full switch

README.md

@@ -51,4 +51,4 @@ The **PatchCamelyon (PCam)** benchmark is a challenging image classification dat
 
 The submission on kaggle with the model trained on this notebook is 
 
-```Public score: 0.9626```
+```Public score: 0.9733```

app.py

@@ -1,30 +1,40 @@
 import torch
 import torchvision
+import numpy as np
 from torch import nn
 from torchvision import transforms
 from torchvision.datasets import PCAM
 import gradio as gr
 from PIL import Image
+import albumentations as A
 
 # ---------------------------------
 # 1. Load model
 # ---------------------------------
 torch.manual_seed(42)
 torch.cuda.manual_seed_all(42)
-model = torch.load("results/pcam/19_06_2025_17_32_18/model_3.pt", map_location="cpu", weights_only=False)
+model = torch.load("results/pcam/20_06_2025_10_49_24/model_4.pt", map_location="cpu", weights_only=False)
 model.eval()
 
 # ---------------------------------
 # 2. Define transform and dataset
 # ---------------------------------
-mean_stdev = [torch.Tensor([0.6981, 0.5428, 0.6933]), torch.Tensor([0.2222, 0.2665, 0.1985])]
 
-models_paths = ['results/pcam/16_06_2025_21_34_05/model_5.pt']
-transform = transforms.Compose([
-    transforms.Resize((224, 224)),
-    transforms.ToTensor(),
-    transforms.Normalize(mean_stdev[0], mean_stdev[1])
-])
+a_transform = A.Compose([
+                A.Resize(224, 224),
+                A.Normalize(normalization="image_per_channel", p=1.0),
+                A.ToTensorV2()
+            ])
+class AlbumentationsToPytorchTransform:
+    def __init__(self, albumentations_transform):
+        self.albumentations_transform = albumentations_transform
+
+    def __call__(self, img):
+        img = np.array(img)  # Convert PIL to NumPy
+        transformed = self.albumentations_transform(image=img)
+        return transformed["image"].to(torch.float32)  # Return tensor
+    
+transform = AlbumentationsToPytorchTransform(a_transform)
 
 def load_datasets(dataset_choice):
     return [PCAM(root="data/", split=dataset_choice, download=True, transform=transform),
@@ -49,7 +59,7 @@ def get_sample(index: int, dataset_choice: str):
     with torch.no_grad():
         output = model(image_tensor.unsqueeze(0)).squeeze()
         probability = torch.sigmoid(output)
-        predicted_label = "Tumor" if probability >= 0.4096705 else "No Tumor"
+        predicted_label = "Tumor" if probability >= 0.4458489 else "No Tumor"
         true_label = "Tumor" if ground_truth == 1 else "No Tumor"
         error_label = ""
         if predicted_label != true_label:

results/pcam/20_06_2025_10_49_24/model_4.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ac9c1eef2742e3f95e354cca4062486a95e3e3952dc49863cef5b7217f5ae9d
+size 54427294