so has anyone got this pytorch crf to work yet ?

[yuhujia]

Looks like this implementation does not work on real data. If anyone has gotten it to work (aka has good performance and (optional) deployed to production) pls correct me here.

[kmkurn]

I have used this myself for production (when I was still working for my former employer). Sorry it doesn't work for you.

[yuhujia]

great ! can you post scripts/data/steps to reproduce your performance ?

[kmkurn]

Unfortunately no, because it was all proprietary. But it was basically an implementation of this paper.

[yuhujia]

Steps to reproduce numbers mentioned in the paper would be great.

[kmkurn]

I'm not sure what steps you need other than those described in the paper.

[robinsongh381]

I have made good products with REAL data :smile:

[elsheikh21]

@robinsongh381 @kmkurn @yuhujia anyone got it to work or not yet?

[kmkurn]

@elsheikh21 Have you tried it yourself? I personally have used this, and it worked.

[elsheikh21]

I personally tried to use it, however, I could not. I sent you an email

[yuye2133]

There is a problem with the latest version of pytorch(1.4), log-likehihood will overflow.

[elsheikh21]

To all future readers, I would like to start by thanking @kmkurn for helping me out, and here is how to utilize his module, I hope it makes everything clear for you (future reader)

So, how to make this work? Firstly, install the module pip install pytorch-crf Secondly, in your script from torchcrf import CRF

Thirdly, add the CRF layer after the model's final layer, self.crf = CRF(num_classes, batch_first=True) where number of tags/output calsses

Forth, in the training loop

for epoch in range(epochs):
   self.model.train()
   for batch_idx, samples in enumerate(train_dataset):
      loss = -self.model.log_probs(inputs, labels)
      sample_loss.backward()
      self.optimizer.step()
   # Compute val_loss

But what is model.log_probs(inputs, labels)?

    """
     Build a simple BiLSTM 
    """
    def forward(self, x):
        ''' Leave model forward function untouched '''
        embeddings = self.word_embedding(x)  # [Samples_Num, Seq_Len]
        embeddings = self.dropout(embeddings)
        o, _ = self.lstm(embeddings)        # output shape: [Samples_Num, Seq_Len, Tags_Num]
        o = self.dropout(o)
        logits = self.classifier(o)
        return logits

    def log_probs(self, x, tags, mask=None):
        # mask = (x != 0).float()  # assuming that padding index is 0
        emissions = self(x)  # Compute Emission scores
        return self.crf(emissions, tags, mask=mask)

Lastly, how to predict outputs as per model? predictions = self.model.predict(inputs) # returns predicted classes, no need to invoke argmax(..)

what is model.predict(inputs)?

    def predict(self, x):
        emissions = self(x)
        return self.crf.decode(emissions)

I was using Python3.6.4 and PyTorch 1,4,0

[kmkurn]

The library seems to work just fine, so I'm closing this issue.