Update README with verification status of EasyCom components and add results from recent experiments comparing open-stream replay and causal-prefix methods. Introduce new metrics CSV and detailed README for the EasyCom results, including current performance metrics and required follow-up actions before publication.

README.md

@@ -30,17 +30,39 @@ classification head.
 | File | Status | Purpose |
 |---|---|---|
 | `smoke_test.py` | Verified on one NVIDIA L20 | Text and 12-second EasyCom audio-video inference |
-| `easycom_manifest.py` | Implemented, awaiting dataset-side run | Discover one-minute clips, preserve session timecodes, and check 5 Hz targets |
+| `easycom_manifest.py` | Verified on EasyCom | Discover one-minute clips, preserve session timecodes, and check 5 Hz targets |
 | `qwen_omni_features.py` | Implemented, not executed here | Load only the 3B Thinker and extract fused states from one AV window |
-| `streaming_turn_model.py` | Implemented | Produce one causal Omni state every 200 ms, with causal-prefix and open-stream modes |
+| `streaming_turn_model.py` | Verified on EasyCom | Produce one causal Omni state every 200 ms, with causal-prefix and open-stream modes |
-| `TurnPredictionHead` | Defined, untrained | Map `[2048]` to `[10, 3]` logits |
+| `TurnPredictionHead` | Trained and evaluated | Map `[2048]` to `[10, 3]` logits |
-| `train_turn_head.py` | Implemented | Train the frozen-Omni anticipation head with the official EasyCom split |
+| `train_turn_head.py` | Verified on EasyCom | Train the frozen-Omni anticipation head with the official EasyCom split |
-| `infer_turn_head.py` | Implemented | Save EgoSpeak-format predictions and reproduce per-horizon AP/mAP |
+| `infer_turn_head.py` | Verified on EasyCom | Save EgoSpeak-format predictions and reproduce per-horizon AP/mAP |
 
-The official full model, Thinker input path, embedded audio, video decoding,
+The full Thinker input path, embedded audio, direct video decoding,
-and FlashAttention-2 have passed a 12-second EasyCom smoke test on one NVIDIA
+FlashAttention-2, 5 Hz streaming extraction, cache replay, target alignment,
-L20. Streaming extraction, cache equivalence, target alignment, training, and
+head training, and EgoSpeak-compatible evaluation have now been exercised on
-paper metrics remain unverified.
+EasyCom. Ego4D feature extraction and multi-seed EasyCom evaluation remain in
+progress.
+
+## Current EasyCom results
+
+The current single-run comparison was recorded on 2026-08-19. Both modes use
+the frozen Qwen2.5-Omni-3B Thinker, 2 FPS video, 5 Hz predictions, a 20-second
+causal context budget, the same `10 x 3` head architecture, and EgoSpeak's
+official EasyCom train/test split.
+
+| Representation | EgoSpeak full-timeline mean mAP | Strict causal mean mAP |
+|---|---:|---:|
+| Open-stream replay | 58.316 | 58.326 |
+| Causal-prefix | 58.830 | 58.841 |
+
+Causal-prefix is 0.514 mAP points higher in the exact Table 2 protocol, but
+the difference is concentrated at 0.2--0.6 seconds. Open-stream is slightly
+higher at 1.6--2.0 seconds. These are not yet final paper values because the
+head-training variance has not been measured across multiple seeds.
+
+The full per-horizon and per-class snapshot, machine-readable CSV, provenance,
+and follow-up checks are in
+[`results/easycom_open_stream_vs_causal_prefix_2026-08-19/`](results/easycom_open_stream_vs_causal_prefix_2026-08-19/README.md).
 
 ## Why Omni-only instead of Omni + RNN
 

results/easycom_open_stream_vs_causal_prefix_2026-08-19/README.md

@@ -0,0 +1,86 @@
+# EasyCom: open-stream replay versus causal-prefix
+
+Recorded on **2026-08-19**. These are the current single-run results reported
+from the remote EasyCom experiments. They are an experiment snapshot, not yet
+the final multi-seed paper numbers.
+
+## Experiment contract
+
+- Dataset: EasyCom using EgoSpeak's official split.
+- Train sessions: 7, 11, 10, 6, 9, 4, 8, 5, and 12.
+- Test sessions: 3, 2, and 1.
+- Model: frozen Qwen2.5-Omni-3B Thinker plus the trainable `10 x 3` head.
+- Prediction rate: 5 Hz, one state every 0.2 seconds.
+- Video sampling: 2 FPS.
+- Causal context budget: 20 seconds.
+- Targets: ten future horizons from 0.2 through 2.0 seconds.
+- Metric: frame AP for all three classes, followed by per-horizon mAP and the
+  mean over all ten horizons.
+
+The two representations are trained and evaluated separately. A head trained
+on one representation must not be reused on the other.
+
+## Summary
+
+All values below are percentages.
+
+| Representation | EgoSpeak full-timeline mean mAP | Strict causal mean mAP | Difference from open stream |
+|---|---:|---:|---:|
+| Open-stream replay | 58.316 | 58.326 | — |
+| Causal-prefix | 58.830 | 58.841 | +0.514 / +0.515 |
+
+For context, the EgoSpeak Table 2 EasyCom reference values currently recorded
+in the project README are 58.7 for Transformer A+V and 60.6 for GRU A+V.
+Accordingly, causal-prefix is 0.130 points above the Transformer reference,
+whereas open-stream replay is 0.384 points below it. Neither reaches the GRU
+A+V result.
+
+## Full-timeline result by horizon
+
+| Horizon | Time | Open-stream mAP | Causal-prefix mAP | Causal minus open |
+|---:|---:|---:|---:|---:|
+| 1 | 0.2 s | 69.756 | 72.149 | +2.393 |
+| 2 | 0.4 s | 65.163 | 66.904 | +1.741 |
+| 3 | 0.6 s | 61.578 | 62.468 | +0.890 |
+| 4 | 0.8 s | 58.783 | 59.214 | +0.431 |
+| 5 | 1.0 s | 56.913 | 57.037 | +0.124 |
+| 6 | 1.2 s | 55.727 | 55.789 | +0.062 |
+| 7 | 1.4 s | 54.819 | 54.845 | +0.026 |
+| 8 | 1.6 s | 53.957 | 53.905 | -0.052 |
+| 9 | 1.8 s | 53.459 | 53.265 | -0.194 |
+| 10 | 2.0 s | 53.002 | 52.723 | -0.279 |
+
+The complete per-class AP values for both evaluation protocols are stored in
+[`metrics.csv`](metrics.csv). [`summary.json`](summary.json) contains the
+machine-readable aggregate comparison.
+
+## Current interpretation
+
+1. Causal-prefix is 0.514 mAP points better overall in the exact EgoSpeak
+   full-timeline protocol.
+2. Almost all of that advantage occurs at 0.2--0.6 seconds. At 1.2--1.4
+   seconds the difference is negligible, and open-stream is slightly better
+   at 1.6--2.0 seconds.
+3. The full-timeline and strict-causal means differ by only 0.010--0.011
+   points, so unavailable pre-session rows are not driving the comparison.
+4. The short-horizon pattern is consistent with block-local audio processing,
+   partial-block AV/TMRoPE organization, or other 2-second boundary effects
+   weakening precise speech-onset/offset cues. It does not look like a general
+   loss of longer-horizon conversational information.
+5. Previous smoke tests showed exact open-stream cache replay but a lower
+   cosine against the continuous-window representation. This supports a real
+   representation-semantics difference rather than a cache-copy error.
+
+## Required follow-up before publication
+
+- Repeat head training with at least three identical seeds for both modes and
+  report mean and standard deviation.
+- Group feature cosine and AP/mAP by the ten 0.2-second phases inside each
+  2-second block to test for a block-boundary effect.
+- Compare samples before and after the 20-second memory budget fills to isolate
+  input segmentation from rolling-memory behavior.
+- Archive the original remote `metrics.json`, checkpoint configuration, git
+  commit, Transformers version, and Qwen model revision alongside the final
+  paper results. The numbers in this snapshot were transcribed from the
+  reported console tables.
+

results/easycom_open_stream_vs_causal_prefix_2026-08-19/metrics.csv

@@ -0,0 +1,42 @@
+representation,protocol,horizon,seconds,background_ap_percent,camera_wearer_ap_percent,other_people_ap_percent,map_percent
+open_stream_replay,egospeak_full_timeline,1,0.2,53.605,74.331,81.331,69.756
+open_stream_replay,egospeak_full_timeline,2,0.4,47.066,69.402,79.020,65.163
+open_stream_replay,egospeak_full_timeline,3,0.6,42.615,64.600,77.520,61.578
+open_stream_replay,egospeak_full_timeline,4,0.8,39.463,60.641,76.244,58.783
+open_stream_replay,egospeak_full_timeline,5,1.0,38.002,57.371,75.367,56.913
+open_stream_replay,egospeak_full_timeline,6,1.2,37.398,55.258,74.523,55.727
+open_stream_replay,egospeak_full_timeline,7,1.4,36.933,53.400,74.126,54.819
+open_stream_replay,egospeak_full_timeline,8,1.6,36.653,51.502,73.716,53.957
+open_stream_replay,egospeak_full_timeline,9,1.8,36.527,50.346,73.504,53.459
+open_stream_replay,egospeak_full_timeline,10,2.0,36.673,49.089,73.246,53.002
+open_stream_replay,causal_valid_only,1,0.2,53.605,74.331,81.336,69.757
+open_stream_replay,causal_valid_only,2,0.4,47.066,69.402,79.029,65.166
+open_stream_replay,causal_valid_only,3,0.6,42.615,64.607,77.532,61.585
+open_stream_replay,causal_valid_only,4,0.8,39.463,60.652,76.259,58.792
+open_stream_replay,causal_valid_only,5,1.0,38.002,57.386,75.385,56.924
+open_stream_replay,causal_valid_only,6,1.2,37.398,55.277,74.543,55.739
+open_stream_replay,causal_valid_only,7,1.4,36.933,53.421,74.148,54.834
+open_stream_replay,causal_valid_only,8,1.6,36.656,51.522,73.740,53.973
+open_stream_replay,causal_valid_only,9,1.8,36.529,50.365,73.530,53.475
+open_stream_replay,causal_valid_only,10,2.0,36.676,49.107,73.275,53.019
+causal_prefix,egospeak_full_timeline,1,0.2,56.402,75.843,84.203,72.149
+causal_prefix,egospeak_full_timeline,2,0.4,48.713,70.695,81.305,66.904
+causal_prefix,egospeak_full_timeline,3,0.6,43.007,65.451,78.946,62.468
+causal_prefix,egospeak_full_timeline,4,0.8,39.482,61.051,77.107,59.214
+causal_prefix,egospeak_full_timeline,5,1.0,37.773,57.423,75.915,57.037
+causal_prefix,egospeak_full_timeline,6,1.2,37.197,54.929,75.241,55.789
+causal_prefix,egospeak_full_timeline,7,1.4,36.884,53.060,74.590,54.845
+causal_prefix,egospeak_full_timeline,8,1.6,36.514,50.916,74.283,53.905
+causal_prefix,egospeak_full_timeline,9,1.8,36.424,49.471,73.901,53.265
+causal_prefix,egospeak_full_timeline,10,2.0,36.095,48.407,73.667,52.723
+causal_prefix,causal_valid_only,1,0.2,56.402,75.843,84.209,72.151
+causal_prefix,causal_valid_only,2,0.4,48.713,70.695,81.315,66.908
+causal_prefix,causal_valid_only,3,0.6,43.007,65.458,78.960,62.475
+causal_prefix,causal_valid_only,4,0.8,39.482,61.062,77.123,59.223
+causal_prefix,causal_valid_only,5,1.0,37.773,57.438,75.934,57.048
+causal_prefix,causal_valid_only,6,1.2,37.197,54.947,75.262,55.802
+causal_prefix,causal_valid_only,7,1.4,36.884,53.081,74.612,54.859
+causal_prefix,causal_valid_only,8,1.6,36.517,50.936,74.308,53.920
+causal_prefix,causal_valid_only,9,1.8,36.426,49.489,73.928,53.281
+causal_prefix,causal_valid_only,10,2.0,36.098,48.425,73.697,52.740
+

results/easycom_open_stream_vs_causal_prefix_2026-08-19/summary.json

@@ -0,0 +1,74 @@
+{
+  "schema_version": 1,
+  "recorded_date": "2026-08-19",
+  "status": "single_run_snapshot",
+  "dataset": "EasyCom",
+  "split": {
+    "train_sessions": [7, 11, 10, 6, 9, 4, 8, 5, 12],
+    "test_sessions": [3, 2, 1]
+  },
+  "model": "Qwen2.5-Omni-3B Thinker",
+  "head": "MLP 2048 -> 1024 -> 30, reshaped to 10 x 3",
+  "prediction_step_seconds": 0.2,
+  "video_fps": 2,
+  "memory_seconds": 20,
+  "metrics_percent": {
+    "open_stream_replay": {
+      "egospeak_full_timeline_mean_map": 58.316,
+      "causal_valid_only_mean_map": 58.326,
+      "egospeak_full_timeline_horizon_map": [
+        69.756,
+        65.163,
+        61.578,
+        58.783,
+        56.913,
+        55.727,
+        54.819,
+        53.957,
+        53.459,
+        53.002
+      ]
+    },
+    "causal_prefix": {
+      "egospeak_full_timeline_mean_map": 58.830,
+      "causal_valid_only_mean_map": 58.841,
+      "egospeak_full_timeline_horizon_map": [
+        72.149,
+        66.904,
+        62.468,
+        59.214,
+        57.037,
+        55.789,
+        54.845,
+        53.905,
+        53.265,
+        52.723
+      ]
+    },
+    "causal_prefix_minus_open_stream": {
+      "egospeak_full_timeline_mean_map": 0.514,
+      "causal_valid_only_mean_map": 0.515,
+      "egospeak_full_timeline_horizon_map": [
+        2.393,
+        1.741,
+        0.890,
+        0.431,
+        0.124,
+        0.062,
+        0.026,
+        -0.052,
+        -0.194,
+        -0.279
+      ]
+    }
+  },
+  "egospeak_table2_reference_percent": {
+    "easycom_transformer_audio_video_mean_map": 58.7,
+    "easycom_gru_audio_video_mean_map": 60.6
+  },
+  "provenance": {
+    "source": "User-reported remote infer_turn_head.py console tables",
+    "raw_metrics_json_archived_here": false,
+    "multi_seed": false
+  }
+}