Implementation of Recurrent Convolutional Neural Network for Object Recognition.
Refer to requirements.txt
Here is the architecture of the model.
Refer to the paper below for more details.
the implementation is a little bit different from the original paper.
• Instead of LRN, using BN here and removing dropout layers except the last linear layer.
• As BN layers introduce extra parameters, the feed-forward filter size in layer 1 is modified to 3 x 3.
• For simplification, using Tencorp (there exsits api in pytorch) for test here instead of the original nine crops.
usage: train.py [-h] [-n K] [-b BATCH_SIZE] [-e EPOCH] [-s SAVE_DIR] [-l TRAINING_LOG]
Train RCNN
optional arguments: -h, --help show this help message and exit -n K the parameter K for RCNN -b BATCH_SIZE the batch size in just one gpu, * GPU_COUNT -e EPOCH the training epoch -s SAVE_DIR the model parameters to be saved -l TRAINING_LOG the logs to be saved
With Data Augmentation
| Network | No. of Parameters | Testing Error (%) |
|---|---|---|
| RCNN-96 | 0.67 M | 7.29 |
| RCNN-128 | 1.19 M | 6.76 |
| RCNN-160 | 1.86 M | 6.66 |
The result improves slightly compared with that in 4.2.2.Maybe the hyper-parameter can be adjusted to get a better model.
