def load_data():
from rnn import load_data
dev_data, _ = load_data(config,
0,
'dev',
chunking=config.value("chunking", ""),
seq_ordering="sorted",
shuffle_frames_of_nseqs=0)
eval_data, _ = load_data(config,
0,
'eval',
chunking=config.value("chunking", ""),
seq_ordering="sorted",
shuffle_frames_of_nseqs=0)
train_data, _ = load_data(config, 0, 'train')
return dev_data, eval_data, train_data
def create_engine():
dev_data, eval_data, train_data = load_data()
engine = Engine()
engine.init_train_from_config(config, train_data, dev_data, eval_data)
engine.init_train_epoch()
engine.train_batches = engine.train_data.generate_batches(
recurrent_net=engine.network.recurrent,
batch_size=engine.batch_size,
max_seqs=engine.max_seqs,
max_seq_length=engine.max_seq_length,
seq_drop=engine.seq_drop,
shuffle_batches=engine.shuffle_batches,
used_data_keys=engine.network.used_data_keys)
engine.updater.set_learning_rate(engine.learning_rate,
session=engine.tf_session)
engine.updater.init_optimizer_vars(session=engine.tf_session)
return engine
def hello_world():
"""Main index page of the site
:return html_template taking in the generated text, if any"""
text = ""
if request.method == "POST":
# Retrieve form data
model = request.form.get("model_type")
history = False
if request.form.get("history"):
history = True
seed_text = request.form.get("seed-text")
if seed_text == "":
seed_text = "<"
# Load the data
datapath = WINE_DATA if model == "wine" else JOKE_DATA
seq, itoc, ctoi = rnn.load_data(datapath)
if history:
# Generate text from the same model at different points in training
filepaths = WINES if model == "wine" else JOKES
models = []
for filepath in filepaths:
models.append(torch.load(filepath))
for model in models:
text += model.generate_text(ctoi,
itoc,
seq.shape[1],
start_phrase=seed_text)
text += "\n\n"
else:
# Generate text from the best model of the given type
path = WINE if model == "wine" else JOKE
model = torch.load(path)
text += model.generate_text(ctoi,
itoc,
seq.shape[1],
start_phrase=seed_text)
return render_template("main.html", text=text)
NUM_LAYERS = 1
EMB_SIZE = 128
HID_SIZE = 256
PAD_CHAR = '#'
JOKES = "jokes/jokes.txt"
WINE = "wine_data/wine_reviews_short.txt"
dev = 'cpu'
if torch.cuda.is_available():
dev = 'cuda'
# Change this to JOKES to train on and generate jokes
filename = JOKES
seq, itoc, ctoi = rnn.load_data(filename)
if filename == JOKES:
X_train, X_test = train_test_split(seq, test_size=0.25, shuffle=False)
else:
# Decrease the text set even more to decrease training time
X_train, X_test = train_test_split(seq, test_size=0.8, shuffle=False)
X_test, _ = train_test_split(X_test, test_size=0.98, shuffle=False)
charnn = rnn.CharNN(len(itoc),
num_layers=NUM_LAYERS,
emb_size=EMB_SIZE,
hid_size=HID_SIZE,
pad_idx=ctoi[PAD_CHAR],
max_len=seq.shape[1],
device=dev)