def train(data, model, num_frontends, resume_epoch, model_dir):
cb = SaveModelsAndTerminateEarly()
cb.set_params(model_dir, resume_epoch)
X_train, y_train = data[0]
X_val, y_val = data[1]
if 1 < num_frontends:
X_train = [X_train] * num_frontends
X_val = [X_val] * num_frontends
try:
model.fit(
X_train, y_train, validation_data=(X_val, y_val),
nb_epoch=10000, batch_size=128, callbacks=[cb],
show_accuracy=True)
except EarlyTermination:
pass
def train(data, model, num_frontends, resume_epoch, model_dir):
cb = SaveModelsAndTerminateEarly()
cb.set_params(model_dir, resume_epoch)
X_train, y_train = data[0]
X_val, y_val = data[1]
if 1 < num_frontends:
X_train = [X_train] * num_frontends
X_val = [X_val] * num_frontends
try:
model.fit(X_train,
y_train,
validation_data=(X_val, y_val),
nb_epoch=10000,
batch_size=128,
callbacks=[cb],
show_accuracy=True)
except EarlyTermination:
pass
def train(pipeline_func_name, network_func_name, train_set, val_set,
save_dir):
# Init the data transformers.
f = getattr(pipelines, pipeline_func_name)
data_pipe, class_pipe, need_to_embed = f()
data_pipe = TransformerPipeline(data_pipe)
class_pipe = TransformerPipeline(class_pipe)
# Transform the data.
(X_train, y_train), (X_val, y_val) = \
fit_transform(train_set, val_set, data_pipe, class_pipe)
# Build the model, now that we know its input/output dimensions.
print 'X_train.shape:', X_train.shape
print 'y_train.shape:', y_train.shape
if need_to_embed:
input_dim = max(map(max, X_train)) + 1
else:
input_dim = X_train.shape[-1]
output_dim = y_train.shape[-1]
print 'Input dim:', input_dim
print 'Output dim:', output_dim
kmodel, nb_frontends = \
construct(network_func_name, input_dim, output_dim)
# Save the model configuration.
if os.path.exists(save_dir):
shutil.rmtree(save_dir)
os.makedirs(save_dir)
d = {
'pipeline_func_name': pipeline_func_name,
'data_pipe': data_pipe,
'class_pipe': class_pipe,
'network_func_name': network_func_name,
'input_dim': input_dim,
'output_dim': output_dim,
}
fn = os.path.join(save_dir, 'model.pkl')
cPickle.dump(d, open(fn, 'wb'))
# Train the model, saving checkpoints.
print 'Training...'
if nb_frontends != 1:
X_train = [X_train] * nb_frontends
X_val = [X_val] * nb_frontends
cb = SaveModelsAndTerminateEarly()
cb.set_params(save_dir)
try:
kmodel.fit(
X_train, y_train, validation_data=(X_val, y_val),
nb_epoch=10000, batch_size=128, callbacks=[cb],
show_accuracy=True, verbose=True)
except EarlyTermination:
logging.info('Terminated training early.')
pass
print 'Done training.'
return Model(data_pipe, kmodel, class_pipe, nb_frontends)
