Some question about office31 sets

[hongshen-zhang]

Your work is marvelous! I am a new-comer in transfer learning. I could test your code in digit sets(mnist, usps,svhn). However, I tried for several days but I still can't test your code in office31 sets (amazon,webcam,dslr). Could you please give me an concrete example on how to test your code in office31 sets?
(Just like python scripts/train_digits.py --source amazon --target webcam --dann)
Thank you very much!

[stes]

Hi @jkljk67890 , thanks for your interest! I'm afraid the current code base does not have a standalone and easy-to-run example for Office31. You'd need to wrap a model like a Resnet from torchvision and adapt it for domain adaptation; this primarily boils down to adjusting the forward function to return a feature vector and the logits, since the solvers internally call it like this:

salad/salad/solver/da/coral.py

Line 25 in 5f55d6f

src_e, src_p = self.model(src_x)

I'd suggest to adapt the digit example to perform this task; please let me know if you need further help, in case you're up for implementing this, feel free to open a PR afterwards!

[hongshen-zhang]

Thank you for your request! I have tried my best to change the code, but I met a question that

$ python train_OFFICE.py --gpu 0 --source amazon --target webcam --assoc
Start Experiments
Registered DigitModel as "classifier"
Registered CrossEntropyLoss as "ce" with weight 1.0
Registered AccuracyScore as "acc_s" with weight None
Registered AccuracyScore as "acc_t" with weight None
Registered AssociativeLoss as "assoc" with weight 1.0
Registered Adam as ""
0%| | 0/100 [00:00<?, ?it/sTraceback (most recent call last): | 0/8 [00:00<?, ?it/s]
File "train_OFFICE.py", line 133, in
experiment.optimize()
File "/home/zhanghongshen/good/salad-master/scripts/salad/solver/base.py", line 459, in optimize
history, terminate = self._epoch(epoch)
File "/home/zhanghongshen/good/salad-master/scripts/salad/solver/base.py", line 423, in _epoch
total_losses = self._step(batch)
File "/home/zhanghongshen/good/salad-master/scripts/salad/solver/base.py", line 391, in _step
loss, loss_dict = self._loss(optimizer, batch)
File "/home/zhanghongshen/good/salad-master/scripts/salad/solver/base.py", line 366, in _loss
loss_args = self.agg_lossoptim
File "/home/zhanghongshen/good/salad-master/scripts/salad/solver/da/association.py", line 38, in call
(xs, ys), (xt, yt) = batch
ValueError: too many values to unpack (expected 2)
(uses assoc).

I tried for several algorithm and their error are similar.

(use teach: File "/home/zhanghongshen/good/salad-master/scripts/salad/solver/da/ensembling.py", line 20, in call
(x_stud_xs, ys), (x_stud_xt,x_teach_xt,yt) = batch
ValueError: too many values to unpack (expected 2)）

It seems that there is something wrong with batch, but I don't know where to change.
Could you please tell me where should I change?
(I print the batch,it was too large, and it can be run for several times and then meet an error(too many values).)