CODEWITHSUNDEEP
pythontensorflow
Tilvan
Tilvan

Reputation: 55

DRQN - Prefix tensor must be either a scalar or vector, but saw tensor

In following this tutorial, I am receiving the following error:

ValueError: prefix tensor must be either a scalar or vector, but saw tensor: Tensor("Placeholder_2:0", dtype=int32)

The error originates from these lines:

# Take the output from the final convolutional layer and send it to a recurrent layer
# The input must be reshaped into [batch x trace x units] for rnn processing, and then returned to
# [batch x units] when sent through the upper levels
self.batch_size = tf.placeholder(dtype=tf.int32)
self.convFlat = tf.reshape(slim.flatten(self.conv4), [self.batch_size, self.trainLength, h_size])
# !!!!This is the line where error city happens!!!!
self.state_in = rnn_cell.zero_state(self.batch_size, tf.float32)

After the network is initialized:

mainQN = Qnetwork(h_size, cell, 'main')

This error is still present when solely running the code in a python console so the error is consistent.

I will post more of the code if that will be helpful

Upvotes: 2

Views: 1267

Answers (2)

xingzhang ren
xingzhang ren

Reputation: 106

There is another solution to solve this problem.

Change

self.batch_size = tf.placeholder(dtype=tf.int32)

TO

self.batch_size = tf.placeholder(dtype=tf.int32, [])

Upvotes: 3

xingzhang ren
xingzhang ren

Reputation: 106

I met the same problem with the version of tensorflow is 1.2.+.

When i changed it to 1.1.0, the problem resolved.

I think it because the API of rnn_cell.zero_state makes arg batch_size must be a scalar or vector, but not tensor.

So, if you change batch_size to scalar, e.g. 128, the problem also could be resolved.

Upvotes: 1

Related Questions