zhongyy / unequal-training-for-deep-face-recognition-with-long-tailed-noisy-data Goto Github PK

Hi

Can you share the pre-trained and dataset on to to gdrive or dropbox ? we are finding it difficult to get baidu.

Cheers

Hi, I just found the parameter noise_beta of NR loss, but I did not found the parameter t, 安and the parameter does not change dynamically as mentioned in the paper, it is a constant number. If I'm wrong, please correct me. Thank you.

in train_NR.py line 301
cross_entropy = cross_entropy / args.batch_size/2

could you explain why the loss should be divided by the batch size and then multiply by 2? thank you

Traceback (most recent call last):
File "train_debug_soft_gs_y0.py", line 530, in
main()
File "train_debug_soft_gs_y0.py", line 527, in main
train_net(args)
File "train_debug_soft_gs_y0.py", line 521, in train_net
epoch_end_callback = epoch_cb )
File "/home/yaoxi/anaconda2/lib/python2.7/site-packages/mxnet/module/base_module.py", line 520, in fit
next_data_batch = next(data_iter)
File "/DATA1/yaoxi/Unequal-Training-for-Deep-Face-Recognition-with-Long-Tailed-Noisy-Data/data_longtail_gs.py", line 425, in next
self.reset()
File "/DATA1/yaoxi/Unequal-Training-for-Deep-Face-Recognition-with-Long-Tailed-Noisy-Data/data_longtail_gs.py", line 377, in reset
self.interclass_reset()
File "/DATA1/yaoxi/Unequal-Training-for-Deep-Face-Recognition-with-Long-Tailed-Noisy-Data/data_longtail_gs.py", line 339, in interclass_reset
id_sel = self.pick_interclass(embeddings, nrof_images_per_class, self.batchsize_id) # shape=(T,3)
File "/DATA1/yaoxi/Unequal-Training-for-Deep-Face-Recognition-with-Long-Tailed-Noisy-Data/data_longtail_gs.py", line 210, in pick_interclass
centers[i] = np.mean([embeddings[iself.images_per_identity], embeddings[iself.images_per_identity+1],embeddings[i*self.images_per_identity+2]], axis=0)
IndexError: index 3600 is out of bounds for axis 0 with size 3600

Hi.
I can't find the code of generating the denoised head data or generating the 0_noiselist.txt in stage2. Could you share and upload the code?
Thank you so much.