jerbarnes
How to get the result in paper
I use parameter_search to get the best parameter, but can't reach the best result proposed in your paper, is there any guide to help me get the result ?
Hi, sure I'd love to help. Can you give me a bit more information about the system you're running the experiments on, the embeddings you use, and the current results?
I run the experiment on torch 0.2.0_3 , and use the google.txt for English word embeddings, sg-300-es for Spanish word embeddings，bilingual dictionary from Bingliu，but I just get the best result when batch size is 80 and alpha 0.8，with f1=0.540.
Besides this, I think I should tell you that I always get the best results at epoch 1, is it wired?
Besides this, I think I should tell you that I always get the best results at epoch 1, is it wired?
I had the same problem trying to run blse.py and I think I found the main reason this happens.
This line:
out = F.softmax(self.clf(x_proj))
should be replaced by this one:
out = self.clf(x_proj)
Since the criterion being used for calculating the loss is nn.CrossEntropyLoss() and according to pytorch docs:
This criterion combines nn.LogSoftmax() and nn.NLLLoss() in one single class
Which means that softmax is being calculated twice on the model outputs, making it hard for any learning to occur.
Be happy to hear if you think I am missing something @jbarnesspain. Thank you for the article and implementation!
