def search_model(experiment_label, steps, batch_size=32):
""" This is where we put everythin together.
We get the dataset, build the Training and Experiment objects, and run the experiment.
The experiments logs are generated in ~/minos/experiment_label
We use the CpuEnvironment to have the experiment run on the cpu, with 2 parralel processes.
We could use GpuEnvironment to use GPUs, and specify which GPUs to use, and how many tasks
per GPU
"""
batch_iterator, test_batch_iterator, nb_classes = get_reuters_dataset(
batch_size, max_words)
layout = build_layout(max_words, nb_classes)
training = Training(Objective('categorical_crossentropy'),
Optimizer(optimizer='Adam'),
Metric('categorical_accuracy'),
epoch_stopping_condition(), batch_size)
parameters = custom_experiment_parameters()
experiment = Experiment(experiment_label,
layout,
training,
batch_iterator,
test_batch_iterator,
CpuEnvironment(n_jobs=2),
parameters=parameters)
run_ga_search_experiment(experiment,
population_size=100,
generations=steps,
resume=False,
log_level='DEBUG')
def train_multi_gpu(max_words=1000, batch_size=32):
batch_iterator, test_batch_iterator, nb_classes = get_reuters_dataset(
batch_size, max_words)
experiment = create_experiment(max_words, nb_classes, batch_size)
blueprint = create_random_blueprint(experiment)
devices = ['/gpu:0', '/gpu:1']
trainer = ModelTrainer(batch_iterator, test_batch_iterator)
with tempfile.TemporaryDirectory() as tmp_dir:
model, history, _duration = trainer.train(blueprint,
devices,
save_best_model=True,
model_filename=join(
tmp_dir, 'model'))
metric = get_associated_validation_metric(
blueprint.training.metric.metric)
epoch = numpy.argmax(history.history[metric])
score = history.history[metric][epoch]
print('Final training score %r after %d epoch' % (score, epoch))
test_size = 10
y_true = numpy.argmax(test_batch_iterator.y[0][:test_size], axis=1)
y_pred = numpy.argmax(model.predict(
test_batch_iterator.X[0][:test_size]),
axis=1)
evaluation = numpy.mean(y_true == y_pred)
print('Final evaluation score %f' % evaluation)
print('Predictions (true, pred) %r' %
list(zip(y_true.tolist(), y_pred.tolist())))