b2txt25/CLAUDE.md

# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

This repository contains the code and data for "An Accurate and Rapidly Calibrating Speech Neuroprosthesis" published in the New England Journal of Medicine (2024). It implements a brain-to-text system that converts neural signals from speech motor cortex into text using RNN models and n-gram language models.

## Development Environment Setup

### Main Environment (b2txt25)
```bash
./setup.sh
conda activate b2txt25
```

### Language Model Environment (b2txt25_lm)
```bash
./setup_lm.sh
conda activate b2txt25_lm
```

**Important**: The project requires two separate conda environments due to conflicting PyTorch versions:
- `b2txt25`: PyTorch with CUDA 12.6 for model training/evaluation
- `b2txt25_lm`: PyTorch 1.13.1 for Kaldi-based n-gram language models

### Redis Setup
Redis is required for inter-process communication. Install on Ubuntu:
```bash
curl -fsSL https://packages.redis.io/gpg | sudo gpg --dearmor -o /usr/share/keyrings/redis-archive-keyring.gpg
echo "deb [signed-by=/usr/share/keyrings/redis-archive-keyring.gpg] https://packages.redis.io/deb $(lsb_release -cs) main" | sudo tee /etc/apt/sources.list.d/redis.list
sudo apt-get update && sudo apt-get install redis
sudo systemctl disable redis-server
```

## Architecture Overview

### High-Level System Flow
1. **Neural Data Input**: 512 features (2 per electrode × 256 electrodes) binned at 20ms resolution
2. **RNN Model**: Converts neural features to phoneme logits via CTC loss
3. **Language Model**: Decodes phoneme logits to words using n-gram models + OPT rescoring
4. **Redis Communication**: Coordinates between RNN inference and language model processes

### Key Components

#### Model Training (`model_training/`)
- **Core Script**: `train_model.py` (loads config from `rnn_args.yaml`)
- **Model Architecture**: `rnn_model.py` - 5-layer GRU with 768 hidden units
- **Trainer**: `rnn_trainer.py` - Custom PyTorch trainer with CTC loss
- **Evaluation**: `evaluate_model.py` - Inference pipeline with Redis communication

#### Language Model (`language_model/`)
- **Standalone Server**: `language-model-standalone.py` - Redis-based LM server
- **Kaldi Integration**: Uses custom C++ bindings for efficient n-gram decoding
- **OPT Rescoring**: Facebook OPT 6.7B for language model rescoring
- **Build System**: Complex CMake-based build for Kaldi/SRILM integration

#### Utilities (`nejm_b2txt_utils/`)
- **General Utils**: `general_utils.py` - Shared utility functions
- **Package**: Installed via `setup.py` as `nejm_b2txt_utils`

#### Analysis (`analyses/`)
- **Jupyter Notebooks**: `figure_2.ipynb`, `figure_4.ipynb` for paper figures

## Common Development Tasks

### Training a Model
```bash
conda activate b2txt25
cd model_training
python train_model.py
```

### Running Evaluation Pipeline
1. Start Redis server:
   ```bash
   redis-server
   ```

2. Start language model (separate terminal):
   ```bash
   conda activate b2txt25_lm
   python language_model/language-model-standalone.py --lm_path language_model/pretrained_language_models/openwebtext_1gram_lm_sil --do_opt --nbest 100 --acoustic_scale 0.325 --blank_penalty 90 --alpha 0.55 --redis_ip localhost --gpu_number 0
   ```

3. Run evaluation (separate terminal):
   ```bash
   conda activate b2txt25
   cd model_training
   python evaluate_model.py --model_path ../data/t15_pretrained_rnn_baseline --data_dir ../data/hdf5_data_final --eval_type test --gpu_number 1
   ```

4. Shutdown Redis:
   ```bash
   redis-cli shutdown
   ```

### Building Language Model from Scratch
```bash
# Build SRILM (in language_model/srilm-1.7.3/)
export SRILM=$PWD
make MAKE_PIC=yes World

# Build Kaldi components (in language_model/runtime/server/x86/)
mkdir build && cd build
cmake .. && make -j8
```

## Data Structure

### Neural Data Format
- **File Type**: HDF5 files in `data/hdf5_data_final/`
- **Features**: 512 neural features per 20ms bin:
  - 0-64: ventral 6v threshold crossings
  - 65-128: area 4 threshold crossings
  - 129-192: 55b threshold crossings
  - 193-256: dorsal 6v threshold crossings
  - 257-320: ventral 6v spike band power
  - 321-384: area 4 spike band power
  - 385-448: 55b spike band power
  - 449-512: dorsal 6v spike band power

### Data Loading
Use `load_h5py_file()` in `model_training/evaluate_model_helpers.py` as reference for HDF5 data loading.

## Important Notes

- **GPU Requirements**: OPT 6.7B requires ~12.4GB VRAM; RTX 4090s recommended
- **Memory Requirements**: 3-gram LM needs ~60GB RAM, 5-gram needs ~300GB RAM
- **Environment Isolation**: Always use correct conda environment for each component
- **Redis Dependency**: Many scripts require Redis server to be running
- **Build Dependencies**: CMake ≥3.14 and GCC ≥10.1 required for language model builds

## XLA Optimizations (TPU-Friendly Model)

The RNN model has been optimized for XLA compilation and TPU training while preserving the original model architecture. These optimizations improve compilation speed and reduce memory usage on TPUs.

### Applied XLA Optimizations

#### 1. Dynamic Shape Operations → Static Operations
**Problem**: XLA compiler struggles with dynamic tensor shapes and indexing
**Solution**: Replace dynamic operations with XLA-friendly alternatives

```python
# Before (XLA-unfriendly):
day_weights = torch.stack([self.day_weights[i] for i in day_idx], dim=0)
day_biases = torch.cat([self.day_biases[i] for i in day_idx], dim=0).unsqueeze(1)

# After (XLA-friendly):
all_day_weights = torch.stack(list(self.day_weights), dim=0)  # Static stack
all_day_biases = torch.stack([bias.squeeze(0) for bias in self.day_biases], dim=0)
day_weights = torch.index_select(all_day_weights, 0, day_idx)  # Static gather
day_biases = torch.index_select(all_day_biases, 0, day_idx).unsqueeze(1)
```

#### 2. Matrix Operations → XLA Primitives
**Problem**: Complex einsum operations are less optimized than native XLA ops
**Solution**: Use batch matrix multiplication (bmm) for better XLA performance

```python
# Before:
x = torch.einsum("btd,bdk->btk", x, day_weights) + day_biases

# After (XLA-optimized):
x = torch.bmm(x, day_weights.to(x.dtype)) + day_biases.to(x.dtype)  # bmm + dtype consistency
```

#### 5. Mixed Precision Dtype Consistency (Comprehensive Fix)
**Problem**: Mixed precision training causes dtype mismatches throughout the adversarial training pipeline
**Error**: `Status: INVALID_ARGUMENT: Call parameter must match argument; got parameter 0 shape: f32[32,7168], argument shape: bf16[32,7168]`

**Root Cause Analysis**: The error occurred at dimension 7168 = 512 * 14, indicating patch processing with patch_size=14. The dtype mismatch cascaded through multiple layers:
1. Initial bmm operations in day-specific transformations
2. Adversarial training residual connections between models
3. Patch processing operations (unfold, permute, reshape)
4. Gradient Reversal Layer (GRL) operations
5. Hidden state initialization in adversarial training helper methods

**Comprehensive Solution**: Implement dtype consistency across the entire adversarial training data flow:

```python
# Fix 1: Basic bmm operations with dtype consistency
x = torch.bmm(x, day_weights.to(x.dtype)) + day_biases.to(x.dtype)

# Fix 2: Patch processing with explicit dtype preservation
if self.patch_size > 0:
    original_dtype = x.dtype  # Preserve original dtype for XLA/TPU compatibility
    x = x.unsqueeze(1)
    x = x.permute(0, 3, 1, 2)
    x_unfold = x.unfold(3, self.patch_size, self.patch_stride)
    x_unfold = x_unfold.squeeze(2)
    x_unfold = x_unfold.permute(0, 2, 3, 1)
    x = x_unfold.reshape(batch_size, x_unfold.size(1), -1)
    # Ensure dtype consistency after patch processing operations
    x = x.to(original_dtype)

# Fix 3: Adversarial training residual connections
noise_output = noise_output.to(x_processed.dtype)
denoised_input = x_processed - noise_output

# Fix 4: Gradient Reversal Layer dtype handling
noisy_input = gradient_reverse(noise_output, grl_lambda) if grl_lambda else noise_output
# Ensure dtype consistency after GRL (preserves input dtype but explicit check)
noisy_input = noisy_input.to(x_processed.dtype)

# Fix 5: Hidden state dtype consistency in helper methods
# In _clean_forward_with_processed_input:
if states is None:
    states = self.clean_speech_model.h0.expand(3, batch_size, self.clean_speech_model.n_units).contiguous()
    # Ensure hidden states match input dtype for mixed precision training
    states = states.to(x_processed.dtype)

# In _noisy_forward_with_processed_input:
if states is None:
    states = self.noisy_speech_model.h0.expand(2, batch_size, self.noisy_speech_model.n_units).contiguous()
    # Ensure hidden states match input dtype for mixed precision training
    states = states.to(x_processed.dtype)
```

**Key Implementation Details**:
- **GradientReversalFn**: Preserves input dtype automatically (identity forward, gradient reversal backward)
- **Patch Processing**: Explicit dtype preservation prevents unfold operations from changing precision
- **Residual Connections**: All tensor arithmetic operations ensure matching dtypes
- **Helper Methods**: Hidden state initialization matches processed input dtype
- **Data Flow**: NoiseModel → GRL → NoisySpeechModel maintains dtype consistency throughout

#### 3. Hidden State Initialization
**Problem**: Dynamic batch size allocation causes XLA recompilation
**Solution**: Use static shapes and avoid x.shape[0] in tensor creation

```python
# Before:
if states is None:
    states = self.h0.expand(2, x.shape[0], self.input_size).contiguous()

# After (XLA-friendly):
batch_size = x.size(0)  # Extract once
if states is None:
    states = self.h0.expand(2, batch_size, self.input_size).contiguous()
```

#### 4. Return Value Optimization
**Problem**: Complex dictionary returns cause XLA compilation issues
**Solution**: Use tuples instead of dictionaries for cleaner XLA graphs

```python
# Before (XLA-unfriendly):
return {
    'clean_logits': clean_logits,
    'noisy_logits': noisy_logits,
    'noise_output': noise_output
}

# After (XLA-friendly):
return clean_logits, noisy_logits, noise_output  # Simple tuple return
```

### Files Modified for XLA Optimization

- **`model_training_nnn/rnn_model.py`**: Comprehensive XLA optimization with dtype consistency
  - **`GradientReversalFn`**: Added adversarial training gradient reversal layer
  - **`NoiseModel.forward()`**: Dynamic indexing → static gather operations + comprehensive dtype consistency + patch processing dtype preservation
  - **`CleanSpeechModel.forward()`**: Same optimizations + bmm for matrix ops + comprehensive dtype consistency + patch processing dtype preservation
  - **`NoisySpeechModel.forward()`**: Hidden state optimization (no day layers, simplified)
  - **`TripleGRUDecoder.forward()`**: Complex return values → tuple returns + comprehensive adversarial training dtype fixes + residual connection dtype consistency + GRL dtype handling
  - **`TripleGRUDecoder._apply_preprocessing()`**: Static preprocessing operations + dtype consistency + patch processing dtype preservation
  - **`TripleGRUDecoder._clean_forward_with_processed_input()`**: Helper method with hidden state dtype consistency for mixed precision
  - **`TripleGRUDecoder._noisy_forward_with_processed_input()`**: Helper method with hidden state dtype consistency for mixed precision

**Specific Dtype Consistency Fixes Applied**:
1. **Basic Operations**: All `torch.bmm()` operations with `.to(x.dtype)` conversions
2. **Patch Processing**: Explicit dtype preservation through unfold/permute/reshape operations
3. **Adversarial Training**: Residual connections with `.to(x_processed.dtype)` conversions
4. **Gradient Reversal**: Dtype consistency after GRL operations
5. **Hidden States**: All hidden state initialization with `.to(x_processed.dtype)` conversions
6. **Data Flow**: End-to-end dtype consistency in NoiseModel → GRL → NoisySpeechModel pipeline

**Error Resolved**: `f32[32,7168] vs bf16[32,7168]` dtype mismatch in mixed precision TPU training

### Benefits of XLA Optimizations

1. **Faster Compilation**: Static shapes allow XLA to pre-compile optimized kernels
2. **Better Memory Usage**: Reduced dynamic allocation during training
3. **Improved TPU Utilization**: XLA primitives map directly to TPU matrix units
4. **Consistent Performance**: Eliminates recompilation caused by dynamic shapes

### Testing and Validation

Created test scripts to verify model consistency:
- **`test_xla_model.py`**: Comprehensive model validation testing
- **`quick_test_xla.py`**: Fast verification of basic functionality

**Important**: These optimizations preserve the exact model architecture and mathematical operations. Only the implementation has been made XLA-friendly.

### Usage Notes

- All original model interfaces remain unchanged
- Both 'inference' and 'full' modes are supported
- Backward compatibility with existing training scripts is maintained
- TPU training should now show improved compilation times and memory efficiency

### Troubleshooting Dtype Issues in Mixed Precision Training

**Common Error Pattern**:
```
Status: INVALID_ARGUMENT: Call parameter must match argument; got parameter 0 shape: f32[X,Y], argument shape: bf16[X,Y]
```

**Diagnosis Steps**:
1. **Identify Operation**: Look at the tensor dimensions to identify which operation is failing
   - `7168 = 512 * 14`: Patch processing operation with patch_size=14
   - `512`: Basic neural features
   - Other patterns may indicate different operations

2. **Check Data Flow**: Trace the tensor through the adversarial training pipeline
   - Input → NoiseModel → residual connection → CleanSpeechModel
   - Input → NoiseModel → GRL → NoisySpeechModel

3. **Verify Dtype Consistency**: Ensure all operations maintain input dtype
   - Use `.to(x.dtype)` for all operand tensors
   - Preserve dtype through complex operations (unfold, permute, reshape)
   - Match hidden state dtype to input tensor dtype

**Quick Fix Template**:
```python
# For any tensor operation between tensors a and b:
result = operation(a, b.to(a.dtype))

# For complex operations that might change dtype:
original_dtype = tensor.dtype
tensor = complex_operation(tensor)
tensor = tensor.to(original_dtype)

# For hidden state initialization:
states = states.to(input_tensor.dtype)
```

## Competition Context
This codebase also serves as baseline for the Brain-to-Text '25 Competition on Kaggle, providing reference implementations for neural signal decoding.