Modular ResNet implementation using tensorflow. PreActivation approach is used for all residual or bottleneck-blocks.
Currently, the ResNet-110 for CIFAR-10 is implemented. The first goal is to match the reported results of the paper on CIFAR-10 (see table below).
The training set is split into 45k training images and 5k validation images to determine the training parameters and global_step for early stopping. Afterwards, it is retrained using the full training dataset. Results are reported on the test set.
| Dataset | Layers | Results | Optimizer | Reported Results |
|---|---|---|---|---|
| CIFAR-10 | 110 | 83.28% | Adam | 93.67 % |
| CIFAR-10 | 110 | 91.58% | Momentum-SGD | 93.67 % |
- batch size 32 due to GPU-limitations
- Learning rate schedule adapted due to different batch size