Incubator-mxnet: what does mx.sym.Embedding() do for vocab index data

[amitkb3]

Hi,
I am developing for text classification using cnn based on the python example provided. I needed to understand what does mx.sym.Embedding() do. I have looked at https://github.com/dmlc/mxnet/issues/2176 but could not follow.

Here is my steps. I take a sentence "Mxnet is cool" and first i pad it to fixed length 56 and then convert each word to vocabulary index. Hence the input to cnn will look like [ 1, 3450, 2303, 0, 0 , ...,0]

This is fed into mx.sym.Embedding() as per the exmple:
embed_layer = mx.sym.Embedding(data=input_x, input_dim=vocab_size, output_dim=num_embed, name='vocab_embed')

here vocab_size is 18765 and num_embed is 300

My question is what is num_embed and what will be the output of mx.sym.Embedding(). How will the result look like.

[zihaolucky]

use infer_shape to help you debug/understand.

# network definition.
input_x = mx.sym.variable('input_x')
embed_layer = mx.sym.Embedding(data=input_x, input_dim=vocab_size, output_dim=num_embed, name='vocab_embed')

# infer_shape
input_shape = {'input_x' : (batch_size, xx, xx)} # choose it by yourself.
embed_layer.infer_shape(**input_shape) 

[amitkb3]

Thanks for help. I am getting below error. I have a training dataset "x_train". I am not able to figure out what the error is.
I appreciate all your help.
code:

network definition.

input_x = mx.sym.variable('x_train')
embed_layer = mx.sym.Embedding(data=input_x, input_dim=vocab_size, output_dim=num_embed, name='vocab_embed')

infer_shape

input_shape = {'input_x' : (batch_size, 1, 5)} # choose it by yourself.
embed_layer.infer_shape(**input_shape)

error:

AttributeError Traceback (most recent call last)
in ()
1 # network definition.
----> 2 input_x = mx.sym.variable('x_train')
3 embed_layer = mx.sym.Embedding(data=input_x, input_dim=vocab_size, output_dim=num_embed, name='vocab_embed')
4
5 # infer_shape

AttributeError: module 'mxnet.symbol' has no attribute 'variable'

[CNevd]

mx.symbol.Variable

[pluskid]

We just pushed some details for the doc of the embedding layers in #3596. They should appear in the online doc once merged. Hope this clarifies the point.

[amitkb3]

thanks a lot!