landiisotta / convae_architecture Goto Github PK

Hey, I got a question about the model you used: why did you decided to go for an Autoencoder rather than a Transformer? I couldn't find any information about that in the paper.

Best regards

I would like to apply the ConvAE method to another EHR dataset. In order to do this, having the vocabulary of medical concepts would be necessary. Would you be willing to share this vocabulary?

Hi, I am interested to know about the pre-trained word2vec using Skip-gram which you have used in the code (patient_representations.py)
May I know, which pre-trained was loaded? (e.g. model = Word2Vec.load(emb_filename))
or did you run the word2vec training?

Thank you.

As far as I can tell the precomputed word2vec embeddings are not being loaded as the model is currently written. In model.py we have :

# load pre-computed embeddings
cnd_emb = pre_embs is not None
cnd_vocab = vocab is not None
if cnd_emb and cnd_vocab:
    weight_mtx = np.zeros((vocab_size, emb_size))
    wfound = 0
    for i in range(vocab_size):
        if i != 0:
            try:
                weight_mtx[i] = pre_embs[vocab[i]]
                wfound += 1
            except KeyError:
                weight_mtx[i] = np.random.normal(
                    scale=0.6, size=(emb_size,))
    print('Found pre-computed embeddings for {0} concepts'.format(
        wfound))
    self.embedding.from_pretrained(torch.FloatTensor(weight_mtx))

The final line does not update the existing self.embedding, instead it returns the pretrained embedding. It needs to be reassigned:

self.embedding = self.embedding.from_pretrained(torch.FloatTensor(weight_mtx))

As currently written the weight_mtx is computed then never used, and the embedding portion of the network starts from random initialisation and is trained as part of the model. Making the above change preloads the computed embeddings, but self.embedding.from_pretrained() also includes freeze = True by default, so with the above change the precomputed embeddings are loaded and then remain fixed during training. In the paper you say "These embedding vectors were then used as input for the ConvAE module and were further refined during the model training." Therefore it seems that what this line should be is:

self.embedding = self.embedding.from_pretrained(torch.FloatTensor(weight_mtx), freeze = False)

I would like to know where can I find the vocabulary mentioned in your paper. I see you have added a snippet of the vocabulary in the github file. But I am trying to implement the vocabulary in my data set. Thank you!