def setUp(self):
self.tm = training_manager.TrainingManager()
self.tm.N_PACKAGES = 1
self.PROJECT_DIR = tb.setup_module()
# load in spm_data
data_path = self.PROJECT_DIR + "/dcm_rnn/resources/template0.pkl"
self.du = tb.load_template(data_path)
self.dr = tfm.DcmRnn()
self.dr.collect_parameters(self.du)
self.tm.prepare_dcm_rnn(self.dr, tag='initializer')
N_SEGMENTS = 1024
N_RECURRENT_STEP = 64
# STEP_SIZE = 0.002 # for 32
# STEP_SIZE = 0.5
# STEP_SIZE = 0.001 # for 64
# STEP_SIZE = 0.001 # 128
# STEP_SIZE = 0.0005 # for 256
STEP_SIZE = 1e-5
# DATA_SHIFT = int(N_RECURRENT_STEP / 4)
DATA_SHIFT = 1
LEARNING_RATE = 0.01 / N_RECURRENT_STEP
print(os.getcwd())
PROJECT_DIR = '/Users/yuanwang/Google_Drive/projects/Gits/DCM_RNN'
data_path = PROJECT_DIR + "/dcm_rnn/resources/template0.pkl"
du = tb.load_template(data_path)
dr = tfm.DcmRnn()
dr.collect_parameters(du)
dr.learning_rate = LEARNING_RATE
dr.shift_data = DATA_SHIFT
dr.n_recurrent_step = N_RECURRENT_STEP
neural_parameter_initial = {
'A': du.get('A'),
'B': du.get('B'),
'C': du.get('C')
}
dr.loss_weighting = {
'prediction': 1.,
'sparsity': 0.1,
'prior': 10,
package["Wxxu"] = w_true[1]
package["Wxu"] = w_true[2]
x_true = du.scan_x(package)
if if_plot:
plt.plot(x_hat, '--')
plt.plot(x_true)
return [x_hat, x_true]
# load in spm_data
print('working directory is ' + os.getcwd())
data_path = os.path.join(PROJECT_DIR, 'experiments', 'calculate_abc_from_x',
'spm_data.pkl')
data_package = tb.load_template(data_path)
data = data_package.data
# RESULT_PATH_DCM_RNN = PROJECT_DIR + "/dcm_rnn/resources/template0.pkl"
# du = tb.load_template(RESULT_PATH_DCM_RNN)
du = data['key']
signal_length = data['x_true_merged'].data.shape[0]
index_range = range(0, int(signal_length * 3 / 4))
# check whether du and spm_data match
np.testing.assert_array_almost_equal(
du.get('u')[index_range], data['u_merged'][index_range])
np.testing.assert_array_almost_equal(
du.get('x')[index_range], data['x_true_merged'][index_range])
# short hand for spm_data
x = data['x_hat_merged'].data[index_range]
noise = np.random.randn(du.get('y').shape[0], du.get('y').shape[1]) * noise_std
du._secured_data['y_noised'] = du._secured_data['y'] + noise
for i in range(du.get('n_node') + 1):
plt.subplot(4, 1, i + 1)
if i == 0:
plt.plot(du.get('u'))
else:
plt.plot(du.get('y_noised')[:, i - 1])
# save data after finding a good u
core = du.collect_parameter_core('y_noised')
# pickle.dump(core, open(CORE_PATH, 'wb'))
## do down sample and up sample, make the data used for DCM-RNN inference
# load data
core = tb.load_template(CORE_PATH)
du_original = tb.DataUnit()
du_original.load_parameter_core(core)
du_original.recover_data_unit()
du = du_original.resample_data_unit()
pickle.dump(du, open(SAVE_PATH_PKL, 'wb'))
# create DCM structure for SPM DCM
core = tb.load_template(CORE_PATH)
du = tb.DataUnit()
du.load_parameter_core(core)
du.recover_data_unit()
mask = du.create_support_mask()
down_sample_rate_u = 4
down_sample_rate_y = 128
MAX_EPOCHS = 32 * 5 CHECK_STEPS = 1 N_RECURRENT_STEP = 192 DATA_SHIFT = 2 MAX_BACK_TRACK = 16 MAX_CHANGE = 0.001 BATCH_RANDOM_DROP_RATE = 0.5 TARGET_T_DELTA = 1. / 16 N_CONFOUNDS = 19 # load data spm_data = sio.loadmat(RAW_DATA_PATH) spm_data['stimulus_names'] = ['Photic', 'Motion', 'Attention'] spm_data['node_names'] = ['V1', 'V5', 'SPC'] spm_data['u'] = spm_data['u'].todense() du = tb.load_template(TEMPLATE_PATH) # process spm_data, # up sample u and y to 16 frame/second shape = list(spm_data['u'].shape) n_time_point = shape[0] spm_data['u_upsampled'] = du.resample(spm_data['u'], shape, order=3) spm_data['y_upsampled'] = du.resample(spm_data['y'], shape, order=3) n_segments = mth.ceil((n_time_point - N_RECURRENT_STEP) / DATA_SHIFT) # assume the observation model is # y = DCM_RNN(u) + Confounds * weights + noise # Confounds are the first cosine transfer basis (19) confounds = idct(np.eye(n_time_point)[:, :N_CONFOUNDS], axis=0, norm='ortho') # plt.plot(confounds)
CONDITION = 'h1_s0_n0'
EXPERIMENT_PATH = os.path.join(PROJECT_DIR, 'experiments',
'compare_estimation_with_simulated_data')
DATA_PATH = os.path.join(EXPERIMENT_PATH, 'data')
RESULT_PATH = os.path.join(EXPERIMENT_PATH, 'results')
IMAGE_PATH = os.path.join(EXPERIMENT_PATH, 'images')
CORE_PATH = os.path.join(DATA_PATH, 'core.pkl')
DCM_RNN_RESULT_PATH = os.path.join(RESULT_PATH,
'estimation_' + CONDITION + '.pkl')
SPM_RESULT_PATH = os.path.join(RESULT_PATH, 'saved_data_' + CONDITION + '.mat')
# show estimation curves
core = tb.load_template(CORE_PATH)
du = tb.DataUnit()
du.load_parameter_core(core)
du.recover_data_unit()
if CONDITION == 'h1_s1_n1':
y_true = du.get('y_noised')[::128]
else:
y_true = du.get('y')[::128]
y_true.shape
du_rnn = pickle.load(open(DCM_RNN_RESULT_PATH, 'rb'))
y_rnn = du_rnn.get('y')[::32]
y_rnn.shape
spm = sio.loadmat(SPM_RESULT_PATH)
spm['b'] = np.rollaxis(spm['b'], 2) # correct matlab-python transfer error
'gamma': False,
'tao': False,
'epsilon': False,
'V0': False,
'TE': False,
'r0': False,
'theta0': False,
'x_h_coupling': False
}
EXPERIMENT_PATH = os.path.join(PROJECT_DIR, 'experiments', 'compare_estimation_with_simulated_data')
DATA_PATH = os.path.join(EXPERIMENT_PATH, 'data', 'du_DCM_RNN.pkl')
SAVE_PATH = os.path.join(EXPERIMENT_PATH, 'results', 'estimation_' + CONDITION + '.pkl')
# load data
du = tb.load_template(DATA_PATH)
if SETTINGS[CONDITION]['if_noised_y']:
du._secured_data['y'] = du.get('y_noised')
n_segments = mth.ceil(len(du.get('y')) - N_RECURRENT_STEP / DATA_SHIFT)
# specify initialization values, loss weighting factors, and mask (support of effective connectivity)
x_parameter_initial = {}
x_parameter_initial['A'] = - np.eye(du.get('n_node'))
x_parameter_initial['B'] = [np.zeros((du.get('n_node'), du.get('n_node'))) for _ in range(du.get('n_stimuli'))]
x_parameter_initial['C'] = np.zeros((du.get('n_node'), du.get('n_stimuli')))
x_parameter_initial['C'][0, 0] = 1.
x_parameter_initial['C'][1, 1] = 1.
h_parameter_initial = du.get('hemodynamic_parameter')
loss_weighting = {'prediction': 1., 'sparsity': 1., 'prior': 1., 'Wxx': 1., 'Wxxu': 1., 'Wxu': 1.}
mask = du.create_support_mask()