movie_type = 'movie'
spatial_downsample_factor = 4
# time_downsample_factor = 5
n_lag = 6
split_type = 'even'
saved_nns_dir = './temp/nets'
responses = load_responses(exp_type)
movies = load_movies(exp_type, movie_type,
downsample_factor=spatial_downsample_factor)
train_test_splits = map(lambda r: train_test_split(r, movies, split_type,
train_frac=0.7,
to_leave_out=0),
responses)
nns = []
for i, response in enumerate(responses):
name = response.name
print 'Mouse %s' % name
print 'Splitting out training and test data...'
tr_rsp, val_rsp, tr_mov, val_mov = train_test_splits[i]
tr_rsp = smooth_responses(tr_rsp)
val_rsp = smooth_responses(val_rsp)
network = fit_NN(tr_mov, tr_rsp, val_mov, val_rsp, n_lag)
nns.append(network)
network.save(os.path.join(saved_nns_dir, '%d.net' % i))
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)
movie_type = 'movie'
spatial_downsample_factor = 4
# time_downsample_factor = 5
n_lag = 6
split_type = 'even'
saved_nns_dir = './temp/nets'
responses = load_responses(exp_type)
movies = load_movies(exp_type,
movie_type,
downsample_factor=spatial_downsample_factor)
train_test_splits = map(
lambda r: train_test_split(
r, movies, split_type, train_frac=0.7, to_leave_out=0), responses)
nns = []
for i, response in enumerate(responses):
name = response.name
print 'Mouse %s' % name
print 'Splitting out training and test data...'
tr_rsp, val_rsp, tr_mov, val_mov = train_test_splits[i]
tr_rsp = smooth_responses(tr_rsp)
val_rsp = smooth_responses(val_rsp)
network = fit_NN(tr_mov, tr_rsp, val_mov, val_rsp, n_lag)
nns.append(network)
network.save(os.path.join(saved_nns_dir, '%d.net' % i))
# Preprocessing
from src.data_manip_utils import smooth_responses
locs_dirn = [os.path.join(DATA_DIR, '%s_dir.npy' % c) for c in MICE_NAMES]
data_dirn = map(lambda (n, loc): Response(n, loc), zip(MICE_NAMES, locs_dirn))
locs_ori = [os.path.join(DATA_DIR, '%s_ori.npy' % c) for c in MICE_NAMES]
data_ori = map(lambda (n, loc): Response(n, loc), zip(MICE_NAMES, locs_ori))
dirs_rad = np.radians(DIRECTIONS)
sigma0 = 2 * np.pi / len(DIRECTIONS) # initial tuning curve width
for index, (m_dir, m_ori) in enumerate(zip(data_dirn, data_ori)):
name = MICE_NAMES[index]
print 'Mouse %s' % name
# Smooth the responses first.
m_dir = smooth_responses(m_dir)
m_ori = smooth_responses(m_ori)
N = m_dir.data.shape[1]
# Get average response over all trials, time
m_dir.avg = np.mean(m_dir.data, axis=(2, 3))
m_ori.avg = np.mean(m_ori.data, axis=(2, 3))
# Obtain tuning curves
init_thetas = [
dirs_rad[np.argmax(m_dir.avg[:, i])] % np.pi for i in range(N)
]
init_cs = np.min(m_dir.avg, axis=0)
init_ws = np.max(m_dir.avg, axis=0) - np.min(m_dir.avg, axis=0)
init_sigmas = np.ones(N) * sigma0
from src.data_manip_utils import smooth_responses
n_clusters = 3
exp_type = 'natural'
if exp_type == 'natural':
from src.params.naturalmovies.datafile_params import PLOTS_DIR
elif exp_type == 'grating':
from src.params.grating.datafile_params import PLOTS_DIR
data = load_responses(exp_type)
for index, m in enumerate(data):
print 'Mouse %s' % m.name
if exp_type == 'natural':
m.data = m.data[:5,:,:,:]
data[index] = smooth_responses(m)
m = data[index]
m.clustering = NeuronClustering(n_clusters, signal_correlation)
m.clustering.fit(m.data)
m.data = m.clustering.divnorm(m.data)
S, N, L, R = m.data.shape
# Computing signal correlation.
m.signal_corr = signal_correlation(m.data)
# Computing noise correlation.
m.noise_corr = noise_correlation(m.data)
# Hierarchical clustering to rearrange the correlation matrix.