mobilenet.py

"""
    MobileNet for ImageNet-1K, implemented in PyTorch.
    Original paper: 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
    https://arxiv.org/abs/1704.04861.
"""

__all__ = ['MobileNet', 'mobilenet_w1', 'mobilenet_w3d4', 'mobilenet_wd2', 'mobilenet_wd4', 'get_mobilenet']

import os
import torch.nn as nn
from common import conv3x3_block, dwsconv3x3_block


class MobileNet(nn.Module):
    """
    MobileNet model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
    https://arxiv.org/abs/1704.04861.

    Parameters:
    ----------
    channels : list of list of int
        Number of output channels for each unit.
    first_stage_stride : bool
        Whether stride is used at the first stage.
    in_channels : int, default 3
        Number of input channels.
    in_size : tuple of two ints, default (224, 224)
        Spatial size of the expected input image.
    num_classes : int, default 1000
        Number of classification classes.
    """
    def __init__(self,
                 channels,
                 first_stage_stride,
                 in_channels=3,
                 in_size=(224, 224),
                 num_classes=1000,
                 dropout=0.0,
                 global_pool=True):
        super(MobileNet, self).__init__()
        self.in_size = in_size
        self.num_classes = num_classes

        self.features = nn.Sequential()
        init_block_channels = channels[0][0]
        self.features.add_module("init_block", conv3x3_block(
            in_channels=in_channels,
            out_channels=init_block_channels,
            stride=2))
        in_channels = init_block_channels
        for i, channels_per_stage in enumerate(channels[1:]):
            stage = nn.Sequential()
            for j, out_channels in enumerate(channels_per_stage):
                stride = 2 if (j == 0) and ((i != 0) or first_stage_stride) else 1
                stage.add_module("unit{}".format(j + 1), dwsconv3x3_block(
                    in_channels=in_channels,
                    out_channels=out_channels,
                    stride=stride))
                in_channels = out_channels
            self.features.add_module("stage{}".format(i + 1), stage)
        if global_pool:
            pass
        else:
            self.features.add_module("final_pool", nn.AvgPool2d(
                kernel_size=7,
                stride=1))

        self.dropout = dropout
        if dropout > 0:
            self.dropout_layer = nn.Dropout(p=dropout)

        # self.output = nn.Linear(
        #     in_features=in_channels,
        #     out_features=num_classes)
        self.classifier = nn.Linear(
            in_features=in_channels,
            out_features=num_classes)

        self._init_params()

    def _init_params(self):
        for name, module in self.named_modules():
            if 'dw_conv.conv' in name:
                nn.init.kaiming_normal_(module.weight, mode='fan_in')
            elif name == 'init_block.conv' or 'pw_conv.conv' in name:
                nn.init.kaiming_normal_(module.weight, mode='fan_out')
            elif 'bn' in name:
                nn.init.constant_(module.weight, 1)
                nn.init.constant_(module.bias, 0)
            elif 'output' in name:
                nn.init.kaiming_normal_(module.weight, mode='fan_out')
                nn.init.constant_(module.bias, 0)

    def forward(self, x):
        x = self.features(x)
        x = x.view(x.size(0), -1)
        # x = self.output(x)
        if self.dropout > 0:
            x = self.dropout_layer(x)
        x = self.classifier(x)
        return x


def get_mobilenet(width_scale,
                  **kwargs):
    """
    Create MobileNet model with specific parameters.

    Parameters:
    ----------
    width_scale : float
        Scale factor for width of layers. The same as depth depth_multiplier
    """
    channels = [[32], [64], [128, 128], [256, 256], [512, 512, 512, 512, 512, 512], [1024, 1024]]
    first_stage_stride = False

    if width_scale != 1.0:
        channels = [[int(cij * width_scale) for cij in ci] for ci in channels]

    net = MobileNet(
        channels=channels,
        first_stage_stride=first_stage_stride,
        **kwargs)

    return net


def mobilenet_w1(**kwargs):
    """
    1.0 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
    https://arxiv.org/abs/1704.04861.
    """
    return get_mobilenet(width_scale=1.0, **kwargs)


def mobilenet_w3d4(**kwargs):
    """
    0.75 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
    https://arxiv.org/abs/1704.04861.
    """
    return get_mobilenet(width_scale=0.75, **kwargs)


def mobilenet_wd2(**kwargs):
    """
    0.5 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
    https://arxiv.org/abs/1704.04861.
    """
    return get_mobilenet(width_scale=0.5, **kwargs)


def mobilenet_wd4(**kwargs):
    """
    0.25 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,'
    https://arxiv.org/abs/1704.04861.
    """
    return get_mobilenet(width_scale=0.25, **kwargs)


def _calc_width(net):
    import numpy as np
    net_params = filter(lambda p: p.requires_grad, net.parameters())
    weight_count = 0
    for param in net_params:
        weight_count += np.prod(param.size())
    return weight_count


def _test():
    import torch

    pretrained = False

    models = [
        mobilenet_w1,
        mobilenet_w3d4,
        mobilenet_wd2,
        mobilenet_wd4,
    ]

    for model in models:

        net = model(pretrained=pretrained)

        # net.train()
        net.eval()
        weight_count = _calc_width(net)
        print("m={}, {}".format(model.__name__, weight_count))
        assert (model != mobilenet_w1 or weight_count == 4231976)
        assert (model != mobilenet_w3d4 or weight_count == 2585560)
        assert (model != mobilenet_wd2 or weight_count == 1331592)
        assert (model != mobilenet_wd4 or weight_count == 470072)

        x = torch.randn(1, 3, 224, 224)
        y = net(x)
        y.sum().backward()
        assert (tuple(y.size()) == (1, 1000))


if __name__ == "__main__":
    # _test()
    net = mobilenet_w1(in_size=128)
    import torch
    net.load_state_dict(torch.load("mobilenet_v1_128_tf2torch.pth"), strict=False)