tensorflow logits and labels must be same size

Question

I'm quite new to tensorflow and python, and currently trying to modify the MNIST for expert tutorial for a 240x320x3 image. I have 2 .py script

tfrecord_reeader.py

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

data_path = 'train.tfrecords'  # address to save the hdf5 file

def read_data():
    with tf.Session() as sess:
        feature = {'train/image': tf.FixedLenFeature([], tf.string),
                   'train/label': tf.FixedLenFeature([], tf.int64)}

        # Create a list of filenames and pass it to a queue
        filename_queue = tf.train.string_input_producer([data_path], num_epochs=1)

        # Define a reader and read the next record
        reader = tf.TFRecordReader()
        _, serialized_example = reader.read(filename_queue)

        # Decode the record read by the reader
        features = tf.parse_single_example(serialized_example, features=feature)

        # Convert the image data from string back to the numbers
        image = tf.decode_raw(features['train/image'], tf.float32)

        # Cast label data into int32
        label = tf.cast(features['train/label'], tf.int32)

        # Reshape image data into the original shape
        image = tf.reshape(image, [240, 320, 3])

    sess.close()
    return image, label

def next_batch(image, label, batchSize):
    imageBatch, labelBatch = tf.train.shuffle_batch([image, label], batch_size=batchSize, capacity=30, num_threads=1,
                                            min_after_dequeue=10)
    return imageBatch, labelBatch

train.py

import tensorflow as tf
from random import shuffle
import glob
import sys
#import cv2
from tfrecord_reader import read_data, next_batch
import argparse # For passing arguments
import numpy as np
import math
import time

IMAGE_WIDTH = 240
IMAGE_HEIGHT = 320
IMAGE_DEPTH = 3
IMAGE_SIZE = 240*320*3
NUM_CLASSES = 5
BATCH_SIZE = 50

# Creates a weight tensor sized by shape
def weight_variable(shape):
  initial = tf.truncated_normal(shape, stddev=0.1)
  return tf.Variable(initial)

# Creates a bias tensor sized by shape
def bias_variable(shape):
  initial = tf.constant(0.1, shape=shape)
  return tf.Variable(initial)

def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')

def main(argv):
    # Perform training
    x = tf.placeholder(tf.float32, [None, IMAGE_SIZE])    # 240*320=76800
    W = tf.Variable(tf.zeros([IMAGE_SIZE, NUM_CLASSES]))
    b = tf.Variable(tf.zeros([NUM_CLASSES]))
    y = tf.matmul(x, W) + b

    # Define loss and optimizer
    y_ = tf.placeholder(tf.float32, [None, NUM_CLASSES])  # Desired output

    # First convolutional layer
    W_conv1 = weight_variable([5, 5, IMAGE_DEPTH, 32])
    b_conv1 = bias_variable([32])

    x_image = tf.reshape(x, [-1, IMAGE_WIDTH, IMAGE_HEIGHT, IMAGE_DEPTH])

    h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
    h_pool1 = max_pool_2x2(h_conv1)

    # Second convolutional layer
    W_conv2 = weight_variable([5, 5, 32, 64])
    b_conv2 = bias_variable([64])
    h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
    h_pool2 = max_pool_2x2(h_conv2)

    # First fully connected layer
    W_fc1 = weight_variable([60 * 80 * 64, 1024])
    b_fc1 = bias_variable([1024])

    # Flatten the layer
    h_pool2_flat = tf.reshape(h_pool2, [-1, 60 * 80 * 64])
    h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

    # Drop out layer
    keep_prob = tf.placeholder(tf.float32)
    h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

    # Second fully connected layer
    W_fc2 = weight_variable([1024, NUM_CLASSES])
    b_fc2 = bias_variable([NUM_CLASSES])

    # Output layer
    y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
    # print(y_conv.shape)
    # print(y_conv.get_shape)

    # Get the loss
    cross_entropy = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

    # Minimize the loss
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

    # Read all data from tfrecord file
    imageList, labelList = read_data()
    imageBatch, labelBatch = next_batch(imageList, labelList, BATCH_SIZE)

    correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

    with tf.Session() as sess:
        sess.run(tf.local_variables_initializer())
        sess.run(tf.global_variables_initializer())

        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(coord=coord)

        train_images, train_labels = sess.run([imageBatch, labelBatch])
        train_images = np.reshape(train_images, (-1, IMAGE_SIZE))
        train_labels = np.reshape(train_labels, (-1, NUM_CLASSES))

        sess.run(train_step, feed_dict = {x: train_images, y_: train_labels, keep_prob: 1.0})

        coord.request_stop()
        coord.join(threads)
    sess.close()

if __name__ == '__main__':
  parser = argparse.ArgumentParser()
  FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

When I run the program, I'm getting

InvalidArgumentError (see above for traceback): logits and labels must be same size: logits_size=[50,5] labels_size=[10,5]
     [[Node: SoftmaxCrossEntropyWithLogits = SoftmaxCrossEntropyWithLogits[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](Reshape_2, Reshape_3)]]

I've done several hours of search on this problem, but could not see why the logits are not matching label size. If I change batchsize to 10, the labels size will become [2,5] as if it's always being divided by 5. Can someone help me out here?

Answer 1

Most likely your labels are single integer values rather than one-hot vectors, so your labelBatch is a vector of size [50] containing single numbers like "1" or "4". Now, when you reshape them using train_labels = np.reshape(train_labels, (-1, NUM_CLASSES)) you're changing the shape to [10, 5].

The tf.nn.softmax_cross_entropy_with_logits function expects labels to be "one-hot" encodings of the labels (this means that a label of 3 translates into a vector of size 5 with a 1 in position 3 and zeros elsewhere). You can achieve this using the tf.nn.one_hot function, but an easier way to do it is instead to use the tf.nn.sparse_softmax_cross_entropy_with_logits function which is designed to work with these single-valued labels. To achieve this, you'll need to change these line:

y_ = tf.placeholder(tf.float32, [None]) # Desired output

cross_entropy = tf.reduce_mean( tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

And get rid of the train_labels = np.reshape(train_labels, (-1, NUM_CLASSES)) line.

(By the way, you don't actually need to use placeholders when reading data in this way - you can just directly use the output tensors.)