def test_mlp(sizes, learning_rate=0.15, n_epochs=100, batch_size=600, L1_reg=1.0e-4, L2_reg=0.0e-4):
n_ex, n_in, n_hidden, n_out = sizes
datasets = load_data(n_ex, n_in, n_out)
regmodel = RegressionModel(
"abc",
n_hidden=n_hidden,
learning_rate=learning_rate,
n_epochs=n_epochs,
batch_size=batch_size,
L1_reg=L1_reg,
L2_reg=L2_reg,
)
regmodel.setup_with_data(datasets)
regmodel.train()
print(map(lambda d: d.get_value(), regmodel.regression.params))
print(regmodel.y_pred(), datasets[2][1].get_value())
"""
def test_mlp(sizes,
learning_rate=0.15,
n_epochs=100,
batch_size=600,
L1_reg=1.0e-4,
L2_reg=0.0e-4):
n_ex, n_in, n_hidden, n_out = sizes
datasets = load_data(n_ex, n_in, n_out)
regmodel = RegressionModel('abc',
n_hidden=n_hidden,
learning_rate=learning_rate,
n_epochs=n_epochs,
batch_size=batch_size,
L1_reg=L1_reg,
L2_reg=L2_reg)
regmodel.setup_with_data(datasets)
regmodel.train()
print(map(lambda d: d.get_value(), regmodel.regression.params))
print(regmodel.y_pred(), datasets[2][1].get_value())
"""
def fit_mlp_models(n_hidden=25,
learning_rate=0.15,
n_epochs=100,
batch_size=600,
L1_reg=0.0e-9,
L2_reg=0.0e-9):
exp_type = 'natural'
movie_type = 'movie'
spatial_downsample_factor = 4
n_lag = 13
saved_models_dir = './temp/mlp-models-%d' % n_hidden
predicted_responses_dir = './temp/mlp-predicted-responses-%s' % n_hidden
if not os.path.isdir(saved_models_dir):
os.makedirs(saved_models_dir)
else:
shutil.rmtree(saved_models_dir)
os.makedirs(saved_models_dir)
if not os.path.isdir(predicted_responses_dir):
os.makedirs(predicted_responses_dir)
else:
shutil.rmtree(predicted_responses_dir)
os.makedirs(predicted_responses_dir)
print '%d hidden layer neurons, %d epochs to train for' % (n_hidden,
n_epochs)
responses = load_responses(exp_type)
movies = load_movies(exp_type,
movie_type,
downsample_factor=spatial_downsample_factor)
mlp_training_errors = []
for i, response in enumerate(responses):
name = response.name
print 'Mouse %s' % name
print 'Splitting out training and test data...'
tr_rsp, tst_rsp, tr_mov, tst_mov = train_test_split(response,
movies,
'even',
train_frac=0.8)
print 'Splitting out training and validation data...'
tr_rsp, val_rsp, tr_mov, val_mov = train_test_split(tr_rsp,
tr_mov,
'even',
train_frac=0.9)
tr_rsp = smooth_responses(tr_rsp)
val_rsp = smooth_responses(val_rsp)
tst_rsp = smooth_responses(tst_rsp)
train_set_x, train_set_y = window_matrices(tr_rsp, tr_mov, n_lag)
valid_set_x, valid_set_y = window_matrices(val_rsp, val_mov, n_lag)
test_set_x, test_set_y = window_matrices(tst_rsp, tst_mov, n_lag)
model = RegressionModel(model_name=name,
n_hidden=n_hidden,
learning_rate=learning_rate,
n_epochs=n_epochs,
batch_size=batch_size,
L1_reg=L1_reg,
L2_reg=L2_reg)
model.setup_with_data([(train_set_x, train_set_y),
(valid_set_x, valid_set_y),
(test_set_x, test_set_y)])
test_error = model.train()
mlp_training_errors.append(test_error)
predicted = model.y_pred()
np.save(os.path.join(predicted_responses_dir, 'pred_%s' % name),
predicted)
with open(os.path.join(saved_models_dir, 'mlp_%s' % name), 'wb') as f:
pickle.dump(model.regression.params, f)
with open(os.path.join(saved_models_dir, 'train_errors'), 'wb') as f:
pickle.dump(mlp_training_errors, f)
def fit_mlp_models(n_hidden=25, learning_rate=0.15, n_epochs=100,
batch_size=600, L1_reg=0.0e-9, L2_reg = 0.0e-9):
exp_type = 'natural'
movie_type = 'movie'
spatial_downsample_factor = 4
n_lag = 13
saved_models_dir = './temp/mlp-models-%d' % n_hidden
predicted_responses_dir = './temp/mlp-predicted-responses-%s' % n_hidden
if not os.path.isdir(saved_models_dir):
os.makedirs(saved_models_dir)
else:
shutil.rmtree(saved_models_dir)
os.makedirs(saved_models_dir)
if not os.path.isdir(predicted_responses_dir):
os.makedirs(predicted_responses_dir)
else:
shutil.rmtree(predicted_responses_dir)
os.makedirs(predicted_responses_dir)
print '%d hidden layer neurons, %d epochs to train for'%(n_hidden,n_epochs)
responses = load_responses(exp_type)
movies = load_movies(exp_type, movie_type,
downsample_factor=spatial_downsample_factor)
mlp_training_errors = []
for i, response in enumerate(responses):
name = response.name
print 'Mouse %s' % name
print 'Splitting out training and test data...'
tr_rsp, tst_rsp, tr_mov, tst_mov = train_test_split(response, movies,
'even', train_frac=0.8)
print 'Splitting out training and validation data...'
tr_rsp, val_rsp, tr_mov, val_mov = train_test_split(tr_rsp, tr_mov,
'even', train_frac=0.9)
tr_rsp = smooth_responses(tr_rsp)
val_rsp = smooth_responses(val_rsp)
tst_rsp = smooth_responses(tst_rsp)
train_set_x, train_set_y = window_matrices(tr_rsp, tr_mov, n_lag)
valid_set_x, valid_set_y = window_matrices(val_rsp, val_mov, n_lag)
test_set_x, test_set_y = window_matrices(tst_rsp, tst_mov, n_lag)
model = RegressionModel(model_name=name, n_hidden=n_hidden,
learning_rate=learning_rate,
n_epochs=n_epochs, batch_size=batch_size,
L1_reg=L1_reg, L2_reg=L2_reg)
model.setup_with_data([(train_set_x, train_set_y),
(valid_set_x, valid_set_y),
(test_set_x, test_set_y)])
test_error = model.train()
mlp_training_errors.append(test_error)
predicted = model.y_pred()
np.save(os.path.join(predicted_responses_dir, 'pred_%s' % name),
predicted)
with open(os.path.join(saved_models_dir,'mlp_%s' % name), 'wb') as f:
pickle.dump(model.regression.params, f)
with open(os.path.join(saved_models_dir, 'train_errors'), 'wb') as f:
pickle.dump(mlp_training_errors, f)