sequence_size = 100
doc_scanner = DocumentScanner(filename, sequence_size)
d, d_o = doc_scanner.build_character_mappings()
vocab_size = len(d)
# Create the LSTM
lstm = LSTM(vocab_size, hidden_size)
# Build zero valued matrices of the output and state
h = np.zeros([1, hidden_size])
state = np.zeros([1, hidden_size])
# Build initial Adam parameters that will be used and update during training
adam_params = lstm.build_adam_params()
# Train the LSTM and sample every 100 epochs
for i in range(20000):
X, labels = lstm.vectorize_sequence(doc_scanner.next_sequence(), d)
loss, state, h, adam_params = lstm.bptt(X, labels, state, h, adam_params)
if i % 100 == 0:
print "Current loss at epoch {}: {}".format(i, loss)
print "Sample text:"
x = X[0]
sample_text = lstm.sample(1000, x, state, h)
print "".join(map(lambda r: d_o[r], sample_text))
print "\n"
# Close the document being trained on
doc_scanner.close_file()
