{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.11.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":112509,"databundleVersionId":14254895,"sourceType":"competition"}],"dockerImageVersionId":31154,"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"# I would like to thank @takaito for his generous sharing (https://www.kaggle.com/code/takaito/csiro-img2bio-training-notebook)\n\nimport timm\nimport torch\nimport torch.nn as nn\nfrom torch.utils.data import Dataset, DataLoader\nfrom PIL import Image\nimport torchvision.transforms as T\nimport pytorch_lightning as pl\nfrom pytorch_lightning.callbacks import ModelCheckpoint, EarlyStopping, LearningRateMonitor\nimport random\nimport os\nimport numpy as np\nimport pandas as pd\nfrom sklearn.model_selection import KFold","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== Load Datafile ========\ntrain_df = pd.read_csv('/kaggle/input/csiro-biomass/train.csv')\npivot_df = train_df.pivot(\n    index='image_path',\n    columns='target_name',\n    values='target'\n).reset_index()\nmeta_cols = [\"image_path\", \"State\", \"Pre_GSHH_NDVI\", \"Height_Ave_cm\"]\nother_cols = train_df[meta_cols].drop_duplicates('image_path').set_index('image_path')\ntrain_df = pivot_df.merge(other_cols, on='image_path', how='left')\ndummies = pd.get_dummies(train_df[\"State\"], prefix=\"State\", dummy_na=False).astype(int)\ntrain_df = pd.concat([train_df.drop(columns=[\"State\"]), dummies], axis=1)\n    \ntrain_df.head()","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== Seed ========\ndef seed_everything(seed: int = 114514):\n    pl.seed_everything(seed, workers=True)\n    random.seed(seed)\n    np.random.seed(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed_all(seed)\n    os.environ[\"PYTHONHASHSEED\"] = str(seed)\n    # \n    torch.backends.cudnn.deterministic = True\n    torch.backends.cudnn.benchmark = False\nseed_everything(114514)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== Dataset ========\nclass ImageRegressionDataset(Dataset):\n    def __init__(self, df, full_image_aug=None, per_half_transform=None):\n        self.df = df\n        self.full_image_aug = full_image_aug\n        self.per_half_transform = per_half_transform\n\n    def __len__(self):\n        return len(self.df)\n\n    def __getitem__(self, idx):\n        row = self.df.iloc[idx]\n        image = Image.open('/kaggle/input/csiro-biomass/' + row[\"image_path\"]).convert(\"RGB\")\n    \n        base = self.full_image_aug(image) if self.full_image_aug is not None else image\n    \n        left = base.crop((0, 0, 1000, 1000))\n        right = base.crop((1000, 0, 2000, 1000))\n    \n        if self.per_half_transform is not None:\n            left = self.per_half_transform(left)\n            right = self.per_half_transform(right)\n        else:\n            left = F.to_tensor(left)\n            right = F.to_tensor(right)\n    \n        imgs = torch.stack([left, right], dim=0)   # (2, C, H, W)\n    \n        targets = torch.tensor([\n            row[\"Dry_Green_g\"],\n            row[\"Dry_Dead_g\"],\n            row[\"Dry_Clover_g\"],\n            row[\"GDM_g\"],\n            row[\"Dry_Total_g\"]\n        ], dtype=torch.float32)\n    \n        return imgs, targets","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== Loss Funtion ========\nclass WeightedR2Loss(nn.Module):\n    def __init__(self, weights=None, eps=1e-8):\n        super().__init__()\n        if weights is None:\n            weights = torch.tensor([0.1, 0.1, 0.1, 0.2, 0.5])\n        self.register_buffer('weights', weights)\n        self.eps = eps\n\n    def forward(self, y_pred, y_true):\n        \"\"\"\n        y_pred: (B, 3)\n        y_true: (B, 5)\n        \"\"\"\n        DG = y_pred[:,0]\n        GDM = y_pred[:,1]\n        DT = y_pred[:,2]\n        DD = DT - GDM\n        DC = GDM - DG\n        y_hat = torch.stack([DG, DD, DC, GDM, DT], dim=1)\n        \n        y_mean = torch.mean(y_true, dim=0, keepdim=True)\n        ss_res = torch.sum((y_true - y_hat) ** 2, dim=0)\n        ss_tot = torch.sum((y_true - y_mean) ** 2, dim=0)\n        r2 = 1 - ss_res / (ss_tot + self.eps)\n\n        weighted_r2 = torch.sum(self.weights * r2)\n        loss = 1 - weighted_r2\n\n        return loss","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== Model ========\nclass MultiRegressionModel(pl.LightningModule):\n    def __init__(\n        self,\n        model_name=\"efficientnet_b2\",\n        pretrained=False,\n        lr=5e-3,\n        output_dim=3,\n        hidden_dim=1536,\n        open_last_n_blocks=1,\n        train_bn=False,\n    ):\n        super().__init__()\n        self.save_hyperparameters()\n        \n        self.backbone = timm.create_model(\n            model_name, \n            pretrained=pretrained, \n            num_classes=0, \n            global_pool=\"avg\"\n        )\n        feat_dim = getattr(self.backbone, \"num_features\", None)\n        if feat_dim is None:\n            x_dummy = torch.zeros(1, 3, 1000, 1000)\n            with torch.no_grad():\n                feat_dim = self.backbone(x_dummy).shape[-1]\n        self.head = nn.Sequential(\n            nn.Linear(feat_dim * 2, hidden_dim),\n            nn.ReLU(inplace=True),\n            nn.Linear(hidden_dim, output_dim),\n        )\n\n        self.criterion = WeightedR2Loss()\n        self.val_outputs = []\n        self._freeze_backbone(open_last_n_blocks=open_last_n_blocks, train_bn=train_bn)\n\n    # Helpers\n    def _set_requires_grad(self, module, flag: bool):\n        for p in module.parameters():\n            p.requires_grad = flag\n            \n    def _freeze_backbone(self, open_last_n_blocks=1, train_bn=False):\n        self._set_requires_grad(self.backbone, False)\n\n        if hasattr(self.backbone, \"blocks\"):\n            blocks = self.backbone.blocks\n            n = len(blocks)\n            assert open_last_n_blocks >= 1 and open_last_n_blocks <= n, \\\n                f\"Expect open_last_n_blocks between [1, {n}], get {open_last_n_blocks}\"\n            for blk in blocks[n - open_last_n_blocks:]:\n                self._set_requires_grad(blk, True)\n\n        if not train_bn:\n            for m in self.backbone.modules():\n                if isinstance(m, (nn.BatchNorm2d, nn.SyncBatchNorm)):\n                    if all(not p.requires_grad for p in m.parameters()):\n                        m.eval()\n\n        self._set_requires_grad(self.head, True)\n\n\n    def forward(self, x_imgs):\n        if x_imgs.ndim == 5:\n            B, T, C, H, W = x_imgs.shape\n            assert T == 2, f\"Expect 2 pics, get T={T}\"\n            x = x_imgs.view(B * T, C, H, W)\n            feats = self.backbone(x)\n            F = feats.shape[-1]\n            feats = feats.view(B, T, F)\n            fused = torch.cat([feats[:, 0, :], feats[:, 1, :]], dim=1)\n        elif x_imgs.ndim == 4:\n            T, C, H, W = x_imgs.shape\n            assert T == 2, f\"Expect 2 pics, get T={T}\"\n            feats = self.backbone(x_imgs)\n            fused = torch.cat([feats[0], feats[1]], dim=0).unsqueeze(0)\n        else:\n            raise ValueError(\"SHAPE should be train/val-(B, 2, C, H, W) or single-sample val-(2, C, H, W)\")\n\n        y_hat = self.head(fused)\n        return y_hat\n\n    def training_step(self, batch, batch_idx):\n        x_img, y = batch\n        y_hat = self(x_img)\n        loss = self.criterion(y_hat, y)\n        self.log(\"train_loss\", loss, on_step=False, on_epoch=True)\n        return loss\n\n    def validation_step(self, batch, batch_idx):\n        x_img, y = batch\n        y_hat = self(x_img)\n        loss = self.criterion(y_hat, y)\n        self.val_outputs.append((y_hat.detach().cpu(), y.detach().cpu()))\n        self.log(\"val_loss\", loss, on_step=False, on_epoch=True)\n        return loss\n\n    def on_validation_epoch_end(self):\n        y_hats = torch.cat([x[0] for x in self.val_outputs], dim=0)\n        y_trues = torch.cat([x[1] for x in self.val_outputs], dim=0)\n        self.val_outputs.clear()\n\n        DG = y_hats[:, 0]\n        GDM = y_hats[:, 1]\n        DT = y_hats[:, 2]\n        DD = DT - GDM\n        DC = GDM - DG\n        y_hat_full = torch.stack([DG, DD, DC, GDM, DT], dim=1)\n\n        y_mean = torch.mean(y_trues, dim=0, keepdim=True)\n        ss_res = torch.sum((y_trues - y_hat_full) ** 2, dim=0)\n        ss_tot = torch.sum((y_trues - y_mean) ** 2, dim=0)\n        r2 = 1 - ss_res / (ss_tot + 1e-8)\n\n        weights = self.criterion.weights.cpu()\n        weighted_r2 = torch.sum(weights * r2)\n\n        for i, name in enumerate([\"DG\", \"DD\", \"DC\", \"GDM\", \"DT\"]):\n            self.log(f\"val_r2_{name}\", r2[i], prog_bar=True, on_epoch=True)\n        self.log(\"val_weighted_r2\", weighted_r2, prog_bar=True, on_epoch=True)\n\n    def configure_optimizers(self):\n        optimizer = torch.optim.AdamW(self.parameters(), lr=self.hparams.lr)\n        scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(\n            optimizer,\n            mode=\"min\",\n            factor=0.5,\n            patience=2,\n            threshold=0.001,\n            min_lr=1e-7,\n            verbose=False,\n        )\n        return {\n            \"optimizer\": optimizer,\n            \"lr_scheduler\": {\n                \"scheduler\": scheduler,\n                \"monitor\": \"val_loss\",\n                \"interval\": \"epoch\",\n                \"frequency\": 1,\n            },\n        }","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== DataModule ========\nclass ImageRegressionDataModule(pl.LightningDataModule):\n    def __init__(self, train_df, valid_df, batch_size=8, num_workers=4, img_size=1000):\n        super().__init__()\n        self.train_df = train_df\n        self.valid_df = valid_df\n        self.batch_size = batch_size\n        self.num_workers = num_workers\n        self.img_size = img_size\n\n    def setup(self, stage=None):\n        full_image_aug_train = T.Compose([\n            T.RandomHorizontalFlip(p=0.5),\n            T.RandomVerticalFlip(p=0.5),\n            T.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.05),\n        ])\n        full_image_aug_val = None\n\n        per_half_transform_train = T.Compose([\n            T.Resize((self.img_size, self.img_size)),\n            T.ToTensor(),\n            T.Normalize([0.485, 0.456, 0.406],\n                        [0.229, 0.224, 0.225])\n        ])\n        per_half_transform_val = T.Compose([\n            T.Resize((self.img_size, self.img_size)),\n            T.ToTensor(),\n            T.Normalize([0.485, 0.456, 0.406],\n                        [0.229, 0.224, 0.225])\n        ])\n\n        self.train_dataset = ImageRegressionDataset(\n            self.train_df,\n            full_image_aug=full_image_aug_train,\n            per_half_transform=per_half_transform_train\n        )\n        self.val_dataset = ImageRegressionDataset(\n            self.valid_df,\n            full_image_aug=full_image_aug_val,\n            per_half_transform=per_half_transform_val\n        )\n\n    def train_dataloader(self):\n        return DataLoader(self.train_dataset, batch_size=self.batch_size,\n                          shuffle=True, num_workers=self.num_workers)\n\n    def val_dataloader(self):\n        return DataLoader(self.val_dataset, batch_size=self.batch_size,\n                          shuffle=False, num_workers=self.num_workers)","metadata":{"trusted":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"# ======== Train ========\nkf = KFold(n_splits=3, random_state=114514, shuffle=True)\nfor fold, (train_index, valid_index) in enumerate(kf.split(train_df)):\n    datamodule = ImageRegressionDataModule(train_df.iloc[train_index], train_df.iloc[valid_index])\n    model = MultiRegressionModel(model_name=\"efficientnet_b2\", pretrained=True, lr=5e-3)\n    \n    checkpoint_callback = ModelCheckpoint(\n        monitor=\"val_weighted_r2\",\n        save_top_k=1,\n        mode=\"max\",\n        filename=f\"best_model_fold{fold}\"\n    )\n    lr_monitor_callback = LearningRateMonitor(logging_interval=\"epoch\")\n    early_stop_callback = EarlyStopping(\n        monitor=\"val_weighted_r2\",\n        patience=5,\n        mode=\"max\",\n        min_delta=0.0002,\n        verbose=True\n    )\n    \n    trainer = pl.Trainer(\n        max_epochs=64,\n        accelerator=\"gpu\" if torch.cuda.is_available() else \"cpu\",\n        callbacks=[checkpoint_callback, lr_monitor_callback, early_stop_callback],\n        precision=\"16-mixed\",\n        log_every_n_steps=15,\n    )\n    trainer.fit(model, datamodule=datamodule)\n    torch.save(model.state_dict(), f\"model_fold{fold}.pth\")","metadata":{"trusted":true},"outputs":[],"execution_count":null}]}