The invention of ViT (Imaginative and prescient Transformer) causes us to suppose that CNNs are out of date. However is that this actually true?
It’s broadly believed that the spectacular efficiency of ViT comes primarily from its transformer-based structure. Nonetheless, researchers from Meta argued that it’s not completely true. If we take a better take a look at the architectural design, ViT launched radical adjustments not solely to the construction of the community but additionally to the mannequin configurations. Meta’s researchers thought that maybe it isn’t the construction that makes ViT superior, however its configuration. So as to show this, they tried to use the ViT configuration parameters to the ResNet structure from 2015.
— They usually discovered their thesis true.
On this article I’m going to speak about ConvNeXt which was first proposed within the paper titled “A ConvNet for the 2020s” written by Liu et al. [1] again in 2022. Right here I’ll additionally attempt to implement it myself from scratch with PyTorch as a way to get higher understanding of the adjustments comprised of the unique ResNet. In actual fact, the precise ConvNeXt implementation is out there of their GitHub repository [2], however I discover it too complicated to clarify line by line. Thus, I made a decision to write down it down by myself in order that I can clarify it with my model, which I imagine is extra beginner-friendly. Disclaimer on, my implementation may not completely replicate the unique one, however I feel it’s nonetheless good to think about my code as a useful resource to study. So, after studying my article I like to recommend you test the unique code particularly if you happen to’re planning to make use of ConvNeXt on your venture.
The Hyperparameter Tuning
What the authors basically did within the analysis was hyperparameter tuning on the ResNet mannequin. Usually talking, there have been 5 features they experimented with: macro design, ResNeXt, inverted bottleneck, massive kernel, and micro design. We will see the experimental outcomes on these features within the following determine.
There have been two ResNet variants used of their experiments: ResNet-50 and ResNet-200 (proven in purple and grey, respectively). Let’s now deal with the outcomes obtained from tuning the ResNet-50 structure. Based mostly on the determine, we will see that this mannequin initially obtained 78.8% accuracy on ImageNet dataset. They tuned this mannequin till ultimately it reached 82.0%, surpassing the state-of-the-art Swin-T structure which solely achieved 81.3% (the orange bar). This tuned model of the ResNet mannequin is the one so-called ConvNeXt proposed within the paper. Their experiments on ResNet-200 affirm that the earlier outcomes are legitimate since its tuned model, i.e., ConvNeXt-B, additionally efficiently surpasses the efficiency of Swin-B (the bigger variant of Swin-T).
Macro Design
The primary change made on the unique ResNet was the macro design. If we take a better take a look at Determine 2 beneath, we will see {that a} ResNet mannequin basically consists of 4 essential levels, specifically conv2_x, conv3_x, conv4_x and conv5_x, which every of them additionally contains a number of bottleneck blocks. Speaking extra particularly about ResNet-50, the bottleneck blocks in every stage is repeated 3, 4, 6 and three occasions, respectively. Afterward, I’ll refer to those numbers as stage ratio.
The authors of the ConvNeXt paper tried to alter this stage ratio based on the Swin-T structure, i.e., 1:1:3:1. Properly, it’s truly 2:2:6:2 if you happen to see the architectural particulars from the unique Swin Transformer paper in Determine 3, however it’s principally only a derivation from the identical ratio. By making use of this configuration, authors obtained 0.6% enchancment (from 78.8% to 79.4%). Thus, they determined to make use of 1:1:3:1 stage ratio for the upcoming experiments.
Nonetheless associated to macro design, adjustments have been additionally made to the primary convolution layer of ResNet. In the event you return to Determine 2 (the conv1 row), you’ll see that it initially makes use of 7×7 kernel with stride 2, which reduces the picture dimension from 224×224 to 112×112. Being impressed by Swin Transformer, authors additionally wished to deal with the enter picture as non-overlapping patches. Thus, they modified the kernel dimension to 4×4 and the stride to 4. This concept was truly adopted from the unique ViT, the place it makes use of 16×16 kernel with stride 16. One factor you must know in ConvNeXt is that the ensuing patches are handled as a typical picture slightly than a sequence. With this modification, the accuracy barely improved from 79.4% to 79.5%. Therefore, authors used this configuration for the primary convolution layer within the subsequent experiments.
ResNeXt-ification
Because the macro design is finished, the following factor authors did was to undertake the ResNeXt structure, which was first proposed in a paper titled “Aggregated Residual Transformations for Deep Neural Networks” [5]. The thought of ResNeXt is that it principally applies group convolution to the bottleneck blocks of the ResNet structure. In case you’re not but acquainted with group convolution, it basically works by separating enter channels into teams and performing convolution operations inside every group independently, permitting quicker computation because the variety of teams will increase. ConvNeXt adopts this concept by setting the variety of teams to be the identical because the variety of kernels. This method, which is often referred to as depthwise convolution, permits the community to acquire the bottom doable computational complexity. Nonetheless, it is very important word that rising the variety of convolution teams like this results in a discount in accuracy because it lowers the mannequin capability to study. Thus, the drop in accuracy to 78.3% was anticipated.
That wasn’t the top of the ResNeXt-ification part, although. In actual fact, the ResNeXt paper offers us a steering that if we enhance the variety of teams, we additionally have to increase the width of the community, i.e., add extra channels. Thus, ConvNeXt authors readjusted the variety of kernels based mostly on the one utilized in Swin-T. You possibly can see in Determine 2 and three that ResNet initially makes use of 64, 128, 256 and 512 kernels in every stage, whereas Swin-T makes use of 96, 192, 384, and 768. Such a rise within the mannequin width permits the community to considerably push the accuracy to 80.5%.
Inverted Bottleneck
Nonetheless with Determine 2, additionally it is seen that ResNet-50, ResNet-101, and ResNet-152 share the very same bottleneck construction. For example, the block at stage conv5_x consists of three convolution layers with 512, 512, and 2048 kernels, the place the enter of the primary convolution is both 1024 (coming from the conv4_x stage) or 2048 (from the earlier block within the conv5_x stage itself). These ResNet variations basically observe the broad → slender → broad construction, which is the explanation that this block is named bottleneck. As an alternative of utilizing a construction like this, ConvNeXt employs the inverted model of bottleneck, the place it follows the slender → broad → slender construction adopted from the feed-forward layer of the Transformer structure. In Determine 4 beneath (a) is the bottleneck block utilized in ResNet and (b) is the so-called inverted bottleneck block. Through the use of this construction, the mannequin accuracy elevated from 80.5% to 80.6%.
Kernel Measurement
The following exploration was performed on the kernel dimension contained in the inverted bottleneck block. Earlier than experimenting with completely different kernel sizes, additional modification was performed to the construction of the block, the place authors swapped the order of the primary and second layer such that the depthwise convolution is now positioned at first of the block as seen in Determine 4 (c). Due to this modification, the block is now referred to as ConvNeXt block because it not fully resembles the unique inverted bottleneck construction. This concept was truly adopted from Transformer, the place the MSA (Multihead Self-Consideration) layer is positioned earlier than the MLP layers. Within the case of ConvNeXt, the depthwise convolution acts because the substitute of MSA, whereas the linear layers in MLP Transformers are changed by pointwise convolutions. Merely shifting up the depthwise convolution like this lowered the accuracy from 80.6% to 79.9%. Nonetheless, that is acceptable as a result of the present experiment set remains to be ongoing.
Experiments on the kernel dimension was then utilized solely on the depthwise convolution layer, leaving the remaining pointwise convolutions unchanged. Right here authors tried to make use of completely different kernel sizes, the place they discovered that 7×7 labored finest because it efficiently recovered the accuracy again to 80.6% with decrease computational complexity (4.6 vs 4.2 GFLOPS). Curiously, this kernel dimension matches the window dimensions within the Swin Transformer structure, which corresponds to the patch dimension used within the self-attention mechanism. You possibly can truly see this in Determine 3 the place the window sizes in Swin Transformer variants are all 7×7.
Micro Design
The ultimate side tuned within the paper is the so-called micro design, which basically refers back to the issues associated to the intricate particulars of the community. Just like the earlier ones, the parameters used listed below are primarily additionally adopted from Transformers. Authors initially changed ReLU with GELU. Regardless that with this substitute the accuracy remained the identical (80.6%), however they determined to go together with this activation operate for the following experiments. The accuracy lastly elevated after the variety of activation capabilities was lowered. As an alternative of making use of GELU after every convolution layer within the ConvNeXt block, this activation operate was positioned solely between the 2 pointwise convolutions. This modification allowed the community to spice up the accuracy as much as 81.3%, at which level this rating was already on par with the Swin-T structure whereas nonetheless having decrease GFLOPS (4.2 vs 4.5).
Subsequent, it’s a frequent observe to make use of Conv-BN-ReLU construction in CNN-based structure, which is strictly what ResNet implements as properly. As an alternative of following this conference, authors determined to implement solely a single batch normalization layer, which is positioned earlier than the primary pointwise convolution layer. This transformation improved the accuracy to 81.4%, surpassing the accuracy of Swin-T by a bit of bit. Regardless of this achievement, parameter tuning was nonetheless continued by changing batch norm with layer norm, which once more raised the accuracy by 0.1% to 81.5%. All of the modifications associated to micro design resulted within the structure proven in Determine 5 (the rightmost picture). Right here you possibly can see how a ConvNeXt block differs from Swin Transformer and ResNet blocks.
The very last thing the authors did associated to the micro design was making use of separate downsampling layers. Within the authentic ResNet structure, the spatial dimension of a tensor reduces by half once we transfer from one stage to a different. You possibly can see in Determine 2 that originally ResNet accepts enter of dimension 224×224 which then shrinks to 112×112, 56×56, 28×28, 14×14, and seven×7 at stage conv1, conv2_x, conv3_x, conv4_x and conv5_x, respectively. Particularly in conv2_x and the following ones, the spatial dimension discount is finished by altering the stride parameter of the pointwise convolution to 2. As an alternative of doing so, ConvNeXt performs downsampling by putting one other convolution layer proper earlier than the element-wise summation operation inside the block. The kernel dimension and stride of this layer are set to 2, simulating a non-overlapping sliding window. In actual fact, it’s talked about within the paper that utilizing this separate downsampling layer brought on the accuracy to degrade as an alternative. Nonetheless, authors managed to resolve this subject by making use of extra layer normalization layers at a number of elements of the community, i.e., earlier than every downsampling layer, after the stem stage and after the worldwide common pooling layer (proper earlier than the ultimate output layer). With this tuning, authors efficiently boosted the accuracy to 82.0%, which is way larger than Swin-T (81.3%) whereas nonetheless having the very same GFLOPS (4.5).
And that’s principally all of the modifications made on the unique ResNet to create the ConvNeXt structure. Don’t fear if it nonetheless feels a bit unclear for now — I imagine issues will grow to be clearer as we get into the code.
ConvNeXt Implementation
Determine 6 beneath shows the small print of the complete ConvNeXt-T structure which we’ll later implement each single of its elements one after the other. Right here you may also see the way it differs from ResNet-50 and Swin-T, the 2 fashions which can be similar to ConvNeXt-T.
In the case of the implementation, the very first thing we have to do is to import the required modules. The one two we import listed below are the bottom torch module and its nn submodule for loading neural community layers.
# Codeblock 1
import torch
import torch.nn as nnConvNeXt Block
Now let’s begin with the ConvNeXt block. You possibly can see in Determine 6 that the block constructions in res2, res3, res4, and res5 levels are principally the identical, by which all of these correspond to the rightmost illustration in Determine 5. Thanks to those equivalent constructions, we will implement them in a single class and use it repeatedly. Have a look at the Codeblock 2a and 2b beneath to see how I try this.
# Codeblock 2a
class ConvNeXtBlock(nn.Module):
def __init__(self, num_channels): #(1)
tremendous().__init__()
hidden_channels = num_channels * 4 #(2)
self.conv0 = nn.Conv2d(in_channels=num_channels, #(3)
out_channels=num_channels, #(4)
kernel_size=7, #(5)
stride=1,
padding=3, #(6)
teams=num_channels) #(7)
self.norm = nn.LayerNorm(normalized_shape=num_channels) #(8)
self.conv1 = nn.Conv2d(in_channels=num_channels, #(9)
out_channels=hidden_channels,
kernel_size=1,
stride=1,
padding=0)
self.gelu = nn.GELU() #(10)
self.conv2 = nn.Conv2d(in_channels=hidden_channels, #(11)
out_channels=num_channels,
kernel_size=1,
stride=1,
padding=0)I made a decision to call this class ConvNeXtBlock. You possibly can see at line #(1) within the above codeblock that this class accepts num_channels as the one parameter, by which it denotes each the variety of enter and output channels. Do not forget that a ConvNeXt block follows the sample of the inverted bottleneck construction, i.e., slender → broad → slender. In the event you take a better take a look at Determine 6, you’ll discover that the broad half is 4 occasions bigger than the slender half. Thus, we set the worth of the hidden_channels variable accordingly (#(2)).
Subsequent, we initialize 3 convolution layers which I check with them as conv0, conv1 and conv2. Each single of those convolution layers has their very own specs. For conv0, we set the variety of enter and output channels to be the identical, which is the explanation that each its in_channels and out_channels parameters are set to num_channels (#(3–4)). We set the kernel dimension of this layer to 7×7 (#(5)). Given this specification, we have to set the padding dimension to three in an effort to retain the spatial dimension (#(6)). Don’t neglect to set the teams parameter to num_channels as a result of we wish this to be a depthwise convolution layer (#(7)). Then again, the conv1 layer (#(9)) is accountable to extend the variety of picture channels, whereas the following conv2 layer (#(11)) is employed to shrink the tensor again to the unique channel depend. It is very important word that conv1 and conv2 are each utilizing 1×1 kernel dimension, which basically implies that it solely works by combining info alongside the channel dimension. Moreover, right here we additionally have to initialize layer norm (#(8)) and GELU activation operate (#(10)) because the substitute for batch norm and ReLU.
As all layers required within the ConvNeXtBlock have been initialized, what we have to do subsequent is to outline the circulation of the tensor within the ahead() technique beneath.
# Codeblock 2b
def ahead(self, x):
residual = x #(1)
print(f'x & residualt: {x.dimension()}')
x = self.conv0(x)
print(f'after conv0t: {x.dimension()}')
x = x.permute(0, 2, 3, 1) #(2)
print(f'after permutet: {x.dimension()}')
x = self.norm(x)
print(f'after normt: {x.dimension()}')
x = x.permute(0, 3, 1, 2) #(3)
print(f'after permutet: {x.dimension()}')
x = self.conv1(x)
print(f'after conv1t: {x.dimension()}')
x = self.gelu(x)
print(f'after gelut: {x.dimension()}')
x = self.conv2(x)
print(f'after conv2t: {x.dimension()}')
x = x + residual #(4)
print(f'after summationt: {x.dimension()}')
return xWhat we principally do within the above code is simply passing the tensor to every layer we outlined earlier sequentially. Nonetheless, there are two issues I would like to focus on right here. First, we have to retailer the unique enter tensor to the residual variable (#(1)), by which it’ll skip over all operations inside the ConvNeXt block. Secondly, do not forget that layer norm is often used for sequential knowledge, the place it sometimes has a distinct form from that of picture knowledge. As a consequence of this purpose, we have to regulate the tensor dimension such that the form turns into (N, H, W, C) (#(2)) earlier than we truly carry out the layer normalization operation. Afterwards, don’t neglect to permute this tensor again to (N, C, H, W) (#(3)). The ensuing tensor is then handed by the remaining layers earlier than being summed with the residual connection (#(4)).
To test if our ConvNeXtBlock class works correctly, we will check it utilizing the Codeblock 3 beneath. Right here we’re going to simulate the block utilized in res2 stage. So, we set the num_channels parameter to 96 (#(1)) and create a dummy tensor which we assume as a batch of single picture of dimension 56×56 (#(2)).
# Codeblock 3
convnext_block_test = ConvNeXtBlock(num_channels=96) #(1)
x_test = torch.rand(1, 96, 56, 56) #(2)
out_test = convnext_block_test(x_test)Under is what the ensuing output seems like. Speaking concerning the inside circulation, it looks like all layers we stacked earlier work correctly. At line #(1) within the output beneath we will see that the tensor dimension modified to 1×56×56×96 (N, H, W, C) after being permuted. This tensor dimension then modified again to 1×96×56×56 (N, C, H, W) after the second permute operation (#(2)). Subsequent, the conv1 layer efficiently expanded the variety of channels to be 4 occasions higher than the enter (#(3)) which was then lowered again to the unique channel depend (#(4)). Right here you possibly can see that the tensor form on the first and the final layer are precisely the identical, permitting us to stack a number of ConvNeXt blocks as many as we wish.
# Codeblock 3 Output
x & residual : torch.Measurement([1, 96, 56, 56])
after conv0 : torch.Measurement([1, 96, 56, 56])
after permute : torch.Measurement([1, 56, 56, 96]) #(1)
after norm : torch.Measurement([1, 56, 56, 96])
after permute : torch.Measurement([1, 96, 56, 56]) #(2)
after conv1 : torch.Measurement([1, 384, 56, 56]) #(3)
after gelu : torch.Measurement([1, 384, 56, 56])
after conv2 : torch.Measurement([1, 96, 56, 56]) #(4)
after summation : torch.Measurement([1, 96, 56, 56])ConvNeXt Block Transition
The following element I need to implement is the one I check with because the ConvNeXt block transition. The thought of this block is definitely much like the ConvNeXt block we applied earlier, besides that this transition block is used once we are about to maneuver from a stage to the following one. Extra particularly, this block will later be employed as the primary ConvNeXt block in every stage (besides res2). The explanation I implement it in separate class is that there are some intricate particulars that differ from the ConvNeXt block. Moreover, it’s value noting that the time period transition will not be formally used within the paper. Moderately, it’s simply the phrase I take advantage of by myself to explain this concept. — I truly additionally used this system again once I write concerning the smaller ResNet model, i.e., ResNet-18 and ResNet-34. Click on on the hyperlink at reference quantity [6] on the finish of this text if you happen to’re to learn that one.
# Codeblock 4a
class ConvNeXtBlockTransition(nn.Module):
def __init__(self, in_channels, out_channels): #(1)
tremendous().__init__()
hidden_channels = out_channels * 4
self.projection = nn.Conv2d(in_channels=in_channels, #(2)
out_channels=out_channels,
kernel_size=1,
stride=2,
padding=0)
self.conv0 = nn.Conv2d(in_channels=in_channels,
out_channels=out_channels,
kernel_size=7,
stride=1,
padding=3,
teams=in_channels)
self.norm0 = nn.LayerNorm(normalized_shape=out_channels)
self.conv1 = nn.Conv2d(in_channels=out_channels,
out_channels=hidden_channels,
kernel_size=1,
stride=1,
padding=0)
self.gelu = nn.GELU()
self.conv2 = nn.Conv2d(in_channels=hidden_channels,
out_channels=out_channels,
kernel_size=1,
stride=1,
padding=0)
self.norm1 = nn.LayerNorm(normalized_shape=out_channels) #(3)
self.downsample = nn.Conv2d(in_channels=out_channels, #(4)
out_channels=out_channels,
kernel_size=2,
stride=2)The primary distinction you would possibly discover right here is the enter of the __init__() technique, which on this case we separate the variety of enter and output channels into two parameters as seen at line #(1) in Codeblock 4a. That is basically performed as a result of we’d like this block to take the output tensor from the earlier stage which has completely different variety of channels from that of the one to be generated within the subsequent stage. Referring to Determine 6, for instance, if we have been to create the primary ConvNeXt block in res3 stage, we have to configure it such that it accepts a tensor of 96 channels from res2 and returns one other tensor with 192 channels.
Secondly, right here we implement the separate downsample layer I defined earlier (#(4)) alongside the corresponding layer norm to be positioned earlier than it (#(3)). Because the identify suggests, this layer is employed to cut back the spatial dimension of the picture by half.
Third, we initialize the so-called projection layer at line #(2). Within the ConvNeXtBlock we created earlier, this layer will not be essential as a result of the enter and output tensor is strictly the identical. Within the case of transition block, the picture spatial dimension is lowered by half, whereas on the similar time the variety of output channels is doubled. This projection layer is accountable to regulate the dimension of the residual connection in an effort to match it with the one from the principle circulation, permitting element-wise operation to be carried out.
The ahead() technique within the Codeblock 4b beneath can be much like the one belongs to the ConvNeXtBlock class, besides that right here the residual connection must be processed with the projection layer (#(1)) whereas the principle tensor requires to be downsampled (#(2)) earlier than the summation is finished at line #(3).
# Codeblock 4b
def ahead(self, x):
print(f'originaltt: {x.dimension()}')
residual = self.projection(x) #(1)
print(f'residual after projt: {residual.dimension()}')
x = self.conv0(x)
print(f'after conv0tt: {x.dimension()}')
x = x.permute(0, 2, 3, 1)
print(f'after permutett: {x.dimension()}')
x = self.norm0(x)
print(f'after norm1tt: {x.dimension()}')
x = x.permute(0, 3, 1, 2)
print(f'after permutett: {x.dimension()}')
x = self.conv1(x)
print(f'after conv1tt: {x.dimension()}')
x = self.gelu(x)
print(f'after gelutt: {x.dimension()}')
x = self.conv2(x)
print(f'after conv2tt: {x.dimension()}')
x = x.permute(0, 2, 3, 1)
print(f'after permutett: {x.dimension()}')
x = self.norm1(x)
print(f'after norm1tt: {x.dimension()}')
x = x.permute(0, 3, 1, 2)
print(f'after permutett: {x.dimension()}')
x = self.downsample(x) #(2)
print(f'after downsamplet: {x.dimension()}')
x = x + residual #(3)
print(f'after summationtt: {x.dimension()}')
return xNow let’s check the ConvNeXtBlockTransition class above utilizing the next codeblock. Suppose we’re about to implement the primary ConvNeXt block in stage res3. To take action, we will merely instantiate the transition block with in_channels=96 and out_channels=192 earlier than ultimately passing a dummy tensor of dimension 1×96×56×56 by it.
# Codeblock 5
convnext_block_transition_test = ConvNeXtBlockTransition(in_channels=96,
out_channels=192)
x_test = torch.rand(1, 96, 56, 56)
out_test = convnext_block_transition_test(x_test)# Codeblock 5 Output
authentic : torch.Measurement([1, 96, 56, 56])
residual after proj : torch.Measurement([1, 192, 28, 28]) #(1)
after conv0 : torch.Measurement([1, 192, 56, 56]) #(2)
after permute : torch.Measurement([1, 56, 56, 192])
after norm0 : torch.Measurement([1, 56, 56, 192])
after permute : torch.Measurement([1, 192, 56, 56])
after conv1 : torch.Measurement([1, 768, 56, 56])
after gelu : torch.Measurement([1, 768, 56, 56])
after conv2 : torch.Measurement([1, 192, 56, 56]) #(3)
after permute : torch.Measurement([1, 56, 56, 192])
after norm1 : torch.Measurement([1, 56, 56, 192])
after permute : torch.Measurement([1, 192, 56, 56])
after downsample : torch.Measurement([1, 192, 28, 28]) #(4)
after summation : torch.Measurement([1, 192, 28, 28]) #(5)You possibly can see within the ensuing output that our projection layer straight maps the 1×96×56×56 residual tensor to 1×192×28×28 as proven at line #(1). In the meantime, the principle tensor x must be processed by the opposite layers we initialized earlier to attain this form. The steps we carried out from line #(2) to #(3) on the x tensor are principally the identical as these within the ConvNeXtBlock class. At this level we already obtained the variety of channels matches our want (192). The spatial dimension is then lowered after the tensor being processed by the downsample layer (#(4)). Because the tensor dimensions of x and residual have matched, we will lastly carry out the element-wise summation (#(5)).
The Complete ConvNeXt Structure
As we obtained ConvNeXtBlock and ConvNeXtBlockTransition lessons prepared to make use of, we will now begin to assemble the complete ConvNeXt structure. Earlier than we try this, I want to introduce some config parameters first. See the Codeblock 6 beneath.
# Codeblock 6
IN_CHANNELS = 3 #(1)
IMAGE_SIZE = 224 #(2)
NUM_BLOCKS = [3, 3, 9, 3] #(3)
OUT_CHANNELS = [96, 192, 384, 768] #(4)
NUM_CLASSES = 1000 #(5)The primary one is the dimension of the enter picture. As proven at line #(1) and #(2), right here we set in_channels to three and image_size to 224 since by default ConvNeXt accepts a batch of RGB photographs of that dimension. The following ones are associated to the mannequin configuration. On this case, I set the variety of ConvNeXt blocks of every stage to [3, 3, 9, 3] (#(3)) and the corresponding variety of output channels to [96, 192, 384, 768] (#(4)) since I need to implement the ConvNeXt-T variant. You possibly can truly change these numbers based on the configuration supplied by the unique paper proven in Determine 7. Lastly, we set the variety of neurons of the output channel to 1000, which corresponds to the variety of lessons within the dataset we practice the mannequin on (#(5)).
We are going to now implement the complete structure within the ConvNeXt class proven in Codeblock 7a and 7b beneath. The next __init__() technique may appear a bit difficult at look, however don’t fear as I’ll clarify it totally.
# Codeblock 7a
class ConvNeXt(nn.Module):
def __init__(self):
tremendous().__init__()
self.stem = nn.Conv2d(in_channels=IN_CHANNELS, #(1)
out_channels=OUT_CHANNELS[0],
kernel_size=4,
stride=4,
)
self.normstem = nn.LayerNorm(normalized_shape=OUT_CHANNELS[0]) #(2)
#(3)
self.res2 = nn.ModuleList()
for _ in vary(NUM_BLOCKS[0]):
self.res2.append(ConvNeXtBlock(num_channels=OUT_CHANNELS[0]))
#(4)
self.res3 = nn.ModuleList([ConvNeXtBlockTransition(in_channels=OUT_CHANNELS[0],
out_channels=OUT_CHANNELS[1])])
for _ in vary(NUM_BLOCKS[1]-1):
self.res3.append(ConvNeXtBlock(num_channels=OUT_CHANNELS[1]))
#(5)
self.res4 = nn.ModuleList([ConvNeXtBlockTransition(in_channels=OUT_CHANNELS[1],
out_channels=OUT_CHANNELS[2])])
for _ in vary(NUM_BLOCKS[2]-1):
self.res4.append(ConvNeXtBlock(num_channels=OUT_CHANNELS[2]))
#(6)
self.res5 = nn.ModuleList([ConvNeXtBlockTransition(in_channels=OUT_CHANNELS[2],
out_channels=OUT_CHANNELS[3])])
for _ in vary(NUM_BLOCKS[3]-1):
self.res5.append(ConvNeXtBlock(num_channels=OUT_CHANNELS[3]))
self.avgpool = nn.AdaptiveAvgPool2d(output_size=(1,1)) #(7)
self.normpool = nn.LayerNorm(normalized_shape=OUT_CHANNELS[3]) #(8)
self.fc = nn.Linear(in_features=OUT_CHANNELS[3], #(9)
out_features=NUM_CLASSES)
self.relu = nn.ReLU()The very first thing we do right here is initializing the stem stage (#(1)), which is actually only a convolution layer with 4×4 kernel dimension and stride 4. This configuration will successfully scale back the picture dimension to be 4 occasions smaller, the place each single pixel within the output tensor represents a 4×4 patch within the enter tensor. For the following levels, we have to wrap the corresponding ConvNeXt blocks with nn.ModuleList(). For stage res3 (#(4)), res4 (#(5)) and res5 (#(6)) we place ConvNeXtBlockTransition at first of every record as a “bridge” between levels. We don’t do that for stage res2 because the tensor produced by the stem stage is already appropriate with it (#(3)). Subsequent, we initialize an nn.AdaptiveAvgPool2d layer, which will likely be used to cut back the spatial dimensions of the tensor to 1×1 by computing the imply throughout every channel (#(7)). In actual fact, that is the very same course of utilized by ResNet to organize the tensor from the final convolution layer in order that it matches the form required by the following output layer (#(9)). Moreover, don’t neglect to initialize two layer normalization layers which I check with as normstem (#(2)) and normpool (#(8)), by which these two layers will then be positioned proper after the stem stage and the avgpool layer.
The ahead() technique is fairly easy. All we have to do within the following code is simply to put the layers one after one other. Needless to say because the ConvNeXt blocks are saved in lists, we have to name them iteratively with loops as seen at line #(1–4). Moreover, don’t neglect to reshape the tensor produced by the nn.AdaptiveAvgPool2d layer (#(5)) in order that it will likely be appropriate with the following fully-connected layer (#(6)).
# Codeblock 7b
def ahead(self, x):
print(f'originalt: {x.dimension()}')
x = self.relu(self.stem(x))
print(f'after stemt: {x.dimension()}')
x = x.permute(0, 2, 3, 1)
print(f'after permutet: {x.dimension()}')
x = self.normstem(x)
print(f'after normstemt: {x.dimension()}')
x = x.permute(0, 3, 1, 2)
print(f'after permutet: {x.dimension()}')
print()
for i, block in enumerate(self.res2): #(1)
x = block(x)
print(f'after res2 #{i}t: {x.dimension()}')
print()
for i, block in enumerate(self.res3): #(2)
x = block(x)
print(f'after res3 #{i}t: {x.dimension()}')
print()
for i, block in enumerate(self.res4): #(3)
x = block(x)
print(f'after res4 #{i}t: {x.dimension()}')
print()
for i, block in enumerate(self.res5): #(4)
x = block(x)
print(f'after res5 #{i}t: {x.dimension()}')
print()
x = self.avgpool(x)
print(f'after avgpoolt: {x.dimension()}')
x = x.permute(0, 2, 3, 1)
print(f'after permutet: {x.dimension()}')
x = self.normpool(x)
print(f'after normpoolt: {x.dimension()}')
x = x.permute(0, 3, 1, 2)
print(f'after permutet: {x.dimension()}')
x = x.reshape(x.form[0], -1) #(5)
print(f'after reshapet: {x.dimension()}')
x = self.fc(x)
print(f'after fct: {x.dimension()}') #(6)
return xNow for the second of reality, let’s see if we now have appropriately applied the complete ConvNeXt mannequin by operating the next code. Right here I attempt to cross a tensor of dimension 1×3×224×224 to the community, simulating a batch of a single RGB picture of dimension 224×224.
# Codeblock 8
convnext_test = ConvNeXt()
x_test = torch.rand(1, IN_CHANNELS, IMAGE_SIZE, IMAGE_SIZE)
out_test = convnext_test(x_test)You possibly can see within the following output that it seems like our implementation is right because the habits of the community aligns with the architectural design proven in Determine 6. The spatial dimension of the picture steadily will get smaller as we get deeper into the community, and on the similar time the variety of channels will increase as an alternative due to the ConvNeXtBlockTransition blocks we positioned at first of stage res3 (#(1)), res4 (#(2)), and res5 (#(3)). The avgpool layer then appropriately downsampled the spatial dimension to 1×1 (#(4)), permitting it to be related to the output layer (#(5)).
# Codeblock 8 Output
authentic : torch.Measurement([1, 3, 224, 224])
after stem : torch.Measurement([1, 96, 56, 56])
after permute : torch.Measurement([1, 56, 56, 96])
after normstem : torch.Measurement([1, 56, 56, 96])
after permute : torch.Measurement([1, 96, 56, 56])
after res2 #0 : torch.Measurement([1, 96, 56, 56])
after res2 #1 : torch.Measurement([1, 96, 56, 56])
after res2 #2 : torch.Measurement([1, 96, 56, 56])
after res3 #0 : torch.Measurement([1, 192, 28, 28]) #(1)
after res3 #1 : torch.Measurement([1, 192, 28, 28])
after res3 #2 : torch.Measurement([1, 192, 28, 28])
after res4 #0 : torch.Measurement([1, 384, 14, 14]) #(2)
after res4 #1 : torch.Measurement([1, 384, 14, 14])
after res4 #2 : torch.Measurement([1, 384, 14, 14])
after res4 #3 : torch.Measurement([1, 384, 14, 14])
after res4 #4 : torch.Measurement([1, 384, 14, 14])
after res4 #5 : torch.Measurement([1, 384, 14, 14])
after res4 #6 : torch.Measurement([1, 384, 14, 14])
after res4 #7 : torch.Measurement([1, 384, 14, 14])
after res4 #8 : torch.Measurement([1, 384, 14, 14])
after res5 #0 : torch.Measurement([1, 768, 7, 7]) #(3)
after res5 #1 : torch.Measurement([1, 768, 7, 7])
after res5 #2 : torch.Measurement([1, 768, 7, 7])
after avgpool : torch.Measurement([1, 768, 1, 1]) #(4)
after permute : torch.Measurement([1, 1, 1, 768])
after normpool : torch.Measurement([1, 1, 1, 768])
after permute : torch.Measurement([1, 768, 1, 1])
after reshape : torch.Measurement([1, 768])
after fc : torch.Measurement([1, 1000]) #(5)Ending
Properly, that was just about every part concerning the principle and the implementation of the ConvNeXt structure. Once more, I do acknowledge that the code I show above may not totally seize every part since this text is meant to cowl the overall concept of the mannequin. So, I extremely suggest you learn the unique implementation by Meta’s researchers [2] if you wish to know extra concerning the intricate particulars.
I hope you discover this text helpful. Thanks for studying!
P.S. the pocket book used on this article is out there on my GitHub repo. See the hyperlink at reference quantity [7].
References
[1] Zhuang Liu et al. A ConvNet for the 2020s. Arxiv. https://arxiv.org/pdf/2201.03545 [Accessed January 18, 2025].
[2] facebookresearch. ConvNeXt. GitHub. https://github.com/facebookresearch/ConvNeXt/blob/essential/fashions/convnext.py [Accessed January 18, 2025].
[3] Kaiming He et al. Deep Residual Studying for Picture Recognition. Arxiv. https://arxiv.org/pdf/1512.03385 [Accessed January 18, 2025].
[4] Ze Liu et al. Swin Transformer: Hierarchical Imaginative and prescient Transformer utilizing Shifted Home windows. Arxiv. https://arxiv.org/pdf/2103.14030 [Accessed January 18, 2025].
[5] Saining Xie et al. Aggregated Residual Transformations for Deep Neural Networks. Arxiv. https://arxiv.org/pdf/1611.05431 [Accessed January 18, 2025].
[6] Muhammad Ardi. Paper Walkthrough: Residual Community (ResNet). Python in Plain English. https://python.plainenglish.io/paper-walkthrough-residual-network-resnet-62af58d1c521 [Accessed January 19, 2025].
[7] MuhammadArdiPutra. The CNN That Challenges ViT — ConvNeXt. GitHub. https://github.com/MuhammadArdiPutra/medium_articles/blob/essential/Thepercent20CNNpercent20Thatpercent20Challengespercent20ViTpercent20-%20ConvNeXt.ipynb [Accessed January 24, 2025].
