Tensorflow summery merge error : Shape [-1,784] has negative dimensions

Question

I am trying to get summary of a training process of the neural net below.

import tensorflow as tf 
import numpy as n         


        

Answer 1

From one comment of the deleted answer, from the original poster:

I actually build a neural net under with tf.Graph() as g. I removed the interactive session and started session as with tf.Session(g) as sess. It fixed the problem.

The graph g was not marked as the default graph that way, thus the session (tf.InteractiveSession in the original code) would use another graph instead.

Note that I stumbled upon here because of the same error message. In my case, I had accidentally something like this:

input_data = tf.placeholder(tf.float32, shape=(None, 50))
input_data = tf.tanh(input_data)
session.run(..., feed_dict={input_data: ...})

I.e. I didn't feed the placeholder. It seems that some other tensor operations can then result in this confusing error as internally an undefined dimension is represented as -1.

Answer 2

This has may have to do with the InteractiveSession initialization.

I initialized it at the beginning and then it worked - then initialized the global variables within the session.

I am unable to reproduce the error with the old code, which makes it unpredictable or caching settings somewhere.

import tensorflow as tf
sess = tf.InteractiveSession()


from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

x = tf.placeholder(tf.float32, [None, 784])

W = tf.Variable(tf.zeros([784,10]))

b = tf.Variable(tf.zeros([10]))

y = tf.nn.softmax(tf.matmul(x, W)+b)

y_ = tf.placeholder(tf.float32, [None,10])



cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y), reduction_indices=[1]))
train_step = tf.train.GradientDescentOptimizer(0.05).minimize(cross_entropy)
sess.run(tf.global_variables_initializer())


for _ in range(1000):
    batch_xs, batch_ys = mnist.train.next_batch(100)
    #print batch_xs.shape, batch_ys.shape
    sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

Answer 3

I was also having this problem. Searching around the basic consensus is to check for problems somewhere else in your code.

What fixed it for me was I was doing a sess.run(summary_op) without feeding in data for my placeholders.

Tensorflow seems to be a bit strange with placeholders, often they won't mind you not feeding them if you're trying to evaluate part of the graph that is independent of them. Here though, it did.