- Create fairseq_compat.py as drop-in replacement for fairseq (which cannot
build on Windows). Implements minimal Wav2Vec2Model with exact architecture
matching the XLSR-300M checkpoint key structure.
- Fix ConvFeatureExtraction to produce correct key naming (conv_layers.X.0/2.1)
- Add weight_norm to positional convolution for weight_g/weight_v keys
- Update wav2vec2_Nes2Net_X.py to use fairseq_compat fallback and dynamic
XLSR checkpoint path via environment variable
- Update audio inference wrapper with correct model args (SE_ratio=[1],
pool_func='mean') and softmax-based score normalization
- Add Colab/Kaggle notebook for full pipeline including video model hosting
- Add GPU test scripts for RTX 5050 (sm_120/Blackwell) verification
- Update task tracker and documentation with current progress
Tested: Full Nes2Net model loads with zero missing keys, runs inference
on RTX 5050 GPU using 2.43 GB VRAM.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
feat: initial project setup with audio/video detection modules and fusion
