jamescalam

import numpy as np
import matplotlib.pyplot as plt
from matplotlib import animation, cm
from mpl_toolkits.mplot3d import Axes3D


# create a figure
fig = plt.figure()
# initialise 3D Axes
ax = Axes3D(fig)

jamescalam

import requests
import re

# import Meditations
response = requests.get('http://classics.mit.edu/Antoninus/meditations.mb.txt')
data = response.text

# clean the text
data = data.split("Translated by George Long")[1].replace("-", "").split("THE END")[0]
data = re.sub("BOOK [A-Z]+\n", "", data)

jamescalam

char2idx = {c:i for i, c in enumerate(vocab)}
idx2char = np.array(vocab)

jamescalam

model = tf.keras.Sequential([
  tf.keras.layers.Embedding(len(vocab), EMBED_DIM,
                            batch_input_shape=[BATCH_SIZE, None]),
  tf.keras.layers.GRU(UNITS, return_sequences=True,
                      stateful=True,
                      dropout=0.1),
  tf.keras.layers.Dense(len(vocab))

jamescalam

checkpoint = tf.keras.callbacks.ModelCheckpoint(
    filepath='./training_checkpoints/ckpt_{epoch}',
    save_weights_only=True
)

history = model.fit(dataset, epochs=EPOCHS, callbacks=[checkpoint_callback])

jamescalam

model = build_model(len(vocab), EMBED_DIM, UNITS, 1)
model.load_weights(tf.train.latest_checkpoint('./training_checkpoints')
model.build(tf.TensorShape([1, None]))

jamescalam

meditations = "From "  # initialize our meditations text
# convert to indices
input_eval = tf.expand_dims([char2idx[c] for x in meditations], 0)
# initialize states
model.reset_states()
# loop through, generating 100K characters
for i in range(100000):
    y_hat = model(input_eval)  # make a prediction
    y_hat = tf.squeeze(y_hat, 0)  # remove batch dimension
    predicted_idx = tf.random.categorical(y_hat, num_samples=1)[-1,0].numpy()

jamescalam

char_dataset = tf.data.Dataset.from_tensor_slices(data_idx)

jamescalam

sequences = char_dataset.batch(SEQ_LEN+1, drop_remainder=True)

jamescalam

def split_input_output(chunk):
    return chunk[:-1], chunk[1:]

dataset = sequences.map(split_input_output)