\n\n- This notebook explores a baseline solution with efficientnet for the CSIRO Image2BioMass Challenge aiming to predict the coverage of vegetation for a patch of land from images.\n\n- The Train Dataset provides 357 images, each of which record 5 targets - `'Dry_Clover_g' 'Dry_Dead_g' 'Dry_Green_g' 'Dry_Total_g' 'GDM_g'` \n\n- The Competition metric is individual-r2 score for each of the 5 targets","metadata":{}},{"cell_type":"code","source":"SUBMISSION = False\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport seaborn as sns\n\n\nimport os\nimport sys\nfrom tqdm import tqdm \nimport sys\n\n#------------\n#--- data ---\n#------------\nDATA_DIR = r\"/kaggle/input/csiro-biomass\"\nTRAIN_DIR = os.path.join(DATA_DIR, 'train')\nTEST_DIR = os.path.join(DATA_DIR, 'test')\ndf = pd.read_csv(os.path.join(DATA_DIR, \"train.csv\"))\ndf_test = pd.read_csv(os.path.join(DATA_DIR, \"test.csv\"))\ndf_sample = pd.read_csv(os.path.join(DATA_DIR, \"sample_submission.csv\"))\nTARGET = \"target\"\n\n#-----------------\n#--- cosmetics ---\n#-----------------\nSNS_CMAP = 'BrBG'\nfrom rich.console import Console\nrc = Console(force_jupyter=False, color_system=\"truecolor\")\n\nimport warnings\nwarnings.filterwarnings('ignore')\n%load_ext autoreload\n%autoreload 2\n%matplotlib inline\n\nsns.set()\nsns.set_palette(SNS_CMAP)\n# --- dark mode---\n# plt.style.use(\"dark_background\")\n# plt.rcParams['grid.color'] = '#444444'\ncolors = sns.palettes.color_palette(SNS_CMAP)\npd.options.mode.chained_assignment = None\n\ndef rprint(text:str, rule:bool = False)->None:\n rc.print(text)\n if rule:\n rc.rule(\"***\")\n\ndef df_head_regression(df, target_col=None, palette=\"viridis\", n=5, alpha=0.5):\n \"\"\"\n Highlights rows based on a continuous target column using gradient background colors.\n\n Parameters:\n - df: DataFrame to style\n - target_col: column name to base row color on (default: last column)\n - palette: Seaborn/Matplotlib palette name or list of colors\n - n: number of rows to display\n - alpha: transparency level (0 = fully transparent, 1 = opaque)\n \"\"\"\n if target_col is None:\n target_col = df.columns[-1]\n\n df_show = df.head(n)\n\n # Normalize target values between 0 and 1\n target_vals = df_show[target_col].astype(float)\n norm = (target_vals - target_vals.min()) / (target_vals.max() - target_vals.min() + 1e-9)\n\n # Get color palette\n cmap = plt.get_cmap(palette)\n\n def row_style(row):\n val = row[target_col]\n # Normalize this row’s target value\n val_norm = (val - target_vals.min()) / (target_vals.max() - target_vals.min() + 1e-9)\n # Get RGBA\n r, g, b, _ = cmap(val_norm)\n rgba = (int(r * 255), int(g * 255), int(b * 255), alpha)\n return [f'background-color: rgba{rgba}'] * len(row)\n\n return df_show.style.apply(row_style, axis=1)\n\ndf_head_regression(df, TARGET, SNS_CMAP, n=8)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:11:47.852790Z","iopub.execute_input":"2025-11-30T08:11:47.852950Z","iopub.status.idle":"2025-11-30T08:11:50.866911Z","shell.execute_reply.started":"2025-11-30T08:11:47.852934Z","shell.execute_reply":"2025-11-30T08:11:50.866025Z"},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"agg_df = df.agg([\"nunique\", \"unique\", lambda x:x.isna().sum(), \"dtypes\"]).T\nagg_df['unique'] = agg_df['unique'].apply(lambda x: x if len(x)<10 else x[:10]) \nagg_df.style.apply(lambda s: [f'background-color: rgba({colors[0][0]*255}, {colors[0][1]*255}, {colors[0][2]*255}, 0.3)' if i % 2 == 0 else f'background-color: rgba({colors[5][0]*255}, {colors[5][1]*255}, {colors[5][2]*255}, 0.3)' for i in range(len(s))])","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-05T16:23:26.239982Z","iopub.execute_input":"2025-11-05T16:23:26.241129Z","iopub.status.idle":"2025-11-05T16:23:26.305209Z","shell.execute_reply.started":"2025-11-05T16:23:26.241081Z","shell.execute_reply":"2025-11-05T16:23:26.304088Z"},"jupyter":{"source_hidden":true},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#
1. EDA
","metadata":{"_kg_hide-input":true}},{"cell_type":"code","source":"if not SUBMISSION:\n df_grouped = df.groupby('image_path')[[TARGET]].sum()\n df_img = df_grouped.reset_index().merge(\n df[['image_path', 'Species']], how='left', on='image_path')\n \n def visualize_multiclass_from_df(df, cols:int = 2):\n species = df['Species'].unique()\n rprint(f\"found {len(species)} unique species\")\n rows = (len(species)+cols-1)//cols\n fig, ax = plt.subplots(rows, cols, figsize=(15, 4*rows))\n for idx, species_name in enumerate(species):\n img_path = os.path.join(DATA_DIR, df[df['Species']==species_name].iloc[0][\"image_path\"])\n img = cv2.imread(img_path)\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n ax.flatten()[idx].imshow(img)\n # ax.flatten()[idx].axis('off')\n ax.flatten()[idx].set_title(species_name)\n \n for axes in ax.flat:\n axes.grid(False)\n axes.set_xticks([])\n axes.set_yticks([])\n axes.set_xticklabels([])\n axes.set_yticklabels([])\n axes.set_frame_on(False)\n \n plt.tight_layout()\n plt.show()\n \n visualize_multiclass_from_df(df_img)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-10-30T19:58:56.025171Z","iopub.execute_input":"2025-10-30T19:58:56.025519Z","iopub.status.idle":"2025-10-30T19:59:02.777189Z","shell.execute_reply.started":"2025-10-30T19:58:56.025493Z","shell.execute_reply":"2025-10-30T19:59:02.775552Z"},"jupyter":{"source_hidden":true},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if not SUBMISSION:\n fig, ax = plt.subplots(1, 2, figsize = (15, 7))\n sns.histplot(\n data=df,\n x=TARGET,\n bins=20,\n hue=\"target_name\",\n multiple=\"stack\",\n palette=SNS_CMAP,\n ax=ax[0]\n )\n sns.boxplot(\n df,\n y=TARGET,\n x='target_name',\n ax=ax[1],\n palette = SNS_CMAP,\n )\n plt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-05T07:02:53.745816Z","iopub.execute_input":"2025-11-05T07:02:53.746191Z","iopub.status.idle":"2025-11-05T07:02:54.411203Z","shell.execute_reply.started":"2025-11-05T07:02:53.746168Z","shell.execute_reply":"2025-11-05T07:02:54.410307Z"},"jupyter":{"source_hidden":true},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"code","source":"import cv2\nfrom skimage.filters import sobel\nfrom skimage.measure import shannon_entropy\nimport matplotlib.gridspec as gridspec\n\ndef extract_feat(df: pd.DataFrame, data_dir: str) -> pd.DataFrame:\n # Make a copy to avoid modifying the original DataFrame\n df = df.copy()\n\n # Prepare new columns\n df[\"green_ratio\"] = np.nan\n df[\"luminance\"] = np.nan\n df[\"edge_density\"] = np.nan\n df[\"color_entropy\"] = np.nan\n\n # Iterate over rows with tqdm for progress\n for i, df_row in tqdm(df.iterrows(), total=len(df), desc=\"Extracting features\"):\n img_path = os.path.join(data_dir, df_row[\"image_path\"])\n img = cv2.imread(str(img_path))\n if img is None:\n continue\n img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)\n img = img.astype(np.float32) / 255.0\n\n R, G, B = img[..., 0], img[..., 1], img[..., 2]\n\n # 1. Green ratio\n mask_green = (G > B) & (G > R) & (G > 0.2)\n df.at[i, \"green_ratio\"] = mask_green.mean()\n\n # 2. Luminance\n df.at[i, \"luminance\"] = np.mean(0.299 * R + 0.587 * G + 0.114 * B)\n\n # 3. Edge density\n gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)\n edge_map = sobel(gray / 255.0)\n df.at[i, \"edge_density\"] = edge_map.mean()\n\n # 4. Shannon entropy\n df.at[i, \"color_entropy\"] = shannon_entropy(gray)\n\n return df\n\ndef custom_plot(df):\n fig = plt.figure(figsize=(15, 14))\n outer = gridspec.GridSpec(2, 2, figure=fig, wspace=0.3, hspace=0.35)\n \n # --- Helper to create a joint-like set of axes inside an outer cell\n def make_joint_axes(subplot_spec, main_height_ratio=4, main_width_ratio=4):\n \"\"\"\n Given an outer subplot_spec, returns (ax_main, ax_marg_x, ax_marg_y)\n where:\n - ax_main is the central joint plot,\n - ax_marg_x is the marginal on top (shares x with main),\n - ax_marg_y is the marginal on the right (shares y with main).\n \"\"\"\n inner = gridspec.GridSpecFromSubplotSpec(\n 2, 2,\n subplot_spec=subplot_spec,\n height_ratios=[1, main_height_ratio],\n width_ratios=[main_width_ratio, 1],\n hspace=0.05,\n wspace=0.05\n )\n # create main first so top/right can share axes\n ax_main = fig.add_subplot(inner[2]) # row=1,col=0 index=2\n ax_marg_x = fig.add_subplot(inner[0], sharex=ax_main) # row=0,col=0 index=0\n ax_marg_y = fig.add_subplot(inner[3], sharey=ax_main) # row=1,col=1 index=3\n \n # clean up marginal ticks so they don't overlap clutter\n plt.setp(ax_marg_x.get_xticklabels(), visible=False)\n ax_marg_x.tick_params(axis=\"x\", which=\"both\", bottom=False, top=False)\n plt.setp(ax_marg_y.get_yticklabels(), visible=False)\n ax_marg_y.tick_params(axis=\"y\", which=\"both\", left=False, right=False)\n \n return ax_main, ax_marg_x, ax_marg_y\n \n # -------------------------\n # Top-left joint: green_ratio vs TARGET\n ax_main_1, ax_top_1, ax_right_1 = make_joint_axes(outer[0])\n \n sns.kdeplot(\n data=df, x=\"green_ratio\", y=\"Dry_Green_g\",\n fill=True, cmap=\"Greens\", thresh=0.05, ax=ax_main_1, levels=6\n )\n sns.kdeplot(data=df, x=\"green_ratio\", fill=True, ax=ax_top_1, linewidth=1)\n sns.kdeplot(data=df, y=\"Dry_Green_g\", fill=True, ax=ax_right_1, linewidth=1)\n \n ax_main_1.set_xlabel(\"green_ratio\")\n ax_main_1.set_ylabel(\"Dry_Green_g\")\n ax_top_1.set_title(\"Green ratio vs \" + str(\"Dry_Green_g\"))\n \n # -------------------------\n # Bottom-left joint: luminance vs TARGET\n ax_main_2, ax_top_2, ax_right_2 = make_joint_axes(outer[2])\n \n sns.kdeplot(\n data=df, x=\"luminance\", y=\"Dry_Green_g\",\n fill=True, cmap=\"Greens\", thresh=0.05, ax=ax_main_2, levels=6\n )\n sns.kdeplot(data=df, x=\"luminance\", fill=True, ax=ax_top_2, linewidth=1)\n sns.kdeplot(data=df, y=\"Dry_Green_g\", fill=True, ax=ax_right_2, linewidth=1)\n \n ax_main_2.set_xlabel(\"luminance\")\n ax_main_2.set_ylabel(\"Dry_Green_g\")\n ax_top_2.set_title(\"Luminance vs \" + str(\"Dry_Green_g\"))\n \n # -------------------------\n # Top-right: scatter edge_density vs TARGET (placed in outer[1])\n ax_scatter1 = fig.add_subplot(outer[1])\n sns.scatterplot(data=df, x=\"edge_density\", y=\"Dry_Clover_g\", hue=\"Dry_Green_g\", ax=ax_scatter1, s=20, alpha=0.6, palette=SNS_CMAP)\n ax_scatter1.set_title(\"Edge density vs \" + str(\"Dry_Clover_g\"))\n \n # Bottom-right: scatter color_entropy vs TARGET (placed in outer[3])\n ax_scatter2 = fig.add_subplot(outer[3])\n sns.scatterplot(data=df[df[\"color_entropy\"]>14], x=\"color_entropy\", y=\"GDM_g\", hue=\"Dry_Green_g\", ax=ax_scatter2, s=20, alpha=0.6, palette=SNS_CMAP)\n ax_scatter2.set_title(\"Color entropy vs \" + str(\"GDM_g\"))\n \n plt.tight_layout()\n plt.show()\n\ndf = df[[\"image_path\", \"target_name\", \"target\"]]\ndf = df.pivot_table(\n index=[\"image_path\"],\n columns='target_name',\n values=\"target\").reset_index()\n\ndf = extract_feat(df, DATA_DIR)\ncustom_plot(df)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-26T15:50:59.822612Z","iopub.execute_input":"2025-11-26T15:50:59.823183Z","iopub.status.idle":"2025-11-26T15:52:11.175461Z","shell.execute_reply.started":"2025-11-26T15:50:59.823156Z","shell.execute_reply":"2025-11-26T15:52:11.174694Z"},"jupyter":{"source_hidden":true},"_kg_hide-input":true},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#
2. EfficientNet Baseline
","metadata":{}},{"cell_type":"code","source":"!pip install -q lightning\nimport torch\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader, random_split, Dataset\nfrom typing import Union, Optional, Callable\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.base import BaseEstimator\nfrom sklearn.model_selection import train_test_split\nfrom PIL import Image\n\nimport lightning as L","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:11:50.868342Z","iopub.execute_input":"2025-11-30T08:11:50.868723Z","iopub.status.idle":"2025-11-30T08:13:14.874140Z","shell.execute_reply.started":"2025-11-30T08:11:50.868705Z","shell.execute_reply":"2025-11-30T08:13:14.873316Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"class CSIRODataModule(L.LightningDataModule):\n\n IMG_SIZE:int = 384\n VAL_SPLIT:float = 0.1\n NUM_WORKERS = 4\n\n TRAIN_TRANSFORMS = transforms.Compose([\n transforms.Resize((IMG_SIZE, IMG_SIZE)),\n transforms.RandomAffine(\n degrees=5, # very slight rotation\n translate=(0.05, 0.05), # ±5% translation\n scale=None,\n shear=None,\n ),\n transforms.RandomHorizontalFlip(p=0.5),\n \n #slight color transforms (since color is one of our imp signals)\n transforms.ColorJitter(\n brightness=0.10,\n contrast=0.10,\n saturation=0.05,\n hue=0.02,\n ),\n \n # light sharpness + noise\n transforms.RandomAdjustSharpness(sharpness_factor=1.1, p=0.3),\n transforms.GaussianBlur(kernel_size=3, sigma=(0.1, 1.0)),\n \n transforms.ToTensor(),\n transforms.Normalize(mean=(0.485,0.456,0.406),\n std=(0.229,0.224,0.225)),\n ])\n \n VAL_TRANSFORMS = transforms.Compose([\n transforms.Resize((IMG_SIZE, IMG_SIZE)),\n transforms.ToTensor(),\n transforms.Normalize(mean=(0.485,0.456,0.406),\n std=(0.229,0.224,0.225)),\n ])\n\n TARGETS:list[str] = ['Dry_Clover_g', 'Dry_Dead_g', 'Dry_Green_g', 'Dry_Total_g', 'GDM_g']\n\n def __init__(\n self,\n data_dir: str,\n df: Union[pd.DataFrame, str],\n batch_size:int = 128\n )->None:\n super().__init__()\n self.data_dir = data_dir\n self.df = df\n self.batch_size = batch_size\n\n # --dataclass\n class ImageDataset(Dataset):\n def __init__(self, image_dir, df, transform, targets, scaler, test=False):\n self.samples = list(df.itertuples(index=False))\n self.image_dir = image_dir\n self.transform = transform\n self.targets = targets\n self.scaler = scaler\n self.test = test\n\n def __len__(self):\n return len(self.samples)\n\n def __getitem__(self, idx):\n row = self.samples[idx]\n\n # image\n img_path = os.path.join(self.image_dir, row.image_path)\n image = Image.open(img_path).convert(\"RGB\")\n image = self.transform(image)\n\n # targets\n if not self.test:\n t = torch.tensor([getattr(row, t) for t in self.targets], dtype=torch.float32)\n \n if self.scaler:\n t = torch.tensor(self.scaler.transform([t.numpy()])[0], dtype=torch.float32)\n else:\n t = torch.tensor([0, 0, 0, 0, 0])\n \n return image, t\n #---\n\n def setup(self, stage=None):\n\n if stage == \"fit\" or stage is None:\n df_train, df_val = train_test_split(\n self.df, test_size=self.VAL_SPLIT, random_state=42, shuffle=True)\n self.target_scaler = StandardScaler()\n self.target_scaler.fit(df_train[CSIRODataModule.TARGETS])\n \n self.train_dataset = self.ImageDataset(\n self.data_dir, df_train, self.TRAIN_TRANSFORMS, CSIRODataModule.TARGETS, self.target_scaler\n )\n self.val_dataset = self.ImageDataset(\n self.data_dir, df_val, self.VAL_TRANSFORMS, CSIRODataModule.TARGETS, self.target_scaler\n )\n rprint(f\"FIT {self.__class__} with \\n --num_workers--: {self.NUM_WORKERS} \\n --batch_size--: {self.batch_size}, \\n [purple]train_images[/]: {len(df_train)} \\n [purple]val_images[/]: {len(df_val)}\\n image_size: {self.IMG_SIZE}\")\n\n def train_dataloader(self):\n return DataLoader(self.train_dataset, \n batch_size=self.batch_size, \n shuffle=True,\n num_workers=self.NUM_WORKERS)\n\n def val_dataloader(self):\n return DataLoader(self.val_dataset, \n batch_size=self.batch_size,\n num_workers=self.NUM_WORKERS)\n\ndf = pd.read_csv(os.path.join(DATA_DIR, \"train.csv\"))\ndf = df[[\"image_path\", \"target_name\", \"target\"]]\ndf = df.pivot_table(\n index=[\"image_path\"],\n columns='target_name',\n values=\"target\").reset_index()\n\ndf_test = pd.read_csv(os.path.join(DATA_DIR, \"test.csv\"))\ndf_test = df_test[[\"image_path\"]].drop_duplicates().reset_index(drop=True)\n\nfor t in CSIRODataModule.TARGETS:\n df_test[t] = 0","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:13:14.875098Z","iopub.execute_input":"2025-11-30T08:13:14.875656Z","iopub.status.idle":"2025-11-30T08:13:14.968116Z","shell.execute_reply.started":"2025-11-30T08:13:14.875614Z","shell.execute_reply":"2025-11-30T08:13:14.967497Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"#==================================\n#== augmentations ===\n#==================================\nif not SUBMISSION:\n from copy import deepcopy\n \n img_path = os.path.join(DATA_DIR, df.iloc[0][\"image_path\"])\n img = Image.open(img_path).convert(\"RGB\")\n vis_transforms = deepcopy(CSIRODataModule.TRAIN_TRANSFORMS.transforms[1:-2])\n \n ncols = 2\n nrows = (len(vis_transforms)+ncols-1)//ncols\n fig, ax = plt.subplots(nrows, ncols, figsize=(15, 6*nrows))\n ax = ax.flatten()\n ax[0].axis(\"off\")\n ax[0].set_title(\"Original\")\n ax[0].imshow(img)\n for i, transform in enumerate(vis_transforms):\n try:\n transform.p = 1.0\n except:\n pass\n aug_img = transform(img)\n ax[i+1].imshow(aug_img)\n ax[i+1].set_title(type(transform).__name__)\n ax[i+1].axis(\"off\")\n \n plt.tight_layout()\n plt.show()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-29T12:30:46.344974Z","iopub.execute_input":"2025-11-29T12:30:46.346046Z","iopub.status.idle":"2025-11-29T12:30:49.416948Z","shell.execute_reply.started":"2025-11-29T12:30:46.346019Z","shell.execute_reply":"2025-11-29T12:30:49.414939Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":".import torch.nn as nn\n\nclass EfficientNetModule(L.LightningModule):\n\n TARGET_WEIGHTS = torch.tensor(\n [0.1, 0.1, 0.1, 0.5, 0.2],\n dtype=torch.float32\n )\n\n #========================\n class EfficientNetCore(nn.Module):\n def __init__(self):\n super().__init__()\n backbone = efficientnet_b0(weights=EfficientNet_B0_Weights.IMAGENET1K_V1)\n in_feat = backbone.classifier[1].in_features\n backbone.classifier = nn.Identity()\n \n self.backbone = backbone\n self.head = nn.Linear(in_feat, 5)\n \n def forward(self, x):\n x = self.backbone(x)\n x = self.head(x)\n return x\n #========================\n \n def __init__(self,\n lr:float = 1e-2,\n target_scaler = None,\n ):\n super().__init__()\n self.model = self.EfficientNetCore()\n \n self.lr:float = lr\n self.loss_fn:nn.Module = nn.MSELoss()\n self.target_scaler = target_scaler\n self.register_buffer(\"target_weights_device\", self.TARGET_WEIGHTS)\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n # ===== Custom Weighted Global R2 =====\n def weighted_r2(self, y_true, y_pred):\n # unscale before calculating metric\n if self.target_scaler:\n # y_true = torch.tensor(self.target_scaler.inverse_transform(y_true.cpu()), device=y_true.device)\n # y_pred = torch.tensor(self.target_scaler.inverse_transform(y_pred.cpu()), device=y_pred.device)\n y_true = torch.from_numpy(\n self.target_scaler.inverse_transform(y_true.cpu())\n ).float().to(y_true.device)\n\n y_pred = torch.from_numpy(\n self.target_scaler.inverse_transform(y_pred.cpu())\n ).float().to(y_true.device)\n \n w = self.target_weights_device\n w = w.unsqueeze(0)\n\n ss_res = torch.sum(w * (y_true - y_pred) ** 2)\n ss_tot = torch.sum(w * (y_true - torch.mean(y_true, dim=0, keepdim=True)) ** 2)\n\n return 1 - ss_res / (ss_tot + 1e-8)\n\n def configure_optimizers(self):\n optimizer = torch.optim.Adam(self.parameters(), lr=self.lr)\n scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, 20, 0.05*self.lr)\n return [optimizer], [scheduler]\n\n def training_step(self, batch, batch_idx):\n x, y = batch\n preds = self(x)\n loss = self.loss_fn(preds, y)\n self.log(f\"train_loss\", loss, prog_bar=True)\n return loss\n \n def validation_step(self, batch, batch_idx):\n x, y = batch\n preds = self(x)\n val_loss = self.loss_fn(preds, y)\n r2_metric = self.weighted_r2(y, preds)\n self.log_dict({\n \"val_loss\": val_loss,\n \"r2\": r2_metric,\n })\n\n def offload_wts(self, path=\"model.pt\"):\n torch.save(self.model.state_dict(), path)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:13:39.283493Z","iopub.execute_input":"2025-11-30T08:13:39.283835Z","iopub.status.idle":"2025-11-30T08:13:39.338020Z","shell.execute_reply.started":"2025-11-30T08:13:39.283793Z","shell.execute_reply":"2025-11-30T08:13:39.337293Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"lt_data_module = CSIRODataModule(DATA_DIR, df, batch_size=64)\nlt_data_module.setup()\nmodel = EfficientNetModule(target_scaler = lt_data_module.target_scaler)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:13:45.583253Z","iopub.execute_input":"2025-11-30T08:13:45.583880Z","iopub.status.idle":"2025-11-30T08:13:45.988920Z","shell.execute_reply.started":"2025-11-30T08:13:45.583856Z","shell.execute_reply":"2025-11-30T08:13:45.988203Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"if not SUBMISSION:\n from lightning.pytorch.loggers import WandbLogger\n from kaggle_secrets import UserSecretsClient\n import wandb\n \n wandb_key = UserSecretsClient().get_secret(\"WANDB\")\n wandb.login(key=wandb_key)\n \n wandb_logger = WandbLogger(project=\"kaggle-CSIRO\")\n\nelse:\n wandb_logger = None","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:13:46.060715Z","iopub.execute_input":"2025-11-30T08:13:46.060976Z","iopub.status.idle":"2025-11-30T08:13:58.535053Z","shell.execute_reply.started":"2025-11-30T08:13:46.060957Z","shell.execute_reply":"2025-11-30T08:13:58.534281Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"from lightning.pytorch.callbacks import DeviceStatsMonitor\n\ntrainer = L.Trainer(\n logger=wandb_logger,\n # profiler=\"simple\",\n # fast_dev_run=2,\n devices=1,\n precision=\"16-mixed\",\n max_epochs=20,\n\n callbacks=[DeviceStatsMonitor()]\n)\ntrainer.fit(\n model,\n lt_data_module\n)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:13:58.536119Z","iopub.execute_input":"2025-11-30T08:13:58.536556Z","iopub.status.idle":"2025-11-30T08:20:28.102601Z","shell.execute_reply.started":"2025-11-30T08:13:58.536538Z","shell.execute_reply":"2025-11-30T08:20:28.101968Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"model.offload_wts()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:20:28.103506Z","iopub.execute_input":"2025-11-30T08:20:28.103808Z","iopub.status.idle":"2025-11-30T08:20:28.202474Z","shell.execute_reply.started":"2025-11-30T08:20:28.103779Z","shell.execute_reply":"2025-11-30T08:20:28.201924Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"test_dataset = lt_data_module.ImageDataset(DATA_DIR,\n df_test,\n lt_data_module.VAL_TRANSFORMS,None,None,test=True)\n\ntest_dl = DataLoader(\n test_dataset,\n batch_size=32,\n shuffle=False,\n num_workers=4,\n)\n\nmodel.eval()\nall_preds = []\nwith torch.no_grad():\n for imgs, _ in tqdm(test_dl):\n preds = model(imgs)\n if model.target_scaler is not None:\n preds = model.target_scaler.inverse_transform(preds.cpu().numpy())\n else:\n preds = preds.cpu().numpy()\n\n all_preds.append(preds)\n\nall_preds = np.vstack(all_preds)","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:20:28.204048Z","iopub.execute_input":"2025-11-30T08:20:28.204306Z","iopub.status.idle":"2025-11-30T08:20:28.776438Z","shell.execute_reply.started":"2025-11-30T08:20:28.204289Z","shell.execute_reply":"2025-11-30T08:20:28.775630Z"}},"outputs":[],"execution_count":null},{"cell_type":"code","source":"%%capture\nstacked_preds = np.vstack(all_preds)\ndf_test[CSIRODataModule.TARGETS] = stacked_preds\n# wandb.finish()","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:20:28.777579Z","iopub.execute_input":"2025-11-30T08:20:28.777925Z","iopub.status.idle":"2025-11-30T08:20:28.846329Z","shell.execute_reply.started":"2025-11-30T08:20:28.777902Z","shell.execute_reply":"2025-11-30T08:20:28.845680Z"}},"outputs":[],"execution_count":null},{"cell_type":"markdown","source":"#
3. Submission
","metadata":{}},{"cell_type":"code","source":"df_res = df_test.melt(\n id_vars=['image_path'],\n value_vars=['Dry_Clover_g', 'Dry_Dead_g', 'Dry_Green_g', 'Dry_Total_g', 'GDM_g'],\n var_name='target_name',\n value_name='target'\n)\n\ndf_res[\"sample_id\"] = df_res.apply(lambda x: f\"{x['image_path'][5:-4]}__{x['target_name']}\", axis=1)\ndf_res = df_res[['sample_id', 'target']]\n\ndf_sub = pd.read_csv(r\"/kaggle/input/csiro-biomass/sample_submission.csv\")\ndf_sub = df_sub.drop('target', axis=1)\ndf_sub = df_sub.merge(right=df_res)\n\ndf_sub.to_csv(\"submission.csv\", index=False)\ndf_sub","metadata":{"trusted":true,"execution":{"iopub.status.busy":"2025-11-30T08:20:28.847151Z","iopub.execute_input":"2025-11-30T08:20:28.847403Z","iopub.status.idle":"2025-11-30T08:20:28.937907Z","shell.execute_reply.started":"2025-11-30T08:20:28.847386Z","shell.execute_reply":"2025-11-30T08:20:28.937246Z"}},"outputs":[],"execution_count":null}]}