 Xin Wang
|
b5ce64702a
|
Add Ego4D local AV ablation workflow
|
2026-08-28 00:25:54 +08:00 |
|
|
Xin Wang
|
8b2fbd2622
|
Merge remote-tracking branch 'aliyun/main'
|
2026-08-28 00:21:47 +08:00 |
|
|
Xin Wang
|
5f45e7c115
|
Add EasyCom interaction weak-label pipeline
|
2026-08-28 00:07:54 +08:00 |
|
|
Xin Wang
|
c4feeb47f4
|
Fix transformers version
|
2026-08-27 15:24:59 +08:00 |
|
|
Xin Wang
|
694473b475
|
Fix bbox offset
|
2026-08-27 15:08:38 +08:00 |
|
|
Xin Wang
|
11f197724f
|
Add pooled local AV baselines and research notes
|
2026-08-27 14:19:33 +08:00 |
|
|
Xin Wang
|
d34725778f
|
Avoid upstream Omni loading-info speaker crash
|
2026-08-27 13:50:02 +08:00 |
|
|
Xin Wang
|
5e34563739
|
Support upstream Omni class for ROMA loading
|
2026-08-27 13:45:25 +08:00 |
|
|
Xin Wang
|
655d27a4ca
|
Add ROMA frozen backbone extraction support
|
2026-08-27 13:35:59 +08:00 |
|
|
Xin Wang
|
1925ff71a7
|
Record preliminary transition auxiliary results
|
2026-08-27 09:25:08 +08:00 |
|
|
Xin Wang
|
5923b89d60
|
Add transition-aware R1 and R2 objectives
|
2026-08-26 17:18:50 +08:00 |
|
|
Xin Wang
|
0590df1e88
|
Record EasyCom loss structure ablation
|
2026-08-26 16:30:48 +08:00 |
|
|
Xin Wang
|
f35b9108ac
|
Add overlap-aware turn prediction objectives
|
2026-08-26 15:03:56 +08:00 |
|
|
Xin Wang
|
f45ea9cbc6
|
Add higher-capacity last-four-layer fusion
|
2026-08-26 13:06:18 +08:00 |
|
|
Xin Wang
|
43e0e14cf5
|
Add learnable last-four-layer query fusion
|
2026-08-26 10:29:43 +08:00 |
|
|
Xin Wang
|
dbef1fc173
|
Add final-layer mixture readout
|
2026-08-26 09:56:54 +08:00 |
|
|
Xin Wang
|
ca6dd88923
|
Add native block phase diagnostic
|
2026-08-26 09:10:21 +08:00 |
|
|
Xin Wang
|
56c8fe0a0e
|
Fix final LoRA checkpoint report
|
2026-08-25 09:45:47 +08:00 |
|
|
Xin Wang
|
aa4278e13e
|
Fix final-layer LoRA target matching
|
2026-08-24 15:38:45 +08:00 |
|
|
Xin Wang
|
7fdda297d3
|
Save periodic LoRA weight checkpoints
|
2026-08-24 15:31:44 +08:00 |
|
|
Xin Wang
|
7acab44ea2
|
Add low-drift LoRA training controls
|
2026-08-24 15:24:53 +08:00 |
|
|
Xin Wang
|
b7856ece4d
|
Archive EasyCom sparse supervision results
|
2026-08-24 12:02:49 +08:00 |
|
|
Xin Wang
|
edd37dc36a
|
Add offline sparse supervision control
|
2026-08-24 11:45:11 +08:00 |
|
|
Xin Wang
|
43cf1f386e
|
Add frozen Thinker sparse head-only control
|
2026-08-24 11:32:03 +08:00 |
|
|
Xin Wang
|
f6dbce30e4
|
Reuse global history frames in overlap working memory
|
2026-08-24 11:17:39 +08:00 |
|
|
Xin Wang
|
1731f2b6f7
|
Optimize overlap working-memory audio preprocessing
|
2026-08-24 10:18:55 +08:00 |
|
|
Xin Wang
|
20446c0bd8
|
Add overlap working-memory ablation and LoRA epoch2 record
|
2026-08-24 09:52:56 +08:00 |
|
|
Xin Wang
|
3cc0c3da85
|
Add frozen Thinker cache-path control
|
2026-08-24 01:07:42 +08:00 |
|
|
Xin Wang
|
c5e2e044a8
|
Add dense-cache evaluation for Omni LoRA
|
2026-08-24 00:27:42 +08:00 |
|
|
Xin Wang
|
a755ab4fc3
|
Add candidate improvement for explicit working and long memory representation in Omni LoRA
|
2026-08-23 23:49:57 +08:00 |
|
|
Xin Wang
|
909ab66c09
|
Add adapter checkpoint support for continued training in Omni LoRA
|
2026-08-23 23:49:09 +08:00 |
|
|
Xin Wang
|
7cdb2d90c2
|
Add multi-GPU DDP for Omni LoRA training
|
2026-08-23 22:41:39 +08:00 |
|
|
Xin Wang
|
cb4abfdf82
|
Add transition-sparse Omni AV caching
|
2026-08-23 18:28:07 +08:00 |
|
|
Xin Wang
|
8b373cd19b
|
Optimize online-tail audio preprocessing
|
2026-08-23 16:49:10 +08:00 |
|
|
Xin Wang
|
c844673984
|
Add TorchCodec video backend
|
2026-08-23 16:41:05 +08:00 |
|
|
Xin Wang
|
b875a7512e
|
Optimize Omni audio preprocessing
|
2026-08-23 16:26:01 +08:00 |
|
|
Xin Wang
|
85ed8a670c
|
Batch Omni AV cache extraction
|
2026-08-23 16:03:07 +08:00 |
|
|
Xin Wang
|
c2162e1170
|
Add two-GPU AV producer for Omni LoRA
|
2026-08-23 15:29:26 +08:00 |
|
|
Xin Wang
|
dc932cc165
|
Add online current-tail LoRA benchmark
|
2026-08-23 15:10:16 +08:00 |
|
|
Xin Wang
|
1069a0beee
|
Add cached AV pipeline for Omni turn LoRA
|
2026-08-23 14:40:03 +08:00 |
|
|
Xin Wang
|
2eb61e30de
|
Add README for Omni-only Turn-Taking Experiment detailing project goals, architecture, current status, and EasyCom results. Include comparisons with EgoSpeak and clarify dataset handling and evaluation processes.
|
2026-08-19 23:09:56 +08:00 |
|
|
Xin Wang
|
2e773706bd
|
Enhance EasyCom and Ego4D integration by adding dataset support and manifest handling. Update normalization functions to accommodate both datasets, and modify training and inference scripts to require manifest files for Ego4D. Improve argument parsing and validation for dataset-specific operations, ensuring proper session management and error handling. Update documentation to reflect changes in dataset usage and evaluation processes.
|
2026-08-19 20:49:28 +08:00 |
|
|
Xin Wang
|
92bf93c96b
|
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.
|
2026-08-19 11:29:27 +08:00 |
|
|
Xin Wang
|
5fa61eed23
|
Add extract_ego4d_batch.py for efficient EgoSpeak clip extraction. This script loads the Qwen model once and processes selected manifest rows sequentially, ensuring each clip is handled independently. Update README with usage instructions and details on the extraction process, including handling of completed outputs and memory management.
|
2026-08-18 22:22:37 +08:00 |
|
|
Xin Wang
|
609b2781d4
|
Add profiling capabilities for processor components in streaming model. Introduce _patch_component_call to time component invocations and enhance _install_profiling_hooks with new processor stages. Refactor tensor conversion and device transfer functions for improved clarity and performance. Update profiling summary to include new stages and advice for optimization.
|
2026-08-18 11:20:55 +08:00 |
|
|
Xin Wang
|
43f47e8f3a
|
Add caching mechanism for completed AV blocks in streaming model. Introduce CompletedBlockPieceCache class to manage reusable blocks during strict-window replay, enhancing efficiency. Update profiling hooks for cache operations and refine README documentation to reflect changes in block handling and replay strategies.
|
2026-08-18 11:02:42 +08:00 |
|
|
Xin Wang
|
d610305355
|
Add profiling script for GPU bottleneck analysis in EasyCom extraction. Introduce command-line options for GPU ID, log directory, and sampling period. Implement stage profiling to measure extraction performance and synchronize CUDA operations. Enhance CausalSourceAV class with new parameters for decoder threads and open readers, optimizing video decoding process.
|
2026-08-18 10:30:19 +08:00 |
|
|
Xin Wang
|
24cb92f1f4
|
Add support for NVDEC backend in video processing and enhance README documentation. Introduce a new script for extracting EasyCom sessions on specific GPUs, detailing session management and error handling. Update requirements.txt to include PyNvVideoCodec as an optional dependency. Refactor streaming_turn_model_directly.py to accommodate multiple decoder backends and improve frame indexing with ffprobe.
|
2026-08-18 08:41:54 +08:00 |
|
|
Xin Wang
|
e4fb62ba60
|
Add script for extracting EasyCom sessions with three workers per GPU. This script manages session assignments, ensures serialized worker launches, and includes error handling for session readiness and logging. Enhances the EasyCom pipeline for improved session management and feature extraction.
|
2026-08-17 23:43:34 +08:00 |
|
|
Xin Wang
|
d815b59def
|
Refactor README and streaming_turn_model.py to clarify fixed-grid frame ownership and update documentation. Remove merge conflict markers and enhance explanations regarding session launching conditions and audio slicing behavior.
|
2026-08-17 22:59:02 +08:00 |
|