mht
|
0fdbcedc8e
|
Fix: ResNet BatchNorm discrepancies and comparison script issues
- **C++ ResNet BatchNorm Fix**: Remove manual float64 BatchNorm computation and use standard float32 forward() to match Python behavior
- Replace manual BatchNorm calculation with bn->forward(x) for both training and eval modes
- This resolves major discrepancies in Layer1-4 outputs, achieving perfect cosine similarity (1.0000)
- **Python Comparison Script Fix**: Remove unsupported output layer requests that caused "output_layer is wrong" errors
- Only request layers that Python ResNet actually supports: ['conv1', 'layer1', 'layer2', 'layer3', 'layer4']
- Remove requests for intermediate debug layers (bn1_output, conv1_pre_bn, etc.)
- Update layer mapping logic to gracefully skip unsupported layers
- Fix Conv1 mapping from 'conv1_pre_bn' to 'conv1'
- **Test Models Fix**: Remove invalid debug method calls that caused build errors
- Remove calls to non-existent debug_get_conv3_1t_output() and debug_get_conv4_1t_output()
- Restore correct BBRegressor processing using valid methods (get_iou_feat, get_modulation, predict_iou)
**Results:**
- ResNet Layer1-4 and Features now have perfect cosine similarity (1.0000)
- Build errors resolved, comparison script runs successfully
- BatchNorm running_mean, running_var, and num_batches_tracked fixes working correctly
- Conv1 still has issues (0.5291 cosine similarity) - separate investigation needed
|
3 weeks ago |
