every time when I train the model I receive a different result - why? - tensorflow

Question

I don't understand what I do wrong - every time when I launch this code I receive a different result.
I figured out that the result will be different when I change the batch size, but the accuracy should not depend on batch size.
And charts look completely wrong then I expected
Could somebody point me on my mistake?
%reload_ext autoreload
%autoreload 2
%matplotlib inline
## Load dataset
import tensorflow as tf
tf.config.set_soft_device_placement(True)
tf.debugging.set_log_device_placement(True)
import tensorflow_datasets as tfds # must be 2.1
import matplotlib.pyplot as plt
builder = tfds.builder('beans')
info = builder.info
builder.download_and_prepare()
datasets = builder.as_dataset()
raw_train_dataset, raw_test_dataset = datasets['train'], datasets['test']
## Build a model
def get_model(image_width:int, image_height:int, num_classes:int):
model = tf.keras.Sequential()
#layer 1
model.add(tf.keras.layers.Conv2D(filters=96, kernel_size=(11,11), strides= 4, padding= 'valid',
activation=tf.keras.activations.relu, input_shape=(image_width, image_height ,3)))
model.add(tf.keras.layers.MaxPooling2D(pool_size=(3, 3), strides=(2,2),padding= 'valid'))
# layer 2
model.add(tf.keras.layers.Conv2D(filters=256, kernel_size=(5,5), strides= 1, padding='same',
activation=tf.keras.activations.relu))
model.add(tf.keras.layers.MaxPooling2D(pool_size=(3, 3), strides=(2,2), padding= 'valid'))
# layer 3
model.add(tf.keras.layers.Conv2D(filters=384, kernel_size=(3,3), strides= 1, padding= 'same', activation=tf.keras.activations.relu))
model.add(tf.keras.layers.Conv2D(filters=384, kernel_size=(3,3), strides= 1, padding= 'same', activation=tf.keras.activations.relu))
model.add(tf.keras.layers.Conv2D(filters=256, kernel_size=(3,3), strides= 1, padding= 'same', activation=tf.keras.activations.relu))
model.add(tf.keras.layers.MaxPooling2D(pool_size=(3, 3), strides=(2,2), padding= 'valid'))
# layer4 4
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(4096, activation=tf.keras.activations.relu))
model.add(tf.keras.layers.Dropout(0.5))
# layer4 4
model.add(tf.keras.layers.Dense(4096, activation=tf.keras.activations.relu))
model.add(tf.keras.layers.Dropout(0.5))
# layer4 5
model.add(tf.keras.layers.Dense(1000, activation = 'relu'))
model.add(tf.keras.layers.Dropout(0.5))
# layer4 6
model.add(tf.keras.layers.Dense(num_classes, activation = 'softmax'))
return model
IMAGE_width = 500
IMAGE_height = 500
NUM_CLASSES = info.features["label"].num_classes
CLASS_NAMES = info.features["label"].names
model = get_model(IMAGE_width, IMAGE_height, NUM_CLASSES)
result = model.compile( loss=tf.keras.losses.CategoricalCrossentropy(),
optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
metrics=['accuracy'])
## Train
def prepare_record(record):
image = record['image']
# image = tf.image.resize(image, (image_width,image_height))
# image = tf.cast(image, tf.int32)
label = record['label']
return image, label
train_dataset = raw_train_dataset.map(prepare_record, num_parallel_calls=tf.data.experimental.AUTOTUNE).shuffle(1034).batch(517).prefetch(tf.data.experimental.AUTOTUNE)
for train_image_batch, train_label_batch in train_dataset:
train_one_hot_y = tf.one_hot(train_label_batch, NUM_CLASSES )
history = model.fit(train_image_batch, train_one_hot_y, epochs=10, verbose=0,validation_split=0.2)
plt.plot(history.history['accuracy'])
plt.plot(history.history['val_accuracy'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'val'], loc='upper left')
plt.show()
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'val'], loc='upper left')
plt.show()

Answer 1

In the fit function there is a parameter called shuffle:
Boolean (whether to shuffle the training data before each epoch) or str (for 'batch'). 'batch' is a special option for dealing with the limitations of HDF5 data; it shuffles in batch-sized chunks. Has no effect when steps_per_epoch is not None.
If you set it to False, the results should be equal.
Another, probably preferable, way would be to use tf.random.set_seed(seed), so that the shuffling is always performed in the same way (see docs).