{"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"name":"python","version":"3.10.13","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"},"kaggle":{"accelerator":"nvidiaTeslaT4","dataSources":[{"sourceId":71549,"databundleVersionId":8561470,"sourceType":"competition"},{"sourceId":9114148,"sourceType":"datasetVersion","datasetId":5501147},{"sourceId":9423242,"sourceType":"datasetVersion","datasetId":5705584},{"sourceId":9482835,"sourceType":"datasetVersion","datasetId":5762105},{"sourceId":9484766,"sourceType":"datasetVersion","datasetId":5723776},{"sourceId":9516581,"sourceType":"datasetVersion","datasetId":5786322},{"sourceId":9578924,"sourceType":"datasetVersion","datasetId":5801157}],"dockerImageVersionId":30733,"isInternetEnabled":false,"language":"python","sourceType":"notebook","isGpuEnabled":true}},"nbformat_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"# Import Libralies","metadata":{}},{"cell_type":"code","source":"import os\nimport sys\nsys.path.append('/kaggle/input/timm-3d/')\nsys.path.append('/kaggle/input/lhwcv-rsna2024-code/')\nos.environ['OMP_NUM_THREADS'] = '1'\n\nimport pandas as pd\nimport numpy as np\nimport torch\nimport tqdm\nimport random\nimport gc\nimport pickle\n\nfrom concurrent.futures.thread import ThreadPoolExecutor\n\nfrom infer_scripts.v2.dataset import RSNA24DatasetTest_LHW_keypoint_3D_Axial, RSNA24DatasetTest_LHW_V2, \\\n RSNA24DatasetTest_LHW_keypoint_3D_Saggital\nfrom infer_scripts.v2.models import RSNA24Model_Keypoint_3D_Sag_V2\nfrom infer_scripts.v2.dataset import RSNA24DatasetTest_LHW_V2, v2_collate_fn\nfrom infer_scripts.v2.models import HybridModel_V2\nfrom infer_scripts.v20.models import build_v20_sag_model\nfrom infer_scripts.v24.infer_utils import *\n\nimport warnings\nwarnings.filterwarnings('ignore', category=pd.errors.SettingWithCopyWarning)\n","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:41.753428Z","iopub.execute_input":"2024-10-08T16:36:41.754065Z","iopub.status.idle":"2024-10-08T16:36:51.055596Z","shell.execute_reply.started":"2024-10-08T16:36:41.754029Z","shell.execute_reply":"2024-10-08T16:36:51.054796Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def setup_seed(seed, deterministic=True):\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.cuda.manual_seed_all(seed)\n torch.backends.cudnn.deterministic = deterministic\n torch.backends.cudnn.benchmark = False","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.057071Z","iopub.execute_input":"2024-10-08T16:36:51.057499Z","iopub.status.idle":"2024-10-08T16:36:51.062982Z","shell.execute_reply.started":"2024-10-08T16:36:51.057476Z","shell.execute_reply":"2024-10-08T16:36:51.062099Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"SEED = 8620\nsetup_seed(SEED, deterministic=True)","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.064026Z","iopub.execute_input":"2024-10-08T16:36:51.064311Z","iopub.status.idle":"2024-10-08T16:36:51.086981Z","shell.execute_reply.started":"2024-10-08T16:36:51.064288Z","shell.execute_reply":"2024-10-08T16:36:51.086132Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"data_root = '/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification/'\ntest_series_descriptions_fn = 'test_series_descriptions.csv'\nimage_dir = 'test_images'\n\n# test_series_descriptions_fn = 'train_series_descriptions.csv'\n# image_dir = 'train_images'\n\nN_WORKERS = os.cpu_count()\nbatch_size = 4\n\nkeypoint_dir = './keypoints_pred/'\nCACHE_DIR = './cache/'\n\nmodel_dir = \"/kaggle/input/lhwcv-rsna2024-final-models/\"\nmodel_dir2 = \"/kaggle/input/lhwcv-rsna2024-final-models3/\"\nmodel_dir2_old = \"/kaggle/input/lhwcv-rsna2024-final-models2/\"","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.089251Z","iopub.execute_input":"2024-10-08T16:36:51.089575Z","iopub.status.idle":"2024-10-08T16:36:51.095051Z","shell.execute_reply.started":"2024-10-08T16:36:51.089546Z","shell.execute_reply":"2024-10-08T16:36:51.094207Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"os.makedirs(keypoint_dir, exist_ok=True)\nos.makedirs(CACHE_DIR, exist_ok=True)\n","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.095958Z","iopub.execute_input":"2024-10-08T16:36:51.096223Z","iopub.status.idle":"2024-10-08T16:36:51.102644Z","shell.execute_reply.started":"2024-10-08T16:36:51.096202Z","shell.execute_reply":"2024-10-08T16:36:51.101744Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"CONDITIONS = [\n 'spinal_canal_stenosis', \n 'left_neural_foraminal_narrowing', \n 'right_neural_foraminal_narrowing',\n 'left_subarticular_stenosis',\n 'right_subarticular_stenosis'\n]\n\nLEVELS = [\n 'l1_l2',\n 'l2_l3',\n 'l3_l4',\n 'l4_l5',\n 'l5_s1',\n]\ndf = pd.read_csv(f'{data_root}/{test_series_descriptions_fn}')\nstudy_ids = list(df['study_id'].unique())\n\nrow_names = []\nfor si in study_ids:\n for cond in CONDITIONS:\n for level in LEVELS:\n row_names.append(str(si) + '_' + cond + '_' + level)","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.103710Z","iopub.execute_input":"2024-10-08T16:36:51.103992Z","iopub.status.idle":"2024-10-08T16:36:51.122571Z","shell.execute_reply.started":"2024-10-08T16:36:51.103971Z","shell.execute_reply":"2024-10-08T16:36:51.121914Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sample_sub = pd.read_csv(f'{data_root}/sample_submission.csv')\nLABELS = list(sample_sub.columns[1:])","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.123667Z","iopub.execute_input":"2024-10-08T16:36:51.123945Z","iopub.status.idle":"2024-10-08T16:36:51.142371Z","shell.execute_reply.started":"2024-10-08T16:36:51.123924Z","shell.execute_reply":"2024-10-08T16:36:51.132283Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 1. keypoint detection","metadata":{}},{"cell_type":"code","source":"\ndef infer_axial_3d_keypoints_v2(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers=8,\n is_parallel=False):\n \"\"\"\n this has moved to infer_axial_3d_keypoints_v24 for dataloader speed\n \"\"\"\n model_cfgs = [\n {\n 'sub_dir': 'keypoint_3d_v2_axial/densenet161_lr_0.0006',\n 'backbone': 'densenet161',\n }\n ]\n v2_folds = 3\n models = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n model = RSNA24Model_Keypoint_3D(backbone,\n num_classes=30,\n pretrained=False)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models.append(model)\n\n dset = RSNA24DatasetTest_LHW_keypoint_3D_Axial(data_root,\n test_series_descriptions_fn,\n study_ids,\n image_dir=data_root + f'/{image_dir}/')\n\n dloader = DataLoader(dset, batch_size=8, num_workers=num_workers)\n\n study_id_to_pred_keypoints = {}\n for volumns, study_ids_, sids, depths in tqdm.tqdm(dloader,desc=f'{device}'):\n bs, _, _, _ = volumns.shape\n volumns = volumns.unsqueeze(1).to(device)\n keypoints = None\n with torch.no_grad():\n with autocast:\n for i in range(len(models)):\n p = models[i](volumns).cpu().numpy().reshape(bs, 10, 3)\n if keypoints is None:\n keypoints = p\n else:\n keypoints += p\n keypoints = keypoints / len(models)\n for idx, study_id in enumerate(study_ids_):\n study_id = int(study_id)\n sid = int(sids[idx])\n d = depths[idx]\n # print(study_id, sid)\n if study_id not in study_id_to_pred_keypoints.keys():\n study_id_to_pred_keypoints[study_id] = {}\n study_id_to_pred_keypoints[study_id][sid] = {\n 'points': keypoints[idx],\n 'd': int(d)\n }\n return study_id_to_pred_keypoints\n\n\ndef infer_axial_3d_keypoints_v24(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers=8,\n is_parallel=False,\n ):\n z_model_cfgs = [\n {\n 'sub_dir': 'keypoint_3d_v24_axial/level_cls/convnext_small.in12k_ft_in1k_384',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n },\n ]\n z_folds = 1\n z_models = []\n for cfg in z_model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(z_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n print('load: ', fn)\n model = Axial_Level_Cls_Model_for_Test(backbone,\n pretrained=False).to(device)\n model.load_state_dict(torch.load(fn))\n model.eval()\n z_models.append(model)\n\n \n #\n xy_model_cfgs = [\n# {\n# 'sub_dir': 'keypoint_3d_v24_axial/axial_2d_keypoints/densenet161_lr_0.0006',\n# 'backbone': 'densenet161',\n# },\n {\n 'sub_dir': 'keypoint_2d_v24_axial/xception65.tf_in1k_lr_0.0006/',\n 'backbone': 'xception65.tf_in1k',\n },\n ]\n xy_folds = 5\n xy_models = []\n for cfg in xy_model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(xy_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n if not os.path.exists(fn):\n fn = os.path.join(model_dir2_old, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n \n print('load: ', fn)\n model = RSNA24Model_Keypoint_2D(backbone,\n pretrained=False,\n num_classes=4,\n ).to(device)\n model.load_state_dict(torch.load(fn))\n model.eval()\n xy_models.append(model)\n\n v24_axial_pred_keypoints_info = {}\n v24_axial_pred_keypoints_info_2 = {}\n v2_axial_pred_keypoints_info = {}\n\n # v2 moved to here for dataloader speed\n boost_v2_dataloader = False\n if boost_v2_dataloader:\n model_cfgs = [\n {\n 'sub_dir': 'keypoint_3d_v2_axial/densenet161_lr_0.0006',\n 'backbone': 'densenet161',\n }\n ]\n v2_folds = 3\n v2_models = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n model = RSNA24Model_Keypoint_3D(backbone,\n num_classes=30,\n pretrained=False).to(device)\n model.load_state_dict(torch.load(fn))\n model.eval()\n v2_models.append(model)\n\n dset = Axial_Level_Dataset_Multi_V24(data_root, study_ids,\n test_series_descriptions_fn=test_series_descriptions_fn,\n image_dir=image_dir)\n dloader = DataLoader(dset, batch_size=1,\n num_workers=num_workers,\n collate_fn=axial_v24_collate_fn)\n\n\n for data_dict in tqdm.tqdm(dloader,desc=f'{device}'):\n study_id, series_id_list, pred_keypoints = axial_v24_infer_z(z_models, data_dict, device)\n study_id = int(study_id)\n pred_keypoints = axial_v24_infer_xy(xy_models, pred_keypoints, data_dict)\n\n if boost_v2_dataloader:\n v2_pred = axial_v2_infer_xyz(v2_models, data_dict)\n v2_axial_pred_keypoints_info[study_id] = v2_pred\n\n v24_axial_pred_keypoints_info[study_id] = {}\n for i, sid in enumerate(series_id_list):\n sid = int(sid)\n v24_axial_pred_keypoints_info[study_id][sid] = {\n 'points': pred_keypoints[i] # 2, 5, 4\n }\n \n if not boost_v2_dataloader:\n v2_axial_pred_keypoints_info = infer_axial_3d_keypoints_v2(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers,\n is_parallel)\n \n return (v24_axial_pred_keypoints_info, v2_axial_pred_keypoints_info)\n\n\ndef _infer_sag_xy(models, img):\n # img: b, 1, h, w\n pts = None\n for i in range(len(models)):\n with torch.no_grad():\n with autocast:\n p = models[i](img)\n if pts is None:\n pts = p\n else:\n pts += p\n pts = pts / len(models)\n return pts\n\n\ndef infer_sag_2d_keypoints_v2(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers=8,\n is_parallel=False):\n model_cfgs_t1 = [\n# {\n# 'sub_dir': 'keypoint_2d_v20_sag_t1/densenet161_lr_0.0006_sag_T1',\n# 'backbone': 'densenet161',\n# },\n {\n 'sub_dir': 'keypoint_2d_v20_sag_t1/fastvit_ma36.apple_dist_in1k_lr_0.0008_sag_T1_not_exclude_hard',\n 'backbone': 'fastvit_ma36.apple_dist_in1k',\n },\n\n ]\n v2_folds = 5\n models_t1 = []\n for cfg in model_cfgs_t1:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n if not os.path.exists(fn):\n fn = os.path.join(model_dir2, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n\n model = RSNA24Model_Keypoint_2D(backbone,\n num_classes=10,\n pretrained=False)\n print(f'{device} -->load: ',fn)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models_t1.append(model)\n #\n model_cfgs_t2 = [\n# {\n# 'sub_dir': 'keypoint_2d_v20_sag_t2/densenet161_lr_0.0006_sag_T2',\n# 'backbone': 'densenet161',\n# },\n {\n 'sub_dir': 'keypoint_2d_v20_sag_t2/convformer_s36.sail_in22k_ft_in1k_384_lr_0.0008_sag_T2_not_exclude_hard',\n 'backbone': 'convformer_s36.sail_in22k_ft_in1k_384',\n },\n ]\n v2_folds = 5\n models_t2 = []\n for cfg in model_cfgs_t2:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n if not os.path.exists(fn):\n fn = os.path.join(model_dir2, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n\n model = RSNA24Model_Keypoint_2D(backbone,\n num_classes=10,\n pretrained=False)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models_t2.append(model)\n\n #\n dset = RSNA24DatasetTest_LHW_V2(data_root, study_ids,\n image_dir=data_root + f'/{image_dir}/',\n test_series_descriptions_fn=test_series_descriptions_fn,\n with_axial=False,\n cache_dir=CACHE_DIR)\n print(f'{device} build data done!')\n dloader = DataLoader(dset, batch_size=8, num_workers=num_workers)\n\n study_id_to_pred_keypoints = {}\n\n for tensor_dict in tqdm.tqdm(dloader, desc=f'{device}'):\n imgs = tensor_dict['img']\n study_ids = tensor_dict['study_id']\n bs, _, _, _ = imgs.shape\n imgs = imgs.to(device)\n s_t1 = imgs[:, :10]\n s_t1 = s_t1[:, 5:6] # take the center\n s_t2 = imgs[:, 10: 20]\n s_t2 = s_t2[:, 5: 6]\n\n keypoints_t1 = _infer_sag_xy(models_t1, s_t1).reshape(bs, 1, 5, 2)\n keypoints_t2 = _infer_sag_xy(models_t2, s_t2).reshape(bs, 1, 5, 2)\n keypoints = torch.cat((keypoints_t1, keypoints_t2), dim=1).cpu().numpy()\n\n for idx, study_id in enumerate(study_ids):\n study_id = int(study_id)\n study_id_to_pred_keypoints[study_id] = keypoints[idx] * 512\n return study_id_to_pred_keypoints\n\n\ndef infer_sag_3d_keypoints_v2(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers=8,\n is_parallel=False,\n ):\n model_cfgs = [\n {\n 'sub_dir': 'keypoint_3d_v2_sag/densenet161_lr_0.0006/',\n 'backbone': 'densenet161',\n }\n ]\n v2_folds = 3\n models = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n print(f'{device} -->load: ', fn)\n model = RSNA24Model_Keypoint_3D_Sag_V2(model_name=backbone,\n pretrained=False)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models.append(model)\n #\n dset = RSNA24DatasetTest_LHW_keypoint_3D_Saggital(data_root,\n test_series_descriptions_fn,\n study_ids,\n image_dir=data_root + f'{image_dir}',\n cache_dir=CACHE_DIR)\n\n dloader = DataLoader(dset, batch_size=8, num_workers=num_workers)\n\n study_id_to_pred_keypoints_sag = {}\n for volumns, study_ids_ in tqdm.tqdm(dloader, desc=f'{device}'):\n with torch.no_grad():\n volumns = volumns.to(device)\n with autocast:\n keypoints = None\n for i in range(len(models)):\n y = models[i](volumns)\n bs = y.shape[0]\n if keypoints is None:\n keypoints = y.cpu().numpy().reshape(bs, 3, 5, 3)\n else:\n keypoints += y.cpu().numpy().reshape(bs, 3, 5, 3)\n keypoints = keypoints / len(models)\n for idx, study_id in enumerate(study_ids_):\n study_id = int(study_id)\n study_id_to_pred_keypoints_sag[study_id] = {\n 'points': keypoints[idx],\n }\n return study_id_to_pred_keypoints_sag\n\n\ndef infer_sag_3d_t1_keypoints_v20(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers=8,\n is_parallel=False):\n with_cascade = False\n if with_cascade:\n model_cfgs_t1_2d = [\n {\n 'sub_dir': 'keypoint_2d_v20_sag_t1/densenet161_lr_0.0006_sag_T1',\n 'backbone': 'densenet161',\n }\n ]\n v2_folds = 5\n models_t1_2d = []\n for cfg in model_cfgs_t1_2d:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n model = RSNA24Model_Keypoint_2D(backbone,\n num_classes=10,\n pretrained=False)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models_t1_2d.append(model)\n\n #\n pred_sag_keypoints_infos_3d_t1 = {}\n\n model_cfgs = [\n# {\n# 'sub_dir': 'keypoint_3d_v20_sag_t1/densenet161_lr_0.0006_sag_T1',\n# 'backbone': 'densenet161',\n# },\n {\n 'sub_dir': 'keypoint_3d_v20_sag_t1/gluon_resnet152_v1s_lr_0.0006_sag_T1_not_exclude_hard',\n 'backbone': 'gluon_resnet152_v1s',\n }\n ]\n folds = 5\n models = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n if not os.path.exists(fn):\n fn = os.path.join(model_dir2, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n model = RSNA24Model_Keypoint_3D(backbone,\n in_chans=1,\n pretrained=False,\n num_classes=30).to(device)\n model.load_state_dict(torch.load(fn))\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models.append(model)\n\n dset = Sag_3D_Point_Dataset_V24(data_root,\n study_ids,\n test_series_descriptions_fn=test_series_descriptions_fn,\n image_dir=image_dir,\n series_description='Sagittal T1',\n with_origin_arr=True if with_cascade else False,\n cache_dir=CACHE_DIR\n )\n\n dloader = DataLoader(dset, batch_size=1 if with_cascade else 8,\n num_workers=num_workers)\n\n for tensor_dict in tqdm.tqdm(dloader,desc=f'{device}'):\n with torch.no_grad():\n x = tensor_dict['imgs'].to(device)\n study_id_list = tensor_dict['study_id']\n series_id_list = tensor_dict['series_id']\n with autocast:\n y = None\n for i in range(len(models)):\n p = models[i](x)\n if y is None:\n y = p\n else:\n y += p\n y = y / len(models)\n bs = y.shape[0]\n for b in range(bs):\n study_id = int(study_id_list[b])\n series_id = int(series_id_list[b])\n pts = y[b].reshape(-1, 5, 3).cpu().numpy()\n\n if with_cascade:\n origin_depth = int(tensor_dict['origin_depth'][b])\n origin_imgs = tensor_dict['origin_imgs'][b].to(device)\n sub_imgs = []\n for n in range(pts.shape[0]):\n for level in range(5):\n z = pts[n, level, 2] * origin_depth\n z = int(np.round(z))\n if z < 0:\n z = 0\n if z > origin_depth - 1:\n z = origin_depth - 1\n\n img = origin_imgs[z].unsqueeze(0).unsqueeze(0)\n sub_imgs.append(img)\n sub_imgs = torch.cat(sub_imgs, dim=0) # 2*5, 1, 512, 512\n xy_pred = _infer_sag_xy(models_t1_2d, sub_imgs).reshape(-1, 5, 5, 2).cpu().numpy()\n for n in range(pts.shape[0]):\n for level in range(5):\n pts[n, level, :2] = xy_pred[n, level, level]\n\n if study_id not in pred_sag_keypoints_infos_3d_t1:\n pred_sag_keypoints_infos_3d_t1[study_id] = {}\n pred_sag_keypoints_infos_3d_t1[study_id][series_id] = pts\n\n return pred_sag_keypoints_infos_3d_t1\n\n\ndef infer_sag_3d_t2_keypoints_v20(data_root,\n study_ids,\n model_dir,\n model_dir2,\n device,\n num_workers=8,\n is_parallel=False):\n with_cascade = False\n if with_cascade:\n model_cfgs_t1_2d = [\n {\n 'sub_dir': 'keypoint_2d_v20_sag_t2/densenet161_lr_0.0006_sag_T2',\n 'backbone': 'densenet161',\n }\n ]\n v2_folds = 5\n models_t2_2d = []\n for cfg in model_cfgs_t1_2d:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(v2_folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n model = RSNA24Model_Keypoint_2D(backbone,\n num_classes=10,\n pretrained=False)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models_t2_2d.append(model)\n\n #\n pred_sag_keypoints_infos_3d_t2 = {}\n\n model_cfgs = [\n# {\n# 'sub_dir': 'keypoint_3d_v20_sag_t2/densenet161_lr_0.0006_sag_T2',\n# 'backbone': 'densenet161',\n# },\n {\n 'sub_dir': 'keypoint_3d_v20_sag_t2/gluon_resnet152_v1s_lr_0.0006_sag_T2_not_exclude_hard',\n 'backbone': 'gluon_resnet152_v1s',\n }\n ]\n folds = 5\n models = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n for fold in range(folds):\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n if not os.path.exists(fn):\n fn = os.path.join(model_dir2, sub_dir)\n fn = fn + f'/best_fold_{fold}_ema.pt'\n model = RSNA24Model_Keypoint_3D(backbone,\n in_chans=1,\n pretrained=False,\n num_classes=15).to(device)\n model.load_state_dict(torch.load(fn))\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models.append(model)\n\n dset = Sag_3D_Point_Dataset_V24(data_root,\n study_ids,\n test_series_descriptions_fn=test_series_descriptions_fn,\n image_dir=image_dir,\n series_description='Sagittal T2/STIR',\n cache_dir=CACHE_DIR,\n with_origin_arr=True if with_cascade else False\n )\n\n dloader = DataLoader(dset, batch_size=1 if with_cascade else 8,\n num_workers=num_workers)\n\n for tensor_dict in tqdm.tqdm(dloader,desc=f'{device}'):\n with torch.no_grad():\n x = tensor_dict['imgs'].to(device)\n study_id_list = tensor_dict['study_id']\n series_id_list = tensor_dict['series_id']\n with autocast:\n y = None\n for i in range(len(models)):\n p = models[i](x)\n if y is None:\n y = p\n else:\n y += p\n y = y / len(models)\n bs = y.shape[0]\n for b in range(bs):\n study_id = int(study_id_list[b])\n series_id = int(series_id_list[b])\n pts = y[b].reshape(-1, 5, 3).cpu().numpy()\n\n if with_cascade:\n origin_depth = int(tensor_dict['origin_depth'][b])\n origin_imgs = tensor_dict['origin_imgs'][b].to(device)\n sub_imgs = []\n for n in range(pts.shape[0]):\n for level in range(5):\n z = pts[n, level, 2] * origin_depth\n z = int(np.round(z))\n if z < 0:\n z = 0\n if z > origin_depth - 1:\n z = origin_depth - 1\n\n img = origin_imgs[z].unsqueeze(0).unsqueeze(0)\n sub_imgs.append(img)\n sub_imgs = torch.cat(sub_imgs, dim=0) # 2*5, 1, 512, 512\n xy_pred = _infer_sag_xy(models_t2_2d, sub_imgs).reshape(-1, 5, 5, 2).cpu().numpy()\n for n in range(pts.shape[0]):\n for level in range(5):\n pts[n, level, :2] = xy_pred[n, level, level]\n\n if study_id not in pred_sag_keypoints_infos_3d_t2:\n pred_sag_keypoints_infos_3d_t2[study_id] = {}\n pred_sag_keypoints_infos_3d_t2[study_id][series_id] = pts\n\n return pred_sag_keypoints_infos_3d_t2","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.144116Z","iopub.execute_input":"2024-10-08T16:36:51.144460Z","iopub.status.idle":"2024-10-08T16:36:51.235864Z","shell.execute_reply.started":"2024-10-08T16:36:51.144428Z","shell.execute_reply":"2024-10-08T16:36:51.235128Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"\ndef infer_all_keypoints(data_root,\n model_dir,\n model_dir2,\n with_v2_sag_center_slice_2d=True,\n with_v2_axial_3d=True,\n with_v2_sag_3d=True,\n with_v20_sag_t1_3d=True,\n with_v20_sag_t2_3d=True,\n with_v24_axial_3d=True\n ):\n print('with_v2_sag_center_slice_2d: ', with_v2_sag_center_slice_2d)\n print('with_v2_axial_3d: ', with_v2_axial_3d)\n print('with_v2_sag_3d: ', with_v2_sag_3d)\n print('with_v20_sag_t1_3d: ', with_v20_sag_t1_3d)\n print('with_v20_sag_t2_3d: ', with_v20_sag_t2_3d)\n print('with_v24_axial_3d: ', with_v24_axial_3d)\n\n df = pd.read_csv(f'{data_root}/{test_series_descriptions_fn}')\n study_ids = df['study_id'].unique().tolist()\n # study_ids = study_ids[:16]\n\n ret_dict = {}\n device_0 = torch.device('cuda:0')\n\n\n is_parallel = torch.cuda.device_count() > 1 and len(study_ids) >=2\n print('is_parallel: ', is_parallel)\n if with_v2_axial_3d or with_v24_axial_3d:\n pred_dict_v24, pred_dict_v2 = \\\n infer_axial_3d_keypoints_v24(data_root, study_ids, model_dir,model_dir2,\n device_0,N_WORKERS,is_parallel)\n\n ret_dict['axial_3d_keypoints_v2'] = pred_dict_v2\n ret_dict['axial_3d_keypoints_v24'] = pred_dict_v24\n\n\n if with_v2_sag_3d:\n print('infer_sag_3d_keypoints_v2...')\n pred_dict = infer_sag_3d_keypoints_v2(data_root, study_ids,\n model_dir, model_dir2, device_0, N_WORKERS,\n is_parallel)\n ret_dict['sag_3d_keypoints_v2'] = pred_dict\n\n\n if with_v2_sag_center_slice_2d:\n print('infer_sag_2d_keypoints_v2...')\n pred_dict = infer_sag_2d_keypoints_v2(data_root, study_ids,\n model_dir, model_dir2, device_0, N_WORKERS,\n is_parallel)\n ret_dict['sag_keypoints_v2'] = pred_dict\n\n\n if with_v20_sag_t1_3d:\n print('infer_sag_3d_t1_keypoints_v20...')\n pred_dict = infer_sag_3d_t1_keypoints_v20(data_root, study_ids,\n model_dir, model_dir2, device_0, N_WORKERS,\n is_parallel)\n ret_dict['sag_t1_3d_keypoints_v20'] = pred_dict\n\n if with_v20_sag_t2_3d:\n print('infer_sag_3d_t2_keypoints_v20..')\n pred_dict = infer_sag_3d_t2_keypoints_v20(data_root, study_ids,\n model_dir, model_dir2, device_0, N_WORKERS,\n is_parallel)\n ret_dict['sag_t2_3d_keypoints_v20'] = pred_dict\n\n return ret_dict","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.237083Z","iopub.execute_input":"2024-10-08T16:36:51.237457Z","iopub.status.idle":"2024-10-08T16:36:51.249853Z","shell.execute_reply.started":"2024-10-08T16:36:51.237426Z","shell.execute_reply":"2024-10-08T16:36:51.248938Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"\nif True:\n \n save_dir = keypoint_dir\n\n sag_keypoints_v2 = True\n axial_3d_keypoints_v2 = True\n sag_3d_keypoints_v2 = True\n\n sag_t1_3d_keypoints_v20 = True\n sag_t2_3d_keypoints_v20 = True\n axial_3d_keypoints_v24 = True\n\n sag_keypoints_v2_save_fn = f'{save_dir}/sag_keypoints_v2.pkl'\n axial_3d_keypoints_v2_save_fn = f'{save_dir}/axial_3d_keypoints_v2.pkl'\n sag_3d_keypoints_v2_save_fn = f'{save_dir}/sag_3d_keypoints_v2.pkl'\n sag_t1_3d_keypoints_v20_save_fn = f'{save_dir}/sag_t1_3d_keypoints_v20.pkl'\n sag_t2_3d_keypoints_v20_save_fn = f'{save_dir}/sag_t2_3d_keypoints_v20.pkl'\n axial_3d_keypoints_v24_fn = f'{save_dir}/axial_3d_keypoints_v24.pkl'\n\n if os.path.exists(sag_keypoints_v2_save_fn):\n sag_keypoints_v2 = False\n if os.path.exists(axial_3d_keypoints_v2_save_fn):\n axial_3d_keypoints_v2 = False\n if os.path.exists(sag_3d_keypoints_v2_save_fn):\n sag_3d_keypoints_v2 = False\n\n if os.path.exists(sag_t1_3d_keypoints_v20_save_fn):\n sag_t1_3d_keypoints_v20 = False\n if os.path.exists(sag_t2_3d_keypoints_v20_save_fn):\n sag_t2_3d_keypoints_v20 = False\n if os.path.exists(axial_3d_keypoints_v24_fn):\n axial_3d_keypoints_v24 = False\n\n ret_dict = infer_all_keypoints(data_root,\n model_dir,\n model_dir2,\n with_v2_sag_center_slice_2d=sag_keypoints_v2,\n with_v2_axial_3d=axial_3d_keypoints_v2,\n with_v2_sag_3d=sag_3d_keypoints_v2,\n with_v20_sag_t1_3d=sag_t1_3d_keypoints_v20,\n with_v20_sag_t2_3d=sag_t2_3d_keypoints_v20,\n with_v24_axial_3d=axial_3d_keypoints_v24, )\n\n for k, v in ret_dict.items():\n with open(f'{save_dir}/{k}.pkl', 'wb') as file_handle:\n pickle.dump(v, file_handle)","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:36:51.253224Z","iopub.execute_input":"2024-10-08T16:36:51.253624Z","iopub.status.idle":"2024-10-08T16:39:28.032853Z","shell.execute_reply.started":"2024-10-08T16:36:51.253600Z","shell.execute_reply":"2024-10-08T16:39:28.031581Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 2. condition classification","metadata":{}},{"cell_type":"code","source":"def infer_v2(data_root,\n study_ids,\n model_dir,\n save_dir,\n device,\n num_workers=8,\n folds=None,\n is_parallel=False):\n model_cfgs = [\n {\n 'sub_dir': 'v2_cond/pvt_v2_b2.in1k_axial_pvt_v2_b2.in1k_axial_size_256_sag_size_128',\n 'backbone_sag': 'pvt_v2_b2.in1k',\n 'backbone_axial': 'pvt_v2_b2.in1k',\n 'weight': 0.3469085736266617,\n },\n\n {\n 'sub_dir': 'v2_cond/convnext_small.in12k_ft_in1k_384_axial_densenet161_axial_size_256_sag_size_128',\n 'backbone_sag': 'convnext_small.in12k_ft_in1k_384',\n 'backbone_axial': 'densenet161',\n 'weight': 0.20715647998779071,\n },\n {\n 'sub_dir': 'v2_cond/convnext_nano.in12k_ft_in1k_axial_convnext_nano.in12k_ft_in1k_axial_size_256_sag_size_128',\n 'backbone_sag': 'convnext_nano.in12k_ft_in1k',\n 'backbone_axial': 'convnext_nano.in12k_ft_in1k',\n 'weight': 0.1927349675794976,\n },\n {\n 'sub_dir': 'v2_cond/pvt_v2_b1.in1k_axial_pvt_v2_b1.in1k_axial_size_256_sag_size_128',\n 'backbone_sag': 'pvt_v2_b1.in1k',\n 'backbone_axial': 'pvt_v2_b1.in1k',\n 'weight': 0.13677624328967222,\n },\n {\n 'sub_dir': 'v2_cond/pvt_v2_b1.in1k_axial_densenet161_axial_size_256_sag_size_128',\n 'backbone_sag': 'pvt_v2_b1.in1k',\n 'backbone_axial': 'densenet161',\n 'weight': 0.11642373551637777,\n },\n\n ]\n v2_folds = list(range(5))\n if folds is not None:\n v2_folds = folds\n\n models = []\n weights = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone_sag = cfg['backbone_sag']\n backbone_axial = cfg['backbone_axial']\n w = cfg['weight']\n for fold in v2_folds:\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_wll_model_fold-{fold}_score.pt'\n print('load: ', fn)\n model = HybridModel_V2(backbone_sag,\n backbone_axial,\n pretrained=False)\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n models.append(model)\n weights.append(w)\n\n weights_sum = np.sum(weights)\n #\n with open(f'{keypoint_dir}/sag_keypoints_v2.pkl', 'rb') as f:\n sag_keypoints_v2 = pickle.load(f)\n\n with open(f'{keypoint_dir}/axial_3d_keypoints_v2.pkl', 'rb') as f:\n axial_3d_keypoints_v2 = pickle.load(f)\n\n valid_ds = RSNA24DatasetTest_LHW_V2(data_root,\n study_ids,\n test_series_descriptions_fn,\n study_id_to_pred_keypoints_sag=sag_keypoints_v2,\n study_id_to_pred_keypoints_axial=axial_3d_keypoints_v2,\n image_dir=data_root + f'{image_dir}/',\n cache_dir=CACHE_DIR)\n valid_dl = DataLoader(\n valid_ds,\n batch_size=4,\n shuffle=False,\n pin_memory=False,\n drop_last=False,\n num_workers=num_workers,\n collate_fn=v2_collate_fn\n )\n\n all_preds = {}\n for tensor_d in tqdm.tqdm(valid_dl):\n with torch.no_grad():\n x = tensor_d['img'].to(device)\n axial_imgs = tensor_d['axial_imgs'].to(device)\n n_list = tensor_d['n_list']\n cond = tensor_d['cond'].to(device)\n study_id_list = tensor_d['study_id']\n preds = None\n with autocast:\n for i in range(len(models)):\n p = models[i](x, axial_imgs, cond, n_list)\n if preds is None:\n preds = (p.float() * weights[i])\n else:\n preds += (p.float() * weights[i])\n preds = preds / weights_sum\n bs, _ = preds.shape\n # bs, 5_cond, 5_level, 3\n preds = preds.reshape(bs, 5, 5, 3).cpu()\n for b in range(bs):\n study_id = int(study_id_list[b])\n all_preds[study_id] = preds[b]\n\n return all_preds\n\n\ndef infer_v24( # df_valid,\n data_root,\n study_ids,\n model_dir,\n save_dir,\n device,\n num_workers=8,\n folds=None,\n is_parallel=False,\n):\n def _forward_tta(model, x, cond):\n # return model(x, cond)\n x_flip = torch.flip(x, dims=[-1, ]) # A.HorizontalFlip(p=0.5),\n x0 = model(x, cond)\n x1 = model(x_flip, cond)\n return (x0 + x1) / 2\n\n # t1\n\n model_cfgs = [\n {\n 'sub_dir': 'v20_cond_t1/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_8620_h128_w128_legacy',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n 'with_cond': True,\n 'z_imgs': 3,\n 'with_gru': False,\n 'with_level_lstm': False,\n 'data_key': '3_128_128',\n 'weight': 0.20538912558481712,\n },\n {\n 'sub_dir': 'v20_cond_t1/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_8620_h72_w128',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': False,\n 'with_level_lstm': False,\n 'data_key': '3_72_128',\n 'weight': 0.18552050200612183,\n },\n {\n 'sub_dir': 'v20_cond_t1/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_8620_h72_w128_with_gru',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': True,\n 'with_level_lstm': False,\n 'data_key': '3_72_128',\n 'weight': 0.306110664432234,\n },\n {\n 'sub_dir': 'v20_cond_t1/convnext_tiny.in12k_ft_in1k_384_z_imgs_3_seed_8620_h128_w128_with_gru',\n 'backbone': 'convnext_tiny.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': True,\n 'with_level_lstm': False,\n 'data_key': '3_128_128',\n 'weight': 0.30297970797682705,\n }\n\n ]\n v24_folds = list(range(5))\n if folds is not None:\n v24_folds = folds\n\n models = []\n data_keys = []\n weights = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n with_cond = cfg['with_cond']\n with_gru = cfg['with_gru']\n with_level_lstm = cfg['with_level_lstm']\n z_imgs = cfg['z_imgs']\n data_key = cfg['data_key']\n w = cfg['weight']\n for fold in v24_folds:\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_wll_model_fold-{fold}_ema.pt'\n print('load: ', fn)\n model = build_v20_sag_model(\n backbone,\n with_cond=with_cond,\n with_gru=with_gru,\n with_level_lstm=with_level_lstm,\n z_imgs=z_imgs,\n pretrained=False\n )\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n\n models.append(model)\n weights.append(w)\n data_keys.append(data_key)\n weights_sum = np.sum(weights)\n with open(f'{keypoint_dir}/sag_t1_3d_keypoints_v20.pkl', 'rb') as f:\n pred_sag_keypoints_infos_3d_t1 = pickle.load(f)\n\n other_crop_size_list = [\n (3, 72, 128),\n ]\n dset = Sag_T1_Dataset_V24(data_root,\n study_ids,\n pred_sag_keypoints_infos_3d_t1,\n test_series_descriptions_fn=test_series_descriptions_fn,\n image_dir=image_dir,\n crop_size_h=128,\n crop_size_w=128,\n cache_dir=CACHE_DIR,\n other_crop_size_list=other_crop_size_list\n )\n dloader = DataLoader(\n dset,\n batch_size=batch_size,\n shuffle=False,\n pin_memory=False,\n drop_last=False,\n num_workers=num_workers,\n )\n all_preds_t1 = {}\n sag_t1_preds = []\n for tensor_d in tqdm.tqdm(dloader):\n with torch.no_grad():\n for k in tensor_d.keys():\n if k not in ['study_id']:\n tensor_d[k] = tensor_d[k].to(device)\n study_id_list = tensor_d['study_id']\n cond = tensor_d['cond']\n preds = None\n with autocast:\n for i in range(len(models)):\n p = _forward_tta(models[i], tensor_d[data_keys[i]], cond)\n if preds is None:\n preds = (p.float() * weights[i])\n else:\n preds += (p.float() * weights[i])\n preds = preds / weights_sum\n bs, _ = preds.shape\n # bs, n_cond, 5_level, 3\n preds = preds.reshape(bs, -1, 5, 3).cpu()\n sag_t1_preds.append(preds)\n for b in range(bs):\n study_id = int(study_id_list[b])\n all_preds_t1[study_id] = preds[b]\n sag_t1_preds = torch.cat(sag_t1_preds, dim=0)\n\n # t2\n model_cfgs = [\n {\n 'sub_dir': 'v20_cond_t2_part2/rexnetr_200.sw_in12k_ft_in1k_z_imgs_3_seed_8620_h128_w128_level_lstm',\n 'backbone': 'rexnetr_200.sw_in12k_ft_in1k',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': False,\n 'with_level_lstm': True,\n 'data_key': '3_128_128',\n 'weight': 0.2523387488397007,\n },\n\n {\n 'sub_dir': 'v20_cond_t2/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_8620_h128_w128_with_gru',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': True,\n 'with_level_lstm': False,\n 'data_key': '3_128_128',\n 'weight': 0.24855049213914635,\n },\n {\n 'sub_dir': 'v20_cond_t2_part2/pvt_v2_b5.in1k_z_imgs_5_seed_8620_h128_w128_level_lstm',\n 'backbone': 'pvt_v2_b5.in1k',\n 'with_cond': False,\n 'z_imgs': 5,\n 'with_gru': False,\n 'with_level_lstm': True,\n 'data_key': '5_128_128',\n 'weight': 0.19532993129183054,\n },\n {\n 'sub_dir': 'v20_cond_t2/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_8620_h64_w96_level_lstm',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': False,\n 'with_level_lstm': True,\n 'data_key': '3_64_96',\n 'weight': 0.14061503567169062,\n },\n {\n 'sub_dir': 'v20_cond_t2_part2/convnext_tiny.in12k_ft_in1k_384_z_imgs_3_seed_8620_h64_w96_level_lstm',\n 'backbone': 'convnext_tiny.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 3,\n 'with_gru': False,\n 'with_level_lstm': True,\n 'data_key': '3_64_96',\n 'weight': 0.090578570264412,\n },\n\n {\n 'sub_dir': 'v20_cond_t2_part2/convnext_tiny.in12k_ft_in1k_384_z_imgs_5_seed_8620_h128_w128_level_lstm',\n 'backbone': 'convnext_tiny.in12k_ft_in1k_384',\n 'with_cond': False,\n 'z_imgs': 5,\n 'with_gru': False,\n 'with_level_lstm': True,\n 'data_key': '5_128_128',\n 'weight': 0.0725872217932197,\n },\n ]\n v24_folds = list(range(5))\n if folds is not None:\n v24_folds = folds\n\n models = []\n data_keys = []\n weights = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n with_cond = cfg['with_cond']\n with_gru = cfg['with_gru']\n with_level_lstm = cfg['with_level_lstm']\n z_imgs = cfg['z_imgs']\n data_key = cfg['data_key']\n w = cfg['weight']\n for fold in v24_folds:\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_wll_model_fold-{fold}_ema.pt'\n print('load: ', fn)\n model = build_v20_sag_model(\n backbone,\n with_cond=with_cond,\n with_gru=with_gru,\n with_level_lstm=with_level_lstm,\n z_imgs=z_imgs,\n pretrained=False\n )\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n\n models.append(model)\n weights.append(w)\n data_keys.append(data_key)\n weights_sum = np.sum(weights)\n\n with open(f'{keypoint_dir}/sag_t2_3d_keypoints_v20.pkl', 'rb') as f:\n pred_sag_keypoints_infos_3d_t2 = pickle.load(f)\n\n# with open(f'{keypoint_dir}/sag_3d_keypoints_v2.pkl', 'rb') as f:\n# pred_sag_keypoints_infos_3d = pickle.load(f)\n\n# for study_id in pred_sag_keypoints_infos_3d_t2.keys():\n# pred_keypoints = pred_sag_keypoints_infos_3d[study_id]['points'].reshape(15, 3)\n# t2_pred_keypoints = pred_keypoints[:5]\n# t2_pred_keypoints[:, :2] = t2_pred_keypoints[:, :2] / 4.0\n# t2_pred_keypoints[:, 2] = t2_pred_keypoints[:, 2] / 16.0\n# for sid in pred_sag_keypoints_infos_3d_t2[study_id].keys():\n# pred_sag_keypoints_infos_3d_t2[study_id][sid] = t2_pred_keypoints\n\n other_crop_size_list = [\n (3, 64, 96),\n (5, 128, 128),\n ]\n \n dset = Sag_T2_Dataset_V24(data_root,\n study_ids,\n pred_sag_keypoints_infos_3d_t2,\n test_series_descriptions_fn=test_series_descriptions_fn,\n image_dir=image_dir,\n crop_size_h=128,\n crop_size_w=128,\n cache_dir=CACHE_DIR,\n other_crop_size_list=other_crop_size_list\n )\n dloader = DataLoader(\n dset,\n batch_size=batch_size,\n shuffle=False,\n pin_memory=False,\n drop_last=False,\n num_workers=num_workers,\n )\n all_preds_t2 = {}\n sag_t2_preds = []\n for tensor_d in tqdm.tqdm(dloader):\n with torch.no_grad():\n for k in tensor_d.keys():\n if k not in ['study_id']:\n tensor_d[k] = tensor_d[k].to(device)\n study_id_list = tensor_d['study_id']\n cond = None # tensor_d['cond']\n preds = None\n with autocast:\n for i in range(len(models)):\n # p = _forward_tta(models[i], tensor_d['img'], cond)\n p = _forward_tta(models[i], tensor_d[data_keys[i]], cond)\n if preds is None:\n preds = (p.float() * weights[i])\n else:\n preds += (p.float() * weights[i])\n preds = preds / weights_sum\n bs, _ = preds.shape\n # bs, n_cond, 5_level, 3\n preds = preds.reshape(bs, -1, 5, 3).cpu()\n sag_t2_preds.append(preds)\n for b in range(bs):\n study_id = int(study_id_list[b])\n all_preds_t2[study_id] = preds[b]\n sag_t2_preds = torch.cat(sag_t2_preds, dim=0)\n # return all_preds_t2\n # axial\n\n model_cfgs = [\n {\n 'sub_dir': 'v24_cond_axial/convnext_small.in12k_ft_in1k_384_z_imgs_3',\n 'backbone': 'convnext_small.in12k_ft_in1k_384',\n 'axial_in_channels': 1,\n 'weight': 0.19731044421522836,\n 'z_imgs': 3,\n 'data_key': 'axial_imgs_3_128_128'\n },\n {\n 'sub_dir': 'v24_cond_axial/convnext_tiny.in12k_ft_in1k_384_z_imgs_3',\n 'backbone': 'convnext_tiny.in12k_ft_in1k',\n 'axial_in_channels': 1,\n 'weight': 0.30013472174512595,\n 'z_imgs': 3,\n 'data_key': 'axial_imgs_3_128_128'\n },\n {\n 'sub_dir': 'v24_cond_axial/densenet161_z_imgs_3',\n 'backbone': 'densenet161',\n 'axial_in_channels': 1,\n 'weight': 0.1627137568736331,\n 'z_imgs': 3,\n 'data_key': 'axial_imgs_3_128_128'\n },\n {\n 'sub_dir': 'v24_cond_axial/pvt_v2_b1.in1k_z_imgs_3',\n 'backbone': 'pvt_v2_b1.in1k',\n 'axial_in_channels': 1,\n 'weight': 1.0,\n 'z_imgs': 0.0987451411551922,\n 'data_key': 'axial_imgs_3_128_128'\n },\n {\n 'sub_dir': 'v24_cond_axial/rexnetr_200.sw_in12k_ft_in1k_z_imgs_5',\n 'backbone': 'rexnetr_200.sw_in12k_ft_in1k',\n 'axial_in_channels': 3,\n 'weight': 0.2410959360108204,\n 'z_imgs': 5,\n 'data_key': 'axial_imgs_5_128_128'\n },\n\n ]\n v24_folds = list(range(5))\n if folds is not None:\n v24_folds = folds\n\n models = []\n data_keys = []\n weights = []\n for cfg in model_cfgs:\n sub_dir = cfg['sub_dir']\n backbone = cfg['backbone']\n data_key = cfg['data_key']\n axial_in_channels = cfg['axial_in_channels']\n w = cfg['weight']\n for fold in v24_folds:\n fn = os.path.join(model_dir, sub_dir)\n fn = fn + f'/best_wll_model_fold-{fold}_ema.pt'\n print('load: ', fn)\n model = Axial_HybridModel_24(backbone,\n backbone,\n pretrained=False,\n axial_in_channels=axial_in_channels\n )\n model.load_state_dict(torch.load(fn))\n model.to(device)\n model.eval()\n if is_parallel:\n model = nn.DataParallel(model)\n\n models.append(model)\n weights.append(w)\n data_keys.append(data_key)\n weights_sum = np.sum(weights)\n\n with open(f'{keypoint_dir}/axial_3d_keypoints_v24.pkl', 'rb') as f:\n axial_3d_keypoints_v24 = pickle.load(f)\n\n other_crop_size_list = [\n (5, 128, 128),\n ]\n dset = Axial_Cond_Dataset_Multi_V24(data_root,\n study_ids,\n axial_3d_keypoints_v24,\n pred_sag_keypoints_infos_3d_t2,\n test_series_descriptions_fn=test_series_descriptions_fn,\n image_dir=image_dir,\n z_imgs=3,\n other_z_imgs_list=other_crop_size_list,\n cache_dir=CACHE_DIR,\n )\n dloader = DataLoader(\n dset,\n batch_size=batch_size,\n shuffle=False,\n pin_memory=False,\n drop_last=False,\n num_workers=num_workers,\n )\n all_preds_axial = {}\n axial_preds = []\n for tensor_d in tqdm.tqdm(dloader):\n with torch.no_grad():\n for k in tensor_d.keys():\n if k not in ['study_id']:\n tensor_d[k] = tensor_d[k].to(device)\n study_id_list = tensor_d['study_id']\n preds = None\n with autocast:\n for i in range(len(models)):\n p = models[i](tensor_d['sag_t2'], tensor_d[data_keys[i]])\n if preds is None:\n preds = (p.float() * weights[i])\n else:\n preds += (p.float() * weights[i])\n preds = preds / weights_sum\n bs, _ = preds.shape\n # bs, n_cond, 5_level, 3\n preds = preds.reshape(bs, -1, 5, 3).cpu()\n axial_preds.append(preds)\n for b in range(bs):\n study_id = int(study_id_list[b])\n all_preds_axial[study_id] = preds[b]\n\n axial_preds = torch.cat(axial_preds, dim=0)\n all_preds = torch.cat((sag_t2_preds, sag_t1_preds, axial_preds), dim=1)\n #all_preds = torch.cat((sag_t2_preds, sag_t1_preds), dim=1)\n all_preds_dict = {}\n for i, study_id in enumerate(study_ids):\n all_preds_dict[study_id] = all_preds[i]\n return all_preds_dict","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:39:28.034613Z","iopub.execute_input":"2024-10-08T16:39:28.034976Z","iopub.status.idle":"2024-10-08T16:39:28.104311Z","shell.execute_reply.started":"2024-10-08T16:39:28.034943Z","shell.execute_reply":"2024-10-08T16:39:28.103336Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def infer_all(data_root,\n model_dir,\n ):\n df = pd.read_csv(f'{data_root}/{test_series_descriptions_fn}')\n study_ids = df['study_id'].unique().tolist()\n \n device_0 = torch.device('cuda:0')\n\n is_parallel = torch.cuda.device_count() > 1 and len(study_ids) >=2\n print('is_parallel: ', is_parallel)\n folds = None\n \n pred_dict_v24 = infer_v24(data_root, study_ids, model_dir,\n save_dir, device_0, N_WORKERS, folds, is_parallel)\n\n\n pred_dict_v2 = infer_v2(data_root, study_ids, model_dir,\n save_dir, device_0, N_WORKERS, folds, is_parallel)\n\n \n print(len(study_ids))\n fold_preds = []\n for study_id in study_ids:\n pred = pred_dict_v2[study_id]\n fold_preds.append(pred.unsqueeze(0))\n fold_preds = torch.cat(fold_preds, dim=0)\n \n print('fold_preds shape: ', fold_preds.shape)\n\n fold_preds_v24 = []\n for study_id in study_ids:\n pred = pred_dict_v24[study_id]\n fold_preds_v24.append(pred.unsqueeze(0))\n fold_preds_v24 = torch.cat(fold_preds_v24, dim=0)\n\n fold_preds_en = fold_preds\n fold_preds_en[:, 0:1] = 0.5896 * fold_preds[:, 0:1] + 0.4104 * fold_preds_v24[:, 0:1]\n fold_preds_en[:, 1:3] = 0.09 * fold_preds[:, 1:3] + 0.91 * fold_preds_v24[:, 1:3]\n fold_preds_en[:, 3:5] = 0.38 * fold_preds[:, 3:5] + 0.62 * fold_preds_v24[:, 3:5]\n\n fold_preds_en = nn.Softmax(dim=-1)(fold_preds_en).cpu().numpy()\n return fold_preds_en","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:39:28.105498Z","iopub.execute_input":"2024-10-08T16:39:28.106951Z","iopub.status.idle":"2024-10-08T16:39:28.119039Z","shell.execute_reply.started":"2024-10-08T16:39:28.106925Z","shell.execute_reply":"2024-10-08T16:39:28.118157Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# def infer_v2_seed666(data_root,\n# study_ids,\n# model_dir,\n# save_dir,\n# device,\n# num_workers=8,\n# folds=None,\n# is_parallel=False):\n# model_cfgs = [\n# {\n# 'sub_dir': 'v2_cond/pvt_v2_b1.in1k_axial_pvt_v2_b1.in1k_axial_size_256_sag_size_128',\n# 'backbone_sag': 'pvt_v2_b1.in1k',\n# 'backbone_axial': 'pvt_v2_b1.in1k',\n# 'weight': 0.34092462492421793,\n# },\n# {\n# 'sub_dir': 'v2_cond/pvt_v2_b1.in1k_axial_densenet161_axial_size_256_sag_size_128',\n# 'backbone_sag': 'pvt_v2_b1.in1k',\n# 'backbone_axial': 'densenet161',\n# 'weight': 0.2458979925774557,\n# },\n# {\n# 'sub_dir': 'v2_cond/pvt_v2_b2.in1k_axial_pvt_v2_b2.in1k_axial_size_256_sag_size_128',\n# 'backbone_sag': 'pvt_v2_b2.in1k',\n# 'backbone_axial': 'pvt_v2_b2.in1k',\n# 'weight': 0.21641664109925535,\n# },\n# {\n# 'sub_dir': 'v2_cond/convnext_nano.in12k_ft_in1k_axial_convnext_nano.in12k_ft_in1k_axial_size_256_sag_size_128',\n# 'backbone_sag': 'convnext_nano.in12k_ft_in1k',\n# 'backbone_axial': 'convnext_nano.in12k_ft_in1k',\n# 'weight': 0.19676074139907107,\n# },\n\n# ]\n# v2_folds = list(range(5))\n# if folds is not None:\n# v2_folds = folds\n\n# models = []\n# weights = []\n# for cfg in model_cfgs:\n# sub_dir = cfg['sub_dir']\n# backbone_sag = cfg['backbone_sag']\n# backbone_axial = cfg['backbone_axial']\n# w = cfg['weight']\n# for fold in v2_folds:\n# fn = os.path.join(model_dir, sub_dir)\n# fn = fn + f'/best_wll_model_fold-{fold}_score.pt'\n# print('load: ', fn)\n# model = HybridModel_V2(backbone_sag,\n# backbone_axial,\n# pretrained=False,\n# fea_dim=512)\n# model.load_state_dict(torch.load(fn))\n# model.to(device)\n# model.eval()\n# if is_parallel:\n# model = nn.DataParallel(model)\n# models.append(model)\n# weights.append(w)\n\n# weights_sum = np.sum(weights)\n# #\n# with open(f'{keypoint_dir}/sag_keypoints_v2.pkl', 'rb') as f:\n# sag_keypoints_v2 = pickle.load(f)\n\n# with open(f'{keypoint_dir}/axial_3d_keypoints_v2.pkl', 'rb') as f:\n# axial_3d_keypoints_v2 = pickle.load(f)\n# #\n# # with open(f'{keypoint_dir}/sag_t1_3d_keypoints_v20.pkl', 'rb') as f:\n# # pred_sag_keypoints_infos_3d_t1 = pickle.load(f)\n\n# valid_ds = RSNA24DatasetTest_LHW_V2(data_root,\n# study_ids,\n# test_series_descriptions_fn,\n# study_id_to_pred_keypoints_sag=sag_keypoints_v2,\n# study_id_to_pred_keypoints_axial=axial_3d_keypoints_v2,\n# # pred_sag_keypoints_infos_3d_t1=pred_sag_keypoints_infos_3d_t1,\n# image_dir=data_root + f'{image_dir}/',\n# cache_dir=CACHE_DIR)\n# valid_dl = DataLoader(\n# valid_ds,\n# batch_size=4,\n# shuffle=False,\n# pin_memory=False,\n# drop_last=False,\n# num_workers=num_workers,\n# collate_fn=v2_collate_fn\n# )\n\n# all_preds = {}\n# for tensor_d in tqdm.tqdm(valid_dl):\n# with torch.no_grad():\n# x = tensor_d['img'].to(device)\n# axial_imgs = tensor_d['axial_imgs'].to(device)\n# n_list = tensor_d['n_list']\n# cond = tensor_d['cond'].to(device)\n# study_id_list = tensor_d['study_id']\n# preds = None\n# with autocast:\n# for i in range(len(models)):\n# p = models[i](x, axial_imgs, cond, n_list)\n# if preds is None:\n# preds = (p.float() * weights[i])\n# else:\n# preds += (p.float() * weights[i])\n# preds = preds / weights_sum\n# bs, _ = preds.shape\n# # bs, 5_cond, 5_level, 3\n# preds = preds.reshape(bs, 5, 5, 3).cpu()\n# for b in range(bs):\n# study_id = int(study_id_list[b])\n# all_preds[study_id] = preds[b]\n\n# return all_preds\n\n\n# def infer_v24_seed666( # df_valid,\n# data_root,\n# study_ids,\n# model_dir,\n# save_dir,\n# device,\n# num_workers=8,\n# folds=None,\n# is_parallel=False,\n# ):\n# def _forward_tta(model, x, cond):\n# # return model(x, cond)\n# x_flip = torch.flip(x, dims=[-1, ]) # A.HorizontalFlip(p=0.5),\n# x0 = model(x, cond)\n# x1 = model(x_flip, cond)\n# return (x0 + x1) / 2\n\n# # t1\n# model_cfgs = [\n# {\n# 'sub_dir': 'v20_cond_t1/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_666_h128_w128',\n# 'backbone': 'convnext_small.in12k_ft_in1k_384',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': False,\n# 'with_level_lstm': False,\n# 'data_key': '3_128_128',\n# 'weight': 0.27957959035255875,\n# },\n# {\n# 'sub_dir': 'v20_cond_t1/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_666_h72_w128_with_gru',\n# 'backbone': 'convnext_small.in12k_ft_in1k_384',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': True,\n# 'with_level_lstm': False,\n# 'data_key': '3_72_128',\n# 'weight': 0.2299256035844461,\n# },\n# {\n# 'sub_dir': 'v20_cond_t1/rexnetr_200.sw_in12k_ft_in1k_z_imgs_3_seed_666_h72_w128_with_gru',\n# 'backbone': 'rexnetr_200.sw_in12k_ft_in1k',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': True,\n# 'with_level_lstm': False,\n# 'data_key': '3_72_128',\n# 'weight': 0.22004214729630475,\n# },\n# {\n# 'sub_dir': 'v20_cond_t1/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_666_h72_w128',\n# 'backbone': 'convnext_small.in12k_ft_in1k_384',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': False,\n# 'with_level_lstm': False,\n# 'data_key': '3_72_128',\n# 'weight': 0.13829760915067793,\n# },\n# {\n# 'sub_dir': 'v20_cond_t1/convnext_tiny.in12k_ft_in1k_384_z_imgs_3_seed_666_h72_w128_with_gru',\n# 'backbone': 'convnext_tiny.in12k_ft_in1k_384',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': True,\n# 'with_level_lstm': False,\n# 'data_key': '3_72_128',\n# 'weight': 0.13215504961601252,\n# }\n\n# ]\n# v24_folds = list(range(5))\n# if folds is not None:\n# v24_folds = folds\n\n# models = []\n# data_keys = []\n# weights = []\n# for cfg in model_cfgs:\n# sub_dir = cfg['sub_dir']\n# backbone = cfg['backbone']\n# with_cond = cfg['with_cond']\n# with_gru = cfg['with_gru']\n# with_level_lstm = cfg['with_level_lstm']\n# z_imgs = cfg['z_imgs']\n# data_key = cfg['data_key']\n# w = cfg['weight']\n# for fold in v24_folds:\n# fn = os.path.join(model_dir, sub_dir)\n# fn = fn + f'/best_wll_model_fold-{fold}_ema.pt'\n# print('load: ', fn)\n# model = build_v20_sag_model(\n# backbone,\n# with_cond=with_cond,\n# with_gru=with_gru,\n# with_level_lstm=with_level_lstm,\n# z_imgs=z_imgs,\n# pretrained=False\n# )\n# model.load_state_dict(torch.load(fn))\n# model.to(device)\n# model.eval()\n# if is_parallel:\n# model = nn.DataParallel(model)\n\n# models.append(model)\n# weights.append(w)\n# data_keys.append(data_key)\n# weights_sum = np.sum(weights)\n# with open(f'{keypoint_dir}/sag_t1_3d_keypoints_v20.pkl', 'rb') as f:\n# pred_sag_keypoints_infos_3d_t1 = pickle.load(f)\n\n# other_crop_size_list = [\n# (3, 72, 128),\n# ]\n# dset = Sag_T1_Dataset_V24(data_root,\n# study_ids,\n# pred_sag_keypoints_infos_3d_t1,\n# test_series_descriptions_fn=test_series_descriptions_fn,\n# image_dir=image_dir,\n# crop_size_h=128,\n# crop_size_w=128,\n# cache_dir=CACHE_DIR,\n# other_crop_size_list=other_crop_size_list\n# )\n# dloader = DataLoader(\n# dset,\n# batch_size=batch_size,\n# shuffle=False,\n# pin_memory=False,\n# drop_last=False,\n# num_workers=num_workers,\n# )\n# all_preds_t1 = {}\n# sag_t1_preds = []\n# for tensor_d in tqdm.tqdm(dloader):\n# with torch.no_grad():\n# for k in tensor_d.keys():\n# if k not in ['study_id']:\n# tensor_d[k] = tensor_d[k].to(device)\n# study_id_list = tensor_d['study_id']\n# cond = tensor_d['cond']\n# preds = None\n# with autocast:\n# for i in range(len(models)):\n# p = _forward_tta(models[i], tensor_d[data_keys[i]], cond)\n# if preds is None:\n# preds = (p.float() * weights[i])\n# else:\n# preds += (p.float() * weights[i])\n# preds = preds / weights_sum\n# bs, _ = preds.shape\n# # bs, n_cond, 5_level, 3\n# preds = preds.reshape(bs, -1, 5, 3).cpu()\n# sag_t1_preds.append(preds)\n# for b in range(bs):\n# study_id = int(study_id_list[b])\n# all_preds_t1[study_id] = preds[b]\n# sag_t1_preds = torch.cat(sag_t1_preds, dim=0)\n\n# # t2\n# model_cfgs = [\n# {\n# 'sub_dir': 'v20_cond_t2/rexnetr_200.sw_in12k_ft_in1k_z_imgs_3_seed_666_h128_w128_level_lstm',\n# 'backbone': 'rexnetr_200.sw_in12k_ft_in1k',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': False,\n# 'with_level_lstm': True,\n# 'data_key': '3_128_128',\n# 'weight': 0.2676011522464479,\n# },\n\n# {\n# 'sub_dir': 'v20_cond_t2/convnext_small.in12k_ft_in1k_384_z_imgs_3_seed_666_h128_w128_with_gru',\n# 'backbone': 'convnext_small.in12k_ft_in1k_384',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': True,\n# 'with_level_lstm': False,\n# 'data_key': '3_128_128',\n# 'weight': 0.25537269026075604,\n# },\n# {\n# 'sub_dir': 'v20_cond_t2/pvt_v2_b1.in1k_z_imgs_3_seed_666_h64_w96_with_gru',\n# 'backbone': 'pvt_v2_b1.in1k',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': True,\n# 'with_level_lstm': True,\n# 'data_key': '3_64_96',\n# 'weight': 0.20531756213739447,\n# },\n# {\n# 'sub_dir': 'v20_cond_t2/convnext_tiny.in12k_ft_in1k_z_imgs_5_seed_666_h128_w128_level_lstm',\n# 'backbone': 'convnext_tiny.in12k_ft_in1k',\n# 'with_cond': False,\n# 'z_imgs': 5,\n# 'with_gru': False,\n# 'with_level_lstm': True,\n# 'data_key': '5_128_128',\n# 'weight': 0.16238787669016758,\n# },\n# {\n# 'sub_dir': 'v20_cond_t2/convnext_tiny.in12k_ft_in1k_z_imgs_3_seed_666_h64_w96_level_lstm',\n# 'backbone': 'convnext_tiny.in12k_ft_in1k',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': False,\n# 'with_level_lstm': True,\n# 'data_key': '3_64_96',\n# 'weight': 0.08318722948976684,\n# },\n\n# {\n# 'sub_dir': 'v20_cond_t2/pvt_v2_b2.in1k_z_imgs_3_seed_666_h128_w128_level_lstm',\n# 'backbone': 'pvt_v2_b2.in1k',\n# 'with_cond': False,\n# 'z_imgs': 3,\n# 'with_gru': False,\n# 'with_level_lstm': True,\n# 'data_key': '3_128_128',\n# 'weight': 0.02613348917546713,\n# },\n# ]\n# v24_folds = list(range(5))\n# if folds is not None:\n# v24_folds = folds\n\n# models = []\n# data_keys = []\n# weights = []\n# for cfg in model_cfgs:\n# sub_dir = cfg['sub_dir']\n# backbone = cfg['backbone']\n# with_cond = cfg['with_cond']\n# with_gru = cfg['with_gru']\n# with_level_lstm = cfg['with_level_lstm']\n# z_imgs = cfg['z_imgs']\n# data_key = cfg['data_key']\n# w = cfg['weight']\n# for fold in v24_folds:\n# fn = os.path.join(model_dir, sub_dir)\n# fn = fn + f'/best_wll_model_fold-{fold}_ema.pt'\n# print('load: ', fn)\n# model = build_v20_sag_model(\n# backbone,\n# with_cond=with_cond,\n# with_gru=with_gru,\n# with_level_lstm=with_level_lstm,\n# z_imgs=z_imgs,\n# pretrained=False\n# )\n# model.load_state_dict(torch.load(fn))\n# model.to(device)\n# model.eval()\n# if is_parallel:\n# model = nn.DataParallel(model)\n\n# models.append(model)\n# weights.append(w)\n# data_keys.append(data_key)\n# weights_sum = np.sum(weights)\n\n# with open(f'{keypoint_dir}/sag_t2_3d_keypoints_v20.pkl', 'rb') as f:\n# pred_sag_keypoints_infos_3d_t2 = pickle.load(f)\n\n# other_crop_size_list = [\n# (3, 64, 96),\n# (5, 128, 128),\n# ]\n\n# dset = Sag_T2_Dataset_V24(data_root,\n# study_ids,\n# pred_sag_keypoints_infos_3d_t2,\n# test_series_descriptions_fn=test_series_descriptions_fn,\n# image_dir=image_dir,\n# crop_size_h=128,\n# crop_size_w=128,\n# cache_dir=CACHE_DIR,\n# other_crop_size_list=other_crop_size_list\n# )\n# dloader = DataLoader(\n# dset,\n# batch_size=batch_size,\n# shuffle=False,\n# pin_memory=False,\n# drop_last=False,\n# num_workers=num_workers,\n# )\n# all_preds_t2 = {}\n# sag_t2_preds = []\n# for tensor_d in tqdm.tqdm(dloader):\n# with torch.no_grad():\n# for k in tensor_d.keys():\n# if k not in ['study_id']:\n# tensor_d[k] = tensor_d[k].to(device)\n# study_id_list = tensor_d['study_id']\n# cond = None # tensor_d['cond']\n# preds = None\n# with autocast:\n# for i in range(len(models)):\n# # p = _forward_tta(models[i], tensor_d['img'], cond)\n# p = _forward_tta(models[i], tensor_d[data_keys[i]], cond)\n# if preds is None:\n# preds = (p.float() * weights[i])\n# else:\n# preds += (p.float() * weights[i])\n# preds = preds / weights_sum\n# bs, _ = preds.shape\n# # bs, n_cond, 5_level, 3\n# preds = preds.reshape(bs, -1, 5, 3).cpu()\n# sag_t2_preds.append(preds)\n# for b in range(bs):\n# study_id = int(study_id_list[b])\n# all_preds_t2[study_id] = preds[b]\n# sag_t2_preds = torch.cat(sag_t2_preds, dim=0)\n# # return all_preds_t2\n\n# # axial\n# model_cfgs = [\n# {\n# 'sub_dir': 'v24_cond_axial/rexnetr_200.sw_in12k_ft_in1k_z_imgs_5_666',\n# 'backbone': 'rexnetr_200.sw_in12k_ft_in1k',\n# 'axial_in_channels': 3,\n# 'weight': 0.44292661444289017,\n# 'z_imgs': 5,\n# 'data_key': 'axial_imgs_5_128_128'\n# },\n# {\n# 'sub_dir': 'v24_cond_axial/convnext_tiny.in12k_ft_in1k_384_z_imgs_3_666',\n# 'backbone': 'convnext_tiny.in12k_ft_in1k',\n# 'axial_in_channels': 1,\n# 'weight': 0.29758735999956626,\n# 'z_imgs': 3,\n# 'data_key': 'axial_imgs_3_128_128'\n# },\n\n# {\n# 'sub_dir': 'v24_cond_axial/convnext_small.in12k_ft_in1k_384_z_imgs_3_666',\n# 'backbone': 'convnext_small.in12k_ft_in1k_384',\n# 'axial_in_channels': 1,\n# 'weight': 0.2384201128638661,\n# 'z_imgs': 3,\n# 'data_key': 'axial_imgs_3_128_128'\n# },\n# {\n# 'sub_dir': 'v24_cond_axial/densenet161_z_imgs_3_666',\n# 'backbone': 'densenet161',\n# 'axial_in_channels': 1,\n# 'weight': 0.021065912693677462,\n# 'z_imgs': 3,\n# 'data_key': 'axial_imgs_3_128_128'\n# },\n\n# ]\n# v24_folds = list(range(5))\n# if folds is not None:\n# v24_folds = folds\n\n# models = []\n# data_keys = []\n# weights = []\n# for cfg in model_cfgs:\n# sub_dir = cfg['sub_dir']\n# backbone = cfg['backbone']\n# data_key = cfg['data_key']\n# axial_in_channels = cfg['axial_in_channels']\n# w = cfg['weight']\n# for fold in v24_folds:\n# fn = os.path.join(model_dir, sub_dir)\n# fn = fn + f'/best_wll_model_fold-{fold}_ema.pt'\n# print('load: ', fn)\n# model = Axial_HybridModel_24(backbone,\n# backbone,\n# pretrained=False,\n# axial_in_channels=axial_in_channels\n# )\n# model.load_state_dict(torch.load(fn))\n# model.to(device)\n# model.eval()\n# if is_parallel:\n# model = nn.DataParallel(model)\n\n# models.append(model)\n# weights.append(w)\n# data_keys.append(data_key)\n# weights_sum = np.sum(weights)\n\n# with open(f'{keypoint_dir}/axial_3d_keypoints_v24.pkl', 'rb') as f:\n# axial_3d_keypoints_v24 = pickle.load(f)\n\n# other_crop_size_list = [\n# (5, 128, 128),\n# ]\n\n# dset = Axial_Cond_Dataset_Multi_V24(data_root,\n# study_ids,\n# axial_3d_keypoints_v24,\n# pred_sag_keypoints_infos_3d_t2,\n# test_series_descriptions_fn=test_series_descriptions_fn,\n# image_dir=image_dir,\n# z_imgs=3,\n# other_z_imgs_list=other_crop_size_list,\n# cache_dir=CACHE_DIR,\n# )\n\n# dloader = DataLoader(\n# dset,\n# batch_size=batch_size,\n# shuffle=False,\n# pin_memory=False,\n# drop_last=False,\n# num_workers=num_workers,\n# )\n# all_preds_axial = {}\n# axial_preds = []\n# for tensor_d in tqdm.tqdm(dloader):\n# with torch.no_grad():\n# for k in tensor_d.keys():\n# if k not in ['study_id']:\n# tensor_d[k] = tensor_d[k].to(device)\n# study_id_list = tensor_d['study_id']\n# preds = None\n# with autocast:\n# for i in range(len(models)):\n# p = models[i](tensor_d['sag_t2'], tensor_d[data_keys[i]])\n# if preds is None:\n# preds = (p.float() * weights[i])\n# else:\n# preds += (p.float() * weights[i])\n# preds = preds / weights_sum\n# bs, _ = preds.shape\n# # bs, n_cond, 5_level, 3\n# preds = preds.reshape(bs, -1, 5, 3).cpu()\n# axial_preds.append(preds)\n# for b in range(bs):\n# study_id = int(study_id_list[b])\n# all_preds_axial[study_id] = preds[b]\n\n# axial_preds = torch.cat(axial_preds, dim=0)\n# all_preds = torch.cat((sag_t2_preds, sag_t1_preds, axial_preds), dim=1)\n# all_preds_dict = {}\n# for i, study_id in enumerate(study_ids):\n# all_preds_dict[study_id] = all_preds[i]\n# return all_preds_dict","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:39:28.120566Z","iopub.execute_input":"2024-10-08T16:39:28.120902Z","iopub.status.idle":"2024-10-08T16:39:28.145549Z","shell.execute_reply.started":"2024-10-08T16:39:28.120873Z","shell.execute_reply":"2024-10-08T16:39:28.144583Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def infer_all_666(data_root,\n model_dir,\n ):\n df = pd.read_csv(f'{data_root}/{test_series_descriptions_fn}')\n study_ids = df['study_id'].unique().tolist()\n \n device_0 = torch.device('cuda:0')\n\n is_parallel = torch.cuda.device_count() > 1 and len(study_ids) >=2\n print('is_parallel: ', is_parallel)\n folds = None\n \n pred_dict_v24 = infer_v2_seed666(data_root, study_ids, model_dir,\n save_dir, device_0, N_WORKERS, folds, is_parallel)\n\n\n pred_dict_v2 = infer_v24_seed666(data_root, study_ids, model_dir,\n save_dir, device_0, N_WORKERS, folds, is_parallel)\n\n \n print(len(study_ids))\n fold_preds = []\n for study_id in study_ids:\n pred = pred_dict_v2[study_id]\n fold_preds.append(pred.unsqueeze(0))\n fold_preds = torch.cat(fold_preds, dim=0)\n \n print('fold_preds shape: ', fold_preds.shape)\n\n fold_preds_v24 = []\n for study_id in study_ids:\n pred = pred_dict_v24[study_id]\n fold_preds_v24.append(pred.unsqueeze(0))\n fold_preds_v24 = torch.cat(fold_preds_v24, dim=0)\n\n fold_preds_en = fold_preds\n fold_preds_en = fold_preds\n fold_preds_en[:, 0:1] = 0.55 * fold_preds[:, 0:1] + 0.45 * fold_preds_v24[:, 0:1]\n fold_preds_en[:, 1:3] = 0.1 * fold_preds[:, 1:3] + 0.9 * fold_preds_v24[:, 1:3]\n fold_preds_en[:, 3:5] = 0.38 * fold_preds[:, 3:5] + 0.62 * fold_preds_v24[:, 3:5]\n\n\n fold_preds_en = nn.Softmax(dim=-1)(fold_preds_en).cpu().numpy()\n return fold_preds_en","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:39:28.146783Z","iopub.execute_input":"2024-10-08T16:39:28.147056Z","iopub.status.idle":"2024-10-08T16:39:28.158128Z","shell.execute_reply.started":"2024-10-08T16:39:28.147034Z","shell.execute_reply":"2024-10-08T16:39:28.157356Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#lhwcv_preds_666 = infer_all_666(data_root, model_dir2).reshape(-1, 3)\nlhwcv_preds = infer_all(data_root, model_dir).reshape(-1, 3)\n#lhwcv_preds = lhwcv_preds * 0.55 + lhwcv_preds_666 * 0.45","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:39:28.159023Z","iopub.execute_input":"2024-10-08T16:39:28.159297Z","iopub.status.idle":"2024-10-08T16:46:15.051493Z","shell.execute_reply.started":"2024-10-08T16:39:28.159274Z","shell.execute_reply":"2024-10-08T16:46:15.050408Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!rm -r ./cache/\n!rm -r ./keypoints_pred/","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:46:15.053566Z","iopub.execute_input":"2024-10-08T16:46:15.053990Z","iopub.status.idle":"2024-10-08T16:46:17.206919Z","shell.execute_reply.started":"2024-10-08T16:46:15.053949Z","shell.execute_reply":"2024-10-08T16:46:17.205472Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 3. make submission","metadata":{}},{"cell_type":"code","source":"sub = pd.DataFrame()\nsub['row_id'] = row_names\nsub[LABELS] = lhwcv_preds\n\nsub.to_csv('submission.csv', index=False)\nsub.head(25)","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:46:17.208712Z","iopub.execute_input":"2024-10-08T16:46:17.209045Z","iopub.status.idle":"2024-10-08T16:46:17.241815Z","shell.execute_reply.started":"2024-10-08T16:46:17.209011Z","shell.execute_reply":"2024-10-08T16:46:17.240834Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"previous_submit_df = sub\nprevious_submit_df = previous_submit_df.set_index('row_id')\nprint('previous_submit_df', previous_submit_df.shape)\nprint(previous_submit_df.dtypes)\nprint('')\n###########################################\n# heng submission\n\ntry: \n import natsort \nexcept:\n #%pip install natsort\n !pip install natsort --no-index --find-links=file://///kaggle/input/heng-rnas2024-final-01/\n \nimport sys\nimport os \nsys.path.append('/kaggle/input/heng-rnas2024-final-01')\n\nfrom data import *\nimport pandas as pd\nfrom natsort import natsorted\nimport matplotlib\nimport matplotlib.pyplot as plt\n\npd.set_option('display.width', 1000)\n\nprint('IMPORT OK !!!!!!!!!!!!!')","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:46:17.243174Z","iopub.execute_input":"2024-10-08T16:46:17.243534Z","iopub.status.idle":"2024-10-08T16:46:30.567097Z","shell.execute_reply.started":"2024-10-08T16:46:17.243500Z","shell.execute_reply":"2024-10-08T16:46:30.566063Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"KAGGLE_DATA_DIR =\\\n '/kaggle/input/rsna-2024-lumbar-spine-degenerative-classification'\n # '/home/hp/work/2024/kaggle/rsna2024-lumbar-spine/data/kaggle/rsna-2024-lumbar-spine-degenerative-classification'\n \n\nMODE = 'submit' # local #submit\nDEVICE = 'cuda' # 'cpu' 'cuda'\n\n\n\nif MODE == 'local':\n IMAGE_DIR = f'{KAGGLE_DATA_DIR}/train_images'\n valid_df = pd.read_csv(f'{KAGGLE_DATA_DIR}/train_series_descriptions.csv')\n\nif MODE == 'submit':\n IMAGE_DIR = f'{KAGGLE_DATA_DIR}/test_images'\n valid_df = pd.read_csv(f'{KAGGLE_DATA_DIR}/test_series_descriptions.csv')\n\nWEIGHT_DIR =\\\n '/kaggle/input/heng-rnas2024-final-01'\n #'/home/hp/work/2024/kaggle/rsna2024-lumbar-spine/code/team-share-002/final-01/010/scs_model_weight'\n\nscs_cfg = dotdict(\n image_size=320,\n arch='pvt_v2_b4',\n checkpoint=[\n f'{WEIGHT_DIR}/scs_model_weight/fold0-00008857-00010420-swa.pth',\n f'{WEIGHT_DIR}/scs_model_weight/fold1-00008416-00009994-swa.pth',\n f'{WEIGHT_DIR}/scs_model_weight/fold2-00012259-00013858-swa.pth',\n f'{WEIGHT_DIR}/scs_model_weight/fold3-00009486-00014229-swa.pth',\n f'{WEIGHT_DIR}/scs_model_weight/fold4-00006276-00012029-swa.pth',\n ],\n\n IMAGE_DIR= IMAGE_DIR,\n MODE = MODE,\n DEVICE = DEVICE,\n)\nnfn_cfg = dotdict(\n image_size=320,\n arch='pvt_v2_b4',\n checkpoint=[\n f'{WEIGHT_DIR}/nfn_model_weight/fold0-00032592.fix-name.pth',\n f'{WEIGHT_DIR}/nfn_model_weight/fold1-00032144.pth',\n f'{WEIGHT_DIR}/nfn_model_weight/fold2-00036524.pth',\n f'{WEIGHT_DIR}/nfn_model_weight/fold3-00032619.pth',\n f'{WEIGHT_DIR}/nfn_model_weight/fold4-00029488.pth',\n f'{WEIGHT_DIR}/fold2-fix-flip-aug-00015086.pth',\n f'{WEIGHT_DIR}/fold3-fix-flip-aug-000031047.pth',\n ],\n\n IMAGE_DIR= IMAGE_DIR,\n MODE = MODE,\n DEVICE = DEVICE,\n)\nprint(valid_df)\nprint('SETTING OK !!!!!!!!!!!!!')","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:46:30.568785Z","iopub.execute_input":"2024-10-08T16:46:30.569822Z","iopub.status.idle":"2024-10-08T16:46:30.589368Z","shell.execute_reply.started":"2024-10-08T16:46:30.569774Z","shell.execute_reply":"2024-10-08T16:46:30.588577Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"from run_scs_submit import *\nfrom run_nfn_submit import *\n\n#scs_df = run_scs(valid_df, scs_cfg)\n#scs_result = run_scs(valid_df, scs_cfg)\n#scs_df = make_scs_submit(scs_result)\n\nnfn_result = run_nfn(valid_df, nfn_cfg)\nnfn_df = make_nfn_submit(nfn_result)\n\n#----\n# example to merge scs_df to you your submit csv\n\n#submit_df = make_dummy_submit(valid_df) #your submission csv\nsubmit_df = previous_submit_df\n\n#merge\n#submit_df.loc[scs_df.index, grade_col] = \\\n# 0.8*submit_df.loc[scs_df.index, grade_col]+ 0.2*scs_df\nsubmit_df.loc[nfn_df.index, grade_col] = \\\n 0.75*submit_df.loc[nfn_df.index, grade_col]+ 0.25*nfn_df\n\n\nsubmit_df = submit_df.reset_index(drop=False)\nsubmit_df.to_csv('submission.csv', index=False)\n\nprint('** FINAL SUBMIT **')\nprint(submit_df.head(30))\nprint(submit_df.shape)\n\nprint('SUBMIT OK!!!!!')","metadata":{"execution":{"iopub.status.busy":"2024-10-08T16:46:30.590349Z","iopub.execute_input":"2024-10-08T16:46:30.590674Z","iopub.status.idle":"2024-10-08T16:46:55.614354Z","shell.execute_reply.started":"2024-10-08T16:46:30.590620Z","shell.execute_reply":"2024-10-08T16:46:55.613427Z"},"trusted":true},"execution_count":null,"outputs":[]}]}