captions.shape}”) return captions We can do the testing by passing dummy tensors through it. See the Codeblock 24 below for the details. In this case, images is basically just a tensor of random numbers having the dimension of 1×3×384×384 (#(1)), while captions is a tensor of size 1×30 containing random integers (#(2)). # Codeblock 24 encoder_decoder = EncoderDecoder() images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE) #(1) captions = torch.randint(0, VOCAB_SIZE, (BATCH_SIZE, SEQ_LENGTH)) #(2) captions = encoder_decoder(images, captions, look_ahead_mask) Below is what the output looks like. We can see here that our input images and captions successfully went through all layers in the network, which basically means that the CPTR model we created is now ready to actually be trained on image captioning datasets. # Codeblock 24 Output images : torch.Size([1, 3, 384, 384]) captions : torch.Size([1, 30]) after encoder : torch.Size([1, 576, 768]) after decoder : torch.Size([1, 30, 10000]) Ending That was pretty much everything about the theory and implementation of the CaPtion TransformeR architecture. Let me know what deep learning architecture I should implement next. Feel free to leave a comment if you spot any mistakes in this article! The code used in this article is available in my GitHub repo. Here’s the link to my previous article about image captioning, Vision Transformer (ViT), and the original Transformer. References [1] Wei Liu et al. CPTR: Full Transformer Network for Image Captioning. Arxiv. https://arxiv.org/pdf/2101.10804 [Accessed November 16, 2024]. [2] Oriol Vinyals et al. Show and Tell: A Neural Image Caption Generator. Arxiv. https://arxiv.org/pdf/1411.4555 [Accessed December 3, 2024]. [3] Image originally created by author based on: Alexey Dosovitskiy et al. An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale. Arxiv. https://arxiv.org/pdf/2010.11929 [Accessed December 3, 2024]. [4] Image originally created by author based on [6]. [5] Image originally created by author based on [1]. [6] Ashish Vaswani et al. Attention Is All You Need. Arxiv. https://arxiv.org/pdf/1706.03762 [Accessed December 3, 2024].

Introduction

In my previous article, I discussed one of the earliest Deep Learning approaches for image captioning. If you’re interested in reading it, you can find the link to that article at the end of this one.

Today, I would like to talk about Image Captioning again, but this time with the more advanced neural network architecture. The deep learning I am going to talk about is the one proposed in the paper titled “CPTR: Full Transformer Network for Image Captioning,” written by Liu et al. back in 2021 [1]. Specifically, here I will reproduce the model proposed in the paper and explain the underlying theory behind the architecture. However, keep in mind that I won’t actually demonstrate the training process since I only want to focus on the model architecture.

The idea behind CPTR

In fact, the main idea of the CPTR architecture is exactly the same as the earlier image captioning model, as both use the encoder-decoder structure. Previously, in the paper titled “Show and Tell: A Neural Image Caption Generator” [2], the models used are GoogLeNet (a.k.a. Inception V1) and LSTM for the two components, respectively. The illustration of the model proposed in the Show and Tell paper is shown in the following figure.

Figure 1. The neural network architecture for image captioning proposed in the Show and Tell paper [2].

Despite having the same encoder-decoder structure, what makes CPTR different from the previous approach is the basis of the encoder and the decoder themselves. In CPTR, we combine the encoder part of the ViT (Vision Transformer) model with the decoder part of the original Transformer model. The use of transformer-based architecture for both components is essentially where the name CPTR comes from: CaPtion TransformeR.

Note that the discussions in this article are going to be highly related to ViT and Transformer, so I highly recommend you read my previous article about these two topics if you’re not yet familiar with them. You can find the links at the end of this article.

Figure 2 shows what the original ViT architecture looks like. Everything inside the green box is the encoder part of the architecture to be adopted as the CPTR encoder.

Figure 2. The Vision Transformer (ViT) architecture [3].

Next, Figure 3 displays the original Transformer architecture. The components enclosed in the blue box are the layers that we are going to implement in the CPTR decoder.

Figure 3. The original Transformer architecture [4].

If we combine the components inside the green and blue boxes above, we are going to obtain the architecture shown in Figure 4 below. This is exactly what the CPTR model we are going to implement looks like. The idea here is that the ViT Encoder (green) works by encoding the input image into a specific tensor representation which will then be used as the basis of the Transformer Decoder (blue) to generate the corresponding caption.

Figure 4. The CPTR architecture [5].

That’s pretty much everything you need to know for now. I’ll explain more about the details as we go through the implementation.

Module imports & parameter configuration

As always, the first thing we need to do in the code is to import the required modules. In this case, we only import torch and torch.nn since we are about to implement the model from scratch.

# Codeblock 1
import torch
import torch.nn as nn

Next, we are going to initialize some parameters in Codeblock 2. If you have read my previous article about image captioning with GoogLeNet and LSTM, you’ll notice that here, we got a lot more parameters to initialize. In this article, I want to reproduce the CPTR model as closely as possible to the original one, so the parameters mentioned in the paper will be used in this implementation.

# Codeblock 2
BATCH_SIZE         = 1              #(1)

IMAGE_SIZE         = 384            #(2)
IN_CHANNELS        = 3              #(3)

SEQ_LENGTH         = 30             #(4)
VOCAB_SIZE         = 10000          #(5)

EMBED_DIM          = 768            #(6)
PATCH_SIZE         = 16             #(7)
NUM_PATCHES        = (IMAGE_SIZE//PATCH_SIZE) ** 2  #(8)
NUM_ENCODER_BLOCKS = 12             #(9)
NUM_DECODER_BLOCKS = 4              #(10)
NUM_HEADS          = 12             #(11)
HIDDEN_DIM         = EMBED_DIM * 4  #(12)
DROP_PROB          = 0.1            #(13)

The first parameter I want to explain is the BATCH_SIZE, which is written at the line marked with #(1). The number assigned to this variable is not quite important in our case since we are not actually going to train this model. This parameter is set to 1 because, by default, PyTorch treats input tensors as a batch of samples. Here I assume that we only have a single sample in a batch. 

Next, remember that in the case of image captioning we are dealing with images and texts simultaneously. This essentially means that we need to set the parameters for the two. It is mentioned in the paper that the model accepts an RGB image of size 384×384 for the encoder input. Hence, we assign the values for IMAGE_SIZE and IN_CHANNELS variables based on this information (#(2) and #(3)). On the other hand, the paper does not mention the parameters for the captions. So, here I assume that the length of the caption is no more than 30 words (#(4)), with the vocabulary size estimated at 10000 unique words (#(5)).

The remaining parameters are related to the model configuration. Here we set the EMBED_DIM variable to 768 (#(6)). In the encoder side, this number indicates the length of the feature vector that represents each 16×16 image patch (#(7)). The same concept also applies to the decoder side, but in that case the feature vector will represent a single word in the caption. Talking more specifically about the PATCH_SIZE parameter, we are going to use the value to compute the total number of patches in the input image. Since the image has the size of 384×384, there will be 576 patches in total (#(8)).

When it comes to using an encoder-decoder architecture, it is possible to specify the number of encoder and decoder blocks to be used. Using more blocks typically allows the model to perform better in terms of the accuracy, yet in return, it will require more computational power. The authors of this paper decided to stack 12 encoder blocks (#(9)) and 4 decoder blocks (#(10)). Next, since CPTR is a transformer-based model, it is necessary to specify the number of attention heads within the attention blocks inside the encoders and the decoders, which in this case authors use 12 attention heads (#(11)). The value for the HIDDEN_DIM parameter is not mentioned anywhere in the paper. However, according to the ViT and the Transformer paper, this parameter is configured to be 4 times larger than EMBED_DIM (#(12)). The dropout rate is not mentioned in the paper either. Hence, I arbitrarily set DROP_PROB to 0.1 (#(13)).

Encoder

As the modules and parameters have been set up, now that we will get into the encoder part of the network. In this section we are going to implement and explain every single component inside the green box in Figure 4 one by one.

Patch embedding

Figure 5. Dividing the input image into patches and converting them into vectors [5].

You can see in Figure 5 above that the first step to be done is dividing the input image into patches. This is essentially done because instead of focusing on local patterns like CNNs, ViT captures global context by learning the relationships between these patches. We can model this process with the Patcher class shown in the Codeblock 3 below. For the sake of simplicity, here I also include the process inside the patch embedding block within the same class.

# Codeblock 3
class Patcher(nn.Module):
   def __init__(self):
       super().__init__()

       #(1)
       self.unfold = nn.Unfold(kernel_size=PATCH_SIZE, stride=PATCH_SIZE)

       #(2)
       self.linear_projection = nn.Linear(in_features=IN_CHANNELS*PATCH_SIZE*PATCH_SIZE,
                                          out_features=EMBED_DIM)
      
   def forward(self, images):
       print(f'imagestt: {images.size()}')
       images = self.unfold(images)  #(3)
       print(f'after unfoldt: {images.size()}')
      
       images = images.permute(0, 2, 1)  #(4)
       print(f'after permutet: {images.size()}')
      
       features = self.linear_projection(images)  #(5)
       print(f'after lin projt: {features.size()}')
      
       return features

The patching itself is done using the nn.Unfold layer (#(1)). Here we need to set both the kernel_size and stride parameters to PATCH_SIZE (16) so that the resulting patches do not overlap with each other. This layer also automatically flattens these patches once it is applied to the input image. Meanwhile, the nn.Linear layer (#(2)) is employed to perform linear projection, i.e., the process done by the patch embedding block. By setting the out_features parameter to EMBED_DIM, this layer will map every single flattened patch into a feature vector of length 768.

The entire process should make more sense once you read the forward() method. You can see at line #(3) in the same codeblock that the input image is directly processed by the unfold layer. Next, we need to process the resulting tensor with the permute() method (#(4)) to swap the first and the second axis before feeding it to the linear_projection layer (#(5)). Additionally, here I also print out the tensor dimension after each layer so that you can better understand the transformation made at each step.

In order to check if our Patcher class works properly, we can just pass a dummy tensor through the network. Look at the Codeblock 4 below to see how I do it.

# Codeblock 4
patcher  = Patcher()

images   = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)
features = patcher(images)
# Codeblock 4 Output
images         : torch.Size([1, 3, 384, 384])
after unfold   : torch.Size([1, 768, 576])  #(1)
after permute  : torch.Size([1, 576, 768])  #(2)
after lin proj : torch.Size([1, 576, 768])  #(3)

The tensor I passed above represents an RGB image of size 384×384. Here we can see that after the unfold operation is performed, the tensor dimension changed to 1×768×576 (#(1)), denoting the flattened 3×16×16 patch for each of the 576 patches. Unfortunately, this output shape does not match what we need. Remember that in ViT, we perceive image patches as a sequence, so we need to swap the 1st and 2nd axes because typically, the 1st dimension of a tensor represents the temporal axis, while the 2nd one represents the feature vector of each timestep. As the permute() operation is performed, our tensor is now having the dimension of 1×576×768 (#(2)). Lastly, we pass this tensor through the linear projection layer, which the resulting tensor shape remains the same since we set the EMBED_DIM parameter to the same size (768) (#(3)). Despite having the same dimension, the information contained in the final tensor should be richer thanks to the transformation applied by the trainable weights of the linear projection layer.

Learnable positional embedding

Figure 6. Injecting the learnable positional embeddings into the embedded image patches [5].

After the input image has successfully been converted into a sequence of patches, the next thing to do is to inject the so-called positional embedding tensor. This is essentially done because a transformer without positional embedding is permutation-invariant, meaning that it treats the input sequence as if their order does not matter. Interestingly, since an image is not a literal sequence, we should set the positional embedding to be learnable such that it will be able to somewhat reorder the patch sequence that it thinks works best in representing the spatial information. However, keep in mind that the term “reordering” here does not mean that we physically rearrange the sequence. Rather, it does so by adjusting the embedding weights.

The implementation is pretty simple. All we need to do is just to initialize a tensor using nn.Parameter which the dimension is set to match with the output from the Patcher model, i.e., 576×768. Also, don’t forget to write requires_grad=True just to ensure that the tensor is trainable. Look at the Codeblock 5 below for the details.

# Codeblock 5
class LearnableEmbedding(nn.Module):
   def __init__(self):
       super().__init__()
       self.learnable_embedding = nn.Parameter(torch.randn(size=(NUM_PATCHES, EMBED_DIM)),
                                               requires_grad=True)
      
   def forward(self):
       pos_embed = self.learnable_embedding
       print(f'learnable embeddingt: {pos_embed.size()}')
      
       return pos_embed

Now let’s run the following codeblock to see whether our LearnableEmbedding class works properly. You can see in the printed output that it successfully created the positional embedding tensor as expected.

# Codeblock 6
learnable_embedding = LearnableEmbedding()

pos_embed = learnable_embedding()
# Codeblock 6 Output
learnable embedding : torch.Size([576, 768])

The main encoder block

Figure 7. The main encoder block [5].

The next thing we are going to do is to construct the main encoder block displayed in the Figure 7 above. Here you can see that this block consists of several sub-components, namely self-attention, layer norm, FFN (Feed-Forward Network), and another layer norm. The Codeblock 7a below shows how I initialize these layers inside the __init__() method of the EncoderBlock class.

# Codeblock 7a
class EncoderBlock(nn.Module):
   def __init__(self):
       super().__init__()
      
       #(1)
       self.self_attention = nn.MultiheadAttention(embed_dim=EMBED_DIM,
                                                   num_heads=NUM_HEADS,
                                                   batch_first=True,  #(2)
                                                   dropout=DROP_PROB)
      
       self.layer_norm_0 = nn.LayerNorm(EMBED_DIM)  #(3)
      
       self.ffn = nn.Sequential(  #(4)
           nn.Linear(in_features=EMBED_DIM, out_features=HIDDEN_DIM),
           nn.GELU(),
           nn.Dropout(p=DROP_PROB),
           nn.Linear(in_features=HIDDEN_DIM, out_features=EMBED_DIM),
       )
      
       self.layer_norm_1 = nn.LayerNorm(EMBED_DIM)  #(5)

I’ve previously mentioned that the idea of ViT is to capture the relationships between patches within an image. This process is done by the multihead attention layer I initialize at line #(1) in the above codeblock. One thing to keep in mind here is that we need to set the batch_first parameter to True (#(2)). This is essentially done so that the attention layer will be compatible with our tensor shape, in which the batch dimension (batch_size) is at the 0th axis of the tensor. Next, the two layer normalization layers need to be initialized separately, as shown at line #(3) and #(5). Lastly, we initialize the FFN block at line #(4), which the layers stacked using nn.Sequential follows the structure defined in the following equation.

Figure 8. The operations done inside the FFN block [1].

As the __init__() method is complete, we will now continue with the forward() method. Let’s take a look at the Codeblock 7b below.

# Codeblock 7b
   def forward(self, features):  #(1)
      
       residual = features  #(2)
       print(f'features & residualt: {residual.size()}')
      
       #(3)
       features, self_attn_weights = self.self_attention(query=features,
                                                         key=features,
                                                         value=features)
       print(f'after self attentiont: {features.size()}')
       print(f"self attn weightst: {self_attn_weights.shape}")
      
       features = self.layer_norm_0(features + residual)  #(4)
       print(f'after normtt: {features.size()}')
      

       residual = features
       print(f'nfeatures & residualt: {residual.size()}')
      
       features = self.ffn(features)  #(5)
       print(f'after ffntt: {features.size()}')
      
       features = self.layer_norm_1(features + residual)
       print(f'after normtt: {features.size()}')
      
       return features

Here you can see that the input tensor is named features (#(1)). I name it this way because the input of the EncoderBlock is the image that has already been processed with Patcher and LearnableEmbedding, instead of a raw image. Before doing anything, notice in the encoder block that there is a branch separated from the main flow which then returns back to the normalization layer. This branch is commonly known as a residual connection. To implement this, we need to store the original input tensor to the residual variable as I demonstrate at line #(2). As the input tensor has been copied, now we are ready to process the original input with the multihead attention layer (#(3)). Since this is a self-attention (not a cross-attention), the query, key, and value inputs for this layer are all derived from the features tensor. Next, the layer normalization operation is then performed at line #(4), which the input for this layer already contains information from the attention block as well as the residual connection. The remaining steps are basically the same as what I just explained, except that here we replace the self-attention block with FFN (#(5)).

In the following codeblock, I’ll test the EncoderBlock class by passing a dummy tensor of size 1×576×768, simulating an output tensor from the previous operations.

# Codeblock 8
encoder_block = EncoderBlock()

features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM)
features = encoder_block(features)

Below is what the tensor dimension looks like throughout the entire process inside the model.

# Codeblock 8 Output
features & residual  : torch.Size([1, 576, 768])  #(1)
after self attention : torch.Size([1, 576, 768])
self attn weights    : torch.Size([1, 576, 576])  #(2)
after norm           : torch.Size([1, 576, 768])

features & residual  : torch.Size([1, 576, 768])
after ffn            : torch.Size([1, 576, 768])  #(3)
after norm           : torch.Size([1, 576, 768])  #(4)

Here you can see that the final output tensor (#(4)) has the same size as the input (#(1)), allowing us to stack multiple encoder blocks without having to worry about messing up the tensor dimensions. Not only that, the size of the tensor also appears to be unchanged from the beginning all the way to the last layer. In fact, there are actually lots of transformations performed inside the attention block, but we just can’t see it since the entire process is done internally by the nn.MultiheadAttention layer. One of the tensors produced in the layer that we can observe is the attention weight (#(2)). This weight matrix, which has the size of 576×576, is responsible for storing information regarding the relationships between one patch and every other patch in the image. Furthermore, changes in tensor dimension actually also happened inside the FFN layer. The feature vector of each patch which has the initial length of 768 changed to 3072 and immediately shrunk back to 768 again (#(3)). However, this transformation is not printed since the process is wrapped with nn.Sequential back at line #(4) in Codeblock 7a.

ViT encoder

Figure 9. The entire ViT Encoder in the CPTR architecture [5].

As we have finished implementing all encoder components, now that we will assemble them to construct the actual ViT Encoder. We are going to do it in the Encoder class in Codeblock 9.

# Codeblock 9
class Encoder(nn.Module):
   def __init__(self):
       super().__init__()
       self.patcher = Patcher()  #(1)
       self.learnable_embedding = LearnableEmbedding()  #(2)

       #(3)
       self.encoder_blocks = nn.ModuleList(EncoderBlock() for _ in range(NUM_ENCODER_BLOCKS))
  
   def forward(self, images):  #(4)
       print(f'imagesttt: {images.size()}')
      
       features = self.patcher(images)  #(5)
       print(f'after patchertt: {features.size()}')
      
       features = features + self.learnable_embedding()  #(6)
       print(f'after learn embedt: {features.size()}')
      
       for i, encoder_block in enumerate(self.encoder_blocks):
           features = encoder_block(features)  #(7)
           print(f"after encoder block #{i}t: {features.shape}")

       return features

Inside the __init__() method, what we need to do is to initialize all components we created earlier, i.e., Patcher (#(1)), LearnableEmbedding (#(2)), and EncoderBlock (#(3)). In this case, the EncoderBlock is initialized inside nn.ModuleList since we want to repeat it NUM_ENCODER_BLOCKS (12) times. To the forward() method, it initially works by accepting raw image as the input (#(4)). We then process it with the patcher layer (#(5)) to divide the image into small patches and transform them with the linear projection operation. The learnable positional embedding tensor is then injected into the resulting output by element-wise addition (#(6)). Lastly, we pass it into the 12 encoder blocks sequentially with a simple for loop (#(7)).

Now, in Codeblock 10, I am going to pass a dummy image through the entire encoder. Note that since I want to focus on the flow of this Encoder class, I re-run the previous classes we created earlier with the print() functions commented out so that the outputs will look neat.

# Codeblock 10
encoder = Encoder()

images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)
features = encoder(images)

And below is what the flow of the tensor looks like. Here, we can see that our dummy input image successfully passed through all layers in the network, including the encoder blocks that we repeat 12 times. The resulting output tensor is now context-aware, meaning that it already contains information about the relationships between patches within the image. Therefore, this tensor is now ready to be processed further with the decoder, which will later be discussed in the subsequent section.

# Codeblock 10 Output
images                  : torch.Size([1, 3, 384, 384])
after patcher           : torch.Size([1, 576, 768])
after learn embed       : torch.Size([1, 576, 768])
after encoder block #0  : torch.Size([1, 576, 768])
after encoder block #1  : torch.Size([1, 576, 768])
after encoder block #2  : torch.Size([1, 576, 768])
after encoder block #3  : torch.Size([1, 576, 768])
after encoder block #4  : torch.Size([1, 576, 768])
after encoder block #5  : torch.Size([1, 576, 768])
after encoder block #6  : torch.Size([1, 576, 768])
after encoder block #7  : torch.Size([1, 576, 768])
after encoder block #8  : torch.Size([1, 576, 768])
after encoder block #9  : torch.Size([1, 576, 768])
after encoder block #10 : torch.Size([1, 576, 768])
after encoder block #11 : torch.Size([1, 576, 768])

ViT encoder (alternative)

I want to show you something before we talk about the decoder. If you think that our approach above is too complicated, it is actually possible for you to use nn.TransformerEncoderLayer from PyTorch so that you don’t need to implement the EncoderBlock class from scratch. To do so, I am going to reimplement the Encoder class, but this time I’ll name it EncoderTorch.

# Codeblock 11
class EncoderTorch(nn.Module):
   def __init__(self):
       super().__init__()
       self.patcher = Patcher()
       self.learnable_embedding = LearnableEmbedding()
      
       #(1)
       encoder_block = nn.TransformerEncoderLayer(d_model=EMBED_DIM,
                                                  nhead=NUM_HEADS,
                                                  dim_feedforward=HIDDEN_DIM,
                                                  dropout=DROP_PROB,
                                                  batch_first=True)
      
       #(2)
       self.encoder_blocks = nn.TransformerEncoder(encoder_layer=encoder_block,
                                                   num_layers=NUM_ENCODER_BLOCKS)
  
   def forward(self, images):
       print(f'imagesttt: {images.size()}')
      
       features = self.patcher(images)
       print(f'after patchertt: {features.size()}')
      
       features = features + self.learnable_embedding()
       print(f'after learn embedt: {features.size()}')
      
       features = self.encoder_blocks(features)  #(3)
       print(f'after encoder blockst: {features.size()}')

       return features

What we basically do in the above codeblock is that instead of using the EncoderBlock class, here we use nn.TransformerEncoderLayer (#(1)), which will automatically create a single encoder block based on the parameters we pass to it. To repeat it multiple times, we can just use nn.TransformerEncoder and pass a number to the num_layers parameter (#(2)). With this approach, we don’t necessarily need to write the forward pass in a loop like what we did earlier (#(3)).

The testing code in the Codeblock 12 below is exactly the same as the one in Codeblock 10, except that here I use the EncoderTorch class. You can also see here that the output is basically the same as the previous one.

# Codeblock 12
encoder_torch = EncoderTorch()

images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)
features = encoder_torch(images)
# Codeblock 12 Output
images               : torch.Size([1, 3, 384, 384])
after patcher        : torch.Size([1, 576, 768])
after learn embed    : torch.Size([1, 576, 768])
after encoder blocks : torch.Size([1, 576, 768])

Decoder

As we have successfully created the encoder part of the CPTR architecture, now that we will talk about the decoder. In this section I am going to implement every single component inside the blue box in Figure 4. Based on the figure, we can see that the decoder accepts two inputs, i.e., the image caption ground truth (the lower part of the blue box) and the sequence of embedded patches produced by the encoder (the arrow coming from the green box). It is important to know that the architecture drawn in Figure 4 is intended to illustrate the training phase, where the entire caption ground truth is fed into the decoder. Later in the inference phase, we only provide a (Beginning of Sentence) token for the caption input. The decoder will then predict each word sequentially based on the given image and the previously generated words. This process is commonly known as an autoregressive mechanism.

Sinusoidal positional embedding

Figure 10. Where the sinusoidal positional embedding component is located in the decoder [5].

If you take a look at the CPTR model, you’ll see that the first step in the decoder is to convert each word into the corresponding feature vector representation using the word embedding block. However, since this step is very easy, we are going to implement it later. Now let’s assume that this word vectorization process is already done, so we can move to the positional embedding part.

As I’ve mentioned earlier, since transformer is permutation-invariant by nature, we need to apply positional embedding to the input sequence. Different from the previous one, here we use the so-called sinusoidal positional embedding. We can think of it like a method to label each word vector by assigning numbers obtained from a sinusoidal wave. By doing so, we can expect our model to understand word orders thanks to the information given by the wave patterns.

If you go back to Codeblock 6 Output, you’ll see that the positional embedding tensor in the encoder has the size of NUM_PATCHES × EMBED_DIM (576×768). What we basically want to do in the decoder is to create a tensor having the size of SEQ_LENGTH × EMBED_DIM (30×768), which the values are computed based on the equation shown in Figure 11. This tensor is then set to be non-trainable because a sequence of words must maintain a fixed order to preserve its meaning.

Figure 11. The equation for creating sinusoidal positional encoding proposed in the Transformer paper [6].

Here I want to explain the following code quickly because I actually have discussed this more thoroughly in my previous article about Transformer. Generally speaking, what we basically do here is to create the sine and cosine wave using torch.sin() (#(1)) and torch.cos() (#(2)). The resulting two tensors are then merged using the code at line #(3) and #(4).

# Codeblock 13
class SinusoidalEmbedding(nn.Module):
   def forward(self):
       pos = torch.arange(SEQ_LENGTH).reshape(SEQ_LENGTH, 1)
       print(f"postt: {pos.shape}")
      
       i = torch.arange(0, EMBED_DIM, 2)
       denominator = torch.pow(10000, i/EMBED_DIM)
       print(f"denominatort: {denominator.shape}")
      
       even_pos_embed = torch.sin(pos/denominator)  #(1)
       odd_pos_embed  = torch.cos(pos/denominator)  #(2)
       print(f"even_pos_embedt: {even_pos_embed.shape}")
      
       stacked = torch.stack([even_pos_embed, odd_pos_embed], dim=2)  #(3)
       print(f"stackedtt: {stacked.shape}")

       pos_embed = torch.flatten(stacked, start_dim=1, end_dim=2)  #(4)
       print(f"pos_embedt: {pos_embed.shape}")
      
       return pos_embed

Now we can check if the SinusoidalEmbedding class above works properly by running the Codeblock 14 below. As expected earlier, here you can see that the resulting tensor has the size of 30×768. This dimension matches with the tensor obtained by the process done in the word embedding block, allowing them to be summed in an element-wise manner.

# Codeblock 14
sinusoidal_embedding = SinusoidalEmbedding()
pos_embed = sinusoidal_embedding()
# Codeblock 14 Output
pos            : torch.Size([30, 1])
denominator    : torch.Size([384])
even_pos_embed : torch.Size([30, 384])
stacked        : torch.Size([30, 384, 2])
pos_embed      : torch.Size([30, 768])

Look-ahead mask

Figure 12. A look-ahead mask needs to be applied to the masked-self attention layer [5].

The next thing I am going to talk about in the decoder is the masked self-attention layer highlighted in the above figure. I am not going to code the attention mechanism from scratch. Rather, I’ll only implement the so-called look-ahead mask, which will be useful for the self-attention layer so that it doesn’t attend to the subsequent words in the caption during the training phase.

The way to do it is pretty easy, what we need to do is just to create a triangular matrix which the size is set to match with the attention weight matrix, i.e., SEQ_LENGTH × SEQ_LENGTH (30×30). Look at the create_mask()function below for the details.

# Codeblock 15
def create_mask(seq_length):
   mask = torch.tril(torch.ones((seq_length, seq_length)))  #(1)
   mask[mask == 0] = -float('inf')  #(2)
   mask[mask == 1] = 0  #(3)
   return mask

Even though creating a triangular matrix can simply be done with torch.tril() and torch.ones() (#(1)), but here we need to make a little modification by changing the 0 values to -inf (#(2)) and the 1s to 0 (#(3)). This is essentially done because the nn.MultiheadAttention layer applies the mask by element-wise addition. By assigning -inf to the subsequent words, the attention mechanism will completely ignore them. Again, the internal process inside an attention layer has also been discussed in detail in my previous article about transformer.

Now I am going to run the function with seq_length=7 so that you can see what the mask actually looks like. Later in the complete flow, we need to set the seq_length parameter to SEQ_LENGTH (30) so that it matches with the actual caption length.

# Codeblock 16
mask_example = create_mask(seq_length=7)
mask_example
# Codeblock 16 Output
tensor([[0., -inf, -inf, -inf, -inf, -inf, -inf],
       [0., 0., -inf, -inf, -inf, -inf, -inf],
       [0., 0., 0., -inf, -inf, -inf, -inf],
       [0., 0., 0., 0., -inf, -inf, -inf],
       [0., 0., 0., 0., 0., -inf, -inf],
       [0., 0., 0., 0., 0., 0., -inf],
       [0., 0., 0., 0., 0., 0., 0.]])

The main decoder block

Figure 13. The main decoder block [5].

We can see in the above figure that the structure of the decoder block is a bit longer than that of the encoder block. It seems like everything is nearly the same, except that the decoder part has a cross-attention mechanism and an additional layer normalization step placed after it. This cross-attention layer can actually be perceived as the bridge between the encoder and the decoder, as it is employed to capture the relationships between each word in the caption and every single patch in the input image. The two arrows coming from the encoder are the key and value inputs for the attention layer, whereas the query is derived from the previous layer in the decoder itself. Look at the Codeblock 17a and 17b below to see the implementation of the entire decoder block.

# Codeblock 17a
class DecoderBlock(nn.Module):
   def __init__(self):
       super().__init__()
      
       #(1)
       self.self_attention = nn.MultiheadAttention(embed_dim=EMBED_DIM,
                                                   num_heads=NUM_HEADS,
                                                   batch_first=True,
                                                   dropout=DROP_PROB)
       #(2)
       self.layer_norm_0 = nn.LayerNorm(EMBED_DIM)
       #(3)
       self.cross_attention = nn.MultiheadAttention(embed_dim=EMBED_DIM,
                                                    num_heads=NUM_HEADS,
                                                    batch_first=True,
                                                    dropout=DROP_PROB)

       #(4)
       self.layer_norm_1 = nn.LayerNorm(EMBED_DIM)
      
       #(5)      
       self.ffn = nn.Sequential(
           nn.Linear(in_features=EMBED_DIM, out_features=HIDDEN_DIM),
           nn.GELU(),
           nn.Dropout(p=DROP_PROB),
           nn.Linear(in_features=HIDDEN_DIM, out_features=EMBED_DIM),
       )
      
       #(6)
       self.layer_norm_2 = nn.LayerNorm(EMBED_DIM)

In the __init__() method, we first initialize both self-attention (#(1)) and cross-attention (#(3)) layers with nn.MultiheadAttention. These two layers appear to be exactly the same now, but later you’ll see the difference in the forward() method. The three layer normalization operations are initialized separately as shown at line #(2), #(4) and #(6), since each of them will contain different normalization parameters. Lastly, the ffn layer (#(5)) is exactly the same as the one in the encoder, which basically follows the equation back in Figure 8.

Talking about the forward() method below, it initially works by accepting three inputs: features, captions, and attn_mask, which each of them denotes the tensor coming from the encoder, the tensor from the decoder itself, and a look-ahead mask, respectively (#(1)). The remaining steps are somewhat similar to that of the EncoderBlock, except that here we repeat the multihead attention block twice. The first attention mechanism takes captions as the query, key, and value parameters (#(2)). This is essentially done because we want the layer to capture the context within the captions tensor itself — hence the name self-attention. Here we also need to pass the attn_mask parameter to this layer so that it cannot see the subsequent words during the training phase. The second attention mechanism is different (#(3)). Since we want to combine the information from the encoder and the decoder, we need to pass the captions tensor as the query, whereas the features tensor will be passed as the key and value — hence the name cross-attention. A look-ahead mask is not necessary in the cross-attention layer since later in the inference phase the model will be able to see the entire input image at once rather than looking at the patches one by one. As the tensor has been processed by the two attention layers, we will then pass it through the feed forward network (#(4)). Lastly, don’t forget to create the residual connections and apply the layer normalization steps after each sub-component.

# Codeblock 17b
   def forward(self, features, captions, attn_mask):  #(1)
       print(f"attn_masktt: {attn_mask.shape}")
       residual = captions
       print(f"captions & residualt: {captions.shape}")
      
       #(2)
       captions, self_attn_weights = self.self_attention(query=captions,
                                                         key=captions,
                                                         value=captions,
                                                         attn_mask=attn_mask)
       print(f"after self attentiont: {captions.shape}")
       print(f"self attn weightst: {self_attn_weights.shape}")
      
       captions = self.layer_norm_0(captions + residual)
       print(f"after normtt: {captions.shape}")
      
      
       print(f"nfeaturestt: {features.shape}")
       residual = captions
       print(f"captions & residualt: {captions.shape}")
      
       #(3)
       captions, cross_attn_weights = self.cross_attention(query=captions,
                                                           key=features,
                                                           value=features)
       print(f"after cross attentiont: {captions.shape}")
       print(f"cross attn weightst: {cross_attn_weights.shape}")
      
       captions = self.layer_norm_1(captions + residual)
       print(f"after normtt: {captions.shape}")
      
       residual = captions
       print(f"ncaptions & residualt: {captions.shape}")
      
       captions = self.ffn(captions)  #(4)
       print(f"after ffntt: {captions.shape}")
      
       captions = self.layer_norm_2(captions + residual)
       print(f"after normtt: {captions.shape}")
      
       return captions

As the DecoderBlock class is completed, we can now test it with the following code.

# Codeblock 18
decoder_block = DecoderBlock()

features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM)  #(1)
captions = torch.randn(BATCH_SIZE, SEQ_LENGTH, EMBED_DIM)   #(2)
look_ahead_mask = create_mask(seq_length=SEQ_LENGTH)  #(3)

captions = decoder_block(features, captions, look_ahead_mask)

Here we assume that features is a tensor containing a sequence of patch embeddings produced by the encoder (#(1)), while captions is a sequence of embedded words (#(2)). The seq_length parameter of the look-ahead mask is set to SEQ_LENGTH (30) to match it to the number of words in the caption (#(3)). The tensor dimensions after each step are displayed in the following output.

# Codeblock 18 Output
attn_mask             : torch.Size([30, 30])
captions & residual   : torch.Size([1, 30, 768])
after self attention  : torch.Size([1, 30, 768])
self attn weights     : torch.Size([1, 30, 30])    #(1)
after norm            : torch.Size([1, 30, 768])

features              : torch.Size([1, 576, 768])
captions & residual   : torch.Size([1, 30, 768])
after cross attention : torch.Size([1, 30, 768])
cross attn weights    : torch.Size([1, 30, 576])   #(2)
after norm            : torch.Size([1, 30, 768])

captions & residual   : torch.Size([1, 30, 768])
after ffn             : torch.Size([1, 30, 768])
after norm            : torch.Size([1, 30, 768])

Here we can see that our DecoderBlock class works properly as it successfully processed the input tensors all the way to the last layer in the network. Here I want you to take a closer look at the attention weights at lines #(1) and #(2). Based on these two lines, we can confirm that our decoder implementation is correct since the attention weight produced by the self-attention layer has the size of 30×30 (#(1)), which basically means that this layer really captured the context within the input caption. Meanwhile, the attention weight matrix generated by the cross-attention layer has the size of 30×576 (#(2)), indicating that it successfully captured the relationships between the words and the patches. This essentially implies that after cross-attention operation is performed, the resulting captions tensor has been enriched with the information from the image.

Transformer decoder

Figure 14. The entire Transformer Decoder in the CPTR architecture [5].

Now that we have successfully created all components for the entire decoder, what I am going to do next is to put them together into a single class. Look at the Codeblock 19a and 19b below to see how I do that.

# Codeblock 19a
class Decoder(nn.Module):
   def __init__(self):
       super().__init__()

       #(1)
       self.embedding = nn.Embedding(num_embeddings=VOCAB_SIZE,
                                     embedding_dim=EMBED_DIM)

       #(2)
       self.sinusoidal_embedding = SinusoidalEmbedding()

       #(3)
       self.decoder_blocks = nn.ModuleList(DecoderBlock() for _ in range(NUM_DECODER_BLOCKS))

       #(4)
       self.linear = nn.Linear(in_features=EMBED_DIM,
                               out_features=VOCAB_SIZE)

If you compare this Decoder class with the Encoder class from codeblock 9, you’ll notice that they are somewhat similar in terms of the structure. In the encoder, we convert image patches into vectors using Patcher, while in the decoder we convert every single word in the caption into a vector using the nn.Embedding layer (#(1)), which I haven’t explained earlier. Afterward, we initialize the positional embedding layer, where for the decoder we use the sinusoidal rather than the trainable one (#(2)). Next, we stack multiple decoder blocks using nn.ModuleList (#(3)). The linear layer written at line #(4), which doesn’t exist in the encoder, is necessary to be implemented here since it will be responsible to map each of the embedded words into a vector of length VOCAB_SIZE (10000). Later on, this vector will contain the logit of every word in the dictionary, and what we need to do afterward is just to take the index containing the highest value, i.e., the most likely word to be predicted.

The flow of the tensors within the forward() method itself is also pretty similar to the one in the Encoder class. In the Codeblock 19b below we pass features, captions, and attn_mask as the input (#(1)). Keep in mind that in this case the captions tensor contains the raw word sequence, so we need to vectorize these words with the embedding layer beforehand (#(2)). Next, we inject the sinusoidal positional embedding tensor using the code at line #(3) before eventually passing it through the four decoder blocks sequentially (#(4)). Finally, we pass the resulting tensor through the last linear layer to obtain the prediction logits (#(5)).

# Codeblock 19b
   def forward(self, features, captions, attn_mask):  #(1)
       print(f"featurestt: {features.shape}")
       print(f"captionstt: {captions.shape}")
      
       captions = self.embedding(captions)  #(2)
       print(f"after embeddingtt: {captions.shape}")
      
       captions = captions + self.sinusoidal_embedding()  #(3)
       print(f"after sin embedtt: {captions.shape}")
      
       for i, decoder_block in enumerate(self.decoder_blocks):
           captions = decoder_block(features, captions, attn_mask)  #(4)
           print(f"after decoder block #{i}t: {captions.shape}")
      
       captions = self.linear(captions)  #(5)
       print(f"after lineartt: {captions.shape}")
      
       return captions

At this point you might be wondering why we don’t implement the softmax activation function as drawn in the illustration. This is essentially because during the training phase, softmax is typically included within the loss function, whereas in the inference phase, the index of the largest value will remain the same regardless of whether softmax is applied.

Now let’s run the following testing code to check whether there are errors in our implementation. Previously I mentioned that the captions input of the Decoder class is a raw word sequence. To simulate this, we can simply create a sequence of random integers ranging between 0 and VOCAB_SIZE (10000) with the length of SEQ_LENGTH (30) words (#(1)).

# Codeblock 20
decoder = Decoder()

features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM)
captions = torch.randint(0, VOCAB_SIZE, (BATCH_SIZE, SEQ_LENGTH))  #(1)

captions = decoder(features, captions, look_ahead_mask)

And below is what the resulting output looks like. Here you can see in the last line that the linear layer produced a tensor of size 30×10000, indicating that our decoder model is now capable of predicting the logit scores for each word in the vocabulary across all 30 sequence positions.

# Codeblock 20 Output
features               : torch.Size([1, 576, 768])
captions               : torch.Size([1, 30])
after embedding        : torch.Size([1, 30, 768])
after sin embed        : torch.Size([1, 30, 768])
after decoder block #0 : torch.Size([1, 30, 768])
after decoder block #1 : torch.Size([1, 30, 768])
after decoder block #2 : torch.Size([1, 30, 768])
after decoder block #3 : torch.Size([1, 30, 768])
after linear           : torch.Size([1, 30, 10000])

Transformer decoder (alternative)

It is actually also possible to make the code simpler by replacing the DecoderBlock class with the nn.TransformerDecoderLayer, just like what we did in the ViT Encoder. Below is what the code looks like if we use this approach instead.

# Codeblock 21
class DecoderTorch(nn.Module):
   def __init__(self):
       super().__init__()
       self.embedding = nn.Embedding(num_embeddings=VOCAB_SIZE,
                                     embedding_dim=EMBED_DIM)
      
       self.sinusoidal_embedding = SinusoidalEmbedding()
      
       #(1)
       decoder_block = nn.TransformerDecoderLayer(d_model=EMBED_DIM,
                                                  nhead=NUM_HEADS,
                                                  dim_feedforward=HIDDEN_DIM,
                                                  dropout=DROP_PROB,
                                                  batch_first=True)
      
       #(2)
       self.decoder_blocks = nn.TransformerDecoder(decoder_layer=decoder_block,
                                                   num_layers=NUM_DECODER_BLOCKS)
      
       self.linear = nn.Linear(in_features=EMBED_DIM,
                               out_features=VOCAB_SIZE)
      
   def forward(self, features, captions, tgt_mask):
       print(f"featurestt: {features.shape}")
       print(f"captionstt: {captions.shape}")
      
       captions = self.embedding(captions)
       print(f"after embeddingtt: {captions.shape}")
      
       captions = captions + self.sinusoidal_embedding()
       print(f"after sin embedtt: {captions.shape}")
      
       #(3)
       captions = self.decoder_blocks(tgt=captions,
                                      memory=features,
                                      tgt_mask=tgt_mask)
       print(f"after decoder blockst: {captions.shape}")
      
       captions = self.linear(captions)
       print(f"after lineartt: {captions.shape}")
      
       return captions

The main difference you will see in the __init__() method is the use of nn.TransformerDecoderLayer and nn.TransformerDecoder at line #(1) and #(2), where the former is used to initialize a single decoder block, and the latter is for repeating the block multiple times. Next, the forward() method is mostly similar to the one in the Decoder class, except that the forward propagation on the decoder blocks is automatically repeated four times without needing to be put inside a loop (#(3)). One thing that you need to pay attention to in the decoder_blocks layer is that the tensor coming from the encoder (features) must be passed as the argument for the memory parameter. Meanwhile, the tensor from the decoder itself (captions) has to be passed as the input to the tgt parameter.

The testing code for the DecoderTorch model below is basically the same as the one written in Codeblock 20. Here you can see that this model also generates the final output tensor of size 30×10000.

# Codeblock 22
decoder_torch = DecoderTorch()

features = torch.randn(BATCH_SIZE, NUM_PATCHES, EMBED_DIM)
captions = torch.randint(0, VOCAB_SIZE, (BATCH_SIZE, SEQ_LENGTH))

captions = decoder_torch(features, captions, look_ahead_mask)
# Codeblock 22 Output
features             : torch.Size([1, 576, 768])
captions             : torch.Size([1, 30])
after embedding      : torch.Size([1, 30, 768])
after sin embed      : torch.Size([1, 30, 768])
after decoder blocks : torch.Size([1, 30, 768])
after linear         : torch.Size([1, 30, 10000])

The entire CPTR model

Finally, it’s time to put the encoder and the decoder part we just created into a single class to actually construct the CPTR architecture. You can see in Codeblock 23 below that the implementation is very simple. All we need to do here is just to initialize the encoder (#(1)) and the decoder (#(2)) components, then pass the raw images and the corresponding caption ground truths as well as the look-ahead mask to the forward() method (#(3)). Additionally, it is also possible for you to replace the Encoder and the Decoder with EncoderTorch and DecoderTorch, respectively.

# Codeblock 23
class EncoderDecoder(nn.Module):
   def __init__(self):
       super().__init__()
       self.encoder = Encoder()  #EncoderTorch()  #(1)
       self.decoder = Decoder()  #DecoderTorch()  #(2)
      
   def forward(self, images, captions, look_ahead_mask):  #(3)
       print(f"imagesttt: {images.shape}")
       print(f"captionstt: {captions.shape}")
      
       features = self.encoder(images)
       print(f"after encodertt: {features.shape}")
      
       captions = self.decoder(features, captions, look_ahead_mask)
       print(f"after decodertt: {captions.shape}")
      
       return captions

We can do the testing by passing dummy tensors through it. See the Codeblock 24 below for the details. In this case, images is basically just a tensor of random numbers having the dimension of 1×3×384×384 (#(1)), while captions is a tensor of size 1×30 containing random integers (#(2)).

# Codeblock 24
encoder_decoder = EncoderDecoder()

images = torch.randn(BATCH_SIZE, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)  #(1)
captions = torch.randint(0, VOCAB_SIZE, (BATCH_SIZE, SEQ_LENGTH))  #(2)

captions = encoder_decoder(images, captions, look_ahead_mask)

Below is what the output looks like. We can see here that our input images and captions successfully went through all layers in the network, which basically means that the CPTR model we created is now ready to actually be trained on image captioning datasets.

# Codeblock 24 Output
images         : torch.Size([1, 3, 384, 384])
captions       : torch.Size([1, 30])
after encoder  : torch.Size([1, 576, 768])
after decoder  : torch.Size([1, 30, 10000])

Ending

That was pretty much everything about the theory and implementation of the CaPtion TransformeR architecture. Let me know what deep learning architecture I should implement next. Feel free to leave a comment if you spot any mistakes in this article!

The code used in this article is available in my GitHub repo. Here’s the link to my previous article about image captioning, Vision Transformer (ViT), and the original Transformer.

References

[1] Wei Liu et al. CPTR: Full Transformer Network for Image Captioning. Arxiv. https://arxiv.org/pdf/2101.10804 [Accessed November 16, 2024].

[2] Oriol Vinyals et al. Show and Tell: A Neural Image Caption Generator. Arxiv. https://arxiv.org/pdf/1411.4555 [Accessed December 3, 2024].

[3] Image originally created by author based on: Alexey Dosovitskiy et al. An Image is Worth 16×16 Words: Transformers for Image Recognition at Scale. Arxiv. https://arxiv.org/pdf/2010.11929 [Accessed December 3, 2024].

[4] Image originally created by author based on [6].

[5] Image originally created by author based on [1].

[6] Ashish Vaswani et al. Attention Is All You Need. Arxiv. https://arxiv.org/pdf/1706.03762 [Accessed December 3, 2024].

Shape
Shape
Stay Ahead

Explore More Insights

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy,  bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

Shape

AMD acquires MEXT to add predictive memory optimization to its AI stack

Gogia said predictive memory tiering addresses inefficiencies that often leave expensive DRAM underutilized, but cautioned that optimization should complement, rather than replace, sound infrastructure design. “Predictive tiering attacks the waste inside that reflex,” he said, referring to the tendency to address performance challenges by purchasing more memory instead of improving

Read More »

HPE product barrage targets AI networks, agents, management

HPE Mist integration Continuing that integration theme, HPE said it will integrate Juniper’s natural language Mist AI into HPE Aruba Central and vice versa, all fed by its core AIOps Marvis AI engine. Marvis collects telemetry and user state data from Juniper’s routers, switches, access points, firewalls, and applications to detect and

Read More »

MRV lets EPCIC contract for Coral North FLNG project

Mozambique Rovuma Venture (MRV) SpA  has let an engineering, procurement, construction, installation, and commissioning (EPCIC) contract to Technip Energies, in partnership with JGC and Samsung Heavy Industries, for the Coral North FLNG project offshore Mozambique. Under the contract, JGC France and Technip Energies, through their joint venture, will be primarily responsible for the engineering and procurement of the FLNG topside infrastructure as well as overall project management. Samsung Heavy Industries will undertake the engineering, procurement, and construction of the FLNG hull and the fabrication of the topside modules. The Coral North project includes construction of a new FLNG vessel with a production capacity of about 3.6 million tonnes/year (tpy). The vessel will be installed in Coral gas field, about 50 km offshore northern Mozambique. Coral North is designed as an enhanced replica of Coral Sul, the first development in Mozambique’s Area 4 offshore gas block, and is expected to double Coral hub’s capacity to 7 million tpy.  Coral FLNG SA is a special-purpose entity incorporated by Area 4 partners Eni SPA (operator), China National Petroleum Corp. (CNPC), Empresa Nacional de Hidrocarbonetos (ENH), Galp Energia SGPS SA, and Korea Gas Corp. (KOGAS).

Read More »

EIA forecasts prolonged oil market tightening amid Hormuz shipping disruptions

In its June 2026 Short-Term Energy Outlook (STEO), the US Energy Information Administration (EIA) assumes the Strait of Hormuz will remain effectively closed in the near term, with oil shipments resuming in third–quarter 2026. The agency expects it will take until early 2027 for traffic through the waterway to return to pre-conflict levels. Some Middle East oil production is expected to remain disrupted beyond the forecast period. Global oil producers in the Middle East reduced crude oil production by more than 11 million b/d in May compared with pre-conflict levels because of limited shipping traffic through the strait. EIA estimates production shut-ins averaged 11.3 million b/d in May and forecasts disruptions of 11.34 million b/d in June before easing to 10.11 million b/d in the third quarter and 5.70 million b/d in the fourth quarter. Stay updated on oil price volatility, shipping disruptions, LNG market analysis, and production output at OGJ’s Iran war content hub. As a result, EIA expects global oil inventories to fall by an average of 6.3 million b/d in second-quarter 2026 and 7.6 million b/d in third-quarter 2026. OECD commercial inventories are forecast to fall to just under 2.3 billion bbl by December 2026, the lowest level since 2003. On a days-of-supply basis, OECD inventories are expected to fall to 50 days by yearend 2026. Brent crude oil averaged $107/bbl in May, down $10/bbl from April. EIA forecasts Brent prices will average about $105/bbl in June and July before declining to an average of $89/bbl in fourth-quarter 2026 as oil flows gradually resume. Brent is forecast to average $95/bbl in 2026 and $79/bbl in 2027. High fuel prices, reduced fuel availability, and government initiatives have lowered oil demand, EIA said. The agency now forecasts global oil demand will decline by 1.1 million b/d in 2026 compared

Read More »

EnQuest looks to boost production by over 130% with Malaysian asset deals

EnQuest PLC subsidiary EnQuest Petroleum Production Malaysia Ltd. has agreed to acquire interests in producing upstream assets in Peninsular Malaysia and Sarawak. The company expects to leverage its integrated technical capabilities and experience in managing brownfield and late-life assets to support continued operations and redevelopment. Under three separate transaction packages with Petronas Carigali Sdn. Bhd. and E&P Malaysia Venture Sdn. Bhd., EnQuest will acquire interests in four offshore production sharing contracts (PSCs) for a maximum total consideration of up to $833 million. As part of the agreements, EnQuest Petroleum Production Malaysia will assume operatorship and participating interests in the Balingian PSC (Package 1, 90% participating interest), SK8 PSC (Package 1, 100% interest), and D35 PSC (Package 2, 50% interest), and will hold a nonoperated participating interest in the PM6/12 PSC (Package 3, 30% interest). The transaction also includes participation by Terengganu-based TI Exploration & Production Sdn. Bhd. (TI EP), which will hold a nonoperated interest in the PM6/12 PSC. TI EP is a joint venture between TI Petroleum Sdn. Bhd., a subsidiary of state-owned Terengganu Inc., and Ping Petroleum Ltd., an independent upstream company. On a 2025 net participating interest basis, the acquired interests are expected to add about 57,400 boe/d of production (47% liquids, 53% gas). This would increase EnQuest’s group production to more than 100,000 boe/d, representing a 134% increase compared with its 2025 production. The assets are expected to support production at around 100,000 boe/d through the end of the decade, the company said. EnQuest would also add 138 MMboe of 2P reserves and 208 MMboe of 2C resources (net WI). The acquisitions are expected to close by yearend, subject to customary conditions, including the waiver or expiry of applicable pre-emption rights associated with Package 2.  Package 1, Package 2, and Package 3 are subject to separate acquisition

Read More »

EIA: US crude inventories down 7.2 million bbl

US crude oil inventories for the week ended June 5, excluding the Strategic Petroleum Reserve, decreased by 7.2 million bbl from the previous week, according to data from the US Energy Information Administration (EIA). At 426.5 million bbl, US crude oil inventories are about 5% below the 5-year average for this time of year, the EIA report indicated. EIA said total motor gasoline inventories increased by 200,000 bbl from last week and are about 6% below the 5-year average for this time of year. Finished gasoline inventories increased while blending components inventories decreased last week. Distillate fuel inventories decreased by 200,000 bbl last week and are about 13% below the 5-year average for this time of year. Propane-propylene inventories increased by 1.1 million bbl from last week and are about 35% above the 5-year average for this time of year, EIA said. US crude oil refinery inputs averaged 17.0 million b/d for the week ended June 5, which was 80,000 b/d more than the previous week’s average. Refineries operated at 95.3% of capacity. Gasoline production increased, averaging 9.7 million b/d. Distillate fuel production increased, averaging 5.2 million b/d. US crude oil imports averaged 5.9 million b/d, down 500,000 b/d from the previous week. Over the last 4 weeks, crude oil imports averaged about 5.9 million b/d, 5.8% less than the same 4-week period last year. Total motor gasoline imports averaged 714,000 b/d. Distillate fuel imports averaged 130,000 b/d.

Read More »

Shell discovers oil in Namibia’s Orange basin

Shell discovered oil at the Merlin-1X exploration well in Orange basin 250 km off the southern coast of Namibia. Merlin-1X, spudded on Apr. 8, 2026, is the tenth well drilled in Petroleum Exploration License 39 (PEL 0039). The well penetrated the Coniacian play and has delivered the most promising subsurface results to date in PEL 0039, indicating good reservoir quality with light oil and limited associated gas, the operator said in a release June 9. Shell said additional drilling late this year is under consideration as part of a broader exploratory appraisal program. PEL 0039 covers 12,000 sq km. Over the last 4 years, 10 wells have been drilled in the license: Graff-1X, La Rona-1X, Jonker-1X, Graff-1A, Lesedi-1X, Cullinan-1X, Jonker-1A, Jonker-2A, Enigma-1X, and Merlin-1X. Shell is operator of PEL 0039 with 45% working interest. Partners are QatarEnergy (45%) and the National Petroleum Corp. of Namibia (NAMCOR) (10%).

Read More »

Equinor to farm out 50% stake in Itaimbezinho block offshore Brazil

@import url(‘https://fonts.googleapis.com/css2?family=Inter:[email protected]&display=swap’); .ebm-page__main h1, .ebm-page__main h2, .ebm-page__main h3, .ebm-page__main h4, .ebm-page__main h5, .ebm-page__main h6 { font-family: Inter; } body { line-height: 150%; letter-spacing: 0.025em; } button, .ebm-button-wrapper { font-family: Inter; } .label-style { text-transform: uppercase; color: var(–color-grey); font-weight: 600; font-size: 0.75rem; } .caption-style { font-size: 0.75rem; opacity: .6; } #onetrust-pc-sdk [id*=btn-handler], #onetrust-pc-sdk [class*=btn-handler] { background-color: #c19a06 !important; border-color: #c19a06 !important; } #onetrust-policy a, #onetrust-pc-sdk a, #ot-pc-content a { color: #c19a06 !important; } #onetrust-consent-sdk #onetrust-pc-sdk .ot-active-menu { border-color: #c19a06 !important; } #onetrust-consent-sdk #onetrust-accept-btn-handler, #onetrust-banner-sdk #onetrust-reject-all-handler, #onetrust-consent-sdk #onetrust-pc-btn-handler.cookie-setting-link { background-color: #c19a06 !important; border-color: #c19a06 !important; } #onetrust-consent-sdk .onetrust-pc-btn-handler { color: #c19a06 !important; border-color: #c19a06 !important; } Equinor Brasil Energia Ltda. has agreed to farm out a 50% interest in the Itaimbezinho exploration block in Brazil’s offshore Campos basin to Petróleo Brasileiro SA (Petrobras). Equinor was awarded a 100% stake in the block during Brazil’s National Agency of Petroleum, Natural Gas and Biofuels (ANP) 3rd Cycle pre-salt bidding round in 2025. Upon completion of the deal with Petrobras, Equinor will retain a 50% interest and remain operator, while Pré-Sal Petróleo SA (PPSA) will continue as manager of the production-sharing contract. The agreement builds on Equinor and Petrobras’ broader partnership in the basin. The companies jointly acquired the Jaspe exploration block in the same bidding round, with Petrobras as operator (60%) and Equinor holding 40%. The companies also partner on the Raia natural gas project, where Equinor, as operator, began a 6-well drilling campaign in March.   Completion of the transaction remains subject to customary regulatory and governmental approvals.

Read More »

MSI’s Strategic Shift: From Server Vendor to Full-Spectrum AI Infrastructure Provider

The 100 kW rack figure places MSI’s offering squarely in the world of AI-era rack densities, where conventional air cooling becomes increasingly difficult or inefficient. The announcement also suggests that MSI is aligning with hyperscale and large cloud design principles, particularly through ORv3 and 48V power distribution. The company is moving from the “we have servers that can be liquid cooled” message, to “we can participate in rack-level AI infrastructure design.” The EIA air-cooled architecture, by contrast, is designed for more conventional data center environments. MSI says its 19-inch, 48RU EIA air-cooled rack supports standard deployments and can be configured with 16 2U2N multi-node systems, with AMD EPYC 9005 and Intel Xeon 6 platform options. That split matters because the AI infrastructure market is not moving in one uniform direction. Hyperscalers, neoclouds, and AI factories may move aggressively into ORv3, liquid cooling, busbar power, and rack-scale designs. Enterprise data centers, managed service providers, and colocation customers often need to work within existing 19-inch rack footprints and existing facility constraints. MSI wants to supply both markets. The CG681-S6093: MSI’s Flagship Liquid-Cooled AI Server The centerpiece of MSI’s NVIDIA-based AI server announcement is the CG681-S6093, a 6U liquid-cooled AI server based on NVIDIA MGX architecture. MSI says the system supports dual AMD EPYC processors and up to eight NVIDIA RTX PRO 6000 Blackwell Server Edition Liquid Cooled GPUs. It also supports 32 DDR5 DIMMs and NVIDIA ConnectX-8 SuperNICs with up to 8×400Gbps networking. This system is a direct entry into high-density AI inference, HPC, simulation, graphics, video, and physical AI workloads. The server is not positioned only for frontier model training. Instead, MSI appears to be aiming at the expanding middle of the AI infrastructure market: large inference clusters, visual computing, simulation, industrial AI, scientific computing, and agentic AI workloads. The next

Read More »

Cooling at AI Scale: Inside Motivair’s Blueprint for the Liquid-Cooled Data Center

BUFFALO, N.Y. — In the race to build AI infrastructure, the industry often focuses on GPUs, power availability, and the massive capital investments reshaping the digital infrastructure landscape. But a walk through Motivair’s manufacturing facility in Buffalo, as provided on the eve of the Motivair-Schneider Electric Global Press Event’s tour of the nearby Terawulf Lake Mariner AI campus, offers a reminder that another critical component of the AI boom is being built one coolant distribution unit at a time. During a recent Data Center Frontier Show podcast recorded at Motivair’s Buffalo headquarters, CEO Rich Whitmore described a reality that is becoming bedrock across the industry: Liquid cooling is now very far from being an emerging technology. It is now a prerequisite for deploying the most advanced AI systems. “You cannot deploy AI servers—at least the cutting-edge AI servers—without liquid cooling,” Whitmore said. That observation may be obvious to infrastructure veterans. Yet it points to a larger shift now underway across the data center ecosystem. As AI workloads drive rack densities beyond the practical limits of air cooling, thermal infrastructure has moved from a supporting role to a primary design consideration. For Whitmore and Motivair, that transition did not begin with ChatGPT. From Supercomputing to Commercial AI Long before AI became the defining growth story of the data center sector, Motivair was developing liquid cooling systems for high-performance computing and supercomputing environments. Whitmore describes today’s AI market as less of a technological revolution than a commercialization of capabilities that have existed for years inside elite computing environments. “We cut our teeth in high-performance computing and supercomputing,” Whitmore explained. “What we’re seeing today as we go into the AI era is really a commercialization of traditional supercomputing.” That experience has positioned Motivair differently than many newer entrants rushing into the liquid cooling

Read More »

From Components to AI Factories: Peter Panfil Says the Future of Data Centers Is All About Integration at Scale

ORLANDO, Fla. — For years, the data center industry optimized individual systems: power distribution, cooling, racks, UPS equipment, and mechanical infrastructure. In the AI era, according to Vertiv Distinguished Engineer and Vice President of Technical Business Development Peter Panfil, that approach is no longer sufficient. Speaking during Wednesday morning’s keynote at the 2026 7×24 Exchange Spring Conference, Panfil presented a vision in which the data center itself becomes a single, tightly orchestrated computing appliance—truly an “AI factory” whose success depends less on standalone components than on the seamless interaction between them. Throughout his presentation, titled “Scale at Speed: How Massively Parallel Compute GPUs Are Revolutionizing Data Center Design,” Panfil repeatedly returned to a single imperative: the AI infrastructure race is increasingly defined by execution velocity. “If you think you’re going big enough, go bigger,” he told attendees. “If you think you’re going fast enough, you’re going to have to go faster.” For an industry gathered under the conference’s overarching theme of future-proofing AI infrastructure, Panfil’s message suggested something subtly different. Rather than trying to predict the future, operators should build systems capable of adapting to it. “I would much rather be future ready,” he said, “than future proof.” Speed Becomes the New Competitive Metric One of the keynote’s recurring themes was that deployment speed has become an economic variable in its own right. Panfil argued that hyperscalers and AI providers increasingly view time-to-capacity as directly tied to business value, making delays in construction or commissioning far more expensive than traditional infrastructure inefficiencies. “The cost of speeding up has real benefits right now,” he observed. That urgency is changing the way facilities are assembled. Rather than coordinating numerous independent contractors and subsystem vendors on-site, Panfil described an emerging model built around highly standardized, factory-produced HAC [hot aisle containment] modules—or “hacks”—that arrive

Read More »

Beyond the GPU: Cisco Says AI’s Biggest Challenge May Be the Network That Connects It All

For much of the AI boom, the industry’s attention has centered on GPUs, power availability, and liquid cooling. But according to Cisco Senior Business Development Manager Robin Olds, another critical constraint is rapidly moving to the forefront: the network itself. Speaking with Data Center Frontier on the show floor at Fiber Connect 2026, Olds argued that AI represents a once-in-a-generation shift comparable to the birth of the commercial internet, fundamentally changing traffic patterns and forcing service providers, data center operators, hyperscalers, and emerging neoclouds to rethink infrastructure design. “It’s really like the internet when it was created,” Olds said. “We’re at another intersection in time where we could really see things happening.” AI Is Rewriting the Bandwidth Equation The most significant change may not be compute density alone but the sustained demand AI places on transport networks. According to Olds, service providers are already seeing AI traffic account for roughly 30% of utilization on backbone infrastructure; a dramatic increase from less than 1% only two years ago. As AI workloads continue to proliferate, those utilization levels are expected to rise further. The next wave of agentic AI could amplify that trend. Unlike consumer chatbots, which generate bursty request patterns, autonomous AI agents continuously interact with applications and external services, creating more persistent traffic flows. “Everything’s about chatbots,” Olds observed. “It’s very spiky—up, down. Agentic AI is going to maintain utilization because now I have agents working on my behalf.” For data center developers, network operators, and cloud providers alike, that implies planning not just for peak demand but for elevated baseline utilization across metro and long-haul infrastructure. Compressing the Network Stack Cisco’s response centers on architectural simplification. Olds highlighted the company’s Agile Services Networking framework, which combines router and optical networking technologies with coherent optics to converge functions that historically

Read More »

Emerging Power Strategies Transforming AI Data Center Development

This deal follows the industry trend shown in the Flex/EP² acquisition. One transaction targets engineered control and protection systems; the other targets high-power conductive components and value-added services around them. The clarity here is that data center growth is creating investable demand not just for chips and campuses, but for the industrial companies that make the electrical buildout possible. AZIO AI and EVTV: EV maker to AI infrastructure After multiple announcements of deals between the two companies, and the end of May, they decided a corporate merger was in order. The AZIO AI–EVTV merger turns Envirotech Vehicles Inc., from an electric-vehicle company into a prospective power-backed AI infrastructure and data center platform. The core impact is not just an increase in AI servers; it is the combination of compute hardware, controlled land, behind-the-meter power, fiber, modular deployment, and cooling validation building on the AI data center growth story. AZIO describes the deal as part of EVTV’s strategic transformation into an AI infrastructure and compute platform focused on domestic AI deployment, data center operations, and long-term compute capacity expansion. For data center ambitions, the most important element is power access. AZIO and EVTV say about 11 MW of power capacity has been identified at EVTV’s existing site, with hardware orders placed for an initial 6 MW deployment. They are also discussing long-term rights tied to as much as 500 MW of additional same-site capacity. That matters because AI data center development is increasingly constrained by power availability rather than just real estate or server supply. And with the modular infrastructure the company develops, having that power availability is the key to future site growth. EVTV later said it had expanded its controlled development footprint to more than 548 acres, secured dedicated fiber for current and future operations, and was advancing natural

Read More »

Amazon claims its data centers are 7x more water-efficient than the industry average

“Amazon is on the leading edge, but it’s not a secret recipe,” he said. What sets the company apart is scale, execution, facility design, geographic mix, and its aggressive pursuit of energy goals. Others are doing the similar things, if through different avenues: Microsoft is investing in closed-loop cooling systems that dramatically reduce evaporative water loss. Google is heavily focused on reclaimed water and using AI to optimize data centers. Meta has long relied on outside-air cooling. And overall, the industry is moving toward liquid cooling for dense AI deployments, “which changes the water equation again,” said Kimball. One of the big variables is location: Climate influences water efficiency, so where a company builds its infrastructure is as important as its cooling methods. Further, power-consumptive AI changes the discussion, he emphasized; traditional enterprise workloads and dense AI training clusters create very different thermal profiles.

Read More »

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs).  In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

Read More »

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

Read More »

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

Read More »

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle

Read More »