tensorlm
Generate Shakespeare poems with 4 lines of code.
Installation
tensorlm is written in / for Python 3.4+ and TensorFlow 1.1+
pip3 install tensorlm
Basic Usage
Use the CharLM or WordLM class:
import tensorflow as tf
from tensorlm import CharLM
with tf.Session() as session:
# Create a new model. You can also use WordLM
model = CharLM(session, "datasets/sherlock/tinytrain.txt", max_vocab_size=96,
neurons_per_layer=100, num_layers=3, num_timesteps=15)
# Train it
model.train(session, max_epochs=10, max_steps=500)
# Let it generate a text
generated = model.sample(session, "The ", num_steps=100)
print("The " + generated)
This should output something like:
The ee e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e e
Command Line Usage
Train: python3 -m tensorlm.cli --train=True --level=char --train_text_path=datasets/sherlock/tinytrain.txt --max_vocab_size=96 --neurons_per_layer=100 --num_layers=2 --batch_size=10 --num_timesteps=15 --save_dir=out/model --max_epochs=300 --save_interval_hours=0.5
Sample: python3 -m tensorlm.cli --sample=True --level=char --neurons_per_layer=400 --num_layers=3 --num_timesteps=160 --save_dir=out/model
Evaluate: python3 -m tensorlm.cli --evaluate=True --level=char --evaluate_text_path=datasets/sherlock/tinyvalid.txt --neurons_per_layer=400 --num_layers=3 --batch_size=10 --num_timesteps=160 --save_dir=out/model
See python3 -m tensorlm.cli --help for all options.
Advanced Usage
Custom Input Data
The inputs and targets don't have to be text. GeneratingLSTM only expects token ids, so you can use any data type for the sequences, as long as you can encode the data to integer ids.
# We use integer ids from 0 to 19, so the vocab size is 20. The range of ids must always start
# at zero.
batch_inputs = np.array([[1, 2, 3, 4], [15, 16, 17, 18]]) # 2 batches, 4 time steps each
batch_targets = np.array([[2, 3, 4, 5], [16, 17, 18, 19]])
# Create the model in a TensorFlow graph
model = GeneratingLSTM(vocab_size=20, neurons_per_layer=10, num_layers=2, max_batch_size=2)
# Initialize all defined TF Variables
session.run(tf.global_variables_initializer())
for _ in range(5000):
model.train_step(session, batch_inputs, batch_targets)
sampled = model.sample_ids(session, [15], num_steps=3)
print("Sampled: " + str(sampled))
This should output something like:
Sampled: [16, 18, 19]
Custom Training, Dropout etc.
Use the GeneratingLSTM class directly. This class is agnostic to the dataset type. It expects integer ids and returns integer ids.
import tensorflow as tf
from tensorlm import Vocabulary, Dataset, GeneratingLSTM
BATCH_SIZE = 20
NUM_TIMESTEPS = 15
with tf.Session() as session:
# Generate a token -> id vocabulary based on the text
vocab = Vocabulary.create_from_text("datasets/sherlock/tinytrain.txt", max_vocab_size=96,
level="char")
# Obtain input and target batches from the text file
dataset = Dataset("datasets/sherlock/tinytrain.txt", vocab, BATCH_SIZE, NUM_TIMESTEPS)
# Create the model in a TensorFlow graph
model = GeneratingLSTM(vocab_size=vocab.get_size(), neurons_per_layer=100, num_layers=2,
max_batch_size=BATCH_SIZE, output_keep_prob=0.5)
# Initialize all defined TF Variables
session.run(tf.global_variables_initializer())
# Do the training
epoch = 1
step = 1
for epoch in range(20):
for inputs, targets in dataset:
loss = model.train_step(session, inputs, targets)
if step % 100 == 0:
# Evaluate from time to time
dev_dataset = Dataset("datasets/sherlock/tinyvalid.txt", vocab,
batch_size=BATCH_SIZE, num_timesteps=NUM_TIMESTEPS)
dev_loss = model.evaluate(session, dev_dataset)
print("Epoch: %d, Step: %d, Train Loss: %f, Dev Loss: %f" % (
epoch, step, loss, dev_loss))
# Sample from the model from time to time
print("Sampled: \"The " + model.sample_text(session, vocab, "The ") + "\"")
step += 1
This should output something like:
Epoch: 3, Step: 100, Train Loss: 3.824941, Dev Loss: 3.778008
Sampled: "The "
Epoch: 7, Step: 200, Train Loss: 2.832825, Dev Loss: 2.896187
Sampled: "The "
Epoch: 11, Step: 300, Train Loss: 2.778579, Dev Loss: 2.830176
Sampled: "The eee "
Epoch: 15, Step: 400, Train Loss: 2.655153, Dev Loss: 2.684828
Sampled: "The ee e e e e e e e e e e e e e e e e e e e e e e e e e e e "
Epoch: 19, Step: 500, Train Loss: 2.444502, Dev Loss: 2.479753
Sampled: "The an an an on on on on on on on on on on on on on on on on on on on on on o"
1 Dec 15, 2021
1 Nov 23, 2021
3 Nov 11, 2022
1 Jan 25, 2022
27 Dec 20, 2022
9.6k Jan 2, 2023
23 Oct 17, 2022
272 Dec 23, 2022
40 Dec 10, 2022
10 Dec 10, 2022
178 Jan 06, 2023
4 Mar 12, 2022
80 Oct 31, 2022
8 Oct 03, 2022
54 Jun 22, 2022
39 Nov 21, 2022
11 Sep 10, 2022
517 Dec 28, 2022
15 Feb 10, 2022
78 Dec 18, 2022
50 Jan 05, 2023
610 Jan 05, 2023
320 Dec 12, 2022
54 Jan 03, 2023
16 Dec 27, 2022
87 Jan 01, 2023
183 Jan 05, 2023
97 Nov 29, 2022