{"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"},"kaggle":{"accelerator":"none","dataSources":[{"sourceId":59093,"databundleVersionId":7469972,"sourceType":"competition"}],"isInternetEnabled":true,"language":"python","sourceType":"notebook","isGpuEnabled":false}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"code","source":"import os\nimport gc\nimport random\nimport shutil\nfrom time import time\nimport typing as tp\nfrom pathlib import Path\n\nimport numpy as np\nimport pandas as pd\n\nfrom tqdm import tqdm\nfrom sklearn.model_selection import StratifiedGroupKFold\nimport torch\nfrom torch import nn\nfrom torch import optim\nfrom torch.optim import lr_scheduler\nfrom torch.cuda import amp\nimport timm\nimport albumentations as A\nfrom albumentations.pytorch import ToTensorV2\n\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0\"\n\n# ROOT = Path.cwd().parent\n# INPUT = ROOT / \"input\"\n# OUTPUT = ROOT / \"output\"\n# INPUT = r\"D:\\lwh\\data\\hms\"\n# OUTPUT = ROOT / \"output\"\n# SRC = ROOT / \"src\"\n\n\n\n\nclass HMSHBACSpecModel(nn.Module):\n\n    def __init__(\n            self,\n            model_name: str,\n            pretrained: bool,\n            in_channels: int,\n            num_classes: int,\n        ):\n        super().__init__()\n        self.model = timm.create_model(\n            model_name=model_name, pretrained=pretrained,\n            num_classes=num_classes, in_chans=in_channels)\n\n    def forward(self, x):\n        h = self.model(x)\n\n        return h\n\n\nFilePath = tp.Union[str, Path]\nLabel = tp.Union[int, float, np.ndarray]\n\n\nclass HMSHBACSpecDataset(torch.utils.data.Dataset):\n\n    def __init__(\n            self,\n            image_paths: tp.Sequence[FilePath],\n            labels: tp.Sequence[Label],\n            transform: A.Compose,\n    ):\n        self.image_paths = image_paths\n        self.labels = labels\n        self.transform = transform\n\n    def __len__(self):\n        return len(self.image_paths)\n\n    def __getitem__(self, index: int):\n        img_path = self.image_paths[index]\n        label = self.labels[index]\n\n        img = np.load(img_path)  # shape: (Hz, Time) = (400, 300)\n\n        # log transform\n        img = np.clip(img, np.exp(-4), np.exp(8))\n        img = np.log(img)\n\n        # normalize per image\n        eps = 1e-6\n        img_mean = img.mean(axis=(0, 1))\n        img = img - img_mean\n        img_std = img.std(axis=(0, 1))\n        img = img / (img_std + eps)\n\n        img = img[..., None]  # shape: (Hz, Time) -> (Hz, Time, Channel)\n        img = self._apply_transform(img)\n\n        return {\"data\": img, \"target\": label}\n\n    def _apply_transform(self, img: np.ndarray):\n        \"\"\"apply transform to image and mask\"\"\"\n        transformed = self.transform(image=img)\n        img = transformed[\"image\"]\n        return img\n\n\nclass KLDivLossWithLogits(nn.KLDivLoss):\n\n    def __init__(self):\n        super().__init__(reduction=\"batchmean\")\n\n    def forward(self, y, t):\n        y = nn.functional.log_softmax(y, dim=1)\n        loss = super().forward(y, t)\n\n        return loss\n\n\nclass KLDivLossWithLogitsForVal(nn.KLDivLoss):\n\n    def __init__(self):\n        \"\"\"\"\"\"\n        super().__init__(reduction=\"batchmean\")\n        self.log_prob_list = []\n        self.label_list = []\n\n    def forward(self, y, t):\n        y = nn.functional.log_softmax(y, dim=1)\n        self.log_prob_list.append(y.numpy())\n        self.label_list.append(t.numpy())\n\n    def compute(self):\n        log_prob = np.concatenate(self.log_prob_list, axis=0)\n        label = np.concatenate(self.label_list, axis=0)\n        final_metric = super().forward(\n            torch.from_numpy(log_prob),\n            torch.from_numpy(label)\n        ).item()\n        self.log_prob_list = []\n        self.label_list = []\n\n        return final_metric\n\n\nclass CFG:\n    model_name = \"efficientnet_b2\"\n    img_size = 512\n    max_epoch = 9\n    batch_size = 16\n    lr = 1.0e-03\n    weight_decay = 1.0e-02\n    es_patience =  5\n    seed = 1086\n    deterministic = True\n    enable_amp = True\n    device = \"cuda\"\n\n\ndef set_random_seed(seed: int = 42, deterministic: bool = False):\n    \"\"\"Set seeds\"\"\"\n    random.seed(seed)\n    np.random.seed(seed)\n    os.environ[\"PYTHONHASHSEED\"] = str(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)  # type: ignore\n    torch.backends.cudnn.deterministic = deterministic  # type: ignore\n\n\ndef to_device(\n        tensors: tp.Union[tp.Tuple[torch.Tensor], tp.Dict[str, torch.Tensor]],\n        device: torch.device, *args, **kwargs\n):\n    if isinstance(tensors, tuple):\n        return (t.to(device, *args, **kwargs) for t in tensors)\n    elif isinstance(tensors, dict):\n        return {\n            k: t.to(device, *args, **kwargs) for k, t in tensors.items()}\n    else:\n        return tensors.to(device, *args, **kwargs)\n\n\ndef get_path_label(val_fold, train_all: pd.DataFrame):\n    \"\"\"Get file path and target info.\"\"\"\n\n    train_idx = train_all[train_all[\"fold\"] != val_fold].index.values\n    val_idx = train_all[train_all[\"fold\"] == val_fold].index.values\n    img_paths = []\n    labels = train_all[CLASSES].values\n    for label_id in train_all[\"label_id\"].values:\n        img_path = TRAIN_SPEC_SPLIT + f\"/{label_id}.npy\"\n        img_paths.append(img_path)\n\n    train_data = {\n        \"image_paths\": [img_paths[idx] for idx in train_idx],\n        \"labels\": [labels[idx].astype(\"float32\") for idx in train_idx]}\n\n    val_data = {\n        \"image_paths\": [img_paths[idx] for idx in val_idx],\n        \"labels\": [labels[idx].astype(\"float32\") for idx in val_idx]}\n\n    return train_data, val_data, train_idx, val_idx\n\n\ndef get_transforms(CFG):\n    train_transform = A.Compose([\n        A.Resize(p=1.0, height=CFG.img_size, width=CFG.img_size),\n        ToTensorV2(p=1.0)\n    ])\n    val_transform = A.Compose([\n        A.Resize(p=1.0, height=CFG.img_size, width=CFG.img_size),\n        ToTensorV2(p=1.0)\n    ])\n    return train_transform, val_transform\n\n\ndef train_one_fold(CFG, val_fold, train_all, output_path):\n    \"\"\"Main\"\"\"\n    torch.backends.cudnn.benchmark = True\n    set_random_seed(CFG.seed, deterministic=CFG.deterministic)\n    device = torch.device(CFG.device)\n\n    train_path_label, val_path_label, _, _ = get_path_label(val_fold, train_all)\n    train_transform, val_transform = get_transforms(CFG)\n\n    train_dataset = HMSHBACSpecDataset(**train_path_label, transform=train_transform)\n    val_dataset = HMSHBACSpecDataset(**val_path_label, transform=val_transform)\n\n    train_loader = torch.utils.data.DataLoader(\n        train_dataset, batch_size=CFG.batch_size, num_workers=4, shuffle=True, drop_last=True)\n    val_loader = torch.utils.data.DataLoader(\n        val_dataset, batch_size=CFG.batch_size, num_workers=4, shuffle=False, drop_last=False)\n\n    model = HMSHBACSpecModel(\n        model_name=CFG.model_name, pretrained=True, num_classes=6, in_channels=1)\n    model.to(device)\n\n    optimizer = optim.AdamW(params=model.parameters(), lr=CFG.lr, weight_decay=CFG.weight_decay)\n    scheduler = lr_scheduler.OneCycleLR(\n        optimizer=optimizer, epochs=CFG.max_epoch,\n        pct_start=0.0, steps_per_epoch=len(train_loader),\n        max_lr=CFG.lr, div_factor=25, final_div_factor=4.0e-01\n    )\n\n    loss_func = KLDivLossWithLogits()\n    loss_func.to(device)\n    loss_func_val = KLDivLossWithLogitsForVal()\n\n    use_amp = CFG.enable_amp\n    scaler = amp.GradScaler(enabled=use_amp)\n\n    best_val_loss = 1.0e+09\n    best_epoch = 0\n    train_loss = 0\n\n    for epoch in range(1, CFG.max_epoch + 1):\n        epoch_start = time()\n        model.train()\n        for batch in tqdm(train_loader):\n            batch = to_device(batch, device)\n            x, t = batch[\"data\"], batch[\"target\"]\n\n            optimizer.zero_grad()\n            with amp.autocast(use_amp):\n                y = model(x)\n                loss = loss_func(y, t)\n            scaler.scale(loss).backward()\n            scaler.step(optimizer)\n            scaler.update()\n            scheduler.step()\n            train_loss += loss.item()\n\n        train_loss /= len(train_loader)\n\n        model.eval()\n        for batch in val_loader:\n            x, t = batch[\"data\"], batch[\"target\"]\n            x = to_device(x, device)\n            with torch.no_grad(), amp.autocast(use_amp):\n                y = model(x)\n            y = y.detach().cpu().to(torch.float32)\n            loss_func_val(y, t)\n        val_loss = loss_func_val.compute()\n        if val_loss < best_val_loss:\n            best_epoch = epoch\n            best_val_loss = val_loss\n            # print(\"save model\")\n            torch.save(model.state_dict(), str(output_path / f'snapshot_epoch_{epoch}.pth'))\n\n        elapsed_time = time() - epoch_start\n        print(\n            f\"[epoch {epoch}] train loss: {train_loss: .6f}, val loss: {val_loss: .6f}, elapsed_time: {elapsed_time: .3f}\")\n\n        if epoch - best_epoch > CFG.es_patience:\n            print(\"Early Stopping!\")\n            break\n\n        train_loss = 0\n\n    return val_fold, best_epoch, best_val_loss\n\n\n\n\nif __name__ == '__main__':\n    # DATA = r\"D:\\lwh\\data\\hms\\hms-harmful-brain-activity-classification\"\n    TRAIN_SPEC = r\"D:\\lwh\\data\\hms\\hms-harmful-brain-activity-classification\\train_spectrograms\"\n    TEST_SPEC = r\"D:\\lwh\\data\\hms\\hms-harmful-brain-activity-classification\\test_spectrograms\"\n\n    # TMP = ROOT / \"tmp\"\n    TRAIN_SPEC_SPLIT = r\"D:\\lwh\\data\\hms\\hms-harmful-brain-activity-classification\\train_spectrograms_split\"\n    TEST_SPEC_SPLIT = r\"D:\\lwh\\data\\hms\\hms-harmful-brain-activity-classification\\test_spectrograms_split\"\n    # TMP.mkdir(exist_ok=True)\n    # TRAIN_SPEC_SPLIT.mkdir(exist_ok=True)\n    # TEST_SPEC_SPLIT.mkdir(exist_ok=True)\n\n    RANDAM_SEED = 1086\n    CLASSES = [\"seizure_vote\", \"lpd_vote\", \"gpd_vote\", \"lrda_vote\", \"grda_vote\", \"other_vote\"]\n    N_CLASSES = len(CLASSES)\n    FOLDS = [0, 1, 2, 3, 4]\n    N_FOLDS = len(FOLDS)\n\n    train = pd.read_csv(r\"D:\\lwh\\data\\hms\\hms-harmful-brain-activity-classification\\train.csv\")\n\n    # convert vote to probability\n    train[CLASSES] /= train[CLASSES].sum(axis=1).values[:, None]\n\n\n    # 数据类型名：eeg_id ,eeg_sub_id, eeg_label_offset_seconds, spectrogram_id, spectrogram_sub_id, spectrogram_label_offset_seconds,\n    # label_id,patient_id,expert_consensus,     seizure_vote,lpd_vote,gpd_vote,lrda_vote,grda_vote,other_vote\n    train = train.groupby(\"spectrogram_id\").head(1).reset_index(drop=True)\n    # print(train.shape)\n\n    sgkf = StratifiedGroupKFold(n_splits=N_FOLDS, shuffle=True, random_state=RANDAM_SEED)\n\n    train[\"fold\"] = -1\n\n    for fold_id, (_, val_idx) in enumerate(\n            sgkf.split(train, y=train[\"expert_consensus\"], groups=train[\"patient_id\"])\n    ):\n        train.loc[val_idx, \"fold\"] = fold_id\n    train.groupby(\"fold\")[CLASSES].sum()\n    train.groupby(\"spectrogram_id\")\n    # for k in train[\"label_id\"]:\n    #     print(k)\n    #\n    # for spec_id, df in tqdm(train.groupby(\"spectrogram_id\")):\n    #     # print(f\"{spec_id}.parquet\")\n    #     # print(TRAIN_SPEC + f\"/{spec_id}.parquet\")\n    #     spec = pd.read_parquet(TRAIN_SPEC + f\"/{spec_id}.parquet\")\n    #\n    #     spec_arr = spec.fillna(0).values[:, 1:].T.astype(\"float32\")  # (Hz, Time) = (400, 300)\n    #\n    #     for spec_offset, label_id in df[\n    #         [\"spectrogram_label_offset_seconds\", \"label_id\"]\n    #     ].values:\n    #         spec_offset = int(spec_offset)\n    #         # print(label_id)\n    #         label_id = int(label_id)\n    #         print(label_id)\n    #         spec_offset = spec_offset // 2\n    #         split_spec_arr = spec_arr[:, spec_offset: spec_offset + 300]\n    #         np.save(TRAIN_SPEC_SPLIT + f\"/{label_id}.npy\", split_spec_arr)\n\n    score_list = []\n    for fold_id in FOLDS:\n        output_path = Path(f\"fold{fold_id}\")\n        output_path.mkdir(exist_ok=True)\n        print(f\"[fold{fold_id}]\")\n        score_list.append(train_one_fold(CFG, fold_id, train, output_path))\n\n    print(score_list)\n\n    best_log_list = []\n    for (fold_id, best_epoch, _) in score_list:\n\n        exp_dir_path = Path(f\"fold{fold_id}\")\n        best_model_path = exp_dir_path / f\"snapshot_epoch_{best_epoch}.pth\"\n        copy_to = f\"./best_model_fold{fold_id}.pth\"\n        shutil.copy(best_model_path, copy_to)\n\n        for p in exp_dir_path.glob(\"*.pth\"):\n            p.unlink()\n\n\n    def run_inference_loop(model, loader, device):\n        model.to(device)\n        model.eval()\n        pred_list = []\n        with torch.no_grad():\n            for batch in tqdm(loader):\n                x = to_device(batch[\"data\"], device)\n                y = model(x)\n                pred_list.append(y.softmax(dim=1).detach().cpu().numpy())\n\n        pred_arr = np.concatenate(pred_list)\n        del pred_list\n        return pred_arr\n\n\n    label_arr = train[CLASSES].values\n    oof_pred_arr = np.zeros((len(train), N_CLASSES))\n    score_list = []\n\n\n    for fold_id in range(N_FOLDS):\n        print(f\"\\n[fold {fold_id}]\")\n        device = torch.device(CFG.device)\n\n        # # get_dataloader\n        _, val_path_label, _, val_idx = get_path_label(fold_id, train)\n        _, val_transform = get_transforms(CFG)\n        val_dataset = HMSHBACSpecDataset(**val_path_label, transform=val_transform)\n        val_loader = torch.utils.data.DataLoader(\n            val_dataset, batch_size=CFG.batch_size, num_workers=4, shuffle=False, drop_last=False)\n\n        # # get model\n        model_path = f\"./best_model_fold{fold_id}.pth\"\n        model = HMSHBACSpecModel(\n            model_name=CFG.model_name, pretrained=False, num_classes=6, in_channels=1)\n        model.load_state_dict(torch.load(model_path, map_location=device))\n\n        # # inference\n        val_pred = run_inference_loop(model, val_loader, device)\n        oof_pred_arr[val_idx] = val_pred\n\n        del val_idx, val_path_label\n        del model, val_loader\n        torch.cuda.empty_cache()\n        gc.collect()\n\n    # sys.path.append('/kaggle/input/kaggle-kl-div')\n    from common.kaggle_kl_div import score\n\n    true = train[[\"label_id\"] + CLASSES].copy()\n\n    oof = pd.DataFrame(oof_pred_arr, columns=CLASSES)\n    oof.insert(0, \"label_id\", train[\"label_id\"])\n\n    cv_score = score(solution=true, submission=oof, row_id_column_name='label_id')\n    print('CV Score KL-Div for ResNet34d',cv_score)","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}