Keras: Getting different accuracy using model.train_on_batch() and model.fit(). What could be the reason and how to fix that?

Question

I have two identical models with same parameters. Both of these are trained on MNIST dataset. First one is trained using model.fit() and the second one is trained using model.train_on_batch(). The second model is giving less accuracy. I want to know what could be the reason for that and how to fix it?

Data preperation:

batch_size = 150
num_classes = 10
epochs = 12

# input image dimensions
img_rows, img_cols = 28, 28

# the data, split between train and test sets
(x_train, y_train), (x_test, y_test) = mnist.load_data()

if K.image_data_format() == 'channels_first':
    x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
    x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
    input_shape = (1, img_rows, img_cols)
else:
    x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
    x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
    input_shape = (img_rows, img_cols, 1)

x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# convert class vectors to binary class matrices
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

MODEL 1:

model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Conv2D(128, (3, 3), activation='relu'))
model.add(Conv2D(256, (3, 3), activation='relu'))
model.add(Conv2D(128, (3, 3), activation='relu'))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Conv2D(32, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))

model.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])

model.fit(x_train, y_train,
          batch_size=batch_size,
          epochs=epochs,
          verbose=1,
          validation_data=(x_test, y_test))
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

MODEL 1 ACCURACY:

Test loss: 0.023489486496470636 Test accuracy: 0.9924

MODEL 2:

model2 = Sequential()
model2.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model2.add(Conv2D(64, (3, 3), activation='relu'))
model2.add(Conv2D(128, (3, 3), activation='relu'))
model2.add(Conv2D(256, (3, 3), activation='relu'))
model2.add(Conv2D(128, (3, 3), activation='relu'))
model2.add(Conv2D(64, (3, 3), activation='relu'))
model2.add(Conv2D(64, (3, 3), activation='relu'))
model2.add(Conv2D(32, (3, 3), activation='relu'))
model2.add(MaxPooling2D(pool_size=(2, 2)))
model2.add(Dropout(0.25))
model2.add(Flatten())
model2.add(Dense(128, activation='relu'))
model2.add(Dropout(0.5))
model2.add(Dense(num_classes, activation='softmax'))

model2.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])

batch_size2 = 150
epochs2 = 12
step_epoch = x_train.shape[0] // batch_size2

def next_batch_train(i):
  return x_train[i:i+batch_size2,:,:,:], y_train[i:i+batch_size2,:]

iter_num = 0
epoch_num = 0
model_outputs = []
loss_history  = []

while epoch_num < epochs2:
  while iter_num < step_epoch:
    x,y = next_batch_train(iter_num)
    loss_history += model2.train_on_batch(x,y)

    iter_num += 1

  print("EPOCH {} FINISHED".format(epoch_num + 1))
  epoch_num += 1
  iter_num = 0 # reset counter


score = model2.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

MODEL 2 ACCURACY:

Test loss: 0.5577236003954947 Test accuracy: 0.9387

Answer 1

Four sources of difference:

1. fit() uses shuffle=True by default, this includes the very first epoch (and subsequent ones)
2. You don't use a random seed; see my answer here
3. You have step_epoch number of batches, but iterate over step_epoch - 1 ; change < to <=
4. Your next_batch_train slicing is way off; here's what it's doing vs what it needs to be doing:
   • x_train[0:128] --> x_train[1:129] --> x_train[2:130] -->...
   • x_train[0:128] --> x_train[128:256] --> x_train[256:384] -->...

To remedy, you should include a shuffling step in your model2 's train loop - or use fit with shuffle=False (not recommended). Also, a tip: 64, 128, 256, 128, 64 Conv2D filters is a pretty bad arrangement; what you're doing is upsampling greatly, in a sense "fabricating data" - if you're going to use more filters, also increase their strides proportionally so that the total tensor size between the layers remains ~same (or less).

All mentioned fixes + updated seed function below; run it for 1 epoch, 12 takes too long - if 1 works so will 12. Can keep your original model if you'd like, but I recommend testing with one below, as it's significantly faster.

---

import tensorflow as tf
import numpy as np
import random

def reset_seeds():
    np.random.seed(1)
    random.seed(2)
    if tf.__version__[0] == '2':
        tf.random.set_seed(3)
    else:
        tf.set_random_seed(3)
    print("RANDOM SEEDS RESET")

reset_seeds()
model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Conv2D(32, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))
model.add(Flatten())
model.add(Dense(128, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(num_classes, activation='softmax'))

model.compile(loss=keras.losses.categorical_crossentropy,
              optimizer=keras.optimizers.Adadelta(),
              metrics=['accuracy'])

def next_batch_train(i):
  return (x_train[i*batch_size2:(i+1)*batch_size2,:,:,:], 
          y_train[i*batch_size2:(i+1)*batch_size2,:])

iter_num = 0
epoch_num = 0
model_outputs = []
loss_history  = []

while epoch_num < epochs2:
  while iter_num < step_epoch:
    x,y = next_batch_train(iter_num)
    loss_history += model2.train_on_batch(x,y)

    iter_num += 1

  print("EPOCH {} FINISHED".format(epoch_num + 1))
  epoch_num += 1
  iter_num = 0 # reset counter

score = model2.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])

---

Better alternative: use shuffling

class TrainBatches():
    def __init__(self, x_train, y_train, batch_size):
        self.x_train=x_train
        self.y_train=y_train
        self.batch_size=batch_size

        self.indices = [i for i in range(len(x_train))]
        self.counter = 0

    def get_next(self):
        start = self.indices[self.counter] * self.batch_size
        end   = start + self.batch_size
        self.counter += 1
        return self.x_train[start:end], self.y_train[start:end]

    def shuffle(self):
        np.random.shuffle(self.indices)
        print("BATCHES SHUFFLED")

train_batches = TrainBatches(x_train, y_train, batch_size)

while epoch_num < epochs2:
  while iter_num <= step_epoch:
    x, y = train_batches.get_next()
    loss_history += model2.train_on_batch(x,y)

    iter_num += 1

  train_batches.shuffle()
  train_batches.counter = 0
  print("EPOCH {} FINISHED".format(epoch_num + 1))
  epoch_num += 1
  iter_num = 0 # reset counter

Note that this won't guarantee your results will agree with fit() , as fit() may shuffle differently (even with a random seed) - but the implementation is in fact correct. Above also doesn't shuffle at the first epoch (easy to change).

Answer 2

One difference between those two models that I noticed is that in your second model, you didn't shuffle your training data after each epoch. .fit() will shuffle your training data in default.