def train(batch_size, lr, epochs, period):
assert period >= batch_size and period % batch_size == 0
params, sqrs, vs = adam_init_params()
w = params[0]
b = params[1]
total_loss = [np.mean(square_loss(net(X, w, b), y).asnumpy())]
t = 0
for epoch in range(1, epochs + 1):
if epoch > 2:
lr *= 0.1
for batch_i, data, label in data_iter(batch_size):
with autograd.record():
output = net(data, w, b)
loss = square_loss(output, label)
loss.backward()
t += 1
optimizer.adam(params, sqrs, vs, batch_size, lr, t)
if batch_i * batch_size % period == 0:
total_loss.append(np.mean(square_loss(net(X,w ,b), y).asnumpy()))
print("Batch size %d, Learning rate %f, Epoch %d, loss %.4e" % (
batch_size, lr, epoch, total_loss[-1]
))
print("w:", np.reshape(w.asnumpy(), (1, -1)),
"b:", b.asnumpy()[0])
print("Total loss length:", len(total_loss))
x_axis = np.linspace(0, epochs, len(total_loss), endpoint=True)
plt.semilogy(x_axis, total_loss)
plt.xlabel("Epochs")
plt.ylabel("Loss")
plt.show()
def __init__(self, x, y, l, window, opt, lr, init_emb, dim_emb, dim_hidden, n_vocab, L2_reg, unit,
sim='cos', n_layers=1, activation=tanh):
self.tr_inputs = [x, y, l]
self.pr_inputs = [x, y, l]
self.x = x # 1D: batch_size * l * 2, 2D: window; elem=word_id
self.y = y # 1D: batch_size; elem=label
self.l = l # scalar: elem=sentence length
batch_size = y.shape[0]
n_cands = x.shape[0] / batch_size / l
self.pad = build_shared_zeros((1, dim_emb))
if init_emb is None:
self.emb = theano.shared(sample_weights(n_vocab - 1, dim_emb))
else:
self.emb = theano.shared(init_emb)
self.E = T.concatenate([self.pad, self.emb], 0)
self.W_out = theano.shared(sample_weights(dim_hidden, dim_hidden))
self.params = [self.emb, self.W_out]
""" Input Layer """
e = self.E[x] # e: 1D: batch_size * l * 2, 2D: window, 3D: dim_emb
x_in = e.reshape((batch_size * n_cands, l, -1))
""" Intermediate Layer """
# h: 1D: n_batch * n_cands, 2D: dim_emb
h, params = cnn.layers(x_in, window, dim_emb, dim_hidden, n_layers, activation)
self.params.extend(params)
""" Output Layer """
h = h.reshape((batch_size, n_cands, -1))
h_1 = h[T.arange(batch_size), 0]
h_2 = h[T.arange(batch_size), 1:]
if sim == 'cos':
y_score = cosign_similarity(h_1, h_2)
else:
y_score = T.batched_dot(T.dot(h_1, self.W_out), h_2.dimshuffle(0, 2, 1))
y_score_hat = T.max(y_score, 1)
""" Objective Function """
self.nll = max_margin_loss(y_score_hat, y_score[T.arange(batch_size), y])
self.L2_sqr = regularization(self.params)
self.cost = self.nll + L2_reg * self.L2_sqr / 2.
""" Optimization """
if opt == 'adagrad':
self.update = ada_grad(cost=self.cost, params=self.params, lr=lr)
elif opt == 'ada_delta':
self.update = ada_delta(cost=self.cost, params=self.params)
elif opt == 'adam':
self.update = adam(cost=self.cost, params=self.params, lr=lr)
else:
self.update = sgd(cost=self.cost, params=self.params, lr=lr)
""" Predicts """
y_hat = T.argmax(y_score, 1)
""" Check Accuracies """
self.correct = T.eq(y_hat, y)
def run_bm(bm_fname, optz_cfg=None, verbose=True, misc=None):
# load bm_fname
bm = importlib.import_module(bm_fname.rsplit('.', 1)[0])
e = bm.e; decorate_stind(e)
thts_init = bm.thts_init
compare = bm.compare
if optz_cfg is None: optz_cfg = bm.optz_cfg
# optz_detail
optz_detail = get_optz_detail(bm_fname, optz_cfg)
print('\n===== OPTZ: %s =====' % optz_detail)
# run experiments
for alg_str in compare:
print('[%s] ' % alg_str, end='')
alg = importlib.import_module(compare[alg_str].rsplit('.',1)[0])
alg.init(e)
misc_arg = {'misc':misc} if alg_str == 'ours2' else {}
# run adam
grad_func = lambda thts, e=e: alg.elbo_grad(e, thts, **misc_arg)
thts_res = optimizer.adam(grad_func, thts_init,
iter_n = optz_cfg['iter_n'],
lr = optz_cfg['lr'],
sample_n_grad = optz_cfg['sample_n_grad'],
sample_n_var = optz_cfg['sample_n_var'],
verbose = verbose,
**misc_arg)
# save res to file
save_res(optz_detail, alg_str, thts_res)
def ScbowTrain(file):
corpus , word_to_id , id_to_word = preprocess(file)
contexts, target = create_context_target(corpus, window_size = 1)
vocab_size = len(word_to_id)
target = one_hot_v(target,vocab_size)
contexts = one_hot_v(contexts,vocab_size)
model = scbow(vocab_size , hidden_size)
optimizer = adam()
train = Trainer(model, optimizer)
train.fit(contexts, target, max_epoch, batch_size)
train.plot()
word_vecs = model.word_vecs
for word_id, word in id_to_word.items():
print(word,word_vecs[word_id])
C = co_mat(corpus,vocab_size,window_size = 1)
ms('彼女',word_to_id,id_to_word,C,top = 10)
def build_model(self, lr=0.001, dropout=None):
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype='int32') # step * samples
x_mask = T.matrix('x_mask', dtype='float32') # step * samples
y = T.matrix('y', dtype='int32') # sample * emb
ctx = T.tensor3('ctx', dtype='float32') # sample * annotation * dim
n_timesteps = x.shape[0]
n_samples = x.shape[1]
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
ctx0 = ctx
ctx_mean = ctx0.mean(1)
init_state = T.dot(ctx_mean, self.W_hidden_init) + self.b_hidden_init
init_memory = T.dot(ctx_mean, self.W_memory_init) + self.b_memory_init
# proj : lstm hidden 들의 리스트
proj = self.lstm_layer(emb,
mask=x_mask,
context=ctx,
init_state=init_state,
init_memory=init_memory)
proj_h = proj[0]
# hidden 들의 평균
proj_h = (proj_h * x_mask[:, :, None]).sum(axis=0)
proj_h = proj_h / x_mask.sum(axis=0)[:, None] # sample * dim
# 마지막 hidden
#proj_h = proj_h[-1] # sample * dim
if dropout is not None:
proj_h = dropout_layer(proj_h, use_noise, trng, dropout)
output = T.dot(proj_h, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, x_mask, ctx, y, cost, updates, prediction
def update(self, lr=0.00001, weight_decay=0.0004):
'''
# mini-batch SGD
self.weights *= (1 - weight_decay)
self.bias *= (1 - weight_decay)
self.weights -= lr * self.d_weights
self.bias -= lr * self.d_bias
'''
# adam optimizer
self.weights, self.config_w = adam(self.weights * (1 - weight_decay),
self.d_weights,
config=self.config_w)
self.bias, self.config_b = adam(self.bias * (1 - weight_decay),
self.d_bias,
config=self.config_b)
# clear gradients
self.d_weights = np.zeros(self.weights.shape)
self.d_bias = np.zeros(self.bias.shape)
def build_model(self, lr=0.001):
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype = 'int32')
x_mask = T.matrix('x_mask', dtype='float32')
y = T.matrix('y', dtype = 'int32')
img = T.matrix('img', dtype = 'float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
init_state = T.dot(img, self.W_img_emb) + self.b_img_emb
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
# proj : gru hidden 들의 리스트
proj = self.gru_layer(emb, init_state, mask=x_mask)
# hidden 들의 평균
proj = (proj * x_mask[:, :, None]).sum(axis=0)
proj = proj / x_mask.sum(axis=0)[:, None] # sample * dim
# 마지막 hidden
#proj = proj[-1] # sample * dim
output = T.dot(proj, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, x_mask, img, y, cost, updates, prediction
def build_model(self, lr=0.001):
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype='int32')
x_mask = T.matrix('x_mask', dtype='float32')
y = T.matrix('y', dtype='int32')
img = T.matrix('img', dtype='float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
init_state = T.dot(img, self.W_img_emb) + self.b_img_emb
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
# proj : gru hidden 들의 리스트
proj = self.gru_layer(emb, init_state, mask=x_mask)
# hidden 들의 평균
#proj = (proj * x_mask[:, :, None]).sum(axis=0)
#proj = proj / x_mask.sum(axis=0)[:, None] # sample * dim
# 마지막 hidden
proj = proj[-1] # sample * dim
output = T.dot(proj, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return x, x_mask, img, y, cost, updates, prediction
def build_model(self, lr=0.001, dropout=None):
def concatenate(tensor_list, axis=0):
concat_size = sum(tt.shape[axis] for tt in tensor_list)
output_shape = ()
for k in range(axis):
output_shape += (tensor_list[0].shape[k],)
output_shape += (concat_size,)
for k in range(axis + 1, tensor_list[0].ndim):
output_shape += (tensor_list[0].shape[k],)
out = tensor.zeros(output_shape)
offset = 0
for tt in tensor_list:
indices = ()
for k in range(axis):
indices += (slice(None),)
indices += (slice(offset, offset + tt.shape[axis]),)
for k in range(axis + 1, tensor_list[0].ndim):
indices += (slice(None),)
out = tensor.set_subtensor(out[indices], tt)
offset += tt.shape[axis]
return out
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype = 'int32') # step * samples
x_mask = T.matrix('x_mask', dtype='float32') # step * samples
y = T.matrix('y', dtype = 'int32') # sample * emb
ctx = T.tensor3('ctx', dtype = 'float32') # sample * annotation * dim
n_timesteps = x.shape[0]
n_samples = x.shape[1]
xr = x[::-1]
xr_mask = x_mask[::-1]
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
embr = self.W_emb[xr.flatten()]
embr = embr.reshape([n_timesteps, n_samples, self.dim_word])
ctx0 = ctx
ctx_mean = ctx0.mean(1)
init_state = T.dot(ctx_mean, self.W_ctx_init) + self.b_ctx_init
# proj : gru hidden 들의 리스트
proj = self.gru_layer(emb, mask=x_mask, context=ctx, init_state=init_state)
proj_h = proj[0]
projr = self.gru_layer(embr, mask=xr_mask, context=ctx, init_state=init_state)
projr_h = projr[0]
concat_proj_h = concatenate([proj_h, projr_h[::-1]], axis=proj_h.ndim-1)
# step_ctx : step * samples * (dim*2)
concat_proj_h = (concat_proj_h * x_mask[:,:,None]).sum(0) / x_mask.sum(0)[:,None]
# step_ctx_mean : samples * (dim*2)
if dropout is not None :
concat_proj_h = dropout_layer(concat_proj_h, use_noise, trng, dropout)
output = T.dot(concat_proj_h, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, x_mask, ctx, y, cost, updates, prediction
def build_model(self, lr=0.001, dropout=None):
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype = 'int32') # step * samples
x_mask = T.matrix('x_mask', dtype='float32') # step * samples
y = T.matrix('y', dtype = 'int32') # sample * emb
ctx = T.tensor3('ctx', dtype = 'float32') # sample * annotation * dim
n_timesteps = x.shape[0]
n_samples = x.shape[1]
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
ctx0 = ctx
ctx_mean = ctx0.mean(1)
init_state = T.dot(ctx_mean, self.W_hidden_init) + self.b_hidden_init
init_memory = T.dot(ctx_mean, self.W_memory_init) + self.b_memory_init
# proj : lstm hidden 들의 리스트
proj = self.lstm_layer(emb, mask=x_mask, context=ctx, init_state=init_state, init_memory=init_memory)
proj_h = proj[0]
# hidden 들의 평균
proj_h = (proj_h * x_mask[:, :, None]).sum(axis=0)
proj_h = proj_h / x_mask.sum(axis=0)[:, None] # sample * dim
# 마지막 hidden
#proj_h = proj_h[-1] # sample * dim
if dropout is not None :
proj_h = dropout_layer(proj_h, use_noise, trng, dropout)
output = T.dot(proj_h, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, x_mask, ctx, y, cost, updates, prediction
def build_model(self, lr=0.001, dropout=None):
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.tensor3('x', dtype = 'float32') # step * sample * 5555
y = T.matrix('y', dtype = 'int32')
img = T.tensor3('img', dtype = 'float32') # 1*sample * 4096
# T.set_subtensor(img_t3[0], img)
# emb=theano.tensor.concatenate([img_t3,x])
emb = x
embr = x[::-1]
# proj : gru hidden 들의 리스트
proj = self.gru_layer(emb, img)
projr = self.gru_cond_layer(embr, img)
proj = concatenate([proj, projr[::-1]], axis=proj.ndim-1)
# hidden 들의 평균
proj = proj.mean(axis=0)
# 마지막 hidden
#proj = proj[-1] # sample * dim
if dropout is not None :
proj = dropout_layer(proj, use_noise, trng, dropout)
output = T.dot(proj, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
'''
decay_c = 0.000001
# add L2 regularization costs
if decay_c > 0.:
decay_c = theano.shared(np.float32(decay_c), name='decay_c')
weight_decay = 0.
for vv in self.params:
weight_decay += (vv ** 2).sum()
weight_decay *= decay_c
cost += weight_decay
'''
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, img, y, cost, updates, prediction, probs
def build_model(self, lr=0.001, dropout=None):
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.tensor3('x', dtype = 'float32')
y = T.matrix('y', dtype = 'int32')
img = T.matrix('img', dtype = 'float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
init_state = T.dot(img, self.W_hidden_init) + self.b_hidden_init
init_memory = T.dot(img, self.W_memory_init) + self.b_memory_init
emb = x
embr = x.swapaxes(0,1)[::-1].swapaxes(0,1)
# proj : gru hidden 들의 리스트
proj = self.lstm_layer(emb, init_state=init_state, init_memory=init_memory)[0]
projr = self.lstm_layer(embr, init_state=init_state, init_memory=init_memory)[0]
proj = concatenate([proj, projr[::-1]], axis=proj.ndim-1)
# hidden 들의 평균
proj = proj.mean(axis=0)
# 마지막 hidden
#proj = proj[-1] # sample * dim
if dropout is not None :
proj = dropout_layer(proj, use_noise, trng, dropout)
output = T.dot(proj, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-8
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
'''
decay_c = 0.000001
# add L2 regularization costs
if decay_c > 0.:
decay_c = theano.shared(np.float32(decay_c), name='decay_c')
weight_decay = 0.
for vv in self.params:
weight_decay += (vv ** 2).sum()
weight_decay *= decay_c
cost += weight_decay
'''
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, img, y, cost, updates, prediction
b_h3 = shared_normal((2, n_h3), sigma = 0)
b_h4 = shared_normal((2, n_h4), sigma = 0)
X = binomial(X) #internal binarization
#model calls
[dout_prob_y, dout_dual_recon_err] = model_NG_ACE(X, batch_size, gaussian_err, 0.2, 0.5) #with dropout
[prob_y, dual_recon_err] = model_NG_ACE(X, batch_size, gaussian_err, 0., 0.) #without dropout
y_model = T.argmax(prob_y, axis=1) #model labels
#dropout classification err
dout_class_err = T.nnet.categorical_crossentropy(dout_prob_y, Y).sum()
#optimizer call
cost = dout_class_err + dout_dual_recon_err
params = [W_h, W_h2, W_h3, W_h4, W_o, b_h, b_h2, b_h3, b_h4]
updates, norm_grad = adam(cost, params, lr = learning_rate, data_part = float(batch_size) / P)
#givens
s_trX, s_teX, s_trY, s_teY = shared(trX), shared(teX), shared(trY), shared(teY)
tr_batch_X = s_trX[start : end]
tr_batch_Y = s_trY[start : end]
te_batch_X = s_teX[start : end]
te_batch_Y = s_teY[start : end]
#train & test functions
mode = theano.compile.get_default_mode()
train = theano.function(inputs=[start, end, learning_rate], outputs= [dout_class_err, dual_recon_err, y_model, norm_grad], updates=updates, givens = {X : tr_batch_X, Y : tr_batch_Y}, allow_input_downcast=True, mode = mode)
test = theano.function(inputs=[start, end], outputs= [dual_recon_err, y_model], givens = {X : te_batch_X}, allow_input_downcast=True, mode = mode)
#main loop over epochs
tr_len = len(trY)
def main():
logging.info("️start loading setting data.")
settings = LearningDataSettings(args.train_setting_file)
logging.info("☑ loading setting data complete.")
vector_size = settings.input_unit
hidden_unit_num = settings.hidden_unit
class_num = settings.class_unit
hidden_layer_value_num = args.division_num
logging.info(
"input_vector(n):%d, hidden_unit(m):%d, class_num(K):%d, div_num:%d" %
(vector_size, hidden_unit_num, class_num, hidden_layer_value_num))
drbm = DRBM.load_from_json(settings.initial_model,
args.division_num,
args.sparse,
sparse_learning_rate=args.sparse_learning_rate,
sparse_adamax=args.sparse_adamax)
logging.info("initial model: {}".format(str(drbm)))
if args.datasize_limit != 0 and not args.generative_model:
settings.training_data = settings.training_data.restore_minibatch(
args.datasize_limit, random=False)
gen_drbm = None
if args.kl_divergence:
gen_drbm = DRBM.load_from_json(args.kl_divergence)
logging.info("generative model: {}".format(str(gen_drbm)))
elif args.generative_model:
gen_drbm = DRBM(settings.gen_input,
settings.gen_hidden,
settings.gen_class,
0,
random_bias=True)
logging.info("generated generative model: {}".format(str(gen_drbm)))
value, target = gen_drbm.stick_break(args.datasize_limit)
settings.training_data = Categorical(value, target, class_num)
settings.test_data = Categorical(np.array([]), np.array([]), class_num)
opt = None
if args.optimizer == "momentum":
logging.info("optimize method: momentum")
opt = optimizer.momentum(vector_size, hidden_unit_num, class_num)
elif args.optimizer == "adam":
logging.info("optimize method: adam")
opt = optimizer.adam(vector_size, hidden_unit_num, class_num)
else:
logging.info("optimize method: adamax")
opt = optimizer.adamax(vector_size, hidden_unit_num, class_num)
logging.info("train started.")
start_time = time.time()
learning_result = drbm.train(
settings.training_data,
settings.test_data,
args.learning_num,
args.minibatch_size,
opt,
test_interval=args.test_interval,
correct_rate=args.correct_rate,
gen_drbm=gen_drbm,
)
end_time = time.time()
logging.info("☑ train complete. time: {} sec".format(end_time -
start_time))
now = datetime.datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
hidden_layer = "s" if args.sparse else "d"
filename_template = "{}_{}_{}{}_v{}h{}c{}_%s.json".format(
now, args.filename_prefix, hidden_layer, drbm.div_num,
drbm.num_visible, drbm.num_hidden, drbm.num_class)
learning_result.save(
os.path.join(args.result_directory, filename_template % "log"))
drbm.save(os.path.join(args.result_directory,
filename_template % "params"))
logging.info("☑ parameters dumped.")
def main(method, LR_start, Binarize_weight_only):
# BN parameters
name = "mnist"
print("dataset = " + str(name))
print("Binarize_weight_only=" + str(Binarize_weight_only))
print("Method = " + str(method))
# alpha is the exponential moving average factor
alpha = .1
print("alpha = " + str(alpha))
epsilon = 1e-4
print("epsilon = " + str(epsilon))
batch_size = 100
print("batch_size = " + str(batch_size))
num_epochs = 50
print("num_epochs = " + str(num_epochs))
# network structure
num_units = 2048
print("num_units = " + str(num_units))
n_hidden_layers = 3
print("n_hidden_layers = " + str(n_hidden_layers))
print("LR_start = " + str(LR_start))
LR_decay = 0.1
print("LR_decay=" + str(LR_decay))
if Binarize_weight_only == "w":
activation = lasagne.nonlinearities.rectify
else:
activation = lab.binary_tanh_unit
print("activation = " + str(activation))
print('Loading MNIST dataset...')
train_set = MNIST(which_set='train', start=0, stop=50000, center=True)
valid_set = MNIST(which_set='train', start=50000, stop=60000, center=True)
test_set = MNIST(which_set='test', center=True)
# bc01 format
train_set.X = train_set.X.reshape(-1, 1, 28, 28)
valid_set.X = valid_set.X.reshape(-1, 1, 28, 28)
test_set.X = test_set.X.reshape(-1, 1, 28, 28)
# flatten targets
train_set.y = np.hstack(train_set.y)
valid_set.y = np.hstack(valid_set.y)
test_set.y = np.hstack(test_set.y)
# Onehot the targets
train_set.y = np.float32(np.eye(10)[train_set.y])
valid_set.y = np.float32(np.eye(10)[valid_set.y])
test_set.y = np.float32(np.eye(10)[test_set.y])
# for hinge loss
train_set.y = 2 * train_set.y - 1.
valid_set.y = 2 * valid_set.y - 1.
test_set.y = 2 * test_set.y - 1.
print('Building the MLP...')
# Prepare Theano variables for inputs and targets
input = T.tensor4('inputs')
target = T.matrix('targets')
LR = T.scalar('LR', dtype=theano.config.floatX)
mlp = lasagne.layers.InputLayer(shape=(None, 1, 28, 28), input_var=input)
for k in range(n_hidden_layers):
mlp = lab.DenseLayer(mlp,
nonlinearity=lasagne.nonlinearities.identity,
num_units=num_units,
method=method)
mlp = batch_norm.BatchNormLayer(mlp, epsilon=epsilon, alpha=alpha)
mlp = lasagne.layers.NonlinearityLayer(mlp, nonlinearity=activation)
mlp = lab.DenseLayer(mlp,
nonlinearity=lasagne.nonlinearities.identity,
num_units=10,
method=method)
mlp = batch_norm.BatchNormLayer(mlp, epsilon=epsilon, alpha=alpha)
train_output = lasagne.layers.get_output(mlp, deterministic=False)
# squared hinge loss
loss = T.mean(T.sqr(T.maximum(0., 1. - target * train_output)))
if method != "FPN":
# W updates
W = lasagne.layers.get_all_params(mlp, binary=True)
W_grads = lab.compute_grads(loss, mlp)
updates = optimizer.adam(loss_or_grads=W_grads,
params=W,
learning_rate=LR)
updates = lab.clipping_scaling(updates, mlp)
# other parameters updates
params = lasagne.layers.get_all_params(mlp,
trainable=True,
binary=False)
updates = OrderedDict(updates.items() + optimizer.adam(
loss_or_grads=loss, params=params, learning_rate=LR).items())
## update 2nd moment, can get from the adam optimizer also
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(mlp, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp] = acc_tag_temp * beta2 + W_grads[
idx] * W_grads[idx] * (1 - beta2)
idx = idx + 1
updates = OrderedDict(updates.items() + updates3.items())
else:
params = lasagne.layers.get_all_params(mlp, trainable=True)
updates = optimizer.adam(loss_or_grads=loss,
params=params,
learning_rate=LR)
test_output = lasagne.layers.get_output(mlp, deterministic=True)
test_loss = T.mean(T.sqr(T.maximum(0., 1. - target * test_output)))
test_err = T.mean(T.neq(T.argmax(test_output, axis=1),
T.argmax(target, axis=1)),
dtype=theano.config.floatX)
# Compile a function performing a training step on a mini-batch (by giving the updates dictionary)
# and returning the corresponding training loss:
train_fn = theano.function([input, target, LR], loss, updates=updates)
val_fn = theano.function([input, target], [test_loss, test_err])
print('Training...')
lab.train(name, method, train_fn, val_fn, batch_size, LR_start, LR_decay,
num_epochs, train_set.X, train_set.y, valid_set.X, valid_set.y,
test_set.X, test_set.y)
def main(method,LR_start, SEQ_LENGTH):
lasagne.random.set_rng(np.random.RandomState(1))
name = "linux"
print("dataset = "+str(name))
print("Method = "+str(method))
# Sequence Length
SEQ_LENGTH = SEQ_LENGTH
# SEQ_LENGTH = 100 #can have diffvalues 50, 100, 200
print("SEQ_LENGTH = "+str(SEQ_LENGTH))
# Number of units in the two hidden (LSTM) layers
N_HIDDEN = 512
print("N_HIDDEN = "+str(N_HIDDEN))
# All gradients above this will be clipped
GRAD_CLIP=5. #### this clip the gradients at every time step, while T.clip clips the sum of gradients as a whole
print("GRAD_CLIP ="+str(GRAD_CLIP))
# Number of epochs to train the net
num_epochs = 200
print("num_epochs = "+str(num_epochs))
# Batch Size
batch_size = 100
print("batch_size = "+str(batch_size))
print("LR_start = "+str(LR_start))
LR_decay = 0.98
print("LR_decay="+str(LR_decay))
activation = lasagne.nonlinearities.tanh
## load data, change data file dir
with open('data/linux_input.txt', 'r') as f:
in_text = f.read()
generation_phrase = "Copyright (C) 1992, 1998-2004 Linus Torvalds, Ingo Molnar\n *\n * This file contains the interrupt probing code and driver APIs.\n */\n\n#include"
#This snippet loads the text file and creates dictionaries to
#encode characters into a vector-space representation and vice-versa.
chars = list(set(in_text))
data_size, vocab_size = len(in_text), len(chars)
char_to_ix = { ch:i for i,ch in enumerate(chars) }
ix_to_char = { i:ch for i,ch in enumerate(chars) }
num_splits = [0.9, 0.05, 0.05]
num_splits_all = np.floor(data_size/batch_size/SEQ_LENGTH)
num_train = np.floor(num_splits_all*num_splits[0])
num_val = np.floor(num_splits_all*num_splits[1])
num_test = num_splits_all - num_train - num_val
train_X = in_text[0:(num_train*batch_size*SEQ_LENGTH+1).astype('int32')]
val_X = in_text[(num_train*batch_size*SEQ_LENGTH).astype('int32'):((num_train+num_val)*batch_size*SEQ_LENGTH+1).astype('int32')]
test_X = in_text[((num_train+num_val)*batch_size*SEQ_LENGTH).astype('int32'):(num_splits_all*batch_size*SEQ_LENGTH+1).astype('int32')]
## build model
print('Building the model...')
# input = T.tensor3('inputs')
target = T.imatrix('target')
LR = T.scalar('LR', dtype=theano.config.floatX)
# (batch size, SEQ_LENGTH, num_features)
l_in = lasagne.layers.InputLayer(shape=(None, None, vocab_size))
l_forward_2 = laq.LSTMLayer(
l_in,
num_units=N_HIDDEN,
grad_clipping=GRAD_CLIP,
peepholes=False,
nonlinearity=activation, ### change this activation can change the hidden layer to binary
method=method) ### batch_size*SEQ_LENGTH*N_HIDDEN
l_shp = lasagne.layers.ReshapeLayer(l_forward_2, (-1, N_HIDDEN)) ## (batch_size*SEQ_LENGTH, N_HIDDEN)
l_out = lasagne.layers.DenseLayer(l_shp, num_units=vocab_size, W = lasagne.init.Normal(), nonlinearity=lasagne.nonlinearities.softmax)
batchsize, seqlen, _ = l_in.input_var.shape
train_output = lasagne.layers.get_output(l_out, deterministic=False)
loss = T.nnet.categorical_crossentropy(train_output,target.flatten()).mean()
if method!= "FPN":
# W updates
W = lasagne.layers.get_all_params(l_out, quantized=True)
W_grads = laq.compute_grads(loss,l_out)
updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR, epsilon=1e-8)
updates = laq.clipping_scaling(updates,l_out)
# other parameters updates
params = lasagne.layers.get_all_params(l_out, trainable=True, quantized=False)
updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR, epsilon = 1e-8).items())
## update the ternary matrix
ternary_weights = laq.get_quantized_weights(loss, l_out)
updates2 = OrderedDict()
idx = 0
tt_tag = lasagne.layers.get_all_params(l_out, tt=True)
for tt_tag_temp in tt_tag:
updates2[tt_tag_temp]= ternary_weights[idx]
idx = idx+1
updates = OrderedDict(updates.items() + updates2.items())
## update 2nd momentum
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(l_out, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2)
idx = idx+1
updates = OrderedDict(updates.items() + updates3.items())
else:
params_other = lasagne.layers.get_all_params(l_out, trainable=True)
W_grads = [theano.grad(loss, wrt=l_forward_2.W_in_to_ingate), theano.grad(loss, wrt=l_forward_2.W_hid_to_ingate),
theano.grad(loss, wrt=l_forward_2.W_in_to_forgetgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_forgetgate),
theano.grad(loss, wrt=l_forward_2.W_in_to_cell),theano.grad(loss, wrt=l_forward_2.W_hid_to_cell),
theano.grad(loss, wrt=l_forward_2.W_in_to_outgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_outgate)]
updates = optimizer.adam(loss_or_grads=loss, params=params_other, learning_rate=LR)
test_output = lasagne.layers.get_output(l_out, deterministic=True)
test_loss = T.nnet.categorical_crossentropy(test_output,target.flatten()).mean()
train_fn = theano.function([l_in.input_var, target, LR], loss, updates=updates, allow_input_downcast=True)
val_fn = theano.function([l_in.input_var, target], test_loss, allow_input_downcast=True)
print('Training...')
X_train = train_X
X_val = val_X
X_test = test_X
def gen_data(pp, batch_size,SEQ_LENGTH, data, return_target=True):
x = np.zeros((batch_size,SEQ_LENGTH,vocab_size)) ###### 128*100*85
y = np.zeros((batch_size, SEQ_LENGTH))
for n in range(batch_size):
# ptr = n
for i in range(SEQ_LENGTH):
x[n,i,char_to_ix[data[pp[n]*SEQ_LENGTH+i]]] = 1.
y[n,i] = char_to_ix[data[pp[n]*SEQ_LENGTH+i+1]]
return x, np.array(y,dtype='int32')
in_text = X_train+X_val+X_test
chars = list(set(in_text))
data_size, vocab_size = len(in_text), len(chars)
char_to_ix = { ch:i for i,ch in enumerate(chars) }
ix_to_char = { i:ch for i,ch in enumerate(chars) }
def train_epoch(X,LR):
loss = 0
batches = len(X)/batch_size/SEQ_LENGTH
num_seq = len(X)/SEQ_LENGTH
shuffled_ind = range(num_seq)
np.random.shuffle(shuffled_ind)
for i in range(batches):
tmp_ind = shuffled_ind[i*batch_size:(i+1)*batch_size]
xx,yy = gen_data(tmp_ind,batch_size,SEQ_LENGTH, X)
new_loss = train_fn(xx,yy,LR)
loss+=new_loss
loss=loss/batches
return loss
# This function tests the model a full epoch (on the whole dataset)
def val_epoch(X):
# err = 0
loss = 0
batches = len(X)/batch_size/SEQ_LENGTH
num_seq = len(X)/SEQ_LENGTH
ind = range(num_seq)
for i in range(batches):
tmp_ind = ind[i*batch_size:(i+1)*batch_size]
xx, yy = gen_data(tmp_ind, batch_size, SEQ_LENGTH, X)
new_loss = val_fn(xx,yy)
loss += new_loss
loss = loss/batches
return loss
best_val_loss=100
best_epoch = 1
LR = LR_start
# iterate over epochs:
for epoch in range(1,num_epochs+1):
start_time = time.time()
train_loss = train_epoch(X_train, LR)
val_loss = val_epoch(X_val)
# test if validation error went down
if val_loss <= best_val_loss:
best_val_loss = val_loss
best_epoch = epoch
test_loss = val_epoch(X_test)
# all_params = lasagne.layers.get_all_params(l_out)
# np.savez("{0}/{1}_seq{2}_lr{3}_hid{4}_{5}.npz".format(method, name, SEQ_LENGTH, LR_start, N_HIDDEN, method), *all_params)
epoch_duration = time.time() - start_time
# Then we print the results for this epoch:
print(" Epoch "+str(epoch)+" of "+str(num_epochs)+" took "+str(epoch_duration)+"s")
print(" LR: "+str(LR))
print(" training loss: "+str(train_loss))
print(" validation loss: "+str(val_loss))
print(" best epoch: "+str(best_epoch))
print(" test loss: "+str(test_loss))
with open("{0}/{1}_seq{2}_lr{3}_hid{4}_{5}.txt".format(method, name, SEQ_LENGTH, LR_start, N_HIDDEN, method), "a") as myfile:
myfile.write("{0} {1:.3f} {2:.3f} {3:.3f} {4:.3f}\n".format(epoch, train_loss, val_loss,
test_loss, epoch_duration))
# learning rate update scheme
if epoch>10:
LR *= LR_decay
def main(method,LR_start,Binarize_weight_only, SEQ_LENGTH):
lasagne.random.set_rng(np.random.RandomState(1))
name = "linux"
print("dataset = "+str(name))
print("Binarize_weight_only="+str(Binarize_weight_only))
print("Method = "+str(method))
# Sequence Length
SEQ_LENGTH = SEQ_LENGTH
# SEQ_LENGTH = 100 #can have diffvalues 50, 100, 200
print("SEQ_LENGTH = "+str(SEQ_LENGTH))
# Number of units in the two hidden (LSTM) layers
N_HIDDEN = 512
print("N_HIDDEN = "+str(N_HIDDEN))
# All gradients above this will be clipped
GRAD_CLIP=5. #### this clip the gradients at every time step, while T.clip clips the sum of gradients as a whole
print("GRAD_CLIP ="+str(GRAD_CLIP))
# Number of epochs to train the net
num_epochs = 200
print("num_epochs = "+str(num_epochs))
# Batch Size
batch_size = 100
print("batch_size = "+str(batch_size))
print("LR_start = "+str(LR_start))
LR_decay = 0.98
print("LR_decay="+str(LR_decay))
if Binarize_weight_only =="w":
activation = lasagne.nonlinearities.tanh
else:
activation = lab.binary_tanh_unit
print("activation = "+ str(activation))
name = name+"_"+Binarize_weight_only
## load data, change data file dir
with open('data/linux_input.txt', 'r') as f:
in_text = f.read()
generation_phrase = "Copyright (C) 1992, 1998-2004 Linus Torvalds, Ingo Molnar\n *\n * This file contains the interrupt probing code and driver APIs.\n */\n\n#include"
#This snippet loads the text file and creates dictionaries to
#encode characters into a vector-space representation and vice-versa.
chars = list(set(in_text))
data_size, vocab_size = len(in_text), len(chars)
char_to_ix = { ch:i for i,ch in enumerate(chars) }
ix_to_char = { i:ch for i,ch in enumerate(chars) }
num_splits = [0.9, 0.05, 0.05]
num_splits_all = np.floor(data_size/batch_size/SEQ_LENGTH)
num_train = np.floor(num_splits_all*num_splits[0])
num_val = np.floor(num_splits_all*num_splits[1])
num_test = num_splits_all - num_train - num_val
train_X = in_text[0:(num_train*batch_size*SEQ_LENGTH+1).astype('int32')]
val_X = in_text[(num_train*batch_size*SEQ_LENGTH).astype('int32'):((num_train+num_val)*batch_size*SEQ_LENGTH+1).astype('int32')]
test_X = in_text[((num_train+num_val)*batch_size*SEQ_LENGTH).astype('int32'):(num_splits_all*batch_size*SEQ_LENGTH+1).astype('int32')]
## build model
print('Building the model...')
# input = T.tensor3('inputs')
target = T.imatrix('target')
LR = T.scalar('LR', dtype=theano.config.floatX)
# (batch size, SEQ_LENGTH, num_features)
l_in = lasagne.layers.InputLayer(shape=(None, None, vocab_size))
l_forward_2 = lab.LSTMLayer(
l_in,
num_units=N_HIDDEN,
grad_clipping=GRAD_CLIP,
peepholes=False,
nonlinearity=activation, ### change this activation can change the hidden layer to binary
method=method) ### batch_size*SEQ_LENGTH*N_HIDDEN
l_shp = lasagne.layers.ReshapeLayer(l_forward_2, (-1, N_HIDDEN)) ## (batch_size*SEQ_LENGTH, N_HIDDEN)
l_out = lasagne.layers.DenseLayer(l_shp, num_units=vocab_size, W = lasagne.init.Normal(), nonlinearity=lasagne.nonlinearities.softmax)
batchsize, seqlen, _ = l_in.input_var.shape
l_shp1 = lasagne.layers.ReshapeLayer(l_out, (batchsize, seqlen, vocab_size))
l_out1 = lasagne.layers.SliceLayer(l_shp1, -1, 1)
train_output = lasagne.layers.get_output(l_out, deterministic=False)
loss = T.nnet.categorical_crossentropy(train_output,target.flatten()).mean()
if method!= "FPN":
# W updates
W = lasagne.layers.get_all_params(l_out, binary=True)
W_grads = lab.compute_grads(loss,l_out)
updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR, epsilon = 1e-8) ### can choose different methods to update
updates = lab.clipping_scaling(updates,l_out)
# other parameters updates
params = lasagne.layers.get_all_params(l_out, trainable=True, binary=False)
updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR, epsilon = 1e-8).items())
## update 2 momentum
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(l_out, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
# updates3[acc_tag_temp]=updates.keys()[idx]
updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2)
idx = idx+1
updates = OrderedDict(updates.items() + updates3.items())
else:
params_other = lasagne.layers.get_all_params(l_out, trainable=True)
W_grads = [theano.grad(loss, wrt=l_forward_2.W_in_to_ingate), theano.grad(loss, wrt=l_forward_2.W_hid_to_ingate),
theano.grad(loss, wrt=l_forward_2.W_in_to_fotgetgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_forgetgate),
theano.grad(loss, wrt=l_forward_2.W_in_to_cell),theano.grad(loss, wrt=l_forward_2.W_hid_to_cell),
theano.grad(loss, wrt=l_forward_2.W_in_to_outgate),theano.grad(loss, wrt=l_forward_2.W_hid_to_outgate)]
updates = optimizer.adam(loss_or_grads=loss, params=params_other, learning_rate=LR)
test_output = lasagne.layers.get_output(l_out, deterministic=True)
test_loss = T.nnet.categorical_crossentropy(test_output,target.flatten()).mean()
train_fn = theano.function([l_in.input_var, target, LR], [loss, W_grads[5]], updates=updates, allow_input_downcast=True)
val_fn = theano.function([l_in.input_var, target], test_loss, allow_input_downcast=True)
probs = theano.function([l_in.input_var],lasagne.layers.get_output(l_out1), allow_input_downcast=True)
print('Training...')
lab.train(
name, method,
train_fn,val_fn,
batch_size,
SEQ_LENGTH,
N_HIDDEN,
LR_start,LR_decay,
num_epochs,
train_X,
val_X,
test_X)
def build_model(self, lr=0.001, dropout=None):
def concatenate(tensor_list, axis=0):
concat_size = sum(tt.shape[axis] for tt in tensor_list)
output_shape = ()
for k in range(axis):
output_shape += (tensor_list[0].shape[k],)
output_shape += (concat_size,)
for k in range(axis + 1, tensor_list[0].ndim):
output_shape += (tensor_list[0].shape[k],)
out = T.zeros(output_shape)
offset = 0
for tt in tensor_list:
indices = ()
for k in range(axis):
indices += (slice(None),)
indices += (slice(offset, offset + tt.shape[axis]),)
for k in range(axis + 1, tensor_list[0].ndim):
indices += (slice(None),)
out = T.set_subtensor(out[indices], tt)
offset += tt.shape[axis]
return out
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype = 'int32')
x_mask = T.matrix('x_mask', dtype='float32')
y = T.matrix('y', dtype = 'int32')
img = T.matrix('img', dtype = 'float32')
n_timesteps = x.shape[0]
n_samples = x.shape[1]
init_state = T.dot(img, self.W_img_emb) + self.b_img_emb
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
xr = x[::-1]
xr_mask = x_mask[::-1]
embr = self.W_emb[xr.flatten()]
embr = embr.reshape([n_timesteps, n_samples, self.dim_word])
proj = self.gru_layer(emb, init_state, mask=x_mask)
projr = self.gru_layer(embr, init_state, mask=xr_mask)
proj = concatenate([proj, projr[::-1]], axis=proj.ndim-1)
# hidden 들의 평균
proj = (proj * x_mask[:, :, None]).sum(axis=0)
proj = proj / x_mask.sum(axis=0)[:, None] # sample * dim
# 마지막 hidden
#proj = proj[-1] # sample * dim
if dropout is not None :
proj = dropout_layer(proj, use_noise, trng, dropout)
output = T.dot(proj, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, x_mask, img, y, cost, updates, prediction
def main(method, LR_start, Binarize_weight_only):
name = "svhn"
print("dataset = " + str(name))
print("Binarize_weight_only=" + str(Binarize_weight_only))
print("Method = " + str(method))
# alpha is the exponential moving average factor
alpha = .1
print("alpha = " + str(alpha))
epsilon = 1e-4
print("epsilon = " + str(epsilon))
# Training parameters
batch_size = 50
print("batch_size = " + str(batch_size))
num_epochs = 50
print("num_epochs = " + str(num_epochs))
print("LR_start = " + str(LR_start))
LR_decay = 0.1
print("LR_decay=" + str(LR_decay))
# BTW, LR decay might good for the BN moving average...
if Binarize_weight_only == "w":
activation = lasagne.nonlinearities.rectify
else:
activation = lab.binary_tanh_unit
print("activation = " + str(activation))
## number of filters in the first convolutional layer
K = 64
print("K=" + str(K))
print('Building the CNN...')
# Prepare Theano variables for inputs and targets
input = T.tensor4('inputs')
target = T.matrix('targets')
LR = T.scalar('LR', dtype=theano.config.floatX)
l_in = lasagne.layers.InputLayer(shape=(None, 3, 32, 32), input_var=input)
# 128C3-128C3-P2
l_cnn1 = lab.Conv2DLayer(l_in,
num_filters=K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_bn1 = batch_norm.BatchNormLayer(l_cnn1, epsilon=epsilon, alpha=alpha)
l_nl1 = lasagne.layers.NonlinearityLayer(l_bn1, nonlinearity=activation)
l_cnn2 = lab.Conv2DLayer(l_nl1,
num_filters=K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_mp1 = lasagne.layers.MaxPool2DLayer(l_cnn2, pool_size=(2, 2))
l_bn2 = batch_norm.BatchNormLayer(l_mp1, epsilon=epsilon, alpha=alpha)
l_nl2 = lasagne.layers.NonlinearityLayer(l_bn2, nonlinearity=activation)
# 256C3-256C3-P2
l_cnn3 = lab.Conv2DLayer(l_nl2,
num_filters=2 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_bn3 = batch_norm.BatchNormLayer(l_cnn3, epsilon=epsilon, alpha=alpha)
l_nl3 = lasagne.layers.NonlinearityLayer(l_bn3, nonlinearity=activation)
l_cnn4 = lab.Conv2DLayer(l_nl3,
num_filters=2 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_mp2 = lasagne.layers.MaxPool2DLayer(l_cnn4, pool_size=(2, 2))
l_bn4 = batch_norm.BatchNormLayer(l_mp2, epsilon=epsilon, alpha=alpha)
l_nl4 = lasagne.layers.NonlinearityLayer(l_bn4, nonlinearity=activation)
# 512C3-512C3-P2
l_cnn5 = lab.Conv2DLayer(l_nl4,
num_filters=4 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_bn5 = batch_norm.BatchNormLayer(l_cnn5, epsilon=epsilon, alpha=alpha)
l_nl5 = lasagne.layers.NonlinearityLayer(l_bn5, nonlinearity=activation)
l_cnn6 = lab.Conv2DLayer(l_nl5,
num_filters=4 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_mp3 = lasagne.layers.MaxPool2DLayer(l_cnn6, pool_size=(2, 2))
l_bn6 = batch_norm.BatchNormLayer(l_mp3, epsilon=epsilon, alpha=alpha)
l_nl6 = lasagne.layers.NonlinearityLayer(l_bn6, nonlinearity=activation)
# print(cnn.output_shape)
# 1024FP-1024FP-10FP
l_dn1 = lab.DenseLayer(l_nl6,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method=method)
l_bn7 = batch_norm.BatchNormLayer(l_dn1, epsilon=epsilon, alpha=alpha)
l_nl7 = lasagne.layers.NonlinearityLayer(l_bn7, nonlinearity=activation)
l_dn2 = lab.DenseLayer(l_nl7,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method=method)
l_bn8 = batch_norm.BatchNormLayer(l_dn2, epsilon=epsilon, alpha=alpha)
l_nl8 = lasagne.layers.NonlinearityLayer(l_bn8, nonlinearity=activation)
l_dn3 = lab.DenseLayer(l_nl8,
nonlinearity=lasagne.nonlinearities.identity,
num_units=10,
method=method)
l_out = batch_norm.BatchNormLayer(l_dn3, epsilon=epsilon, alpha=alpha)
train_output = lasagne.layers.get_output(l_out, deterministic=False)
# squared hinge loss
loss = T.mean(T.sqr(T.maximum(0., 1. - target * train_output)))
if method != "FPN":
# W updates
W = lasagne.layers.get_all_params(l_out, binary=True)
W_grads = lab.compute_grads(loss, l_out)
updates = optimizer.adam(loss_or_grads=W_grads,
params=W,
learning_rate=LR)
updates = lab.clipping_scaling(updates, l_out)
# other parameters updates
params = lasagne.layers.get_all_params(l_out,
trainable=True,
binary=False)
updates = OrderedDict(updates.items() + optimizer.adam(
loss_or_grads=loss, params=params, learning_rate=LR).items())
## update 2nd moment, can get from the adam optimizer also
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(l_out, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp] = acc_tag_temp * beta2 + W_grads[
idx] * W_grads[idx] * (1 - beta2)
idx = idx + 1
updates = OrderedDict(updates.items() + updates3.items())
else:
params = lasagne.layers.get_all_params(l_out, trainable=True)
updates = optimizer.adam(loss_or_grads=loss,
params=params,
learning_rate=LR)
test_output = lasagne.layers.get_output(l_out, deterministic=True)
test_loss = T.mean(T.sqr(T.maximum(0., 1. - target * test_output)))
test_err = T.mean(T.neq(T.argmax(test_output, axis=1),
T.argmax(target, axis=1)),
dtype=theano.config.floatX)
# Compile a function performing a training step on a mini-batch (by giving the updates dictionary)
# and returning the corresponding training loss:
train_fn = theano.function([input, target, LR], loss, updates=updates)
val_fn = theano.function([input, target], [test_loss, test_err])
## load data
print('Loading SVHN dataset')
train_set = SVHN(
which_set='splitted_train',
# which_set= 'valid',
path="${SVHN_LOCAL_PATH}",
axes=['b', 'c', 0, 1])
valid_set = SVHN(which_set='valid',
path="${SVHN_LOCAL_PATH}",
axes=['b', 'c', 0, 1])
test_set = SVHN(which_set='test',
path="${SVHN_LOCAL_PATH}",
axes=['b', 'c', 0, 1])
# bc01 format
# print train_set.X.shape
train_set.X = np.reshape(train_set.X, (-1, 3, 32, 32))
valid_set.X = np.reshape(valid_set.X, (-1, 3, 32, 32))
test_set.X = np.reshape(test_set.X, (-1, 3, 32, 32))
train_set.y = np.array(train_set.y).flatten()
valid_set.y = np.array(valid_set.y).flatten()
test_set.y = np.array(test_set.y).flatten()
# Onehot the targets
train_set.y = np.float32(np.eye(10)[train_set.y])
valid_set.y = np.float32(np.eye(10)[valid_set.y])
test_set.y = np.float32(np.eye(10)[test_set.y])
# for hinge loss
train_set.y = 2 * train_set.y - 1.
valid_set.y = 2 * valid_set.y - 1.
test_set.y = 2 * test_set.y - 1.
print('Training...')
# ipdb.set_trace()
lab.train(name, method, train_fn, val_fn, batch_size, LR_start, LR_decay,
num_epochs, train_set.X, train_set.y, valid_set.X, valid_set.y,
test_set.X, test_set.y)
def main(method, LR_start):
name = "svhn"
print("dataset = " + str(name))
print("Method = " + str(method))
# alpha is the exponential moving average factor
alpha = .1
print("alpha = " + str(alpha))
epsilon = 1e-4
print("epsilon = " + str(epsilon))
# Training parameters
batch_size = 50
print("batch_size = " + str(batch_size))
num_epochs = 50
print("num_epochs = " + str(num_epochs))
print("LR_start = " + str(LR_start))
LR_decay = 0.1
print("LR_decay=" + str(LR_decay))
# BTW, LR decay might good for the BN moving average...
activation = lasagne.nonlinearities.rectify
# number of filters in the first convolutional layer
K = 64
print("K=" + str(K))
print('Building the CNN...')
# Prepare Theano variables for inputs and targets
input = T.tensor4('inputs')
target = T.matrix('targets')
LR = T.scalar('LR', dtype=theano.config.floatX)
l_in = lasagne.layers.InputLayer(shape=(None, 3, 32, 32), input_var=input)
# 128C3-128C3-P2
l_cnn1 = laq.Conv2DLayer(l_in,
num_filters=K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_bn1 = batch_norm.BatchNormLayer(l_cnn1, epsilon=epsilon, alpha=alpha)
l_nl1 = lasagne.layers.NonlinearityLayer(l_bn1, nonlinearity=activation)
l_cnn2 = laq.Conv2DLayer(l_nl1,
num_filters=K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_mp1 = lasagne.layers.MaxPool2DLayer(l_cnn2, pool_size=(2, 2))
l_bn2 = batch_norm.BatchNormLayer(l_mp1, epsilon=epsilon, alpha=alpha)
l_nl2 = lasagne.layers.NonlinearityLayer(l_bn2, nonlinearity=activation)
# 256C3-256C3-P2
l_cnn3 = laq.Conv2DLayer(l_nl2,
num_filters=2 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_bn3 = batch_norm.BatchNormLayer(l_cnn3, epsilon=epsilon, alpha=alpha)
l_nl3 = lasagne.layers.NonlinearityLayer(l_bn3, nonlinearity=activation)
l_cnn4 = laq.Conv2DLayer(l_nl3,
num_filters=2 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_mp2 = lasagne.layers.MaxPool2DLayer(l_cnn4, pool_size=(2, 2))
l_bn4 = batch_norm.BatchNormLayer(l_mp2, epsilon=epsilon, alpha=alpha)
l_nl4 = lasagne.layers.NonlinearityLayer(l_bn4, nonlinearity=activation)
# 512C3-512C3-P2
l_cnn5 = laq.Conv2DLayer(l_nl4,
num_filters=4 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_bn5 = batch_norm.BatchNormLayer(l_cnn5, epsilon=epsilon, alpha=alpha)
l_nl5 = lasagne.layers.NonlinearityLayer(l_bn5, nonlinearity=activation)
l_cnn6 = laq.Conv2DLayer(l_nl5,
num_filters=4 * K,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method=method)
l_mp3 = lasagne.layers.MaxPool2DLayer(l_cnn6, pool_size=(2, 2))
l_bn6 = batch_norm.BatchNormLayer(l_mp3, epsilon=epsilon, alpha=alpha)
l_nl6 = lasagne.layers.NonlinearityLayer(l_bn6, nonlinearity=activation)
# print(cnn.output_shape)
# 1024FP-1024FP-10FP
l_dn1 = laq.DenseLayer(l_nl6,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method=method)
l_bn7 = batch_norm.BatchNormLayer(l_dn1, epsilon=epsilon, alpha=alpha)
l_nl7 = lasagne.layers.NonlinearityLayer(l_bn7, nonlinearity=activation)
l_dn2 = laq.DenseLayer(l_nl7,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method=method)
l_bn8 = batch_norm.BatchNormLayer(l_dn2, epsilon=epsilon, alpha=alpha)
l_nl8 = lasagne.layers.NonlinearityLayer(l_bn8, nonlinearity=activation)
l_dn3 = laq.DenseLayer(l_nl8,
nonlinearity=lasagne.nonlinearities.identity,
num_units=10,
method=method)
l_out = batch_norm.BatchNormLayer(l_dn3, epsilon=epsilon, alpha=alpha)
train_output = lasagne.layers.get_output(l_out, deterministic=False)
# squared hinge loss
loss = T.mean(T.sqr(T.maximum(0., 1. - target * train_output)))
if method != "FPN":
# W updates
W = lasagne.layers.get_all_params(l_out, quantized=True)
W_grads = laq.compute_grads(loss, l_out)
updates = optimizer.adam(loss_or_grads=W_grads,
params=W,
learning_rate=LR)
updates = laq.clipping_scaling(updates, l_out)
# other parameters updates
params = lasagne.layers.get_all_params(l_out,
trainable=True,
quantized=False)
updates = OrderedDict(updates.items() + optimizer.adam(
loss_or_grads=loss, params=params, learning_rate=LR).items())
## update 2nd moment, can get from the adam optimizer also
ternary_weights = laq.get_quantized_weights(loss, l_out)
updates2 = OrderedDict()
idx = 0
tt_tag = lasagne.layers.get_all_params(l_out, tt=True)
for tt_tag_temp in tt_tag:
updates2[tt_tag_temp] = ternary_weights[idx]
idx = idx + 1
updates = OrderedDict(updates.items() + updates2.items())
## update 2nd momentum
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(l_out, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp] = acc_tag_temp * beta2 + W_grads[
idx] * W_grads[idx] * (1 - beta2)
idx = idx + 1
updates = OrderedDict(updates.items() + updates3.items())
else:
params = lasagne.layers.get_all_params(l_out, trainable=True)
updates = optimizer.adam(loss_or_grads=loss,
params=params,
learning_rate=LR)
test_output = lasagne.layers.get_output(l_out, deterministic=True)
test_loss = T.mean(T.sqr(T.maximum(0., 1. - target * test_output)))
test_err = T.mean(T.neq(T.argmax(test_output, axis=1),
T.argmax(target, axis=1)),
dtype=theano.config.floatX)
train_fn = theano.function([input, target, LR], loss, updates=updates)
val_fn = theano.function([input, target], [test_loss, test_err])
## load data
print('Loading SVHN dataset')
train_set = SVHN(
which_set='splitted_train',
# which_set= 'valid',
path="${SVHN_LOCAL_PATH}",
axes=['b', 'c', 0, 1])
valid_set = SVHN(which_set='valid',
path="${SVHN_LOCAL_PATH}",
axes=['b', 'c', 0, 1])
test_set = SVHN(which_set='test',
path="${SVHN_LOCAL_PATH}",
axes=['b', 'c', 0, 1])
# bc01 format
# print train_set.X.shape
train_set.X = np.reshape(train_set.X, (-1, 3, 32, 32))
valid_set.X = np.reshape(valid_set.X, (-1, 3, 32, 32))
test_set.X = np.reshape(test_set.X, (-1, 3, 32, 32))
train_set.y = np.array(train_set.y).flatten()
valid_set.y = np.array(valid_set.y).flatten()
test_set.y = np.array(test_set.y).flatten()
# Onehot the targets
train_set.y = np.float32(np.eye(10)[train_set.y])
valid_set.y = np.float32(np.eye(10)[valid_set.y])
test_set.y = np.float32(np.eye(10)[test_set.y])
# for hinge loss
train_set.y = 2 * train_set.y - 1.
valid_set.y = 2 * valid_set.y - 1.
test_set.y = 2 * test_set.y - 1.
print('Training...')
X_train = train_set.X
y_train = train_set.y
X_val = valid_set.X
y_val = valid_set.y
X_test = test_set.X
y_test = test_set.y
# This function trains the model a full epoch (on the whole dataset)
def train_epoch(X, y, LR):
loss = 0
batches = len(X) / batch_size
# move shuffle here to save memory
# k = 5
# batches = int(batches/k)*k
shuffled_range = range(len(X))
np.random.shuffle(shuffled_range)
for i in range(batches):
tmp_ind = shuffled_range[i * batch_size:(i + 1) * batch_size]
newloss = train_fn(X[tmp_ind], y[tmp_ind], LR)
loss += newloss
loss /= batches
return loss
# This function tests the model a full epoch (on the whole dataset)
def val_epoch(X, y):
err = 0
loss = 0
batches = len(X) / batch_size
for i in range(batches):
new_loss, new_err = val_fn(X[i * batch_size:(i + 1) * batch_size],
y[i * batch_size:(i + 1) * batch_size])
err += new_err
loss += new_loss
err = err / batches * 100
loss /= batches
return err, loss
best_val_err = 100
best_epoch = 1
LR = LR_start
# We iterate over epochs:
for epoch in range(1, num_epochs + 1):
start_time = time.time()
train_loss = train_epoch(X_train, y_train, LR)
val_err, val_loss = val_epoch(X_val, y_val)
# test if validation error went down
if val_err <= best_val_err:
best_val_err = val_err
best_epoch = epoch
test_err, test_loss = val_epoch(X_test, y_test)
epoch_duration = time.time() - start_time
# Then we print the results for this epoch:
print("Epoch " + str(epoch) + " of " + str(num_epochs) + " took " +
str(epoch_duration) + "s")
print(" LR: " + str(LR))
print(" training loss: " + str(train_loss))
print(" validation loss: " + str(val_loss))
print(" validation error rate: " + str(val_err) + "%")
print(" best epoch: " + str(best_epoch))
print(" best validation error rate: " + str(best_val_err) + "%")
print(" test loss: " + str(test_loss))
print(" test error rate: " + str(test_err) + "%")
with open(
"{0}/{1}_lr{2}_{3}.txt".format(method, name, LR_start, method),
"a") as myfile:
myfile.write(
"{0} {1:.5f} {2:.5f} {3:.5f} {4:.5f} {5:.5f} {6:.5f} {7:.5f}\n"
.format(epoch, train_loss, val_loss, test_loss, val_err,
test_err, epoch_duration, LR))
## Learning rate update scheme
if epoch == 15 or epoch == 25:
LR *= LR_decay
def main(method,LR_start):
name = "cifar100"
print("dataset = "+str(name))
print("Method = "+str(method))
# alpha is the exponential moving average factor
alpha = .1
print("alpha = "+str(alpha))
epsilon = 1e-4
print("epsilon = "+str(epsilon))
# Training parameters
batch_size = 100
print("batch_size = "+str(batch_size))
num_epochs = 200
print("num_epochs = "+str(num_epochs))
print("LR_start = "+str(LR_start))
LR_decay = 0.5
print("LR_decay="+str(LR_decay))
activation = lasagne.nonlinearities.rectify
train_set_size = 45000
print("train_set_size = "+str(train_set_size))
print('Loading CIFAR-100 dataset...')
preprocessor = serial.load("${PYLEARN2_DATA_PATH}/cifar100/pylearn2_gcn_whitened/preprocessor.pkl")
train_set = ZCA_Dataset(
preprocessed_dataset=serial.load("${PYLEARN2_DATA_PATH}/cifar100/pylearn2_gcn_whitened/train.pkl"),
preprocessor = preprocessor,
start=0, stop = train_set_size)
valid_set = ZCA_Dataset(
preprocessed_dataset= serial.load("${PYLEARN2_DATA_PATH}/cifar100/pylearn2_gcn_whitened/train.pkl"),
preprocessor = preprocessor,
start=45000, stop = 50000)
test_set = ZCA_Dataset(
preprocessed_dataset= serial.load("${PYLEARN2_DATA_PATH}/cifar100/pylearn2_gcn_whitened/test.pkl"),
preprocessor = preprocessor)
# bc01 format
train_set.X = train_set.X.reshape(-1,3,32,32)
valid_set.X = valid_set.X.reshape(-1,3,32,32)
test_set.X = test_set.X.reshape(-1,3,32,32)
# flatten targets
train_set.y = np.int32(np.hstack(train_set.y))
valid_set.y = np.int32(np.hstack(valid_set.y))
test_set.y = np.int32(np.hstack(test_set.y))
print('Building the CNN...')
# Prepare Theano variables for inputs and targets
input = T.tensor4('inputs')
target = T.ivector('targets')
LR = T.scalar('LR', dtype=theano.config.floatX)
l_in = lasagne.layers.InputLayer(
shape=(None, 3, 32, 32),
input_var=input)
# 128C3-128C3-P2
l_cnn1 = laq.Conv2DLayer(
l_in,
num_filters=128,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_bn1 = batch_norm.BatchNormLayer(
l_cnn1,
epsilon=epsilon,
alpha=alpha)
l_nl1 = lasagne.layers.NonlinearityLayer(
l_bn1,
nonlinearity = activation)
l_cnn2 = laq.Conv2DLayer(
l_nl1,
num_filters=128,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_mp1 = lasagne.layers.MaxPool2DLayer(l_cnn2, pool_size=(2, 2))
l_bn2 = batch_norm.BatchNormLayer(
l_mp1,
epsilon=epsilon,
alpha=alpha)
l_nl2 = lasagne.layers.NonlinearityLayer(
l_bn2,
nonlinearity = activation)
# 256C3-256C3-P2
l_cnn3 = laq.Conv2DLayer(
l_nl2,
num_filters=256,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_bn3 = batch_norm.BatchNormLayer(
l_cnn3,
epsilon=epsilon,
alpha=alpha)
l_nl3 = lasagne.layers.NonlinearityLayer(
l_bn3,
nonlinearity = activation)
l_cnn4 = laq.Conv2DLayer(
l_nl3,
num_filters=256,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_mp2 = lasagne.layers.MaxPool2DLayer(l_cnn4, pool_size=(2, 2))
l_bn4 = batch_norm.BatchNormLayer(
l_mp2,
epsilon=epsilon,
alpha=alpha)
l_nl4 = lasagne.layers.NonlinearityLayer(
l_bn4,
nonlinearity = activation)
# 512C3-512C3-P2
l_cnn5 = laq.Conv2DLayer(
l_nl4,
num_filters=512,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_bn5 = batch_norm.BatchNormLayer(
l_cnn5,
epsilon=epsilon,
alpha=alpha)
l_nl5 = lasagne.layers.NonlinearityLayer(
l_bn5,
nonlinearity = activation)
l_cnn6 = laq.Conv2DLayer(
l_nl5,
num_filters=512,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_mp3 = lasagne.layers.MaxPool2DLayer(l_cnn6, pool_size=(2, 2))
l_bn6 = batch_norm.BatchNormLayer(
l_mp3,
epsilon=epsilon,
alpha=alpha)
l_nl6 = lasagne.layers.NonlinearityLayer(
l_bn6,
nonlinearity = activation)
# print(cnn.output_shape)
# 1024FP-1024FP-10FP
l_dn1 = laq.DenseLayer(
l_nl6,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method = method)
l_bn7 = batch_norm.BatchNormLayer(
l_dn1,
epsilon=epsilon,
alpha=alpha)
l_nl7 = lasagne.layers.NonlinearityLayer(
l_bn7,
nonlinearity = activation)
l_dn2 = laq.DenseLayer(
l_nl7,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method = method)
l_bn8 = batch_norm.BatchNormLayer(
l_dn2,
epsilon=epsilon,
alpha=alpha)
l_nl8 = lasagne.layers.NonlinearityLayer(
l_bn8,
nonlinearity = activation)
l_dn3 = laq.DenseLayer(
l_nl8,
nonlinearity=lasagne.nonlinearities.identity,
num_units=100,
method = method)
l_out = lasagne.layers.NonlinearityLayer(l_dn3, nonlinearity=lasagne.nonlinearities.softmax)
train_output = lasagne.layers.get_output(l_out, deterministic=False)
loss = categorical_crossentropy(train_output, target).mean()
if method!="FPN":
# W updates
W = lasagne.layers.get_all_params(l_out, quantized=True)
W_grads = laq.compute_grads(loss,l_out)
updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR)
updates = laq.clipping_scaling(updates,l_out)
# other parameters updates
params = lasagne.layers.get_all_params(l_out, trainable=True, quantized=False)
updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR).items())
## update 2nd moment, can get from the adam optimizer also
ternary_weights = laq.get_quantized_weights(loss, l_out)
updates2 = OrderedDict()
idx = 0
tt_tag = lasagne.layers.get_all_params(l_out, tt=True)
for tt_tag_temp in tt_tag:
updates2[tt_tag_temp]= ternary_weights[idx]
idx = idx+1
updates = OrderedDict(updates.items() + updates2.items())
## update 2nd momentum
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(l_out, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2)
idx = idx+1
updates = OrderedDict(updates.items() + updates3.items())
else:
params = lasagne.layers.get_all_params(l_out, trainable=True)
updates = optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR)
test_output = lasagne.layers.get_output(l_out, deterministic=True)
test_loss = categorical_crossentropy(test_output, target).mean()
test_err = T.mean(T.neq(T.argmax(test_output, axis=1), target),dtype=theano.config.floatX)
train_fn = theano.function([input, target, LR], loss, updates=updates)
val_fn = theano.function([input, target], [test_loss, test_err])
print('Training...')
X_train = train_set.X
y_train = train_set.y
X_val = valid_set.X
y_val = valid_set.y
X_test = test_set.X
y_test = test_set.y
# This function trains the model a full epoch (on the whole dataset)
def train_epoch(X,y,LR):
loss = 0
batches = len(X)/batch_size
shuffled_range = range(len(X))
np.random.shuffle(shuffled_range)
for i in range(batches):
tmp_ind = shuffled_range[i*batch_size:(i+1)*batch_size]
newloss = train_fn(X[tmp_ind],y[tmp_ind],LR)
loss +=newloss
loss/=batches
return loss
# This function tests the model a full epoch (on the whole dataset)
def val_epoch(X,y):
err = 0
loss = 0
batches = len(X)/batch_size
for i in range(batches):
new_loss, new_err = val_fn(X[i*batch_size:(i+1)*batch_size], y[i*batch_size:(i+1)*batch_size])
err += new_err
loss += new_loss
err = err / batches * 100
loss /= batches
return err, loss
best_val_err = 100
best_epoch = 1
LR = LR_start
# We iterate over epochs:
for epoch in range(1, num_epochs+1):
start_time = time.time()
train_loss = train_epoch(X_train,y_train,LR)
val_err, val_loss = val_epoch(X_val,y_val)
# test if validation error went down
if val_err <= best_val_err:
best_val_err = val_err
best_epoch = epoch
test_err, test_loss = val_epoch(X_test,y_test)
epoch_duration = time.time() - start_time
# Then we print the results for this epoch:
print("Epoch "+str(epoch)+" of "+str(num_epochs)+" took "+str(epoch_duration)+"s")
print(" LR: "+str(LR))
print(" training loss: "+str(train_loss))
print(" validation loss: "+str(val_loss))
print(" validation error rate: "+str(val_err)+"%")
print(" best epoch: "+str(best_epoch))
print(" best validation error rate: "+str(best_val_err)+"%")
print(" test loss: "+str(test_loss))
print(" test error rate: "+str(test_err)+"%")
with open("{0}/{1}_lr{2}_{3}.txt".format(method, name, LR_start, method), "a") as myfile:
myfile.write("{0} {1:.5f} {2:.5f} {3:.5f} {4:.5f} {5:.5f} {6:.5f} {7:.5f}\n".format(epoch,
train_loss, val_loss, test_loss, val_err, test_err, epoch_duration, LR))
if epoch % 15 ==0:
LR*=LR_decay
def __init__(self,
x,
y,
l,
window,
opt,
lr,
init_emb,
dim_emb,
dim_hidden,
n_vocab,
L2_reg,
unit,
sim='cos',
n_layers=1,
activation=tanh):
self.tr_inputs = [x, y, l]
self.pr_inputs = [x, y, l]
self.x = x # 1D: batch_size * l * 2, 2D: window; elem=word_id
self.y = y # 1D: batch_size; elem=label
self.l = l # scalar: elem=sentence length
batch_size = y.shape[0]
n_cands = x.shape[0] / batch_size / l
self.pad = build_shared_zeros((1, dim_emb))
if init_emb is None:
self.emb = theano.shared(sample_weights(n_vocab - 1, dim_emb))
else:
self.emb = theano.shared(init_emb)
self.E = T.concatenate([self.pad, self.emb], 0)
self.W_out = theano.shared(sample_weights(dim_hidden, dim_hidden))
self.params = [self.emb, self.W_out]
""" Input Layer """
e = self.E[x] # e: 1D: batch_size * l * 2, 2D: window, 3D: dim_emb
x_in = e.reshape((batch_size * n_cands, l, -1))
""" Intermediate Layer """
# h: 1D: n_batch * n_cands, 2D: dim_emb
h, params = cnn.layers(x_in, window, dim_emb, dim_hidden, n_layers,
activation)
self.params.extend(params)
""" Output Layer """
h = h.reshape((batch_size, n_cands, -1))
h_1 = h[T.arange(batch_size), 0]
h_2 = h[T.arange(batch_size), 1:]
if sim == 'cos':
y_score = cosign_similarity(h_1, h_2)
else:
y_score = T.batched_dot(T.dot(h_1, self.W_out),
h_2.dimshuffle(0, 2, 1))
y_score_hat = T.max(y_score, 1)
""" Objective Function """
self.nll = max_margin_loss(y_score_hat, y_score[T.arange(batch_size),
y])
self.L2_sqr = regularization(self.params)
self.cost = self.nll + L2_reg * self.L2_sqr / 2.
""" Optimization """
if opt == 'adagrad':
self.update = ada_grad(cost=self.cost, params=self.params, lr=lr)
elif opt == 'ada_delta':
self.update = ada_delta(cost=self.cost, params=self.params)
elif opt == 'adam':
self.update = adam(cost=self.cost, params=self.params, lr=lr)
else:
self.update = sgd(cost=self.cost, params=self.params, lr=lr)
""" Predicts """
y_hat = T.argmax(y_score, 1)
""" Check Accuracies """
self.correct = T.eq(y_hat, y)
def main(method,LR_start):
# BN parameters
name = "mnist"
print("dataset = "+str(name))
print("Method = "+str(method))
# alpha is the exponential moving average factor
alpha = .1
print("alpha = "+str(alpha))
epsilon = 1e-4
print("epsilon = "+str(epsilon))
batch_size = 100
print("batch_size = "+str(batch_size))
num_epochs = 50
print("num_epochs = "+str(num_epochs))
# network structure
num_units = 2048
print("num_units = "+str(num_units))
n_hidden_layers = 3
print("n_hidden_layers = "+str(n_hidden_layers))
print("LR_start = "+str(LR_start))
LR_decay = 0.1
print("LR_decay="+str(LR_decay))
activation = lasagne.nonlinearities.rectify
print('Loading MNIST dataset...')
train_set = MNIST(which_set= 'train', start=0, stop = 50000, center = True)
valid_set = MNIST(which_set= 'train', start=50000, stop = 60000, center = True)
test_set = MNIST(which_set= 'test', center = True)
# bc01 format
train_set.X = train_set.X.reshape(-1, 1, 28, 28)
valid_set.X = valid_set.X.reshape(-1, 1, 28, 28)
test_set.X = test_set.X.reshape(-1, 1, 28, 28)
# flatten targets
train_set.y = np.hstack(train_set.y)
valid_set.y = np.hstack(valid_set.y)
test_set.y = np.hstack(test_set.y)
# Onehot the targets
train_set.y = np.float32(np.eye(10)[train_set.y])
valid_set.y = np.float32(np.eye(10)[valid_set.y])
test_set.y = np.float32(np.eye(10)[test_set.y])
# for hinge loss
train_set.y = 2* train_set.y - 1.
valid_set.y = 2* valid_set.y - 1.
test_set.y = 2* test_set.y - 1.
print('Building the MLP...')
# Prepare Theano variables for inputs and targets
input = T.tensor4('inputs')
target = T.matrix('targets')
LR = T.scalar('LR', dtype=theano.config.floatX)
mlp = lasagne.layers.InputLayer(
shape=(None, 1, 28, 28),
input_var=input)
for k in range(n_hidden_layers):
mlp = laq.DenseLayer(
mlp,
nonlinearity=lasagne.nonlinearities.identity,
num_units=num_units,
method = method)
mlp = batch_norm.BatchNormLayer(
mlp,
epsilon=epsilon,
alpha=alpha)
mlp = lasagne.layers.NonlinearityLayer(
mlp,
nonlinearity = activation)
mlp = laq.DenseLayer(
mlp,
nonlinearity=lasagne.nonlinearities.identity,
num_units=10,
method = method)
mlp = batch_norm.BatchNormLayer(
mlp,
epsilon=epsilon,
alpha=alpha)
train_output = lasagne.layers.get_output(mlp, deterministic=False)
# squared hinge loss
loss = T.mean(T.sqr(T.maximum(0.,1.-target*train_output)))
if method!="FPN":
# W updates
W = lasagne.layers.get_all_params(mlp, quantized=True)
W_grads = laq.compute_grads(loss,mlp)
updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR)
updates = laq.clipping_scaling(updates,mlp)
# other parameters updates
params = lasagne.layers.get_all_params(mlp, trainable=True, quantized=False)
updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params,
learning_rate=LR, epsilon = 1e-8).items())
## update the ternary matrix
ternary_weights = laq.get_quantized_weights(loss, mlp)
updates2 = OrderedDict()
idx = 0
tt_tag = lasagne.layers.get_all_params(mlp, tt=True)
for tt_tag_temp in tt_tag:
updates2[tt_tag_temp]= ternary_weights[idx]
idx = idx+1
updates = OrderedDict(updates.items() + updates2.items())
## update 2nd momentum
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(mlp, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2)
idx = idx+1
updates = OrderedDict(updates.items() + updates3.items())
else:
params = lasagne.layers.get_all_params(mlp, trainable=True)
updates = optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR)
test_output = lasagne.layers.get_output(mlp, deterministic=True)
test_loss = T.mean(T.sqr(T.maximum(0.,1.-target*test_output)))
test_err = T.mean(T.neq(T.argmax(test_output, axis=1), T.argmax(target, axis=1)),dtype=theano.config.floatX)
train_fn = theano.function([input, target, LR], loss, updates=updates)
val_fn = theano.function([input, target], [test_loss, test_err])
print('Training...')
X_train = train_set.X
y_train = train_set.y
X_val = valid_set.X
y_val = valid_set.y
X_test = test_set.X
y_test = test_set.y
# This function trains the model a full epoch (on the whole dataset)
def train_epoch(X,y,LR):
loss = 0
batches = len(X)/batch_size
shuffled_range = range(len(X))
np.random.shuffle(shuffled_range)
for i in range(batches):
tmp_ind = shuffled_range[i*batch_size:(i+1)*batch_size]
newloss = train_fn(X[tmp_ind],y[tmp_ind],LR)
loss +=newloss
loss/=batches
return loss
# This function tests the model a full epoch (on the whole dataset)
def val_epoch(X,y):
err = 0
loss = 0
batches = len(X)/batch_size
for i in range(batches):
new_loss, new_err = val_fn(X[i*batch_size:(i+1)*batch_size], y[i*batch_size:(i+1)*batch_size])
err += new_err
loss += new_loss
err = err / batches * 100
loss /= batches
return err, loss
best_val_err = 100
best_epoch = 1
LR = LR_start
# We iterate over epochs:
for epoch in range(1, num_epochs+1):
start_time = time.time()
train_loss = train_epoch(X_train,y_train,LR)
val_err, val_loss = val_epoch(X_val,y_val)
# test if validation error went down
if val_err <= best_val_err:
best_val_err = val_err
best_epoch = epoch
test_err, test_loss = val_epoch(X_test,y_test)
all_params = lasagne.layers.get_all_params(mlp)
np.savez('{0}/{1}_lr{2}_{3}.npz'.format(method, name, LR_start, method), *all_params)
epoch_duration = time.time() - start_time
# Then we print the results for this epoch:
print("Epoch "+str(epoch)+" of "+str(num_epochs)+" took "+str(epoch_duration)+"s")
print(" LR: "+str(LR))
print(" training loss: "+str(train_loss))
print(" validation loss: "+str(val_loss))
print(" validation error rate: "+str(val_err)+"%")
print(" best epoch: "+str(best_epoch))
print(" best validation error rate: "+str(best_val_err)+"%")
print(" test loss: "+str(test_loss))
print(" test error rate: "+str(test_err)+"%")
with open("{0}/{1}_lr{2}_{3}.txt".format(method,name, LR_start, method), "a") as myfile:
myfile.write("{0} {1:.5f} {2:.5f} {3:.5f} {4:.5f} {5:.5f} {6:.5f} {7:.5f}\n".format(epoch,
train_loss, val_loss, test_loss, val_err, test_err, epoch_duration, LR))
# Learning rate update scheme
if epoch == 15 or epoch==25:
LR*=LR_decay
def main(method,LR_start,Binarize_weight_only):
name = "cifar"
print("dataset = "+str(name))
print("Binarize_weight_only="+str(Binarize_weight_only))
print("Method = "+str(method))
# alpha is the exponential moving average factor
alpha = .1
print("alpha = "+str(alpha))
epsilon = 1e-4
print("epsilon = "+str(epsilon))
# Training parameters
batch_size = 50
print("batch_size = "+str(batch_size))
num_epochs = 200
print("num_epochs = "+str(num_epochs))
print("LR_start = "+str(LR_start))
LR_decay = 0.5
print("LR_decay="+str(LR_decay))
if Binarize_weight_only =="w":
activation = lasagne.nonlinearities.rectify
else:
activation = lab.binary_tanh_unit
print("activation = "+ str(activation))
train_set_size = 45000
print("train_set_size = "+str(train_set_size))
print('Loading CIFAR-10 dataset...')
preprocessor = serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/preprocessor.pkl")
train_set = ZCA_Dataset(
preprocessed_dataset=serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/train.pkl"),
preprocessor = preprocessor,
start=0, stop = train_set_size)
valid_set = ZCA_Dataset(
preprocessed_dataset= serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/train.pkl"),
preprocessor = preprocessor,
start=45000, stop = 50000)
test_set = ZCA_Dataset(
preprocessed_dataset= serial.load("${PYLEARN2_DATA_PATH}/cifar10/pylearn2_gcn_whitened/test.pkl"),
preprocessor = preprocessor)
# bc01 format
train_set.X = train_set.X.reshape(-1,3,32,32)
valid_set.X = valid_set.X.reshape(-1,3,32,32)
test_set.X = test_set.X.reshape(-1,3,32,32)
# flatten targets
train_set.y = np.hstack(train_set.y)
valid_set.y = np.hstack(valid_set.y)
test_set.y = np.hstack(test_set.y)
# Onehot the targets
train_set.y = np.float32(np.eye(10)[train_set.y])
valid_set.y = np.float32(np.eye(10)[valid_set.y])
test_set.y = np.float32(np.eye(10)[test_set.y])
# for hinge loss
train_set.y = 2* train_set.y - 1.
valid_set.y = 2* valid_set.y - 1.
test_set.y = 2* test_set.y - 1.
print('Building the CNN...')
# Prepare Theano variables for inputs and targets
input = T.tensor4('inputs')
target = T.matrix('targets')
LR = T.scalar('LR', dtype=theano.config.floatX)
l_in = lasagne.layers.InputLayer(
shape=(None, 3, 32, 32),
input_var=input)
# 128C3-128C3-P2
l_cnn1 = lab.Conv2DLayer(
l_in,
num_filters=128,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_bn1 = batch_norm.BatchNormLayer(
l_cnn1,
epsilon=epsilon,
alpha=alpha)
l_nl1 = lasagne.layers.NonlinearityLayer(
l_bn1,
nonlinearity = activation)
l_cnn2 = lab.Conv2DLayer(
l_nl1,
num_filters=128,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_mp1 = lasagne.layers.MaxPool2DLayer(l_cnn2, pool_size=(2, 2))
l_bn2 = batch_norm.BatchNormLayer(
l_mp1,
epsilon=epsilon,
alpha=alpha)
l_nl2 = lasagne.layers.NonlinearityLayer(
l_bn2,
nonlinearity = activation)
# 256C3-256C3-P2
l_cnn3 = lab.Conv2DLayer(
l_nl2,
num_filters=256,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_bn3 = batch_norm.BatchNormLayer(
l_cnn3,
epsilon=epsilon,
alpha=alpha)
l_nl3 = lasagne.layers.NonlinearityLayer(
l_bn3,
nonlinearity = activation)
l_cnn4 = lab.Conv2DLayer(
l_nl3,
num_filters=256,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_mp2 = lasagne.layers.MaxPool2DLayer(l_cnn4, pool_size=(2, 2))
l_bn4 = batch_norm.BatchNormLayer(
l_mp2,
epsilon=epsilon,
alpha=alpha)
l_nl4 = lasagne.layers.NonlinearityLayer(
l_bn4,
nonlinearity = activation)
# 512C3-512C3-P2
l_cnn5 = lab.Conv2DLayer(
l_nl4,
num_filters=512,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_bn5 = batch_norm.BatchNormLayer(
l_cnn5,
epsilon=epsilon,
alpha=alpha)
l_nl5 = lasagne.layers.NonlinearityLayer(
l_bn5,
nonlinearity = activation)
l_cnn6 = lab.Conv2DLayer(
l_nl5,
num_filters=512,
filter_size=(3, 3),
pad=1,
nonlinearity=lasagne.nonlinearities.identity,
method = method)
l_mp3 = lasagne.layers.MaxPool2DLayer(l_cnn6, pool_size=(2, 2))
l_bn6 = batch_norm.BatchNormLayer(
l_mp3,
epsilon=epsilon,
alpha=alpha)
l_nl6 = lasagne.layers.NonlinearityLayer(
l_bn6,
nonlinearity = activation)
# print(cnn.output_shape)
# 1024FP-1024FP-10FP
l_dn1 = lab.DenseLayer(
l_nl6,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method = method)
l_bn7 = batch_norm.BatchNormLayer(
l_dn1,
epsilon=epsilon,
alpha=alpha)
l_nl7 = lasagne.layers.NonlinearityLayer(
l_bn7,
nonlinearity = activation)
l_dn2 = lab.DenseLayer(
l_nl7,
nonlinearity=lasagne.nonlinearities.identity,
num_units=1024,
method = method)
l_bn8 = batch_norm.BatchNormLayer(
l_dn2,
epsilon=epsilon,
alpha=alpha)
l_nl8 = lasagne.layers.NonlinearityLayer(
l_bn8,
nonlinearity = activation)
l_dn3 = lab.DenseLayer(
l_nl8,
nonlinearity=lasagne.nonlinearities.identity,
num_units=10,
method = method)
l_out = batch_norm.BatchNormLayer(
l_dn3,
epsilon=epsilon,
alpha=alpha)
train_output = lasagne.layers.get_output(l_out, deterministic=False)
# squared hinge loss
loss = T.mean(T.sqr(T.maximum(0.,1.-target*train_output)))
if method!="FPN":
# W updates
W = lasagne.layers.get_all_params(l_out, binary=True)
W_grads = lab.compute_grads(loss,l_out)
updates = optimizer.adam(loss_or_grads=W_grads, params=W, learning_rate=LR)
updates = lab.clipping_scaling(updates,l_out)
# other parameters updates
params = lasagne.layers.get_all_params(l_out, trainable=True, binary=False)
updates = OrderedDict(updates.items() + optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR).items())
## update 2nd moment, can get from the adam optimizer also
updates3 = OrderedDict()
acc_tag = lasagne.layers.get_all_params(l_out, acc=True)
idx = 0
beta2 = 0.999
for acc_tag_temp in acc_tag:
updates3[acc_tag_temp]= acc_tag_temp*beta2 + W_grads[idx]*W_grads[idx]*(1-beta2)
idx = idx+1
updates = OrderedDict(updates.items() + updates3.items())
else:
params = lasagne.layers.get_all_params(l_out, trainable=True)
updates = optimizer.adam(loss_or_grads=loss, params=params, learning_rate=LR)
test_output = lasagne.layers.get_output(l_out, deterministic=True)
test_loss = T.mean(T.sqr(T.maximum(0.,1.-target*test_output)))
test_err = T.mean(T.neq(T.argmax(test_output, axis=1), T.argmax(target, axis=1)),dtype=theano.config.floatX)
# Compile a function performing a training step on a mini-batch (by giving the updates dictionary)
# and returning the corresponding training loss:
train_fn = theano.function([input, target, LR], loss, updates=updates)
val_fn = theano.function([input, target], [test_loss, test_err])
print('Training...')
lab.train(
name, method,
train_fn,val_fn,
batch_size,
LR_start,LR_decay,
num_epochs,
train_set.X,train_set.y,
valid_set.X,valid_set.y,
test_set.X,test_set.y)
def build_model(self, lr=0.001, dropout=None):
def concatenate(tensor_list, axis=0):
concat_size = sum(tt.shape[axis] for tt in tensor_list)
output_shape = ()
for k in range(axis):
output_shape += (tensor_list[0].shape[k], )
output_shape += (concat_size, )
for k in range(axis + 1, tensor_list[0].ndim):
output_shape += (tensor_list[0].shape[k], )
out = tensor.zeros(output_shape)
offset = 0
for tt in tensor_list:
indices = ()
for k in range(axis):
indices += (slice(None), )
indices += (slice(offset, offset + tt.shape[axis]), )
for k in range(axis + 1, tensor_list[0].ndim):
indices += (slice(None), )
out = tensor.set_subtensor(out[indices], tt)
offset += tt.shape[axis]
return out
trng = RandomStreams(1234)
use_noise = theano.shared(np.float32(0.))
# description string: #words x #samples
x = T.matrix('x', dtype='int32') # step * samples
x_mask = T.matrix('x_mask', dtype='float32') # step * samples
y = T.matrix('y', dtype='int32') # sample * emb
ctx = T.tensor3('ctx', dtype='float32') # sample * annotation * dim
n_timesteps = x.shape[0]
n_samples = x.shape[1]
xr = x[::-1]
xr_mask = x_mask[::-1]
emb = self.W_emb[x.flatten()]
emb = emb.reshape([n_timesteps, n_samples, self.dim_word])
embr = self.W_emb[xr.flatten()]
embr = embr.reshape([n_timesteps, n_samples, self.dim_word])
ctx0 = ctx
ctx_mean = ctx0.mean(1)
init_state = T.dot(ctx_mean, self.W_ctx_init) + self.b_ctx_init
# proj : gru hidden 들의 리스트
proj = self.gru_layer(emb,
mask=x_mask,
context=ctx,
init_state=init_state)
proj_h = proj[0]
projr = self.gru_layer(embr,
mask=xr_mask,
context=ctx,
init_state=init_state)
projr_h = projr[0]
concat_proj_h = concatenate([proj_h, projr_h[::-1]],
axis=proj_h.ndim - 1)
# step_ctx : step * samples * (dim*2)
concat_proj_h = (concat_proj_h *
x_mask[:, :, None]).sum(0) / x_mask.sum(0)[:, None]
# step_ctx_mean : samples * (dim*2)
if dropout is not None:
concat_proj_h = dropout_layer(concat_proj_h, use_noise, trng,
dropout)
output = T.dot(concat_proj_h, self.W_pred) + self.b_pred
probs = T.nnet.softmax(output)
prediction = probs.argmax(axis=1)
## avoid NaN
epsilon = 1.0e-9
probs = T.clip(probs, epsilon, 1.0 - epsilon)
probs /= probs.sum(axis=-1, keepdims=True)
## avoid NaN
cost = T.nnet.categorical_crossentropy(probs, y)
cost = T.mean(cost)
updates = optimizer.adam(cost=cost, params=self.params, lr=lr)
return trng, use_noise, x, x_mask, ctx, y, cost, updates, prediction