What does it mean when the loss starts going up again? - tensorflow

Question

I am running the code from https://www.tensorflow.org/tutorials/text/text_generation. I will copy it at the bottom of the question. If I change the EPOCHS line to
EPOCHS = 100
something odd happens to the loss. It starts by going down, as in:
Epoch 1/100
172/172 [==============================] - 301s 2s/step - loss: 2.7219
Epoch 2/100
172/172 [==============================] - 328s 2s/step - loss: 1.9963
Epoch 3/100
172/172 [==============================] - 344s 2s/step - loss: 1.7313
Epoch 4/100
172/172 [==============================] - 321s 2s/step - loss: 1.5778
Epoch 5/100
172/172 [==============================] - 325s 2s/step - loss: 1.4840
reaching it's lowest level at Epoch 46/100 when the loss is 0.6233. It then goes back up again finishing with:
Epoch 96/100
172/172 [==============================] - 292s 2s/step - loss: 0.8749
Epoch 97/100
172/172 [==============================] - 292s 2s/step - loss: 0.8933
Epoch 98/100
172/172 [==============================] - 292s 2s/step - loss: 0.9073
Epoch 99/100
172/172 [==============================] - 292s 2s/step - loss: 0.9181
Epoch 100/100
172/172 [==============================] - 292s 2s/step - loss: 0.9298
Why is it doing this and what does it mean?
import tensorflow as tf
import numpy as np
import os
import time
path_to_file = tf.keras.utils.get_file('shakespeare.txt', 'https://storage.googleapis.com/download.tensorflow.org/data/shakespeare.txt')
# Read, then decode for py2 compat.
text = open(path_to_file, 'rb').read().decode(encoding='utf-8')
# length of text is the number of characters in it
print('Length of text: {} characters'.format(len(text)))
# Take a look at the first 250 characters in text
print(text[:250])
# The unique characters in the file
vocab = sorted(set(text))
print('{} unique characters'.format(len(vocab)))
# Creating a mapping from unique characters to indices
char2idx = {u:i for i, u in enumerate(vocab)}
idx2char = np.array(vocab)
text_as_int = np.array([char2idx[c] for c in text])
print('{')
for char,_ in zip(char2idx, range(20)):
print(' {:4s}: {:3d},'.format(repr(char), char2idx[char]))
print(' ...\n}')
# Show how the first 13 characters from the text are mapped to integers
print('{} ---- characters mapped to int ---- > {}'.format(repr(text[:13]), text_as_int[:13]))
# The maximum length sentence you want for a single input in characters
seq_length = 100
examples_per_epoch = len(text)//(seq_length+1)
# Create training examples / targets
char_dataset = tf.data.Dataset.from_tensor_slices(text_as_int)
for i in char_dataset.take(5):
print(idx2char[i.numpy()])
sequences = char_dataset.batch(seq_length+1, drop_remainder=True)
for item in sequences.take(5):
print(repr(''.join(idx2char[item.numpy()])))
def split_input_target(chunk):
input_text = chunk[:-1]
target_text = chunk[1:]
return input_text, target_text
dataset = sequences.map(split_input_target)
for input_example, target_example in dataset.take(1):
print('Input data: ', repr(''.join(idx2char[input_example.numpy()])))
print('Target data:', repr(''.join(idx2char[target_example.numpy()])))
for i, (input_idx, target_idx) in enumerate(zip(input_example[:5], target_example[:5])):
print("Step {:4d}".format(i))
print(" input: {} ({:s})".format(input_idx, repr(idx2char[input_idx])))
print(" expected output: {} ({:s})".format(target_idx, repr(idx2char[target_idx])))
# Batch size
BATCH_SIZE = 64
# Buffer size to shuffle the dataset
# (TF data is designed to work with possibly infinite sequences,
# so it doesn't attempt to shuffle the entire sequence in memory. Instead,
# it maintains a buffer in which it shuffles elements).
BUFFER_SIZE = 10000
dataset = dataset.shuffle(BUFFER_SIZE).batch(BATCH_SIZE, drop_remainder=True)
dataset
# Length of the vocabulary in chars
vocab_size = len(vocab)
# The embedding dimension
embedding_dim = 256
# Number of RNN units
rnn_units = 1024
def build_model(vocab_size, embedding_dim, rnn_units, batch_size):
model = tf.keras.Sequential([
tf.keras.layers.Embedding(vocab_size, embedding_dim,
batch_input_shape=[batch_size, None]),
tf.keras.layers.GRU(rnn_units,
return_sequences=True,
stateful=True,
recurrent_initializer='glorot_uniform'),
tf.keras.layers.Dense(vocab_size)
])
return model
model = build_model(
vocab_size=len(vocab),
embedding_dim=embedding_dim,
rnn_units=rnn_units,
batch_size=BATCH_SIZE)
for input_example_batch, target_example_batch in dataset.take(1):
example_batch_predictions = model(input_example_batch)
print(example_batch_predictions.shape, "# (batch_size, sequence_length, vocab_size)")
model.summary()
sampled_indices = tf.random.categorical(example_batch_predictions[0], num_samples=1)
sampled_indices = tf.squeeze(sampled_indices,axis=-1).numpy()
sampled_indices
print("Input: \n", repr("".join(idx2char[input_example_batch[0]])))
print()
print("Next Char Predictions: \n", repr("".join(idx2char[sampled_indices ])))
def loss(labels, logits):
return tf.keras.losses.sparse_categorical_crossentropy(labels, logits, from_logits=True)
example_batch_loss = loss(target_example_batch, example_batch_predictions)
print("Prediction shape: ", example_batch_predictions.shape, " # (batch_size, sequence_length, vocab_size)")
print("scalar_loss: ", example_batch_loss.numpy().mean())
model.compile(optimizer='adam', loss=loss)
# Directory where the checkpoints will be saved
checkpoint_dir = './training_checkpoints'
# Name of the checkpoint files
checkpoint_prefix = os.path.join(checkpoint_dir, "ckpt_{epoch}")
checkpoint_callback = tf.keras.callbacks.ModelCheckpoint(
filepath=checkpoint_prefix,
save_weights_only=True)
EPOCHS = 100
history = model.fit(dataset, epochs=EPOCHS, callbacks=[checkpoint_callback])
tf.train.latest_checkpoint(checkpoint_dir)
model = build_model(vocab_size, embedding_dim, rnn_units, batch_size=1)
model.load_weights(tf.train.latest_checkpoint(checkpoint_dir))
model.build(tf.TensorShape([1, None]))
model.summary()
def generate_text(model, start_string):
# Evaluation step (generating text using the learned model)
# Number of characters to generate
num_generate = 1000
# Converting our start string to numbers (vectorizing)
input_eval = [char2idx[s] for s in start_string]
input_eval = tf.expand_dims(input_eval, 0)
# Empty string to store our results
text_generated = []
# Low temperature results in more predictable text.
# Higher temperature results in more surprising text.
# Experiment to find the best setting.
temperature = 1.0
# Here batch size == 1
model.reset_states()
for i in range(num_generate):
predictions = model(input_eval)
# remove the batch dimension
predictions = tf.squeeze(predictions, 0)
# using a categorical distribution to predict the character returned by the model
predictions = predictions / temperature
predicted_id = tf.random.categorical(predictions, num_samples=1)[-1,0].numpy()
# Pass the predicted character as the next input to the model
# along with the previous hidden state
input_eval = tf.expand_dims([predicted_id], 0)
text_generated.append(idx2char[predicted_id])
return (start_string + ''.join(text_generated))
print(generate_text(model, start_string=u"ROMEO: "))

Answer 1

This particular model can't fit any better than this, since it is limited to its architecture and only one symbol generation per step.
A loss steadily going up after some epochs is a usual thing indicating your model overtrains, and there is no point in training any further.
You could tune hyperparameters to (possibly) make some minor improvements.
Edit:
To tune embedding dimensions, rnn units, and sequence length change those values:
seq_length = 100
embedding_dim = 256
rnn_units = 1024
To tune learning rate replace this lane:
model.compile(optimizer='adam', loss=loss)
with this one:
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.005), loss=loss)
Also, you can add arbitrary layers to build_model function.
Here is an example with an extra GRU layer:
def build_model(vocab_size, embedding_dim, rnn_units, batch_size):
model = tf.keras.Sequential([
tf.keras.layers.Embedding(vocab_size, embedding_dim,
batch_input_shape=[batch_size, None]),
tf.keras.layers.GRU(rnn_units,
return_sequences=True,
stateful=True,
recurrent_initializer='glorot_uniform'),
tf.keras.layers.GRU(rnn_units,
return_sequences=True,
stateful=True,
recurrent_initializer='glorot_uniform'),
tf.keras.layers.Dense(vocab_size)
])
return model