Welcome back to the Tiny Giant series — a series where I share what I learned about MobileNet architectures. In the past two articles I covered MobileNetV1 and MobileNetV2. Check out references [1] and [2] if you’re interested in reading them. In today’s article I would like to continue with the next version of the model: MobileNetV3.

MobileNetV3 was first proposed in a paper titled “Searching for MobileNetV3” written by Howard et al. in 2019 [3]. Just a quick review: the main idea of the first MobileNet version was replacing full-convolutions with depthwise separable convolutions, which reduced the number of params by nearly 90% compared to its standard CNN counterpart. In the second MobileNet version, the authors introduced the so-called inverted residual and linear bottleneck mechanisms, which they integrated into the original MobileNetV1 building blocks. Now in the third MobileNet version, the authors attempted to push the performance of the network even further by incorporating Squeeze-and-Excitation (SE) modules and hard activation functions into the building blocks. Additionally, the overall structure of MobileNetV3 itself is partially designed using NAS (Neural Architecture Search), in which it essentially works somewhat like a parameter tuning that operates on the architectural level by maximizing accuracy while minimizing latency. However, note that in this article I won’t go into how NAS works in detail. Instead, I will focus on the final design of MobileNetV3 proposed in the paper.

---

The Detailed MobileNetV3 Architecture

The authors propose two variants of this model which they refer to as MobileNetV3-Large and MobileNetV3-Small. You can see the details of the two architectures in Figure 1 below.

Figure 1. The MobileNetV3-Large (left) and MobileNetV3-Small (right) architectures [3].

Taking a closer look at the architecture, we can see that the two networks mainly consist of bneck (bottleneck) blocks. The configuration of the blocks themselves is described in columns exp size, #out, SE, NL, and s. The internal structure of these blocks as well as the corresponding parameter configurations will be discussed further in the following subsection.

---

The Bottleneck

MobileNetV3 uses the modified version of the building blocks used in MobileNetV2. As I’ve mentioned earlier, what makes the two different is the presence of SE module and the use of hard activation function. You can see the two building blocks in Figure 2, with MobileNetV2 at the top and MobileNetV3 at the bottom.

Figure 2. The MobileNetV2 (top) and MobileNetV3 (bottom) building blocks [3].

Notice that the first two convolution layers in both building blocks are basically the same: a pointwise convolution followed by a depthwise convolution. The former is used for expanding the number of channels to *exp size *(expansion size), whereas the latter is responsible to process each channel of the resulting tensor independently. The only difference between the two building blocks lies in the activation functions used, which they refer to as NL (Nonlinearity). In MobileNetV2, the activation functions placed after the two convolution layers are set fixed to ReLU6, whereas in MobileNetV3 it can either be ReLU6 or hard-swish. The RE and HS you saw earlier in Figure 1 basically refer to these two types of activations.

Next, in MobileNetV3 we place the SE module after the depthwise convolution layer. If you’re not yet familiar with SE module, it is essentially a kind of building block we can attach in any kind of CNN-based model. This component is useful for giving weights to different channels, allowing the model to pay more attention to the important channels only. I actually have a separate article discussing the SE module in detail. Click on the link at reference number [4] if you want to read that one. It is important to note that the SE module used here is slightly different, in that the last FC layer uses hard-sigmoid rather than the standard sigmoid activation function. (I’ll talk more about the hard activations used in MobileNetV3 later in the subsequent subsection.) In fact, the SE module itself is not always included in every bottleneck block. If you go back to Figure 1, you’ll notice that some of the bottleneck blocks have a checkmark in the SE column, indicating that the SE module is applied. On the other hand, some blocks don’t include the module, which might probably be because the NAS process didn’t find any performance improvement from using SE modules in those blocks.

As the SE module has been connected, we need to place another pointwise convolution, which is responsible to adjust the number of output channels according to the #out column in Figure 1. This pointwise convolution does not include any activation function, aligning with the linear bottleneck design originally introduced in MobileNetV2. I actually need to clarify something here. If you take a look at the MobileNetV2 building block in Figure 2 above, you’ll notice that the last pointwise convolution has a ReLU6 placed on it. I believe this is a mistake made by the authors, because according to the MobileNetV2 paper [6], the ReLU6 should be in the first pointwise convolution at the beginning of the block instead.

Last but not least, notice that there is also a residual connection that skips across all layers in the bottleneck block. This connection is only present when the output tensor has the exact same dimensions as the input, i.e., when the number of input and output channels is the same and when the s (stride) is 1.

Hard-Sigmoid and Hard-Swish

The activation functions used in MobileNetV3 are not commonly found in other deep learning models. To start with, let’s look at the hard-sigmoid activation first, which is the one used in the SE module as a replacement for the conventional sigmoid. Take a look at Figure 3 below to see the difference between the two.

Figure 3. The sigmoid and the hard-sigmoid activation functions [3].

Here you might probably be wondering, why don’t we just use the conventional sigmoid? Why do we really need to use piecewise linear function that appears less smooth instead? To answer this question, we need to understand the mathematical definition of a sigmoid function in advance, which I provide in Figure 4 below.

Figure 4. The equation of the standard sigmoid function [5].

We can clearly see in the above figure that the sigmoid function originally involves an exponential term in the denominator. In fact, this term causes the function to be computationally expensive, which in turn makes the activation function less suitable for low-power devices. Not only that, the output of the sigmoid function itself is a high-precision floating-point value, which is also not preferable for low-power devices due to their limited support for handling such values.

If you look at Figure 3 again, you might think that the hard-sigmoid function is directly derived from the original sigmoid. In fact, that’s actually not quite right. Despite having a similar shape, hard-sigmoid is basically constructed using ReLU6 instead, which can formally be expressed in Figure 5 below. Here you can see that the equation is much simpler as it only consists of basic arithmetic operations and clipping, allowing it to be processed much faster.

Figure 5. The equation of the hard sigmoid function [5].

The next activation function we are going to utilize in MobileNetV3 is the so-called hard-swish, which will be implemented after each of the first two convolution layers in the bottleneck block. Just like sigmoid and hard-sigmoid, the graph of the hard-swish function appears to be similar to the original one.

Figure 6. The swish and hard-swish activation functions [3].

The original swish function itself can mathematically be expressed in the equation in Figure 7. Again, since the equation involves sigmoid, it will definitely slow down the computation. Hence, to speed up the process, we can simply replace the sigmoid function with hard-sigmoid we just discussed. By doing so, we now have the hard version of the swish activation function as shown in Figure 8.

Figure 7. The equation of the swish activation function [5]. Figure 8. The equation of the hard-swish activation function [5].

---

Some Experimental Results

Before we get into the experimental results, you need to know that there are two parameters in MobileNetV3 that allow us to adjust the model size according to our needs. These two parameters are width multiplier and input resolution, which in MobileNetV1 are known as α and ρ, respectively. Although we can technically adjust the value for the two freely, the authors already provided several numbers we can use. For the width multiplier, we can set it to either 0.35, 0.5, 0.75, 1.0, or 1.25, where using a value smaller than 1.0 causes the model to have fewer number of channels than those disclosed in Figure 1, effectively reducing the model size. For instance, if we set this parameter to 0.35, then the model will only have 35% of its default width (i.e., channel count) throughout the entire network.

Meanwhile, the input resolution can either be 96, 128, 160, 192, 224, or 256, which as the name suggests, it directly controls the spatial dimension of the input image. It is worth noting that even though using a small input size reduces the number of operations during inference, it does not affect the model size at all. So, if your objective is to reduce model size, you need to adjust the width multiplier, whereas if your goal is to lower computational cost, you can play around with both the width multiplier and input resolution.

Now looking at the experimental results in Figure 9, we can clearly see that MobileNetV3 outperforms MobileNetV2 in terms of accuracy at similar latency. The MobileNetV3-Small of default configuration (i.e., width multiplier 1.0 and input resolution 224×224) indeed has a lower accuracy than the largest MobileNetV2 variant. But if you take the default MobileNetV3-Large into account, it got an easy win over the largest MobileNetV2 both in terms of accuracy and latency. Additionally, we can still push the accuracy of MobileNetV3 even further by enlarging the model size by 1.25 times (the blue datapoint at the top right), but keep in mind that doing so significantly sacrifices computational speed.

Figure 9. Performance comparison between MobileNetV3-Large, MobileNetV3-Small, and MobileNetV2 [3].

The authors also conducted a comparative analysis with other lightweight models, of which the results are shown in the table in Figure 10.

Figure 10. Performance comparison of MobileNetV3 with other lightweight models [3].

The rows of the table above are divided into two groups, where the upper group is used to compare models with complexity similar to MobileNetV3-Large, while the lower group consists of models comparable to MobileNetV3-Small. Here you can see that both V3-Large and V3-Small obtained the best accuracy on ImageNet within their respective groups. It is worth noting that although MnasNet-A1 and V3-Large have the exact same accuracy, the number of operations (MAdds) of the former model is higher, which results in higher latency, as seen in columns P-1, P-2, and P-3 (measured in milliseconds). In case you’re wondering, the labels P-1, P-2, and P-3 essentially correspond to different Google Pixel series used to test the actual computational speed. Next, it is necessary to acknowledge that both MobileNetV3 variants have the highest parameter count (the params column) compared to other models in their group. However, this does not seem to be a major concern for the authors as the primary goal of MobileNetV3 is to minimize computational latency, even if that means having a slightly bigger model.

The next experiment the authors conducted was about the effects of value quantization, i.e., a technique that reduces the precision of floating-point numbers to speed up computation. While the networks already incorporate hard activation functions, which are compatible with quantized values, this experiment takes quantization a step further by applying it to the entire network to see how much the speed improves. The experimental results when value quantization was applied are shown in Figure 11 below.

Figure 11. The accuracy and latency of MobileNetV2 and MobileNetV3 when using quantized values [3].

If you compare the results of V2 and V3 in Figure 11 with the corresponding models in Figure 10, you’ll notice that there is a decrease in latency, proving that the use of low-precision numbers does improve computational speed. However, it is important to keep in mind that this also leads to a decrease in accuracy.

---

MobileNetV3 Implementation

I think all the explanations above cover pretty much everything you need to know about the theory behind MobileNetV3. Now in this section I am going to bring you into the most fun part of this article: implementing MobileNetV3 from scratch.

As always, the very first thing we do is importing the required modules.

# Codeblock 1
import torch
import torch.nn as nn

Afterwards, we need to initialize the configurable parameters of the model, namely WIDTH_MULTIPLIER, INPUT_RESOLUTION, and NUM_CLASSES, as shown in Codeblock 2 below. I believe the first two variables are straightforward as I’ve explained them thoroughly in the previous section. Here I decided to assign default values for the two. You can definitely change these numbers based on the values provided in the paper if you want to adjust the complexity of the model. Next, the third variable corresponds to the number of output neurons in the classification head. Here I set it to 1000 because the model is originally trained on the ImageNet-1K dataset. It is worth noting that the MobileNetV3 architecture is actually not limited to classification tasks only. Instead, it can also be used for object detection and semantic segmentation as demonstrated in the paper. However, since the focus of this article is to implement the backbone, let’s just use the standard classification head for the output layer to keep things simple.

# Codeblock 2
WIDTH_MULTIPLIER = 1.0
INPUT_RESOLUTION = 224
NUM_CLASSES      = 1000

What we are going to do next is to wrap the repeating components into separate classes. By doing this, we will later be able to simply instantiate them whenever needed instead of rewriting the same code over and over again. Now let’s begin with the Squeeze-and-Excitation module first.

---

The Squeeze-and-Excitation Module

The implementation of this component is shown in Codeblock 3. I am not going to get very deep into the code since it is almost exactly the same as the one in my previous article [4]. However, generally speaking, this code works by representing each input channel with a single number (line #(1)), processing the resulting vector with a sequence of linear layers (#(2–3)), then converting it into a weight vector (#(4)). Keep in mind that in the original SE module we typically use the standard sigmoid activation function to obtain the weight vector, but here in MobileNetV3 we use hard-sigmoid instead. This weight vector will then be multiplied with the original tensor, which by doing so we can reduce the influence of channels that do not give contribution to the final output (#(5)).

# Codeblock 3
class SEModule(nn.Module):
def __init__(self, num_channels, r):
super().__init__()

self.global_pooling = nn.AdaptiveAvgPool2d(output_size=(1,1))
self.fc0 = nn.Linear(in_features=num_channels,
out_features=num_channels//r,
bias=False)
self.relu6 = nn.ReLU6()
self.fc1 = nn.Linear(in_features=num_channels//r,
out_features=num_channels,
bias=False)
self.hardsigmoid = nn.Hardsigmoid()

def forward(self, x):
print(f'original\t\t: {x.size()}')

squeezed = self.global_pooling(x)              #(1)
print(f'after avgpool\t\t: {squeezed.size()}')

squeezed = torch.flatten(squeezed, 1)
print(f'after flatten\t\t: {squeezed.size()}')

excited = self.fc0(squeezed)                   #(2)
print(f'after fc0\t\t: {excited.size()}')

excited = self.relu6(excited)
print(f'after relu6\t\t: {excited.size()}')

excited = self.fc1(excited)                    #(3)
print(f'after fc1\t\t: {excited.size()}')

excited = self.hardsigmoid(excited)            #(4)
print(f'after hardsigmoid\t: {excited.size()}')

excited = excited[:, :, None, None]
print(f'after reshape\t\t: {excited.size()}')

scaled = x * excited                           #(5)
print(f'after scaling\t\t: {scaled.size()}')

return scaled

Now let’s check if the above code works properly by creating an SEModule instance and passing a dummy tensor through it. See Codeblock 4 below for the details. Here I configure the SE module to accept a 512-channel image for the input. Meanwhile, the r (reduction ratio) parameter is set to 4, meaning that the vector length between the two FC layers is going to be 4 times smaller than that of its input and output. It might be worth knowing that this number is different from the one mentioned in the original Squeeze-and-Excitation paper [7], where r = 16 is said to be the sweet spot for balancing accuracy and complexity.

# Codeblock 4
semodule = SEModule(num_channels=512, r=4)
x = torch.randn(1, 512, 28, 28)

out = semodule(x)

If the code above produces the following output, it confirms that our SE module implementation is correct as it successfully passed the input tensor through all layers within the entire SE module.

# Codeblock 4 Output
original          : torch.Size([1, 512, 28, 28])
after avgpool     : torch.Size([1, 512, 1, 1])
after flatten     : torch.Size([1, 512])
after fc0         : torch.Size([1, 128])
after relu6       : torch.Size([1, 128])
after fc1         : torch.Size([1, 512])
after hardsigmoid : torch.Size([1, 512])
after reshape     : torch.Size([1, 512, 1, 1])
after scaling     : torch.Size([1, 512, 28, 28])

---

The Convolution Block

The next component I am going to create is the one wrapped in the ConvBlock class, which the detailed implementation can be seen in Codeblock 5. In fact, this is actually just a standard convolution layer, but we don’t simply use nn.Conv2d because in CNN we typically use the Conv-BN-ReLU structure. Hence, it will be convenient if we just group these three layers together within a single class. However, instead of actually following this standard structure, we are going to customize it to match the requirements for the MobileNetV3 architecture.

# Codeblock 5
class ConvBlock(nn.Module):
def __init__(self,
in_channels,             #(1)
out_channels,            #(2)
kernel_size,             #(3)
stride,                  #(4)
padding,                 #(5)
groups=1,                #(6)
batchnorm=True,          #(7)
activation=nn.ReLU6()):  #(8)
super().__init__()

bias = False if batchnorm else True    #(9)

self.conv = nn.Conv2d(in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=bias)
self.bn = nn.BatchNorm2d(num_features=out_channels) if batchnorm else nn.Identity()  #(10)
self.activation = activation

def forward(self, x):    #(11)
print(f'original\t\t: {x.size()}')

x = self.conv(x)
print(f'after conv\t\t: {x.size()}')

x = self.bn(x)
print(f'after bn\t\t: {x.size()}')

x = self.activation(x)
print(f'after activation\t: {x.size()}')

return x

There are several parameters you need to pass to instantiate a ConvBlock instance. The first five ones (#(1–5)) are pretty straightforward as they are basically just the standard parameters for the nn.Conv2d layer. Here I set the groups parameter to be configurable (#(6)) so that this class can be flexibly used not only for standard convolutions but also for depthwise convolutions. Next, at line #(7) I create a parameter called batchnorm, which determines whether or not a ConvBlock instance implements a batch normalization layer. This is essentially done because there are some cases where we don’t implement this layer, i.e., in the last two convolutions with NBN label (which stands for no batch normalization) in Figure 1. The last parameter we have here is the activation function (#(8)). Later on, there will be cases that require us to set it to either nn.ReLU6(), nn.Hardswish() or nn.Identity() (no activation).

Inside the __init__() method, there are two things happening if we change the input argument for the batchnorm parameter. When we set it to True, firstly, the bias term of the convolution layer will be deactivated (#(9)), and secondly, bn will be an nn.BatchNorm2d() layer (#(10)). The bias term will not be used in this case because applying batch normalization after convolution will cancel it out. So, there is basically no point of utilizing bias in the first place. Meanwhile, if we set the batchnorm parameter to False, the bias variable is going to be True since in this situation it will not be canceled out. The bn itself will just be an identity layer, meaning that it won’t do anything to the tensor.

Regarding the forward() method (#(11)), I don’t think I need to explain anything because what we do here is just passing a tensor through the layers sequentially. Now let’s just move on to Codeblock 6 to see whether our ConvBlock implementation is correct. Here I try to create two ConvBlock instances, where the first one uses default batchnorm and activation, whereas the second one omits the batch normalization layer (#(1)) and uses hard-swish activation function (#(2)). Instead of passing a tensor through them, here I want you to see in the resulting output that our code correctly implements both structures according to the input arguments we pass.

# Codeblock 6
convblock1 = ConvBlock(in_channels=64,
out_channels=128,
kernel_size=3,
stride=2,
padding=1)

convblock2 = ConvBlock(in_channels=64,
out_channels=128,
kernel_size=3,
stride=2,
padding=1,
batchnorm=False,             #(1)
activation=nn.Hardswish())   #(2)

print(convblock1)
print('')
print(convblock2)

# Codeblock 6 Output
ConvBlock(
(conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
(bn): BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): ReLU6()
)

ConvBlock(
(conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1))
(bn): Identity()
(activation): Hardswish()
)

---

The Bottleneck

As the SEModule and the ConvBlock are done, we can now move on to the main component of the MobileNetV3 architecture: the bottleneck. What we essentially do in the bottleneck is just placing one layer after another which the general structure is shown earlier in Figure 2. In the case of MobileNetV2, it only consists of three convolution layers, whereas here in MobileNetV3 we have an additional SE block placed between the second and the third convolutions. Look at Codeblock 7a and 7b to see how I implement the bottleneck block for MobileNetV3.

# Codeblock 7a
class Bottleneck(nn.Module):
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
exp_size,     #(1)
se,           #(2)
activation):
super().__init__()

self.add = in_channels == out_channels and stride == 1    #(3)

self.conv0 = ConvBlock(in_channels=in_channels,    #(4)
out_channels=exp_size,    #(5)
kernel_size=1,    #(6)
stride=1,
padding=0,
activation=activation)

self.conv1 = ConvBlock(in_channels=exp_size,    #(7)
out_channels=exp_size,    #(8)
kernel_size=kernel_size,    #(9)
stride=stride,
padding=padding,
groups=exp_size,    #(10)
activation=activation)

self.semodule = SEModule(num_channels=exp_size, r=4) if se else nn.Identity()    #(11)

self.conv2 = ConvBlock(in_channels=exp_size,    #(12)
out_channels=out_channels,    #(13)
kernel_size=1,    #(14)
stride=1,
padding=0,
activation=nn.Identity())    #(15)

The input parameters of the Bottleneck class look similar to those of the ConvBlock class at a glance. This definitely makes sense because we will indeed use them to instantiate ConvBlock instances inside the Bottleneck. However, if you take a closer look at them again, you will notice that there are some other parameters you haven’t seen before, namely se (#(1)) and exp_size (#(2)). Later on, the input arguments for these parameters will be obtained from the configuration provided in the table in Figure 1.

Inside the __init__() method, what we need to do first is to check whether the input and output tensor dimensions are the same using the code at line #(3). By doing this, we will have our add variable containing either True or False. This dimensionality checking is important because we need to decide whether or not we perform element-wise summation between the two to implement the skip-connection that skips through all layers within the bottleneck block.

Next, let’s now instantiate the layers themselves, of which the first two are a pointwise convolution (conv0) and a depthwise convolution (conv1). For conv0, we need to set the kernel size to 1×1 (#(6)), whereas for conv1 the kernel size should match the one in the input argument (#(9)), which can either be 3×3 or 5×5. It is necessary to apply padding in the ConvBlock to prevent the image size from shrinking after every convolution operation. For kernel sizes of 1×1, 3×3, and 5×5, the required padding values are 0, 1, and 2, respectively. Talking about the number of channels, conv0 is responsible to expand it from in_channels to exp_size (#(4–5)). Meanwhile, the number of input and output channels of conv1 are exactly the same (#(7–8)). In addition to the conv1 layer, the groups parameter should be set to exp_size (#(10)) because we want each input channel to be processed independently of each other.

After the first two convolution layers are done, what we need to instantiate next is the Squeeze-and-Excitation module (#(11)). Here we need to set the input channel count to exp_size, matching with the tensor size produced by the conv1 layer. Remember that SE module is not always used, hence the instantiation of this component should be done within a condition, where it will actually be instantiated only when the se parameter is True. Otherwise, it will just be an identity layer.

Finally, the last convolution layer (conv2) is responsible to map the number of output channels from exp_size to out_channels (#(12–13)). Just like the conv0 layer, this one is also a pointwise convolution, hence we set the kernel size to 1×1 (#(14)) so that it only focuses on aggregating information along the channel dimension. The activation function of this layer is set fixed to nn.Identity() (#(15)) because here we will implement the idea of linear bottleneck.

And that’s pretty much everything for the layers within the bottleneck block. All we need to do afterwards is to create the flow of the network in the forward() method as shown in Codeblock 7b below.

# Codeblock 7b
def forward(self, x):
residual = x
print(f'original\t\t: {x.size()}')

x = self.conv0(x)
print(f'after conv0\t\t: {x.size()}')

x = self.conv1(x)
print(f'after conv1\t\t: {x.size()}')

x = self.semodule(x)
print(f'after semodule\t\t: {x.size()}')

x = self.conv2(x)
print(f'after conv2\t\t: {x.size()}')

if self.add:
x += residual
print(f'after summation\t\t: {x.size()}')

return x

Now I would like to test the Bottleneck class we just created by simulating the third row of the MobileNetV3-Large architecture in the table in Figure 1. Look at the Codeblock 8 below to see how I do this. If you go back to the architectural details, you will notice that this bottleneck accepts a tensor of size 16×112×112 (#(7)). In this case, the bottleneck block is configured to expand the number of channels to 64 (#(3)) before eventually shrinking it to 24 (#(1)). The kernel size of the depthwise convolution is set to 3×3 (#(2)) and the stride is set to 2 (#(4)) which will reduce the spatial dimension by half. Here we use ReLU6 for the activation function (#(6)) of the first two convolutions. Lastly, SE module will not be implemented (#(5)) since there is no checkmark in the SE column in the table.

# Codeblock 8
bottleneck = Bottleneck(in_channels=16,
out_channels=24,   #(1)
kernel_size=3,     #(2)
exp_size=64,       #(3)
stride=2,          #(4)
padding=1,
se=False,          #(5)
activation=nn.ReLU6())  #(6)

x = torch.randn(1, 16, 112, 112)           #(7)
out = bottleneck(x)

If you run the above code, the following output should appear on your screen.

# Codeblock 8 Output
original        : torch.Size([1, 16, 112, 112])
after conv0     : torch.Size([1, 64, 112, 112])
after conv1     : torch.Size([1, 64, 56, 56])
after semodule  : torch.Size([1, 64, 56, 56])
after conv2     : torch.Size([1, 24, 56, 56])

This output confirms that our implementation is correct in terms of the tensor shape, where the spatial dimension halves from 112×112 to 56×56 while the number of channels correctly expands from 16 to 64 and then reduces from 64 to 24. Talking more specifically about the SE module, we can see in the above output that the tensor is still passed through this component despite we have set the se parameter to False. In fact, if you try to print out the detailed architecture of this bottleneck like what I do in Codeblock 9, you will see that semodule is just an identity layer, which effectively makes this structure behave as if we’re passing the output of conv1 directly to conv2.

# Codeblock 9
bottleneck

# Codeblock 9 Output
Bottleneck(
(conv0): ConvBlock(
(conv): Conv2d(16, 64, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): ReLU6()
)
(conv1): ConvBlock(
(conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), groups=64, bias=False)
(bn): BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): ReLU6()
)
(semodule): Identity()
(conv2): ConvBlock(
(conv): Conv2d(64, 24, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(24, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): Identity()
)
)

The above bottleneck is going to behave differently if we instantiate it with the se parameter set to True. In Codeblock 10 below, I try to create the bottleneck block in the fifth row in the MobileNetV3-Large architecture. In this case, if you print out the detailed structure, you will see that semodule consists of all layers in the SEModule class we created earlier instead of just being an identity layer like before.

# Codeblock 10
bottleneck = Bottleneck(in_channels=24,
out_channels=40,
kernel_size=5,
exp_size=72,
stride=2,
padding=2,
se=True,
activation=nn.ReLU6())

bottleneck

# Codeblock 10 Output
Bottleneck(
(conv0): ConvBlock(
(conv): Conv2d(24, 72, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(72, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): ReLU6()
)
(conv1): ConvBlock(
(conv): Conv2d(72, 72, kernel_size=(5, 5), stride=(2, 2), padding=(2, 2), groups=72, bias=False)
(bn): BatchNorm2d(72, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): ReLU6()
)
(semodule): SEModule(
(global_pooling): AdaptiveAvgPool2d(output_size=(1, 1))
(fc0): Linear(in_features=72, out_features=18, bias=False)
(relu6): ReLU6()
(fc1): Linear(in_features=18, out_features=72, bias=False)
(hardsigmoid): Hardsigmoid()
)
(conv2): ConvBlock(
(conv): Conv2d(72, 40, kernel_size=(1, 1), stride=(1, 1), bias=False)
(bn): BatchNorm2d(40, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
(activation): Identity()
)
)

---

The Complete MobileNetV3

As all components have been created, what we need to do next is to construct the main class of the MobileNetV3 model. But before doing so, I would like to initialize a list that stores the input arguments used for instantiating the bottleneck blocks as shown in Codeblock 11 below. Keep in mind that these arguments are written according to the MobileNetV3-Large version. You’ll need to adjust the values in the BOTTLENECKS list if you want to create the small version instead.

# Codeblock 11
HS = nn.Hardswish()
RE = nn.ReLU6()

BOTTLENECKS = [[16,  16,  3, 16,  False, RE, 1, 1],
[16,  24,  3, 64,  False, RE, 2, 1],
[24,  24,  3, 72,  False, RE, 1, 1],
[24,  40,  5, 72,  True,  RE, 2, 2],
[40,  40,  5, 120, True,  RE, 1, 2],
[40,  40,  5, 120, True,  RE, 1, 2],
[40,  80,  3, 240, False, HS, 2, 1],
[80,  80,  3, 200, False, HS, 1, 1],
[80,  80,  3, 184, False, HS, 1, 1],
[80,  80,  3, 184, False, HS, 1, 1],
[80,  112, 3, 480, True,  HS, 1, 1],
[112, 112, 3, 672, True,  HS, 1, 1],
[112, 160, 5, 672, True,  HS, 2, 2],
[160, 160, 5, 960, True,  HS, 1, 2],
[160, 160, 5, 960, True,  HS, 1, 2]]

The arguments listed above are structured in the following order (from left to right): in channels, out channels, kernel size, expansion size, SE, activation, stride, and padding. Keep in mind that padding is not explicitly stated in the original table, but I include it here because it is required as an input when instantiating the bottleneck blocks.

Now let’s actually create the MobileNetV3 class. See the code implementation in Codeblocks 12a and 12b below.

# Codeblock 12a
class MobileNetV3(nn.Module):
def __init__(self):
super().__init__()

self.first_conv = ConvBlock(in_channels=3,    #(1)
out_channels=int(WIDTH_MULTIPLIER*16),
kernel_size=3,
stride=2,
padding=1,
activation=nn.Hardswish())

self.blocks = nn.ModuleList([])    #(2)
for config in BOTTLENECKS:         #(3)
in_channels, out_channels, kernel_size, exp_size, se, activation, stride, padding = config
self.blocks.append(Bottleneck(in_channels=int(WIDTH_MULTIPLIER*in_channels),
out_channels=int(WIDTH_MULTIPLIER*out_channels),
kernel_size=kernel_size,
exp_size=int(WIDTH_MULTIPLIER*exp_size),
stride=stride,
padding=padding,
se=se,
activation=activation))

self.second_conv = ConvBlock(in_channels=int(WIDTH_MULTIPLIER*160), #(4)
out_channels=int(WIDTH_MULTIPLIER*960),
kernel_size=1,
stride=1,
padding=0,
activation=nn.Hardswish())

self.avgpool = nn.AdaptiveAvgPool2d(output_size=(1,1))              #(5)

self.third_conv = ConvBlock(in_channels=int(WIDTH_MULTIPLIER*960),  #(6)
out_channels=int(WIDTH_MULTIPLIER*1280),
kernel_size=1,
stride=1,
padding=0,
batchnorm=False,
activation=nn.Hardswish())

self.dropout = nn.Dropout(p=0.8)    #(7)

self.output = ConvBlock(in_channels=int(WIDTH_MULTIPLIER*1280),     #(8)
out_channels=int(NUM_CLASSES),              #(9)
kernel_size=1,
stride=1,
padding=0,
batchnorm=False,
activation=nn.Identity())

Notice in Figure 1 that we initially start from the standard convolution layer. In the above codeblock, I refer to this layer as first_conv (#(1)). It is worth noting that the input arguments for this layer are not included in the BOTTLENECKS list, hence we need to define them manually. Remember to multiply the channel counts at each step by WIDTH_MULTIPLIER since we want the model size to be adjustable through that variable. Next, we initialize a placeholder named blocks for storing all the bottleneck blocks (#(2)). With a simple loop at line #(3), we will iterate through all items in the BOTTLENECKS list to actually instantiate the bottleneck blocks and append them one by one to blocks. In fact, this loop constructs the majority of the layers in the network, as it covers nearly all components listed in the table.

As the sequence of bottleneck blocks is done, we will now continue with the next convolution layer, which I refer to as second_conv (#(4)). Again, since the configuration parameters for this layer are not stored in the BOTTLENECKS list, we need to manually hard-code them. The output of this layer will then be passed through a global average pooling layer (#(5)) which will drop the spatial dimension to 1×1. Afterwards, we connect this layer to two consecutive pointwise convolutions (#(6) and #(8)) with a dropout layer in between (#(7)).

Talking more specifically about the two convolutions, it is important to know that applying a 1×1 convolution on a tensor that has a 1×1 spatial dimension is essentially equivalent to applying an FC layer to a flattened tensor, where the number of channels will correspond to the number of neurons. This is the reason that I set the output channel count of the last layer equal to the number of classes in the dataset (#(9)). The batchnorm parameter of both third_conv and output layers are set to False, as suggested in the architecture.

Meanwhile, the activation function of third_conv is set to nn.Hardswish(), whereas the output layer uses nn.Identity(), which is equivalent to not applying any activation function at all. This is essentially done because during training softmax is already included in the loss function (nn.CrossEntropyLoss()). Later in the inference phase, we need to replace nn.Identity() with nn.Softmax() in the output layer so that the model will directly return the probability score of each class.

Next, let’s take a look at the forward() method below, which I won’t explain any further since I think it is pretty easy to understand.

# Codeblock 12b
def forward(self, x):
print(f'original\t\t: {x.size()}')

x = self.first_conv(x)
print(f'after first_conv\t: {x.size()}')

for i, block in enumerate(self.blocks):
x = block(x)
print(f"after bottleneck #{i}\t: {x.shape}")

x = self.second_conv(x)
print(f'after second_conv\t: {x.size()}')

x = self.avgpool(x)
print(f'after avgpool\t\t: {x.size()}')

x = self.third_conv(x)
print(f'after third_conv\t: {x.size()}')

x = self.dropout(x)
print(f'after dropout\t\t: {x.size()}')

x = self.output(x)
print(f'after output\t\t: {x.size()}')

x = torch.flatten(x, start_dim=1)
print(f'after flatten\t\t: {x.size()}')

return x

The code in Codeblock 13 demonstrates how we initialize a MobileNetV3 instance and pass a dummy tensor through it. Remember that here we use the default input resolution, so we can basically think of the tensor as a batch of a single RGB image of size 224×224.

# Codeblock 13
mobilenetv3 = MobileNetV3()

x = torch.randn(1, 3, INPUT_RESOLUTION, INPUT_RESOLUTION)
out = mobilenetv3(x)

And below is what the resulting output looks like, in which the tensor dimension after each block matches exactly with the MobileNetV3-Large architecture in Figure 1.

# Codeblock 13 Output
original             : torch.Size([1, 3, 224, 224])
after first_conv     : torch.Size([1, 16, 112, 112])
after bottleneck #0  : torch.Size([1, 16, 112, 112])
after bottleneck #1  : torch.Size([1, 24, 56, 56])
after bottleneck #2  : torch.Size([1, 24, 56, 56])
after bottleneck #3  : torch.Size([1, 40, 28, 28])
after bottleneck #4  : torch.Size([1, 40, 28, 28])
after bottleneck #5  : torch.Size([1, 40, 28, 28])
after bottleneck #6  : torch.Size([1, 80, 14, 14])
after bottleneck #7  : torch.Size([1, 80, 14, 14])
after bottleneck #8  : torch.Size([1, 80, 14, 14])
after bottleneck #9  : torch.Size([1, 80, 14, 14])
after bottleneck #10 : torch.Size([1, 112, 14, 14])
after bottleneck #11 : torch.Size([1, 112, 14, 14])
after bottleneck #12 : torch.Size([1, 160, 7, 7])
after bottleneck #13 : torch.Size([1, 160, 7, 7])
after bottleneck #14 : torch.Size([1, 160, 7, 7])
after second_conv    : torch.Size([1, 960, 7, 7])
after avgpool        : torch.Size([1, 960, 1, 1])
after third_conv     : torch.Size([1, 1280, 1, 1])
after dropout        : torch.Size([1, 1280, 1, 1])
after output         : torch.Size([1, 1000, 1, 1])
after flatten        : torch.Size([1, 1000])

In order to ensure that our implementation is correct, we can print out the number of parameters contained in the model using the following code.

# Codeblock 14
total_params = sum(p.numel() for p in mobilenetv3.parameters())
total_params

# Codeblock 14 Output
5476416

Here you can see that this model contains around 5.5 million parameters, in which this is approximately the same as the one disclosed in the original paper (see Figure 10). Furthermore, the parameter count given in the PyTorch documentation is also similar to this number as you can see in Figure 12 below. Based on these facts, I believe I can confirm that our MobileNetV3-Large implementation is correct.

Figure 12. The details of the MobileNetV3-Large model from the official PyTorch documentation [8].

---

Ending

Well, that’s pretty much everything about the MobileNetV3 architecture. Here I encourage you to actually train this model from scratch on any datasets you want. Not only that, I also want you to play around with the parameter configurations of the bottleneck blocks to see whether we can still improve the performance of MobileNetV3 even further. By the way, the code used in this article is also available in my GitHub repo, which you can find in the link at reference number [9].

Thank you for reading. Feel free to reach me through LinkedIn [10] if you spot any mistake in my explanation or in the code. See ya in my next article!

---

References

[1] Muhammad Ardi. MobileNetV1 Paper Walkthrough: The Tiny Giant. AI Advances. https://medium.com/ai-advances/mobilenetv1-paper-walkthrough-the-tiny-giant-987196f40cd5 [Accessed October 24, 2025].

[2] Muhammad Ardi. MobileNetV2 Paper Walkthrough: The Smarter Tiny Giant. Towards Data Science. https://towardsdatascience.com/mobilenetv2-paper-walkthrough-the-smarter-tiny-giant/ [Accessed October 24, 2025].

[3] Andrew Howard et al. Searching for MobileNetV3. Arxiv. https://arxiv.org/abs/1905.02244 [Accessed May 1, 2025].

[4] Muhammad Ardi. SENet Paper Walkthrough: The Channel-Wise Attention. AI Advances. https://medium.com/ai-advances/senet-paper-walkthrough-the-channel-wise-attention-8ac72b9cc252 [Accessed October 24, 2025].

[5] Image created originally by author.

[6] Mark Sandler et al. MobileNetV2: Inverted Residuals and Linear Bottlenecks. Arxiv. https://arxiv.org/abs/1801.04381 [Accessed May 12, 2025].

[7] Jie Hu et al. Squeeze and Excitation Networks. Arxiv. https://arxiv.org/abs/1709.01507 [Accessed May 12, 2025].

[8] Mobilenet_v3_large. PyTorch. https://docs.pytorch.org/vision/main/models/generated/torchvision.models.mobilenet_v3_large.html#torchvision.models.mobilenet_v3_large [Accessed May 12, 2025].

[9] MuhammadArdiPutra. The Tiny Giant Getting Even Smarter — MobileNetV3. GitHub. https://github.com/MuhammadArdiPutra/medium_articles/blob/main/The%20Tiny%20Giant%20Getting%20Even%20Smarter%20-%20MobileNetV3.ipynb [Accessed May 12, 2025].

[10] Muhammad Ardi Putra. LinkedIn. https://www.linkedin.com/in/muhammad-ardi-putra-879528152/ [Accessed May 12, 2025].