Persist evals to /data CSV and upload per-submission JSONL to dataset repo (peiranli0930/VisEval); add UI feedback

app.py

@@ -1,8 +1,10 @@
 import csv
 import itertools
+import json
 import os
+import uuid
 from datetime import datetime
-import os
+from io import BytesIO
 from typing import Dict, List, Tuple
 
 import gradio as gr
@@ -13,6 +15,8 @@ except Exception:  # optional dependency at runtime
 
 
 BASE_DIR = os.path.dirname(__file__)
+# Persistent local storage inside HF Spaces
+PERSIST_DIR = os.environ.get("PERSIST_DIR", "/data")
 TASK_CONFIG = {
     "Scene Composition & Object Insertion": {
         "folder": "scene_composition_and_object_insertion",
@@ -100,11 +104,31 @@ def _format_pair_header(_pair: Dict[str, str]) -> str:
     return ""
 
 
-def _append_evaluation(task_name: str, pair: Dict[str, str], scores: Dict[str, int]) -> None:
-    csv_path = _csv_path_for_task(task_name, "evaluation_results.csv")
+def _build_eval_row(pair: Dict[str, str], scores: Dict[str, int]) -> Dict[str, object]:
+    row = {
+        "eval_date": datetime.utcnow().isoformat(),
+        "test_id": pair["test_id"],
+        "model1_name": pair["model1_name"],
+        "model2_name": pair["model2_name"],
+        "org_img": pair["org_img"],
+        "model1_res": pair["model1_res"],
+        "model2_res": pair["model2_res"],
+        "model1_path": pair["model1_path"],
+        "model2_path": pair["model2_path"],
+    }
+    row.update(scores)
+    return row
+
+
+def _local_persist_csv_path(task_name: str) -> str:
+    folder = TASK_CONFIG[task_name]["folder"]
+    return os.path.join(PERSIST_DIR, folder, "evaluation_results.csv")
+
+
+def _append_local_persist_csv(task_name: str, row: Dict[str, object]) -> bool:
+    csv_path = _local_persist_csv_path(task_name)
     os.makedirs(os.path.dirname(csv_path), exist_ok=True)
     csv_exists = os.path.exists(csv_path)
-
     fieldnames = [
         "eval_date",
         "test_id",
@@ -115,64 +139,55 @@ def _append_evaluation(task_name: str, pair: Dict[str, str], scores: Dict[str, i
         "model2_res",
         "model1_path",
         "model2_path",
-        # Per-image scores for Model A (输出A)
         "model1_physical_interaction_fidelity_score",
         "model1_optical_effect_accuracy_score",
         "model1_semantic_functional_alignment_score",
         "model1_overall_photorealism_score",
-        # Per-image scores for Model B (输出B)
         "model2_physical_interaction_fidelity_score",
         "model2_optical_effect_accuracy_score",
         "model2_semantic_functional_alignment_score",
         "model2_overall_photorealism_score",
     ]
+    try:
+        with open(csv_path, "a", newline="", encoding="utf-8") as csv_file:
+            writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
+            if not csv_exists:
+                writer.writeheader()
+            writer.writerow(row)
+        return True
+    except Exception:
+        return False
+
 
-    with open(csv_path, "a", newline="", encoding="utf-8") as csv_file:
-        writer = csv.DictWriter(csv_file, fieldnames=fieldnames)
-        if not csv_exists:
-            writer.writeheader()
-
-        row = {
-            "eval_date": datetime.utcnow().isoformat(),
-            "test_id": pair["test_id"],
-            "model1_name": pair["model1_name"],
-            "model2_name": pair["model2_name"],
-            "org_img": pair["org_img"],
-            "model1_res": pair["model1_res"],
-            "model2_res": pair["model2_res"],
-            "model1_path": pair["model1_path"],
-            "model2_path": pair["model2_path"],
-        }
-        row.update(scores)
-        writer.writerow(row)
-
-    # Optionally push updated CSV to the Space repo if credentials are available
-    _try_push_to_hub(csv_path)
-
-
-def _try_push_to_hub(csv_path: str) -> None:
-    """Attempt to commit the CSV to the current Space repo if HF_TOKEN and SPACE_ID exist.
-    Safe no-op if huggingface_hub isn't available or env vars are missing.
+def _upload_eval_record_to_dataset(task_name: str, row: Dict[str, object]) -> bool:
+    """Upload a single-eval JSONL record to a dataset repo.
+    Repo is taken from EVAL_REPO_ID env or defaults to 'peiranli0930/VisEval'.
     """
     if HfApi is None:
-        return
+        return False
     token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACEHUB_API_TOKEN")
-    space_id = os.environ.get("SPACE_ID")
-    if not token or not space_id:
-        return
+    repo_id = os.environ.get("EVAL_REPO_ID", "peiranli0930/VisEval")
+    if not token or not repo_id:
+        return False
     try:
+        from huggingface_hub import CommitOperationAdd
+
         api = HfApi(token=token)
-        rel_path = os.path.relpath(csv_path, BASE_DIR).replace(os.sep, "/")
-        api.upload_file(
-            path_or_fileobj=csv_path,
-            path_in_repo=rel_path,
-            repo_id=space_id,
-            repo_type="space",
-            commit_message=f"Update eval CSV: {rel_path} at {datetime.utcnow().isoformat()}",
+        date_prefix = datetime.utcnow().strftime("%Y-%m-%d")
+        folder = TASK_CONFIG[task_name]["folder"]
+        uid = str(uuid.uuid4())
+        path_in_repo = f"submissions/{folder}/{date_prefix}/{uid}.jsonl"
+        payload = (json.dumps(row, ensure_ascii=False) + "\n").encode("utf-8")
+        operations = [CommitOperationAdd(path_in_repo=path_in_repo, path_or_fileobj=BytesIO(payload))]
+        api.create_commit(
+            repo_id=repo_id,
+            repo_type="dataset",
+            operations=operations,
+            commit_message=f"Add eval {folder} {row.get('test_id')} {uid}",
         )
+        return True
     except Exception:
-        # Silently ignore push errors to avoid breaking the UI flow
-        pass
+        return False
 
 
 def on_task_change(task_name: str, _state_pairs: List[Dict[str, str]]):
@@ -243,10 +258,14 @@ def on_submit(
         "model2_semantic_functional_alignment_score": int(b_semantic_score),
         "model2_overall_photorealism_score": int(b_overall_score),
     }
-    _append_evaluation(task_name, pair, score_map)
+    row = _build_eval_row(pair, score_map)
+    ok_local = _append_local_persist_csv(task_name, row)
+    ok_hub = _upload_eval_record_to_dataset(task_name, row)
 
     next_index = min(index + 1, len(pairs) - 1)
     info = f"已保存 Test ID {pair['test_id']} 的评价结果。"
+    info += " 本地持久化" + ("成功" if ok_local else "失败") + "。"
+    info += " 上传Hub" + ("成功" if ok_hub else "失败") + "。"
 
     if next_index != index:
         pair = pairs[next_index]