how to reproduce office-31's W->A result in paper

[sunbaigui]

I've tried to reproduce the result of office-31, but there's no good. I only got 74.6 in W->A, which is far away from 76.8 as claimed in the paper. Could you please kindly help me to see if there's any trick I've missed.

[ly19965]

Similar un-reproduced result is occurred in the A->W of office-31. We achieve only 94.088 performance, which is far away from 96.1 reported in your paper. Could you please kindly help me to see if there's any trick I've missed. Log is as follows:

arch = resnet50 gpu_id = 6 dset = office s_dset_path = data/list/office/amazon_31.txt t_dset_path = data/list/office/webcam_31.txt output_dir = log/office31 workers = 4 epochs = 80 iters_per_epoch = 500 print_freq = 100 batch_size = 32 lr = 0.01 momentum = 0.9 weight_decay = 0.001 seed = 1 lambda0 = 0.25 MI = 0.1

epoch = 00, best_acc1 = 86.038, best_acc1 = 86.038 epoch = 01, best_acc1 = 90.314, best_acc1 = 90.314 epoch = 02, best_acc1 = 91.447, best_acc1 = 91.447 epoch = 03, best_acc1 = 92.075, best_acc1 = 92.075 epoch = 04, best_acc1 = 92.075, best_acc1 = 92.075 epoch = 05, best_acc1 = 92.704, best_acc1 = 92.704 epoch = 06, best_acc1 = 93.333, best_acc1 = 93.333 epoch = 07, best_acc1 = 93.711, best_acc1 = 93.711 epoch = 08, best_acc1 = 93.711, best_acc1 = 93.711 epoch = 09, best_acc1 = 93.333, best_acc1 = 93.711 epoch = 10, best_acc1 = 93.836, best_acc1 = 93.836 epoch = 11, best_acc1 = 93.333, best_acc1 = 93.836 epoch = 12, best_acc1 = 93.459, best_acc1 = 93.836 epoch = 13, best_acc1 = 93.585, best_acc1 = 93.836 epoch = 14, best_acc1 = 94.088, best_acc1 = 94.088 epoch = 15, best_acc1 = 93.333, best_acc1 = 94.088 epoch = 16, best_acc1 = 93.333, best_acc1 = 94.088 epoch = 17, best_acc1 = 93.208, best_acc1 = 94.088 epoch = 18, best_acc1 = 92.956, best_acc1 = 94.088 epoch = 19, best_acc1 = 92.579, best_acc1 = 94.088 epoch = 20, best_acc1 = 92.956, best_acc1 = 94.088 epoch = 21, best_acc1 = 92.956, best_acc1 = 94.088 epoch = 22, best_acc1 = 93.082, best_acc1 = 94.088 epoch = 23, best_acc1 = 92.830, best_acc1 = 94.088 epoch = 24, best_acc1 = 92.956, best_acc1 = 94.088 epoch = 25, best_acc1 = 92.453, best_acc1 = 94.088 epoch = 26, best_acc1 = 92.579, best_acc1 = 94.088 epoch = 27, best_acc1 = 92.579, best_acc1 = 94.088 epoch = 28, best_acc1 = 92.579, best_acc1 = 94.088 epoch = 29, best_acc1 = 92.830, best_acc1 = 94.088 epoch = 30, best_acc1 = 92.830, best_acc1 = 94.088 epoch = 31, best_acc1 = 92.830, best_acc1 = 94.088 epoch = 32, best_acc1 = 92.830, best_acc1 = 94.088 epoch = 33, best_acc1 = 92.956, best_acc1 = 94.088 epoch = 34, best_acc1 = 92.830, best_acc1 = 94.088 epoch = 35, best_acc1 = 92.704, best_acc1 = 94.088 epoch = 36, best_acc1 = 92.704, best_acc1 = 94.088 epoch = 37, best_acc1 = 92.453, best_acc1 = 94.088 epoch = 38, best_acc1 = 92.956, best_acc1 = 94.088 epoch = 39, best_acc1 = 92.453, best_acc1 = 94.088 epoch = 40, best_acc1 = 93.208, best_acc1 = 94.088

[Xiemixue]

It seems that you are trying to reproduce the results of "BSP+TSA". The BSP in our paper is short for BSP+CDAN. We run "BSP+TSA" based on the open code at https://github.com/thuml/Transfer-Learning-Library/tree/master/examples/domain_adaptation/classification/cdan.py. You can add BSP loss and our TSA loss on the basis.

[sunbaigui]

It seems that you are trying to reproduce the results of "BSP+TSA". The BSP in our paper is short for BSP+CDAN. We run "BSP+TSA" based on the open code at https://github.com/thuml/Transfer-Learning-Library/tree/master/examples/domain_adaptation/classification/cdan.py. You can add BSP loss and our TSA loss on the basis.

followed your suggestion, still can not reproduce， could you please release the code for this task ?