Update app.py

app.py

@@ -3,6 +3,8 @@
 Loads `best_ema_v2.pt` from iamcode6/dermnet-skin23-eva02 on startup and serves
 predictions via Gradio. Runs on CPU (free HF Space tier).
 """
+import os
+
 import gradio as gr
 import torch
 import timm
@@ -10,6 +12,9 @@ from huggingface_hub import hf_hub_download
 from PIL import Image
 from timm.data import create_transform
 
+# Use both vCPUs on the free Space tier
+torch.set_num_threads(int(os.environ.get("TORCH_NUM_THREADS", 2)))
+
 # ==============================================================================
 # Load model from HF Hub at startup
 # ==============================================================================
@@ -46,7 +51,17 @@ val_tf = create_transform(
 
 
 # ==============================================================================
-# Inference (with light TTA: identity + hflip)
+# Warm up the model (first forward pass is JIT-slower; pre-pay it at startup)
+# ==============================================================================
+print("Warming up model...")
+with torch.no_grad():
+    dummy = torch.zeros(1, 3, IMG_SIZE, IMG_SIZE)
+    _ = model(dummy)
+print("Warmup done — ready to serve.")
+
+
+# ==============================================================================
+# Inference — single forward pass (no TTA on free CPU; would 2x latency)
 # ==============================================================================
 @torch.no_grad()
 def predict(image):
@@ -56,12 +71,9 @@ def predict(image):
         image = Image.fromarray(image)
     image = image.convert("RGB")
 
-    x = val_tf(image).unsqueeze(0)
-    x_flip = torch.flip(x, dims=[-1])
-    batch = torch.cat([x, x_flip], dim=0).to(device)
-
-    logits = model(batch)
-    probs = torch.softmax(logits, dim=-1).mean(dim=0)
+    x = val_tf(image).unsqueeze(0).to(device)
+    logits = model(x)
+    probs = torch.softmax(logits, dim=-1).squeeze(0)
     return {idx_to_class[i]: float(probs[i]) for i in range(NUM_CLASSES)}
 
 
@@ -75,7 +87,7 @@ DESCRIPTION = """# DermNet-Skin23 Classifier
 **Single-model accuracy**: **81.48%** acc / **0.7969** macro F1 on a 3,856-image val split.
 **Full 5-model ensemble reaches 82.86% / 0.8113** — see the ensemble repo linked below.
 
-Upload a clinical or dermoscopy photo and the model returns the top-5 most likely categories with calibrated confidence.
+Upload a clinical or dermoscopy photo, click **Classify**, and the model returns the top-5 most likely categories with calibrated confidence. **Inference takes ~30-45 seconds** on the free CPU tier — large vision transformer at 448×448 resolution.
 
 ---
 
@@ -91,7 +103,7 @@ LINKS = """
 **Trained on**: AMD Instinct MI300X (192 GB HBM3) via DigitalOcean, ROCm 7.0, PyTorch with HIP.
 """
 
-with gr.Blocks(title="DermNet-Skin23 Classifier", theme=gr.themes.Soft()) as demo:
+with gr.Blocks(title="DermNet-Skin23 Classifier") as demo:
     gr.Markdown(DESCRIPTION)
 
     with gr.Row():
@@ -103,9 +115,10 @@ with gr.Blocks(title="DermNet-Skin23 Classifier", theme=gr.themes.Soft()) as dem
 
     gr.Markdown(LINKS)
 
+    # Only run on explicit button click — avoids racing the Gradio queue
+    # if a user uploads multiple images quickly
     predict_btn.click(predict, inputs=input_image, outputs=output_label)
-    input_image.change(predict, inputs=input_image, outputs=output_label)
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(theme=gr.themes.Soft())