grad = T.grad(ls, params, disconnected_inputs='warn')
grad_fn = T.function([network.input_layer.input_var, out_var],
grad,
allow_input_downcast=True)
loss_fn = T.function([network.input_layer.input_var, out_var],
loss,
allow_input_downcast=True)
out = T.function([network.input_layer.input_var],
prediction,
allow_input_downcast=True)
Xs, Ys = synthesize_data(dim, batch_size, tasks, seed=seed)
train(train_fn, np.concatenate(Xs), np.concatenate(Ys), init_epochs)
print('------- initial training complete ---------------')
init_param_value = np.copy(network.get_param_values())
#Xs, Ys = synthesize_data(dim, 2000, tasks)
task_grads = []
task_grad_samples = []
for i in range(len(Xs)):
task_grads.append([])
task_grad_samples.append([])
total_grads = []
net_weight_values = []
#for i in range(epochs):
# net_weight_values.append(network.get_param_values())
# train(train_fn, np.concatenate(Xs), np.concatenate(Ys), 1)
epochs = 1
for i in range(epochs):
loss = loss.mean()
params = network.get_params(trainable=True)
updates = lasagne.updates.adam(loss, params, learning_rate=0.0005)
train_fn = T.function([network.input_layer.input_var, out_var], loss, updates=updates, allow_input_downcast=True)
ls = TT.mean((prediction - out_var)**2)
grad = T.grad(ls, params, disconnected_inputs='warn')
grad_fn = T.function([network.input_layer.input_var, out_var], grad, allow_input_downcast=True)
loss_fn = T.function([network.input_layer.input_var, out_var], loss, allow_input_downcast=True)
out = T.function([network.input_layer.input_var], prediction, allow_input_downcast=True)
if not load_init_policy:
losses=train(train_fn, np.concatenate(trainingXs), np.concatenate(trainingYs), init_epochs)
joblib.dump(network, 'data/trained/gradient_temp/supervised_split_' + append + '/init_network.pkl', compress=True)
print('------- initial training complete ---------------')
init_param_value = np.copy(network.get_param_values())
#Xs, Ys = synthesize_data(dim, 2000, tasks)
task_grads = []
for i in range(len(trainingXs)):
task_grads.append([])
if not load_split_data:
net_weight_values = []
for i in range(epochs):
net_weight_values.append(network.get_param_values())
train(train_fn, np.concatenate(trainingXs), np.concatenate(trainingYs), 1)
joblib.dump(net_weight_values, 'data/trained/gradient_temp/supervised_split_' + append + '/net_weight_values.pkl', compress=True)
else:
net_weight_values = joblib.load('data/trained/gradient_temp/supervised_split_' + append + '/net_weight_values.pkl')
for i in range(epochs):
network.set_param_values(net_weight_values[i])