airkid
airkid
Add transforms.Resize((112,112)) into dataset transform
Hi @mikelkl , I believe this is a kind of right implementation of calculating F1 score in this task. Have you reproduce the experiment? I can only reach F1 score...
Hi @mikelkl, can you try on the data split update by author? My result is still far away from the paper.
I got the same problem. In the first few epoch the f1 score keeps 0 then grows slowly. Have you tried using lr==0.01 and optimizer=='adam'? Its training curve looks normally...
@meryemmhamdi1 BTW I'm now also trying to run this code. I'd glad to share my progress and preprocessed data. Send me email if you have interests to work it together
> Hi, Thank you for your message. I have tried to play with the parameters and optimizer but the same pattern repeats itself: “loss decreasing, increasing improving but f1 score...
> I think they are using them in lines: pretrained_embedding = Vectors(self.a.webd, ".", unk_init=partial(torch.nn.init.uniform_, a=-0.15, b=0.15)) WordsField.build_vocab(train_set.WORDS, dev_set.WORDS, vectors=pretrained_embedding) to convert input words directly to vectors. Did you try weights...
I just found the code will crash if I set "loss_alpha">1.0. Have you got the same problem? @lx865712528
Hi, I ran the newer update train script and found that the update rate for argument is still very low (F1
Hi @ycc1028 , after this modification it still can not reach the performance cause there are still another evaluate problem #6