Tensorboardx: How to draw a picture about attention ?

[travel-go]

Hello, I want to draw a picture about attention which size is batchsrc_lenthtgt_lenth,like a Tensor (20,30,25),so I want to draw one picture which size is 30*25,but I find the function of add_image's image_tensor should be given by a Tensor like(3,H,W).
I am confused that how can I draw this picture with only H-W.
I can copy this dimension three times,Is it right?

[lanpa]

Passing a Tensor of size (20,30,25) should be handled automatically by tensorboardX. The only thing you need to do is normalize your image to [0, 1]. The doc is too old.

[travel-go]

Thank you for your reply.Do you mean to directly send tensor (20, 30, 25) to the add_image?When I do this,I meet a problem.

 12 Traceback (most recent call last):$
 13   File "/home/hfyu/anaconda/lib/python3.6/site-packages/PIL/Image.py", line 2423, in fromarray$
 14     mode, rawmode = _fromarray_typemap[typekey]$
 15 KeyError: ((1, 1, 200), '|u1')$
 16 $
 17 During handling of the above exception, another exception occurred:$
 18 $
 19 Traceback (most recent call last):$
 20   File "train.py", line 29, in <module>$
 21     main(args)$
 22   File "train.py", line 23, in main$
 23     singleprocess_main(args)$
 24   File "/home/hfyu/fairseq-py-0.4.0/singleprocess_train.py", line 86, in main$
 25     train(args, trainer, dataset, epoch, batch_offset)$
 26   File "/home/hfyu/fairseq-py-0.4.0/singleprocess_train.py", line 157, in train$
 27     log_output = trainer.train_step(sample)$
 28   File "/home/hfyu/fairseq-py-0.4.0/fairseq/trainer.py", line 97, in train_step$
 29     loss, sample_sizes, logging_outputs, ooms_fwd = self._forward(sample)$
 30   File "/home/hfyu/fairseq-py-0.4.0/fairseq/trainer.py", line 145, in _forward$
 31     loss, sample_size, logging_output_ = self.criterion(self.model, sample)$
 32   File "/home/hfyu/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__$
 33     result = self.forward(*input, **kwargs)$
 34   File "/home/hfyu/fairseq-py-0.4.0/fairseq/criterions/cross_entropy.py", line 30, in forward$
 35     net_output = model(**sample['net_input'])$
 36   File "/home/hfyu/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__$
 37     result = self.forward(*input, **kwargs)$
 38   File "/home/hfyu/fairseq-py-0.4.0/fairseq/models/fairseq_model.py", line 44, in forward$
 39     decoder_out = self.decoder(prev_output_tokens, encoder_out)$
 40   File "/home/hfyu/anaconda/lib/python3.6/site-packages/torch/nn/modules/module.py", line 491, in __call__$
 41     result = self.forward(*input, **kwargs)$
 42   File "/home/hfyu/fairseq-py-0.4.0/fairseq/models/fconv.py", line 323, in forward$
 43     self.writer.add_image('atten_Image', avg_attn_scores)$
 44   File "/home/hfyu/anaconda/lib/python3.6/site-packages/tensorboardX/writer.py", line 342, in add_image$
 45     self.file_writer.add_summary(image(tag, img_tensor), global_step)$
 46   File "/home/hfyu/anaconda/lib/python3.6/site-packages/tensorboardX/summary.py", line 158, in image$
 47     image = make_image(tensor)$
 48   File "/home/hfyu/anaconda/lib/python3.6/site-packages/tensorboardX/summary.py", line 166, in make_image$
 49     image = Image.fromarray(tensor)$
 50   File "/home/hfyu/anaconda/lib/python3.6/site-packages/PIL/Image.py", line 2426, in fromarray$
 51     raise TypeError("Cannot handle this data type")$
 52 TypeError: Cannot handle this data type$

[lanpa]

Sorry. It will not convert automatically in this case. You can reshape to (20, 1, 30, 25). or stacking it to (20, 3, 30, 25)

[travel-go]

Thank you for your reply,I get it.

[travel-go]

Trouble you,I have another problem,the attention picture I draw is too dark,I think maybe the number of my attention's Tensor is between 0 and 1(max is 1)so ,I hope it can be showed with percentage.
Or what is the cause of this picture too dark?

[travel-go]

@lanpa

[lanpa]

maybe your feature map is just empty. you can check the distribution with add_histogram().

[travel-go]

OK, I'll try it

[travel-go]

oh，I change the Tensor（20,30,25）to （1,1,30,25）like this.because I just want to one picture.
a= torch.randn(20,30,25) a = a[:1,:,:].unsqueeze(0)
Does it have effects?

[travel-go]

Thank you ,I solve it