Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -1,5 +1,4 @@
|
|
1 |
# app.py β HTR Space with Feedback Loop, Memory Post-Correction, and GRPO Export
|
2 |
-
|
3 |
import os, time, json, hashlib, difflib, uuid, csv
|
4 |
from datetime import datetime
|
5 |
from collections import Counter, defaultdict
|
@@ -18,17 +17,17 @@ from jiwer import cer
|
|
18 |
|
19 |
# ---------------- Storage & Paths ----------------
|
20 |
os.makedirs("data", exist_ok=True)
|
21 |
-
FEEDBACK_PATH = "data/feedback.jsonl"
|
22 |
-
MEMORY_RULES_PATH = "data/memory_rules.json"
|
23 |
-
GRPO_EXPORT_PATH = "data/grpo_prefs.jsonl"
|
24 |
-
CSV_EXPORT_PATH = "data/feedback.csv"
|
25 |
|
26 |
# ---------------- Models ----------------
|
27 |
MODEL_PATHS = {
|
28 |
"Model 1 (Complex handwrittings )": ("prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it", Qwen2_5_VLForConditionalGeneration),
|
29 |
"Model 2 (simple and scanned handwritting )": ("nanonets/Nanonets-OCR-s", Qwen2_5_VLForConditionalGeneration),
|
30 |
}
|
31 |
-
|
32 |
|
33 |
MAX_NEW_TOKENS_DEFAULT = 512
|
34 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
@@ -161,7 +160,7 @@ def _compile_rules_from_feedback(min_count: int = 2, max_phrase_len: int = 40):
|
|
161 |
row = json.loads(line)
|
162 |
except Exception:
|
163 |
continue
|
164 |
-
if row.get("reward", 0) < 1:
|
165 |
continue
|
166 |
pred = _safe_text(row.get("prediction", ""))
|
167 |
corr = _safe_text(row.get("correction", "")) or _safe_text(row.get("ground_truth", ""))
|
@@ -266,7 +265,7 @@ def _append_jsonl(path, obj):
|
|
266 |
def _export_csv():
|
267 |
# optional: CSV summary for spreadsheet views
|
268 |
if not os.path.exists(FEEDBACK_PATH):
|
269 |
-
return
|
270 |
rows = []
|
271 |
with open(FEEDBACK_PATH, "r", encoding="utf-8") as f:
|
272 |
for line in f:
|
@@ -285,15 +284,17 @@ def _export_csv():
|
|
285 |
w.writerow(flat)
|
286 |
return CSV_EXPORT_PATH
|
287 |
|
|
|
288 |
def save_feedback(image: Image.Image, model_choice: str, prompt: str,
|
289 |
prediction: str, correction: str, ground_truth: str, reward: int):
|
290 |
"""
|
291 |
reward: 1 = good/accepted, 0 = neutral, -1 = bad
|
292 |
"""
|
293 |
if image is None:
|
294 |
-
|
|
|
295 |
if not prediction and not correction and not ground_truth:
|
296 |
-
return "Nothing to save."
|
297 |
|
298 |
image_hash = _hash_image(image)
|
299 |
# best target = correction, else ground_truth, else prediction
|
@@ -319,7 +320,8 @@ def save_feedback(image: Image.Image, model_choice: str, prompt: str,
|
|
319 |
"cer": float(cer_score) if cer_score is not None else None,
|
320 |
}
|
321 |
_append_jsonl(FEEDBACK_PATH, row)
|
322 |
-
return f"β
Feedback saved (reward={reward})."
|
|
|
323 |
|
324 |
def compile_memory_rules():
|
325 |
_compile_rules_from_feedback(min_count=2, max_phrase_len=60)
|
@@ -357,11 +359,18 @@ def export_grpo_preferences():
|
|
357 |
count += 1
|
358 |
return f"β
Exported {count} GRPO preference pairs to {GRPO_EXPORT_PATH}."
|
359 |
|
360 |
-
|
361 |
-
|
362 |
-
if
|
363 |
-
return
|
364 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
365 |
|
366 |
# ---------------- Evaluation Orchestration ----------------
|
367 |
@spaces.GPU
|
@@ -438,7 +447,7 @@ def main():
|
|
438 |
|
439 |
trainer = GRPOTrainer(
|
440 |
model=model,
|
441 |
-
ref_model=None,
|
442 |
args=cfg,
|
443 |
tokenizer=tok,
|
444 |
train_dataset=ds
|
@@ -460,7 +469,7 @@ def _write_trainer_script():
|
|
460 |
|
461 |
# ---------------- Gradio Interface ----------------
|
462 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
463 |
-
gr.Markdown("## βπΎ wilson Handwritten
|
464 |
|
465 |
model_choice = gr.Radio(choices=list(MODEL_PATHS.keys()),
|
466 |
value=list(MODEL_PATHS.keys())[0],
|
@@ -552,38 +561,61 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
552 |
|
553 |
with gr.Tab("βοΈ Feedback & Memory"):
|
554 |
gr.Markdown("""
|
555 |
-
**Pipeline**
|
556 |
-
1) Save feedback (π / π) and add corrections.
|
557 |
-
2) Click **Build/Refresh Memory** to generate auto-fix rules from positive feedback.
|
558 |
3) Keep **Enable Memory Post-correction** checked on inference/eval tabs.
|
559 |
""")
|
560 |
build_mem_btn = gr.Button("π§ Build/Refresh Memory from Feedback")
|
561 |
mem_status = gr.Markdown()
|
562 |
build_mem_btn.click(fn=compile_memory_rules, outputs=[mem_status])
|
563 |
|
564 |
-
csv_btn = gr.Button("π€ Export Feedback as CSV")
|
565 |
csv_status = gr.Markdown()
|
566 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
567 |
|
568 |
with gr.Tab("π§ͺ GRPO / Dataset"):
|
569 |
gr.Markdown("""
|
570 |
**GRPO Fine-tuning** (run offline or in a training Space):
|
571 |
- Click **Export GRPO Preferences** to produce `data/grpo_prefs.jsonl` of (prompt, chosen, rejected).
|
572 |
- Click **Write Trainer Script** to create `train/grpo_train.py`.
|
573 |
-
- Then run:
|
574 |
```bash
|
575 |
pip install trl accelerate peft transformers datasets
|
576 |
python train/grpo_train.py
|
577 |
-
|
|
|
578 |
Set `BASE_MODEL`/`OUTPUT_DIR` env vars if you like.
|
579 |
-
|
580 |
-
|
581 |
-
|
582 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
583 |
|
584 |
-
|
585 |
-
write_script_status = gr.Markdown()
|
586 |
-
write_script_btn.click(fn=lambda: f"β
Trainer script written to `{_write_trainer_script()}`", outputs=[write_script_status])
|
587 |
|
588 |
if __name__ == "__main__":
|
589 |
demo.queue(max_size=50).launch(share=True)
|
|
|
1 |
# app.py β HTR Space with Feedback Loop, Memory Post-Correction, and GRPO Export
|
|
|
2 |
import os, time, json, hashlib, difflib, uuid, csv
|
3 |
from datetime import datetime
|
4 |
from collections import Counter, defaultdict
|
|
|
17 |
|
18 |
# ---------------- Storage & Paths ----------------
|
19 |
os.makedirs("data", exist_ok=True)
|
20 |
+
FEEDBACK_PATH = "data/feedback.jsonl" # raw feedback log (per sample)
|
21 |
+
MEMORY_RULES_PATH = "data/memory_rules.json" # compiled post-correction rules
|
22 |
+
GRPO_EXPORT_PATH = "data/grpo_prefs.jsonl" # preference pairs for GRPO
|
23 |
+
CSV_EXPORT_PATH = "data/feedback.csv" # optional tabular export
|
24 |
|
25 |
# ---------------- Models ----------------
|
26 |
MODEL_PATHS = {
|
27 |
"Model 1 (Complex handwrittings )": ("prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it", Qwen2_5_VLForConditionalGeneration),
|
28 |
"Model 2 (simple and scanned handwritting )": ("nanonets/Nanonets-OCR-s", Qwen2_5_VLForConditionalGeneration),
|
29 |
}
|
30 |
+
# Model 3 removed to conserve memory.
|
31 |
|
32 |
MAX_NEW_TOKENS_DEFAULT = 512
|
33 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
|
160 |
row = json.loads(line)
|
161 |
except Exception:
|
162 |
continue
|
163 |
+
if row.get("reward", 0) < 1: # only learn from thumbs-up or explicit 'accepted_correction'
|
164 |
continue
|
165 |
pred = _safe_text(row.get("prediction", ""))
|
166 |
corr = _safe_text(row.get("correction", "")) or _safe_text(row.get("ground_truth", ""))
|
|
|
265 |
def _export_csv():
|
266 |
# optional: CSV summary for spreadsheet views
|
267 |
if not os.path.exists(FEEDBACK_PATH):
|
268 |
+
return None
|
269 |
rows = []
|
270 |
with open(FEEDBACK_PATH, "r", encoding="utf-8") as f:
|
271 |
for line in f:
|
|
|
284 |
w.writerow(flat)
|
285 |
return CSV_EXPORT_PATH
|
286 |
|
287 |
+
# ------------------- MODIFIED -------------------
|
288 |
def save_feedback(image: Image.Image, model_choice: str, prompt: str,
|
289 |
prediction: str, correction: str, ground_truth: str, reward: int):
|
290 |
"""
|
291 |
reward: 1 = good/accepted, 0 = neutral, -1 = bad
|
292 |
"""
|
293 |
if image is None:
|
294 |
+
# Bug Fix: Return a single string, not a tuple
|
295 |
+
return "Please provide the image again to link feedback."
|
296 |
if not prediction and not correction and not ground_truth:
|
297 |
+
return "Nothing to save."
|
298 |
|
299 |
image_hash = _hash_image(image)
|
300 |
# best target = correction, else ground_truth, else prediction
|
|
|
320 |
"cer": float(cer_score) if cer_score is not None else None,
|
321 |
}
|
322 |
_append_jsonl(FEEDBACK_PATH, row)
|
323 |
+
return f"β
Feedback saved (reward={reward})."
|
324 |
+
# ------------------------------------------------
|
325 |
|
326 |
def compile_memory_rules():
|
327 |
_compile_rules_from_feedback(min_count=2, max_phrase_len=60)
|
|
|
359 |
count += 1
|
360 |
return f"β
Exported {count} GRPO preference pairs to {GRPO_EXPORT_PATH}."
|
361 |
|
362 |
+
# ------------------- NEW -------------------
|
363 |
+
def get_grpo_file():
|
364 |
+
if os.path.exists(GRPO_EXPORT_PATH):
|
365 |
+
return GRPO_EXPORT_PATH
|
366 |
+
return None
|
367 |
+
|
368 |
+
def get_csv_file():
|
369 |
+
_export_csv()
|
370 |
+
if os.path.exists(CSV_EXPORT_PATH):
|
371 |
+
return CSV_EXPORT_PATH
|
372 |
+
return None
|
373 |
+
# -------------------------------------------
|
374 |
|
375 |
# ---------------- Evaluation Orchestration ----------------
|
376 |
@spaces.GPU
|
|
|
447 |
|
448 |
trainer = GRPOTrainer(
|
449 |
model=model,
|
450 |
+
ref_model=None, # let TRL create a frozen copy internally
|
451 |
args=cfg,
|
452 |
tokenizer=tok,
|
453 |
train_dataset=ds
|
|
|
469 |
|
470 |
# ---------------- Gradio Interface ----------------
|
471 |
with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
472 |
+
gr.Markdown("## βπΎ wilson Handwritten OCR β with Feedback Loop, Memory & GRPO Export")
|
473 |
|
474 |
model_choice = gr.Radio(choices=list(MODEL_PATHS.keys()),
|
475 |
value=list(MODEL_PATHS.keys())[0],
|
|
|
561 |
|
562 |
with gr.Tab("βοΈ Feedback & Memory"):
|
563 |
gr.Markdown("""
|
564 |
+
**Pipeline**
|
565 |
+
1) Save feedback (π / π) and add corrections.
|
566 |
+
2) Click **Build/Refresh Memory** to generate auto-fix rules from positive feedback.
|
567 |
3) Keep **Enable Memory Post-correction** checked on inference/eval tabs.
|
568 |
""")
|
569 |
build_mem_btn = gr.Button("π§ Build/Refresh Memory from Feedback")
|
570 |
mem_status = gr.Markdown()
|
571 |
build_mem_btn.click(fn=compile_memory_rules, outputs=[mem_status])
|
572 |
|
|
|
573 |
csv_status = gr.Markdown()
|
574 |
+
|
575 |
+
# ------------------- MODIFIED -------------------
|
576 |
+
gr.Markdown("---")
|
577 |
+
gr.Markdown("### β¬οΈ Download Feedback Data")
|
578 |
+
with gr.Row():
|
579 |
+
download_csv_btn = gr.Button("β¬οΈ Download Feedback as CSV")
|
580 |
+
download_csv_file = gr.File(label="CSV File")
|
581 |
+
download_csv_btn.click(fn=get_csv_file, outputs=download_csv_file)
|
582 |
+
# ------------------------------------------------
|
583 |
|
584 |
with gr.Tab("π§ͺ GRPO / Dataset"):
|
585 |
gr.Markdown("""
|
586 |
**GRPO Fine-tuning** (run offline or in a training Space):
|
587 |
- Click **Export GRPO Preferences** to produce `data/grpo_prefs.jsonl` of (prompt, chosen, rejected).
|
588 |
- Click **Write Trainer Script** to create `train/grpo_train.py`.
|
589 |
+
- Then run:
|
590 |
```bash
|
591 |
pip install trl accelerate peft transformers datasets
|
592 |
python train/grpo_train.py
|
593 |
+
````
|
594 |
+
|
595 |
Set `BASE_MODEL`/`OUTPUT_DIR` env vars if you like.
|
596 |
+
""")
|
597 |
+
grpo\_btn = gr.Button("π¦ Export GRPO Preferences")
|
598 |
+
grpo\_status = gr.Markdown()
|
599 |
+
grpo\_btn.click(fn=export\_grpo\_preferences, outputs=[grpo\_status])
|
600 |
+
|
601 |
+
|
602 |
+
write_script_btn = gr.Button("π Write grpo_train.py")
|
603 |
+
write_script_status = gr.Markdown()
|
604 |
+
write_script_btn.click(fn=lambda: f"β
Trainer script written to `{_write_trainer_script()}`", outputs=[write_script_status])
|
605 |
+
|
606 |
+
|
607 |
+
# \------------------- NEW -------------------
|
608 |
+
|
609 |
+
|
610 |
+
gr.Markdown("---")
|
611 |
+
gr.Markdown("### β¬οΈ Download GRPO Dataset")
|
612 |
+
with gr.Row():
|
613 |
+
download_grpo_btn = gr.Button("β¬οΈ Download GRPO Data (grpo_prefs.jsonl)")
|
614 |
+
download_grpo_file = gr.File(label="GRPO Dataset File")
|
615 |
+
download_grpo_btn.click(fn=get_grpo_file, outputs=download_grpo_file)
|
616 |
+
|
617 |
|
618 |
+
# \-------------------------------------------
|
|
|
|
|
619 |
|
620 |
if __name__ == "__main__":
|
621 |
demo.queue(max_size=50).launch(share=True)
|