import subprocess subprocess.run(["pip", "install", "../input/pycocotools202/pycocotools-2.0.2-cp37-cp37m-linux_x86_64.whl"]) subprocess.run(["tar", "-xvf", "../input/gdcminstall/gdcm.tar"]) subprocess.run(["conda", "install", "../working/gdcm/conda-4.8.4-py37hc8dfbb8_2.tar.bz2"]) subprocess.run(["conda", "install", "../working/gdcm/gdcm-2.8.9-py37h71b2a6d_0.tar.bz2"]) subprocess.run(["conda", "install", "../working/gdcm/libjpeg-turbo-2.0.3-h516909a_1.tar.bz2"]) def get_none_results(image_id_list_matched_midrc, midrc_dict): import sys sys.path.insert(0, '../input/covidtimm/') import numpy as np import pandas as pd import os import cv2 from tqdm import tqdm import torch.nn as nn from torch import optim import torch.nn.functional as F from torchvision import transforms from torch.utils.data import Dataset, DataLoader import torch import timm import pickle import time from torch.cuda.amp import autocast, GradScaler import glob import pydicom from pydicom.pixel_data_handlers.util import apply_voi_lut # https://www.kaggle.com/raddar/convert-dicom-to-np-array-the-correct-way def read_xray(path, voi_lut = True, fix_monochrome = True): dicom = pydicom.read_file(path) # VOI LUT (if available by DICOM device) is used to transform raw DICOM data to "human-friendly" view if voi_lut: data = apply_voi_lut(dicom.pixel_array, dicom) else: data = dicom.pixel_array # depending on this value, X-ray may look inverted - fix that: if fix_monochrome and dicom.PhotometricInterpretation == "MONOCHROME1": data = np.amax(data) - data data = data - np.min(data) data = data / np.max(data) data = (data * 255).astype(np.uint8) return data class TestDataset(Dataset): def __init__(self, label_dict, image_list): self.label_dict=label_dict self.image_list=image_list def __len__(self): return len(self.image_list) def __getitem__(self, index): image_id = self.image_list[index] img = read_xray(self.label_dict[image_id]['img_dir']) img = cv2.resize(img, (1280, 1280)) image = np.zeros((1280,1280,3), dtype=np.uint8) image[:,:,0] = img image[:,:,1] = img image[:,:,2] = img image672 = cv2.resize(image, (672, 672)) image608 = cv2.resize(image, (608, 608)) image672 = image672.transpose(2, 0, 1) image608 = image608.transpose(2, 0, 1) return image672, image608 class TestDataset1(Dataset): def __init__(self, label_dict, image_list): self.label_dict=label_dict self.image_list=image_list def __len__(self): return len(self.image_list) def __getitem__(self, index): image_id = self.image_list[index] image = cv2.imread(self.label_dict[image_id]['img_dir']) image = cv2.resize(image, (1280, 1280)) image672 = cv2.resize(image, (672, 672)) image608 = cv2.resize(image, (608, 608)) image672 = image672.transpose(2, 0, 1) image608 = image608.transpose(2, 0, 1) return image672, image608 class NoneEffb7(nn.Module): def __init__(self, ): super().__init__() self.net = timm.create_model('tf_efficientnet_b7_ns', features_only=True, out_indices=(3, 4), drop_path_rate=0.5, pretrained=False) self.avg_pool = nn.AdaptiveAvgPool2d(1) in_features = 640 self.last_linear = nn.Linear(in_features, 1) self.conv_mask = nn.Conv2d(224, 1, kernel_size=1, stride=1, bias=True) def forward(self, x): x1, x = self.net(x) x_mask = self.conv_mask(x1) x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.last_linear(x) return x, x_mask class NoneEffb8(nn.Module): def __init__(self, ): super().__init__() self.net = timm.create_model('tf_efficientnet_b8', features_only=True, out_indices=(3, 4), drop_path_rate=0.5, pretrained=False) self.avg_pool = nn.AdaptiveAvgPool2d(1) in_features = 704 self.last_linear = nn.Linear(in_features, 1) self.conv_mask = nn.Conv2d(248, 1, kernel_size=1, stride=1, bias=True) def forward(self, x): x1, x = self.net(x) x_mask = self.conv_mask(x1) x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.last_linear(x) return x, x_mask start_time = time.time() # prepare input image_list = sorted(glob.glob('../input/siim-covid19-detection/test/*/*/*.dcm')) study_id_list = [] image_id_list = [] studyid2imageid = {} imageid2studyid = {} label_dict = {} for i in range(len(image_list)): study_id = image_list[i].split('/')[-3] image_id = image_list[i].split('/')[-1][:-4] image_id_list.append(image_id) if study_id not in studyid2imageid: study_id_list.append(study_id) studyid2imageid[study_id] = [image_id] else: studyid2imageid[study_id].append(image_id) imageid2studyid[image_id] = study_id label_dict[image_id] = { 'img_dir': image_list[i], } print(len(image_id_list), len(study_id_list), len(studyid2imageid), len(imageid2studyid)) image_id_list111 = [] for image_id in image_id_list: if image_id not in image_id_list_matched_midrc: image_id_list111.append(image_id) image_id_list = image_id_list111 print(len(image_id_list)) # hyperparameters batch_size = 4 # build model model11_fold0 = NoneEffb8() model11_fold0.load_state_dict(torch.load('../input/covidpretrained4/model11_fold0/epoch10_polyak')) model11_fold0.cuda() model11_fold0.eval() model11_fold1 = NoneEffb8() model11_fold1.load_state_dict(torch.load('../input/covidpretrained4/model11_fold1/epoch10_polyak')) model11_fold1.cuda() model11_fold1.eval() model11_fold2 = NoneEffb8() model11_fold2.load_state_dict(torch.load('../input/covidpretrained4/model11_fold2/epoch10_polyak')) model11_fold2.cuda() model11_fold2.eval() model11_fold3 = NoneEffb8() model11_fold3.load_state_dict(torch.load('../input/covidpretrained4/model11_fold3/epoch10_polyak')) model11_fold3.cuda() model11_fold3.eval() model11_fold4 = NoneEffb8() model11_fold4.load_state_dict(torch.load('../input/covidpretrained4/model11_fold4/epoch10_polyak')) model11_fold4.cuda() model11_fold4.eval() model12_fold0 = NoneEffb7() model12_fold0.load_state_dict(torch.load('../input/covidpretrained4/model12_fold0/epoch6_polyak')) model12_fold0.cuda() model12_fold0.eval() model12_fold1 = NoneEffb7() model12_fold1.load_state_dict(torch.load('../input/covidpretrained4/model12_fold1/epoch6_polyak')) model12_fold1.cuda() model12_fold1.eval() model12_fold2 = NoneEffb7() model12_fold2.load_state_dict(torch.load('../input/covidpretrained4/model12_fold2/epoch6_polyak')) model12_fold2.cuda() model12_fold2.eval() model12_fold3 = NoneEffb7() model12_fold3.load_state_dict(torch.load('../input/covidpretrained4/model12_fold3/epoch6_polyak')) model12_fold3.cuda() model12_fold3.eval() model12_fold4 = NoneEffb7() model12_fold4.load_state_dict(torch.load('../input/covidpretrained4/model12_fold4/epoch6_polyak')) model12_fold4.cuda() model12_fold4.eval() # iterator for testing datagen = TestDataset(label_dict=label_dict, image_list=image_id_list) generator = DataLoader(dataset=datagen, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=False) pred_prob = np.zeros((len(image_id_list), ), dtype=np.float32) for i, (images672, images608) in enumerate(generator): with torch.no_grad(): start = i*batch_size end = start+batch_size if i == len(generator)-1: end = len(generator.dataset) images672 = images672.cuda().float() / 255.0 images608 = images608.cuda().float() / 255.0 with autocast(): logits_model11_fold0, _ = model11_fold0(images608) logits_model11_fold1, _ = model11_fold1(images608) logits_model11_fold2, _ = model11_fold2(images608) logits_model11_fold3, _ = model11_fold3(images608) logits_model11_fold4, _ = model11_fold4(images608) logits_flip_model11_fold0, _ = model11_fold0(images608.flip(3)) logits_flip_model11_fold1, _ = model11_fold1(images608.flip(3)) logits_flip_model11_fold2, _ = model11_fold2(images608.flip(3)) logits_flip_model11_fold3, _ = model11_fold3(images608.flip(3)) logits_flip_model11_fold4, _ = model11_fold4(images608.flip(3)) logits_model12_fold0, _ = model12_fold0(images672) logits_model12_fold1, _ = model12_fold1(images672) logits_model12_fold2, _ = model12_fold2(images672) logits_model12_fold3, _ = model12_fold3(images672) logits_model12_fold4, _ = model12_fold4(images672) logits_flip_model12_fold0, _ = model12_fold0(images672.flip(3)) logits_flip_model12_fold1, _ = model12_fold1(images672.flip(3)) logits_flip_model12_fold2, _ = model12_fold2(images672.flip(3)) logits_flip_model12_fold3, _ = model12_fold3(images672.flip(3)) logits_flip_model12_fold4, _ = model12_fold4(images672.flip(3)) pred_prob[start:end] += logits_model11_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob /= 20.0 # generate submission image_level_dict = {} for image_id in image_id_list: image_level_dict[image_id] = {"opacity": [], "none": 0., "negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(image_id_list)): image_level_dict[image_id_list[i]]["none"] = pred_prob[i] ### re-prediction ### # iterator for testing datagen = TestDataset1(label_dict=midrc_dict, image_list=image_id_list_matched_midrc) generator = DataLoader(dataset=datagen, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=False) pred_prob = np.zeros((len(image_id_list_matched_midrc), ), dtype=np.float32) for i, (images672, images608) in enumerate(generator): with torch.no_grad(): start = i*batch_size end = start+batch_size if i == len(generator)-1: end = len(generator.dataset) images672 = images672.cuda().float() / 255.0 images608 = images608.cuda().float() / 255.0 with autocast(): logits_model11_fold0, _ = model11_fold0(images608) logits_model11_fold1, _ = model11_fold1(images608) logits_model11_fold2, _ = model11_fold2(images608) logits_model11_fold3, _ = model11_fold3(images608) logits_model11_fold4, _ = model11_fold4(images608) logits_flip_model11_fold0, _ = model11_fold0(images608.flip(3)) logits_flip_model11_fold1, _ = model11_fold1(images608.flip(3)) logits_flip_model11_fold2, _ = model11_fold2(images608.flip(3)) logits_flip_model11_fold3, _ = model11_fold3(images608.flip(3)) logits_flip_model11_fold4, _ = model11_fold4(images608.flip(3)) logits_model12_fold0, _ = model12_fold0(images672) logits_model12_fold1, _ = model12_fold1(images672) logits_model12_fold2, _ = model12_fold2(images672) logits_model12_fold3, _ = model12_fold3(images672) logits_model12_fold4, _ = model12_fold4(images672) logits_flip_model12_fold0, _ = model12_fold0(images672.flip(3)) logits_flip_model12_fold1, _ = model12_fold1(images672.flip(3)) logits_flip_model12_fold2, _ = model12_fold2(images672.flip(3)) logits_flip_model12_fold3, _ = model12_fold3(images672.flip(3)) logits_flip_model12_fold4, _ = model12_fold4(images672.flip(3)) pred_prob[start:end] += logits_model11_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model11_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model11_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_model12_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold0.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold1.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold2.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold3.sigmoid().cpu().data.numpy().squeeze() pred_prob[start:end] += logits_flip_model12_fold4.sigmoid().cpu().data.numpy().squeeze() pred_prob /= 20.0 for image_id in image_id_list_matched_midrc: image_level_dict[image_id] = {"opacity": [], "none": 0., "negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(image_id_list_matched_midrc)): image_level_dict[image_id_list_matched_midrc[i]]["none"] = pred_prob[i] end_time = time.time() print(end_time-start_time) return image_level_dict def get_study_level_results(): import sys sys.path.insert(0, '../input/covidtimm/') import numpy as np import pandas as pd import os import cv2 from tqdm import tqdm import torch.nn as nn from torch import optim import torch.nn.functional as F from torchvision import transforms from torch.utils.data import Dataset, DataLoader import torch import timm import pickle import time from torch.cuda.amp import autocast, GradScaler import glob import pydicom from pydicom.pixel_data_handlers.util import apply_voi_lut # https://www.kaggle.com/raddar/convert-dicom-to-np-array-the-correct-way def read_xray(path, voi_lut = True, fix_monochrome = True): dicom = pydicom.read_file(path) # VOI LUT (if available by DICOM device) is used to transform raw DICOM data to "human-friendly" view if voi_lut: data = apply_voi_lut(dicom.pixel_array, dicom) else: data = dicom.pixel_array # depending on this value, X-ray may look inverted - fix that: if fix_monochrome and dicom.PhotometricInterpretation == "MONOCHROME1": data = np.amax(data) - data data = data - np.min(data) data = data / np.max(data) data = (data * 255).astype(np.uint8) return data class TestDataset(Dataset): def __init__(self, label_dict, image_list): self.label_dict=label_dict self.image_list=image_list def __len__(self): return len(self.image_list) def __getitem__(self, index): image_id = self.image_list[index] img = read_xray(self.label_dict[image_id]['img_dir']) img = cv2.resize(img, (1280, 1280)) image = np.zeros((1280,1280,3), dtype=np.uint8) image[:,:,0] = img image[:,:,1] = img image[:,:,2] = img image672 = cv2.resize(image, (672, 672)) image608 = cv2.resize(image, (608, 608)) image512 = cv2.resize(image, (512, 512)) image672 = image672.transpose(2, 0, 1) image608 = image608.transpose(2, 0, 1) image512 = image512.transpose(2, 0, 1) return image672, image608, image512 class TestDataset1(Dataset): def __init__(self, label_dict, image_list): self.label_dict=label_dict self.image_list=image_list def __len__(self): return len(self.image_list) def __getitem__(self, index): image_id = self.image_list[index] image = cv2.imread(self.label_dict[image_id]['img_dir']) image = cv2.resize(image, (1280, 1280)) image672 = cv2.resize(image, (672, 672)) image608 = cv2.resize(image, (608, 608)) image672 = image672.transpose(2, 0, 1) image608 = image608.transpose(2, 0, 1) return image672, image608 class StudyLevelEffb7(nn.Module): def __init__(self, ): super().__init__() self.net = timm.create_model('tf_efficientnet_b7_ns', features_only=True, out_indices=(3, 4), drop_path_rate=0.5, pretrained=False) self.avg_pool = nn.AdaptiveAvgPool2d(1) in_features = 640 self.last_linear = nn.Linear(in_features, 4) self.conv_mask = nn.Conv2d(224, 1, kernel_size=1, stride=1, bias=True) def forward(self, x): x1, x = self.net(x) x_mask = self.conv_mask(x1) x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.last_linear(x) return x, x_mask class StudyLevelEffb8(nn.Module): def __init__(self, ): super().__init__() self.net = timm.create_model('tf_efficientnet_b8', features_only=True, out_indices=(3, 4), drop_path_rate=0.5, pretrained=False) self.avg_pool = nn.AdaptiveAvgPool2d(1) in_features = 704 self.last_linear = nn.Linear(in_features, 4) self.conv_mask = nn.Conv2d(248, 1, kernel_size=1, stride=1, bias=True) def forward(self, x): x1, x = self.net(x) x_mask = self.conv_mask(x1) x = self.avg_pool(x) x = x.view(x.size(0), -1) x = self.last_linear(x) return x, x_mask class XrayNet(nn.Module): def __init__(self, ): super().__init__() self.net = timm.create_model('tf_efficientnet_b3_ns', drop_path_rate=0.3, pretrained=False) self.avg_pool = nn.AdaptiveAvgPool2d(1) def forward(self, x): x = self.net.forward_features(x) x = self.avg_pool(x) x = x.view(x.size(0), -1) return x start_time = time.time() # prepare input image_list = sorted(glob.glob('../input/siim-covid19-detection/test/*/*/*.dcm')) study_id_list = [] image_id_list = [] studyid2imageid = {} imageid2studyid = {} label_dict = {} for i in range(len(image_list)): study_id = image_list[i].split('/')[-3] image_id = image_list[i].split('/')[-1][:-4] image_id_list.append(image_id) if study_id not in studyid2imageid: study_id_list.append(study_id) studyid2imageid[study_id] = [image_id] else: studyid2imageid[study_id].append(image_id) imageid2studyid[image_id] = study_id label_dict[image_id] = { 'img_dir': image_list[i], } print(len(image_id_list), len(study_id_list), len(studyid2imageid), len(imageid2studyid)) # hyperparameters batch_size = 4 # build model model1_fold0 = StudyLevelEffb8() model1_fold0.load_state_dict(torch.load('../input/covidpretrained3/model1_fold0/epoch8_polyak')) model1_fold0.cuda() model1_fold0.eval() model1_fold1 = StudyLevelEffb8() model1_fold1.load_state_dict(torch.load('../input/covidpretrained3/model1_fold1/epoch8_polyak')) model1_fold1.cuda() model1_fold1.eval() model1_fold2 = StudyLevelEffb8() model1_fold2.load_state_dict(torch.load('../input/covidpretrained3/model1_fold2/epoch8_polyak')) model1_fold2.cuda() model1_fold2.eval() model1_fold3 = StudyLevelEffb8() model1_fold3.load_state_dict(torch.load('../input/covidpretrained3/model1_fold3/epoch8_polyak')) model1_fold3.cuda() model1_fold3.eval() model1_fold4 = StudyLevelEffb8() model1_fold4.load_state_dict(torch.load('../input/covidpretrained3/model1_fold4/epoch8_polyak')) model1_fold4.cuda() model1_fold4.eval() model2_fold0 = StudyLevelEffb7() model2_fold0.load_state_dict(torch.load('../input/covidpretrained3/model2_fold0/epoch9_polyak')) model2_fold0.cuda() model2_fold0.eval() model2_fold1 = StudyLevelEffb7() model2_fold1.load_state_dict(torch.load('../input/covidpretrained3/model2_fold1/epoch9_polyak')) model2_fold1.cuda() model2_fold1.eval() model2_fold2 = StudyLevelEffb7() model2_fold2.load_state_dict(torch.load('../input/covidpretrained3/model2_fold2/epoch9_polyak')) model2_fold2.cuda() model2_fold2.eval() model2_fold3 = StudyLevelEffb7() model2_fold3.load_state_dict(torch.load('../input/covidpretrained3/model2_fold3/epoch9_polyak')) model2_fold3.cuda() model2_fold3.eval() model2_fold4 = StudyLevelEffb7() model2_fold4.load_state_dict(torch.load('../input/covidpretrained3/model2_fold4/epoch9_polyak')) model2_fold4.cuda() model2_fold4.eval() model_xray = XrayNet() model_xray.load_state_dict(torch.load('../input/covidnearestneighbors/epoch79')) model_xray = model_xray.cuda() model_xray.eval() # iterator for testing datagen = TestDataset(label_dict=label_dict, image_list=image_id_list) generator = DataLoader(dataset=datagen, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=False) pred_prob = np.zeros((len(image_id_list), 4), dtype=np.float32) valid_feature = np.zeros((len(image_id_list), 1536), dtype=np.float32) for i, (images672, images608, images512) in enumerate(generator): with torch.no_grad(): start = i*batch_size end = start+batch_size if i == len(generator)-1: end = len(generator.dataset) images672 = images672.cuda().float() / 255.0 images608 = images608.cuda().float() / 255.0 images512 = images512.cuda().float() / 255.0 with autocast(): logits_model1_fold0, _ = model1_fold0(images608) logits_model1_fold1, _ = model1_fold1(images608) logits_model1_fold2, _ = model1_fold2(images608) logits_model1_fold3, _ = model1_fold3(images608) logits_model1_fold4, _ = model1_fold4(images608) logits_flip_model1_fold0, _ = model1_fold0(images608.flip(3)) logits_flip_model1_fold1, _ = model1_fold1(images608.flip(3)) logits_flip_model1_fold2, _ = model1_fold2(images608.flip(3)) logits_flip_model1_fold3, _ = model1_fold3(images608.flip(3)) logits_flip_model1_fold4, _ = model1_fold4(images608.flip(3)) logits_model2_fold0, _ = model2_fold0(images672) logits_model2_fold1, _ = model2_fold1(images672) logits_model2_fold2, _ = model2_fold2(images672) logits_model2_fold3, _ = model2_fold3(images672) logits_model2_fold4, _ = model2_fold4(images672) logits_flip_model2_fold0, _ = model2_fold0(images672.flip(3)) logits_flip_model2_fold1, _ = model2_fold1(images672.flip(3)) logits_flip_model2_fold2, _ = model2_fold2(images672.flip(3)) logits_flip_model2_fold3, _ = model2_fold3(images672.flip(3)) logits_flip_model2_fold4, _ = model2_fold4(images672.flip(3)) features = model_xray(images512) pred_prob[start:end] += logits_model1_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold4.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold4.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold4.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold4.sigmoid().cpu().data.numpy() valid_feature[start:end] = F.normalize(features).cpu().data.numpy() pred_prob /= 20.0 # generate submission image_level_dict = {} for image_id in image_id_list: image_level_dict[image_id] = {"opacity": [], "none": 0., "negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(image_id_list)): image_level_dict[image_id_list[i]]["negative"] = pred_prob[i][0] image_level_dict[image_id_list[i]]["typical"] = pred_prob[i][1] image_level_dict[image_id_list[i]]["indeterminate"] = pred_prob[i][2] image_level_dict[image_id_list[i]]["atypical"] = pred_prob[i][3] ### from sklearn.neighbors import NearestNeighbors train_feature_bimcv = np.load('../input/covidnearestneighbors/ext_features_bimcv.npy') image_id_list = np.array(image_id_list) neigh = NearestNeighbors(n_neighbors=1, metric='cosine', n_jobs=-1) neigh.fit(train_feature_bimcv) pred_dist, pred_label = np.squeeze(neigh.kneighbors(X=valid_feature, n_neighbors=1, return_distance=True)) pred_dist = np.squeeze(pred_dist) pred_label = np.squeeze(pred_label) num_unmatch_bimcv = 0 num_matched_bimcv = 0 idx_unmatch_bimcv = [] idx_matched_bimcv = [] for i in range(len(pred_dist)): if pred_dist[i] < 0.01: num_matched_bimcv += 1 idx_matched_bimcv.append(i) else: num_unmatch_bimcv += 1 idx_unmatch_bimcv.append(i) print(num_matched_bimcv, num_unmatch_bimcv) pred_dist_matched_bimcv = pred_dist[idx_matched_bimcv] pred_label_matched_bimcv = pred_label[idx_matched_bimcv] image_id_list_matched_bimcv = list(image_id_list[idx_matched_bimcv]) print(len(image_id_list_matched_bimcv)) bimcv_id_list = list(pd.read_csv('../input/covidnearestneighbors/ext_bimcv.csv')['id'].values) #bimcv_train_id_list = [] bimcv_dict = {} for i in range(len(image_id_list_matched_bimcv)): #bimcv_train_id_list.append(bimcv_id_list[int(pred_label_matched_bimcv[i])]) bimcv_dict[image_id_list_matched_bimcv[i]] = {'bimcv_id': bimcv_id_list[int(pred_label_matched_bimcv[i])]} valid_feature_unmatch_bimcv = valid_feature[idx_unmatch_bimcv] print(valid_feature_unmatch_bimcv.shape) train_feature_midrc = np.load('../input/covidnearestneighbors/ext_features_midrc.npy') neigh = NearestNeighbors(n_neighbors=1, metric='cosine', n_jobs=-1) neigh.fit(train_feature_midrc) pred_dist, pred_label = np.squeeze(neigh.kneighbors(X=valid_feature_unmatch_bimcv, n_neighbors=1, return_distance=True)) pred_dist = np.squeeze(pred_dist) pred_label = np.squeeze(pred_label) num_unmatch_midrc = 0 num_matched_midrc = 0 idx_unmatch_midrc = [] idx_matched_midrc = [] for i in range(len(pred_dist)): if pred_dist[i] < 0.5: num_matched_midrc += 1 idx_matched_midrc.append(i) else: num_unmatch_midrc += 1 idx_unmatch_midrc.append(i) print(num_matched_midrc, num_unmatch_midrc) pred_dist_matched_midrc = pred_dist[idx_matched_midrc] pred_label_matched_midrc = pred_label[idx_matched_midrc] image_id_list_matched_midrc = list(image_id_list[idx_unmatch_bimcv][idx_matched_midrc]) print(len(image_id_list_matched_midrc)) midrc_id_list = list(pd.read_csv('../input/covidnearestneighbors/ext_midrc.csv')['id'].values) #midrc_train_id_list = [] midrc_dict = {} for i in range(len(image_id_list_matched_midrc)): #midrc_train_id_list.append(midrc_id_list[int(pred_label_matched_midrc[i])]) midrc_dict[image_id_list_matched_midrc[i]] = {'img_dir': '../input/ricord-covid19-xray-positive-tests/MIDRC-RICORD/MIDRC-RICORD/' + midrc_id_list[int(pred_label_matched_midrc[i])], 'midrc_id': midrc_id_list[int(pred_label_matched_midrc[i])]} ### re-prediction ### # iterator for testing datagen = TestDataset1(label_dict=midrc_dict, image_list=image_id_list_matched_midrc) generator = DataLoader(dataset=datagen, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=False) pred_prob = np.zeros((len(image_id_list_matched_midrc), 4), dtype=np.float32) for i, (images672, images608) in enumerate(generator): with torch.no_grad(): start = i*batch_size end = start+batch_size if i == len(generator)-1: end = len(generator.dataset) images672 = images672.cuda().float() / 255.0 images608 = images608.cuda().float() / 255.0 with autocast(): logits_model1_fold0, _ = model1_fold0(images608) logits_model1_fold1, _ = model1_fold1(images608) logits_model1_fold2, _ = model1_fold2(images608) logits_model1_fold3, _ = model1_fold3(images608) logits_model1_fold4, _ = model1_fold4(images608) logits_flip_model1_fold0, _ = model1_fold0(images608.flip(3)) logits_flip_model1_fold1, _ = model1_fold1(images608.flip(3)) logits_flip_model1_fold2, _ = model1_fold2(images608.flip(3)) logits_flip_model1_fold3, _ = model1_fold3(images608.flip(3)) logits_flip_model1_fold4, _ = model1_fold4(images608.flip(3)) logits_model2_fold0, _ = model2_fold0(images672) logits_model2_fold1, _ = model2_fold1(images672) logits_model2_fold2, _ = model2_fold2(images672) logits_model2_fold3, _ = model2_fold3(images672) logits_model2_fold4, _ = model2_fold4(images672) logits_flip_model2_fold0, _ = model2_fold0(images672.flip(3)) logits_flip_model2_fold1, _ = model2_fold1(images672.flip(3)) logits_flip_model2_fold2, _ = model2_fold2(images672.flip(3)) logits_flip_model2_fold3, _ = model2_fold3(images672.flip(3)) logits_flip_model2_fold4, _ = model2_fold4(images672.flip(3)) pred_prob[start:end] += logits_model1_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model1_fold4.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model1_fold4.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_model2_fold4.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold0.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold1.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold2.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold3.sigmoid().cpu().data.numpy() pred_prob[start:end] += logits_flip_model2_fold4.sigmoid().cpu().data.numpy() pred_prob /= 20.0 for i in range(len(image_id_list_matched_midrc)): image_level_dict[image_id_list_matched_midrc[i]]["negative"] = pred_prob[i][0] image_level_dict[image_id_list_matched_midrc[i]]["typical"] = pred_prob[i][1] image_level_dict[image_id_list_matched_midrc[i]]["indeterminate"] = pred_prob[i][2] image_level_dict[image_id_list_matched_midrc[i]]["atypical"] = pred_prob[i][3] end_time = time.time() print(end_time-start_time) return image_level_dict, image_id_list_matched_midrc, midrc_dict, image_id_list_matched_bimcv, bimcv_dict def get_detection_results(image_id_list_matched_midrc, midrc_dict): import sys sys.path.insert(0, '../input/covidalbumentations/') sys.path.insert(0, '../input/covideffdet/') sys.path.insert(0, '../input/covidensembleboxes/') sys.path.insert(0, '../input/covidomegaconf/') sys.path.insert(0, '../input/covidtimm/') sys.path.insert(0, '../input/covidyolo/yolov5-5.0/') import numpy as np import pandas as pd import os import cv2 from tqdm import tqdm import torchvision import torch.nn as nn from torch.utils.data import Dataset, DataLoader import torch import pickle import albumentations from effdet import get_efficientdet_config, EfficientDet, DetBenchPredict, load_pretrained from effdet.data.parsers import CocoParserCfg, create_parser from ensemble_boxes import weighted_boxes_fusion import time from torch.cuda.amp import autocast, GradScaler import glob import pydicom from pydicom.pixel_data_handlers.util import apply_voi_lut from utils.general import xyxy2xywh, xywh2xyxy from utils.datasets import letterbox from utils.torch_utils import select_device from utils.general import scale_coords from models.experimental import attempt_load def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False, labels=()): """Runs Non-Maximum Suppression (NMS) on inference results Returns: list of detections, on (n,6) tensor per image [xyxy, conf, cls] """ nc = prediction.shape[2] - 5 # number of classes xc = prediction[..., 4] > conf_thres # candidates # Settings min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height max_det = 100 # maximum number of detections per image max_nms = 30000 # maximum number of boxes into torchvision.ops.nms() time_limit = 10.0 # seconds to quit after redundant = True # require redundant detections multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img) merge = False # use merge-NMS t = time.time() output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0] for xi, x in enumerate(prediction): # image index, image inference # Apply constraints # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height x = x[xc[xi]] # confidence # Cat apriori labels if autolabelling if labels and len(labels[xi]): l = labels[xi] v = torch.zeros((len(l), nc + 5), device=x.device) v[:, :4] = l[:, 1:5] # box v[:, 4] = 1.0 # conf v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls x = torch.cat((x, v), 0) # If none remain process next image if not x.shape[0]: continue # Compute conf x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf # Box (center x, center y, width, height) to (x1, y1, x2, y2) box = xywh2xyxy(x[:, :4]) # Detections matrix nx6 (xyxy, conf, cls) if multi_label: i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1) else: # best class only conf, j = x[:, 5:].max(1, keepdim=True) x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres] # Filter by class if classes is not None: x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)] # Apply finite constraint # if not torch.isfinite(x).all(): # x = x[torch.isfinite(x).all(1)] # Check shape n = x.shape[0] # number of boxes if not n: # no boxes continue elif n > max_nms: # excess boxes x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence # Batched NMS c = x[:, 5:6] * (0 if agnostic else max_wh) # classes boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS if i.shape[0] > max_det: # limit detections i = i[:max_det] if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean) # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4) iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix weights = iou * scores[None] # box weights x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes if redundant: i = i[iou.sum(1) > 1] # require redundancy output[xi] = x[i] if (time.time() - t) > time_limit: print(f'WARNING: NMS time limit {time_limit}s exceeded') break # time limit exceeded return output # https://www.kaggle.com/raddar/convert-dicom-to-np-array-the-correct-way def read_xray(path, voi_lut = True, fix_monochrome = True): dicom = pydicom.read_file(path) # VOI LUT (if available by DICOM device) is used to transform raw DICOM data to "human-friendly" view if voi_lut: data = apply_voi_lut(dicom.pixel_array, dicom) else: data = dicom.pixel_array # depending on this value, X-ray may look inverted - fix that: if fix_monochrome and dicom.PhotometricInterpretation == "MONOCHROME1": data = np.amax(data) - data data = data - np.min(data) data = data / np.max(data) data = (data * 255).astype(np.uint8) return data class TestDataset(Dataset): def __init__(self, label_dict, image_list): self.label_dict=label_dict self.image_list=image_list def __len__(self): return len(self.image_list) def __getitem__(self,index): image_id = self.image_list[index] img = read_xray(self.label_dict[image_id]['img_dir']) h, w = img.shape[0], img.shape[1] img = cv2.resize(img, (1280, 1280)) image = np.zeros((1280,1280,3), dtype=np.uint8) image[:,:,0] = img image[:,:,1] = img image[:,:,2] = img image1024 = cv2.resize(image, (1024, 1024)) image896 = cv2.resize(image, (896, 896)) imageyolo = cv2.resize(image, (1024, 1024)) imageyolo = imageyolo[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 imageyolo = np.ascontiguousarray(imageyolo) return image1024, image896, imageyolo, h, w class TestDataset1(Dataset): def __init__(self, label_dict, image_list): self.label_dict=label_dict self.image_list=image_list def __len__(self): return len(self.image_list) def __getitem__(self,index): image_id = self.image_list[index] image = cv2.imread(self.label_dict[image_id]['img_dir']) h, w = image.shape[0], image.shape[1] image = cv2.resize(image, (1280, 1280)) image1024 = cv2.resize(image, (1024, 1024)) image896 = cv2.resize(image, (896, 896)) imageyolo = cv2.resize(image, (1024, 1024)) imageyolo = imageyolo[:, :, ::-1].transpose(2, 0, 1) # BGR to RGB, to 3x416x416 imageyolo = np.ascontiguousarray(imageyolo) return image1024, image896, imageyolo, h, w def collate_fn(batch): image1024_list = [] image896_list = [] imageyolo_list = [] h_list = [] w_list = [] for i in range(len(batch)): image1024_list.append(batch[i][0]) image896_list.append(batch[i][1]) imageyolo_list.append(batch[i][2]) h_list.append(batch[i][3]) w_list.append(batch[i][4]) image1024_list = np.stack(image1024_list) image896_list = np.stack(image896_list) image1024_list = torch.from_numpy(image1024_list) image896_list = torch.from_numpy(image896_list) imageyolo_list = np.stack(imageyolo_list) imageyolo_list = torch.from_numpy(imageyolo_list) targets = {'h': h_list, 'w': w_list} return image1024_list, image896_list, imageyolo_list, targets def get_effdetd5(): config = get_efficientdet_config('tf_efficientdet_d5') config.image_size = [1024, 1024] config.norm_kwargs = dict(eps=0.001, momentum=0.01) net = EfficientDet(config, pretrained_backbone=False) net.reset_head(num_classes=6) return DetBenchPredict(net) def get_effdetd6(): config = get_efficientdet_config('tf_efficientdet_d6') config.image_size = [896, 896] config.norm_kwargs = dict(eps=0.001, momentum=0.01) net = EfficientDet(config, pretrained_backbone=False) net.reset_head(num_classes=6) return DetBenchPredict(net) start_time = time.time() # prepare input image_list = sorted(glob.glob('../input/siim-covid19-detection/test/*/*/*.dcm')) study_id_list = [] image_id_list = [] studyid2imageid = {} imageid2studyid = {} label_dict = {} for i in range(len(image_list)): study_id = image_list[i].split('/')[-3] image_id = image_list[i].split('/')[-1][:-4] image_id_list.append(image_id) if study_id not in studyid2imageid: study_id_list.append(study_id) studyid2imageid[study_id] = [image_id] else: studyid2imageid[study_id].append(image_id) imageid2studyid[image_id] = study_id label_dict[image_id] = { 'img_dir': image_list[i], } print(len(image_id_list), len(study_id_list), len(studyid2imageid), len(imageid2studyid)) image_id_list111 = [] for image_id in image_id_list: if image_id not in image_id_list_matched_midrc: image_id_list111.append(image_id) image_id_list = image_id_list111 print(len(image_id_list)) # hyperparameters batch_size = 4 # build model model0_fold0 = get_effdetd5() model0_fold0.load_state_dict(torch.load('../input/covidpretrained/effdetd5_fold0/epoch50_polyak')) model0_fold0.cuda() model0_fold0.eval() model0_fold1 = get_effdetd5() model0_fold1.load_state_dict(torch.load('../input/covidpretrained/effdetd5_fold1/epoch50_polyak')) model0_fold1.cuda() model0_fold1.eval() model0_fold2 = get_effdetd5() model0_fold2.load_state_dict(torch.load('../input/covidpretrained/effdetd5_fold2/epoch50_polyak')) model0_fold2.cuda() model0_fold2.eval() model0_fold3 = get_effdetd5() model0_fold3.load_state_dict(torch.load('../input/covidpretrained/effdetd5_fold3/epoch50_polyak')) model0_fold3.cuda() model0_fold3.eval() model0_fold4 = get_effdetd5() model0_fold4.load_state_dict(torch.load('../input/covidpretrained/effdetd5_fold4/epoch50_polyak')) model0_fold4.cuda() model0_fold4.eval() model1_fold0 = get_effdetd6() model1_fold0.load_state_dict(torch.load('../input/covidpretrained/effdetd6_fold0/epoch44_polyak')) model1_fold0.cuda() model1_fold0.eval() model1_fold1 = get_effdetd6() model1_fold1.load_state_dict(torch.load('../input/covidpretrained/effdetd6_fold1/epoch44_polyak')) model1_fold1.cuda() model1_fold1.eval() model1_fold2 = get_effdetd6() model1_fold2.load_state_dict(torch.load('../input/covidpretrained/effdetd6_fold2/epoch44_polyak')) model1_fold2.cuda() model1_fold2.eval() model1_fold3 = get_effdetd6() model1_fold3.load_state_dict(torch.load('../input/covidpretrained/effdetd6_fold3/epoch44_polyak')) model1_fold3.cuda() model1_fold3.eval() model1_fold4 = get_effdetd6() model1_fold4.load_state_dict(torch.load('../input/covidpretrained/effdetd6_fold4/epoch44_polyak')) model1_fold4.cuda() model1_fold4.eval() device = select_device('0') modelyolo1_fold0 = attempt_load('../input/covidpretrained/yolo1_fold0/best.pt', map_location=device) modelyolo1_fold0.eval() modelyolo1_fold1 = attempt_load('../input/covidpretrained/yolo1_fold1/best.pt', map_location=device) modelyolo1_fold1.eval() modelyolo1_fold2 = attempt_load('../input/covidpretrained/yolo1_fold2/best.pt', map_location=device) modelyolo1_fold2.eval() modelyolo1_fold3 = attempt_load('../input/covidpretrained/yolo1_fold3/best.pt', map_location=device) modelyolo1_fold3.eval() modelyolo2_fold0 = attempt_load('../input/covidpretrained/yolo2_fold0/best.pt', map_location=device) modelyolo2_fold0.eval() modelyolo2_fold1 = attempt_load('../input/covidpretrained/yolo2_fold1/best.pt', map_location=device) modelyolo2_fold1.eval() modelyolo2_fold2 = attempt_load('../input/covidpretrained/yolo2_fold2/best.pt', map_location=device) modelyolo2_fold2.eval() modelyolo2_fold3 = attempt_load('../input/covidpretrained/yolo2_fold3/best.pt', map_location=device) modelyolo2_fold3.eval() # iterator for testing datagen = TestDataset(label_dict=label_dict, image_list=image_id_list) generator = DataLoader(dataset=datagen, collate_fn=collate_fn, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=False) detections_all = [] for i, (images1024, images896, imagesyolo, targets) in enumerate(generator): with torch.no_grad(): start = i*batch_size end = start+batch_size if i == len(generator)-1: end = len(generator.dataset) images1024 = images1024.cuda() images1024 = images1024.permute(0,3,1,2).float().div(255) images896 = images896.cuda() images896 = images896.permute(0,3,1,2).float().div(255) imagesyolo = imagesyolo.cuda() imagesyolo = imagesyolo.float().div(255) with autocast(): detections_effdetd5_fold0 = model0_fold0(images1024).cpu().numpy() detections_effdetd5_fold1 = model0_fold1(images1024).cpu().numpy() detections_effdetd5_fold2 = model0_fold2(images1024).cpu().numpy() detections_effdetd5_fold3 = model0_fold3(images1024).cpu().numpy() detections_effdetd5_fold4 = model0_fold4(images1024).cpu().numpy() detections_flip_effdetd5_fold0 = model0_fold0(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold1 = model0_fold1(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold2 = model0_fold2(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold3 = model0_fold3(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold4 = model0_fold4(images1024.flip(3)).cpu().numpy() detections_effdetd6_fold0 = model1_fold0(images896).cpu().numpy() detections_effdetd6_fold1 = model1_fold1(images896).cpu().numpy() detections_effdetd6_fold2 = model1_fold2(images896).cpu().numpy() detections_effdetd6_fold3 = model1_fold3(images896).cpu().numpy() detections_effdetd6_fold4 = model1_fold4(images896).cpu().numpy() detections_flip_effdetd6_fold0 = model1_fold0(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold1 = model1_fold1(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold2 = model1_fold2(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold3 = model1_fold3(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold4 = model1_fold4(images896.flip(3)).cpu().numpy() detections_yolo1_fold0 = modelyolo1_fold0(imagesyolo, augment=False)[0] detections_yolo1_fold0 = non_max_suppression(detections_yolo1_fold0, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hflip_yolo1_fold1 = modelyolo1_fold1(imagesyolo.flip(3), augment=False)[0] detections_hflip_yolo1_fold1 = non_max_suppression(detections_hflip_yolo1_fold1, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_vflip_yolo1_fold2 = modelyolo1_fold2(imagesyolo.flip(2), augment=False)[0] detections_vflip_yolo1_fold2 = non_max_suppression(detections_vflip_yolo1_fold2, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hvflip_yolo1_fold3 = modelyolo1_fold3(imagesyolo.flip(3).flip(2), augment=False)[0] detections_hvflip_yolo1_fold3 = non_max_suppression(detections_hvflip_yolo1_fold3, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_yolo2_fold0 = modelyolo2_fold0(imagesyolo, augment=False)[0] detections_yolo2_fold0 = non_max_suppression(detections_yolo2_fold0, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hflip_yolo2_fold1 = modelyolo2_fold1(imagesyolo.flip(3), augment=False)[0] detections_hflip_yolo2_fold1 = non_max_suppression(detections_hflip_yolo2_fold1, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_vflip_yolo2_fold2 = modelyolo2_fold2(imagesyolo.flip(2), augment=False)[0] detections_vflip_yolo2_fold2 = non_max_suppression(detections_vflip_yolo2_fold2, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hvflip_yolo2_fold3 = modelyolo2_fold3(imagesyolo.flip(3).flip(2), augment=False)[0] detections_hvflip_yolo2_fold3 = non_max_suppression(detections_hvflip_yolo2_fold3, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_flip_effdetd5_fold0[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold0[:,:,[2,0]] detections_flip_effdetd5_fold1[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold1[:,:,[2,0]] detections_flip_effdetd5_fold2[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold2[:,:,[2,0]] detections_flip_effdetd5_fold3[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold3[:,:,[2,0]] detections_flip_effdetd5_fold4[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold4[:,:,[2,0]] detections_flip_effdetd6_fold0[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold0[:,:,[2,0]] detections_flip_effdetd6_fold1[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold1[:,:,[2,0]] detections_flip_effdetd6_fold2[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold2[:,:,[2,0]] detections_flip_effdetd6_fold3[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold3[:,:,[2,0]] detections_flip_effdetd6_fold4[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold4[:,:,[2,0]] detections_effdetd5_fold0[:,:,0] /= 1024 detections_effdetd5_fold0[:,:,2] /= 1024 detections_effdetd5_fold0[:,:,1] /= 1024 detections_effdetd5_fold0[:,:,3] /= 1024 detections_flip_effdetd5_fold0[:,:,0] /= 1024 detections_flip_effdetd5_fold0[:,:,2] /= 1024 detections_flip_effdetd5_fold0[:,:,1] /= 1024 detections_flip_effdetd5_fold0[:,:,3] /= 1024 detections_effdetd5_fold1[:,:,0] /= 1024 detections_effdetd5_fold1[:,:,2] /= 1024 detections_effdetd5_fold1[:,:,1] /= 1024 detections_effdetd5_fold1[:,:,3] /= 1024 detections_flip_effdetd5_fold1[:,:,0] /= 1024 detections_flip_effdetd5_fold1[:,:,2] /= 1024 detections_flip_effdetd5_fold1[:,:,1] /= 1024 detections_flip_effdetd5_fold1[:,:,3] /= 1024 detections_effdetd5_fold2[:,:,0] /= 1024 detections_effdetd5_fold2[:,:,2] /= 1024 detections_effdetd5_fold2[:,:,1] /= 1024 detections_effdetd5_fold2[:,:,3] /= 1024 detections_flip_effdetd5_fold2[:,:,0] /= 1024 detections_flip_effdetd5_fold2[:,:,2] /= 1024 detections_flip_effdetd5_fold2[:,:,1] /= 1024 detections_flip_effdetd5_fold2[:,:,3] /= 1024 detections_effdetd5_fold3[:,:,0] /= 1024 detections_effdetd5_fold3[:,:,2] /= 1024 detections_effdetd5_fold3[:,:,1] /= 1024 detections_effdetd5_fold3[:,:,3] /= 1024 detections_flip_effdetd5_fold3[:,:,0] /= 1024 detections_flip_effdetd5_fold3[:,:,2] /= 1024 detections_flip_effdetd5_fold3[:,:,1] /= 1024 detections_flip_effdetd5_fold3[:,:,3] /= 1024 detections_effdetd5_fold4[:,:,0] /= 1024 detections_effdetd5_fold4[:,:,2] /= 1024 detections_effdetd5_fold4[:,:,1] /= 1024 detections_effdetd5_fold4[:,:,3] /= 1024 detections_flip_effdetd5_fold4[:,:,0] /= 1024 detections_flip_effdetd5_fold4[:,:,2] /= 1024 detections_flip_effdetd5_fold4[:,:,1] /= 1024 detections_flip_effdetd5_fold4[:,:,3] /= 1024 detections_effdetd6_fold0[:,:,0] /= 896 detections_effdetd6_fold0[:,:,2] /= 896 detections_effdetd6_fold0[:,:,1] /= 896 detections_effdetd6_fold0[:,:,3] /= 896 detections_flip_effdetd6_fold0[:,:,0] /= 896 detections_flip_effdetd6_fold0[:,:,2] /= 896 detections_flip_effdetd6_fold0[:,:,1] /= 896 detections_flip_effdetd6_fold0[:,:,3] /= 896 detections_effdetd6_fold1[:,:,0] /= 896 detections_effdetd6_fold1[:,:,2] /= 896 detections_effdetd6_fold1[:,:,1] /= 896 detections_effdetd6_fold1[:,:,3] /= 896 detections_flip_effdetd6_fold1[:,:,0] /= 896 detections_flip_effdetd6_fold1[:,:,2] /= 896 detections_flip_effdetd6_fold1[:,:,1] /= 896 detections_flip_effdetd6_fold1[:,:,3] /= 896 detections_effdetd6_fold2[:,:,0] /= 896 detections_effdetd6_fold2[:,:,2] /= 896 detections_effdetd6_fold2[:,:,1] /= 896 detections_effdetd6_fold2[:,:,3] /= 896 detections_flip_effdetd6_fold2[:,:,0] /= 896 detections_flip_effdetd6_fold2[:,:,2] /= 896 detections_flip_effdetd6_fold2[:,:,1] /= 896 detections_flip_effdetd6_fold2[:,:,3] /= 896 detections_effdetd6_fold3[:,:,0] /= 896 detections_effdetd6_fold3[:,:,2] /= 896 detections_effdetd6_fold3[:,:,1] /= 896 detections_effdetd6_fold3[:,:,3] /= 896 detections_flip_effdetd6_fold3[:,:,0] /= 896 detections_flip_effdetd6_fold3[:,:,2] /= 896 detections_flip_effdetd6_fold3[:,:,1] /= 896 detections_flip_effdetd6_fold3[:,:,3] /= 896 detections_effdetd6_fold4[:,:,0] /= 896 detections_effdetd6_fold4[:,:,2] /= 896 detections_effdetd6_fold4[:,:,1] /= 896 detections_effdetd6_fold4[:,:,3] /= 896 detections_flip_effdetd6_fold4[:,:,0] /= 896 detections_flip_effdetd6_fold4[:,:,2] /= 896 detections_flip_effdetd6_fold4[:,:,1] /= 896 detections_flip_effdetd6_fold4[:,:,3] /= 896 for n in range(len(detections_yolo1_fold0)): detections_yolo1_fold0[n] = detections_yolo1_fold0[n].cpu().numpy() detections_yolo1_fold0[n][:,0] /= 1024 detections_yolo1_fold0[n][:,2] /= 1024 detections_yolo1_fold0[n][:,1] /= 1024 detections_yolo1_fold0[n][:,3] /= 1024 detections_yolo1_fold0[n][:,5] += 1 # effdet labels start from 1 detections_hflip_yolo1_fold1[n] = detections_hflip_yolo1_fold1[n].cpu().numpy() detections_hflip_yolo1_fold1[n][:,[0,2]] = 1024 - detections_hflip_yolo1_fold1[n][:,[2,0]] detections_hflip_yolo1_fold1[n][:,0] /= 1024 detections_hflip_yolo1_fold1[n][:,2] /= 1024 detections_hflip_yolo1_fold1[n][:,1] /= 1024 detections_hflip_yolo1_fold1[n][:,3] /= 1024 detections_hflip_yolo1_fold1[n][:,5] += 1 # effdet labels start from 1 detections_vflip_yolo1_fold2[n] = detections_vflip_yolo1_fold2[n].cpu().numpy() detections_vflip_yolo1_fold2[n][:,[1,3]] = 1024 - detections_vflip_yolo1_fold2[n][:,[3,1]] detections_vflip_yolo1_fold2[n][:,0] /= 1024 detections_vflip_yolo1_fold2[n][:,2] /= 1024 detections_vflip_yolo1_fold2[n][:,1] /= 1024 detections_vflip_yolo1_fold2[n][:,3] /= 1024 detections_vflip_yolo1_fold2[n][:,5] += 1 # effdet labels start from 1 detections_hvflip_yolo1_fold3[n] = detections_hvflip_yolo1_fold3[n].cpu().numpy() detections_hvflip_yolo1_fold3[n][:,[0,2]] = 1024 - detections_hvflip_yolo1_fold3[n][:,[2,0]] detections_hvflip_yolo1_fold3[n][:,[1,3]] = 1024 - detections_hvflip_yolo1_fold3[n][:,[3,1]] detections_hvflip_yolo1_fold3[n][:,0] /= 1024 detections_hvflip_yolo1_fold3[n][:,2] /= 1024 detections_hvflip_yolo1_fold3[n][:,1] /= 1024 detections_hvflip_yolo1_fold3[n][:,3] /= 1024 detections_hvflip_yolo1_fold3[n][:,5] += 1 # effdet labels start from 1 detections_yolo2_fold0[n] = detections_yolo2_fold0[n].cpu().numpy() detections_yolo2_fold0[n][:,0] /= 1024 detections_yolo2_fold0[n][:,2] /= 1024 detections_yolo2_fold0[n][:,1] /= 1024 detections_yolo2_fold0[n][:,3] /= 1024 detections_yolo2_fold0[n][:,5] += 1 # effdet labels start from 1 detections_hflip_yolo2_fold1[n] = detections_hflip_yolo2_fold1[n].cpu().numpy() detections_hflip_yolo2_fold1[n][:,[0,2]] = 1024 - detections_hflip_yolo2_fold1[n][:,[2,0]] detections_hflip_yolo2_fold1[n][:,0] /= 1024 detections_hflip_yolo2_fold1[n][:,2] /= 1024 detections_hflip_yolo2_fold1[n][:,1] /= 1024 detections_hflip_yolo2_fold1[n][:,3] /= 1024 detections_hflip_yolo2_fold1[n][:,5] += 1 # effdet labels start from 1 detections_vflip_yolo2_fold2[n] = detections_vflip_yolo2_fold2[n].cpu().numpy() detections_vflip_yolo2_fold2[n][:,[1,3]] = 1024 - detections_vflip_yolo2_fold2[n][:,[3,1]] detections_vflip_yolo2_fold2[n][:,0] /= 1024 detections_vflip_yolo2_fold2[n][:,2] /= 1024 detections_vflip_yolo2_fold2[n][:,1] /= 1024 detections_vflip_yolo2_fold2[n][:,3] /= 1024 detections_vflip_yolo2_fold2[n][:,5] += 1 # effdet labels start from 1 detections_hvflip_yolo2_fold3[n] = detections_hvflip_yolo2_fold3[n].cpu().numpy() detections_hvflip_yolo2_fold3[n][:,[0,2]] = 1024 - detections_hvflip_yolo2_fold3[n][:,[2,0]] detections_hvflip_yolo2_fold3[n][:,[1,3]] = 1024 - detections_hvflip_yolo2_fold3[n][:,[3,1]] detections_hvflip_yolo2_fold3[n][:,0] /= 1024 detections_hvflip_yolo2_fold3[n][:,2] /= 1024 detections_hvflip_yolo2_fold3[n][:,1] /= 1024 detections_hvflip_yolo2_fold3[n][:,3] /= 1024 detections_hvflip_yolo2_fold3[n][:,5] += 1 # effdet labels start from 1 detections_ensemble = np.zeros(detections_effdetd5_fold0.shape, dtype=np.float32) for n in range(detections_effdetd5_fold0.shape[0]): boxes = [ detections_effdetd5_fold0[n,:,:4].tolist(), detections_effdetd5_fold1[n,:,:4].tolist(), detections_effdetd5_fold2[n,:,:4].tolist(), detections_effdetd5_fold3[n,:,:4].tolist(), detections_effdetd5_fold4[n,:,:4].tolist(), detections_flip_effdetd5_fold0[n,:,:4].tolist(), detections_flip_effdetd5_fold1[n,:,:4].tolist(), detections_flip_effdetd5_fold2[n,:,:4].tolist(), detections_flip_effdetd5_fold3[n,:,:4].tolist(), detections_flip_effdetd5_fold4[n,:,:4].tolist(), detections_effdetd6_fold0[n,:,:4].tolist(), detections_effdetd6_fold1[n,:,:4].tolist(), detections_effdetd6_fold2[n,:,:4].tolist(), detections_effdetd6_fold3[n,:,:4].tolist(), detections_effdetd6_fold4[n,:,:4].tolist(), detections_flip_effdetd6_fold0[n,:,:4].tolist(), detections_flip_effdetd6_fold1[n,:,:4].tolist(), detections_flip_effdetd6_fold2[n,:,:4].tolist(), detections_flip_effdetd6_fold3[n,:,:4].tolist(), detections_flip_effdetd6_fold4[n,:,:4].tolist(), detections_yolo1_fold0[n][:,:4].tolist(), detections_hflip_yolo1_fold1[n][:,:4].tolist(), detections_vflip_yolo1_fold2[n][:,:4].tolist(), detections_hvflip_yolo1_fold3[n][:,:4].tolist(), detections_yolo2_fold0[n][:,:4].tolist(), detections_hflip_yolo2_fold1[n][:,:4].tolist(), detections_vflip_yolo2_fold2[n][:,:4].tolist(), detections_hvflip_yolo2_fold3[n][:,:4].tolist(), ] scores = [ detections_effdetd5_fold0[n,:,4].tolist(), detections_effdetd5_fold1[n,:,4].tolist(), detections_effdetd5_fold2[n,:,4].tolist(), detections_effdetd5_fold3[n,:,4].tolist(), detections_effdetd5_fold4[n,:,4].tolist(), detections_flip_effdetd5_fold0[n,:,4].tolist(), detections_flip_effdetd5_fold1[n,:,4].tolist(), detections_flip_effdetd5_fold2[n,:,4].tolist(), detections_flip_effdetd5_fold3[n,:,4].tolist(), detections_flip_effdetd5_fold4[n,:,4].tolist(), detections_effdetd6_fold0[n,:,4].tolist(), detections_effdetd6_fold1[n,:,4].tolist(), detections_effdetd6_fold2[n,:,4].tolist(), detections_effdetd6_fold3[n,:,4].tolist(), detections_effdetd6_fold4[n,:,4].tolist(), detections_flip_effdetd6_fold0[n,:,4].tolist(), detections_flip_effdetd6_fold1[n,:,4].tolist(), detections_flip_effdetd6_fold2[n,:,4].tolist(), detections_flip_effdetd6_fold3[n,:,4].tolist(), detections_flip_effdetd6_fold4[n,:,4].tolist(), detections_yolo1_fold0[n][:,4].tolist(), detections_hflip_yolo1_fold1[n][:,4].tolist(), detections_vflip_yolo1_fold2[n][:,4].tolist(), detections_hvflip_yolo1_fold3[n][:,4].tolist(), detections_yolo2_fold0[n][:,4].tolist(), detections_hflip_yolo2_fold1[n][:,4].tolist(), detections_vflip_yolo2_fold2[n][:,4].tolist(), detections_hvflip_yolo2_fold3[n][:,4].tolist(), ] labels = [ detections_effdetd5_fold0[n,:,5].tolist(), detections_effdetd5_fold1[n,:,5].tolist(), detections_effdetd5_fold2[n,:,5].tolist(), detections_effdetd5_fold3[n,:,5].tolist(), detections_effdetd5_fold4[n,:,5].tolist(), detections_flip_effdetd5_fold0[n,:,5].tolist(), detections_flip_effdetd5_fold1[n,:,5].tolist(), detections_flip_effdetd5_fold2[n,:,5].tolist(), detections_flip_effdetd5_fold3[n,:,5].tolist(), detections_flip_effdetd5_fold4[n,:,5].tolist(), detections_effdetd6_fold0[n,:,5].tolist(), detections_effdetd6_fold1[n,:,5].tolist(), detections_effdetd6_fold2[n,:,5].tolist(), detections_effdetd6_fold3[n,:,5].tolist(), detections_effdetd6_fold4[n,:,5].tolist(), detections_flip_effdetd6_fold0[n,:,5].tolist(), detections_flip_effdetd6_fold1[n,:,5].tolist(), detections_flip_effdetd6_fold2[n,:,5].tolist(), detections_flip_effdetd6_fold3[n,:,5].tolist(), detections_flip_effdetd6_fold4[n,:,5].tolist(), detections_yolo1_fold0[n][:,5].tolist(), detections_hflip_yolo1_fold1[n][:,5].tolist(), detections_vflip_yolo1_fold2[n][:,5].tolist(), detections_hvflip_yolo1_fold3[n][:,5].tolist(), detections_yolo2_fold0[n][:,5].tolist(), detections_hflip_yolo2_fold1[n][:,5].tolist(), detections_vflip_yolo2_fold2[n][:,5].tolist(), detections_hvflip_yolo2_fold3[n][:,5].tolist(), ] boxes, scores, labels = weighted_boxes_fusion(boxes, scores, labels, iou_thr=0.6) boxes[:,0] *= targets['w'][n] boxes[:,2] *= targets['w'][n] boxes[:,1] *= targets['h'][n] boxes[:,3] *= targets['h'][n] if len(boxes)>=99: detections_ensemble[n,:99,:4] = boxes[:99,:] detections_ensemble[n,:99,4] = scores[:99] detections_ensemble[n,:99,5] = labels[:99] else: detections_ensemble[n,:len(boxes),:4] = boxes detections_ensemble[n,:len(boxes),4] = scores detections_ensemble[n,:len(boxes),5] = labels # Estimate the none class using the topK (K=3) opacity probs. detections_ensemble[n,-1,:4] = np.array([0,0,1,1], dtype=np.float32) non_prob = 1.0 count = 0 for bb in range(detections_ensemble.shape[1]): if detections_ensemble[n,bb,5]==1.0: non_prob *= (1.0-detections_ensemble[n,bb,4]) count += 1 if count>=3: break detections_ensemble[n,-1,4] = non_prob detections_ensemble[n,-1,5] = 2 detections_all.append(detections_ensemble[n]) # generate submission image_level_dict = {} for image_id in image_id_list: image_level_dict[image_id] = {"opacity": [], "none": 0., "negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(image_id_list)): image_detections = detections_all[i] for j in range(len(image_detections)): if image_detections[j,5]==1.0: image_level_dict[image_id_list[i]]["opacity"].append([image_detections[j,4]]+list(image_detections[j,:4])) image_level_dict[image_id_list[i]]["none"] += image_detections[-1,4] for j in range(len(image_detections)): if image_detections[j,5]==3.0: image_level_dict[image_id_list[i]]["negative"] += image_detections[j,4] break for j in range(len(image_detections)): if image_detections[j,5]==4.0: image_level_dict[image_id_list[i]]["typical"] += image_detections[j,4] break for j in range(len(image_detections)): if image_detections[j,5]==5.0: image_level_dict[image_id_list[i]]["indeterminate"] += image_detections[j,4] break for j in range(len(image_detections)): if image_detections[j,5]==6.0: image_level_dict[image_id_list[i]]["atypical"] += image_detections[j,4] break # reprediction # iterator for testing datagen = TestDataset1(label_dict=midrc_dict, image_list=image_id_list_matched_midrc) generator = DataLoader(dataset=datagen, collate_fn=collate_fn, batch_size=batch_size, shuffle=False, num_workers=2, pin_memory=False) detections_all = [] for i, (images1024, images896, imagesyolo, targets) in enumerate(generator): with torch.no_grad(): start = i*batch_size end = start+batch_size if i == len(generator)-1: end = len(generator.dataset) images1024 = images1024.cuda() images1024 = images1024.permute(0,3,1,2).float().div(255) images896 = images896.cuda() images896 = images896.permute(0,3,1,2).float().div(255) imagesyolo = imagesyolo.cuda() imagesyolo = imagesyolo.float().div(255) with autocast(): detections_effdetd5_fold0 = model0_fold0(images1024).cpu().numpy() detections_effdetd5_fold1 = model0_fold1(images1024).cpu().numpy() detections_effdetd5_fold2 = model0_fold2(images1024).cpu().numpy() detections_effdetd5_fold3 = model0_fold3(images1024).cpu().numpy() detections_effdetd5_fold4 = model0_fold4(images1024).cpu().numpy() detections_flip_effdetd5_fold0 = model0_fold0(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold1 = model0_fold1(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold2 = model0_fold2(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold3 = model0_fold3(images1024.flip(3)).cpu().numpy() detections_flip_effdetd5_fold4 = model0_fold4(images1024.flip(3)).cpu().numpy() detections_effdetd6_fold0 = model1_fold0(images896).cpu().numpy() detections_effdetd6_fold1 = model1_fold1(images896).cpu().numpy() detections_effdetd6_fold2 = model1_fold2(images896).cpu().numpy() detections_effdetd6_fold3 = model1_fold3(images896).cpu().numpy() detections_effdetd6_fold4 = model1_fold4(images896).cpu().numpy() detections_flip_effdetd6_fold0 = model1_fold0(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold1 = model1_fold1(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold2 = model1_fold2(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold3 = model1_fold3(images896.flip(3)).cpu().numpy() detections_flip_effdetd6_fold4 = model1_fold4(images896.flip(3)).cpu().numpy() detections_yolo1_fold0 = modelyolo1_fold0(imagesyolo, augment=False)[0] detections_yolo1_fold0 = non_max_suppression(detections_yolo1_fold0, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hflip_yolo1_fold1 = modelyolo1_fold1(imagesyolo.flip(3), augment=False)[0] detections_hflip_yolo1_fold1 = non_max_suppression(detections_hflip_yolo1_fold1, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_vflip_yolo1_fold2 = modelyolo1_fold2(imagesyolo.flip(2), augment=False)[0] detections_vflip_yolo1_fold2 = non_max_suppression(detections_vflip_yolo1_fold2, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hvflip_yolo1_fold3 = modelyolo1_fold3(imagesyolo.flip(3).flip(2), augment=False)[0] detections_hvflip_yolo1_fold3 = non_max_suppression(detections_hvflip_yolo1_fold3, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_yolo2_fold0 = modelyolo2_fold0(imagesyolo, augment=False)[0] detections_yolo2_fold0 = non_max_suppression(detections_yolo2_fold0, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hflip_yolo2_fold1 = modelyolo2_fold1(imagesyolo.flip(3), augment=False)[0] detections_hflip_yolo2_fold1 = non_max_suppression(detections_hflip_yolo2_fold1, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_vflip_yolo2_fold2 = modelyolo2_fold2(imagesyolo.flip(2), augment=False)[0] detections_vflip_yolo2_fold2 = non_max_suppression(detections_vflip_yolo2_fold2, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_hvflip_yolo2_fold3 = modelyolo2_fold3(imagesyolo.flip(3).flip(2), augment=False)[0] detections_hvflip_yolo2_fold3 = non_max_suppression(detections_hvflip_yolo2_fold3, conf_thres=0.001, iou_thres=0.5, multi_label=True) detections_flip_effdetd5_fold0[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold0[:,:,[2,0]] detections_flip_effdetd5_fold1[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold1[:,:,[2,0]] detections_flip_effdetd5_fold2[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold2[:,:,[2,0]] detections_flip_effdetd5_fold3[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold3[:,:,[2,0]] detections_flip_effdetd5_fold4[:,:,[0,2]] = 1024 - detections_flip_effdetd5_fold4[:,:,[2,0]] detections_flip_effdetd6_fold0[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold0[:,:,[2,0]] detections_flip_effdetd6_fold1[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold1[:,:,[2,0]] detections_flip_effdetd6_fold2[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold2[:,:,[2,0]] detections_flip_effdetd6_fold3[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold3[:,:,[2,0]] detections_flip_effdetd6_fold4[:,:,[0,2]] = 896 - detections_flip_effdetd6_fold4[:,:,[2,0]] detections_effdetd5_fold0[:,:,0] /= 1024 detections_effdetd5_fold0[:,:,2] /= 1024 detections_effdetd5_fold0[:,:,1] /= 1024 detections_effdetd5_fold0[:,:,3] /= 1024 detections_flip_effdetd5_fold0[:,:,0] /= 1024 detections_flip_effdetd5_fold0[:,:,2] /= 1024 detections_flip_effdetd5_fold0[:,:,1] /= 1024 detections_flip_effdetd5_fold0[:,:,3] /= 1024 detections_effdetd5_fold1[:,:,0] /= 1024 detections_effdetd5_fold1[:,:,2] /= 1024 detections_effdetd5_fold1[:,:,1] /= 1024 detections_effdetd5_fold1[:,:,3] /= 1024 detections_flip_effdetd5_fold1[:,:,0] /= 1024 detections_flip_effdetd5_fold1[:,:,2] /= 1024 detections_flip_effdetd5_fold1[:,:,1] /= 1024 detections_flip_effdetd5_fold1[:,:,3] /= 1024 detections_effdetd5_fold2[:,:,0] /= 1024 detections_effdetd5_fold2[:,:,2] /= 1024 detections_effdetd5_fold2[:,:,1] /= 1024 detections_effdetd5_fold2[:,:,3] /= 1024 detections_flip_effdetd5_fold2[:,:,0] /= 1024 detections_flip_effdetd5_fold2[:,:,2] /= 1024 detections_flip_effdetd5_fold2[:,:,1] /= 1024 detections_flip_effdetd5_fold2[:,:,3] /= 1024 detections_effdetd5_fold3[:,:,0] /= 1024 detections_effdetd5_fold3[:,:,2] /= 1024 detections_effdetd5_fold3[:,:,1] /= 1024 detections_effdetd5_fold3[:,:,3] /= 1024 detections_flip_effdetd5_fold3[:,:,0] /= 1024 detections_flip_effdetd5_fold3[:,:,2] /= 1024 detections_flip_effdetd5_fold3[:,:,1] /= 1024 detections_flip_effdetd5_fold3[:,:,3] /= 1024 detections_effdetd5_fold4[:,:,0] /= 1024 detections_effdetd5_fold4[:,:,2] /= 1024 detections_effdetd5_fold4[:,:,1] /= 1024 detections_effdetd5_fold4[:,:,3] /= 1024 detections_flip_effdetd5_fold4[:,:,0] /= 1024 detections_flip_effdetd5_fold4[:,:,2] /= 1024 detections_flip_effdetd5_fold4[:,:,1] /= 1024 detections_flip_effdetd5_fold4[:,:,3] /= 1024 detections_effdetd6_fold0[:,:,0] /= 896 detections_effdetd6_fold0[:,:,2] /= 896 detections_effdetd6_fold0[:,:,1] /= 896 detections_effdetd6_fold0[:,:,3] /= 896 detections_flip_effdetd6_fold0[:,:,0] /= 896 detections_flip_effdetd6_fold0[:,:,2] /= 896 detections_flip_effdetd6_fold0[:,:,1] /= 896 detections_flip_effdetd6_fold0[:,:,3] /= 896 detections_effdetd6_fold1[:,:,0] /= 896 detections_effdetd6_fold1[:,:,2] /= 896 detections_effdetd6_fold1[:,:,1] /= 896 detections_effdetd6_fold1[:,:,3] /= 896 detections_flip_effdetd6_fold1[:,:,0] /= 896 detections_flip_effdetd6_fold1[:,:,2] /= 896 detections_flip_effdetd6_fold1[:,:,1] /= 896 detections_flip_effdetd6_fold1[:,:,3] /= 896 detections_effdetd6_fold2[:,:,0] /= 896 detections_effdetd6_fold2[:,:,2] /= 896 detections_effdetd6_fold2[:,:,1] /= 896 detections_effdetd6_fold2[:,:,3] /= 896 detections_flip_effdetd6_fold2[:,:,0] /= 896 detections_flip_effdetd6_fold2[:,:,2] /= 896 detections_flip_effdetd6_fold2[:,:,1] /= 896 detections_flip_effdetd6_fold2[:,:,3] /= 896 detections_effdetd6_fold3[:,:,0] /= 896 detections_effdetd6_fold3[:,:,2] /= 896 detections_effdetd6_fold3[:,:,1] /= 896 detections_effdetd6_fold3[:,:,3] /= 896 detections_flip_effdetd6_fold3[:,:,0] /= 896 detections_flip_effdetd6_fold3[:,:,2] /= 896 detections_flip_effdetd6_fold3[:,:,1] /= 896 detections_flip_effdetd6_fold3[:,:,3] /= 896 detections_effdetd6_fold4[:,:,0] /= 896 detections_effdetd6_fold4[:,:,2] /= 896 detections_effdetd6_fold4[:,:,1] /= 896 detections_effdetd6_fold4[:,:,3] /= 896 detections_flip_effdetd6_fold4[:,:,0] /= 896 detections_flip_effdetd6_fold4[:,:,2] /= 896 detections_flip_effdetd6_fold4[:,:,1] /= 896 detections_flip_effdetd6_fold4[:,:,3] /= 896 for n in range(len(detections_yolo1_fold0)): detections_yolo1_fold0[n] = detections_yolo1_fold0[n].cpu().numpy() detections_yolo1_fold0[n][:,0] /= 1024 detections_yolo1_fold0[n][:,2] /= 1024 detections_yolo1_fold0[n][:,1] /= 1024 detections_yolo1_fold0[n][:,3] /= 1024 detections_yolo1_fold0[n][:,5] += 1 # effdet labels start from 1 detections_hflip_yolo1_fold1[n] = detections_hflip_yolo1_fold1[n].cpu().numpy() detections_hflip_yolo1_fold1[n][:,[0,2]] = 1024 - detections_hflip_yolo1_fold1[n][:,[2,0]] detections_hflip_yolo1_fold1[n][:,0] /= 1024 detections_hflip_yolo1_fold1[n][:,2] /= 1024 detections_hflip_yolo1_fold1[n][:,1] /= 1024 detections_hflip_yolo1_fold1[n][:,3] /= 1024 detections_hflip_yolo1_fold1[n][:,5] += 1 # effdet labels start from 1 detections_vflip_yolo1_fold2[n] = detections_vflip_yolo1_fold2[n].cpu().numpy() detections_vflip_yolo1_fold2[n][:,[1,3]] = 1024 - detections_vflip_yolo1_fold2[n][:,[3,1]] detections_vflip_yolo1_fold2[n][:,0] /= 1024 detections_vflip_yolo1_fold2[n][:,2] /= 1024 detections_vflip_yolo1_fold2[n][:,1] /= 1024 detections_vflip_yolo1_fold2[n][:,3] /= 1024 detections_vflip_yolo1_fold2[n][:,5] += 1 # effdet labels start from 1 detections_hvflip_yolo1_fold3[n] = detections_hvflip_yolo1_fold3[n].cpu().numpy() detections_hvflip_yolo1_fold3[n][:,[0,2]] = 1024 - detections_hvflip_yolo1_fold3[n][:,[2,0]] detections_hvflip_yolo1_fold3[n][:,[1,3]] = 1024 - detections_hvflip_yolo1_fold3[n][:,[3,1]] detections_hvflip_yolo1_fold3[n][:,0] /= 1024 detections_hvflip_yolo1_fold3[n][:,2] /= 1024 detections_hvflip_yolo1_fold3[n][:,1] /= 1024 detections_hvflip_yolo1_fold3[n][:,3] /= 1024 detections_hvflip_yolo1_fold3[n][:,5] += 1 # effdet labels start from 1 detections_yolo2_fold0[n] = detections_yolo2_fold0[n].cpu().numpy() detections_yolo2_fold0[n][:,0] /= 1024 detections_yolo2_fold0[n][:,2] /= 1024 detections_yolo2_fold0[n][:,1] /= 1024 detections_yolo2_fold0[n][:,3] /= 1024 detections_yolo2_fold0[n][:,5] += 1 # effdet labels start from 1 detections_hflip_yolo2_fold1[n] = detections_hflip_yolo2_fold1[n].cpu().numpy() detections_hflip_yolo2_fold1[n][:,[0,2]] = 1024 - detections_hflip_yolo2_fold1[n][:,[2,0]] detections_hflip_yolo2_fold1[n][:,0] /= 1024 detections_hflip_yolo2_fold1[n][:,2] /= 1024 detections_hflip_yolo2_fold1[n][:,1] /= 1024 detections_hflip_yolo2_fold1[n][:,3] /= 1024 detections_hflip_yolo2_fold1[n][:,5] += 1 # effdet labels start from 1 detections_vflip_yolo2_fold2[n] = detections_vflip_yolo2_fold2[n].cpu().numpy() detections_vflip_yolo2_fold2[n][:,[1,3]] = 1024 - detections_vflip_yolo2_fold2[n][:,[3,1]] detections_vflip_yolo2_fold2[n][:,0] /= 1024 detections_vflip_yolo2_fold2[n][:,2] /= 1024 detections_vflip_yolo2_fold2[n][:,1] /= 1024 detections_vflip_yolo2_fold2[n][:,3] /= 1024 detections_vflip_yolo2_fold2[n][:,5] += 1 # effdet labels start from 1 detections_hvflip_yolo2_fold3[n] = detections_hvflip_yolo2_fold3[n].cpu().numpy() detections_hvflip_yolo2_fold3[n][:,[0,2]] = 1024 - detections_hvflip_yolo2_fold3[n][:,[2,0]] detections_hvflip_yolo2_fold3[n][:,[1,3]] = 1024 - detections_hvflip_yolo2_fold3[n][:,[3,1]] detections_hvflip_yolo2_fold3[n][:,0] /= 1024 detections_hvflip_yolo2_fold3[n][:,2] /= 1024 detections_hvflip_yolo2_fold3[n][:,1] /= 1024 detections_hvflip_yolo2_fold3[n][:,3] /= 1024 detections_hvflip_yolo2_fold3[n][:,5] += 1 # effdet labels start from 1 detections_ensemble = np.zeros(detections_effdetd5_fold0.shape, dtype=np.float32) for n in range(detections_effdetd5_fold0.shape[0]): boxes = [ detections_effdetd5_fold0[n,:,:4].tolist(), detections_effdetd5_fold1[n,:,:4].tolist(), detections_effdetd5_fold2[n,:,:4].tolist(), detections_effdetd5_fold3[n,:,:4].tolist(), detections_effdetd5_fold4[n,:,:4].tolist(), detections_flip_effdetd5_fold0[n,:,:4].tolist(), detections_flip_effdetd5_fold1[n,:,:4].tolist(), detections_flip_effdetd5_fold2[n,:,:4].tolist(), detections_flip_effdetd5_fold3[n,:,:4].tolist(), detections_flip_effdetd5_fold4[n,:,:4].tolist(), detections_effdetd6_fold0[n,:,:4].tolist(), detections_effdetd6_fold1[n,:,:4].tolist(), detections_effdetd6_fold2[n,:,:4].tolist(), detections_effdetd6_fold3[n,:,:4].tolist(), detections_effdetd6_fold4[n,:,:4].tolist(), detections_flip_effdetd6_fold0[n,:,:4].tolist(), detections_flip_effdetd6_fold1[n,:,:4].tolist(), detections_flip_effdetd6_fold2[n,:,:4].tolist(), detections_flip_effdetd6_fold3[n,:,:4].tolist(), detections_flip_effdetd6_fold4[n,:,:4].tolist(), detections_yolo1_fold0[n][:,:4].tolist(), detections_hflip_yolo1_fold1[n][:,:4].tolist(), detections_vflip_yolo1_fold2[n][:,:4].tolist(), detections_hvflip_yolo1_fold3[n][:,:4].tolist(), detections_yolo2_fold0[n][:,:4].tolist(), detections_hflip_yolo2_fold1[n][:,:4].tolist(), detections_vflip_yolo2_fold2[n][:,:4].tolist(), detections_hvflip_yolo2_fold3[n][:,:4].tolist(), ] scores = [ detections_effdetd5_fold0[n,:,4].tolist(), detections_effdetd5_fold1[n,:,4].tolist(), detections_effdetd5_fold2[n,:,4].tolist(), detections_effdetd5_fold3[n,:,4].tolist(), detections_effdetd5_fold4[n,:,4].tolist(), detections_flip_effdetd5_fold0[n,:,4].tolist(), detections_flip_effdetd5_fold1[n,:,4].tolist(), detections_flip_effdetd5_fold2[n,:,4].tolist(), detections_flip_effdetd5_fold3[n,:,4].tolist(), detections_flip_effdetd5_fold4[n,:,4].tolist(), detections_effdetd6_fold0[n,:,4].tolist(), detections_effdetd6_fold1[n,:,4].tolist(), detections_effdetd6_fold2[n,:,4].tolist(), detections_effdetd6_fold3[n,:,4].tolist(), detections_effdetd6_fold4[n,:,4].tolist(), detections_flip_effdetd6_fold0[n,:,4].tolist(), detections_flip_effdetd6_fold1[n,:,4].tolist(), detections_flip_effdetd6_fold2[n,:,4].tolist(), detections_flip_effdetd6_fold3[n,:,4].tolist(), detections_flip_effdetd6_fold4[n,:,4].tolist(), detections_yolo1_fold0[n][:,4].tolist(), detections_hflip_yolo1_fold1[n][:,4].tolist(), detections_vflip_yolo1_fold2[n][:,4].tolist(), detections_hvflip_yolo1_fold3[n][:,4].tolist(), detections_yolo2_fold0[n][:,4].tolist(), detections_hflip_yolo2_fold1[n][:,4].tolist(), detections_vflip_yolo2_fold2[n][:,4].tolist(), detections_hvflip_yolo2_fold3[n][:,4].tolist(), ] labels = [ detections_effdetd5_fold0[n,:,5].tolist(), detections_effdetd5_fold1[n,:,5].tolist(), detections_effdetd5_fold2[n,:,5].tolist(), detections_effdetd5_fold3[n,:,5].tolist(), detections_effdetd5_fold4[n,:,5].tolist(), detections_flip_effdetd5_fold0[n,:,5].tolist(), detections_flip_effdetd5_fold1[n,:,5].tolist(), detections_flip_effdetd5_fold2[n,:,5].tolist(), detections_flip_effdetd5_fold3[n,:,5].tolist(), detections_flip_effdetd5_fold4[n,:,5].tolist(), detections_effdetd6_fold0[n,:,5].tolist(), detections_effdetd6_fold1[n,:,5].tolist(), detections_effdetd6_fold2[n,:,5].tolist(), detections_effdetd6_fold3[n,:,5].tolist(), detections_effdetd6_fold4[n,:,5].tolist(), detections_flip_effdetd6_fold0[n,:,5].tolist(), detections_flip_effdetd6_fold1[n,:,5].tolist(), detections_flip_effdetd6_fold2[n,:,5].tolist(), detections_flip_effdetd6_fold3[n,:,5].tolist(), detections_flip_effdetd6_fold4[n,:,5].tolist(), detections_yolo1_fold0[n][:,5].tolist(), detections_hflip_yolo1_fold1[n][:,5].tolist(), detections_vflip_yolo1_fold2[n][:,5].tolist(), detections_hvflip_yolo1_fold3[n][:,5].tolist(), detections_yolo2_fold0[n][:,5].tolist(), detections_hflip_yolo2_fold1[n][:,5].tolist(), detections_vflip_yolo2_fold2[n][:,5].tolist(), detections_hvflip_yolo2_fold3[n][:,5].tolist(), ] boxes, scores, labels = weighted_boxes_fusion(boxes, scores, labels, iou_thr=0.6) boxes[:,0] *= targets['w'][n] boxes[:,2] *= targets['w'][n] boxes[:,1] *= targets['h'][n] boxes[:,3] *= targets['h'][n] if len(boxes)>=99: detections_ensemble[n,:99,:4] = boxes[:99,:] detections_ensemble[n,:99,4] = scores[:99] detections_ensemble[n,:99,5] = labels[:99] else: detections_ensemble[n,:len(boxes),:4] = boxes detections_ensemble[n,:len(boxes),4] = scores detections_ensemble[n,:len(boxes),5] = labels # Estimate the none class using the topK (K=3) opacity probs. detections_ensemble[n,-1,:4] = np.array([0,0,1,1], dtype=np.float32) non_prob = 1.0 count = 0 for bb in range(detections_ensemble.shape[1]): if detections_ensemble[n,bb,5]==1.0: non_prob *= (1.0-detections_ensemble[n,bb,4]) count += 1 if count>=3: break detections_ensemble[n,-1,4] = non_prob detections_ensemble[n,-1,5] = 2 detections_all.append(detections_ensemble[n]) # generate submission for image_id in image_id_list_matched_midrc: image_level_dict[image_id] = {"opacity": [], "none": 0., "negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(image_id_list_matched_midrc)): image_detections = detections_all[i] for j in range(len(image_detections)): if image_detections[j,5]==1.0: image_level_dict[image_id_list_matched_midrc[i]]["opacity"].append([image_detections[j,4]]+list(image_detections[j,:4])) image_level_dict[image_id_list_matched_midrc[i]]["none"] += image_detections[-1,4] for j in range(len(image_detections)): if image_detections[j,5]==3.0: image_level_dict[image_id_list_matched_midrc[i]]["negative"] += image_detections[j,4] break for j in range(len(image_detections)): if image_detections[j,5]==4.0: image_level_dict[image_id_list_matched_midrc[i]]["typical"] += image_detections[j,4] break for j in range(len(image_detections)): if image_detections[j,5]==5.0: image_level_dict[image_id_list_matched_midrc[i]]["indeterminate"] += image_detections[j,4] break for j in range(len(image_detections)): if image_detections[j,5]==6.0: image_level_dict[image_id_list_matched_midrc[i]]["atypical"] += image_detections[j,4] break end_time = time.time() print(end_time-start_time) return image_level_dict def main(): import numpy as np import pandas as pd import os import cv2 from tqdm import tqdm import pickle import time import glob # prepare input image_list = sorted(glob.glob('../input/siim-covid19-detection/test/*/*/*.dcm')) study_id_list = [] image_id_list = [] studyid2imageid = {} imageid2studyid = {} label_dict = {} for i in range(len(image_list)): study_id = image_list[i].split('/')[-3] image_id = image_list[i].split('/')[-1][:-4] image_id_list.append(image_id) if study_id not in studyid2imageid: study_id_list.append(study_id) studyid2imageid[study_id] = [image_id] else: studyid2imageid[study_id].append(image_id) imageid2studyid[image_id] = study_id label_dict[image_id] = { 'img_dir': image_list[i], } print(len(image_id_list), len(study_id_list), len(studyid2imageid), len(imageid2studyid)) image_level_dict1, image_id_list_matched_midrc, midrc_dict, image_id_list_matched_bimcv, bimcv_dict = get_study_level_results() image_level_dict2 = get_none_results(image_id_list_matched_midrc, midrc_dict) image_level_dict3 = get_detection_results(image_id_list_matched_midrc, midrc_dict) pred_none2 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_none3 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_negative1 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_negative3 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_typical1 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_typical3 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_indeterminate1 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_indeterminate3 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_atypical1 = np.zeros((len(image_id_list), ), dtype=np.float32) pred_atypical3 = np.zeros((len(image_id_list), ), dtype=np.float32) for i in range(len(image_id_list)): pred_none2[i] = image_level_dict2[image_id_list[i]]["none"] pred_none3[i] = image_level_dict3[image_id_list[i]]["none"] pred_negative1[i] = image_level_dict1[image_id_list[i]]["negative"] pred_negative3[i] = image_level_dict3[image_id_list[i]]["negative"] pred_typical1[i] = image_level_dict1[image_id_list[i]]["typical"] pred_typical3[i] = image_level_dict3[image_id_list[i]]["typical"] pred_indeterminate1[i] = image_level_dict1[image_id_list[i]]["indeterminate"] pred_indeterminate3[i] = image_level_dict3[image_id_list[i]]["indeterminate"] pred_atypical1[i] = image_level_dict1[image_id_list[i]]["atypical"] pred_atypical3[i] = image_level_dict3[image_id_list[i]]["atypical"] pred_none2 = pred_none2.argsort().argsort() / len(image_id_list) pred_none3 = pred_none3.argsort().argsort() / len(image_id_list) pred_negative1 = pred_negative1.argsort().argsort() / len(image_id_list) pred_negative3 = pred_negative3.argsort().argsort() / len(image_id_list) pred_typical1 = pred_typical1.argsort().argsort() / len(image_id_list) pred_typical3 = pred_typical3.argsort().argsort() / len(image_id_list) pred_indeterminate1 = pred_indeterminate1.argsort().argsort() / len(image_id_list) pred_indeterminate3 = pred_indeterminate3.argsort().argsort() / len(image_id_list) pred_atypical1 = pred_atypical1.argsort().argsort() / len(image_id_list) pred_atypical3 = pred_atypical3.argsort().argsort() / len(image_id_list) ########## df_pred_midrc = pd.read_csv('../input/covidnearestneighbors/df_pred_midrc.csv') id_list_pred_midrc = df_pred_midrc['id'].values negative_list_pred_midrc = df_pred_midrc['negative'].values typical_list_pred_midrc = df_pred_midrc['typical'].values indeterminate_list_pred_midrc = df_pred_midrc['indeterminate'].values atypical_list_pred_midrc = df_pred_midrc['atypical'].values none_list_pred_midrc = df_pred_midrc['none'].values pred_midrc_dict = {} for i in range(len(id_list_pred_midrc)): pred_midrc_dict[id_list_pred_midrc[i]] = {"none": none_list_pred_midrc[i], "negative": negative_list_pred_midrc[i], "typical": typical_list_pred_midrc[i], "indeterminate": indeterminate_list_pred_midrc[i], "atypical": atypical_list_pred_midrc[i]} df_pred_bimcv = pd.read_csv('../input/covidnearestneighbors/df_pred_bimcv.csv') id_list_pred_bimcv = df_pred_bimcv['id'].values negative_list_pred_bimcv = df_pred_bimcv['negative'].values typical_list_pred_bimcv = df_pred_bimcv['typical'].values indeterminate_list_pred_bimcv = df_pred_bimcv['indeterminate'].values atypical_list_pred_bimcv = df_pred_bimcv['atypical'].values none_list_pred_bimcv = df_pred_bimcv['none'].values pred_bimcv_dict = {} for i in range(len(id_list_pred_bimcv)): pred_bimcv_dict[id_list_pred_bimcv[i]] = {"none": none_list_pred_bimcv[i], "negative": negative_list_pred_bimcv[i], "typical": typical_list_pred_bimcv[i], "indeterminate": indeterminate_list_pred_bimcv[i], "atypical": atypical_list_pred_bimcv[i]} ########## image_level_dict = {} for image_id in image_id_list: image_level_dict[image_id] = {"opacity": [], "none": 0., "negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(image_id_list)): image_level_dict[image_id_list[i]]["opacity"] = image_level_dict3[image_id_list[i]]["opacity"] if image_id_list[i] in image_id_list_matched_midrc: image_level_dict[image_id_list[i]]["none"] = 0.65*(0.5*pred_none2[i] + 0.5*pred_none3[i])+0.35*(pred_midrc_dict[midrc_dict[image_id_list[i]]['midrc_id']]['none']) image_level_dict[image_id_list[i]]["negative"] = 0.65*(0.67*pred_negative1[i] + 0.33*pred_negative3[i])+0.35*(pred_midrc_dict[midrc_dict[image_id_list[i]]['midrc_id']]['negative']) image_level_dict[image_id_list[i]]["typical"] = 0.65*(0.67*pred_typical1[i] + 0.33*pred_typical3[i])+0.35*(pred_midrc_dict[midrc_dict[image_id_list[i]]['midrc_id']]['typical']) image_level_dict[image_id_list[i]]["indeterminate"] = 0.65*(0.67*pred_indeterminate1[i] + 0.33*pred_indeterminate3[i])+0.35*(pred_midrc_dict[midrc_dict[image_id_list[i]]['midrc_id']]['indeterminate']) image_level_dict[image_id_list[i]]["atypical"] = 0.65*(0.67*pred_atypical1[i] + 0.33*pred_atypical3[i])+0.35*(pred_midrc_dict[midrc_dict[image_id_list[i]]['midrc_id']]['atypical']) elif image_id_list[i] in image_id_list_matched_bimcv: image_level_dict[image_id_list[i]]["none"] = 0.65*(0.5*pred_none2[i] + 0.5*pred_none3[i])+0.35*(pred_bimcv_dict[bimcv_dict[image_id_list[i]]['bimcv_id']]['none']) image_level_dict[image_id_list[i]]["negative"] = 0.65*(0.67*pred_negative1[i] + 0.33*pred_negative3[i])+0.35*(pred_bimcv_dict[bimcv_dict[image_id_list[i]]['bimcv_id']]['negative']) image_level_dict[image_id_list[i]]["typical"] = 0.65*(0.67*pred_typical1[i] + 0.33*pred_typical3[i])+0.35*(pred_bimcv_dict[bimcv_dict[image_id_list[i]]['bimcv_id']]['typical']) image_level_dict[image_id_list[i]]["indeterminate"] = 0.65*(0.67*pred_indeterminate1[i] + 0.33*pred_indeterminate3[i])+0.35*(pred_bimcv_dict[bimcv_dict[image_id_list[i]]['bimcv_id']]['indeterminate']) image_level_dict[image_id_list[i]]["atypical"] = 0.65*(0.67*pred_atypical1[i] + 0.33*pred_atypical3[i])+0.35*(pred_bimcv_dict[bimcv_dict[image_id_list[i]]['bimcv_id']]['atypical']) else: image_level_dict[image_id_list[i]]["none"] = 0.5*pred_none2[i] + 0.5*pred_none3[i] image_level_dict[image_id_list[i]]["negative"] = 0.67*pred_negative1[i] + 0.33*pred_negative3[i] image_level_dict[image_id_list[i]]["typical"] = 0.67*pred_typical1[i] + 0.33*pred_typical3[i] image_level_dict[image_id_list[i]]["indeterminate"] = 0.67*pred_indeterminate1[i] + 0.33*pred_indeterminate3[i] image_level_dict[image_id_list[i]]["atypical"] = 0.67*pred_atypical1[i] + 0.33*pred_atypical3[i] study_level_dict = {} for study_id in study_id_list: study_level_dict[study_id] = {"negative": 0., "typical": 0., "indeterminate": 0., "atypical": 0.} for i in range(len(study_id_list)): for image_id in studyid2imageid[study_id_list[i]]: study_level_dict[study_id_list[i]]["negative"] += image_level_dict[image_id]["negative"] study_level_dict[study_id_list[i]]["negative"] /= len(studyid2imageid[study_id_list[i]]) for image_id in studyid2imageid[study_id_list[i]]: study_level_dict[study_id_list[i]]["typical"] += image_level_dict[image_id]["typical"] study_level_dict[study_id_list[i]]["typical"] /= len(studyid2imageid[study_id_list[i]]) for image_id in studyid2imageid[study_id_list[i]]: study_level_dict[study_id_list[i]]["indeterminate"] += image_level_dict[image_id]["indeterminate"] study_level_dict[study_id_list[i]]["indeterminate"] /= len(studyid2imageid[study_id_list[i]]) for image_id in studyid2imageid[study_id_list[i]]: study_level_dict[study_id_list[i]]["atypical"] += image_level_dict[image_id]["atypical"] study_level_dict[study_id_list[i]]["atypical"] /= len(studyid2imageid[study_id_list[i]]) id_list = [] PredictionString_list = [] for i in range(len(study_id_list)): id_list.append(study_id_list[i]+'_study') pred_str = "" pred_str += "negative" pred_str += " " pred_str += str(np.float16(study_level_dict[study_id_list[i]]["negative"])) pred_str += " " pred_str += "0 0 1 1" pred_str += " " pred_str += "typical" pred_str += " " pred_str += str(np.float16(study_level_dict[study_id_list[i]]["typical"])) pred_str += " " pred_str += "0 0 1 1" pred_str += " " pred_str += "indeterminate" pred_str += " " pred_str += str(np.float16(study_level_dict[study_id_list[i]]["indeterminate"])) pred_str += " " pred_str += "0 0 1 1" pred_str += " " pred_str += "atypical" pred_str += " " pred_str += str(np.float16(study_level_dict[study_id_list[i]]["atypical"])) pred_str += " " pred_str += "0 0 1 1" pred_str += " " PredictionString_list.append(pred_str) for i in range(len(image_id_list)): id_list.append(image_id_list[i]+'_image') pred_str = "" pred_str += "none" pred_str += " " pred_str += str(np.float16(image_level_dict[image_id_list[i]]["none"])) pred_str += " " pred_str += "0 0 1 1" pred_str += " " for bb in image_level_dict[image_id_list[i]]["opacity"]: pred_str += "opacity" pred_str += " " pred_str += str(np.float16(bb[0])) pred_str += " " pred_str += str(np.float16(bb[1])) pred_str += " " pred_str += str(np.float16(bb[2])) pred_str += " " pred_str += str(np.float16(bb[3])) pred_str += " " pred_str += str(np.float16(bb[4])) pred_str += " " PredictionString_list.append(pred_str) # generate submission sub_df = pd.DataFrame(data={'id': id_list, 'PredictionString': PredictionString_list}) sub_df.to_csv('submission.csv', index=False) if __name__ == "__main__": main()