import random import numpy as np import torch from tqdm import tqdm from torch.utils.data import DataLoader import argparse import os import time from torch import nn from torch.optim import SGD, Optimizer from torch.optim.lr_scheduler import MultiStepLR from torch.utils.data import Dataset import torch.nn.functional as F from glob import glob import utilities torch.autograd.set_detect_anomaly(True) device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') print(device) parser = argparse.ArgumentParser() parser.add_argument('--folder', help='Folder of the runs. Either resnet_pseudo or resnet_pseudo_augment or resnet_pseudo_uncertainty or resnet_pseudo_uncertainty_augment', required=False, default="resnet_pseudo") parser.add_argument('--run_num', help='Number of run', required=False, default="0") args = parser.parse_args() cfolder = str(args.folder) run_num = int(args.run_num) min_max = {0: [-133.38071, 0.0], 1: [-135.69371, 0.0], 2: [-182.77412, -20.55280], 3: [-184.71960, -18.33420], 4: [-136.01241, 27.37697], 5: [0.0, 255.0]} norm_dict = {key: [val[0], val[1] - val[0]] for key, val in min_max.items()} bp = "./" all_files = {0: bp + "test_0_1F.txt", 1: bp + "test_1_2F.txt", 2: bp + "test_2_3C.txt", 3: bp + "test_3_3C2.txt", 4: bp + "test_4_3F.txt", 5: bp + "test_5_5F6F.txt"} time_size = 243 num_classes = 9 # Create Folders time.sleep(np.random.randint(5)) train_output_folder = "results_" + cfolder folder = glob(bp + train_output_folder + "/%06d_202*" % run_num)[0] resnet_to_use = int(folder.split("_")[-1]) use_norm = int(folder.split("_")[-2]) data_inds = (folder.split("_data_inds-")[-1].split("_")[0]).split("-") data_inds = [int(x) for x in data_inds] if not os.path.exists(folder + "/test/"): os.makedirs(folder + "/test/") else: print("Skipped %d: Already Tested." % run_num) exit() # Load Model # if resnet_to_use == 18: from torchvision.models import resnet18 model = resnet18(weights='IMAGENET1K_V1') model.fc = nn.Linear(512, num_classes) model = model.to(device) model.eval() elif resnet_to_use == 34: from torchvision.models import resnet34 model = resnet34(weights='IMAGENET1K_V1') model.fc = nn.Linear(512, num_classes) model = model.to(device) model.eval() elif resnet_to_use == 50: from torchvision.models import resnet50 model = resnet50(weights='IMAGENET1K_V1') model.fc = nn.Linear(2048, num_classes) model = model.to(device) model.eval() elif resnet_to_use == 101: from torchvision.models import resnet101 model = resnet101(weights='IMAGENET1K_V1') model.fc = nn.Linear(2048, num_classes) model = model.to(device) model.eval() elif resnet_to_use == 152: from torchvision.models import resnet152 model = resnet152(weights='IMAGENET1K_V1') model.fc = nn.Linear(2048, num_classes) model = model.to(device) model.eval() else: print("Invalid Resnet Number: ", resnet_to_use) exit() # ---------- # num_subfolder = len(glob(folder + "/[0-9]")) for di, data_ind in enumerate(data_inds): data_set_test = utilities.get_dataset([all_files[data_ind]], norm_dict, norm=use_norm) for i in range(num_subfolder): print("Dealing with Training %d" % i) cur_weight_files = glob(folder + "/%d/weights/*.h5" % i) cur_weight_files.sort() if "weights_base.h5" in cur_weight_files[-1]: cur_weight_files = [cur_weight_files[-1]] + cur_weight_files[:-1] first_base = True else: first_base = False test_log = folder + "/test/test_%d_%d.txt" % (data_ind, i) test_log_rev = folder + "/test/test_%d_%d_rev.txt" % (data_ind, i) for wi, wfile in enumerate(cur_weight_files): print(wfile) # Data Loader # test_loader = DataLoader(data_set_test, batch_size=1, shuffle=False, num_workers=0, pin_memory=True, drop_last=True) # ----------- # model.load_state_dict(torch.load(wfile, map_location=device)) model.eval() if wi == 0 and first_base: save_file_all_vals = folder + "/test/preds_%d_%d_base.npy" % (data_ind, i) else: if first_base: save_file_all_vals = folder + "/test/preds_%d_%d_%d.npy" % (data_ind, i, wi - 1) else: save_file_all_vals = folder + "/test/preds_%d_%d_%d.npy" % (data_ind, i, wi) preds = [] labs = [] inds = [] total_predictions = 0 correct_predictions = 0 with torch.no_grad(): with tqdm(enumerate(test_loader), total=len(test_loader)) as tqdm_eval: for vi, (x_dat, y_dat) in tqdm_eval: y_batch = y_dat.numpy() predictions = model(x_dat.to(device)).float().detach().cpu().numpy() preds.extend(predictions) labs.extend(y_batch) del x_dat number_of_correct_predictions = np.sum(np.argmax(predictions, axis=-1) == y_batch) total_predictions += len(y_batch) correct_predictions += number_of_correct_predictions tqdm_eval.set_postfix(accuracy=correct_predictions / total_predictions) multi_acc = correct_predictions / total_predictions preds = np.array(preds) labs = np.array(labs) np.save(save_file_all_vals, preds) pred_bool = np.argmax(preds, axis=-1) > 0 lab_bool = labs > 0 acc = np.mean(pred_bool == lab_bool) true_positives = np.sum(pred_bool * lab_bool) true_negatives = np.sum((~pred_bool) * (~lab_bool)) false_positives = np.sum(pred_bool * (~lab_bool)) false_negatives = np.sum((~pred_bool) * lab_bool) precision = true_positives/(true_positives + false_positives) recall = true_positives/(true_positives + false_negatives) f1_score = (2 * precision * recall) / (precision + recall) f2_score = ((1 + 2**2) * precision * recall) / (2**2 * precision + recall) f5_score = ((1 + 5**2) * precision * recall) / (5**2 * precision + recall) print('Test Multi Acc: %.5f, Acc: %.5f, F1: %.5f, F2: %.5f, F5: %.5f, Prec: %.5f, Rec: %.5f' % (multi_acc, acc, f1_score, f2_score, f5_score, precision, recall)) with open(test_log, "a") as f: f.write('%.5f\t%.5f\t%.5f\t%.5f\t%.5f\t%.5f\t%.5f\n' % (multi_acc, acc, f1_score, f2_score, f5_score, precision, recall)) # Same thing for reverse F1 Score (Swap positive and negative classes) pred_bool = np.argmax(preds, axis=-1) == 0 lab_bool = labs == 0 acc = np.mean(pred_bool == lab_bool) true_positives = np.sum(pred_bool * lab_bool) true_negatives = np.sum((~pred_bool) * (~lab_bool)) false_positives = np.sum(pred_bool * (~lab_bool)) false_negatives = np.sum((~pred_bool) * lab_bool) precision = true_positives/(true_positives + false_positives) recall = true_positives/(true_positives + false_negatives) f1_score = (2 * precision * recall) / (precision + recall) f2_score = ((1 + 2**2) * precision * recall) / (2**2 * precision + recall) f5_score = ((1 + 5**2) * precision * recall) / (5**2 * precision + recall) print('Test Multi Acc: %.5f, Acc: %.5f, F1: %.5f, F2: %.5f, F5: %.5f, Prec: %.5f, Rec: %.5f' % (multi_acc, acc, f1_score, f2_score, f5_score, precision, recall)) with open(test_log_rev, "a") as f: f.write('%.5f\t%.5f\t%.5f\t%.5f\t%.5f\t%.5f\t%.5f\n' % (multi_acc, acc, f1_score, f2_score, f5_score, precision, recall)) print("Done")