{"cells":[{"metadata":{"trusted":true},"cell_type":"code","source":"from typing import Dict\n\nfrom tempfile import gettempdir\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport pandas as pd\nimport torch\nfrom torch import nn, optim\nfrom torch.utils.data import DataLoader\nfrom torchvision.models.resnet import resnet50, resnet18, resnet34, resnet101\nfrom tqdm import tqdm\n\nimport l5kit\nfrom l5kit.configs import load_config_data\nfrom l5kit.data import LocalDataManager, ChunkedDataset\nfrom l5kit.dataset import AgentDataset, EgoDataset\nfrom l5kit.rasterization import build_rasterizer\nfrom l5kit.evaluation import write_pred_csv, compute_metrics_csv, read_gt_csv, create_chopped_dataset\nfrom l5kit.evaluation.chop_dataset import MIN_FUTURE_STEPS\nfrom l5kit.evaluation.metrics import neg_multi_log_likelihood, time_displace\nfrom l5kit.geometry import transform_points\nfrom l5kit.visualization import PREDICTED_POINTS_COLOR, TARGET_POINTS_COLOR, draw_trajectory\nfrom prettytable import PrettyTable\nfrom pathlib import Path\n\nimport matplotlib.pyplot as plt\n\nimport os\nimport random\nimport time\n\nimport warnings\nwarnings.filterwarnings(\"ignore\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"l5kit.__version__","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def set_seed(seed):\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)\n    \nset_seed(42)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Configs"},{"metadata":{"trusted":true},"cell_type":"code","source":"# --- Lyft configs ---\ncfg = {\n    'format_version': 4,\n    'data_path': \"/kaggle/input/lyft-motion-prediction-autonomous-vehicles\",\n    'model_params': {\n        'model_architecture': 'resnet34',\n        'history_num_frames': 10,\n        'history_step_size': 1,\n        'history_delta_time': 0.1,\n        'future_num_frames': 50,\n        'future_step_size': 1,\n        'future_delta_time': 0.1,\n        'model_name': \"model_resnet34_output\",\n        'lr': 1e-3,\n        'weight_path': \"/kaggle/input/lyft-pretrained-model-hv/model_multi_update_lyft_public.pth\",\n        'train': False,\n        'predict': True\n    },\n\n    'raster_params': {\n        'raster_size': [224, 224],\n        'pixel_size': [0.5, 0.5],\n        'ego_center': [0.25, 0.5],\n        'map_type': 'py_semantic',\n        'satellite_map_key': 'aerial_map/aerial_map.png',\n        'semantic_map_key': 'semantic_map/semantic_map.pb',\n        'dataset_meta_key': 'meta.json',\n        'filter_agents_threshold': 0.5\n    },\n\n    'train_data_loader': {\n        'key': 'scenes/train.zarr',\n        'batch_size': 16,\n        'shuffle': True,\n        'num_workers': 4\n    },\n    \n    'test_data_loader': {\n        'key': 'scenes/test.zarr',\n        'batch_size': 32,\n        'shuffle': False,\n        'num_workers': 4\n    },\n\n    'train_params': {\n        'max_num_steps': 101,\n        'checkpoint_every_n_steps': 20,\n    }\n}","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"Couple of things to note:\n\n - **model_architecture:** you can put 'resnet18', 'resnet34' or 'resnet50'. For the pretrained model we use resnet18 so we need to use 'resnet18' in the config.\n - **weight_path:** path to the pretrained model. If you don't have a pretrained model and want to train from scratch, put **weight_path** = False. \n - **model_name:** the name of the model that will be saved as output, this is only when **train**= True.\n - **train:** True if you want to continue to train the model. Unfortunately due to Kaggle memory constraint if **train**=True then you should put **predict** = False.\n - **predict:** True if you want to predict and submit to Kaggle. Unfortunately due to Kaggle memory constraint if you want to predict then you need  to put **train** = False.\n - **lr:** learning rate of the model, feel free to change as you see fit. In the future I also plan to implement learning rate decay. \n - **raster_size:** specify the size of the image, the default is [224,224]. Increase **raster_size** can improve the score. However the training time will be significantly longer. \n - **batch_size:** number of inputs for one forward pass, again one of the parameters to tune. \n - **max_num_steps:** the number of iterations to train, i.e. number of epochs.\n - **checkpoint_every_n_steps:** the model will be saved at every n steps, again change this number as to how you want to keep track of the model."},{"metadata":{},"cell_type":"markdown","source":"## Load the train and test data"},{"metadata":{"trusted":true},"cell_type":"code","source":"# set env variable for data\nDIR_INPUT = cfg[\"data_path\"]\nos.environ[\"L5KIT_DATA_FOLDER\"] = DIR_INPUT\ndm = LocalDataManager(None)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# ===== INIT TRAIN DATASET============================================================\ntrain_cfg = cfg[\"train_data_loader\"]\nrasterizer = build_rasterizer(cfg, dm)\ntrain_zarr = ChunkedDataset(dm.require(train_cfg[\"key\"])).open()\ntrain_dataset = AgentDataset(cfg, train_zarr, rasterizer)\ntrain_dataloader = DataLoader(train_dataset, shuffle=train_cfg[\"shuffle\"], batch_size=train_cfg[\"batch_size\"], \n                             num_workers=train_cfg[\"num_workers\"])\nprint(\"==================================TRAIN DATA==================================\")\nprint(train_dataset)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"#====== INIT TEST DATASET=============================================================\ntest_cfg = cfg[\"test_data_loader\"]\nrasterizer = build_rasterizer(cfg, dm)\ntest_zarr = ChunkedDataset(dm.require(test_cfg[\"key\"])).open()\ntest_mask = np.load(f\"{DIR_INPUT}/scenes/mask.npz\")[\"arr_0\"]\ntest_dataset = AgentDataset(cfg, test_zarr, rasterizer, agents_mask=test_mask)\ntest_dataloader = DataLoader(test_dataset,shuffle=test_cfg[\"shuffle\"],batch_size=test_cfg[\"batch_size\"],\n                             num_workers=test_cfg[\"num_workers\"])\nprint(\"==================================TEST DATA==================================\")\nprint(test_dataset)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Simple visualization"},{"metadata":{},"cell_type":"markdown","source":"Let us visualize how an input to the model looks like."},{"metadata":{"trusted":true},"cell_type":"code","source":"def visualize_trajectory(dataset, index, title=\"target_positions movement with draw_trajectory\"):\n    data = dataset[index]\n    im = data[\"image\"].transpose(1, 2, 0)\n    im = dataset.rasterizer.to_rgb(im)\n    target_positions_pixels = transform_points(data[\"target_positions\"] + data[\"centroid\"][:2], data[\"world_to_image\"])\n    draw_trajectory(im, target_positions_pixels, TARGET_POINTS_COLOR, radius=1, yaws=data[\"target_yaws\"])\n\n    plt.title(title)\n    plt.imshow(im[::-1])\n    plt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"plt.figure(figsize = (8,6))\nvisualize_trajectory(train_dataset, index=90)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Loss function"},{"metadata":{"trusted":true},"cell_type":"code","source":"# --- Function utils ---\n# Original code from https://github.com/lyft/l5kit/blob/20ab033c01610d711c3d36e1963ecec86e8b85b6/l5kit/l5kit/evaluation/metrics.py\nimport numpy as np\n\nimport torch\nfrom torch import Tensor\n\n\ndef pytorch_neg_multi_log_likelihood_batch(\n    gt: Tensor, pred: Tensor, confidences: Tensor, avails: Tensor\n) -> Tensor:\n    \"\"\"\n    Compute a negative log-likelihood for the multi-modal scenario.\n    log-sum-exp trick is used here to avoid underflow and overflow, For more information about it see:\n    https://en.wikipedia.org/wiki/LogSumExp#log-sum-exp_trick_for_log-domain_calculations\n    https://timvieira.github.io/blog/post/2014/02/11/exp-normalize-trick/\n    https://leimao.github.io/blog/LogSumExp/\n    Args:\n        gt (Tensor): array of shape (bs)x(time)x(2D coords)\n        pred (Tensor): array of shape (bs)x(modes)x(time)x(2D coords)\n        confidences (Tensor): array of shape (bs)x(modes) with a confidence for each mode in each sample\n        avails (Tensor): array of shape (bs)x(time) with the availability for each gt timestep\n    Returns:\n        Tensor: negative log-likelihood for this example, a single float number\n    \"\"\"\n    assert len(pred.shape) == 4, f\"expected 3D (MxTxC) array for pred, got {pred.shape}\"\n    batch_size, num_modes, future_len, num_coords = pred.shape\n\n    assert gt.shape == (batch_size, future_len, num_coords), f\"expected 2D (Time x Coords) array for gt, got {gt.shape}\"\n    assert confidences.shape == (batch_size, num_modes), f\"expected 1D (Modes) array for gt, got {confidences.shape}\"\n    assert torch.allclose(torch.sum(confidences, dim=1), confidences.new_ones((batch_size,))), \"confidences should sum to 1\"\n    assert avails.shape == (batch_size, future_len), f\"expected 1D (Time) array for gt, got {avails.shape}\"\n    # assert all data are valid\n    assert torch.isfinite(pred).all(), \"invalid value found in pred\"\n    assert torch.isfinite(gt).all(), \"invalid value found in gt\"\n    assert torch.isfinite(confidences).all(), \"invalid value found in confidences\"\n    assert torch.isfinite(avails).all(), \"invalid value found in avails\"\n\n    # convert to (batch_size, num_modes, future_len, num_coords)\n    gt = torch.unsqueeze(gt, 1)  # add modes\n    avails = avails[:, None, :, None]  # add modes and cords\n\n    # error (batch_size, num_modes, future_len)\n    error = torch.sum(((gt - pred) * avails) ** 2, dim=-1)  # reduce coords and use availability\n\n    with np.errstate(divide=\"ignore\"):  # when confidence is 0 log goes to -inf, but we're fine with it\n        # error (batch_size, num_modes)\n        error = torch.log(confidences) - 0.5 * torch.sum(error, dim=-1)  # reduce time\n\n    # use max aggregator on modes for numerical stability\n    # error (batch_size, num_modes)\n    max_value, _ = error.max(dim=1, keepdim=True)  # error are negative at this point, so max() gives the minimum one\n    error = -torch.log(torch.sum(torch.exp(error - max_value), dim=-1, keepdim=True)) - max_value  # reduce modes\n    # print(\"error\", error)\n    return torch.mean(error)\n\n\ndef pytorch_neg_multi_log_likelihood_single(\n    gt: Tensor, pred: Tensor, avails: Tensor\n) -> Tensor:\n    \"\"\"\n\n    Args:\n        gt (Tensor): array of shape (bs)x(time)x(2D coords)\n        pred (Tensor): array of shape (bs)x(time)x(2D coords)\n        avails (Tensor): array of shape (bs)x(time) with the availability for each gt timestep\n    Returns:\n        Tensor: negative log-likelihood for this example, a single float number\n    \"\"\"\n    # pred (bs)x(time)x(2D coords) --> (bs)x(mode=1)x(time)x(2D coords)\n    # create confidence (bs)x(mode=1)\n    batch_size, future_len, num_coords = pred.shape\n    confidences = pred.new_ones((batch_size, 1))\n    return pytorch_neg_multi_log_likelihood_batch(gt, pred.unsqueeze(1), confidences, avails)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Model"},{"metadata":{"trusted":true},"cell_type":"code","source":"class LyftMultiModel(nn.Module):\n\n    def __init__(self, cfg: Dict, num_modes=3):\n        super().__init__()\n\n        architecture = cfg[\"model_params\"][\"model_architecture\"]\n        backbone = eval(architecture)(pretrained=True, progress=True)\n        self.backbone = backbone\n\n        num_history_channels = (cfg[\"model_params\"][\"history_num_frames\"] + 1) * 2\n        num_in_channels = 3 + num_history_channels\n\n        self.backbone.conv1 = nn.Conv2d(\n            num_in_channels,\n            self.backbone.conv1.out_channels,\n            kernel_size=self.backbone.conv1.kernel_size,\n            stride=self.backbone.conv1.stride,\n            padding=self.backbone.conv1.padding,\n            bias=False,\n        )\n\n        # This is 512 for resnet18 and resnet34;\n        # And it is 2048 for the other resnets\n        \n        if architecture == \"resnet50\":\n            backbone_out_features = 2048\n        else:\n            backbone_out_features = 512\n\n        # X, Y coords for the future positions (output shape: batch_sizex50x2)\n        self.future_len = cfg[\"model_params\"][\"future_num_frames\"]\n        num_targets = 2 * self.future_len\n\n        # You can add more layers here.\n        self.head = nn.Sequential(\n            # nn.Dropout(0.2),\n            nn.Linear(in_features=backbone_out_features, out_features=4096),\n        )\n\n        self.num_preds = num_targets * num_modes\n        self.num_modes = num_modes\n\n        self.logit = nn.Linear(4096, out_features=self.num_preds + num_modes)\n\n    def forward(self, x):\n        x = self.backbone.conv1(x)\n        x = self.backbone.bn1(x)\n        x = self.backbone.relu(x)\n        x = self.backbone.maxpool(x)\n\n        x = self.backbone.layer1(x)\n        x = self.backbone.layer2(x)\n        x = self.backbone.layer3(x)\n        x = self.backbone.layer4(x)\n\n        x = self.backbone.avgpool(x)\n        x = torch.flatten(x, 1)\n\n        x = self.head(x)\n        x = self.logit(x)\n\n        # pred (batch_size)x(modes)x(time)x(2D coords)\n        # confidences (batch_size)x(modes)\n        bs, _ = x.shape\n        pred, confidences = torch.split(x, self.num_preds, dim=1)\n        pred = pred.view(bs, self.num_modes, self.future_len, 2)\n        assert confidences.shape == (bs, self.num_modes)\n        confidences = torch.softmax(confidences, dim=1)\n        return pred, confidences","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def forward(data, model, device, criterion = pytorch_neg_multi_log_likelihood_batch):\n    inputs = data[\"image\"].to(device)\n    target_availabilities = data[\"target_availabilities\"].to(device)\n    targets = data[\"target_positions\"].to(device)\n    # Forward pass\n    preds, confidences = model(inputs)\n    loss = criterion(targets, preds, confidences, target_availabilities)\n    return loss, preds, confidences","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# ==== INIT MODEL=================\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\nmodel = LyftMultiModel(cfg)\n\n#load weight if there is a pretrained model\nweight_path = cfg[\"model_params\"][\"weight_path\"]\nif weight_path:\n    model.load_state_dict(torch.load(weight_path))\n\nmodel.to(device)\noptimizer = optim.Adam(model.parameters(), lr=cfg[\"model_params\"][\"lr\"])\nprint(f'device {device}')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"print(model)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Training loop"},{"metadata":{},"cell_type":"markdown","source":"Next let us implement the training loop, when the **train** parameter is set to True. "},{"metadata":{"trusted":true},"cell_type":"code","source":"# ==== TRAINING LOOP =========================================================\nif cfg[\"model_params\"][\"train\"]:\n    \n    tr_it = iter(train_dataloader)\n    progress_bar = tqdm(range(cfg[\"train_params\"][\"max_num_steps\"]))\n    num_iter = cfg[\"train_params\"][\"max_num_steps\"]\n    losses_train = []\n    iterations = []\n    metrics = []\n    times = []\n    model_name = cfg[\"model_params\"][\"model_name\"]\n    start = time.time()\n    for i in progress_bar:\n        try:\n            data = next(tr_it)\n        except StopIteration:\n            tr_it = iter(train_dataloader)\n            data = next(tr_it)\n        model.train()\n        torch.set_grad_enabled(True)\n        \n        loss, _, _ = forward(data, model, device)\n\n        # Backward pass\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n\n        losses_train.append(loss.item())\n\n        progress_bar.set_description(f\"loss: {loss.item()} loss(avg): {np.mean(losses_train)}\")\n        if i % cfg['train_params']['checkpoint_every_n_steps'] == 0:\n            torch.save(model.state_dict(), f'{model_name}_{i}.pth')\n            iterations.append(i)\n            metrics.append(np.mean(losses_train))\n            times.append((time.time()-start)/60)\n\n    results = pd.DataFrame({'iterations': iterations, 'metrics (avg)': metrics, 'elapsed_time (mins)': times})\n    results.to_csv(f\"train_metrics_{model_name}_{num_iter}.csv\", index = False)\n    print(f\"Total training time is {(time.time()-start)/60} mins\")\n    print(results.head())","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"## Prediction"},{"metadata":{},"cell_type":"markdown","source":"Finally we implement the inference to submit to Kaggle when **predict** param is set to True."},{"metadata":{"trusted":true},"cell_type":"code","source":"# ==== EVAL LOOP ================================================================\nif cfg[\"model_params\"][\"predict\"]:\n    \n    model.eval()\n    torch.set_grad_enabled(False)\n\n    # store information for evaluation\n    future_coords_offsets_pd = []\n    timestamps = []\n    confidences_list = []\n    agent_ids = []\n\n    progress_bar = tqdm(test_dataloader)\n    \n    for data in progress_bar:\n        \n        _, preds, confidences = forward(data, model, device)\n    \n        #fix for the new environment\n        preds = preds.cpu().numpy()\n        world_from_agents = data[\"world_from_agent\"].numpy()\n        centroids = data[\"centroid\"].numpy()\n        coords_offset = []\n        \n        # convert into world coordinates and compute offsets\n        for idx in range(len(preds)):\n            for mode in range(3):\n                preds[idx, mode, :, :] = transform_points(preds[idx, mode, :, :], world_from_agents[idx]) - centroids[idx][:2]\n    \n        future_coords_offsets_pd.append(preds.copy())\n        confidences_list.append(confidences.cpu().numpy().copy())\n        timestamps.append(data[\"timestamp\"].numpy().copy())\n        agent_ids.append(data[\"track_id\"].numpy().copy()) ","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"#create submission to submit to Kaggle\npred_path = 'submission.csv'\nwrite_pred_csv(pred_path,\n           timestamps=np.concatenate(timestamps),\n           track_ids=np.concatenate(agent_ids),\n           coords=np.concatenate(future_coords_offsets_pd),\n           confs = np.concatenate(confidences_list)\n          )","execution_count":null,"outputs":[]}],"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":4,"nbformat_minor":4}