Training Accuracy not increasing - CNN with Tensorflow

Question

I've recently started working with machine learning using Tensorflow in a Google Colab notebook, working on a network to classify images of food.

My dataset is comprised of exactly 101,000 images and 101 classes - 1000 images per class. The network I developed following this Tensorflow Blog

The code I have developed is as follows:

#image dimensions
batch_size = 32
img_height = 50
img_width = 50

#80% for training, 20% for validating
train_ds = image_dataset_from_directory(data_dir,
                                             shuffle=True,
                                             validation_split=0.2,
                                             subset="training",
                                             seed=123,
                                             batch_size=batch_size,
                                             image_size=(img_height, img_width)
                                             )
val_ds = image_dataset_from_directory(data_dir,
                                             shuffle=True,
                                             validation_split=0.2,
                                             subset="validation",
                                             seed=123,
                                             batch_size=batch_size,
                                             image_size=(img_height, img_width)
                                             )
#autotuning, configuring for performance
AUTOTUNE = tf.data.AUTOTUNE

train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)

#data augmentation layer
data_augmentation = keras.Sequential(
  [
    layers.experimental.preprocessing.RandomFlip("horizontal", 
                                                 input_shape=(img_height, 
                                                              img_width,
                                                              3)),
    layers.experimental.preprocessing.RandomRotation(0.1),
    layers.experimental.preprocessing.RandomZoom(0.1),
  ]
)

#network definition
num_classes = 101

model = Sequential([
  data_augmentation,
  layers.experimental.preprocessing.Rescaling(1./255),
  layers.Conv2D(16, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(32, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(64, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(128, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Conv2D(256, 3, padding='same', activation='relu'),
  layers.MaxPooling2D(),
  layers.Dropout(0.2),
  layers.Flatten(),
  layers.Dense(128, activation='relu'),
  layers.Dense(num_classes, activation='softmax')
])

model.compile(optimizer='adam',
              loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
              metrics=['accuracy'])

After training for 500 epochs, the accuracy seems to be increasing incredibly slow:

epoch 100: 2525/2525 - 19s 8ms/step - loss: 2.8151 - accuracy: 0.3144 - val_loss: 3.1659 - val_accuracy: 0.2549

epoch 500: 2525/2525 - 21s 8ms/step - loss: 2.7349 - accuracy: 0.0333 - val_loss: 3.1260 - val_accuracy: 0.2712

I have tried:

1. Increasing the number of epochs (previously 100)
2. Adding more convolutional layers
3. Changing optimizers
4. Changing my image size (down from 180, previously 244)

So far, the code above offers the best results, but I still wonder,

Is this behaviour expected? Is it a result of having such a big dataset? Or is there any flaw in my code that's possibly hindering the learning process?

Answer 1

in your loss function remove from_logits=True