{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Setups","metadata":{}},{"cell_type":"code","source":"import sys\nsys.path.append(\"../input/pytorchimagemodels0412/pytorch-image-models-0.4.12\")\n\nimport timm\nprint(timm.__version__)\n","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:05:23.075573Z","iopub.execute_input":"2022-09-18T12:05:23.075961Z","iopub.status.idle":"2022-09-18T12:05:27.711013Z","shell.execute_reply.started":"2022-09-18T12:05:23.075856Z","shell.execute_reply":"2022-09-18T12:05:27.709089Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import math\nfrom functools import partial\nfrom pathlib import Path \n\nimport albumentations as A\nfrom albumentations.pytorch import ToTensorV2\nimport cv2\nimport numpy as np\nimport pandas as pd\nfrom sklearn.model_selection import StratifiedKFold\nfrom timm.models.layers import DropPath, to_2tuple, trunc_normal_\nimport torch\nimport torch.cuda.amp as amp\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import Dataset, DataLoader\n\n# plot\nimport matplotlib.pyplot as plt","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:05:27.714222Z","iopub.execute_input":"2022-09-18T12:05:27.715282Z","iopub.status.idle":"2022-09-18T12:05:29.326144Z","shell.execute_reply.started":"2022-09-18T12:05:27.715249Z","shell.execute_reply":"2022-09-18T12:05:29.324425Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Configs","metadata":{}},{"cell_type":"code","source":"HUBMAP_ONLY = False\nWEIGHTS = [\n    '../input/hubhpa-train-sf-mitb2/model_best_loss_0.pth',\n]\nN_FOLDS = 4\nIMG_SIZE = 768\nSEED = 43\nDEVICE = \"cuda\" if torch.cuda.is_available() else \"cpu\"","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:05:47.597284Z","iopub.execute_input":"2022-09-18T12:05:47.597646Z","iopub.status.idle":"2022-09-18T12:05:47.603580Z","shell.execute_reply.started":"2022-09-18T12:05:47.597616Z","shell.execute_reply":"2022-09-18T12:05:47.602085Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Utils","metadata":{}},{"cell_type":"code","source":"def rle_encode_less_memory(img):\n    #the image should be transposed\n    pixels = img.T.flatten()\n    \n    # This simplified method requires first and last pixel to be zero\n    pixels[0] = 0\n    pixels[-1] = 0\n    runs = np.where(pixels[1:] != pixels[:-1])[0] + 2\n    runs[1::2] -= runs[::2]\n    \n    return ' '.join(str(x) for x in runs)","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:05:29.396984Z","iopub.execute_input":"2022-09-18T12:05:29.397569Z","iopub.status.idle":"2022-09-18T12:05:29.405897Z","shell.execute_reply.started":"2022-09-18T12:05:29.397531Z","shell.execute_reply":"2022-09-18T12:05:29.405085Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Dataset","metadata":{}},{"cell_type":"code","source":"class HubmapHpaDataset(Dataset):\n    def __init__(\n        self,\n        df: pd.DataFrame, \n        img_path: Path,\n        return_class: bool = False\n    ):\n        self.df = df\n        self.img_path = img_path\n        self.transform = A.Compose(\n            [\n                A.Resize(IMG_SIZE, IMG_SIZE),\n                A.CenterCrop(IMG_SIZE, IMG_SIZE),\n                A.Normalize(\n                    mean=[0.82784054, 0.80224235, 0.8201268],\n                    std=[0.16691974, 0.19275635, 0.17306973],\n                ),\n                ToTensorV2(transpose_mask=True)\n            ]\n        )\n        self.return_class = return_class\n\n    def __len__(self):\n        return len(self.df)\n\n    def __getitem__(self, idx):\n        im = self.img_path / f\"{self.df['id'].iloc[idx]}.tiff\"\n        img = cv2.imread(str(im))\n        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n        \n        if self.transform:\n            img = self.transform(image=img)['image']\n            \n        if self.return_class:\n            return img, self.df['organ'].iloc[idx]\n        return img","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:09:16.567737Z","iopub.execute_input":"2022-09-18T12:09:16.568740Z","iopub.status.idle":"2022-09-18T12:09:16.579685Z","shell.execute_reply.started":"2022-09-18T12:09:16.568702Z","shell.execute_reply":"2022-09-18T12:09:16.578628Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Model","metadata":{}},{"cell_type":"markdown","source":"## Mix Transformer","metadata":{}},{"cell_type":"code","source":"class DWConv(nn.Module):\n    def __init__(self, dim=768):\n        super(DWConv, self).__init__()\n        self.dwconv = nn.Conv2d(dim, dim, 3, 1, 1, bias=True, groups=dim)\n    \n    def forward(self, x, H, W):\n        B, N, C = x.shape\n        x = x.transpose(1, 2).view(B, C, H, W)\n        x = self.dwconv(x)\n        x = x.flatten(2).transpose(1, 2)\n        \n        return x\n\n\nclass Mlp(nn.Module):\n    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n        super().__init__()\n        out_features = out_features or in_features\n        hidden_features = hidden_features or in_features\n        self.fc1 = nn.Linear(in_features, hidden_features)\n        self.dwconv = DWConv(hidden_features)\n        self.act = act_layer()\n        self.fc2 = nn.Linear(hidden_features, out_features)\n        self.drop = nn.Dropout(drop)\n        \n        self.apply(self._init_weights)\n    \n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out //= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n    \n    def forward(self, x, H, W):\n        x = self.fc1(x)\n        x = self.dwconv(x, H, W)\n        x = self.act(x)\n        x = self.drop(x)\n        x = self.fc2(x)\n        x = self.drop(x)\n        return x\n\n\nclass Attention(nn.Module):\n    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., sr_ratio=1):\n        super().__init__()\n        assert dim % num_heads == 0, f\"dim {dim} should be divided by num_heads {num_heads}.\"\n        \n        self.dim = dim\n        self.num_heads = num_heads\n        head_dim = dim // num_heads\n        self.scale = qk_scale or head_dim ** -0.5\n        \n        self.q = nn.Linear(dim, dim, bias=qkv_bias)\n        self.kv = nn.Linear(dim, dim * 2, bias=qkv_bias)\n        self.attn_drop = nn.Dropout(attn_drop)\n        self.proj = nn.Linear(dim, dim)\n        self.proj_drop = nn.Dropout(proj_drop)\n        \n        self.sr_ratio = sr_ratio\n        if sr_ratio > 1:\n            self.sr = nn.Conv2d(dim, dim, kernel_size=sr_ratio, stride=sr_ratio)\n            self.norm = nn.LayerNorm(dim)\n        \n        self.apply(self._init_weights)\n    \n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out //= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n    \n    def forward(self, x, H, W):\n        B, N, C = x.shape\n        q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3)\n        \n        if self.sr_ratio > 1:\n            x_ = x.permute(0, 2, 1).reshape(B, C, H, W)\n            x_ = self.sr(x_).reshape(B, C, -1).permute(0, 2, 1)\n            x_ = self.norm(x_)\n            kv = self.kv(x_).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n        else:\n            kv = self.kv(x).reshape(B, -1, 2, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n        k, v = kv[0], kv[1]\n        \n        attn = (q @ k.transpose(-2, -1)) * self.scale\n        attn = attn.softmax(dim=-1)\n        attn = self.attn_drop(attn)\n        \n        x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n        x = self.proj(x)\n        x = self.proj_drop(x)\n        \n        return x\n\n\nclass Block(nn.Module):\n    \n    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n                 drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, sr_ratio=1):\n        super().__init__()\n        self.norm1 = norm_layer(dim)\n        self.attn = Attention(\n            dim,\n            num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,\n            attn_drop=attn_drop, proj_drop=drop, sr_ratio=sr_ratio)\n        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n        self.norm2 = norm_layer(dim)\n        mlp_hidden_dim = int(dim * mlp_ratio)\n        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n        \n        self.apply(self._init_weights)\n    \n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out //= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n    \n    def forward(self, x, H, W):\n        x = x + self.drop_path(self.attn(self.norm1(x), H, W))\n        x = x + self.drop_path(self.mlp(self.norm2(x), H, W))\n        \n        return x\n\n\nclass OverlapPatchEmbed(nn.Module):\n    \"\"\" Image to Patch Embedding\n    \"\"\"\n    \n    def __init__(self, img_size=224, patch_size=7, stride=4, in_chans=3, embed_dim=768):\n        super().__init__()\n        img_size = to_2tuple(img_size)\n        patch_size = to_2tuple(patch_size)\n        \n        self.img_size = img_size\n        self.patch_size = patch_size\n        self.H, self.W = img_size[0] // patch_size[0], img_size[1] // patch_size[1]\n        self.num_patches = self.H * self.W\n        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride,\n                              padding=(patch_size[0] // 2, patch_size[1] // 2))\n        self.norm = nn.LayerNorm(embed_dim)\n        \n        self.apply(self._init_weights)\n    \n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out //= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n    \n    def forward(self, x):\n        x = self.proj(x)\n        _, _, H, W = x.shape\n        x = x.flatten(2).transpose(1, 2)\n        x = self.norm(x)\n        \n        return x, H, W\n\n\nclass MixVisionTransformer(nn.Module):\n    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dims=[64, 128, 256, 512],\n                 num_heads=[1, 2, 4, 8], mlp_ratios=[4, 4, 4, 4], qkv_bias=False, qk_scale=None, drop_rate=0.,\n                 attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm,\n                 depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1]):\n        super().__init__()\n        self.num_classes = num_classes\n        self.depths = depths\n        self.embed_dims = embed_dims\n        \n        # patch_embed\n        self.patch_embed1 = OverlapPatchEmbed(img_size=img_size, patch_size=7, stride=4, in_chans=in_chans,\n                                              embed_dim=embed_dims[0])\n        self.patch_embed2 = OverlapPatchEmbed(img_size=img_size // 4, patch_size=3, stride=2, in_chans=embed_dims[0],\n                                              embed_dim=embed_dims[1])\n        self.patch_embed3 = OverlapPatchEmbed(img_size=img_size // 8, patch_size=3, stride=2, in_chans=embed_dims[1],\n                                              embed_dim=embed_dims[2])\n        self.patch_embed4 = OverlapPatchEmbed(img_size=img_size // 16, patch_size=3, stride=2, in_chans=embed_dims[2],\n                                              embed_dim=embed_dims[3])\n        \n        # transformer encoder\n        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule\n        cur = 0\n        self.block1 = nn.ModuleList([Block(\n            dim=embed_dims[0], num_heads=num_heads[0], mlp_ratio=mlp_ratios[0], qkv_bias=qkv_bias, qk_scale=qk_scale,\n            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,\n            sr_ratio=sr_ratios[0])\n            for i in range(depths[0])])\n        self.norm1 = norm_layer(embed_dims[0])\n        \n        cur += depths[0]\n        self.block2 = nn.ModuleList([Block(\n            dim=embed_dims[1], num_heads=num_heads[1], mlp_ratio=mlp_ratios[1], qkv_bias=qkv_bias, qk_scale=qk_scale,\n            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,\n            sr_ratio=sr_ratios[1])\n            for i in range(depths[1])])\n        self.norm2 = norm_layer(embed_dims[1])\n        \n        cur += depths[1]\n        self.block3 = nn.ModuleList([Block(\n            dim=embed_dims[2], num_heads=num_heads[2], mlp_ratio=mlp_ratios[2], qkv_bias=qkv_bias, qk_scale=qk_scale,\n            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,\n            sr_ratio=sr_ratios[2])\n            for i in range(depths[2])])\n        self.norm3 = norm_layer(embed_dims[2])\n        \n        cur += depths[2]\n        self.block4 = nn.ModuleList([Block(\n            dim=embed_dims[3], num_heads=num_heads[3], mlp_ratio=mlp_ratios[3], qkv_bias=qkv_bias, qk_scale=qk_scale,\n            drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[cur + i], norm_layer=norm_layer,\n            sr_ratio=sr_ratios[3])\n            for i in range(depths[3])])\n        self.norm4 = norm_layer(embed_dims[3])\n        \n        # classification head\n        # self.head = nn.Linear(embed_dims[3], num_classes) if num_classes > 0 else nn.Identity()\n        \n        self.apply(self._init_weights)\n    \n    def _init_weights(self, m):\n        if isinstance(m, nn.Linear):\n            trunc_normal_(m.weight, std=.02)\n            if isinstance(m, nn.Linear) and m.bias is not None:\n                nn.init.constant_(m.bias, 0)\n        elif isinstance(m, nn.LayerNorm):\n            nn.init.constant_(m.bias, 0)\n            nn.init.constant_(m.weight, 1.0)\n        elif isinstance(m, nn.Conv2d):\n            fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channels\n            fan_out //= m.groups\n            m.weight.data.normal_(0, math.sqrt(2.0 / fan_out))\n            if m.bias is not None:\n                m.bias.data.zero_()\n    \n    def init_weights(self, pretrained=None):\n        pass\n        # if isinstance(pretrained, str):\n        #     logger = get_root_logger()\n        #     load_checkpoint(self, pretrained, map_location='cpu', strict=False, logger=logger)\n    \n    def reset_drop_path(self, drop_path_rate):\n        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(self.depths))]\n        cur = 0\n        for i in range(self.depths[0]):\n            self.block1[i].drop_path.drop_prob = dpr[cur + i]\n        \n        cur += self.depths[0]\n        for i in range(self.depths[1]):\n            self.block2[i].drop_path.drop_prob = dpr[cur + i]\n        \n        cur += self.depths[1]\n        for i in range(self.depths[2]):\n            self.block3[i].drop_path.drop_prob = dpr[cur + i]\n        \n        cur += self.depths[2]\n        for i in range(self.depths[3]):\n            self.block4[i].drop_path.drop_prob = dpr[cur + i]\n    \n    def freeze_patch_emb(self):\n        self.patch_embed1.requires_grad = False\n    \n    @torch.jit.ignore\n    def no_weight_decay(self):\n        return {'pos_embed1', 'pos_embed2', 'pos_embed3', 'pos_embed4', 'cls_token'}  # has pos_embed may be better\n    \n    def get_classifier(self):\n        return self.head\n    \n    def reset_classifier(self, num_classes, global_pool=''):\n        self.num_classes = num_classes\n        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n    \n    def forward_features(self, x):\n        B = x.shape[0]\n        outs = []\n        \n        # stage 1\n        x, H, W = self.patch_embed1(x)\n        for i, blk in enumerate(self.block1):\n            x = blk(x, H, W)\n        x = self.norm1(x)\n        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()\n        outs.append(x)\n        \n        # stage 2\n        x, H, W = self.patch_embed2(x)\n        for i, blk in enumerate(self.block2):\n            x = blk(x, H, W)\n        x = self.norm2(x)\n        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()\n        outs.append(x)\n        \n        # stage 3\n        x, H, W = self.patch_embed3(x)\n        for i, blk in enumerate(self.block3):\n            x = blk(x, H, W)\n        x = self.norm3(x)\n        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()\n        outs.append(x)\n        \n        # stage 4\n        x, H, W = self.patch_embed4(x)\n        for i, blk in enumerate(self.block4):\n            x = blk(x, H, W)\n        x = self.norm4(x)\n        x = x.reshape(B, H, W, -1).permute(0, 3, 1, 2).contiguous()\n        outs.append(x)\n        \n        return outs\n    \n    def forward(self, x):\n        x = self.forward_features(x)\n        # x = self.head(x)\n        \n        return x\n\n\n\n\nclass mit_b0(MixVisionTransformer):\n    def __init__(self, **kwargs):\n        super(mit_b0, self).__init__(\n            patch_size=4, embed_dims=[32, 64, 160, 256], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],\n            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],\n            drop_rate=0.0, drop_path_rate=0.1)\n\n\nclass mit_b1(MixVisionTransformer):\n    def __init__(self, **kwargs):\n        super(mit_b1, self).__init__(\n            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],\n            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[2, 2, 2, 2], sr_ratios=[8, 4, 2, 1],\n            drop_rate=0.0, drop_path_rate=0.1)\n\n\nclass mit_b2(MixVisionTransformer):\n    def __init__(self, **kwargs):\n        super(mit_b2, self).__init__(\n            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],\n            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 6, 3], sr_ratios=[8, 4, 2, 1],\n            drop_rate=0.0, drop_path_rate=0.1)\n\n\nclass mit_b3(MixVisionTransformer):\n    def __init__(self, **kwargs):\n        super(mit_b3, self).__init__(\n            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],\n            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 4, 18, 3], sr_ratios=[8, 4, 2, 1],\n            drop_rate=0.0, drop_path_rate=0.1)\n\n\nclass mit_b4(MixVisionTransformer):\n    def __init__(self, **kwargs):\n        super(mit_b4, self).__init__(\n            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],\n            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 8, 27, 3], sr_ratios=[8, 4, 2, 1],\n            drop_rate=0.0, drop_path_rate=0.1)\n\nclass mit_b5(MixVisionTransformer):\n    def __init__(self, **kwargs):\n        super(mit_b5, self).__init__(\n            patch_size=4, embed_dims=[64, 128, 320, 512], num_heads=[1, 2, 5, 8], mlp_ratios=[4, 4, 4, 4],\n            qkv_bias=True, norm_layer=partial(nn.LayerNorm, eps=1e-6), depths=[3, 6, 40, 3], sr_ratios=[8, 4, 2, 1],\n            drop_rate=0.0, drop_path_rate=0.1)","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:07:01.788570Z","iopub.execute_input":"2022-09-18T12:07:01.788987Z","iopub.status.idle":"2022-09-18T12:07:01.868384Z","shell.execute_reply.started":"2022-09-18T12:07:01.788950Z","shell.execute_reply":"2022-09-18T12:07:01.867364Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## SegFormer","metadata":{}},{"cell_type":"code","source":"class MixUpSample(nn.Module):\n    def __init__( self, scale_factor=2):\n        super().__init__()\n        self.mixing = nn.Parameter(torch.tensor(0.5))\n        self.scale_factor = scale_factor\n    \n    def forward(self, x):\n        x = self.mixing *F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False) \\\n            + (1-self.mixing )*F.interpolate(x, scale_factor=self.scale_factor, mode='nearest')\n        return x\n    \n    \nclass SegformerDecoder(nn.Module):\n    def __init__(\n            self,\n            encoder_dim = [32, 64, 160, 256],\n            decoder_dim = 256,\n    ):\n        super().__init__()\n        self.mixing = nn.Parameter(torch.FloatTensor([0.5,0.5,0.5,0.5]))\n        self.mlp = nn.ModuleList([\n            nn.Sequential(\n                nn.Conv2d(dim, decoder_dim, 1, padding= 0,  bias=False), #follow mmseg to use conv-bn-relu\n                nn.BatchNorm2d(decoder_dim),\n                nn.ReLU(inplace=True),\n                MixUpSample(2**i) if i!=0 else nn.Identity(),\n            ) for i, dim in enumerate(encoder_dim)])\n        \n        self.fuse = nn.Sequential(\n            nn.Conv2d(len(encoder_dim) * decoder_dim, decoder_dim, 1, padding=0, bias=False),\n            nn.BatchNorm2d(decoder_dim),\n            nn.ReLU(inplace=True),\n            # nn.Conv2d(decoder_dim, decoder_dim, 3, padding=1, bias=False),\n            # nn.BatchNorm2d(decoder_dim),\n            # nn.ReLU(inplace=True),\n        )\n    \n    def forward(self, feature):\n        \n        out = []\n        for i,f in enumerate(feature):\n            f = self.mlp[i](f)\n            out.append(f)\n         \n        x = self.fuse(torch.cat(out, dim = 1))\n        return x, out","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:07:29.370389Z","iopub.execute_input":"2022-09-18T12:07:29.371121Z","iopub.status.idle":"2022-09-18T12:07:29.387557Z","shell.execute_reply.started":"2022-09-18T12:07:29.371073Z","shell.execute_reply":"2022-09-18T12:07:29.386807Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Net","metadata":{}},{"cell_type":"code","source":"class Net(nn.Module):\n    def __init__(self, builder):\n        super(Net, self).__init__()\n        self.dropout = nn.Dropout(0.1)\n        \n        self.encoder = builder()\n        encoder_dim = self.encoder.embed_dims\n        #[64, 128, 320, 512]\n        \n        self.decoder = SegformerDecoder(\n            encoder_dim = encoder_dim,\n            decoder_dim = 320,\n        )\n        self.logit = nn.Sequential(\n            nn.Conv2d(320, 1, kernel_size=1, padding=0),\n        )\n        self.aux = nn.ModuleList([\n            nn.Conv2d(encoder_dim[i], 1, kernel_size=1, padding=0) for i in range(4)\n        ])\n    \n    \n    def forward(self, x, aux:bool=True):\n        encoder = self.encoder(x)\n        last, _ = self.decoder(encoder)\n        last  = self.dropout(last)\n        logit = self.logit(last)\n        logit = F.interpolate(logit, size=None, scale_factor=4, mode='bilinear', align_corners=False)\n        if aux:\n            outputs = [self.aux[i](encoder[i]) for i in range(4)]\n            return logit, outputs\n        else:\n            return logit","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:07:47.641325Z","iopub.execute_input":"2022-09-18T12:07:47.641780Z","iopub.status.idle":"2022-09-18T12:07:47.655320Z","shell.execute_reply.started":"2022-09-18T12:07:47.641740Z","shell.execute_reply":"2022-09-18T12:07:47.653548Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Inference","metadata":{}},{"cell_type":"code","source":"# Test DataFrame\ntest = pd.read_csv('/kaggle/input/hubmap-organ-segmentation/test.csv')\n\ndataset = HubmapHpaDataset(test,\n                       Path(\"../input/hubmap-organ-segmentation/test_images\"))\ndisplay(test)\n\nmodel = torch.load(WEIGHTS[0], map_location=\"cpu\")\nmodel.to(DEVICE)\nmodel.eval();","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:09:26.742057Z","iopub.execute_input":"2022-09-18T12:09:26.742661Z","iopub.status.idle":"2022-09-18T12:09:27.039324Z","shell.execute_reply.started":"2022-09-18T12:09:26.742620Z","shell.execute_reply":"2022-09-18T12:09:27.038344Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Predictions are stored as a list of dictionaries\ntest_rows = []\n\n# Iterate over all test images\nfor row_idx, row in test.iterrows():\n    X = dataset[row_idx]\n    X = X.to(DEVICE).unsqueeze(0)\n    with torch.no_grad():\n        Y_pred = model(X, aux=False).sigmoid().squeeze(0).squeeze(0)\n    Y_pred = Y_pred.cpu().numpy()\n        \n    # Resize and Binarize Mask\n    if HUBMAP_ONLY and row[\"data_source\"] == \"HPA\":\n        test_rows.append({\n            'id': row['id'],\n            'rle': \"\"\n        })\n    else:\n        thres = cv2.threshold((Y_pred*255).astype(np.uint8), 0, 255, cv2.THRESH_OTSU)[0] / 255\n        Y_pred = cv2.resize(Y_pred, [row['img_height'], row['img_width']], interpolation=cv2.INTER_LINEAR).astype(float)\n        Y_binary = (Y_pred > thres).astype(np.int8)\n        test_rows.append({\n            'id': row['id'],\n            'rle': rle_encode_less_memory(Y_binary)\n        })\n        \n        if row_idx == 0:\n            print(f\"thres = {thres:.4f}\")\n            fig, axes = plt.subplots(1, 3, figsize=(12,4))\n            axes[0].imshow(X.squeeze(0).cpu().numpy().transpose(1,2,0))\n            axes[1].imshow(Y_pred)\n            axes[2].imshow(Y_binary)\n            plt.show()","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:09:27.931276Z","iopub.execute_input":"2022-09-18T12:09:27.931657Z","iopub.status.idle":"2022-09-18T12:09:35.275419Z","shell.execute_reply.started":"2022-09-18T12:09:27.931615Z","shell.execute_reply":"2022-09-18T12:09:35.274175Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# Submission","metadata":{}},{"cell_type":"code","source":"submission = pd.DataFrame(test_rows)\ndisplay(submission)","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:09:42.825227Z","iopub.execute_input":"2022-09-18T12:09:42.825590Z","iopub.status.idle":"2022-09-18T12:09:42.837149Z","shell.execute_reply.started":"2022-09-18T12:09:42.825560Z","shell.execute_reply":"2022-09-18T12:09:42.835995Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"submission.to_csv('submission.csv', index=False)","metadata":{"execution":{"iopub.status.busy":"2022-09-18T12:09:45.392528Z","iopub.execute_input":"2022-09-18T12:09:45.392914Z","iopub.status.idle":"2022-09-18T12:09:45.401237Z","shell.execute_reply.started":"2022-09-18T12:09:45.392860Z","shell.execute_reply":"2022-09-18T12:09:45.400183Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}