def get_optimizer(wrt,
fprime,
optimize_method,
step_rate,
momentum=0.0,
decay=0.9,
**kwargs):
'''Get an optimizer.'''
if optimize_method == 'rmsprop':
optimizer = RmsProp(wrt=wrt,
fprime=fprime,
step_rate=step_rate,
decay=decay,
momentum=momentum)
elif optimize_method == 'adam':
optimizer = Adam(wrt=wrt, fprime=fprime, step_rate=step_rate)
elif optimize_method == 'gd':
optimizer = GradientDescent(wrt=wrt,
fprime=fprime,
step_rate=step_rate,
momentum=momentum)
else:
raise ValueError('Can not load predefined optimization method %s' %
optimize_method)
return optimizer
def train(bm, theta_list, method, max_iter=1000, step_rate=0.1):
'''
train a Born Machine.
Args:
bm (QCBM): quantum circuit born machine training strategy.
theta_list (1darray): initial parameters.
method ('Adam'|'L-BFGS-B'):
* L-BFGS-B: efficient, but not noise tolerant.
* Adam: noise tolerant.
max_iter (int): maximum allowed number of iterations.
step_rate (float): learning rate for Adam optimizer.
Returns:
(float, 1darray): final loss and parameters.
'''
theta_list = np.array(theta_list)
if method == 'Adam':
from climin import Adam
optimizer = Adam(wrt=theta_list,
fprime=bm.gradient,
step_rate=step_rate)
for info in optimizer:
step = info['n_iter']
loss = bm.mmd_loss(theta_list)
print('step = %d, loss = %s' % (step, loss))
if step == max_iter:
break
return bm.mmd_loss(theta_list), theta_list
else:
res = minimize(
bm.mmd_loss,
x0=theta_list,
method=method,
jac=bm.gradient,
tol=1e-12,
options={
'maxiter': max_iter,
'disp': 2,
'gtol': 1e-12,
'ftol': 0
},
)
return res.fun, res.x
first_layer=True),
Linear(doutput, 10),
Softmax()
]
'''
# baseline purposes
model = [
Dropout(dropout),
LSTM(1, states, sigma=sigma, fbias=1.5, last_state_only=True),
Linear(states, 10),
Softmax()
]
'''
W = extract_weights(model)
optimizer = Adam(W, dW, learning_rate, momentum=momentum)
print 'Approx. Parameters: ', W.size
for i in optimizer:
if i['n_iter'] > niterations:
break
print str(data.epoch) + '\t' + str(i['n_iter']), '\t',
print logs['loss'][-1], '\t',
print logs['gradient_norm'][-1]
if data.epoch_complete:
inputs, labels = data.fetch_val()
nsamples = inputs.shape[2]
inputs = np.split(inputs, nsamples / batch_size, axis=2)
Linear(output, doutputs), Softmax() ] ''' model = [ Dropout(dropout), LSTM(dinput, states, fbias=0.0, sigma=0.1), Dropout(dropout), LSTM(states, states, fbias=0.0, sigma=0.1), Linear(states, doutputs), Softmax() ] W = extract_weights(model) optimizer = Adam(W, dW, learning_rate, momentum=momentum) config = 'experiment_name = ' + str(experiment_name) + '\n' + 'periods = ' + str(periods) + '\n' + 'vocabulary_size = ' + str(vocabulary_size) + '\n' + 'states = ' + str(states) + '\n' + 'output = ' + str(output) + '\n' + 'dinput = ' + str(dinput) + '\n' + 'doutputs = ' + str(doutputs) + '\n' + 'sequence_length = ' + str(sequence_length) + '\n' + 'batch_size = ' + str(batch_size) + '\n' + 'learning_rate = ' + str(learning_rate) + '\n' + 'niterations = ' + str(niterations) + '\n' + 'momentum = ' + str(momentum) + '\n' + 'forget_every = ' + str(forget_every) + '\n' + 'gradient_clip = ' + str(gradient_clip) + '\n' + 'sample_every = ' + str(sample_every) + '\n' + 'save_every = ' + str(save_every) + '\n' + 'plot_every = ' + str(plot_every) + '\n' + 'full_recurrence = ' + str(full_recurrence) + '\n' + 'learn_state = ' + str(learn_state) + '\n' + 'anneal = ' + str(anneal) + '\n' + 'dynamic_forgetting = ' + str(dynamic_forgetting) + '\n' + 'text = ' + str(text_file) + '\n' config += 'optimizer = ' + str(optimizer.__class__.__name__) + '\n' config += 'dropout = ' + str(dropout) + '\n' f = open(path + 'config.txt', 'w') f.write(config) f.close() print config, 'Approx. Parameters: ', W.size for i in optimizer: if i['n_iter'] > niterations: break