list indices must be integers, not list

[AlexAlifimoff]

Hey all,

Working on my first project in Keras, attempting to reimplement this paper: http://arxiv.org/pdf/1509.00685v2.pdf

My code is:

from future import *

from keras.models import Sequential, Graph
from keras.layers.core import Dense, Activation, Reshape, Lambda, Flatten
from keras.layers.recurrent import GRU
from keras.layers.embeddings import Embedding

import theano.tensor as T
import theano
# define parameters intuitively, then translate to notation
# this is unnecessary, but helps with understanding
# during the initial construction of the network.
word_embedding_size = 20
hidden_layer_size = 10
context_length = 5
vocabulary_size = 3 # need to get this from training data... but using this as a stand-in for testing the model

# We are going to use the same notation as the paper.
D = word_embedding_size  # size of the word embeddings
H = hidden_layer_size # hidden layer size
C = context_length 
V = vocabulary_size

print("Building model...")

# We start with two seperate tensor streams for processing y and x respectively.
# Stream is something I made up, not sure if that's a technical term
model = Graph()

# Elements of model
E = Embedding(V, D, input_length=C)
U = Dense(H, activation='tanh') #GRU(H, activation='tanh')
V_ = Dense(V)

model.add_input('y_c',(C,)) # our context... C is size of context, we don't include size V, since we just pass indexes of words in...
model.inputs['y_c'].input = T.imatrix()
model.add_input('x', (None, ))
model.inputs['x'].input = T.imatrix()

 # y processing side of the model
model.add_node( E, 'y_c~', input='y_c' ) # output dimension (nb_samples, C) --> (nb_samples, C, D)
model.add_node( Reshape( dims=(C*D,) ), 'y_c~reshaped', input='y_c~') # here we reshape to a single     vector of C*D # elements per example
model.add_node( U, 'h', input='y_c~reshaped' )
model.add_node( V_, 'V_', input='h' )

# now we build the bag of words encoder...
model.add_node( Embedding(V, H), 'bow_f', input='x')

model.add_node( Reshape(dims = (H,)), 'bow_f_reshaped', input='bow_f')
model.add_node( Dense(V), 'W', input='bow_f_reshaped' )

def exp(X):
    return T.exp(X)

model.add_node( Lambda(exp, output_shape), 'final', inputs=['V_', 'W'], merge_mode='sum')

# output... we sum the two components of our model
model.add_output('output', input='final')

model.compile('sgd', {'output': 'mse'})

print("Model built and compiled")

y = [ [1, 2, 3, 4, 5] ]
x = [ [2, 3, 4] ]

model.predict( {'y_c': y, 'x': x } )

The model compiles successfully, but I'm getting the following error when I try to test it on a simple example.

Building model...
Model built and compiled
Traceback (most recent call last):
  File "abstractive_sentence_summarizer.py", line 75, in <module>
    model.predict( {'y_c': y, 'x': x } )
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-    packages/keras/models.py", line 695, in predict
    outs = self._predict_loop(self._predict, ins, batch_size, verbose)
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/models.py", line 254, in _predict_loop
    ins_batch = slice_X(ins, batch_ids)
  File "/Library/Frameworks/Python.framework/Versions/2.7/lib/python2.7/site-packages/keras/models.py", line 58, in slice_X
    return [x[start] for x in X]
TypeError: list indices must be integers, not list

Sorry if this is an obvious question. I did my due diligence in researching before posting and couldn't seem to find an answer.

[pasky]

Your inputs must be numpy ndarrays, not plain lists. (There should probably be a sanity check...)

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed after 30 days if no further activity occurs, but feel free to re-open a closed issue if needed.