If you test it with 224x224, the top-1 accuracy will be 81.82%. Note that it is trained with 224x224 but tested with 320x320, so that it is still trainable with a global batch size of 256 on a single machine with 8 1080Ti GPUs. Will release the quantized models after tuning the hyper-parameters and finishing the QAT.Ī deeper RepVGG model achieves 83.55% top-1 accuracy on ImageNet with SE blocks and an input resolution of 320x320 (and a wider version achieves 83.67% accuracy without SE). Such a weight decay also improves the full-precision accuracy. Jfound out that high-performance quantization required a custom weight decay. Just add -custwd to the training command. JTraining with the custom weight decay has been tested. JAn example of using a simple toolbox, torch.quantization, to quantize RepVGG. JA pure-VGG model (without SE) seems to outperform some vision transformer models with a better training scheme. To use it for downstream tasks like semantic segmentation, just discard the aux classifiers and the final FC layer. Loss = criterion(outputs, targets) # Your original code Loss += 0.1 * criterion(pred, targets) # Assume "criterion" is cross-entropy for classification # 'L2': the custom L2 regularization term # '*aux*': the output of auxiliary classifiers # A training-time RepVGGplus outputs a dict. Outputs = model(samples) # Your original code For samples, targets in enumerate(train_data_loader):
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |