[FEATURE] Grouped convolutions #183
Merged
+14,904
−250
Conversation
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
- Add RingBuffer class to manage write/read pointers for Eigen::MatrixXf buffers - Add Reset() method to Conv1D to initialize ring buffer and pre-allocate output - Add get_output() method to Conv1D to access output buffer - Rename process_() to Process() and update to use internal ring buffer - Update ConvNet to use Conv1D internal buffers via Process() and get_output() - Update WaveNet to use Conv1D internal buffers and propagate Reset() - Add comprehensive tests for Conv1D in test_conv1d.cpp Implements issue #145
- Fix Reset() to zero buffer behind starting write position - Fix Rewind() to copy receptive_field (max lookback) samples to start - Set write position after receptive_field when rewinding - Add comprehensive test suite for RingBuffer (12 tests) - Tests cover: construction, reset, write/read, advance, rewind, lookback, etc. Fixes issues with RingBuffer not properly handling max lookback when rewinding.
- Rename _receptive_field member to _max_lookback - Rename SetReceptiveField() to SetMaxLookback() - Update all comments and documentation - Update all test references from receptive_field to max_lookback - More descriptive name for ring buffer context
- Create NAM/ring_buffer.h and NAM/ring_buffer.cpp for RingBuffer class - Create NAM/conv1d.h and NAM/conv1d.cpp for Conv1D class - Remove RingBuffer and Conv1D from NAM/dsp.h and NAM/dsp.cpp - Update includes in convnet.h, wavenet.h, and test files - All tests pass after refactoring
- Remove GetCapacity() from public interface (storage size is internal detail) - Add GetMaxBufferSize() to return the max_buffer_size passed to Reset() - Store max_buffer_size as member variable - Update tests to use GetMaxBufferSize() instead of GetCapacity() - External code should trust that storage is sized correctly
- Assert that write pointer is at least 2 * max_lookback to avoid aliasing - Assert that copy start position is within storage bounds - Remove silent failure condition - now asserts instead of silently skipping - Prevents data corruption from overlapping copy operations
- Renamed method in conv1d.h and conv1d.cpp - Updated all usages in convnet.cpp and wavenet.cpp - Updated all test cases in test_conv1d.cpp - Matches naming convention with Conv1x1::GetOutput()
- Remove _total_written member variable and all references - Remove GetReadPos() method (read_pos calculation now inline in Read()) - Remove test_get_read_pos() test function - Simplify RingBuffer for exclusive use by Conv1D layer - Update tests to work without GetReadPos()
…dd GetHeadOutputs() - Remove head_outputs parameter from both Process() overloads - Add GetHeadOutputs() method to retrieve head outputs from _head_rechannel - Update WaveNet::process() to use GetHeadOutputs() directly from layer arrays - Remove deprecated set_num_frames_() methods from _Layer and _LayerArray - Update _Layer::SetMaxBufferSize() to use Conv1D::SetMaxBufferSize() instead of Reset() - Simplify head input zeroing in LayerArray::Process()
- Remove _DilatedConv wrapper class, use Conv1D directly in _Layer - Refactor LayerArray::Process() to extract common logic into ProcessInner() - Make _Layer::Process() RT-safe by removing resize in Process() - Update _Layer::SetMaxBufferSize() to use Conv1D::SetMaxBufferSize() - Update tests and ConvNet to use SetMaxBufferSize() instead of Reset() - Simplify _Layer::Process() by combining conv and input_mixin processing
- Split WaveNet tests into three files by component: - test_layer.cpp: Tests for individual WaveNet layers - test_layer_array.cpp: Tests for layer arrays - test_full.cpp: Tests for full WaveNet models - Use nested namespaces: test_wavenet::test_layer, test_wavenet::test_layer_array, test_wavenet::test_full - Remove old monolithic test_wavenet.cpp file - Update test runner to include new test files directly - Add 12 new comprehensive tests covering: - Gated and non-gated layers - Different activations - Multi-channel layers - Layer array processing - Receptive field calculations - Full model processing - Edge cases (zero input, different buffer sizes) - Prewarm functionality
- Add Conv1D constructor with shape parameters - Rename Conv1D::Reset() to SetMaxBufferSize(), remove unused sampleRate param - Add Conv1D::has_bias() method - Reorganize RingBuffer public/private interface (move internal methods to private) - Improve test assertions with numerical accuracy checks - Clean up ring buffer tests, remove internal state checks - Remove commented code from WaveNet - Update NOTES with completed tasks
- Refactor ConvNetBlock to add Process() and GetOutput() methods using new Conv1D API - Simplify ConvNet::process() to eliminate _block_vals for Conv1D layers - Simplify buffer management - Conv1D handles its own ring buffers - Fix test expectations in test_conv1d.cpp (corrected weight ordering) - Fix test bug in test_conv1d.cpp (output2 -> output) - Fix ring buffer test assertion - Add comprehensive ConvNet tests (test_convnet.cpp) - Update test runner to include ConvNet tests This completes the refactoring work for issue #145, making ConvNet use Conv1D's new ring buffer API similar to how WaveNet was refactored.
Fixed assertion failure 'DenseBase::resize() does not actually allow to resize' at wavenet.cpp:61. The issue occurred when assigning _z.leftCols() to _output_head.leftCols() for gated layers, where _z has 2*channels rows but _output_head has only channels rows. Fix: Use _z.topRows(channels).leftCols(num_frames) for gated layers to correctly select only the first channels rows that should be copied to _output_head.
Fixed incorrect weight counts in ConvNet tests. The tests were missing bias weights when batchnorm=false (since !batchnorm=true means do_bias=true in ConvNetBlock::set_weights_). Changes: - test_convnet_basic: Added 4 bias weights (2 per block) - 15 to 19 weights - test_convnet_batchnorm: Removed 1 extra bias (batchnorm=true means no bias) - 10 to 9 weights - test_convnet_multiple_blocks: Added 6 bias weights (2 per block) - 23 to 29 weights - test_convnet_zero_input: Added 1 bias weight - 3 to 4 weights - test_convnet_different_buffer_sizes: Added 1 bias weight - 3 to 4 weights - test_convnet_prewarm: Added 6 bias weights (2 per block) - 23 to 29 weights - test_convnet_multiple_calls: Added 1 bias weight - 3 to 4 weights All tests now pass successfully.
Added SetMaxBufferSize and Process methods to ConvNetBlock to manage output buffers independently from Conv1D, allowing proper batchnorm and activation application on block-owned buffers.
- Remove WaveNet::_head_output member variable (never read) - Remove WaveNet::_set_num_frames_ method declaration (never implemented) - Remove _LayerArray::_get_receptive_field() private method (has TODO to remove) - Remove _Head::_head member variable (never used) - Remove entire _Head class (never instantiated in WaveNet)
- Add .noalias() to matrix assignments for better performance - Remove unnecessary _z.setZero() call (matrix is initialized as needed) - Remove redundant comment in SetMaxBufferSize - Add build workaround for conv1d.cpp Eigen warnings
- Add test_process_realtime_safe() to verify no allocations during process() - Add allocation tracking using malloc/free hooks to catch Eigen allocations - Add helper tests to verify allocation tracking works correctly - Test processes buffers of multiple sizes with two layer arrays - Ensures WaveNet::process() is real-time safe (no allocations/frees)
- Add test_conv1d_process_realtime_safe() to test Conv1D with full matrix input - Add test_conv1d_process_block_realtime_safe() to test Conv1D with Block input - Add test_layer_process_realtime_safe() to test Layer::Process() - Add test_layer_array_process_realtime_safe() to test LayerArray::Process() - Tests confirm Conv1D allocates when passed Block expressions - Attempt to fix RingBuffer::Write() to avoid allocations (work in progress)
- Remove Conv1D Block-input real-time test and rely on full-matrix path - Ensure tests and runtime code pass full buffers between layers, slicing only inside - Simplify RingBuffer::Write to take MatrixXf and add note about full-buffer requirement
- Add full-buffer getters for Conv1x1, _Layer, and _LayerArray - Change LayerArray and WaveNet to pass full MatrixXf buffers and slice internally - Simplify RingBuffer::Write API and document the full-buffer requirement - Restore real-time safety test file and keep only full-matrix Conv1D test in runner
- Remove GetOutputHead(num_frames) and GetOutputNextLayer(num_frames) from _Layer - Remove GetLayerOutputs(num_frames) and GetHeadOutputs(num_frames) from _LayerArray - Rename *Full() methods to original names (GetOutputHead, GetOutputNextLayer, etc.) - Update Conv1D and Conv1x1 GetOutput() to return full buffer (no num_frames param) - Update all internal code and tests to use .leftCols(num_frames) on full buffers - All methods now return pre-allocated full buffers; callers slice as needed
- Add groups parameter to Conv1D (defaults to 1 for backward compatibility) - Support grouped convolutions in ConvNet and WaveNet architectures - Update JSON config keys to use 'groups' instead of 'numGroups' - Reorder function arguments to place groups before weights iterator - Remove unused get_num_groups() method Note: This is a work in progress. Some test files may need updates.
- Add groups parameter (defaults to 1) to all _LayerArray and LayerArrayParams constructor calls in test files - Replace magic number 1 with named 'groups' variable for clarity - Update test_layer_array.cpp, test_full.cpp, and test_real_time_safe.cpp - All tests now build successfully
- Add run_allocation_test helper function that takes setup, test, and teardown functions - Add convenience wrappers for zero-allocation and expected-allocation tests - Refactor all existing tests to use the new abstraction - All tests pass successfully
- Merged PReLU activation support from main - Merged gating activations tests and related functionality - Preserved grouped convolution (groups parameter) support - All conflicts resolved, tests passing
…s_ method for improved clarity and consistency in grouped convolution tests.
…blending tests for cleaner test results.
- Implemented multiple tests for grouped convolution, including basic processing, bias handling, multiple groups, kernel size variations, dilation, and channel isolation. - Added assertions to verify correct output and weight calculations for grouped convolutions. - Updated run_tests.cpp to include new test cases for grouped convolution functionality.
- Implemented tests for Conv1D grouped convolution processing, ensuring no memory allocations occur during execution. - Added tests for both grouped and grouped dilated convolution scenarios to validate real-time safety. - Updated run_tests.cpp to include the new test cases for comprehensive coverage of grouped convolution functionality.
- Modified the _Layer constructor and related methods to accept groups_input instead of groups for improved clarity. - Updated all relevant test cases in test_layer and test_real_time_safe to utilize the new groups_input parameter. - Ensured consistency across the codebase by replacing instances of groups with groups_input in the WaveNet implementation.
- Implemented test_layer_grouped_process_realtime_safe() to verify that the Layer::Process() method with grouped convolution does not allocate or free memory during execution. - Updated run_tests.cpp to include the new test case for comprehensive coverage of grouped convolution functionality.
- Removed the BUILD_TYPE environment variable from the GitHub Actions workflow. - Changed the CMake build type to Debug in the build workflow for better debugging during CI. - Updated CMakeLists.txt to ensure assertions are enabled for the run_tests target by undefining NDEBUG in specific build configurations. - Improved clarity in test_conv1d.cpp by refining comments related to weight initialization for grouped convolutions.
- Updated the apply method to override the base class virtual method, changing the input from a vector of floats to a pointer and size for improved performance in batch processing. - Adjusted the loop to iterate over the data using indices instead of range-based for loop for better control and efficiency.
Owner
Author
|
|
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Grouped convolutions in
Conv1Dusing thegroupsargument.Relevant to #172 , but need
Conv1x1before resolving