Claude/fix cuda device selection 01 k6 e xnxy5o5 pn lsqbj vl23 g - #30
Open
nnknk0802 wants to merge 20 commits into
Open
Conversation
This commit introduces a clean interface for the SSSD-ECG diffusion model: New Features: - model_wrapper.py: SSSDECG class with simplified API - model(x, y) for training loss computation - model.generate() for sample generation - Built-in checkpoint save/load functionality - dataset.py: Dedicated Dataset classes - ECGDataset: Generic ECG dataset class - PTBXLDataset: PTB-XL specific dataset - create_dataloaders(): Utility function for easy DataLoader creation - train_new.py: Clean training script using new interface - Progress tracking with tqdm - Flexible configuration via command line or JSON - Better checkpoint management - example_usage.py: Comprehensive usage examples - Basic training loop - Sample generation - Checkpoint operations - Real data handling - README_REFACTORED.md: Japanese documentation - API reference - Usage examples - Troubleshooting guide Benefits: - Cleaner separation of concerns (data, model, training) - Easier to use and extend - Better code organization - Original implementation remains unchanged
…ader-01FjNF9yHomAMfj5bjHrSeNF Refactor SSSD-ECG: Separate data loading from model implementation
Create a standalone, self-contained implementation of SSSD-ECG that can be easily copied to other projects. Structure: - sssd_standalone/: Main package directory - model_wrapper.py: Clean API wrapper for SSSDECG model - dataset.py: Dataset classes (ECGDataset, PTBXLDataset) - models/: Model implementations (SSSD_ECG, S4Model) - utils/: Utility functions for diffusion, training, sampling - config/: Configuration file (config_SSSD_ECG.json) - examples/: Training and usage examples - README.md: Comprehensive documentation - requirements.txt: Dependencies Features: - Simple API for training and inference - Modular design with separated components - Well-documented with usage examples - Easy to integrate into other projects - Includes parameter configuration file
…-setup-01BqMf5XWmL8yFxJuHDcvdZ9 Add standalone SSSD-ECG implementation
- Replace pytorch_lightning.utilities.rank_zero_only with simple local implementation - Add missing dependencies (einops, opt_einsum) to requirements.txt - This makes the standalone implementation truly independent without heavy dependencies The rank_zero_only decorator is now a simple pass-through function for standalone usage. For distributed training, users can modify this decorator as needed.
…-setup-01BqMf5XWmL8yFxJuHDcvdZ9 Fix: Remove pytorch_lightning dependency from standalone implementation
Convert the Jupyter notebook ecg_data_preprocessing.ipynb to a reusable Python module for creating PTB-XL dataloaders. This provides a cleaner API for data loading and makes it easier to integrate PTB-XL data into training pipelines. Changes: - Add ptbxl_dataloader.py: Main module with prepare_ptbxl_data(), create_ptbxl_dataloaders(), and get_ptbxl_dataloaders() functions - Add example_ptbxl_dataloader.py: Example usage scripts demonstrating different use cases - Add __init__.py: Package initialization for proper module structure - Update README.md: Add documentation for using the new Python module alongside the existing Jupyter notebook option The module supports: - Multi-label classification with configurable label types - Customizable sampling rates (100 Hz or 500 Hz) - Train/val/test splits based on stratified folds - Memory-efficient memmap data loading - Flexible batch size and worker configuration
…-01EsSESzvsw2on8AHCaRCHgc Claude/ptbxl dataloader creation 01 es se szvsw2on8 ah ca rc hgc
…ript - Fix bug in prepare_ptbxl_data() where reformat_as_memmap() return value was not being captured, which could cause KeyError when loading data - Add else branch to load already prepared data when recreate_data=False - Add debug_dataloader.py script to help troubleshoot data loading issues and verify PTB-XL data preparation step-by-step This should resolve the 'label_diag_filtered_numeric' KeyError that occurs when the memmap DataFrame is not properly saved/loaded.
…-01EsSESzvsw2on8AHCaRCHgc Fix: Properly handle reformat_as_memmap return value and add debug sc…
- Add check_dataframes.py to inspect saved DataFrames and compare df.pkl vs df_memmap.pkl - Add ptbxl_dataloader_v2.py with workaround that uses df_mapped directly instead of loading via load_dataset(), which may lose label columns - V2 includes detailed logging at each step to help identify where issues occur
…-01EsSESzvsw2on8AHCaRCHgc Add DataFrame checker and v2 dataloader with workaround
When min_cnt=0, prepare_data_ptb_xl() does not create *_filtered_numeric columns, only *_numeric columns. This caused KeyError when trying to access label_diag_filtered_numeric. Changes: - Both ptbxl_dataloader.py and ptbxl_dataloader_v2.py now check if *_filtered_numeric columns exist - If not found, fall back to *_numeric columns - Add informative error messages showing available label columns This allows users to use min_cnt=0 to include all labels without filtering, or min_cnt>0 to filter rare labels.
…-01EsSESzvsw2on8AHCaRCHgc Fix: Support min_cnt=0 by using unfiltered label columns
This commit adds a new generate_jit() method to the SSSDECG model wrapper that uses PyTorch's JIT compilation to speed up inference. Changes: - Added _check_jit_availability() to detect torch.compile or torch.jit - Added _sampling_label_jit() as JIT-compatible sampling function - Added generate_jit() method with same interface as generate() - Model is compiled on first call and cached for subsequent calls - Automatically falls back to regular generate() if JIT unavailable - Updated both src/sssd/model_wrapper.py and sssd_standalone/model_wrapper.py - Added test_generate_jit.py to test the new functionality Benefits: - Significantly faster inference after initial compilation - Supports PyTorch 2.0+ torch.compile and older torch.jit - Backward compatible - falls back gracefully if JIT not available - No changes needed to existing generate() usage
…QS5PbZSk9QawwuX4 Add JIT-compiled generate_jit method for faster inference
Previously, the model was hardcoded to use cuda:0 only. Multiple `.cuda()` calls throughout the codebase prevented users from specifying alternative CUDA devices like cuda:1. Changes: - Modified util.py functions to accept device parameter: * std_normal(): Added device parameter, changed .cuda() to .to(device) * calc_diffusion_step_embedding(): Added device parameter with auto-detection * sampling_label(): Added device parameter for tensor creation * training_loss_label(): Added device parameter for tensor creation - Updated model_wrapper.py to pass device to utility functions in: * forward(): Pass device to training_loss_label() * generate(): Pass device to sampling_label() - Updated SSSD_ECG.py Residual_group.forward() to pass device parameter to calc_diffusion_step_embedding() - Applied changes to both src/sssd/ and sssd_standalone/ directories Usage: # Use cuda:0 (first GPU) model = SSSDECG(config_path="config.json", device="cuda:0") # Use cuda:1 (second GPU) model = SSSDECG(config_path="config.json", device="cuda:1") # Use CPU model = SSSDECG(config_path="config.json", device="cpu") All changes are backward compatible. Default behavior unchanged. Added documentation (CUDA_DEVICE_SELECTION.md) and test script (test_cuda_devices.py) to verify functionality across devices.
…ction Improvements: - Added device validation in __init__ to catch invalid device selections before attempting tensor operations - Check if CUDA is available when requested - Validate device index is within available range - Provide clear error messages listing available devices New helper methods: - list_available_devices(): Returns list of available device strings - print_available_devices(): Prints detailed device information Enhanced documentation: - Added "Checking Available Devices" section - Added comprehensive "Troubleshooting" section covering: * Invalid device ordinal errors * CUDA not available errors * CUDA_VISIBLE_DEVICES usage * Device assignment verification This prevents the cryptic "CUDA error: invalid device ordinal" that occurred when specifying a non-existent device (e.g., cuda:1 when only cuda:0 is available). Example error message before: RuntimeError: CUDA error: invalid device ordinal Example error message after: RuntimeError: CUDA device 'cuda:1' requested but only 1 device(s) available. Available devices: cuda:0, cpu
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.