test transform impact - codes

test_transform_exploration.py

+import os
+
+import torch
+import torchvision.transforms as transforms
+#import torch.nn as nn
+#import torch.optim as optim
+import torch.nn.functional as F
+
+from torchvision.datasets import ImageFolder
+from torch.utils.data import DataLoader, Subset
+
+# from torchvision.transforms import v2
+
+torch.backends.cudnn.deterministic = True
+torch.backends.cudnn.benchmark = False
+
+import numpy as np
+import pandas as pd
+#import matplotlib.pyplot as plt
+#import sklearn.metrics as metrics
+
+from sklearn.metrics import confusion_matrix
+#import seaborn as sns
+
+from itertools import product
+
+import random
+random.seed(42)
+torch.manual_seed(42)
+np.random.seed(42)
+
+# Define a function to get the transformations based on the configuration
+def get_transform(use_flip, rotation):
+    transform_list = []
+    # resize
+    transform_list.append(transforms.Resize(224))
+    # flip
+    if use_flip:
+        transform_list.append(transforms.RandomHorizontalFlip())
+    # rotation
+    transform_list.append(transforms.RandomRotation(degrees=rotation))
+    transform_list.append(transforms.ToTensor())
+    return transforms.Compose(transform_list)
+
+
+def predict(model, data_loader, device):
+    model.eval()
+
+    # save the predictions in this list
+    y_pred = []
+
+    # no gradient needed
+    with torch.no_grad():
+
+        # go over each batch in the loader. We can ignore the targets here
+        for batch, _ in data_loader:
+
+            # Move batch to the GPU
+            batch = batch.to(device)
+
+            # predict probabilities of each class
+            predictions = model(batch)
+
+            # apply a softmax to the predictions
+            predictions = F.softmax(predictions, dim=1)
+
+            # move to the cpu and convert to numpy
+            predictions = predictions.cpu().numpy()
+
+            # save
+            y_pred.append(predictions)
+
+    # stack predictions into a (num_samples, 10) array
+    y_pred = np.vstack(y_pred)
+    return y_pred
+
+
+# MAIN
+if torch.cuda.is_available():
+    device = 'cuda'
+    print('cuda is available')
+else:
+    device = 'cpu'
+    print('cuda is not available, change to cpu')
+    
+
+model_path = '/net/cremi/bformanek/TRDP_II/local_models/'
+model_name = 'transfer_checkpoints_resnet18_adam_amp_criterion_balanced_avp_025.pkl'
+model = torch.load(model_path + model_name, map_location=torch.device('cpu'), weights_only=False)
+model.to(device)
+
+DATA_PATH = '/net/travail/bformanek/MRI_dataset'
+TEST_FOLDER = DATA_PATH + '/test'
+
+# Set up your transformations with a list of options
+use_random_flip = [False, True]
+random_rotation = list(range(0, 180, 10))
+
+BATCH_SIZE = 64
+WORKERS = 8
+
+# Initialize list to store results
+results = []
+for flip, rotation in product(use_random_flip, random_rotation):
+    transform = get_transform(flip, rotation)
+
+    test_set = ImageFolder(TEST_FOLDER, transform = transform)
+    test_loader = DataLoader(test_set, batch_size = BATCH_SIZE,  shuffle = False, num_workers=WORKERS)
+    
+    # compute predictions on the test set
+    y_pred = predict(model, test_loader, device)
+    
+    # find the argmax of each of the predictions
+    y_pred = y_pred.argmax(axis=1)
+
+    # get the true labels and convert to numpy
+    y_true = np.array(test_set.targets)
+    
+    # balanced accuracy    
+    conf_matrix = confusion_matrix(y_true, y_pred)
+    TP = conf_matrix.diagonal()
+    P = conf_matrix.sum(axis=1)
+    balanced_accuracy = sum(TP / P) / len(P)
+    
+    results.append({
+        "use_random_flip": flip,
+        "random_rotation": rotation,
+        "balanced_accuracy": balanced_accuracy
+    })
+    
+    print(f"use_random_flip - {flip}, random_rotation - {rotation}, balanced_accuracy - {balanced_accuracy}")
+
+# Save results to a DataFrame and export to CSV for easy analysis
+df_results = pd.DataFrame(results)
+df_results.to_csv("transform_evaluation_results.csv", index=False)
\ No newline at end of file

transform_evaluation_results.csv

+use_random_flip,random_rotation,balanced_accuracy
+False,0,0.9910891981996583
+False,10,0.9791003824858996
+False,20,0.9555263606009453
+False,30,0.9350323928076885
+False,40,0.920475736344423
+False,50,0.9035698129166397
+False,60,0.8884172084451903
+False,70,0.8242767620826293
+False,80,0.8627538664121938
+False,90,0.8084297581588451
+False,100,0.7491355267513998
+False,110,0.798434987707292
+False,120,0.7858226803376203
+False,130,0.780957605269648
+False,140,0.7319620036389785
+False,150,0.7699308726795506
+False,160,0.7279718128212013
+False,170,0.7794329344544063
+True,0,0.9458076701140617
+True,10,0.9380200646297198
+True,20,0.9179824545578235
+True,30,0.8951788501005208
+True,40,0.8899794965416481
+True,50,0.8776486711418438
+True,60,0.8615161023146182
+True,70,0.8088444130000744
+True,80,0.8430272303480528
+True,90,0.7870537398293254
+True,100,0.7825543670178262
+True,110,0.7304412568002858
+True,120,0.7700303001804264
+True,130,0.7172527098110664
+True,140,0.7275591371037369
+True,150,0.7743519262432214
+True,160,0.728260065937451
+True,170,0.773318541938019