def evaluate_model(training_file, test_file, epochs, augment, batch_size, model_params):
print("Learning rate: {}, dropout rate: {}, {} residual blocks, {} filters, bn: {}, fc: {}, augmenting {}, batch_size: {}".format(model_params['learning_rate'], model_params['dropout_rate'], model_params['residual_blocks'], model_params['filters'], model_params['batch_norm'], model_params['fc_layers'], augment, batch_size))
# Create a deep neural network regression classifier.
# Build custom classifier
classifier = deep_model.estimator(model_params)
results = {}
for epoch in range(epochs):
# Train our model, use the previously function my_input_fn
# Input to training is a file with training example
classifier.train(
input_fn=lambda: deep_model.my_input_fn(training_file, True, 1, augment, batch_size))
if epoch % 10 == 0 or epoch == (epochs - 1):
print("Epoch: {}".format(epoch))
# Evaluate our model
# Return value will contain evaluation_metrics such as: loss & average_loss
# Evaluate on training set
evaluate_result = classifier.evaluate(
input_fn=lambda: deep_model.my_input_fn(training_file, False, 4, False, batch_size), name='train')
print("Evaluation results")
for key in evaluate_result:
print(" {}, was: {}".format(key, evaluate_result[key]))
results['training_loss'] = evaluate_result['loss']
results['training_accuracy'] = evaluate_result['accuracy']
# Evaluate on test set
evaluate_result = classifier.evaluate(
input_fn=lambda: deep_model.my_input_fn(test_file, False, 4, False, batch_size), name='test')
print("Evaluation results")
for key in evaluate_result:
print(" {}, was: {}".format(key, evaluate_result[key]))
results['test_loss'] = evaluate_result['loss']
results['test_accuracy'] = evaluate_result['accuracy']
return results
default=64,
help="Minibatch size")
parser.add_argument('--params',
'-p',
default='params.json',
help="Parameters file")
parser.add_argument('--gamma',
'-g',
default=0.9,
type=float,
help="Gamma, discount factor for future rewards")
args = parser.parse_args()
# Load hyperparameters from file
with open(args.params, 'r') as f:
params = json.load(f)
# Load estimator
estimator = deep_model.estimator(params)
start = datetime.datetime.now()
replay_memory = training_data.training_data()
replay_memory.import_csv(args.input)
#replay_memory.augment()
train(estimator, replay_memory, args.gamma, args.iterations,
args.minibatch_size)
end = datetime.datetime.now()
taken = end - start
print("took {}".format(taken))