Upload residual heatmap guide asset: residual_heatmap_generation_and_upload.md

Reward_finetuning/SAGE-GRPO/docs/residual_heatmap_generation_and_upload.md

@@ -0,0 +1,448 @@
+# Residual / Heatmap Visualization and HF Upload Guide
+
+This guide documents how the token-credit heatmap and residual examples were generated and uploaded to `YFanwang/Backup`.
+
+Run all commands from the SAGE-GRPO repo root:
+
+```bash
+cd /nfs/ywang29/Reward_finetuning/SAGE-GRPO
+```
+
+## 1. What Is Being Visualized
+
+The input signal is the saved VLM visual-gradient credit map in each reward record:
+
+- `credit_map`: flattened visual credit values.
+- `credit_map_shape`: `[T, H, W]`.
+- In the current saved samples, a common native shape is `10 x 2 x 4`.
+
+The renderer normalizes this credit map, upsamples it spatially, and then compares three aggregation levels:
+
+- `Fine-map`: normalized `credit_map`, resized to `spatial_bins x spatial_bins` per frame. We usually use `32 x 32`.
+- `Window3`: `Fine-map` averaged with a `3 x 3` spatial window independently per frame, then renormalized.
+- `Window7`: `Fine-map` averaged with a `7 x 7` spatial window independently per frame, then renormalized.
+
+The heatmap rows are overlays of `Fine-map`, `Window3`, and `Window7` on the selected video frame. They use the same color scale within a panel:
+
+```text
+vmin = 0
+vmax = 98.5th percentile over [fine, window3, window7]
+colormap = turbo
+alpha = 0.60
+```
+
+The residual rows show positive differences:
+
+```text
+Fine > Window3    = max(Fine-map - Window3, 0)
+Window3 > Window7 = max(Window3 - Window7, 0)
+```
+
+Interpretation:
+
+- `Fine > Window3` highlights extra local fine-map credit that the `3 x 3` window suppresses. This is useful for showing fine-map noise or overly local sensitivity.
+- `Window3 > Window7` highlights local detail preserved by `3 x 3` but washed out by `7 x 7`. This is useful for arguing why `3 x 3` can be empirically better than a coarser window.
+
+The representative frame is selected by residual mass:
+
+```text
+top_frame = argmax_t sum_pixels(max(Fine - Window3, 0) + 1.5 * max(Window3 - Window7, 0))
+```
+
+The main script implementing this is:
+
+```text
+scripts/select_spatial_two_object_credit_examples.py
+```
+
+Its map construction comes from:
+
+```text
+scripts/visualize_finemap_noise_tradeoff_examples.py
+```
+
+## 2. Generate Panels From Existing `rerun_records.jsonl`
+
+Use this when you already have rerun payloads under:
+
+```text
+outputs/rerun_vlm_same_gradient_window_smoothing/*/rerun_records.jsonl
+```
+
+Example:
+
+```bash
+python scripts/select_spatial_two_object_credit_examples.py \
+  --input-jsonl outputs/rerun_vlm_same_gradient_window_smoothing/window3_no_pred_first100_maxpix200704/rerun_records.jsonl \
+               outputs/rerun_vlm_same_gradient_window_smoothing/window3_no_pred_next200_maxpix200704/rerun_records.jsonl \
+               outputs/rerun_vlm_same_gradient_window_smoothing/window3_no_pred_remaining195_maxpix200704/rerun_records.jsonl \
+  --output-dir outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled \
+  --num-candidates 500 \
+  --tile-width 260 \
+  --contact-page-size 10 \
+  --contact-thumb-width 1700 \
+  --heat-blur 4 \
+  --residual-blur 7 \
+  --fill-mode object \
+  --min-object-count 2 \
+  --title-prefix "Spatial two-object prompt"
+```
+
+Important outputs:
+
+```text
+<output-dir>/comparison_panels/
+<output-dir>/by_category_top1/original/
+<output-dir>/by_category_top1/fine-map/
+<output-dir>/by_category_top1/window3/
+<output-dir>/by_category_top1/window7/
+<output-dir>/by_category_top1/fine_gt_window3/
+<output-dir>/by_category_top1/window3_gt_window7/
+<output-dir>/candidate_manifest.csv
+<output-dir>/candidate_contact_page_*.png
+```
+
+For the previously generated batch, expected counts were:
+
+```bash
+find outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled/comparison_panels -type f -name '*.png' | wc -l
+find outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled -maxdepth 1 -type f -name 'candidate_contact_page_*.png' | wc -l
+```
+
+Expected:
+
+```text
+500
+50
+```
+
+## 3. Generate Object-Specific Payloads From `rl_samples`
+
+Use this when the rerun pool is too small and you want more examples with explicit object requirements. This route reads saved training reward files, so it does not rerun the VLM.
+
+Source directory:
+
+```text
+outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples
+```
+
+Each `*_rewards.json` file is a list of samples. Each sample contains:
+
+```text
+prompt
+video_path
+sample_idx
+reward_records
+```
+
+Each reward record contains:
+
+```text
+question_id
+question
+target_answer
+pred_sentence
+credit_map
+credit_map_shape
+```
+
+Build a renderer-ready payload pool:
+
+```bash
+python scripts/build_object_requirement_credit_payloads.py \
+  --repo-root . \
+  --rl-samples-root outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples \
+  --output-jsonl outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500/object_requirement_payloads.jsonl \
+  --manifest-csv outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500/object_requirement_payload_manifest.csv \
+  --use-default-excludes \
+  --min-question-object-count 1 \
+  --min-prompt-object-count 2
+```
+
+This filters for:
+
+- reward records that contain valid `credit_map` and `credit_map_shape`;
+- atomic questions with at least one object term;
+- source prompts with at least two object terms;
+- non-camera-only questions, excluding obvious camera/style/lighting questions;
+- keys not already used in earlier selected manifests when `--use-default-excludes` is set.
+
+If the payload pool is very large, take a high-scoring subset before rendering:
+
+```bash
+python scripts/build_object_requirement_credit_payloads.py \
+  --repo-root . \
+  --rl-samples-root outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples \
+  --output-jsonl outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object/object_requirement_payloads_top5000.jsonl \
+  --manifest-csv outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object/object_requirement_payloads_top5000_manifest.csv \
+  --use-default-excludes \
+  --min-question-object-count 1 \
+  --min-prompt-object-count 2 \
+  --top-k 5000
+```
+
+Then render 500 panels from that payload subset:
+
+```bash
+python scripts/select_spatial_two_object_credit_examples.py \
+  --input-jsonl outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object/object_requirement_payloads_top5000.jsonl \
+  --output-dir outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object \
+  --num-candidates 500 \
+  --tile-width 260 \
+  --contact-page-size 10 \
+  --contact-thumb-width 1700 \
+  --heat-blur 4 \
+  --residual-blur 7 \
+  --fill-mode object \
+  --min-object-count 2 \
+  --title-prefix "Object-specific prompt"
+```
+
+Expected final checks:
+
+```bash
+find outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object/comparison_panels -type f -name '*.png' | wc -l
+find outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object -maxdepth 1 -type f -name 'candidate_contact_page_*.png' | wc -l
+
+for d in outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object/by_category_top1/*; do
+  [ -d "$d" ] || continue
+  printf '%s ' "$(basename "$d")"
+  find "$d" -type f -name '*.png' | wc -l
+done
+```
+
+Expected:
+
+```text
+comparison panels: 500
+contact pages: 50
+each by_category_top1 category: 500
+```
+
+## 4. Visual Settings to Adjust
+
+The most useful knobs are:
+
+- `--spatial-bins 32`: display resolution for credit maps. Larger values make overlays look finer but do not create new gradient information.
+- `--heat-blur 4`: visual blur on `Fine-map`, `Window3`, and `Window7` overlay rows.
+- `--residual-blur 7`: visual blur on residual rows. Increase slightly if residual blocks look too square.
+- `--residual-smooth-window 1`: optional pre-display smoothing on residual maps. Keep this at `1` if you want residuals to stay local.
+- `--tile-width 260`: width of each tile in a comparison panel.
+- `--contact-page-size 10`: number of panels per contact sheet.
+
+Use `--fill-mode none` if you want only strict spatial prompts. Use `--fill-mode object` if you want to fill the batch with object-specific prompts when explicit left/right or foreground/background language is not enough.
+
+## 5. Upload to Hugging Face
+
+The target used here is the Hugging Face dataset repo:
+
+```text
+YFanwang/Backup
+```
+
+The token is read from:
+
+```text
+/nfs/ywang29/hf_token.txt
+```
+
+Do not print the token. Read it inside Python:
+
+```python
+token = Path("/nfs/ywang29/hf_token.txt").read_text().strip()
+```
+
+For small directories, `upload_folder` is enough:
+
+```bash
+python - <<'PY'
+from pathlib import Path
+from huggingface_hub import HfApi
+
+repo_id = "YFanwang/Backup"
+repo_type = "dataset"
+token = Path("/nfs/ywang29/hf_token.txt").read_text().strip()
+api = HfApi(token=token)
+
+local = Path("/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object")
+remote = "Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object"
+
+info = api.upload_folder(
+    repo_id=repo_id,
+    repo_type=repo_type,
+    folder_path=str(local),
+    path_in_repo=remote,
+    commit_message=f"Upload {local.name}",
+    token=token,
+    ignore_patterns=[".DS_Store", "__pycache__/*", "*.pyc"],
+)
+print(info.commit_url)
+PY
+```
+
+For large directories or many small PNGs, use chunked commits. This avoids the `504 Gateway Time-out` that can happen when one commit contains thousands of files.
+
+```bash
+python - <<'PY'
+from pathlib import Path
+from huggingface_hub import HfApi, CommitOperationAdd
+from huggingface_hub.utils import disable_progress_bars
+import hashlib
+import json
+import os
+import time
+
+os.environ["HF_HUB_DISABLE_PROGRESS_BARS"] = "1"
+disable_progress_bars()
+
+repo_id = "YFanwang/Backup"
+repo_type = "dataset"
+token = Path("/nfs/ywang29/hf_token.txt").read_text().strip()
+api = HfApi(token=token)
+
+state_path = Path("/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/hf_backup_upload_state_manual.json")
+
+items = [
+    (
+        Path("/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled"),
+        "Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled",
+    ),
+    (
+        Path("/nfs/ywang29/Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object"),
+        "Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object",
+    ),
+]
+
+def load_state():
+    return json.loads(state_path.read_text()) if state_path.exists() else {"done": {}}
+
+def save_state(state):
+    state_path.parent.mkdir(parents=True, exist_ok=True)
+    tmp = state_path.with_suffix(".tmp")
+    tmp.write_text(json.dumps(state, indent=2, sort_keys=True))
+    tmp.replace(state_path)
+
+def iter_files(root):
+    for path in sorted(root.rglob("*")):
+        if not path.is_file():
+            continue
+        if "__pycache__" in path.parts or path.name == ".DS_Store" or path.suffix == ".pyc":
+            continue
+        yield path, path.relative_to(root).as_posix(), path.stat().st_size
+
+def make_chunks(files, max_files=500, max_bytes=120 * 1024 * 1024):
+    chunk, size = [], 0
+    for item in files:
+        if chunk and (len(chunk) >= max_files or size + item[2] > max_bytes):
+            yield chunk
+            chunk, size = [], 0
+        chunk.append(item)
+        size += item[2]
+    if chunk:
+        yield chunk
+
+def chunk_id(remote_root, chunk):
+    h = hashlib.sha1(remote_root.encode())
+    for _path, rel, size in chunk:
+        h.update(rel.encode())
+        h.update(str(size).encode())
+    return h.hexdigest()[:16]
+
+state = load_state()
+done = state.setdefault("done", {})
+
+for local_root, remote_root in items:
+    chunks = list(make_chunks(list(iter_files(local_root))))
+    print(f"{local_root} -> {remote_root}: {len(chunks)} chunks")
+    for idx, chunk in enumerate(chunks, start=1):
+        cid = chunk_id(remote_root, chunk)
+        if done.get(cid):
+            print(f"skip {remote_root} {idx}/{len(chunks)}")
+            continue
+        ops = [
+            CommitOperationAdd(
+                path_in_repo=f"{remote_root}/{rel}",
+                path_or_fileobj=str(path),
+            )
+            for path, rel, _size in chunk
+        ]
+        print(f"upload {remote_root} {idx}/{len(chunks)} files={len(chunk)}")
+        last_error = None
+        for attempt in range(1, 4):
+            try:
+                info = api.create_commit(
+                    repo_id=repo_id,
+                    repo_type=repo_type,
+                    operations=ops,
+                    commit_message=f"Upload backup chunk {idx}/{len(chunks)}: {Path(remote_root).name}",
+                    token=token,
+                )
+                done[cid] = {"commit": info.oid, "files": len(chunk), "remote_root": remote_root}
+                save_state(state)
+                print(f"done {remote_root} {idx}/{len(chunks)} {info.oid}")
+                break
+            except Exception as exc:
+                last_error = repr(exc)
+                print(f"retry {remote_root} {idx}/{len(chunks)} attempt={attempt}: {type(exc).__name__}")
+                time.sleep(15 * attempt)
+        else:
+            raise RuntimeError(last_error)
+PY
+```
+
+## 6. Verify Upload
+
+Avoid listing the whole repo if it is large. Verify only the uploaded subpath:
+
+```bash
+python - <<'PY'
+from pathlib import Path
+from huggingface_hub import HfApi
+
+repo_id = "YFanwang/Backup"
+repo_type = "dataset"
+token = Path("/nfs/ywang29/hf_token.txt").read_text().strip()
+api = HfApi(token=token)
+
+paths = [
+    "Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled",
+    "Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object",
+    "YFanwang",
+]
+
+for path in paths:
+    items = list(api.list_repo_tree(
+        repo_id=repo_id,
+        repo_type=repo_type,
+        token=token,
+        path_in_repo=path,
+        recursive=True,
+    ))
+    files = [item for item in items if item.__class__.__name__ == "RepoFile"]
+    print(path, len(files))
+PY
+```
+
+For the completed upload, the verified counts were:
+
+```text
+spatial_two_object_credit_examples_500_filled: 3551
+object_requirement_credit_examples_more500_strict_object: 3552
+YFanwang: 897
+```
+
+## 7. Remote Locations Used
+
+The uploaded files are under:
+
+```text
+https://huggingface.co/datasets/YFanwang/Backup/tree/main
+```
+
+Repo paths:
+
+```text
+Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/spatial_two_object_credit_examples_500_filled
+Reward_finetuning/SAGE-GRPO/outputs/rerun_vlm_same_gradient_window_smoothing/object_requirement_credit_examples_more500_strict_object
+YFanwang
+```