S-Rajesh/triqa-iqa

1# Copyright (c) Meta Platforms, Inc. and affiliates.2 3# All rights reserved.4 5# This source code is licensed under the license found in the6# LICENSE file in the root directory of this source tree.7 8 9import torch10import torch.nn as nn11import torch.nn.functional as F12from timm.models.layers import trunc_normal_, DropPath13# from timm.models.registry import register_model14 15class Block(nn.Module):16    r""" ConvNeXt Block. There are two equivalent implementations:17    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)18    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back19    We use (2) as we find it slightly faster in PyTorch20    21    Args:22        dim (int): Number of input channels.23        drop_path (float): Stochastic depth rate. Default: 0.024        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.25    """26    def __init__(self, dim, drop_path=0., layer_scale_init_value=1e-6):27        super().__init__()28        self.dwconv = nn.Conv2d(dim, dim, kernel_size=7, padding=3, groups=dim) # depthwise conv29        self.norm = LayerNorm(dim, eps=1e-6)30        self.pwconv1 = nn.Linear(dim, 4 * dim) # pointwise/1x1 convs, implemented with linear layers31        self.act = nn.GELU()32        self.pwconv2 = nn.Linear(4 * dim, dim)33        self.gamma = nn.Parameter(layer_scale_init_value * torch.ones((dim)), 34                                    requires_grad=True) if layer_scale_init_value > 0 else None35        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()36 37    def forward(self, x):38        input = x39        x = self.dwconv(x)40        x = x.permute(0, 2, 3, 1) # (N, C, H, W) -> (N, H, W, C)41        x = self.norm(x)42        x = self.pwconv1(x)43        x = self.act(x)44        x = self.pwconv2(x)45        if self.gamma is not None:46            x = self.gamma * x47        x = x.permute(0, 3, 1, 2) # (N, H, W, C) -> (N, C, H, W)48 49        x = input + self.drop_path(x)50        return x51 52class ConvNeXt(nn.Module):53    r""" ConvNeXt54        A PyTorch impl of : `A ConvNet for the 2020s`  -55          https://arxiv.org/pdf/2201.03545.pdf56 57    Args:58        in_chans (int): Number of input image channels. Default: 359        num_classes (int): Number of classes for classification head. Default: 100060        depths (tuple(int)): Number of blocks at each stage. Default: [3, 3, 9, 3]61        dims (int): Feature dimension at each stage. Default: [96, 192, 384, 768]62        drop_path_rate (float): Stochastic depth rate. Default: 0.63        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.64        head_init_scale (float): Init scaling value for classifier weights and biases. Default: 1.65    """66    def __init__(self, in_chans=3, num_classes=128, 67                 depths=[3, 3, 9, 3], dims=[96, 192, 384, 768], drop_path_rate=0., 68                 layer_scale_init_value=1e-6, head_init_scale=1.,69                 ):70        super().__init__()71 72        self.downsample_layers = nn.ModuleList() # stem and 3 intermediate downsampling conv layers73        stem = nn.Sequential(74            nn.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),75            LayerNorm(dims[0], eps=1e-6, data_format="channels_first")76        )77        self.downsample_layers.append(stem)78        for i in range(3):79            downsample_layer = nn.Sequential(80                    LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),81                    nn.Conv2d(dims[i], dims[i+1], kernel_size=2, stride=2),82            )83            self.downsample_layers.append(downsample_layer)84 85        self.stages = nn.ModuleList() # 4 feature resolution stages, each consisting of multiple residual blocks86        dp_rates=[x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))] 87        cur = 088        for i in range(4):89            stage = nn.Sequential(90                *[Block(dim=dims[i], drop_path=dp_rates[cur + j], 91                layer_scale_init_value=layer_scale_init_value) for j in range(depths[i])]92            )93            self.stages.append(stage)94            cur += depths[i]95 96        self.norm = nn.LayerNorm(dims[-1], eps=1e-6) # final norm layer97        self.head = nn.Linear(dims[-1], num_classes)98 99        self.apply(self._init_weights)100        self.head.weight.data.mul_(head_init_scale)101        self.head.bias.data.mul_(head_init_scale)102 103    def _init_weights(self, m):104        if isinstance(m, (nn.Conv2d, nn.Linear)):105            trunc_normal_(m.weight, std=.02)106            nn.init.constant_(m.bias, 0)107 108    def forward_features(self, x):109        for i in range(4):110            x = self.downsample_layers[i](x)111            x = self.stages[i](x)112        return self.norm(x.mean([-2, -1])) # global average pooling, (N, C, H, W) -> (N, C)113 114    def forward(self, x1):115    # def forward(self, x1,x2,x3):116        x1 = self.forward_features(x1)117        x1 = self.head(x1)118 119        # x2 = self.forward_features(x2)120        # x2 = self.head(x2)121 122        # x3 = self.forward_features(x3)123        # x3 = self.head(x3)124        return x1125        # return x1,x2,x3126 127class LayerNorm(nn.Module):128    r""" LayerNorm that supports two data formats: channels_last (default) or channels_first. 129    The ordering of the dimensions in the inputs. channels_last corresponds to inputs with 130    shape (batch_size, height, width, channels) while channels_first corresponds to inputs 131    with shape (batch_size, channels, height, width).132    """133    def __init__(self, normalized_shape, eps=1e-6, data_format="channels_last"):134        super().__init__()135        self.weight = nn.Parameter(torch.ones(normalized_shape))136        self.bias = nn.Parameter(torch.zeros(normalized_shape))137        self.eps = eps138        self.data_format = data_format139        if self.data_format not in ["channels_last", "channels_first"]:140            raise NotImplementedError 141        self.normalized_shape = (normalized_shape, )142    143    def forward(self, x):144        if self.data_format == "channels_last":145            return F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)146        elif self.data_format == "channels_first":147            u = x.mean(1, keepdim=True)148            s = (x - u).pow(2).mean(1, keepdim=True)149            x = (x - u) / torch.sqrt(s + self.eps)150            x = self.weight[:, None, None] * x + self.bias[:, None, None]151            return x152 153 154model_urls = {155    "convnext_tiny_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_tiny_1k_224_ema.pth",156    "convnext_small_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_small_1k_224_ema.pth",157    "convnext_base_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_base_1k_224_ema.pth",158    "convnext_large_1k": "https://dl.fbaipublicfiles.com/convnext/convnext_large_1k_224_ema.pth",159    "convnext_tiny_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_tiny_22k_224.pth",160    "convnext_small_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_small_22k_224.pth",161    "convnext_base_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_base_22k_224.pth",162    "convnext_large_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_large_22k_224.pth",163    "convnext_xlarge_22k": "https://dl.fbaipublicfiles.com/convnext/convnext_xlarge_22k_224.pth",164}