The parameter cycle_lens specifies the length of the training, and it is adjusted depending on the amount of images in the training set.įor instance, in the DRIVE case, -cycle_lens 20/50 implies that we train for 20 cycles, each cycle running for 50 epochs.Īs CHASE-DB has less training images than DRIVE (8 vs 16), we double the number of cycles in that case. This will store the model weights in experiments/wnet_drive, experiments/wnet_chasedb, experiments/wnet_hrf respectively. im_size 1024 -batch_size 2 -grad_acc_steps 1 -device cuda:0 model_name wnet -save_path wnet_hrf_1024 Python train_cyclical.py -csv_train data/HRF/train.csv -cycle_lens 30/50 model_name wnet -save_path wnet_chasedb -device cuda:0 Python train_cyclical.py -csv_train data/CHASEDB/train.csv -cycle_lens 40/50 model_name wnet -save_path wnet_drive -device cuda:0 Python train_cyclical.py -csv_train data/DRIVE/train.csv -cycle_lens 20/50 To reproduce our results in table 2 of our paper, you need to run: Note that the training defaults to using the CPU, which is feasible due to the small size of our models. You also need to supply the path to save the model. Note: The LES-AV dataset is still public but it now needs to be downloaded manually, please see the comments in get_public_data.py Line 400 forward for details. The HRF dataset will also contain folders called images_resized, manual_resized, mask_resized. Note: The DRIVE dataset will also contain a folder called ZoneB_manual, which is used to evaluate A/V performance around the optic disc.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |