Spaces:
Sleeping
Sleeping
| """ | |
| Histopathologic Cancer Detection: Gradio Inference Space | |
| Model: ConvNeXt-Tiny + Residual Head (locked-test AUC 0.9979) | |
| """ | |
| import torch | |
| import numpy as np | |
| import torch.nn as nn | |
| import timm | |
| import gradio as gr | |
| from huggingface_hub import hf_hub_download | |
| from PIL import Image | |
| from torchvision import transforms | |
| # ────────────────────────────────────────────── | |
| # 1. Model Architecture (must match training) | |
| # ────────────────────────────────────────────── | |
| class ResidualHead(nn.Module): | |
| """Classification head with residual connection - MUST match training architecture.""" | |
| def __init__(self, in_features: int, hidden: int = 256, dropout: float = 0.3): | |
| super().__init__() | |
| self.fc1 = nn.Linear(in_features, hidden) | |
| self.bn1 = nn.BatchNorm1d(hidden) | |
| self.fc2 = nn.Linear(hidden, hidden) | |
| self.bn2 = nn.BatchNorm1d(hidden) | |
| self.fc_out = nn.Linear(hidden, 1) | |
| self.shortcut = nn.Linear(in_features, hidden) | |
| self.gelu = nn.GELU() | |
| self.dropout = nn.Dropout(dropout) | |
| def forward(self, x): | |
| identity = self.shortcut(x) | |
| out = self.fc1(x) | |
| out = self.bn1(out) | |
| out = self.gelu(out) | |
| out = self.dropout(out) | |
| out = self.fc2(out) | |
| out = self.bn2(out) | |
| out = out + identity # Residual connection | |
| out = self.gelu(out) | |
| out = self.dropout(out) | |
| return self.fc_out(out) | |
| class CancerClassifier(nn.Module): | |
| def __init__(self): | |
| super().__init__() | |
| self.backbone = timm.create_model("convnext_tiny", pretrained=False, num_classes=0) | |
| num_features = self.backbone.num_features # 768 | |
| self.head = ResidualHead(num_features) | |
| def forward(self, x): | |
| features = self.backbone(x) | |
| return self.head(features) | |
| # ────────────────────────────────────────────── | |
| # 2. Load Model | |
| # ────────────────────────────────────────────── | |
| device = torch.device("cuda" if torch.cuda.is_available() else "cpu") | |
| model = CancerClassifier() | |
| # Serve the v2 model evaluated in the report: the one the reported metrics | |
| # describe (SHA256 b08447ab760d6e0f…). The previous checkpoint was an epoch-50 | |
| # save from the v1 run, selected on v1 validation AUC; the locked-test figures | |
| # do not apply to it. Same 198 tensor keys, so the architecture is unchanged; | |
| # this file is a bare state_dict, handled by the third branch below. | |
| ckpt_path = hf_hub_download( | |
| repo_id="hakim78/Cancer_pathos", | |
| filename="models/final_model_v2_locked.pth", | |
| ) | |
| state = torch.load(ckpt_path, map_location=device, weights_only=False) | |
| # Handle different checkpoint formats | |
| if isinstance(state, dict) and "model_state_dict" in state: | |
| model.load_state_dict(state["model_state_dict"]) | |
| elif isinstance(state, dict) and "state_dict" in state: | |
| model.load_state_dict(state["state_dict"]) | |
| else: | |
| model.load_state_dict(state) | |
| model.to(device) | |
| model.eval() | |
| # ────────────────────────────────────────────── | |
| # 3. Preprocessing (must match training) | |
| # ────────────────────────────────────────────── | |
| transform = transforms.Compose([ | |
| transforms.Resize((96, 96)), | |
| transforms.ToTensor(), | |
| transforms.Normalize( | |
| mean=[0.485, 0.456, 0.406], | |
| std=[0.229, 0.224, 0.225], | |
| ), | |
| ]) | |
| # ────────────────────────────────────────────── | |
| # 4. Prediction function | |
| # ────────────────────────────────────────────── | |
| def _tta_views(t): | |
| """5 vues geometriques reproduisant get_tta_transforms() d'evaluate.py. | |
| Implementees en ops tensorielles (le Space n'a pas albumentations). | |
| Le patch est 96x96 (carre) : rot90/transpose ne changent pas la forme. | |
| """ | |
| return [ | |
| t, # 1. original | |
| torch.flip(t, dims=[3]), # 2. HorizontalFlip (axe largeur) | |
| torch.flip(t, dims=[2]), # 3. VerticalFlip (axe hauteur) | |
| torch.rot90(t, k=1, dims=[2, 3]), # 4. RandomRotate90 (rotation 90 deg) | |
| t.transpose(2, 3), # 5. Transpose (echange H/W) | |
| ] | |
| def predict(image: Image.Image): | |
| if image is None: | |
| return {"Cancer": 0.0, "Non-Cancer": 0.0} | |
| img = image.convert("RGB") | |
| tensor = transform(img).unsqueeze(0).to(device) | |
| # TTA x5 : moyenne des probabilites sur les 5 vues (cf. evaluate.py) | |
| probs = [torch.sigmoid(model(view).squeeze()).item() for view in _tta_views(tensor)] | |
| prob = float(np.mean(probs)) | |
| return { | |
| "Cancer": round(prob, 4), | |
| "Non-Cancer": round(1 - prob, 4), | |
| } | |
| # ────────────────────────────────────────────── | |
| # 5. Gradio Interface | |
| # ────────────────────────────────────────────── | |
| examples = [] # Add example image paths here if you upload them to the Space | |
| demo = gr.Interface( | |
| fn=predict, | |
| inputs=gr.Image(type="pil", label="Histopathology Patch (96x96)"), | |
| outputs=gr.Label(num_top_classes=2, label="Prediction"), | |
| title="Histopathologic Cancer Detection", | |
| description=( | |
| "Upload a 96×96 histopathologic tissue patch to detect metastatic cancer. " | |
| "Model: ConvNeXt-Tiny + Residual Head, locked-test AUC 0.9979 " | |
| "(95% CI 0.9975–0.9982, n = 33,004). " | |
| "~0.8 s per patch on CPU, 5-fold TTA included." | |
| ), | |
| examples=examples or None, | |
| flagging_mode="never", | |
| ) | |
| if __name__ == "__main__": | |
| demo.launch() | |