cost.name = "nll" cg = ComputationGraph(cost) model = Model(cost) transition_matrix = VariableFilter(theano_name_regex="state_to_state")(cg.parameters) for matr in transition_matrix: matr.set_value(0.98 * numpy.eye(hidden_size_recurrent, dtype=floatX)) from play.utils import regex_final_value extra_updates = [] for name, var in states.items(): update = tensor.switch( start_flag, 0.0 * var, VariableFilter(theano_name_regex=regex_final_value(name))(cg.auxiliary_variables)[0] ) extra_updates.append((var, update)) # Old values for n load_name = "sp_and_f0_1" from blocks.serialization import load main_loop = load(save_dir + "pkl/best_" + load_name + ".pkl") new_params = [] for key, value in model.get_parameter_dict().items(): if key in [
cg = ComputationGraph(cost) model = Model(cost) transition_matrix = VariableFilter(theano_name_regex="state_to_state")( cg.parameters) for matr in transition_matrix: matr.set_value(0.98 * numpy.eye(hidden_size_recurrent, dtype=floatX)) from play.utils import regex_final_value extra_updates = [] for name, var in states.items(): update = tensor.switch( start_flag, 0. * var, VariableFilter(theano_name_regex=regex_final_value(name))( cg.auxiliary_variables)[0]) extra_updates.append((var, update)) ################# # Monitoring vars ################# mean_data = x.mean(axis=(0, 1)).copy(name="data_mean") sigma_data = x.std(axis=(0, 1)).copy(name="data_std") max_data = x.max(axis=(0, 1)).copy(name="data_max") min_data = x.min(axis=(0, 1)).copy(name="data_min") monitoring_variables = [cost, lr] data_monitoring = [mean_data, sigma_data, max_data, min_data]
voiced = tensor.matrix('voiced') start_flag = tensor.scalar('start_flag') sp = tensor.tensor3('sp') f0s = f0.dimshuffle(0,1,'x') voiceds = voiced.dimshuffle(0,1,'x') x = tensor.concatenate([sp, f0s, voiceds], 2) cost_matrix = generator.cost_matrix(x) cg = ComputationGraph(cost_matrix) from blocks.filter import VariableFilter from play.utils import regex_final_value extra_updates = [] for name, var in states.items(): update = VariableFilter(theano_name_regex=regex_final_value(name))(cg.auxiliary_variables)[0] extra_updates.append((var, update)) #print function([f0, sp, voiced], cost_matrix, updates = extra_updates)(x_tr[0],x_tr[1],x_tr[2]) #generator.generate(n_steps=steps, batch_size=n_samples, iterate=True, **states) #states = {} sample = ComputationGraph(generator.generate(n_steps=steps, batch_size=n_samples, iterate=True, **states)) sample_fn = sample.get_theano_function() outputs_bp = sample_fn()[-2] for this_sample in range(n_samples): print "Iteration: ", this_sample
cost = cost_matrix.mean() + 0.*start_flag cost.name = "nll" cg = ComputationGraph(cost) model = Model(cost) transition_matrix = VariableFilter( theano_name_regex="state_to_state")(cg.parameters) for matr in transition_matrix: matr.set_value(0.98*numpy.eye(hidden_size_recurrent, dtype=floatX)) from play.utils import regex_final_value extra_updates = [] for name, var in states.items(): update = tensor.switch(start_flag, 0.*var, VariableFilter(theano_name_regex=regex_final_value(name) )(cg.auxiliary_variables)[0]) extra_updates.append((var, update)) ################# # Monitoring vars ################# mean_data = x.mean(axis = (0,1)).copy(name="data_mean") sigma_data = x.std(axis = (0,1)).copy(name="data_std") max_data = x.max(axis = (0,1)).copy(name="data_max") min_data = x.min(axis = (0,1)).copy(name="data_min") monitoring_variables = [cost, lr] data_monitoring = [mean_data, sigma_data,