Upload app.py with huggingface_hub

app.py

@@ -1,11 +1,10 @@
 import os
 import base64
 import inspect
-import tempfile
 from queue import Queue
 import threading
 import torch
-from PIL import Image, ImageOps
+from PIL import Image
 from transformers import AutoProcessor, MiniCPMV4_6ForConditionalGeneration
 import gradio as gr
 
@@ -22,7 +21,6 @@ GPT_FALLBACK_MODEL_ID = "gpt-5.4-mini"
 GPT_MAX_COMPLETION_TOKENS = 4096
 GPT_REASONING_EFFORT = "none"
 NOTES_PROMPT = "Transcribe the musical notes in this image. Return only the transcription."
-MAX_PREPROCESSED_SIDE = 1800
 
 CAMERA_CAPTURE_JS = """
 function () {
@@ -100,86 +98,6 @@ def supports_keyword(callable_obj, keyword):
     return keyword in signature.parameters
 
 
-def _ensure_horizontal_music(image: Image.Image) -> Image.Image:
-    width, height = image.size
-    if height > width:
-        return image.rotate(90, expand=True)
-    return image
-
-
-def _score_training_orientation(image: Image.Image):
-    try:
-        import cv2
-        import numpy as np
-    except ImportError:
-        width, height = image.size
-        return width / max(height, 1)
-
-    gray = np.array(image.convert("L"))
-    height, width = gray.shape[:2]
-    if width == 0 or height == 0:
-        return -1
-
-    # Normalize out uneven camera lighting before comparing densities.
-    blur_size = min(max(min(width, height) // 6, 15) | 1, 201)
-    background = cv2.medianBlur(gray, blur_size)
-    normalized = cv2.divide(gray, background, scale=255)
-    ink = normalized < 128
-
-    band = max(width // 5, 1)
-    left_density = float(ink[:, :band].mean())
-    right_density = float(ink[:, -band:].mean())
-    # Treble clef + key sig always sit on the left, making that band denser.
-    return (width / max(height, 1)) * 2.0 + (left_density - right_density) * 15.0
-
-
-def _normalize_to_training_orientation(image: Image.Image):
-    candidates = []
-    for rotation in (0, 90, 180, 270):
-        candidate = image.rotate(rotation, expand=True) if rotation else image.copy()
-        candidate = _ensure_horizontal_music(candidate)
-        score = _score_training_orientation(candidate)
-        candidates.append((score, rotation, candidate))
-
-    score, rotation, candidate = max(candidates, key=lambda item: item[0])
-    return candidate, rotation, score
-
-
-def _limit_image_size(image: Image.Image) -> Image.Image:
-    width, height = image.size
-    max_side = max(width, height)
-    if max_side <= MAX_PREPROCESSED_SIDE:
-        return image
-
-    scale = MAX_PREPROCESSED_SIDE / max_side
-    return image.resize(
-        (int(width * scale), int(height * scale)),
-        Image.Resampling.LANCZOS,
-    )
-
-
-def preprocess_sheet_music_image(image_path):
-    original = ImageOps.exif_transpose(Image.open(image_path)).convert("RGB")
-    original_size = original.size
-    image, rotation, orientation_score = _normalize_to_training_orientation(original)
-    image = _limit_image_size(image)
-    processed_size = image.size
-
-    temp_file = tempfile.NamedTemporaryFile(suffix=".png", delete=False)
-    temp_path = temp_file.name
-    temp_file.close()
-    image.save(temp_path, format="PNG", optimize=True)
-    changed = processed_size != original_size
-    report = (
-        f"Preprocessing complete. Original: {original_size[0]}x{original_size[1]}; "
-        f"sent to models: {processed_size[0]}x{processed_size[1]}; "
-        f"selected rotation: {rotation} degrees; "
-        f"orientation score: {orientation_score:.2f}; "
-        f"changed: {'yes' if changed else 'no'}. "
-        "Treble clef/key-signature content is expected at the top-left of the preview."
-    )
-    return image, temp_path, report
-
 print("Loading processor...")
 processor = AutoProcessor.from_pretrained(ORIGINAL_MODEL_ID, trust_remote_code=True)
 
@@ -343,26 +261,20 @@ def warmup_models():
         return
 
     print("Warming up local models...")
-    image, processed_image_path, _ = preprocess_sheet_music_image(warmup_path)
-    try:
-        for name, model in (
-            ("fine-tuned", finetuned_model),
-            ("original", original_model),
-        ):
-            if model is None:
-                print(f"  Skipping {name} warmup; model failed to load.")
-                continue
-            print(f"  Warming {name} model...")
-            try:
-                generate_model_text(model, image, max_new_tokens=8)
-            except Exception as e:
-                print(f"  Warmup failed for {name} model: {type(e).__name__}: {e}")
-        print("Model warmup complete.")
-    finally:
+    image = Image.open(warmup_path).convert("RGB")
+    for name, model in (
+        ("fine-tuned", finetuned_model),
+        ("original", original_model),
+    ):
+        if model is None:
+            print(f"  Skipping {name} warmup; model failed to load.")
+            continue
+        print(f"  Warming {name} model...")
         try:
-            os.unlink(processed_image_path)
-        except OSError:
-            pass
+            generate_model_text(model, image, max_new_tokens=8)
+        except Exception as e:
+            print(f"  Warmup failed for {name} model: {type(e).__name__}: {e}")
+    print("Model warmup complete.")
 
 
 warmup_models()
@@ -483,18 +395,18 @@ def _run_stream(index, stream, updates):
 def predict_all(image_path):
     if image_path is None:
         message = "Please upload an image."
-        yield None, "", message, message, message
+        yield message, message, message
         return
 
-    image, processed_image_path, preprocessing_report = preprocess_sheet_music_image(image_path)
+    image = Image.open(image_path).convert("RGB")
     updates = Queue()
     outputs = ["", "", ""]
-    yield image, preprocessing_report, outputs[0], outputs[1], outputs[2]
+    yield outputs[0], outputs[1], outputs[2]
 
     streams = [
         stream_model_text(finetuned_model, image.copy(), "fine-tuned"),
         stream_model_text(original_model, image.copy(), "original"),
-        stream_gpt_text(processed_image_path),
+        stream_gpt_text(image_path),
     ]
 
     threads = [
@@ -517,16 +429,10 @@ def predict_all(image_path):
             continue
 
         outputs[index] = text
-        yield (image, preprocessing_report, *outputs)
+        yield (*outputs,)
 
-    try:
-        for thread in threads:
-            thread.join()
-    finally:
-        try:
-            os.unlink(processed_image_path)
-        except OSError:
-            pass
+    for thread in threads:
+        thread.join()
 
 
 if HAS_SPACES:
@@ -550,30 +456,17 @@ with gr.Blocks(**blocks_kwargs) as demo:
         """
     )
 
-    gr.Markdown(
-        "Camera: allow camera access, use the rear camera, frame the page normally with the top-left of the music at the top-left of the preview, then tap or click the preview to capture."
-    )
     image_input = gr.Image(
         type="filepath",
-        label="Sheet Music Image - capture the page in normal reading orientation",
+        label="Sheet Music Image",
         sources=["webcam", "upload", "clipboard"],
         webcam_options=gr.WebcamOptions(
             mirror=False,
-            constraints={"facingMode": {"ideal": "environment"}},
+            constraints={"facingMode": "environment"},
         ),
-        placeholder="Use Camera to frame the sheet music, then tap or click the preview.",
+        placeholder="Use the camera to capture the sheet music, then click Transcribe Music.",
         elem_id="sheet-music-input",
     )
-    processed_image_output = gr.Image(
-        type="pil",
-        label="Preprocessed Image Sent to All Models",
-        interactive=False,
-    )
-    preprocessing_report_output = gr.Textbox(
-        label="Preprocessing Confirmation",
-        interactive=False,
-        lines=2,
-    )
     gr.Examples(
         examples=[
             ["examples/000100005-1_1_1.png"],
@@ -601,13 +494,7 @@ with gr.Blocks(**blocks_kwargs) as demo:
     notes_btn.click(
         fn=predict_all,
         inputs=[image_input],
-        outputs=[
-            processed_image_output,
-            preprocessing_report_output,
-            finetuned_output,
-            original_output,
-            gpt_output,
-        ],
+        outputs=[finetuned_output, original_output, gpt_output],
     )
 
 launch_kwargs = {