Create app.py

app.py

@@ -0,0 +1,119 @@
+import torch
+import gradio as gr
+from nilearn import datasets
+from nilearn.connectome import ConnectivityMeasure
+from nilearn.maskers import MultiNiftiMapsMasker
+import numpy as np
+
+#device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+# Force torch to use CPU only
+device = torch.device("cpu")
+
+
+
+try:
+    scripted_model = torch.jit.load("fmri_encoder_commercial.pt", map_location="cpu")
+    
+    # If the model is wrapped in DataParallel, unwrap it
+    if isinstance(scripted_model, torch.nn.DataParallel):
+        scripted_model = scripted_model.module
+    
+    scripted_model.to(device)
+    scripted_model.eval()
+except Exception as e:
+    print(f"Error loading model: {str(e)}")
+    exit(1)
+
+# Fetch atlas (e.g., DiFuMo)
+dim = 64  # Number of ROIs
+try:
+    difumo = datasets.fetch_atlas_difumo(dimension=dim, resolution_mm=2, legacy_format=False)
+    atlas_filename = difumo.maps
+except Exception as e:
+    print(f"Error fetching atlas: {str(e)}")
+    exit(1)
+
+# Create masker to extract features
+masker = MultiNiftiMapsMasker(
+    maps_img=atlas_filename,
+    standardize=True,
+    n_jobs=-1,
+    verbose=0
+)
+
+# Connectivity measure
+connectome_measure = ConnectivityMeasure(kind='correlation', vectorize=True, discard_diagonal=True)
+
+# Modified feature extraction function
+def extract_features_multiple(func_preproc_files):
+    all_features = []
+    if not func_preproc_files:
+        return all_features
+    
+    # Fit the masker on the first subject
+    print("Fitting masker on the first subject...")
+    masker.fit(func_preproc_files[0])
+    
+    for i, sub in enumerate(func_preproc_files):
+        print(f"Processing subject {i+1}...")
+        masked_data = masker.transform(sub)
+        transformed_data = connectome_measure.fit_transform([masked_data])[0]
+        all_features.append(transformed_data)
+    
+    print("All subjects processed.")
+    return all_features
+
+# Prediction function with error handling
+def validate_inputs(features_tensor):
+    if features_tensor.shape[1] != 2016:
+        raise ValueError(f"Expected 2016 features but got {features_tensor.shape[1]}")
+    if features_tensor.dim() != 2:
+        raise ValueError(f"Expected 2D tensor but got {features_tensor.dim()}D")
+
+def predict_autism(fmri_files, age, gender):
+    try:
+        if not fmri_files:
+            return "Please upload at least one valid .nii.gz file."
+        
+        features_list = extract_features_multiple(fmri_files)
+        if not features_list:
+            return "Error: Failed to extract features from the fMRI files."
+        
+        # CORRECTED TENSOR SHAPES
+        age_tensor = torch.tensor([float(age)], dtype=torch.float32).to(device)  # Shape: [1]
+        gender_tensor = torch.tensor([int(gender)], dtype=torch.long).to(device)  # Shape: [1]
+
+        predictions = []
+        for features in features_list:
+            features_tensor = torch.tensor(features, dtype=torch.float32).unsqueeze(0).to(device)
+            validate_inputs(features_tensor)
+            
+            with torch.no_grad():
+                prediction = scripted_model(features_tensor, age_tensor, gender_tensor)
+                probability = torch.sigmoid(prediction).item()
+            
+            result = f"Prediction: {'Autism' if probability > 0.5 else 'No Autism'} (Confidence: {probability:.2%})"
+            predictions.append(result)
+        
+        return "\n".join(predictions)
+    
+    except Exception as e:
+        return f"Error: {str(e)}"
+
+# Gradio interface
+iface = gr.Interface(
+    fn=predict_autism,
+    inputs=[
+        gr.File(label="Upload preprocessed fMRI files (.nii.gz)", file_count="multiple"),
+        gr.Number(label="Age", minimum=0, maximum=120),
+        gr.Radio(["0", "1"], label="Gender (0: Female, 1: Male)"),
+    ],
+    outputs=gr.Text(label="Prediction Result"),
+    title="Autism Prediction from fMRI Data",
+    description="Upload one or more preprocessed fMRI files (.nii.gz) and enter the subject's age and gender to predict autism.",
+    theme="default",
+    flagging_mode="never"
+)
+
+iface.launch()