How to use embeddings-for-trees output? about embeddings-for-trees HOT 12 CLOSED

You can find this function in the commode-utils package. You can also install it with pip.

from embeddings-for-trees.

Hi
Can you please help me with this.
Thanks.

from embeddings-for-trees.

Hi!

.ckpt files correspond to model checkpoints with weights.
If you want to extract embedding you need to write some code :)

1. First of all, you need to load checkpoint. See this example.
2. Since you don't need a decoder, extract embedding and encoder modules:

node_emb = model._TreeLSTM2Seq__embedding
tree_enc = model._TreeLSTM2Seq__encoder

Sorry for not very convenient code, I will look forward to provide more accurate interfaces.
3. Retrieve trees embeddings

batched_trees.ndata["x"] = self.__embedding(batched_trees)
encoded_nodes = self.__encoder(batched_trees)
batched_encoded_nodes, mask = cut_into_segments(
    encoded_nodes, batched_trees.batch_num_nodes(), False
)

4. Now you have embeddings of all nodes in the tree, you can aggregate them by mean for example.

from embeddings-for-trees.

Thanks a lot for detailed explanation. I'll try above steps and will let you know.

from embeddings-for-trees.

Sorry to disturb you again.

batched_encoded_nodes, mask = cut_into_segments(
encoded_nodes, batched_trees.batch_num_nodes(), False
)
Where to put this code patch?

from embeddings-for-trees.

Write after extracting encoded_nodes.
To speed up computation, all trees in the batch are collated into a single tree. This function cuts this tree back, so you will have encoded nodes for each tree in batch and mask to properly slice it.

from embeddings-for-trees.

cut_into_segments function is undefined. Do I need to write this script on own or provided by some library?

from embeddings-for-trees.

Thank you.

from embeddings-for-trees.

Hi,
I created vector embedding for Java files.
Basically, I want to train a model for program repair task, such that it takes input-embeddings of buggy line and learn to generate output for fixed line.
I gave input buggy line and created embeddings as shown below. But I'm confused, how to convert these vectors in source code. As which value in vectors corresponding to which token?

Sorry for this silly question, but if you can provide some pointers it will be really helpful.
Thanks.

from embeddings-for-trees.

Wow, it is quite surprising for me that this even works :)
I use PyTorch as a framework to train the model and you use TensorFlow functions to operate with torch tensors...

Answering your questions:

• You may see that encoded_nodes passed into the decoder module. This module generates a sequence of output_length size. output_logits is a matrix with the shape [sequence length; batch size; vocab size]. So, if you want to generate some code, you should: (1) create a vocabulary with possible tokens for generating (label vocabulary), (2) pass the required output length to the forward method.
• batched_encoded_nodes shape should be something like [batch size; size of max tree in the batch]. The first vector corresponds to the root vector, the following vectors correspond to the tree with respect to their definition.

from embeddings-for-trees.

Alright thanks for help. I'll try with above points.

from embeddings-for-trees.

I close it due to inactivity, but feel free to reopen or create new issues in case of any questions!

from embeddings-for-trees.