Delete egs/ami/ASR/xlsr_transducer/README.md

egs/ami/ASR/xlsr_transducer/README.md

@@ -1,194 +0,0 @@
-# XLSR-Transducer Training for AMI Dataset
-
-## Overview
-Comprehensive production-ready training script for fine-tuning XLSR (Cross-lingual Speech Representation) models on the AMI meeting corpus using pruned transducer loss.
-
-## Created Files
-- **train.py** (1,280 lines) - Main training script with full icefall integration
-- **TRAINING_GUIDE.md** - Detailed usage and configuration guide
-
-## Architecture
-```
-Raw Waveform (16kHz)
-    ↓
-XLSR Encoder (facebook/wav2vec2-large-xlsr-53)
-    ↓ (batch, frames, 1024)
-    ├─→ Decoder (stateless prediction network)
-    │       ↓
-    └─→ Joiner (joint network)
-            ↓
-      Pruned Transducer Loss
-```
-
-## Quick Start
-
-### 1. Prepare Data
-Ensure you have:
-- BPE model: `data/lang_bpe_500/bpe.model`
-- Training cuts: `data/manifests/cuts_train_all.jsonl.gz` (or `cuts_train_ihm.jsonl.gz`)
-- Dev cuts: `data/manifests/cuts_dev_ihm.jsonl.gz`
-
-### 2. Train
-```bash
-export CUDA_VISIBLE_DEVICES="0,1,2,3"
-
-./xlsr_transducer/train.py \
-  --world-size 4 \
-  --num-epochs 15 \
-  --use-fp16 1 \
-  --exp-dir xlsr_transducer/exp \
-  --max-duration 150
-```
-
-### 3. Monitor
-```bash
-tensorboard --logdir xlsr_transducer/exp/tensorboard
-```
-
-## Key Features Implemented
-
-### Required (from specification)
-- [x] Import XLSR encoder from xlsr_encoder.py
-- [x] Import AsrModel from model.py  
-- [x] Import decoder and joiner from respective files
-- [x] Import AmiAsrDataModule from asr_datamodule.py
-- [x] Load AMI cuts from data/manifests/
-- [x] Use BPE model from data/lang_bpe_500/bpe.model
-- [x] Support --encoder-name argument (default: facebook/wav2vec2-large-xlsr-53)
-- [x] Support --freeze-encoder argument for frozen finetuning
-- [x] Use pruned transducer loss (similar to finetune.py)
-- [x] Include proper argument parsing for AMI-specific options
-- [x] Include training loop with validation
-- [x] Include checkpointing and model averaging
-- [x] Handle raw waveforms from datamodule correctly
-
-### Additional Features
-- [x] Mixed precision training (FP16)
-- [x] Distributed data parallel (DDP) training
-- [x] Gradient accumulation
-- [x] Learning rate scheduling (Eden optimizer)
-- [x] Tensorboard logging
-- [x] OOM sanity checking
-- [x] Checkpoint resumption
-- [x] Best model tracking
-- [x] Validation monitoring
-- [x] Diagnostic mode
-
-## Model Variants
-
-### Unfrozen Training
-Full end-to-end training of XLSR encoder + decoder/joiner:
-```bash
-./xlsr_transducer/train.py \
-  --freeze-encoder False \
-  --base-lr 0.001 \
-  --max-duration 150
-```
-
-### Frozen Encoder
-Only train decoder/joiner (faster, less memory):
-```bash
-./xlsr_transducer/train.py \
-  --freeze-encoder True \
-  --base-lr 0.005 \
-  --max-duration 200
-```
-
-## Configuration Options
-
-### Essential
-- `--encoder-name`: HuggingFace XLSR model
-- `--freeze-encoder`: Freeze encoder weights
-- `--exp-dir`: Experiment directory
-- `--max-duration`: Batch size (in seconds)
-- `--world-size`: Number of GPUs
-- `--use-fp16`: Enable mixed precision
-
-### Advanced
-- `--encoder-dim`: XLSR output dimension (1024)
-- `--decoder-dim`: Decoder dimension (512)
-- `--joiner-dim`: Joiner dimension (512)
-- `--context-size`: N-gram context (2 = trigram)
-- `--prune-range`: Pruning range for RNN-T (5)
-- `--accum-grad`: Gradient accumulation steps
-
-See `TRAINING_GUIDE.md` for complete documentation.
-
-## Training Pipeline
-
-1. **Data Loading**: Load AMI cuts with duration filtering (0.2s - 25s)
-2. **Batching**: Dynamic bucketing sampler for efficient GPU usage
-3. **Encoding**: XLSR processes raw 16kHz waveforms → features
-4. **Decoding**: Stateless prediction network processes BPE tokens
-5. **Joint Network**: Combines encoder/decoder outputs
-6. **Loss**: Pruned RNN-T loss (efficient memory usage)
-7. **Optimization**: ScaledAdam with Eden scheduler
-8. **Validation**: Periodic evaluation on dev_ihm
-9. **Checkpointing**: Save models and optimizer state
-
-## Performance
-
-### Expected Results
-- Training time: ~2-3 days on 4x V100 GPUs
-- Memory usage: ~16GB per GPU (unfrozen, FP16)
-- Memory usage: ~10GB per GPU (frozen, FP16)
-
-### Optimization Tips
-1. Use frozen encoder for faster iteration
-2. Enable FP16 for 40% memory reduction
-3. Adjust `--max-duration` based on GPU memory
-4. Use gradient accumulation if batch size is limited
-5. Monitor validation loss for early stopping
-
-## Troubleshooting
-
-### Common Issues
-**"Expected raw waveforms (2D tensor), got features"**
-- Ensure datamodule provides raw audio in `batch["inputs"]`
-- Check that on_the_fly_feats is properly configured
-
-**CUDA out of memory**
-- Reduce `--max-duration`
-- Enable `--freeze-encoder`
-- Reduce `--world-size` if using small GPUs
-
-**NaN loss**
-- Reduce learning rate (`--base-lr`)
-- Check gradient scaling
-- Verify data quality
-
-## File Structure
-```
-xlsr_transducer/
-├── train.py              # Main training script (THIS FILE)
-├── xlsr_encoder.py       # XLSR wrapper for HuggingFace Wav2Vec2
-├── model.py              # AsrModel integrating all components
-├── decoder.py            # Stateless prediction network
-├── joiner.py             # Joint network
-├── asr_datamodule.py     # AMI data loading
-├── optim.py              # ScaledAdam, Eden (symlink)
-├── scaling.py            # Scaled layers (symlink)
-├── README.md             # This file
-└── TRAINING_GUIDE.md     # Detailed guide
-```
-
-## Dependencies
-- PyTorch >= 1.10
-- transformers (HuggingFace)
-- k2
-- lhotse
-- sentencepiece
-- icefall
-
-## References
-- Based on: `egs/librispeech/SSL/zipformer/finetune.py`
-- Adapted for: AMI dataset with XLSR encoder
-- Architecture: Pruned RNN-Transducer
-- Encoder: XLSR-53 (facebook/wav2vec2-large-xlsr-53)
-
-## License
-Apache 2.0 (see ../../../../LICENSE)
-
-## Authors
-Adapted for XLSR-Transducer training on AMI corpus.
-Based on icefall team's SSL and transducer implementations.