Upload 3 files

README.md

@@ -24,7 +24,8 @@ Zero-shot image classification for medical imagery powered by **google/medsiglip
 - Zero-shot predictions using the MedSigLIP vision-language model without fine-tuning.
 - Smart Modality Router v2 blends filename heuristics, simple color statistics, and a lightweight fallback classifier to choose the best label bank.
 - CT, Ultrasound, Musculoskeletal, chest X-ray, brain MRI, fundus, histopathology, skin, cardiovascular, and general label libraries curated from MedSigLIP prompts and clinical references.
-- CPU-optimized inference with single model load, float32 execution on CPU, capped torch threads, cached results, and batched label scoring.
+- CPU-optimized inference with single model load, float32 execution on CPU, capped torch threads via `psutil`, cached results, and batched label scoring.
+- Automatic image downscaling to 448×448 before scoring to keep memory usage predictable.
 - Gradio interface ready for local execution or deployment to Hugging Face Spaces (verified on Gradio 4.44.1+, API disabled by default to avoid schema bugs).
 
 
@@ -104,10 +105,10 @@ Each label file contains 100-200 modality-specific diagnostic phrases reflecting
 
 
 ## Performance Considerations
-- Loads the MedSigLIP processor and model once at startup, keeps the model in `eval()` mode, and pins execution to a single CPU thread with `torch.set_num_threads(1)`.
+- Loads the MedSigLIP processor and model once at startup, keeps the model in `eval()` mode, and limits PyTorch threading with `torch.set_num_threads(min(psutil.cpu_count(logical=False), 4))`.
 - Leverages the `cached_inference` utility (LRU cache of five items) to reuse results for repeated requests without re-running the full forward pass.
-- Splits label scoring into batches of 50 within the cache manager, applies softmax over the concatenated logits, and returns the top five predictions.
-- Executes in float32 on CPU (float16 on GPU when available) to balance precision and memory consumption.
+- Downscales incoming images to 448×448 prior to tokenization and splits label scoring into batches of 50, applying softmax over concatenated logits before returning the top five predictions.
+- Executes the transformer in float32 for deterministic CPU inference while still supporting GPU acceleration when available.
 - Avoids `transformers.pipeline()` to retain full control over preprocessing, batching, and device placement.
 
 

app.py

@@ -4,11 +4,12 @@ from functools import lru_cache
 from pathlib import Path
 from typing import Dict, List, Tuple
 
+import psutil
 import torch
 import gradio as gr
 from transformers import AutoModelForZeroShotImageClassification, AutoProcessor
 
-from utils.cache_manager import cached_inference
+from utils.cache_manager import cached_inference, configure_cache
 from utils.modality_router import detect_modality
 
 
@@ -19,19 +20,25 @@ MODEL_ID = "google/medsiglip-448"
 
 HF_TOKEN = os.getenv("HF_TOKEN")
 
-torch.set_num_threads(1)
+physical_cores = psutil.cpu_count(logical=False) or psutil.cpu_count() or 1
+torch.set_num_threads(min(physical_cores, 4))
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-model_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
 
-processor = AutoProcessor.from_pretrained(MODEL_ID, token=HF_TOKEN)
+processor = AutoProcessor.from_pretrained(
+    MODEL_ID,
+    token=HF_TOKEN,
+    use_fast=True,
+)
 model = AutoModelForZeroShotImageClassification.from_pretrained(
     MODEL_ID,
     token=HF_TOKEN,
-    torch_dtype=model_dtype,
+    torch_dtype=torch.float32,
 ).to(device)
 model.eval()
 
+configure_cache(model, processor)
+
 
 LABEL_OVERRIDES = {
     "xray": "chest_labels.json",
@@ -64,7 +71,7 @@ def classify_medical_image(image_path: str) -> Dict[str, float]:
         return {}
 
     candidate_labels = get_candidate_labels(image_path)
-    scores = cached_inference(image_path, candidate_labels, model, processor)
+    scores = cached_inference(image_path, candidate_labels)
 
     if not scores:
         return {}
@@ -81,18 +88,8 @@ demo = gr.Interface(
     outputs=gr.Label(num_top_classes=5, label="🧠 Top Predictions"),
     title="🩻 MedSigLIP Smart Medical Classifier",
     description="Zero-shot model with automatic label filtering for different modalities.",
-    allow_api=False,
 )
 
 
 if __name__ == "__main__":
-    server_name = os.getenv("SERVER_NAME", "0.0.0.0")
-    port_env = os.getenv("SERVER_PORT") or os.getenv("PORT") or "7860"
-    share_env = os.getenv("GRADIO_SHARE", "false").lower()
-
-    demo.launch(
-        server_name=server_name,
-        server_port=int(port_env),
-        share=share_env in {"1", "true", "yes"},
-        show_api=False,
-    )
+    demo.launch(server_name="0.0.0.0", server_port=7860, queue=True)

requirements.txt

@@ -1,10 +1,11 @@
-torch
-transformers>=4.44.0
-gradio>=4.44.1
-huggingface_hub>=0.24.0
+torch>=2.4.0
+transformers>=4.45.0
+gradio>=4.44.0
+huggingface_hub>=0.25.0
 sentencepiece
 Pillow
 numpy
 scikit-image
 timm
 tensorflow
+psutil