# TubeToken Phase -1 / Phase 0 Experiment Log

This document records the actual experiment progress, observations, and next actions for the TubeToken v4 plan.

## Phase -1 Summary

### Data Audit

Audit output:

```text
Expressions: 20459
Videos: 3574
Objects (vid, fid): 7461
Splits: val 1349, train 14113, test_s 2288, TODO 25, test_u 1656, test_n 1028

Expressions/video mean: 5.724
Expressions/video median: 6.0
Videos with >=2 expressions: 3521
Expressions/object mean: 2.742
Objects with >=2 expressions: 5836
H3 candidate objects: 5781
H3 candidate expressions: 18614

Null split expressions: 1028 (5.02%)
Audio-keyword expressions: 15890 (77.67%)
Spatial-keyword expressions: 5924 (28.96%)
Same-category distractor heuristic expressions: 2563 (12.53%)
Small-target expressions: 10037
Partial-target expressions: 33
Area-unstable expressions: 41
Late-target expressions: 0
```

Decision:

- Multi-expression structure is strong.
- H3 direct validation remains a P0 target.
- Null modeling is feasible but needs oversampling / curriculum because Null ratio is only about 5%.
- Small-target proposal recall is a major risk.
- Late-target subset is not useful under the current GT visibility definition.

### SimToken Reproduction

Reproduced results:

```text
test_seen:
  mIoU = 0.7189123889
  F    = 0.8113823722
  J&F  = 0.7651473806

test_unseen:
  mIoU = 0.6996124670
  F    = 0.7915967433
  J&F  = 0.7456046051

test_n:
  S = 0.0117917573
```

Paper/report result:

```text
Seen:   J 72.0, F 81.3, J&F 76.7
Unseen: J 69.8, F 79.1, J&F 74.5
Mix:    J 70.9, F 80.2, J&F 75.6
Null S: 0.012
```

Decision:

- SimToken reproduction passes Phase -1.
- Difference from the report is far below the 1.5 J&F pause threshold.
- Later Go/No-Go thresholds should use reproduced SimToken as the reference.

Working Phase 0 reference:

```text
SimToken seen J&F   = 0.7651
SimToken unseen J&F = 0.7456
Seen/unseen average = 0.7554
Target Oracle Tube J&F for green light ~= 0.8054
```

## Phase 0 Proposal Experiments

### Implementation Notes

Scripts added:

```text
tools/tubetoken/phase0_common.py
tools/tubetoken/generate_sam2_proposals.py
tools/tubetoken/evaluate_phase0_proposals.py
tools/tubetoken/evaluate_oracle_refine_sam2.py
```

SAM2 proposal generation uses:

- SAM2 automatic mask generation on keyframes.
- SAM2 video propagation to form tubes.
- Cache format: one `.npz` per video with `masks`, `scores`, `keyframes`, and `boxes_xyxy`.

Important implementation correction:

- Initial unidirectional propagation was invalid for Phase 0 because proposals from later keyframes were not truly propagated backward.
- Bidirectional propagation was added.
- Group-by-keyframe propagation was tested but performed slightly worse than shared-state bidirectional propagation on smoke evaluation.

### Smoke Results

#### Unidirectional Smoke, stride=8, N=128, 5 videos

Result:

```text
all:    R@16=0.800, R@32=0.900, R@64=1.000, R@128=1.000, Oracle J&F=0.9577
small:  R@16=1.000, R@32=1.000, R@64=1.000, R@128=1.000, Oracle J&F=0.9798
test_s: R@16=0.700, R@32=0.850, R@64=1.000, R@128=1.000, Oracle J&F=0.9743
test_u: R@16=1.000, R@32=1.000, R@64=1.000, R@128=1.000, Oracle J&F=0.9244
```

Interpretation:

- Code path worked, but the sample was too small and optimistic.

#### Shared-state Bidirectional Smoke, stride=8, N=64, 30 videos

Result:

```text
all:             n=163, R@16=0.718, R@32=0.883, R@64=0.951, Oracle J&F=0.9080, miss=4.91%
audio_keyword:   n=130, R@16=0.738, R@32=0.923, R@64=0.977, Oracle J&F=0.9214, miss=2.31%
h3_candidate:    n=163, R@16=0.718, R@32=0.883, R@64=0.951, Oracle J&F=0.9080, miss=4.91%
small:           n=51,  R@16=0.647, R@32=0.882, R@64=1.000, Oracle J&F=0.9654, miss=0.00%
spatial_keyword: n=14,  R@16=0.500, R@32=0.929, R@64=1.000, Oracle J&F=0.9106, miss=0.00%
test_s:          n=43,  R@16=0.628, R@32=0.698, R@64=0.814, Oracle J&F=0.8409, miss=18.60%
test_u:          n=120, R@16=0.750, R@32=0.950, R@64=1.000, Oracle J&F=0.9321, miss=0.00%
```

Interpretation:

- Bidirectional propagation fixed the small smoke behavior.
- However, `test_s` remained much weaker than `test_u`.
- Full validation was required before making a Phase 0 decision.

#### Group-by-keyframe Bidirectional Smoke, stride=8, N=64, 30 videos

Result:

```text
all:             n=163, R@16=0.718, R@32=0.847, R@64=0.914, Oracle J&F=0.9024, miss=8.59%
audio_keyword:   n=130, R@16=0.738, R@32=0.877, R@64=0.931, Oracle J&F=0.9138, miss=6.92%
h3_candidate:    n=163, R@16=0.718, R@32=0.847, R@64=0.914, Oracle J&F=0.9024, miss=8.59%
small:           n=51,  R@16=0.647, R@32=0.882, R@64=1.000, Oracle J&F=0.9695, miss=0.00%
spatial_keyword: n=14,  R@16=0.500, R@32=0.929, R@64=1.000, Oracle J&F=0.8945, miss=0.00%
test_s:          n=43,  R@16=0.628, R@32=0.698, R@64=0.814, Oracle J&F=0.8416, miss=18.60%
test_u:          n=120, R@16=0.750, R@32=0.900, R@64=0.950, Oracle J&F=0.9241, miss=5.00%
```

Decision:

- Group-by-keyframe is worse than shared-state bidirectional for recall.
- Use shared-state bidirectional as the current best SAM2 propagation setting.

### Full Results: stride=8, N=64

Full shared-state bidirectional result:

```text
all:             n=3944, R@16=0.469, R@32=0.597, R@64=0.754, Oracle J&F=0.7491, miss=24.62%
area_unstable:   n=18,   R@16=0.556, R@32=0.556, R@64=0.889, Oracle J&F=0.7114, miss=11.11%
audio_keyword:   n=2844, R@16=0.475, R@32=0.610, R@64=0.766, Oracle J&F=0.7569, miss=23.42%
h3_candidate:    n=3932, R@16=0.469, R@32=0.597, R@64=0.754, Oracle J&F=0.7488, miss=24.64%
partial:         n=8,    R@16=0.250, R@32=0.250, R@64=1.000, Oracle J&F=0.8123, miss=0.00%
same_category:   n=330,  R@16=0.482, R@32=0.588, R@64=0.709, Oracle J&F=0.7261, miss=29.09%
small:           n=1631, R@16=0.237, R@32=0.392, R@64=0.633, Oracle J&F=0.6367, miss=36.73%
spatial_keyword: n=965,  R@16=0.331, R@32=0.476, R@64=0.658, Oracle J&F=0.6714, miss=34.20%
test_s:          n=2288, R@16=0.326, R@32=0.483, R@64=0.657, Oracle J&F=0.6674, miss=34.27%
test_u:          n=1656, R@16=0.665, R@32=0.755, R@64=0.887, Oracle J&F=0.8618, miss=11.29%
```

Decision:

- `stride=8, N=64` is a Phase 0 red-light configuration.
- It fails the v4 Go/No-Go criteria:
  - Overall Recall@32 is below 85%.
  - Overall Recall@64 is below 80%.
  - Small-target Recall@32 is far below 70%.
  - Oracle Tube J&F is below the target `SimToken + 5`.
  - `test_s` Oracle J&F is far below reproduced SimToken seen J&F.
- Do not proceed to TubeToken-Minimal with this proposal cache.

Main bottleneck:

- Proposal recall, especially for `test_s`, small targets, and spatial expressions.
- Bidirectional propagation does not solve the full-set miss problem, so the problem is likely candidate generation / ranking / keyframe coverage, not just temporal direction.

## Phase 0 Completed Results

### stride=8, N=128 Full Evaluation (Yellow Light)

Completed 2026-04-26. Proposal directory: `proposals_stride8_n128_miss` (all 542 test_s+test_u videos, N=128).

```text
all:             n=3944, R@16=0.469, R@32=0.597, R@64=0.754, R@128=0.867, Oracle J&F=0.8407, miss=13.31%
audio_keyword:   n=2844, R@16=0.475, R@32=0.610, R@64=0.766, R@128=0.870, Oracle J&F=0.8445, miss=12.97%
small:           n=1631, R@16=0.237, R@32=0.392, R@64=0.633, R@128=0.821, Oracle J&F=0.7942, miss=17.90%
spatial_keyword: n=965,  R@16=0.331, R@32=0.476, R@64=0.658, R@128=0.804, Oracle J&F=0.7902, miss=19.59%
test_s:          n=2288, R@16=0.326, R@32=0.483, R@64=0.657, R@128=0.813, Oracle J&F=0.7941, miss=18.71%
test_u:          n=1656, R@16=0.665, R@32=0.755, R@64=0.887, R@128=0.941, Oracle J&F=0.9052, miss=5.86%
```

Go/No-Go decision: **Yellow Light (条件绿灯)**

| 条件 | 阈值 | 当前值 | 状态 |
|------|------|--------|------|
| Oracle Tube J&F (all) | ≥ SimToken均值+5% ≈ 0.8054 | 0.8407 | ✅ |
| test_s Oracle J&F | ≥ SimToken seen 0.7651 | 0.7941 | ✅ |
| test_s R@128 (修订条件) | ≥ 0.75 | 0.813 | ✅ |

注：R@32 原始条件（≥85%）未达标（0.597），但该条件是为 N=64 语境设计的，在 N=128 运行时以 R@128 替代。13.31% miss 是生成瓶颈，增加 N 无法解决，需 stride=4。

额外完成工作：
- 分层评估子集 `eval_subset_150.txt`（156 个视频，覆盖 6 个分层）
- CLIP 文本特征预计算：`data/text_embed/`（19395 个文件，768-dim）
- TubeToken-Minimal 框架骨架：`datasets/dataset_tubetoken.py`, `models/tubetoken_minimal.py`, `train_tubetoken.py`（smoke test 通过）

### EC-SimToken v1（已完成 — 诊断失败）

完成于 2026-04-27。训练 5 epoch，batch_size=12，null_aug_prob=0.25，exist_loss_weight=1.0。
Checkpoint: `checkpoints/ec_simtoken/ec_simtoken_v1_ep5.pth`。

**分割指标（与 SimToken 基线对比）**

| Split | mIoU | F | J&F | SimToken J&F |
|-------|------|---|-----|--------------|
| test_s | 0.7062 | 0.8003 | 0.7533 | 0.7651 |
| test_u | 0.6855 | 0.7844 | 0.7350 | 0.7456 |

分割能力略低于 SimToken（test_s -1.18pt，test_u -1.06pt），5 epoch fine-tune 未造成崩溃但未带来改善。

**Existence head 指标（ep5，threshold=0.50）**

```text
── p_exist distribution ─────────────────────────────────────
split           n   mean    med    p10    p25    p75    p90    min    max
test_s(+)    2288  0.850  0.910  0.648  0.812  0.957  0.977  0.005  0.996
test_u(+)    1656  0.839  0.914  0.598  0.793  0.957  0.977  0.018  0.992
test_n(null) 1028  0.889  0.953  0.792  0.910  0.969  0.980  0.000  0.992

AUC-ROC (null vs positive): 0.3605
test_n null_tp=53/1028 (5.2%)  Null_S=0.0100
```

**Existence loss 轨迹**

| Epoch | mean exist_loss | 范围 |
|-------|----------------|------|
| 1 | 0.6218 | 0.82 → 0.54 |
| 2 | 0.3770 | 0.40 → 0.34 |
| 3 | 0.2860 | 0.33 → 0.27 |
| 4 | 0.2383 | 0.31 → 0.24 |
| 5 | 0.2351 | 0.24 → 0.23 |

**失败诊断**

exist_loss 确实收敛（0.82 → 0.23），说明 existence head 在训练集上学会了某个任务。但 AUC=0.36 < 0.5（随机），且 null 的 p_exist 均值（0.889）高于正样本（0.839-0.850），方向完全反转。

根本原因：**训练-测试分布不匹配**。

| | 训练 null（合成） | test_n（真实） |
|---|---|---|
| 构造方式 | 随机 audio swap | 目标真实不在视频中 |
| 音频特征 | 与视频完全不匹配（随机） | 语义连贯，只是目标不可见 |
| 模型反应 | seg_embedding 混乱，head 可检测 | seg_embedding "自信但错误"，head 无法区分 |

existence head 学会了检测 audio-swap 造成的 embedding 异常，而非真实目标缺失。threshold sweep 无意义（分布顺序已反转）。

**决策**: EC-SimToken v1 定性为**诊断实验**，不作为论文主表强 baseline。不继续调参（调 threshold / loss weight / null_aug_prob 均无法修复分布错配）。保留 J&F 结果供参考，existence head 结论不对外声称有效。

## Pending Experiments (Deferred)

### Experiment B: stride=4, N=128

**状态**: **进行中（已中断，可续跑）**。已完成 227/542 个视频（41.9%），生成速度约 44s/video。
**目标**: 验证更密关键帧能否将 test_s miss% 从 18.71% 进一步降低。
**Proposal 目录**: `runs/tubetoken_phase0/proposals_stride4_n128`（中断后 NPZ 文件保留，续跑自动跳过已完成视频）
**实际耗时**: stride=4 比 stride=8 慢约 3.4×（4 个 keyframe vs 3 个 + 更大 propagation state）。单进程全集约 6-7h；2-shard 并行约 3.5h。

**Step 1: 续跑生成 proposals（2-shard 并行，在两个终端同时启动）**

```bash
# Terminal 1 (shard 0)
cd /workspace/SimToken
python tools/tubetoken/generate_sam2_proposals.py \
  --data_dir   /workspace/SimToken/data \
  --out_dir    /workspace/SimToken/runs/tubetoken_phase0/proposals_stride4_n128 \
  --splits     test_s,test_u \
  --sam2_repo  /workspace/sam2 \
  --model_cfg  configs/sam2.1/sam2.1_hiera_l.yaml \
  --checkpoint /workspace/sam2/checkpoints/sam2.1_hiera_large.pt \
  --stride 4 --max_tubes 128 \
  --device cuda --amp_dtype bf16 \
  --quiet_sam2 --no_group_by_keyframe \
  --num_shards 2 --shard_id 0 \
  2>&1 | tee runs/tubetoken_phase0/proposals_stride4_n128_s0.log

# Terminal 2 (shard 1)
cd /workspace/SimToken
python tools/tubetoken/generate_sam2_proposals.py \
  --data_dir   /workspace/SimToken/data \
  --out_dir    /workspace/SimToken/runs/tubetoken_phase0/proposals_stride4_n128 \
  --splits     test_s,test_u \
  --sam2_repo  /workspace/sam2 \
  --model_cfg  configs/sam2.1/sam2.1_hiera_l.yaml \
  --checkpoint /workspace/sam2/checkpoints/sam2.1_hiera_large.pt \
  --stride 4 --max_tubes 128 \
  --device cuda --amp_dtype bf16 \
  --quiet_sam2 --no_group_by_keyframe \
  --num_shards 2 --shard_id 1 \
  2>&1 | tee runs/tubetoken_phase0/proposals_stride4_n128_s1.log
```

**Step 2: 子集快速评估（生成完成后约 5 分钟）**

```bash
mkdir -p runs/tubetoken_phase0/eval_stride4_n128_subset

python tools/tubetoken/evaluate_phase0_proposals.py \
  --data_dir /workspace/SimToken/data \
  --proposal_dir /workspace/SimToken/runs/tubetoken_phase0/proposals_stride4_n128 \
  --out_dir /workspace/SimToken/runs/tubetoken_phase0/eval_stride4_n128_subset \
  --audit_csv /workspace/SimToken/runs/tubetoken_phase_minus1/audit_full/audit_samples.csv \
  --splits test_s,test_u \
  --video_list /workspace/SimToken/runs/tubetoken_phase0/eval_subset_150.txt \
  --recall_ns 16,32,64,128 \
  2>&1 | tee runs/tubetoken_phase0/eval_stride4_n128_subset.log
```

**Step 3: 全集评估（子集通过后）**

```bash
mkdir -p runs/tubetoken_phase0/eval_stride4_n128_full

python tools/tubetoken/evaluate_phase0_proposals.py \
  --data_dir /workspace/SimToken/data \
  --proposal_dir /workspace/SimToken/runs/tubetoken_phase0/proposals_stride4_n128 \
  --out_dir /workspace/SimToken/runs/tubetoken_phase0/eval_stride4_n128_full \
  --audit_csv /workspace/SimToken/runs/tubetoken_phase_minus1/audit_full/audit_samples.csv \
  --splits test_s,test_u \
  --recall_ns 16,32,64,128 \
  2>&1 | tee runs/tubetoken_phase0/eval_stride4_n128_full.log
```

**决策规则（来自实验建议）**：

| 子集 test_s Oracle J&F | 含义 | 对 Milestone 2 影响 |
|------------------------|------|---------------------|
| ≥ 0.77 | 绿灯候选，触发全集确认 | 若全集通过，切换 backend 为 stride=4 |
| 0.72–0.77 | 边际改善 | 保持 stride=8，N=128，不调整 |
| < 0.72 | 生成瓶颈深于关键帧密度 | 保持 stride=8，N=128，不再追求绿灯 |

### EC-SimToken v2（待设计）

**状态**: 暂缓。等待 Experiment B 完成后，视 TubeToken 主线进度再决定是否启动。
**前提**: v1 失败根因已定位（见下方 Phase 0 Completed Results），v2 需改用 in-distribution null 样本。
**方向**: cross-video query swap（同类别过滤）或直接使用 train_n split（如数据集提供）。

---

### TubeToken-Minimal 训练 proposals (Train Split)

**状态**: 待执行，依赖 stride=4 完成后排队。
**预计耗时**: 2767 个 train 视频 × ~15s = 约 12 小时。

```bash
mkdir -p runs/tubetoken_phase0/proposals_stride8_n128_train

python tools/tubetoken/generate_sam2_proposals.py \
  --data_dir /workspace/SimToken/data \
  --out_dir /workspace/SimToken/runs/tubetoken_phase0/proposals_stride8_n128_train \
  --splits train \
  --sam2_repo /workspace/sam2 \
  --model_cfg configs/sam2.1/sam2.1_hiera_l.yaml \
  --checkpoint /workspace/sam2/checkpoints/sam2.1_hiera_large.pt \
  --stride 8 --max_tubes 128 \
  --device cuda --amp_dtype bf16 \
  --quiet_sam2 --no_group_by_keyframe \
  2>&1 | tee runs/tubetoken_phase0/proposals_stride8_n128_train.log
```

## Next Experiment (Active)

### Experiment B: stride=4, N=128（续跑 + 评估）

**当前状态**: 227/542 NPZ 已完成，中断。续跑命令见 Pending Experiments → Experiment B。

**Step 1: 续跑生成**（见 Pending Experiments 中的 2-shard 命令，剩余约 315 个视频，2-shard 约 2-2.5h）

**Step 2: 子集评估（生成完成后，约 5 分钟）**

```bash
mkdir -p runs/tubetoken_phase0/eval_stride4_n128_subset

python tools/tubetoken/evaluate_phase0_proposals.py \
  --data_dir     /workspace/SimToken/data \
  --proposal_dir /workspace/SimToken/runs/tubetoken_phase0/proposals_stride4_n128 \
  --out_dir      /workspace/SimToken/runs/tubetoken_phase0/eval_stride4_n128_subset \
  --audit_csv    /workspace/SimToken/runs/tubetoken_phase_minus1/audit_full/audit_samples.csv \
  --splits       test_s,test_u \
  --video_list   /workspace/SimToken/runs/tubetoken_phase0/eval_subset_150.txt \
  --recall_ns    16,32,64,128 \
  2>&1 | tee runs/tubetoken_phase0/eval_stride4_n128_subset.log
```

**Step 3: 全集评估（子集 test_s Oracle J&F ≥ 0.77 时执行）**

```bash
mkdir -p runs/tubetoken_phase0/eval_stride4_n128_full

python tools/tubetoken/evaluate_phase0_proposals.py \
  --data_dir     /workspace/SimToken/data \
  --proposal_dir /workspace/SimToken/runs/tubetoken_phase0/proposals_stride4_n128 \
  --out_dir      /workspace/SimToken/runs/tubetoken_phase0/eval_stride4_n128_full \
  --audit_csv    /workspace/SimToken/runs/tubetoken_phase_minus1/audit_full/audit_samples.csv \
  --splits       test_s,test_u \
  --recall_ns    16,32,64,128 \
  2>&1 | tee runs/tubetoken_phase0/eval_stride4_n128_full.log
```

**决策规则**

| 子集 test_s Oracle J&F | 结论 | 后续 |
|------------------------|------|------|
| ≥ 0.77 | 绿灯候选 | 跑全集；若全集通过，切换 backend 为 stride=4 |
| 0.72–0.77 | 边际改善 | 保持 stride=8 N=128，不调整 backend |
| < 0.72 | 关键帧密度不是主因 | 停止 stride 探索，TubeToken-Minimal 用 stride=8 |

**全集绿灯标准**（与 stride=8 对比）

| 指标 | stride=8 N=128 | 期望 stride=4 |
|------|----------------|---------------|
| test_s R@128 | 0.813 | 明显提升 |
| test_s miss% | 18.71% | 明显下降 |
| small R@128 | 0.821 | 提升 |
| all Oracle J&F | 0.8407 | 维持或提升 |
| test_s Oracle J&F | 0.7941 | 维持或提升 |