def run_experiment(experiment_generator, out_dir, test_data, plot_loss_batch=False):
np.random.seed(12345)
results = defaultdict(list)
for i, (model_dict, train_dict, exp_name, value) in enumerate(experiment_generator()):
model = MlpNet(**model_dict)
trainer = Trainer(model, **train_dict)
label = f'{exp_name}={value}'
print(f'{i}. {label}')
start_time = time()
trainer.train_loop()
time_period = time() - start_time
log_data = trainer.logger.logging_data
if plot_loss_batch:
# plot train loss per batch in first epoch
filename = exp_name + str(value) + '_loss_one_batch'
plot_val_loss_per_batch(log_data['loss_batch']['train'], filename, out_dir)
results['model_dict'].append(model_dict)
results['train_dict'].append(train_dict)
results['time'].append(time_period)
results['label'].append(label)
results['log_data'].append(log_data)
# calculate accuracy on test data
acc_metric = LabelAccuracy()
x_test, y_test = test_data
accuracy = acc_metric(model.predict_classes(x_test), y_test)
print('Accuracy on test data: {}'.format(accuracy))
return results
def get_model(model_name,
optimizer='SGDMomentum',
initializer='Xavier',
activation='sigmoid',
loss_fun='crossEntropy',
hidden_units=(512, )):
metrics = [LabelAccuracy()]
optimizer = get_optimizer(optimizer)
initializer = get_initializer(initializer, activation)
loss_fun = get_loss_function(loss_fun)
activation = get_activation(activation)
models = dict(MlpNet=MlpNet(optimizer, initializer, metrics, loss_fun,
activation, hidden_units),
ConvNet=ConvNet(optimizer, initializer, metrics, loss_fun))
m = models[model_name]
print("Creating model {} with parameters count: {}".format(
model_name, m.param_count()))
return m
# calculate accuracy on test data
acc_metric = LabelAccuracy()
x_test, y_test = test_data
accuracy = acc_metric(model.predict_classes(x_test), y_test)
print('Accuracy on test data: {}'.format(accuracy))
return results
if __name__ == "__main__":
train_data, val_data, test_data = load_data(DATA_PATH)
model_dict = {
'optimizer': SGD(),
'initializer': Xavier(),
'metrics': [LabelAccuracy()],
'loss_fun': categorical_cross_entropy,
'activation': Sigmoid,
'hidden_units': (500, )
}
train_dict = {
'train_data':
train_data,
'val_data':
val_data,
'epochs':
30,
'batch_size':
50,
'callbacks': [
def calc_test_accuracy(model, test_data, train_dict):
acc_metric = LabelAccuracy()
x_test, y_test = test_data
model.load_variables(train_dict['callbacks'][-1].save_path)
accuracy = acc_metric(model.predict_classes(x_test), y_test)
print('Accuracy on test data: {}'.format(accuracy))
from settings import DATA_PATH, PROJECT_PATH
from training import Trainer
from utils import load_data, ensure_dir_path_exists
if __name__ == "__main__":
np.random.seed(3)
train_data, val_data, test_data = load_data(DATA_PATH)
out_dir = 'my_nets/simple_net'
ensure_dir_path_exists(os.path.join(PROJECT_PATH, out_dir))
model_dict = {
'optimizer': SGD(),
'initializer': Xavier(),
'metrics': [LabelAccuracy()],
'loss_fun': categorical_cross_entropy,
'activation': Sigmoid,
'hidden_units': (100, )
}
train_dict = {
'train_data':
train_data,
'val_data':
val_data,
'epochs':
10,
'batch_size':
50,
'callbacks': [