model.py

import re
import math
import random
import numpy as np
import collections
from functools import partial
import torch
from torch import nn
from torch.nn import functional as F
from torch.nn import Sequential, BatchNorm1d, BatchNorm2d, Dropout, Module, Linear
from itertools import repeat
from torchvision import transforms
from torchvision.transforms.functional import rgb_to_grayscale


# Parameters for the entire model (stem, all blocks, and head)
GlobalParams = collections.namedtuple('GlobalParams', [
    'width_coefficient', 'depth_coefficient', 'image_size', 'dropout_rate',
    'num_classes', 'batch_norm_momentum', 'batch_norm_epsilon',
    'drop_connect_rate', 'depth_divisor', 'min_depth', 'include_top'])

# Parameters for an individual model block
BlockArgs = collections.namedtuple('BlockArgs', [
    'num_repeat', 'kernel_size', 'stride', 'expand_ratio',
    'input_filters', 'output_filters', 'se_ratio', 'id_skip'])

# Set GlobalParams and BlockArgs's defaults
GlobalParams.__new__.__defaults__ = (None,) * len(GlobalParams._fields)
BlockArgs.__new__.__defaults__ = (None,) * len(BlockArgs._fields)


# An ordinary implementation of Swish function
class Swish(nn.Module):
    def forward(self, x):
        return x * torch.sigmoid(x)


# A memory-efficient implementation of Swish function
class SwishImplementation(torch.autograd.Function):
    @staticmethod
    def forward(ctx, i):
        result = i * torch.sigmoid(i)
        ctx.save_for_backward(i)
        return result

    @staticmethod
    def backward(ctx, grad_output):
        i = ctx.saved_tensors[0]
        sigmoid_i = torch.sigmoid(i)
        return grad_output * (sigmoid_i * (1 + i * (1 - sigmoid_i)))


class MemoryEfficientSwish(nn.Module):
    def forward(self, x):
        return SwishImplementation.apply(x)


def round_filters(filters, global_params):
    """Calculate and round number of filters based on width multiplier.
       Use width_coefficient, depth_divisor and min_depth of global_params.

    Args:
        filters (int): Filters number to be calculated.
        global_params (namedtuple): Global params of the model.

    Returns:
        new_filters: New filters number after calculating.
    """
    multiplier = global_params.width_coefficient
    if not multiplier:
        return filters
    # TODO: modify the params names.
    #       maybe the names (width_divisor,min_width)
    #       are more suitable than (depth_divisor,min_depth).
    divisor = global_params.depth_divisor
    min_depth = global_params.min_depth
    filters *= multiplier
    min_depth = min_depth or divisor  # pay attention to this line when using min_depth
    # follow the formula transferred from official TensorFlow implementation
    new_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor)
    if new_filters < 0.9 * filters:  # prevent rounding by more than 10%
        new_filters += divisor
    return int(new_filters)


def round_repeats(repeats, global_params):
    """Calculate module's repeat number of a block based on depth multiplier.
       Use depth_coefficient of global_params.

    Args:
        repeats (int): num_repeat to be calculated.
        global_params (namedtuple): Global params of the model.

    Returns:
        new repeat: New repeat number after calculating.
    """
    multiplier = global_params.depth_coefficient
    if not multiplier:
        return repeats
    # follow the formula transferred from official TensorFlow implementation
    return int(math.ceil(multiplier * repeats))


def drop_connect(inputs, p, training):
    """Drop connect.

    Args:
        input (tensor: BCWH): Input of this structure.
        p (float: 0.0~1.0): Probability of drop connection.
        training (bool): The running mode.

    Returns:
        output: Output after drop connection.
    """
    assert 0 <= p <= 1, 'p must be in range of [0,1]'

    if not training:
        return inputs

    batch_size = inputs.shape[0]
    keep_prob = 1 - p

    # generate binary_tensor mask according to probability (p for 0, 1-p for 1)
    random_tensor = keep_prob
    random_tensor += torch.rand([batch_size, 1, 1, 1], dtype=inputs.dtype, device=inputs.device)
    binary_tensor = torch.floor(random_tensor)

    output = inputs / keep_prob * binary_tensor
    return output


def get_width_and_height_from_size(x):
    """Obtain height and width from x.

    Args:
        x (int, tuple or list): Data size.

    Returns:
        size: A tuple or list (H,W).
    """
    if isinstance(x, int):
        return x, x
    if isinstance(x, list) or isinstance(x, tuple):
        return x
    else:
        raise TypeError()


def calculate_output_image_size(input_image_size, stride):
    """Calculates the output image size when using Conv2dSamePadding with a stride.
       Necessary for static padding. Thanks to mannatsingh for pointing this out.

    Args:
        input_image_size (int, tuple or list): Size of input image.
        stride (int, tuple or list): Conv2d operation's stride.

    Returns:
        output_image_size: A list [H,W].
    """
    if input_image_size is None:
        return None
    image_height, image_width = get_width_and_height_from_size(input_image_size)
    stride = stride if isinstance(stride, int) else stride[0]
    image_height = int(math.ceil(image_height / stride))
    image_width = int(math.ceil(image_width / stride))
    return [image_height, image_width]


def get_same_padding_conv2d(image_size=None):
    """Chooses static padding if you have specified an image size, and dynamic padding otherwise.
       Static padding is necessary for ONNX exporting of models.

    Args:
        image_size (int or tuple): Size of the image.

    Returns:
        Conv2dDynamicSamePadding or Conv2dStaticSamePadding.
    """
    if image_size is None:
        return Conv2dDynamicSamePadding
    else:
        return partial(Conv2dStaticSamePadding, image_size=image_size)


class Conv2dDynamicSamePadding(nn.Conv2d):
    """2D Convolutions like TensorFlow, for a dynamic image size.
       The padding is operated in forward function by calculating dynamically.
    """

    # Tips for 'SAME' mode padding.
    #     Given the following:
    #         i: width or height
    #         s: stride
    #         k: kernel size
    #         d: dilation
    #         p: padding
    #     Output after Conv2d:
    #         o = floor((i+p-((k-1)*d+1))/s+1)
    # If o equals i, i = floor((i+p-((k-1)*d+1))/s+1),
    # => p = (i-1)*s+((k-1)*d+1)-i

    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, groups=1, bias=True):
        super().__init__(in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias)
        self.stride = self.stride if len(self.stride) == 2 else [self.stride[0]] * 2

    def forward(self, x):
        ih, iw = x.size()[-2:]
        kh, kw = self.weight.size()[-2:]
        sh, sw = self.stride
        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)  # change the output size according to stride ! ! !
        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
        if pad_h > 0 or pad_w > 0:
            x = F.pad(x, [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2])
        return F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)


class Conv2dStaticSamePadding(nn.Conv2d):
    """2D Convolutions like TensorFlow's 'SAME' mode, with the given input image size.
       The padding mudule is calculated in construction function, then used in forward.
    """

    # With the same calculation as Conv2dDynamicSamePadding

    def __init__(self, in_channels, out_channels, kernel_size, stride=1, image_size=None, **kwargs):
        super().__init__(in_channels, out_channels, kernel_size, stride, **kwargs)
        self.stride = self.stride if len(self.stride) == 2 else [self.stride[0]] * 2

        # Calculate padding based on image size and save it
        assert image_size is not None
        ih, iw = (image_size, image_size) if isinstance(image_size, int) else image_size
        kh, kw = self.weight.size()[-2:]
        sh, sw = self.stride
        oh, ow = math.ceil(ih / sh), math.ceil(iw / sw)
        pad_h = max((oh - 1) * self.stride[0] + (kh - 1) * self.dilation[0] + 1 - ih, 0)
        pad_w = max((ow - 1) * self.stride[1] + (kw - 1) * self.dilation[1] + 1 - iw, 0)
        if pad_h > 0 or pad_w > 0:
            self.static_padding = nn.ZeroPad2d((pad_w // 2, pad_w - pad_w // 2,
                                                pad_h // 2, pad_h - pad_h // 2))
        else:
            self.static_padding = nn.Identity()

    def forward(self, x):
        x = self.static_padding(x)
        x = F.conv2d(x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups)
        return x


class BlockDecoder(object):
    """Block Decoder for readability,
       straight from the official TensorFlow repository.
    """

    @staticmethod
    def _decode_block_string(block_string):
        """Get a block through a string notation of arguments.

        Args:
            block_string (str): A string notation of arguments.
                                Examples: 'r1_k3_s11_e1_i32_o16_se0.25_noskip'.

        Returns:
            BlockArgs: The namedtuple defined at the top of this file.
        """
        assert isinstance(block_string, str)

        ops = block_string.split('_')
        options = {}
        for op in ops:
            splits = re.split(r'(\d.*)', op)
            if len(splits) >= 2:
                key, value = splits[:2]
                options[key] = value

        # Check stride
        assert (('s' in options and len(options['s']) == 1) or
                (len(options['s']) == 2 and options['s'][0] == options['s'][1]))

        return BlockArgs(
            num_repeat=int(options['r']),
            kernel_size=int(options['k']),
            stride=[int(options['s'][0])],
            expand_ratio=int(options['e']),
            input_filters=int(options['i']),
            output_filters=int(options['o']),
            se_ratio=float(options['se']) if 'se' in options else None,
            id_skip=('noskip' not in block_string))

    @staticmethod
    def _encode_block_string(block):
        """Encode a block to a string.

        Args:
            block (namedtuple): A BlockArgs type argument.

        Returns:
            block_string: A String form of BlockArgs.
        """
        args = [
            'r%d' % block.num_repeat,
            'k%d' % block.kernel_size,
            's%d%d' % (block.strides[0], block.strides[1]),
            'e%s' % block.expand_ratio,
            'i%d' % block.input_filters,
            'o%d' % block.output_filters
        ]
        if 0 < block.se_ratio <= 1:
            args.append('se%s' % block.se_ratio)
        if block.id_skip is False:
            args.append('noskip')
        return '_'.join(args)

    @staticmethod
    def decode(string_list):
        """Decode a list of string notations to specify blocks inside the network.

        Args:
            string_list (list[str]): A list of strings, each string is a notation of block.

        Returns:
            blocks_args: A list of BlockArgs namedtuples of block args.
        """
        assert isinstance(string_list, list)
        blocks_args = []
        for block_string in string_list:
            blocks_args.append(BlockDecoder._decode_block_string(block_string))
        return blocks_args

    @staticmethod
    def encode(blocks_args):
        """Encode a list of BlockArgs to a list of strings.

        Args:
            blocks_args (list[namedtuples]): A list of BlockArgs namedtuples of block args.

        Returns:
            block_strings: A list of strings, each string is a notation of block.
        """
        block_strings = []
        for block in blocks_args:
            block_strings.append(BlockDecoder._encode_block_string(block))
        return block_strings


def efficientnet_params(model_name):
    """Map EfficientNet model name to parameter coefficients.

    Args:
        model_name (str): Model name to be queried.

    Returns:
        params_dict[model_name]: A (width,depth,res,dropout) tuple.
    """
    """
    params_dict = {
        # Coefficients:   width,depth,res,dropout
        'efficientnet-b0': (1.0, 1.0, 224, 0.2),
        'efficientnet-b1': (1.0, 1.1, 240, 0.2),
        'efficientnet-b2': (1.1, 1.2, 260, 0.3),
        'efficientnet-b3': (1.2, 1.4, 300, 0.3),
        'efficientnet-b4': (1.4, 1.8, 380, 0.4),
        'efficientnet-b5': (1.6, 2.2, 456, 0.4),
        'efficientnet-b6': (1.8, 2.6, 528, 0.5),
        'efficientnet-b7': (2.0, 3.1, 600, 0.5),
        'efficientnet-b8': (2.2, 3.6, 672, 0.5),
        'efficientnet-l2': (4.3, 5.3, 800, 0.5),
    }
    """
    params_dict = {
        # Coefficients:   width,depth,res,dropout
        'efficientnet-b0': (1.0, 1.0, 112, 0.2),
        'efficientnet-b1': (1.0, 1.1, 112, 0.2),
        'efficientnet-b2': (1.1, 1.2, 112, 0.3),
        'efficientnet-b3': (1.2, 1.4, 112, 0.3),
        'efficientnet-b4': (1.4, 1.8, 112, 0.4),
        'efficientnet-b5': (1.6, 2.2, 112, 0.4),
        'efficientnet-b6': (1.8, 2.6, 112, 0.5),
        'efficientnet-b7': (2.0, 3.1, 112, 0.5),
        'efficientnet-b8': (2.2, 3.6, 112, 0.5),
        'efficientnet-l2': (4.3, 5.3, 112, 0.5),
    }
    return params_dict[model_name]


def efficientnet(width_coefficient=None, depth_coefficient=None, image_size=None,
                 dropout_rate=0.2, drop_connect_rate=0.2, num_classes=1000, include_top=True):
    """Create BlockArgs and GlobalParams for efficientnet model.

    Args:
        width_coefficient (float)
        depth_coefficient (float)
        image_size (int)
        dropout_rate (float)
        drop_connect_rate (float)
        num_classes (int)

        Meaning as the name suggests.

    Returns:
        blocks_args, global_params.
    """

    # Blocks args for the whole model(efficientnet-b0 by default)
    # It will be modified in the construction of EfficientNet Class according to model
    blocks_args = [
        'r1_k3_s11_e1_i32_o16_se0.25',
        'r2_k3_s22_e6_i16_o24_se0.25',
        'r2_k5_s22_e6_i24_o40_se0.25',
        'r3_k3_s22_e6_i40_o80_se0.25',
        'r3_k5_s11_e6_i80_o112_se0.25',
        'r4_k5_s22_e6_i112_o192_se0.25',
        'r1_k3_s11_e6_i192_o320_se0.25',
    ]
    blocks_args = BlockDecoder.decode(blocks_args)

    global_params = GlobalParams(
        width_coefficient=width_coefficient,
        depth_coefficient=depth_coefficient,
        image_size=image_size,
        dropout_rate=dropout_rate,

        num_classes=num_classes,
        batch_norm_momentum=0.99,
        batch_norm_epsilon=1e-3,
        drop_connect_rate=drop_connect_rate,
        depth_divisor=8,
        min_depth=None,
        include_top=include_top,
    )

    return blocks_args, global_params


def get_model_params(model_name, override_params):
    """Get the block args and global params for a given model name.

    Args:
        model_name (str): Model's name.
        override_params (dict): A dict to modify global_params.

    Returns:
        blocks_args, global_params
    """
    if model_name.startswith('efficientnet'):
        w, d, s, p = efficientnet_params(model_name)
        # note: all models have drop connect rate = 0.2
        blocks_args, global_params = efficientnet(
            width_coefficient=w, depth_coefficient=d, dropout_rate=p, image_size=s)
    else:
        raise NotImplementedError('model name is not pre-defined: {}'.format(model_name))
    if override_params:
        # ValueError will be raised here if override_params has fields not included in global_params.
        global_params = global_params._replace(**override_params)
    return blocks_args, global_params


class MBConvBlock(nn.Module):
    """Mobile Inverted Residual Bottleneck Block.

    Args:
        block_args (namedtuple): BlockArgs, defined in utils.py.
        global_params (namedtuple): GlobalParam, defined in utils.py.
        image_size (tuple or list): [image_height, image_width].

    References:
        [1] https://arxiv.org/abs/1704.04861 (MobileNet v1)
        [2] https://arxiv.org/abs/1801.04381 (MobileNet v2)
        [3] https://arxiv.org/abs/1905.02244 (MobileNet v3)
    """

    def __init__(self, block_args, global_params, image_size=None):
        super().__init__()
        self._block_args = block_args
        self._bn_mom = 1 - global_params.batch_norm_momentum  # pytorch's difference from tensorflow
        self._bn_eps = global_params.batch_norm_epsilon
        self.has_se = (self._block_args.se_ratio is not None) and (0 < self._block_args.se_ratio <= 1)
        self.id_skip = block_args.id_skip  # whether to use skip connection and drop connect

        # Expansion phase (Inverted Bottleneck)
        inp = self._block_args.input_filters  # number of input channels
        oup = self._block_args.input_filters * self._block_args.expand_ratio  # number of output channels
        if self._block_args.expand_ratio != 1:
            Conv2d = get_same_padding_conv2d(image_size=image_size)
            self._expand_conv = Conv2d(in_channels=inp, out_channels=oup, kernel_size=1, bias=False)
            self._bn0 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
            # image_size = calculate_output_image_size(image_size, 1) <-- this wouldn't modify image_size

        # Depthwise convolution phase
        k = self._block_args.kernel_size
        s = self._block_args.stride
        Conv2d = get_same_padding_conv2d(image_size=image_size)
        self._depthwise_conv = Conv2d(
            in_channels=oup, out_channels=oup, groups=oup,  # groups makes it depthwise
            kernel_size=k, stride=s, bias=False)
        self._bn1 = nn.BatchNorm2d(num_features=oup, momentum=self._bn_mom, eps=self._bn_eps)
        image_size = calculate_output_image_size(image_size, s)

        # Squeeze and Excitation layer, if desired
        if self.has_se:
            Conv2d = get_same_padding_conv2d(image_size=(1, 1))
            num_squeezed_channels = max(1, int(self._block_args.input_filters * self._block_args.se_ratio))
            self._se_reduce = Conv2d(in_channels=oup, out_channels=num_squeezed_channels, kernel_size=1)
            self._se_expand = Conv2d(in_channels=num_squeezed_channels, out_channels=oup, kernel_size=1)

        # Pointwise convolution phase
        final_oup = self._block_args.output_filters
        Conv2d = get_same_padding_conv2d(image_size=image_size)
        self._project_conv = Conv2d(in_channels=oup, out_channels=final_oup, kernel_size=1, bias=False)
        self._bn2 = nn.BatchNorm2d(num_features=final_oup, momentum=self._bn_mom, eps=self._bn_eps)
        self._swish = MemoryEfficientSwish()

    def forward(self, inputs, drop_connect_rate=None):
        """MBConvBlock's forward function.

        Args:
            inputs (tensor): Input tensor.
            drop_connect_rate (bool): Drop connect rate (float, between 0 and 1).

        Returns:
            Output of this block after processing.
        """

        # Expansion and Depthwise Convolution
        x = inputs
        if self._block_args.expand_ratio != 1:
            x = self._expand_conv(inputs)
            x = self._bn0(x)
            x = self._swish(x)

        x = self._depthwise_conv(x)
        x = self._bn1(x)
        x = self._swish(x)

        # Squeeze and Excitation
        if self.has_se:
            x_squeezed = F.adaptive_avg_pool2d(x, 1)
            x_squeezed = self._se_reduce(x_squeezed)
            x_squeezed = self._swish(x_squeezed)
            x_squeezed = self._se_expand(x_squeezed)
            x = torch.sigmoid(x_squeezed) * x

        # Pointwise Convolution
        x = self._project_conv(x)
        x = self._bn2(x)

        # Skip connection and drop connect
        input_filters, output_filters = self._block_args.input_filters, self._block_args.output_filters
        if self.id_skip and self._block_args.stride == 1 and input_filters == output_filters:
            # The combination of skip connection and drop connect brings about stochastic depth.
            if drop_connect_rate:
                x = drop_connect(x, p=drop_connect_rate, training=self.training)
            x = x + inputs  # skip connection
        return x

    def set_swish(self, memory_efficient=True):
        """Sets swish function as memory efficient (for training) or standard (for export).

        Args:
            memory_efficient (bool): Whether to use memory-efficient version of swish.
        """
        self._swish = MemoryEfficientSwish() if memory_efficient else Swish()


class EfficientNet(nn.Module):
    def __init__(self, out_h=7, out_w=7, feat_dim=512, blocks_args=None, global_params=None):
        super().__init__()
        assert isinstance(blocks_args, list), 'blocks_args should be a list'
        assert len(blocks_args) > 0, 'block args must be greater than 0'
        self._global_params = global_params
        self._blocks_args = blocks_args

        # Batch norm parameters
        bn_mom = 1 - self._global_params.batch_norm_momentum
        bn_eps = self._global_params.batch_norm_epsilon

        # Get stem static or dynamic convolution depending on image size
        image_size = global_params.image_size
        Conv2d = get_same_padding_conv2d(image_size=image_size)

        # Stem
        in_channels = 3  # rgb
        out_channels = round_filters(32, self._global_params)  # number of output channels
        self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=1, bias=False)
        self._bn0 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)
        image_size = calculate_output_image_size(image_size, 2)

        # Build blocks
        self._blocks = nn.ModuleList([])
        for block_args in self._blocks_args:

            # Update block input and output filters based on depth multiplier.
            block_args = block_args._replace(
                input_filters=round_filters(block_args.input_filters, self._global_params),
                output_filters=round_filters(block_args.output_filters, self._global_params),
                num_repeat=round_repeats(block_args.num_repeat, self._global_params)
            )

            # The first block needs to take care of stride and filter size increase.
            self._blocks.append(MBConvBlock(block_args, self._global_params, image_size=image_size))
            image_size = calculate_output_image_size(image_size, block_args.stride)
            if block_args.num_repeat > 1:  # modify block_args to keep same output size
                block_args = block_args._replace(input_filters=block_args.output_filters, stride=1)
            for _ in range(block_args.num_repeat - 1):
                self._blocks.append(MBConvBlock(block_args, self._global_params, image_size=image_size))
                # image_size = calculate_output_image_size(image_size, block_args.stride)  # stride = 1

        # Head
        in_channels = block_args.output_filters  # output of final block
        out_channels = round_filters(1280, self._global_params)
        Conv2d = get_same_padding_conv2d(image_size=image_size)
        self._conv_head = Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        self._bn1 = nn.BatchNorm2d(num_features=out_channels, momentum=bn_mom, eps=bn_eps)

        # Final linear layer
        self._avg_pooling = nn.AdaptiveAvgPool2d(1)
        self._dropout = nn.Dropout(self._global_params.dropout_rate)
        self._fc = nn.Linear(out_channels, self._global_params.num_classes)
        self._swish = MemoryEfficientSwish()

    def set_swish(self, memory_efficient=True):
        """Sets swish function as memory efficient (for training) or standard (for export).

        Args:
            memory_efficient (bool): Whether to use memory-efficient version of swish.

        """
        self._swish = MemoryEfficientSwish() if memory_efficient else Swish()
        for block in self._blocks:
            block.set_swish(memory_efficient)

    @classmethod
    def my_encoder(cls, model_name='efficientnet-b0'):
        model = cls.from_name(model_name)

        class Encoder(nn.Module):
            def __init__(self):
                super().__init__()

                self.stem_conv = model._conv_stem
                self.stem_batch_norm = model._bn0
                self.stem_swish = Swish()
                self.blocks = model._blocks

            def forward(self, x):
                # Stem
                x = self.stem_conv(x)
                x = self.stem_batch_norm(x)
                x = self.stem_swish(x)

                # Blocks
                for idx, block in enumerate(self.blocks):
                    drop_connect_rate = 0.2
                    if drop_connect_rate:
                        drop_connect_rate *= idx / len(self.blocks)
                    x = block(x, drop_connect_rate)

                return x

        return Encoder()

    @classmethod
    def from_name(cls, model_name, in_channels=3, **override_params):
        """create an efficientnet model according to name.

        Args:
            model_name (str): Name for efficientnet.
            in_channels (int): Input data's channel number.
            override_params (other key word params):
                Params to override model's global_params.
                Optional key:
                    'width_coefficient', 'depth_coefficient',
                    'image_size', 'dropout_rate',
                    'num_classes', 'batch_norm_momentum',
                    'batch_norm_epsilon', 'drop_connect_rate',
                    'depth_divisor', 'min_depth'

        Returns:
            An efficientnet model.
        """
        blocks_args, global_params = get_model_params(model_name, override_params)
        model = cls(blocks_args=blocks_args, global_params=global_params)
        model._change_in_channels(in_channels)
        return model

    def _change_in_channels(self, in_channels):
        """Adjust model's first convolution layer to in_channels, if in_channels not equals 3.

        Args:
            in_channels (int): Input data's channel number.
        """
        if in_channels != 3:
            Conv2d = get_same_padding_conv2d(image_size=self._global_params.image_size)
            out_channels = round_filters(32, self._global_params)
            self._conv_stem = Conv2d(in_channels, out_channels, kernel_size=3, stride=2, bias=False)


class AbsWeighting(nn.Module):
    def __init__(self):
        super(AbsWeighting, self).__init__()
        self.rep_grad = True
        self.epoch = 0

    def init_param(self):
        r"""Define and initialize some trainable parameters required by specific weighting methods.
        """
        self.train_loss_buffer = np.zeros([self.task_num, 9999])
        self.rep_tasks = {}
        self.rep = {}
        self.grads_buffer = [None] * self.task_num

    def update_epoch(self):
        self.epoch += 1

    def _compute_grad_dim(self):
        self.grad_index = []
        for param in self.get_share_params():
            self.grad_index.append(param.data.numel())
        self.grad_dim = sum(self.grad_index)

    def _grad2vec(self):
        grad = torch.zeros(self.grad_dim)
        count = 0
        for param in self.get_share_params():
            if param.grad is not None:
                beg = 0 if count == 0 else sum(self.grad_index[:count])
                end = sum(self.grad_index[:(count + 1)])
                grad[beg:end] = param.grad.data.view(-1)
            count += 1
        return grad

    def _compute_grad(self, losses, mode, rep_grad=False):
        # if not isinstance(self.rep, dict):
        #     grads = torch.zeros(self.task_num, *self.rep.size()).to(self.device)
        # else:
        grads = [torch.zeros(*self.rep[task].size()) for task in self.task_name]
        for tn, task in enumerate(self.task_name):
            ret_graph = True if (tn + 1) != self.task_num else False
            losses[tn].backward(retain_graph=ret_graph)
            grads[tn] = self.rep_tasks[task].grad.data.clone()
            # grads[tn] = torch.mean(self.rep_tasks[task].grad.data.clone(), dim=0)
        return grads

    def _reset_grad(self, new_grads):
        count = 0
        for param in self.get_share_params():
            if param.grad is not None:
                beg = 0 if count == 0 else sum(self.grad_index[:count])
                end = sum(self.grad_index[:(count + 1)])
                param.grad.data = new_grads[beg:end].contiguous().view(param.data.size()).data.clone()
            count += 1

    def _get_grads(self, losses, mode='backward'):
        r"""This function is used to return the gradients of representations or shared parameters.

        If ``rep_grad`` is ``True``, it returns a list with two elements. The first element is \
        the gradients of the representations with the size of [task_num, batch_size, rep_size]. \
        The second element is the resized gradients with size of [task_num, -1], which means \
        the gradient of each task is resized as a vector.

        If ``rep_grad`` is ``False``, it returns the gradients of the shared parameters with size \
        of [task_num, -1], which means the gradient of each task is resized as a vector.
        """
        if self.rep_grad:
            per_grads = self._compute_grad(losses, mode, rep_grad=True)
            if not isinstance(self.rep, dict):
                grads = per_grads.reshape(self.task_num, self.rep.size()[0], -1).sum(1)
            else:
                try:
                    grads = torch.stack(per_grads).sum(1).view(self.task_num, -1)
                except:
                    raise ValueError('The representation dimensions of different tasks must be consistent')
            return [per_grads, grads]
        else:
            self._compute_grad_dim()
            grads = self._compute_grad(losses, mode)
            return grads

    def _backward_new_grads(self, batch_weight, per_grads=None, grads=None):
        r"""This function is used to reset the gradients and make a backward.

        Args:
            batch_weight (torch.Tensor): A tensor with size of [task_num].
            per_grad (torch.Tensor): It is needed if ``rep_grad`` is True. The gradients of the representations.
            grads (torch.Tensor): It is needed if ``rep_grad`` is False. The gradients of the shared parameters.
        """
        if self.rep_grad:
            if not isinstance(self.rep, dict):
                transformed_grad = torch.einsum('i, i... -> ...', batch_weight, per_grads)
                self.rep.backward(transformed_grad)
            else:
                for tn, task in enumerate(self.task_name):
                    rg = True if (tn + 1) != self.task_num else False
                    self.rep[task].backward(batch_weight[tn] * per_grads[tn], retain_graph=rg)
        else:
            new_grads = torch.einsum('i, i... -> ...', batch_weight, grads)
            self._reset_grad(new_grads)


class MomentumMask(AbsWeighting):
    def __init__(self):
        super(MomentumMask, self).__init__()
        self.rep_grad = True

    def backward(self, losses, idx_group):
        beta = 0.95

        per_grads = self._compute_grad(losses, mode='backward', rep_grad=True)
        per_grads = torch.cat(per_grads)

        inputs = torch.cat(list(self.rep.values()))

        group_len = [0]
        for gop in idx_group:
            group_len.append(len(gop) + group_len[-1])

        grads = []
        tmp_grads = (per_grads * inputs.sign()).detach()
        for i in range(1, len(group_len)):
            cur_grads = tmp_grads[group_len[i - 1]:group_len[i]].sum(0)
            if self.grads_buffer[i - 1] is not None:
                cur_grads = beta * self.grads_buffer[i - 1] + (1 - beta) * cur_grads
            self.grads_buffer[i - 1] = cur_grads
            grads.append(cur_grads)
        grads = torch.stack(grads)

        P = 0.5 * (1 + grads.sum(0) / (grads.abs().sum(0) + 1e-7))
        U = torch.rand_like(P)
        M = P.gt(U).unsqueeze(0).repeat_interleave(self.task_num, dim=0) * grads.gt(0) + \
            P.lt(U).unsqueeze(0).repeat_interleave(self.task_num, dim=0) * grads.lt(0)
        transformed_grad = []
        for i in range(1, len(group_len)):
            cur_per_grads = per_grads[group_len[i - 1]:group_len[i]]
            cur_M = M[i - 1].unsqueeze(0).repeat_interleave(cur_per_grads.size()[0], dim=0)
            cur_M = cur_M.detach()
            transformed_grad.append(cur_per_grads * cur_M)

        for tn, task in enumerate(self.task_name):
            ret_graph = True if tn != len(self.task_name) - 1 else False
            self.rep[task].backward(transformed_grad[tn], retain_graph=ret_graph)
        return None


class STATE(MomentumMask):
    def __init__(self, num_class=29, n_fc=0):
        super().__init__()
        self.baseline_extractor = EfficientNet.my_encoder('efficientnet-b0')
        self.output_layer = Sequential(
            nn.AdaptiveAvgPool2d(1),
            nn.Flatten(),
            nn.Linear(320 * 1 * 1, 256),
            BatchNorm1d(256),
        )

        self.invTrans = transforms.Compose(
            [transforms.Normalize(mean=[0., 0., 0.], std=[1 / 0.229, 1 / 0.224, 1 / 0.225]),
             transforms.Normalize(mean=[-0.485, -0.456, -0.406], std=[1., 1., 1.])])
        dim = 320
        self.norm2 = nn.LayerNorm(dim)
        self.num_heads = 10
        self.proj_qv = nn.Linear(dim, dim * 2, bias=True)
        self.proj_k = nn.Linear(dim, dim, bias=True)
        self.proj_final = nn.Linear(dim, dim)
        self.softmax = nn.Softmax(dim=-1)
        self.mlp = nn.Sequential(
            nn.Linear(dim, dim * 2),
            nn.GELU(),
            nn.Linear(dim * 2, dim)
        )

        self.n_fc = n_fc
        for i in range(self.n_fc):
            setattr(self, 'fc{}'.format(i), torch.nn.Linear(256, num_class))

        self.task_name = ['fc%d' % idx for idx in range(self.n_fc)]
        self.task_num = len(self.task_name)
        self.init_param()

    def smooth_attention(self, rgb, feat, nrows=16, ncols=16):
        pre_r, pre_h = 4, 4
        nrows, ncols = nrows // pre_r, ncols // pre_h
        x = self.invTrans(rgb)
        x = x * 255
        x = x.long()
        x = rgb_to_grayscale(x)
        b, c, r, h = x.shape
        x = F.interpolate(x.double(), (r // pre_r, h // pre_h)).long()
        x.clamp_(0, 255)

        b, c, r, h = x.shape
        x = x.reshape(b, h // nrows, nrows, -1, ncols).transpose(2, 3).reshape(b, -1, nrows * ncols)
        p = torch.nn.functional.one_hot(x, 256).sum(dim=[2])
        p = p / p.sum(dim=2).unsqueeze(2)
        ent = -torch.where(p > 0, p * p.log(), p.new([0.0])).sum(dim=2)

        b, c, h, w = feat.shape
        ent = F.normalize(ent)
        ent = ent.unsqueeze(2)
        ent = torch.cat(c * [ent], dim=2)

        feat = feat.permute(0, 2, 3, 1).contiguous().view(b, h * w, c)
        shortcut = feat
        qv = self.proj_qv(feat).reshape(b, h * w, 2, self.num_heads, c // self.num_heads).permute(2, 0, 3, 1, 4)
        q, v = qv[0], qv[1]

        b, n, c = ent.shape
        k = self.proj_k(ent).reshape(b, n, 1, self.num_heads, c // self.num_heads).permute(2, 0, 3, 1, 4)

        attn = (q @ k.transpose(-2, -1))
        attn = self.softmax(attn)
        feat = (attn @ v).transpose(1, 2).reshape(b, h * w, c)
        feat = self.proj_final(feat)
        feat = shortcut + feat
        feat = feat + self.mlp(self.norm2(feat))
        feat = feat.view(b, h, w, c).permute(0, 3, 1, 2).contiguous()
        return feat, attn

    def get_share_params(self):
        r"""Return the shared parameters of the model.
        """
        return self.output_layer.parameters()

    def forward(self, x):
        rgb = x.clone()
        x_f = self.baseline_extractor(x)
        x_a, _ = self.smooth_attention(rgb, x_f)
        trace = self.output_layer(x_a)
        return trace

    def _prepare_rep(self, rep, idx, same_rep=False):
        task = self.task_name[idx]
        self.rep[task] = rep
        self.rep_tasks[task] = rep.detach().clone()
        self.rep_tasks[task].requires_grad = True
        return self.rep_tasks[task]

    def fc_forward(self, feat, idx):
        return getattr(self, 'fc{}'.format(idx))(feat)