{"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":"code","source":"import numpy as np \nimport pandas as pd\nimport random\nimport os\nimport math\nimport time","metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install timm","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from sklearn.metrics import accuracy_score\nfrom sklearn.utils import class_weight\nfrom PIL import Image as pil_image\nfrom tqdm import tqdm\nimport scipy\n\nimport matplotlib\nimport matplotlib.pyplot as plt\nimport plotly.express as px\nimport plotly.graph_objects as go","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom torch.utils.data import DataLoader\n\nimport timm\nfrom timm.optim import Lookahead, RAdam","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"IMG_SIZE = 512\nSEED = 42\nPROJECT_FOLDER = \"../input/hotel-id-2022-fgvc9/\"\nDATA_FOLDER = \"../input/hotelid-2022-train-images-512x512/\"\nIMAGE_FOLDER = DATA_FOLDER\nOUTPUT_FOLDER = \"\"\n\ntrain_df = pd.read_csv(os.path.join(DATA_FOLDER, 'train.csv'))","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def seed_everything(seed):\n    random.seed(seed)\n    os.environ['PYTHONHASHSEED'] = str(seed)\n    np.random.seed(seed)\n    torch.manual_seed(seed)\n    torch.cuda.manual_seed(seed)\n    torch.backends.cudnn.deterministic = True","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import albumentations as A\nimport albumentations.pytorch as APT\nimport cv2 \n\n\nbase_transform = A.Compose([\n    A.ToFloat(),\n    APT.transforms.ToTensorV2(),\n])\n\ntrain_transform = A.Compose([\n    #A.HorizontalFlip(p=0.2),\n    #A.ShiftScaleRotate(p=0.2, border_mode=cv2.BORDER_CONSTANT),\n    #A.RandomBrightness(p=0.2),\n    A.CoarseDropout(p=1., max_holes=1, \n                    min_height=IMG_SIZE//4, max_height=IMG_SIZE//2,\n                    min_width=IMG_SIZE//4,  max_width=IMG_SIZE//2, \n                    fill_value=(255,0,0)), #Occlusions in test data simulation\n    A.ToFloat(),\n    APT.transforms.ToTensorV2(),\n])\n\nvalid_transform = A.Compose([\n    A.ToFloat(),\n    APT.transforms.ToTensorV2(),\n])\n","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class HotelTrainDataset:\n    def __init__(self, data, transform=None, data_path=\"\"):\n        self.data = data\n        self.data_path = data_path\n        self.data_path += \"images/\"\n        self.transform = transform\n        self.fake_load = False\n\n    def __len__(self):\n        return len(self.data)\n    \n    def __getitem__(self, idx):\n        record = self.data.iloc[idx]\n        image_path = self.data_path + record[\"image_id\"]\n\n        if self.fake_load:\n            image = np.random.randint(0, 255, (32, 32, 3)).astype(np.uint8)\n        else:\n            image = np.array(pil_image.open(image_path)).astype(np.uint8)\n\n        if self.transform:\n            transformed = self.transform(image=image)\n        \n        return {\n            \"image\" : transformed[\"image\"],\n            \"target\" : record['hotel_id_code'],\n        }","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class ArcMarginProduct(nn.Module):\n    r\"\"\"Implement of large margin arc distance: :\n        Args:\n            in_features: size of each input sample\n            out_features: size of each output sample\n            s: norm of input feature\n            m: margin\n            cos(theta + m)\n        \"\"\"\n    def __init__(self, in_features, out_features, s=30.0, m=0.50, easy_margin=False):\n        super(ArcMarginProduct, self).__init__()\n        self.in_features = in_features\n        self.out_features = out_features\n        self.s = s\n        self.m = m\n        self.weight = nn.Parameter(torch.FloatTensor(out_features, in_features))\n        nn.init.xavier_uniform_(self.weight)\n\n        self.easy_margin = easy_margin\n        self.cos_m = math.cos(m)\n        self.sin_m = math.sin(m)\n        self.th = math.cos(math.pi - m)\n        self.mm = math.sin(math.pi - m) * m\n\n    def forward(self, input, label):\n        # --------------------------- cos(theta) & phi(theta) ---------------------------\n        cosine = F.linear(F.normalize(input), F.normalize(self.weight))\n        sine = torch.sqrt((1.0 - torch.pow(cosine, 2)).clamp(0, 1))\n        phi = cosine * self.cos_m - sine * self.sin_m\n        if self.easy_margin:\n            phi = torch.where(cosine > 0, phi, cosine)\n        else:\n            phi = torch.where(cosine > self.th, phi, cosine - self.mm)\n        # --------------------------- convert label to one-hot ---------------------------\n        # one_hot = torch.zeros(cosine.size(), requires_grad=True, device='cuda')\n        one_hot = torch.zeros(cosine.size(), device='cuda')\n        one_hot.scatter_(1, label.view(-1, 1).long(), 1)\n        # -------------torch.where(out_i = {x_i if condition_i else y_i) -------------\n        output = (one_hot * phi) + ((1.0 - one_hot) * cosine)  # you can use torch.where if your torch.__version__ is 0.4\n        output *= self.s\n\n        return output","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class HotelIdModel(nn.Module):\n    def __init__(self, out_features, embed_size=256, backbone_name=\"efficientnet_b3\"):\n        super(HotelIdModel, self).__init__()\n\n        self.embed_size = embed_size\n        self.backbone = timm.create_model(backbone_name, pretrained=True)\n        in_features = self.backbone.get_classifier().in_features\n\n        fc_name, _ = list(self.backbone.named_modules())[-1]\n        if fc_name == 'classifier':\n            self.backbone.classifier = nn.Identity()\n        elif fc_name == 'head.fc':\n            self.backbone.head.fc = nn.Identity()\n        elif fc_name == 'fc':\n            self.backbone.fc = nn.Identity()\n        else:\n            raise Exception(\"unknown classifier layer: \" + fc_name)\n\n        self.arc_face = ArcMarginProduct(self.embed_size, out_features, s=64.0, m=0.50, easy_margin=False) #Changed ArcMargin parameters s=64.0 & m=0.50\n\n        self.post = nn.Sequential(\n            nn.utils.weight_norm(nn.Linear(in_features, self.embed_size*2), dim=None),\n            nn.BatchNorm1d(self.embed_size*2),\n            nn.Dropout(0.2),\n            nn.utils.weight_norm(nn.Linear(self.embed_size*2, self.embed_size)),\n            nn.BatchNorm1d(self.embed_size),\n        )\n\n        print(f\"Model {backbone_name} ArcMarginProduct - Features: {in_features}, Embeds: {self.embed_size}\")\n        \n    def forward(self, input, targets = None):\n        x = self.backbone(input)\n        x = x.view(x.size(0), -1)\n        x = self.post(x)\n        \n        if targets is not None:\n            logits = self.arc_face(x, targets)\n            return logits\n        \n        return x","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def get_embeds(loader, model, bar_desc=\"Generating embeds\"):\n    targets_all = []\n    outputs_all = []\n    \n    model.eval()\n    with torch.no_grad():\n        t = tqdm(loader, desc=bar_desc)\n        for i, sample in enumerate(t):\n            input = sample['image'].to(args.device)\n            target = sample['target'].to(args.device)\n            output = model(input)\n\n            targets_all.extend(target.cpu().numpy())\n            outputs_all.extend(output.detach().cpu().numpy())\n            \n    return targets_all, outputs_all","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from sklearn.metrics.pairwise import cosine_similarity\n    \ndef get_distance_matrix(embeds, base_embeds):\n    distance_matrix = []\n    embeds_dataset = torch.utils.data.TensorDataset(torch.Tensor(embeds))\n    embeds_dataloader = DataLoader(embeds_dataset, num_workers=2, batch_size=1024, shuffle=False)\n    \n    t = tqdm(embeds_dataloader)\n    for i, sample in enumerate(t): \n        distances = cosine_similarity(sample[0].numpy(), base_embeds)\n        distance_matrix.extend(distances)\n        \n    return np.array(distance_matrix)","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def save_checkpoint(model, scheduler, optimizer, epoch, name, loss=None, score=None):\n    checkpoint = {\"epoch\": epoch,\n                  \"model\": model.state_dict(),\n                  \"scheduler\": scheduler.state_dict(),\n                  \"optimizer\": optimizer.state_dict(),\n                  \"loss\": loss,\n                  \"score\": score,\n                  }\n\n    torch.save(checkpoint, f\"{OUTPUT_FOLDER}checkpoint-{name}.pt\")\n\n\ndef load_checkpoint(model, scheduler, optimizer, name):\n    checkpoint = torch.load(f\"{OUTPUT_FOLDER}checkpoint-{name}.pt\")\n\n    model.load_state_dict(checkpoint[\"model\"])\n    scheduler.load_state_dict(checkpoint[\"scheduler\"])\n\n    return model, scheduler, optimizer, checkpoint[\"epoch\"]","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def iterate_loader(loader, epochs):\n    \"\"\"\n    Iterates through data loader with fake load (empty data) instead\n    of reading the real images to speed up. Dataloader has no state_dict\n    so manual iterating is need to get the loader with augmentations to \n    correct state.\n    \"\"\"\n    loader.dataset.fake_load = True\n    with torch.no_grad():\n        for i in range(epochs):\n            t = tqdm(loader, desc=f\"Iterating loader {i+1}/{epochs}\")\n            for j, sample in enumerate(t):\n                images = sample['image']\n                targets = sample['target']\n\n    loader.dataset.fake_load = False","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def train_epoch(args, model, loader, criterion, optimizer, scheduler, epoch):\n    losses = []\n    targets_all = []\n    outputs_all = []\n    \n    model.train()\n    t = tqdm(loader)\n    \n    for i, sample in enumerate(t):\n        optimizer.zero_grad()\n        \n        input = sample['image'].to(args.device)\n        target = sample['target'].to(args.device)\n        \n        output = model(input, target)\n        loss = criterion(output, target)\n        \n        loss.backward()\n        optimizer.step()\n\n        if scheduler is not None:\n            scheduler.step()\n        \n        losses.append(loss.item())\n        targets_all.extend(target.cpu().numpy())\n        outputs_all.extend(torch.sigmoid(output).detach().cpu().numpy())\n        \n        score = accuracy_score(targets_all, np.argmax(outputs_all, axis=1))\n        t.set_description(f\"Training epoch {epoch}/{args.epochs} - Train loss:{loss:0.4f}, score: {score:0.4f}\")\n        \n    return np.mean(losses), score","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def find_closest_match(base_df, distance_matrix, n_matches=5):\n    preds = []\n    N_dist = len(distance_matrix)\n    for i in tqdm(range(N_dist), total=N_dist, desc=\"Getting closest match\"):\n        tmp_df = base_df.copy()\n        tmp_df[\"distance\"] = distance_matrix[i]\n        tmp_df = tmp_df.sort_values(by=[\"distance\", \"hotel_id\"], ascending=False).reset_index(drop=True)\n        preds.extend([tmp_df[\"hotel_id_code\"].unique()[:n_matches]])\n    \n    preds = np.array(preds)\n    return preds","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def calc_metric(y_true, y_pred, n_matches=5):\n    y = np.repeat([y_true], repeats=n_matches, axis=0).T\n    acc_top_1 = (y_pred[:, 0] == y_true).mean()\n    acc_top_5 = (y_pred == y).any(axis=1).mean()\n    print(f\"Accuracy: {acc_top_1:0.4f}, top 5 accuracy: {acc_top_5:0.4f}\")\n    return acc_top_1, acc_top_5","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def test(base_loader, valid_loader, model):\n    base_targets, base_embeds = get_embeds(base_loader, model, \"Generating embeds for train\")\n    valid_targets, valid_embeds = get_embeds(valid_loader, model, \"Generating embeds for test\")\n    distance_matrix = get_distance_matrix(valid_embeds, base_embeds)\n    val_preds = find_closest_match(base_loader.dataset.data, distance_matrix)\n    calc_metric(valid_targets, val_preds)\n    return base_embeds, valid_embeds, valid_targets, val_preds, distance_matrix","metadata":{"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def test_closest_match_tta(args, base_loader, valid_df, tta_transforms, model):\n    base_targets, base_embeds = get_embeds(base_loader, model, \"Generating embeds for train\")\n    distance_matrix = None\n\n    for key in tta_transforms:\n        valid_dataset = HotelTrainDataset(valid_df, tta_transforms[key], data_path=DATA_FOLDER)\n        valid_loader = DataLoader(valid_dataset, num_workers=args.num_workers, batch_size=args.batch_size, shuffle=False)\n        valid_targets, valid_embeds = get_embeds(valid_loader, model, f\"Generating embeds for test {key}\")\n        \n        distances = get_distance_matrix(valid_embeds, base_embeds)\n\n        if distance_matrix is None:\n            distance_matrix = distances\n        else:\n            distance_matrix = np.min(np.dstack((distance_matrix, distances)), axis = 2)\n    \n    val_preds = find_closest_match(base_loader.dataset.data, distance_matrix)\n    calc_metric(valid_targets, val_preds)","metadata":{"execution":{"iopub.status.busy":"2022-05-17T14:02:22.340517Z","iopub.execute_input":"2022-05-17T14:02:22.340878Z","iopub.status.idle":"2022-05-17T14:02:22.351371Z","shell.execute_reply.started":"2022-05-17T14:02:22.340833Z","shell.execute_reply":"2022-05-17T14:02:22.350566Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def sample_data(n_hotels, min_images, max_images):\n    data_df = pd.read_csv(DATA_FOLDER + \"train.csv\")\n    sample_df = data_df.groupby(\"hotel_id\").filter(lambda x: (x[\"image_id\"].nunique() > min_images) & (x[\"image_id\"].nunique() < max_images))\n    sample_df[\"hotel_id_code\"] = sample_df[\"hotel_id\"].astype('category').cat.codes.values.astype(np.int64)\n    sample_df = sample_df[sample_df[\"hotel_id_code\"] < n_hotels]\n\n    print(f\"Subsample with {len(sample_df.hotel_id.unique())} hotels out of {len(data_df.hotel_id.unique())} \" + \n          f\"with total {len(sample_df)} images ({len(sample_df) / len(data_df) * 100:0.2f} %)\")\n    \n    return sample_df","metadata":{"execution":{"iopub.status.busy":"2022-05-17T14:02:23.229651Z","iopub.execute_input":"2022-05-17T14:02:23.229896Z","iopub.status.idle":"2022-05-17T14:02:23.237372Z","shell.execute_reply.started":"2022-05-17T14:02:23.229848Z","shell.execute_reply":"2022-05-17T14:02:23.235106Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# FOR TESTING DIFFERENT SETTING\n#data_df = sample_data(1000, 15, 50)\n\n# FOR FINAL TRAINING\ndata_df = pd.read_csv(DATA_FOLDER + \"train.csv\")\ndata_df[\"hotel_id_code\"] = data_df[\"hotel_id\"].astype('category').cat.codes.values.astype(np.int64)\n\nfig = go.Figure()\nfig.add_trace(go.Histogram(x=data_df[\"hotel_id_code\"]))\nfig.update_xaxes(type=\"category\")\nfig.show()","metadata":{"execution":{"iopub.status.busy":"2022-05-17T14:02:23.968193Z","iopub.execute_input":"2022-05-17T14:02:23.968771Z","iopub.status.idle":"2022-05-17T14:02:24.094233Z","shell.execute_reply.started":"2022-05-17T14:02:23.968737Z","shell.execute_reply":"2022-05-17T14:02:24.093521Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data_df.head()","metadata":{"execution":{"iopub.status.busy":"2022-05-17T14:02:24.688418Z","iopub.execute_input":"2022-05-17T14:02:24.688642Z","iopub.status.idle":"2022-05-17T14:02:24.703513Z","shell.execute_reply.started":"2022-05-17T14:02:24.688615Z","shell.execute_reply":"2022-05-17T14:02:24.702903Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def train_and_validate(args, data_df):\n    model_name = f\"arcmargin-model-{args.backbone_name}-{IMG_SIZE}x{IMG_SIZE}-{args.embed_size}embeds-{args.n_classes}hotels\"\n    print(model_name)\n    #SEED and split\n    seed_everything(seed=SEED)\n    valid_df = data_df.groupby(\"hotel_id\").sample(args.val_samples, random_state=SEED)\n    train_df = data_df[~data_df[\"image_id\"].isin(valid_df[\"image_id\"])]\n\n    #Create model\n    model = HotelIdModel(args.n_classes, args.embed_size, args.backbone_name)\n    model = model.to(args.device)\n\n    #Train data loader\n    train_dataset = HotelTrainDataset(train_df, valid_transform, data_path=DATA_FOLDER)\n    train_loader = DataLoader(train_dataset, num_workers=args.num_workers, batch_size=args.batch_size, shuffle=True, pin_memory=False)\n    #Train without augmentations to generate base embeddings\n    base_dataset = HotelTrainDataset(train_df, valid_transform, data_path=DATA_FOLDER)\n    base_loader = DataLoader(base_dataset, num_workers=args.num_workers, batch_size=args.batch_size, shuffle=False)\n    #Valid loader\n    valid_dataset = HotelTrainDataset(valid_df, valid_transform, data_path=DATA_FOLDER)\n    valid_loader = DataLoader(valid_dataset, num_workers=args.num_workers, batch_size=args.batch_size, shuffle=False)\n\n    print(f\"Base: {len(base_dataset)}\\nValidation: {len(valid_dataset)}\")\n\n    criterion = nn.CrossEntropyLoss()\n    optimizer = Lookahead(torch.optim.AdamW(model.parameters(), lr=args.lr), k=3)\n    scheduler = torch.optim.lr_scheduler.OneCycleLR(\n                    optimizer,\n                    max_lr=args.lr,\n                    epochs=args.epochs,\n                    steps_per_epoch=len(train_loader),\n                    div_factor=10,\n                    final_div_factor=1,\n                    pct_start=0.1,\n                    anneal_strategy=\"cos\",\n                )\n    \n    start_epoch = 1\n\n    if args.continue_from_checkpoint:\n        model, scheduler, optimizer, last_epoch = load_checkpoint(model, scheduler, optimizer, model_name)\n        iterate_loader(train_loader, last_epoch)\n        start_epoch = start_epoch + last_epoch\n\n    torch.cuda.empty_cache()\n\n    for epoch in range(start_epoch, args.epochs +1):\n        train_loss, train_score = train_epoch(args, model, train_loader, criterion, optimizer, scheduler, epoch)\n        save_checkpoint(model, scheduler, optimizer, epoch, model_name, train_loss, train_score)\n        if (epoch == 1):\n            base_embeds, valid_embeds, valid_targets, val_preds, distance_matrix = test(base_loader, valid_loader, model)\n\n    base_embeds, valid_embeds, valid_targets, val_preds, distance_matrix = test(base_loader, valid_loader, model)\n\n    output = {\"base_embeds\": base_embeds,\n              \"valid_embeds\": valid_embeds,\n              \"valid_targets\": valid_targets,\n              \"val_preds\": val_preds,\n              \"distance_matrix\": distance_matrix,\n              \"train_df\" : train_df,\n              \"valid_df\": valid_df,\n              }\n\n    torch.save(output, f\"{OUTPUT_FOLDER}output-{model_name}.pt\")\n    \n    base_dataset   = HotelTrainDataset(data_df, train_transform, data_path=IMAGE_FOLDER)\n    base_loader    = DataLoader(base_dataset, num_workers=args.num_workers, batch_size=args.batch_size*4, shuffle=False)\n    _, base_embeds = get_embeds(base_loader, model, \"Generate embeddings for all images\")\n    data_df[\"embeddings\"] = list(base_embeds)\n    data_df.to_pickle(f\"{OUTPUT_FOLDER}{model_name}_image-embeddings.pkl\")","metadata":{"execution":{"iopub.status.busy":"2022-05-17T14:02:26.812385Z","iopub.execute_input":"2022-05-17T14:02:26.812655Z","iopub.status.idle":"2022-05-17T14:02:26.830849Z","shell.execute_reply.started":"2022-05-17T14:02:26.812625Z","shell.execute_reply":"2022-05-17T14:02:26.82993Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"%%time \n\nclass args:\n    epochs = 10\n    lr = 1e-3\n    batch_size = 16\n    num_workers = 2  \n    embed_size = 2048\n    val_samples = 1\n    backbone_name=\"efficientnet_b1\"\n    n_classes = data_df[\"hotel_id_code\"].nunique()\n    device = ('cuda' if torch.cuda.is_available() else 'cpu')\n    continue_from_checkpoint = False\n\ntrain_and_validate(args, data_df)","metadata":{"execution":{"iopub.status.busy":"2022-05-17T14:02:32.105696Z","iopub.execute_input":"2022-05-17T14:02:32.106439Z"},"trusted":true},"execution_count":null,"outputs":[]}]}