def loss_func(ps1, ps2, s1, s2, po1, po2, o1, o2, mask):
ps1 = layers.concat([1 - ps1, ps1], axis=-1)
ps2 = layers.concat([1 - ps2, ps2], axis=-1)
s1 = layers.unsqueeze(s1, -1)
s2 = layers.unsqueeze(s2, -1)
s1_loss = layers.cross_entropy(ps1, s1)
s1_loss = layers.reduce_sum(s1_loss * mask) / layers.reduce_sum(mask)
s2_loss = layers.cross_entropy(ps2, s2)
s2_loss = layers.reduce_sum(s2_loss * mask) / layers.reduce_sum(mask)
po1, o1 = layers.unsqueeze(po1, -1), layers.unsqueeze(o1, -1)
po1 = layers.concat([1 - po1, po1], axis=-1)
o1_loss = layers.reduce_sum(layers.cross_entropy(po1, o1), 2)
o1_loss = layers.reduce_sum(o1_loss * mask) / layers.reduce_sum(mask)
po2, o2 = layers.unsqueeze(po2, -1), layers.unsqueeze(o2, -1)
po2 = layers.concat([1 - po2, po2], axis=-1)
o2_loss = layers.reduce_sum(layers.cross_entropy(po2, o2), 2)
o2_loss = layers.reduce_sum(o2_loss * mask) / layers.reduce_sum(mask)
loss = (s1_loss + s2_loss) + (o1_loss + o2_loss)
return loss
def rc_model(hidden_size, vocab, args):
emb_shape = [vocab.size(), vocab.embed_dim]
start_labels = layers.data(name="start_lables",
shape=[1],
dtype='float32',
lod_level=1)
end_labels = layers.data(name="end_lables",
shape=[1],
dtype='float32',
lod_level=1)
# stage 1:encode
q_id0 = get_data('q_id0', 1, args)
q_ids = get_data('q_ids', 2, args)
p_ids_name = 'p_ids'
p_ids = get_data('p_ids', 2, args)
p_embs = embedding(p_ids, emb_shape, args)
q_embs = embedding(q_ids, emb_shape, args)
drnn = layers.DynamicRNN()
with drnn.block():
p_emb = drnn.step_input(p_embs)
q_emb = drnn.step_input(q_embs)
p_enc = encoder(p_emb, 'p_enc', hidden_size, args)
q_enc = encoder(q_emb, 'q_enc', hidden_size, args)
# stage 2:match
g_i = attn_flow(q_enc, p_enc, p_ids_name, args)
# stage 3:fusion
m_i = fusion(g_i, args)
drnn.output(m_i, q_enc)
ms, q_encs = drnn()
p_vec = layers.lod_reset(x=ms, y=start_labels)
q_vec = layers.lod_reset(x=q_encs, y=q_id0)
# stage 4:decode
start_probs, end_probs = point_network_decoder(p_vec=p_vec,
q_vec=q_vec,
hidden_size=hidden_size,
args=args)
cost0 = layers.sequence_pool(
layers.cross_entropy(input=start_probs,
label=start_labels,
soft_label=True), 'sum')
cost1 = layers.sequence_pool(
layers.cross_entropy(input=end_probs,
label=end_labels,
soft_label=True), 'sum')
cost0 = layers.mean(cost0)
cost1 = layers.mean(cost1)
cost = cost0 + cost1
cost.persistable = True
feeding_list = ["q_ids", "start_lables", "end_lables", "p_ids", "q_id0"]
return cost, start_probs, end_probs, ms, feeding_list
def _collect_metrics(self, inputs, outputs):
""" Calculate loss function by using inputs and outputs. """
metrics = {}
tgt_len = layers.reduce_sum(
layers.reduce_sum(inputs["tgt_mask"], dim=1) - 1)
tgt_len.stop_gradient = True
label = inputs["tgt_token"][:, 1:]
if self.label_smooth > 0:
one_hot_label = layers.one_hot(label, self.num_token_embeddings)
smooth_label = layers.label_smooth(one_hot_label,
epsilon=self.label_smooth,
dtype=self._dtype)
nll = layers.cross_entropy(outputs["dec_pred"],
smooth_label,
soft_label=True,
ignore_index=self.padding_idx)
else:
nll = layers.cross_entropy(outputs["dec_probs"],
label,
ignore_index=self.padding_idx)
nll = layers.reduce_sum(nll, dim=1)
token_nll = layers.reduce_sum(nll) / tgt_len
nll = layers.reduce_mean(nll)
metrics["nll"] = nll
metrics["token_nll"] = token_nll
loss = nll
if self.num_latent > 0 and self.with_bow:
bow_probs = F.unsqueeze(outputs["bow_probs"], [1])
bow_probs = layers.expand(bow_probs, [1, label.shape[1], 1])
if self.label_smooth > 0:
bow = layers.cross_entropy(bow_probs,
smooth_label,
soft_label=True,
ignore_index=self.padding_idx)
else:
bow = layers.cross_entropy(bow_probs,
label,
ignore_index=self.padding_idx)
bow = layers.reduce_sum(bow, dim=1)
token_bow = layers.reduce_sum(bow) / tgt_len
bow = layers.reduce_mean(bow)
metrics["bow"] = bow
metrics["token_bow"] = token_bow
loss = loss + bow
if self.num_latent > 0 and self.use_discriminator:
dis = 0.0 - (layers.log(outputs["pos_probs"]) +
layers.log(1.0 - outputs["neg_probs"]))
dis = layers.reduce_mean(dis)
metrics["dis"] = dis
loss = loss + dis * self.dis_ratio
metrics["loss"] = loss
metrics["token_num"] = tgt_len
return metrics
def loss_function(s_arc, s_rel, arcs, rels, mask):
"""Loss function"""
arcs = nn.masked_select(arcs, mask)
rels = nn.masked_select(rels, mask)
s_arc = nn.masked_select(s_arc, mask)
s_rel = nn.masked_select(s_rel, mask)
s_rel = nn.index_sample(s_rel, layers.unsqueeze(arcs, 1))
arc_loss = layers.cross_entropy(layers.softmax(s_arc), arcs)
rel_loss = layers.cross_entropy(layers.softmax(s_rel), rels)
loss = layers.reduce_mean(arc_loss + rel_loss)
return loss
def dynamic(train_data, use_cuda=False, use_parallel_exe=False):
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
with fluid.dygraph.guard(place):
fluid.default_startup_program().random_seed = SEED
fluid.default_main_program().random_seed = SEED
dy_layer = DygraphLayer()
adam = fluid.optimizer.Adam(learning_rate=LR,
parameter_list=dy_layer.parameters())
sgd = fluid.optimizer.SGD(learning_rate=LR,
parameter_list=dy_layer.parameters())
for epoch in range(EPOCH_NUM):
image_data, label = train_data[epoch]
var_input = fluid.dygraph.to_variable(image_data)
var_label = fluid.dygraph.to_variable(label)
hidden, prediction = dy_layer(var_input)
if epoch % 2 == 0:
cross_entropy_loss = layers.cross_entropy(
prediction, var_label)
loss = layers.mean(cross_entropy_loss)
loss.backward()
adam.minimize(loss)
else:
softmax_loss = layers.softmax_with_cross_entropy(
prediction, var_label)
loss = layers.mean(softmax_loss)
loss.backward()
sgd.minimize(loss)
dy_layer.clear_gradients()
return hidden.numpy(), prediction.numpy(), loss.numpy()
def node_classify_model(word2id, num_labels, embed_dim=16):
"""Build node classify model.
Args:
word2id(dict): map word(node) to its corresponding index
num_labels: The number of labels.
embed_dim: The dimension of embedding.
"""
nodes = fl.data('nodes', shape=[None, 1], dtype='int64')
labels = fl.data('labels', shape=[None, 1], dtype='int64')
embed_nodes = fl.embedding(input=nodes,
size=[len(word2id), embed_dim],
param_attr=fluid.ParamAttr(name='content'))
embed_nodes.stop_gradient = True
probs = fl.fc(input=embed_nodes, size=num_labels, act='softmax')
predict = fl.argmax(probs, axis=-1)
loss = fl.cross_entropy(input=probs, label=labels)
loss = fl.reduce_mean(loss)
return {
'loss': loss,
'probs': probs,
'predict': predict,
'labels': labels,
}
def not_test_raw_api(self):
prog = Program()
startup_prog = Program()
with program_guard(prog, startup_prog):
image = layers.data(name='x', shape=[784], dtype='float32')
label = layers.data(name='y', shape=[1], dtype='int64')
limit = layers.fill_constant(shape=[1], dtype='int64', value=5)
cond = layers.less_than(x=label, y=limit)
true_image, false_image = split_lod_tensor(input=image, mask=cond)
true_out = layers.create_tensor(dtype='float32')
true_cond = ConditionalBlock([cond])
with true_cond.block():
hidden = layers.fc(input=true_image, size=100, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
layers.assign(input=prob, output=true_out)
false_out = layers.create_tensor(dtype='float32')
false_cond = ConditionalBlock([cond])
with false_cond.block():
hidden = layers.fc(input=false_image, size=200, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
layers.assign(input=prob, output=false_out)
prob = merge_lod_tensor(
in_true=true_out, in_false=false_out, mask=cond, x=image)
loss = layers.cross_entropy(input=prob, label=label)
avg_loss = layers.mean(loss)
optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9)
optimizer.minimize(avg_loss, startup_prog)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
batch_size=10)
place = core.CPUPlace()
exe = Executor(place)
exe.run(startup_prog)
PASS_NUM = 100
for pass_id in range(PASS_NUM):
for data in train_reader():
x_data = np.array([x[0] for x in data]).astype("float32")
y_data = np.array([x[1] for x in data]).astype("int64")
y_data = np.expand_dims(y_data, axis=1)
outs = exe.run(prog,
feed={'x': x_data,
'y': y_data},
fetch_list=[avg_loss])
print(outs[0])
if outs[0] < 1.0:
return
self.assertFalse(True)
def mlp_pretrain_forward(train_program, start_program):
with static.program_guard(train_program,
start_program), utils.unique_name.guard():
batch_size = 4
hidden_size = 1024
sequence_len = 512
input = static.data(name="input",
shape=[batch_size, sequence_len, hidden_size],
dtype='float32')
label = static.data(name="label",
shape=[batch_size, sequence_len, 1],
dtype='float32')
auto.shard_tensor(input,
dist_attr={
"process_mesh": _global_process_mesh,
"dims_mappig": [-1, -1, -1]
})
mlp = MLPLayer(hidden_size=hidden_size,
intermediate_size=4 * hidden_size,
dropout_ratio=0.1,
initializer_range=0.02)
predict = mlp(input)
cost = layers.cross_entropy(input=predict, label=label)
avg_cost = layers.mean(x=cost)
return avg_cost, train_program, start_program
def test_recognize_digits_conv(self):
program = Program()
with program_guard(program, startup_program=Program()):
images = layers.data(name='pixel',
shape=[1, 28, 28],
dtype='float32')
label = layers.data(name='label', shape=[1], dtype='int32')
conv_pool_1 = nets.simple_img_conv_pool(input=images,
filter_size=5,
num_filters=2,
pool_size=2,
pool_stride=2,
act="relu")
conv_pool_2 = nets.simple_img_conv_pool(input=conv_pool_1,
filter_size=5,
num_filters=4,
pool_size=2,
pool_stride=2,
act="relu")
predict = layers.fc(input=conv_pool_2, size=10, act="softmax")
cost = layers.cross_entropy(input=predict, label=label)
avg_cost = layers.mean(cost)
print(str(program))
def test_recognize_digits_conv(self):
program = Program()
with program_guard(program, startup_program=Program()):
images = layers.data(
name='pixel', shape=[1, 28, 28], dtype='float32')
label = layers.data(name='label', shape=[1], dtype='int32')
conv_pool_1 = nets.simple_img_conv_pool(
input=images,
filter_size=5,
num_filters=2,
pool_size=2,
pool_stride=2,
act="relu")
conv_pool_2 = nets.simple_img_conv_pool(
input=conv_pool_1,
filter_size=5,
num_filters=4,
pool_size=2,
pool_stride=2,
act="relu")
predict = layers.fc(input=conv_pool_2, size=10, act="softmax")
cost = layers.cross_entropy(input=predict, label=label)
avg_cost = layers.mean(cost)
print(str(program))
def create_loss_op(self, predict, label, epsilon=1e-7):
"""compute loss with tensor
Args:
predict: model output tensor activated by softmax
label: a non-sparse tensor
Returns:
loss: cross-entropy loss
"""
if self.loss_type == "nl" and self.model_type == "train":
one_hot_label = fluid.one_hot(label, depth=predict.shape[-1])
one_hot_label = FL.squeeze(one_hot_label, axes=[-2])
# log
neg_prob = 1 - predict
log_neg_prob = FL.log(
fluid.layers.clip(neg_prob, min=epsilon, max=1.))
ce_loss = -1 * log_neg_prob * one_hot_label
cost = FL.reduce_sum(ce_loss, dim=-1, keep_dim=True)
else: # PL or evaluation
cost = FL.cross_entropy(predict, label)
loss = FL.mean(cost)
return loss
def train_program(is_sparse):
context = encoder(is_sparse)
rnn_out = train_decoder(context, is_sparse)
label = pd.data(
name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
cost = pd.cross_entropy(input=rnn_out, label=label)
avg_cost = pd.mean(cost)
return avg_cost
def train_program(is_sparse):
context = encoder(is_sparse)
rnn_out = train_decoder(context, is_sparse)
label = pd.data(
name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
cost = pd.cross_entropy(input=rnn_out, label=label)
avg_cost = pd.mean(cost)
return avg_cost
def loss_func(logits, label, trg_sequence_length):
probs = layers.softmax(logits)
loss = layers.cross_entropy(input=probs, label=label)
trg_mask = layers.sequence_mask(trg_sequence_length,
maxlen=layers.shape(logits)[1],
dtype="float32")
avg_cost = layers.reduce_sum(loss * trg_mask) / layers.reduce_sum(trg_mask)
return avg_cost
def test_cross_entropy(self):
program = Program()
with program_guard(program):
x = layers.data(name="x", shape=[30, 10], dtype="float32")
label = layers.data(name="label", shape=[30, 1], dtype="int32")
mode = 'channel'
out = layers.cross_entropy(x, label, False, 4)
self.assertIsNotNone(out)
def test_ifelse(self):
prog = Program()
startup_prog = Program()
with program_guard(prog, startup_prog):
image = layers.data(name='x', shape=[784], dtype='float32')
label = layers.data(name='y', shape=[1], dtype='int64')
limit = layers.fill_constant_batch_size_like(input=label,
dtype='int64',
shape=[1],
value=5.0)
cond = layers.less_than(x=label, y=limit)
ie = layers.IfElse(cond)
with ie.true_block():
true_image = ie.input(image)
hidden = layers.fc(input=true_image, size=100, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
ie.output(prob)
with ie.false_block():
false_image = ie.input(image)
hidden = layers.fc(input=false_image, size=200, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
ie.output(prob)
prob = ie()
loss = layers.cross_entropy(input=prob[0], label=label)
avg_loss = layers.mean(loss)
optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9)
optimizer.minimize(avg_loss, startup_prog)
train_reader = paddle.batch(paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
batch_size=200)
place = core.CPUPlace()
exe = Executor(place)
exe.run(kwargs['startup_program'])
PASS_NUM = 100
for pass_id in range(PASS_NUM):
for data in train_reader():
x_data = np.array(map(lambda x: x[0], data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape((y_data.shape[0], 1))
outs = exe.run(kwargs['main_program'],
feed={
'x': x_data,
'y': y_data
},
fetch_list=[avg_loss])
print outs[0]
if outs[0] < 1.0:
return
self.assertFalse(True)
def _compute_loss_acc(self, pred):
loss = layers.cross_entropy(pred, label=self.label, soft_label=False)
loss = layers.reshape(loss, shape=[self.batch_size, -1])
loss = layers.reduce_mean(loss)
acc = fluid.layers.accuracy(input=pred, label=self.label)
return loss, acc
def learn(self, probs, label, weight=None, length=None):
loss = layers.cross_entropy(input=probs, label=label, soft_label=False)
max_seq_len = layers.shape(probs)[1]
mask = layers.sequence_mask(length, maxlen=max_seq_len, dtype="float32")
loss = loss * mask
loss = layers.reduce_mean(loss, dim=[0])
loss = layers.reduce_sum(loss)
optimizer = fluid.optimizer.Adam(self.lr)
optimizer.minimize(loss)
return loss
def loss_func(logits, label, trg_sequence_length):
probs = layers.softmax(logits)
# 使用交叉熵损失函数
loss = layers.cross_entropy(input=probs, label=label)
# 根据长度生成掩码，并依此剔除 padding 部分计算的损失
trg_mask = layers.sequence_mask(trg_sequence_length,
maxlen=layers.shape(logits)[1],
dtype="float32")
avg_cost = layers.reduce_sum(loss * trg_mask) / layers.reduce_sum(trg_mask)
return avg_cost
def get_losses(self, out, cls_out, mask, gt_labels):
loss_cls = L.mean(L.cross_entropy(cls_out,
gt_labels)) * self.train_cfg['w_cls']
loss_tir = 0
for feat in out[:-1]:
feat = L.squeeze(self.avgpool(feat), axes=[2, 3])
loss_tir += self.triple_loss(feat,
gt_labels) * self.train_cfg['w_tri']
loss = loss_cls + loss_tir
return dict(loss_cls=loss_cls, loss_tir=loss_tir, loss=loss)
def test_ifelse(self):
prog = Program()
startup_prog = Program()
with program_guard(prog, startup_prog):
image = layers.data(name='x', shape=[784], dtype='float32')
label = layers.data(name='y', shape=[1], dtype='int64')
limit = layers.fill_constant_batch_size_like(
input=label, dtype='int64', shape=[1], value=5.0)
cond = layers.less_than(x=label, y=limit)
ie = layers.IfElse(cond)
with ie.true_block():
true_image = ie.input(image)
hidden = layers.fc(input=true_image, size=100, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
ie.output(prob)
with ie.false_block():
false_image = ie.input(image)
hidden = layers.fc(input=false_image, size=200, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
ie.output(prob)
prob = ie()
loss = layers.cross_entropy(input=prob[0], label=label)
avg_loss = layers.mean(loss)
optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9)
optimizer.minimize(avg_loss, startup_prog)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
batch_size=200)
place = core.CPUPlace()
exe = Executor(place)
exe.run(kwargs['startup_program'])
PASS_NUM = 100
for pass_id in range(PASS_NUM):
for data in train_reader():
x_data = np.array(map(lambda x: x[0], data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = y_data.reshape((y_data.shape[0], 1))
outs = exe.run(kwargs['main_program'],
feed={'x': x_data,
'y': y_data},
fetch_list=[avg_loss])
print outs[0]
if outs[0] < 1.0:
return
self.assertFalse(True)
def main():
rnn_out = encoder_decoder()
label = layers.data(name="target_language_next_word",
shape=[1],
dtype='int64',
lod_level=1)
cost = layers.cross_entropy(input=rnn_out, label=label)
avg_cost = fluid.layers.mean(cost)
optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
optimizer.minimize(avg_cost)
# fluid.memory_optimize(fluid.default_main_program())
fluid.release_memory(fluid.default_main_program())
# fix the order of training data
train_data = paddle.batch(paddle.dataset.wmt14.train(dict_size),
batch_size=batch_size)
# train_data = paddle.batch(
# paddle.reader.shuffle(
# paddle.dataset.wmt14.train(dict_size), buf_size=1000),
# batch_size=batch_size)
place = core.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
feed_order = [
'src_word_id', 'target_language_word', 'target_language_next_word'
]
feed_list = [
fluid.default_main_program().global_block().var(var_name)
for var_name in feed_order
]
feeder = fluid.DataFeeder(feed_list, place)
batch_id = 0
for pass_id in xrange(10):
for data in train_data():
outs = exe.run(fluid.default_main_program(),
feed=feeder.feed(data),
fetch_list=[avg_cost])
avg_cost_val = np.array(outs[0])
print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
" avg_cost=" + str(avg_cost_val))
if batch_id > 2:
exit(0)
if math.isnan(float(avg_cost_val)):
sys.exit("got NaN loss, training failed.")
batch_id += 1
def learn(self, act_prob, action, reward, length=None):
"""
update policy model self.model with policy gradient algorithm
"""
self.reward = fluid.layers.py_func(
func=reward_func, x=[action, length], out=reward)
neg_log_prob = layers.cross_entropy(act_prob, action)
cost = neg_log_prob * reward
cost = (layers.reduce_sum(cost) / layers.reduce_sum(length)
) if length is not None else layers.reduce_mean(cost)
optimizer = fluid.optimizer.Adam(self.lr)
optimizer.minimize(cost)
return cost
def learn(self, obs, action, reward):
obs = fluid.dygraph.to_variable(obs)
obs = layers.cast(obs, dtype='float32')
act_prob = self.model(obs)
action = fluid.dygraph.to_variable(action)
reward = fluid.dygraph.to_variable(reward)
log_prob = layers.cross_entropy(act_prob, action)
cost = log_prob * reward
cost = layers.cast(cost, dtype='float32')
cost = layers.reduce_mean(cost)
cost.backward()
self.optimizer.minimize(cost)
self.model.clear_gradients()
return cost
def main():
rnn_out = encoder_decoder()
label = layers.data(
name="target_language_next_word", shape=[1], dtype='int64', lod_level=1)
cost = layers.cross_entropy(input=rnn_out, label=label)
avg_cost = fluid.layers.mean(cost)
optimizer = fluid.optimizer.Adagrad(learning_rate=1e-4)
optimizer.minimize(avg_cost)
# fluid.memory_optimize(fluid.default_main_program())
fluid.release_memory(fluid.default_main_program())
# fix the order of training data
train_data = paddle.batch(
paddle.dataset.wmt14.train(dict_size), batch_size=batch_size)
# train_data = paddle.batch(
# paddle.reader.shuffle(
# paddle.dataset.wmt14.train(dict_size), buf_size=1000),
# batch_size=batch_size)
place = core.CPUPlace()
exe = Executor(place)
exe.run(framework.default_startup_program())
batch_id = 0
for pass_id in xrange(10):
for data in train_data():
word_data = to_lodtensor(map(lambda x: x[0], data), place)
trg_word = to_lodtensor(map(lambda x: x[1], data), place)
trg_word_next = to_lodtensor(map(lambda x: x[2], data), place)
outs = exe.run(fluid.default_main_program(),
feed={
'src_word_id': word_data,
'target_language_word': trg_word,
'target_language_next_word': trg_word_next
},
fetch_list=[avg_cost])
avg_cost_val = np.array(outs[0])
print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
" avg_cost=" + str(avg_cost_val))
if batch_id > 2:
exit(0)
if math.isnan(float(avg_cost_val)):
sys.exit("got NaN loss, training failed.")
batch_id += 1
def train_main(use_cuda):
if use_cuda and not fluid.core.is_compiled_with_cuda():
return
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
context = encoder()
state_cell = decoder_state_cell(context)
rnn_out = decoder_train(state_cell)
label = layers.data(name="target_next_word",
shape=[1],
dtype='int64',
lod_level=1)
cost = layers.cross_entropy(input=rnn_out, label=label)
avg_cost = layers.mean(x=cost)
optimizer = fluid.optimizer.Adagrad(learning_rate=1e-3)
optimizer.minimize(avg_cost)
train_reader = paddle.batch(paddle.reader.shuffle(
paddle.dataset.wmt14.train(dict_size), buf_size=1000),
batch_size=batch_size)
feed_order = ['src_word', 'target_word', 'target_next_word']
exe = Executor(place)
def train_loop(main_program):
exe.run(framework.default_startup_program())
feed_list = [
main_program.global_block().var(var_name)
for var_name in feed_order
]
feeder = fluid.DataFeeder(feed_list, place)
for pass_id in range(1):
for batch_id, data in enumerate(train_reader()):
outs = exe.run(main_program,
feed=feeder.feed(data),
fetch_list=[avg_cost])
avg_cost_val = np.array(outs[0])
print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
" avg_cost=" + str(avg_cost_val))
if batch_id > 3:
break
train_loop(framework.default_main_program())
def test_recognize_digits_mlp(self):
program = Program()
with program_guard(program, startup_program=Program()):
# Change g_program, so the rest layers use `g_program`
images = layers.data(name='pixel', shape=[784], dtype='float32')
label = layers.data(name='label', shape=[1], dtype='int32')
hidden1 = layers.fc(input=images, size=128, act='relu')
hidden2 = layers.fc(input=hidden1, size=64, act='relu')
predict = layers.fc(input=[hidden2, hidden1],
size=10,
act='softmax',
param_attr=["sftmax.w1", "sftmax.w2"])
cost = layers.cross_entropy(input=predict, label=label)
avg_cost = layers.mean(cost)
self.assertIsNotNone(avg_cost)
print(str(program))
def test_word_embedding(self):
program = Program()
with program_guard(program, startup_program=Program()):
dict_size = 10000
embed_size = 32
first_word = layers.data(name='firstw', shape=[1], dtype='int64')
second_word = layers.data(name='secondw', shape=[1], dtype='int64')
third_word = layers.data(name='thirdw', shape=[1], dtype='int64')
forth_word = layers.data(name='forthw', shape=[1], dtype='int64')
next_word = layers.data(name='nextw', shape=[1], dtype='int64')
embed_first = layers.embedding(
input=first_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
embed_second = layers.embedding(
input=second_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
embed_third = layers.embedding(
input=third_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
embed_forth = layers.embedding(
input=forth_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
concat_embed = layers.concat(
input=[embed_first, embed_second, embed_third, embed_forth],
axis=1)
hidden1 = layers.fc(input=concat_embed, size=256, act='sigmoid')
predict_word = layers.fc(input=hidden1,
size=dict_size,
act='softmax')
cost = layers.cross_entropy(input=predict_word, label=next_word)
avg_cost = layers.mean(cost)
self.assertIsNotNone(avg_cost)
print(str(program))
def forward(self, cue, label, return_loss=True):
out = self.conv1(cue)
out = self.norm1(out)
out = self.maxpool(out)
out = self.conv2(out)
out = self.norm2(out)
out = self.avgpool(out)
if return_loss:
cls_out = L.dropout(out, dropout_prob=0.5, is_test=False)
cls_out = self.fc(L.squeeze(cls_out, axes=[2, 3]))
loss_cls = L.mean(L.cross_entropy(cls_out, label))
losses = dict(loss_cls=loss_cls, loss=loss_cls)
return losses
else:
cls_out = self.fc(L.squeeze(out, axes=[2, 3]))
cls_out = L.softmax(cls_out).numpy()[:, 0]
return cls_out
def test_word_embedding(self):
program = Program()
with program_guard(program, startup_program=Program()):
dict_size = 10000
embed_size = 32
first_word = layers.data(name='firstw', shape=[1], dtype='int64')
second_word = layers.data(name='secondw', shape=[1], dtype='int64')
third_word = layers.data(name='thirdw', shape=[1], dtype='int64')
forth_word = layers.data(name='forthw', shape=[1], dtype='int64')
next_word = layers.data(name='nextw', shape=[1], dtype='int64')
embed_first = layers.embedding(
input=first_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
embed_second = layers.embedding(
input=second_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
embed_third = layers.embedding(
input=third_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
embed_forth = layers.embedding(
input=forth_word,
size=[dict_size, embed_size],
dtype='float32',
param_attr='shared_w')
concat_embed = layers.concat(
input=[embed_first, embed_second, embed_third, embed_forth],
axis=1)
hidden1 = layers.fc(input=concat_embed, size=256, act='sigmoid')
predict_word = layers.fc(input=hidden1,
size=dict_size,
act='softmax')
cost = layers.cross_entropy(input=predict_word, label=next_word)
avg_cost = layers.mean(cost)
self.assertIsNotNone(avg_cost)
print(str(program))
def test_recognize_digits_mlp(self):
program = Program()
with program_guard(program, startup_program=Program()):
# Change g_program, so the rest layers use `g_program`
images = layers.data(name='pixel', shape=[784], dtype='float32')
label = layers.data(name='label', shape=[1], dtype='int32')
hidden1 = layers.fc(input=images, size=128, act='relu')
hidden2 = layers.fc(input=hidden1, size=64, act='relu')
predict = layers.fc(input=[hidden2, hidden1],
size=10,
act='softmax',
param_attr=["sftmax.w1", "sftmax.w2"])
cost = layers.cross_entropy(input=predict, label=label)
avg_cost = layers.mean(cost)
self.assertIsNotNone(avg_cost)
print(str(program))
def get_losses(self, out, cls_out, mask, gt_labels):
loss_cls = L.mean(L.cross_entropy(cls_out,
gt_labels)) * self.train_cfg['w_cls']
cue = out[-1] if self.train_cfg['with_mask'] else L.elementwise_mul(
out[-1], L.cast(gt_labels, 'float32'), axis=0)
num_reg = L.cast(
L.reduce_sum(gt_labels) * cue.shape[1] * cue.shape[2] *
cue.shape[3], 'float32')
loss_reg = L.reduce_sum(
L.abs(mask - cue)) / (num_reg + 1e-8) * self.train_cfg['w_reg']
loss_tir = 0
for feat in out[:-1]:
feat = L.squeeze(self.avgpool(feat), axes=[2, 3])
loss_tir += self.triple_loss(feat,
gt_labels) * self.train_cfg['w_tri']
loss = loss_cls + loss_reg + loss_tir
return dict(loss_cls=loss_cls,
loss_reg=loss_reg,
loss_tir=loss_tir,
loss=loss)
def node_classify_model(config):
"""Build node classify model.
"""
nodes = fl.data('nodes', shape=[None, 1], dtype='int64')
labels = fl.data('labels', shape=[None, 1], dtype='int64')
embed_nodes = fl.embedding(input=nodes,
size=[config.num_nodes, config.embed_dim],
param_attr=fluid.ParamAttr(name='weight'))
embed_nodes.stop_gradient = True
probs = fl.fc(input=embed_nodes, size=config.num_labels, act='softmax')
predict = fl.argmax(probs, axis=-1)
loss = fl.cross_entropy(input=probs, label=labels)
loss = fl.reduce_mean(loss)
return {
'loss': loss,
'probs': probs,
'predict': predict,
'labels': labels,
}
def test_raw_api(self):
prog = Program()
startup_prog = Program()
with program_guard(prog, startup_prog):
image = layers.data(name='x', shape=[784], dtype='float32')
label = layers.data(name='y', shape=[1], dtype='int64')
limit = layers.fill_constant_batch_size_like(
input=label, dtype='int64', shape=[1], value=5.0)
cond = layers.less_than(x=label, y=limit)
true_image, false_image = layers.split_lod_tensor(
input=image, mask=cond)
true_out = layers.create_tensor(dtype='float32')
true_cond = layers.ConditionalBlock([true_image])
with true_cond.block():
hidden = layers.fc(input=true_image, size=100, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
layers.assign(input=prob, output=true_out)
false_out = layers.create_tensor(dtype='float32')
false_cond = layers.ConditionalBlock([false_image])
with false_cond.block():
hidden = layers.fc(input=false_image, size=200, act='tanh')
prob = layers.fc(input=hidden, size=10, act='softmax')
layers.assign(input=prob, output=false_out)
prob = layers.merge_lod_tensor(
in_true=true_out, in_false=false_out, mask=cond, x=image)
loss = layers.cross_entropy(input=prob, label=label)
avg_loss = layers.mean(loss)
optimizer = MomentumOptimizer(learning_rate=0.001, momentum=0.9)
optimizer.minimize(avg_loss, startup_prog)
train_reader = paddle.batch(
paddle.reader.shuffle(
paddle.dataset.mnist.train(), buf_size=8192),
batch_size=200)
place = core.CPUPlace()
exe = Executor(place)
exe.run(startup_prog)
PASS_NUM = 100
for pass_id in range(PASS_NUM):
for data in train_reader():
x_data = np.array(map(lambda x: x[0], data)).astype("float32")
y_data = np.array(map(lambda x: x[1], data)).astype("int64")
y_data = np.expand_dims(y_data, axis=1)
outs = exe.run(prog,
feed={'x': x_data,
'y': y_data},
fetch_list=[avg_loss])
print outs[0]
if outs[0] < 1.0:
return
self.assertFalse(True)
def knowledge_seq2seq(config):
""" knowledge seq2seq """
emb_size = config.embed_size
hidden_size = config.hidden_size
input_size = emb_size
num_layers = config.num_layers
bi_direc = config.bidirectional
batch_size = config.batch_size
vocab_size = config.vocab_size
run_type = config.run_type
enc_input = layers.data(name="enc_input",
shape=[1],
dtype='int64',
lod_level=1) #enc_input --> goal
enc_mask = layers.data(name="enc_mask", shape=[-1, 1], dtype='float32')
goal_input = layers.data(name="goal_input",
shape=[1],
dtype='int64',
lod_level=1) #goal_input --> x
cue_input = layers.data(name="cue_input",
shape=[1],
dtype='int64',
lod_level=1) #cue_input --> kg
#cue_mask = layers.data(name='cue_mask', shape=[-1, 1], dtype='float32')
memory_mask = layers.data(name='memory_mask',
shape=[-1, 1],
dtype='float32')
tar_input = layers.data(name='tar_input',
shape=[1],
dtype='int64',
lod_level=1) #tar_input --> y
# tar_mask = layers.data(name="tar_mask", shape=[-1, 1], dtype='float32')
rnn_hidden_size = hidden_size
if bi_direc:
rnn_hidden_size //= 2
enc_out, enc_last_hidden = \
rnn_encoder(enc_input, vocab_size, input_size, rnn_hidden_size, batch_size, num_layers, bi_direc,
dropout=0.0, batch_first=True, name="rnn_enc")
goal_out, goal_last_hidden = \
rnn_encoder(goal_input, vocab_size, input_size, rnn_hidden_size, batch_size, num_layers, bi_direc,
dropout=0.0, batch_first=True, name="rnn_enc1")
context_goal_out = fluid.layers.concat(
input=[enc_last_hidden, goal_last_hidden], axis=2)
context_goal_out = layers.reshape(context_goal_out,
shape=[-1, 1, rnn_hidden_size * 4])
# context_goal_out = layers.squeeze(context_goal_out, axes=[1])
context_goal_out = fluid.layers.fc(context_goal_out,
size=rnn_hidden_size * 2,
bias_attr=False)
context_goal_out = layers.unsqueeze(context_goal_out, axes=[0])
bridge_out = fc(context_goal_out, hidden_size, hidden_size, name="bridge")
bridge_out = layers.tanh(bridge_out)
cue_last_mask = layers.data(name='cue_last_mask',
shape=[-1],
dtype='float32')
knowledge_out, knowledge_last_hidden = \
rnn_encoder(cue_input, vocab_size, input_size, rnn_hidden_size, batch_size, num_layers, bi_direc,
dropout=0.0, batch_first=True, last_mask=cue_last_mask, name="knowledge_enc")
query = layers.slice(bridge_out, axes=[0], starts=[0], ends=[1])
query = layers.squeeze(query, axes=[0])
query = layers.unsqueeze(query, axes=[1])
query = layers.reshape(query, shape=[batch_size, -1, hidden_size])
cue_memory = layers.slice(knowledge_last_hidden,
axes=[0],
starts=[0],
ends=[1])
cue_memory = layers.reshape(cue_memory,
shape=[batch_size, -1, hidden_size])
memory_mask = layers.reshape(memory_mask, shape=[batch_size, 1, -1])
weighted_cue, cue_att = dot_attention(query, cue_memory, mask=memory_mask)
cue_att = layers.reshape(cue_att, shape=[batch_size, -1])
knowledge = weighted_cue
if config.use_posterior:
print("config.use_posterior", config.use_posterior)
target_out, target_last_hidden = \
rnn_encoder(tar_input, vocab_size, input_size, rnn_hidden_size, batch_size, num_layers, bi_direc,
dropout=0.0, batch_first=True, name="knowledge_enc1")
target_goal_out = fluid.layers.concat(
input=[target_last_hidden, goal_last_hidden], axis=2)
target_goal_out = layers.reshape(target_goal_out,
shape=[-1, 1, rnn_hidden_size * 4])
# target_goal_out = layers.squeeze(target_goal_out, axes=[1])
target_goal_out = fluid.layers.fc(target_goal_out,
size=rnn_hidden_size * 2,
bias_attr=False)
target_goal_out = layers.unsqueeze(target_goal_out, axes=[0])
# get attenion
# target_query = layers.slice(target_last_hidden, axes=[0], starts=[0], ends=[1])
target_query = layers.slice(target_goal_out,
axes=[0],
starts=[0],
ends=[1])
target_query = layers.squeeze(target_query, axes=[0])
target_query = layers.unsqueeze(target_query, axes=[1])
target_query = layers.reshape(target_query,
shape=[batch_size, -1, hidden_size])
weight_target, target_att = dot_attention(target_query,
cue_memory,
mask=memory_mask)
target_att = layers.reshape(target_att, shape=[batch_size, -1])
# add to output
knowledge = weight_target
enc_memory_mask = layers.data(name="enc_memory_mask",
shape=[-1, 1],
dtype='float32')
enc_memory_mask = layers.unsqueeze(enc_memory_mask, axes=[1])
# decoder init_hidden, enc_memory, enc_mask
dec_init_hidden = bridge_out
pad_value = fluid.layers.assign(input=np.array([0.0], dtype='float32'))
enc_memory, origl_len_1 = layers.sequence_pad(x=enc_out,
pad_value=pad_value)
enc_memory.persistable = True
gru_unit = GRU_unit(input_size + hidden_size,
hidden_size,
num_layers=num_layers,
dropout=0.0,
name="decoder_gru_unit")
cue_gru_unit = GRU_unit(hidden_size + hidden_size,
hidden_size,
num_layers=num_layers,
dropout=0.0,
name="decoder_cue_gru_unit")
tgt_vocab_size = config.vocab_size
if run_type == "train":
if config.use_bow:
bow_logits = fc(knowledge,
hidden_size,
hidden_size,
name='bow_fc_1')
bow_logits = layers.tanh(bow_logits)
bow_logits = fc(bow_logits,
hidden_size,
tgt_vocab_size,
name='bow_fc_2')
bow_logits = layers.softmax(bow_logits)
bow_label = layers.data(name='bow_label',
shape=[-1, config.max_len],
dtype='int64')
bow_mask = layers.data(name="bow_mask",
shape=[-1, config.max_len],
dtype='float32')
bow_logits = layers.expand(bow_logits, [1, config.max_len, 1])
bow_logits = layers.reshape(bow_logits, shape=[-1, tgt_vocab_size])
bow_label = layers.reshape(bow_label, shape=[-1, 1])
bow_loss = layers.cross_entropy(bow_logits,
bow_label,
soft_label=False)
bow_loss = layers.reshape(bow_loss, shape=[-1, config.max_len])
bow_loss *= bow_mask
bow_loss = layers.reduce_sum(bow_loss, dim=[1])
bow_loss = layers.reduce_mean(bow_loss)
dec_input = layers.data(name="dec_input",
shape=[-1, 1, 1],
dtype='int64')
dec_mask = layers.data(name="dec_mask", shape=[-1, 1], dtype='float32')
dec_knowledge = weight_target
knowledge_goal_out = fluid.layers.concat(
input=[dec_knowledge, target_query], axis=2)
knowledge_goal_out = layers.reshape(knowledge_goal_out,
shape=[-1, 1, rnn_hidden_size * 4])
# knowledge_goal_out = layers.squeeze(knowledge_goal_out, axes=[1])
knowledge_goal_out = fluid.layers.fc(knowledge_goal_out,
size=rnn_hidden_size * 2,
bias_attr=False)
knowledge_goal_out = layers.unsqueeze(knowledge_goal_out, axes=[0])
decoder_logits = \
rnn_decoder(gru_unit, cue_gru_unit, dec_input, input_size, hidden_size, num_layers,
enc_memory, enc_memory_mask, dec_knowledge, vocab_size,
init_hidden=dec_init_hidden, mask=dec_mask, dropout=config.dropout)
target_label = layers.data(name='target_label',
shape=[-1, 1],
dtype='int64')
target_mask = layers.data(name='target_mask',
shape=[-1, 1],
dtype='float32')
decoder_logits = layers.reshape(decoder_logits,
shape=[-1, tgt_vocab_size])
target_label = layers.reshape(target_label, shape=[-1, 1])
nll_loss = layers.cross_entropy(decoder_logits,
target_label,
soft_label=False)
nll_loss = layers.reshape(nll_loss, shape=[batch_size, -1])
nll_loss *= target_mask
nll_loss = layers.reduce_sum(nll_loss, dim=[1])
nll_loss = layers.reduce_mean(nll_loss)
prior_attn = cue_att + 1e-10
posterior_att = target_att
posterior_att.stop_gradient = True
prior_attn = layers.log(prior_attn)
kl_loss = posterior_att * (layers.log(posterior_att + 1e-10) -
prior_attn)
kl_loss = layers.reduce_mean(kl_loss)
kl_and_nll_factor = layers.data(name='kl_and_nll_factor',
shape=[1],
dtype='float32')
kl_and_nll_factor = layers.reshape(kl_and_nll_factor, shape=[-1])
final_loss = bow_loss + kl_loss * kl_and_nll_factor + nll_loss * kl_and_nll_factor
return [bow_loss, kl_loss, nll_loss, final_loss]
elif run_type == "test":
beam_size = config.beam_size
batch_size = config.batch_size
token = layers.fill_constant(shape=[batch_size * beam_size, 1],
value=config.bos_id,
dtype='int64')
token = layers.reshape(token, shape=[-1, 1])
max_decode_len = config.max_dec_len
dec_knowledge = knowledge
INF = 100000000.0
init_score_np = np.ones([beam_size * batch_size],
dtype='float32') * -INF
for i in range(batch_size):
init_score_np[i * beam_size] = 0.0
pre_score = layers.assign(init_score_np)
pos_index_np = np.arange(batch_size).reshape(-1, 1)
pos_index_np = \
np.tile(pos_index_np, (1, beam_size)).reshape(-1).astype('int32') * beam_size
pos_index = layers.assign(pos_index_np)
id_array = []
score_array = []
index_array = []
init_enc_memory = layers.expand(enc_memory, [1, beam_size, 1])
init_enc_memory = layers.reshape(
init_enc_memory, shape=[batch_size * beam_size, -1, hidden_size])
init_enc_mask = layers.expand(enc_memory_mask, [1, beam_size, 1])
init_enc_mask = layers.reshape(init_enc_mask,
shape=[batch_size * beam_size, 1, -1])
dec_knowledge = layers.reshape(dec_knowledge,
shape=[-1, 1, hidden_size])
init_dec_knowledge = layers.expand(dec_knowledge, [1, beam_size, 1])
init_dec_knowledge = layers.reshape(
init_dec_knowledge,
shape=[batch_size * beam_size, -1, hidden_size])
dec_init_hidden = layers.expand(dec_init_hidden, [1, 1, beam_size])
dec_init_hidden = layers.reshape(dec_init_hidden,
shape=[1, -1, hidden_size])
length_average = config.length_average
UNK = config.unk_id
EOS = config.eos_id
for i in range(1, max_decode_len + 1):
dec_emb = get_embedding(token, input_size, vocab_size)
dec_out, dec_last_hidden = \
decoder_step(gru_unit, cue_gru_unit,
dec_emb, dec_init_hidden, input_size, hidden_size,
init_enc_memory, init_enc_mask, init_dec_knowledge, mask=None)
output_in_size = hidden_size + hidden_size
rnnout = layers.dropout(dec_out,
dropout_prob=config.dropout,
is_test=True)
rnnout = fc(rnnout,
output_in_size,
hidden_size,
name='dec_out_fc1')
rnnout = fc(rnnout, hidden_size, vocab_size, name='dec_out_fc2')
log_softmax_output = log_softmax(rnnout)
log_softmax_output = layers.squeeze(log_softmax_output, axes=[1])
if i > 1:
if length_average:
log_softmax_output = layers.elementwise_add(
(log_softmax_output / i),
(pre_score * (1.0 - 1.0 / i)),
axis=0)
else:
log_softmax_output = layers.elementwise_add(
log_softmax_output, pre_score, axis=0)
else:
log_softmax_output = layers.elementwise_add(log_softmax_output,
pre_score,
axis=0)
log_softmax_output = layers.reshape(log_softmax_output,
shape=[batch_size, -1])
topk_score, topk_index = layers.topk(log_softmax_output,
k=beam_size)
topk_score = layers.reshape(topk_score, shape=[-1])
topk_index = layers.reshape(topk_index, shape=[-1])
vocab_var = layers.fill_constant([1],
dtype='int64',
value=vocab_size)
new_token = topk_index % vocab_var
index = topk_index // vocab_var
id_array.append(new_token)
index_array.append(index)
index = index + pos_index
score_array.append(topk_score)
eos_ids = layers.fill_constant([beam_size * batch_size],
dtype='int64',
value=EOS)
unk_ids = layers.fill_constant([beam_size * batch_size],
dtype='int64',
value=UNK)
eos_eq = layers.cast(layers.equal(new_token, eos_ids),
dtype='float32')
topk_score += eos_eq * -100000000.0
unk_eq = layers.cast(layers.equal(new_token, unk_ids),
dtype='float32')
topk_score += unk_eq * -100000000.0
# update
token = new_token
pre_score = topk_score
token = layers.reshape(token, shape=[-1, 1])
index = layers.cast(index, dtype='int32')
dec_last_hidden = layers.squeeze(dec_last_hidden, axes=[0])
dec_init_hidden = layers.gather(dec_last_hidden, index=index)
dec_init_hidden = layers.unsqueeze(dec_init_hidden, axes=[0])
init_enc_memory = layers.gather(init_enc_memory, index)
init_enc_mask = layers.gather(init_enc_mask, index)
init_dec_knowledge = layers.gather(init_dec_knowledge, index)
final_score = layers.concat(score_array, axis=0)
final_ids = layers.concat(id_array, axis=0)
final_index = layers.concat(index_array, axis=0)
final_score = layers.reshape(
final_score, shape=[max_decode_len, beam_size * batch_size])
final_ids = layers.reshape(
final_ids, shape=[max_decode_len, beam_size * batch_size])
final_index = layers.reshape(
final_index, shape=[max_decode_len, beam_size * batch_size])
return final_score, final_ids, final_index
def transformer(
src_vocab_size,
trg_vocab_size,
max_length,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate,
src_pad_idx,
trg_pad_idx,
pos_pad_idx, ):
file_obj = open_recordio_file(
filename=os.environ.get('RECORDIO_FILENAME', '/tmp/wmt16.recordio'),
shapes=[
[batch_size * max_length, 1],
[batch_size * max_length, 1],
[batch_size * max_length, 1],
[batch_size * max_length, 1],
[batch_size, n_head, max_length, max_length],
[batch_size, n_head, max_length, max_length],
[batch_size, n_head, max_length, max_length],
[batch_size * max_length, 1],
[batch_size * max_length, 1],
],
dtypes=[
'int64',
'int64',
'int64',
'int64',
'float32',
'float32',
'float32',
'int64',
'float32',
],
lod_levels=[0] * 9)
src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, trg_slf_attn_bias, trg_src_attn_bias, gold, weights = fluid.layers.read_file(
file_obj)
enc_input = prepare_encoder(
src_word,
src_pos,
src_vocab_size,
d_model,
src_pad_idx,
max_length,
dropout_rate, )
enc_output = encoder(
enc_input,
src_slf_attn_bias,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate, )
dec_input = prepare_decoder(
trg_word,
trg_pos,
trg_vocab_size,
d_model,
trg_pad_idx,
max_length,
dropout_rate, )
dec_output = decoder(
dec_input,
enc_output,
trg_slf_attn_bias,
trg_src_attn_bias,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate, )
# TODO(guosheng): Share the weight matrix between the embedding layers and
# the pre-softmax linear transformation.
predict = layers.reshape(
x=layers.fc(input=dec_output,
size=trg_vocab_size,
param_attr=fluid.initializer.Xavier(uniform=False),
bias_attr=False,
num_flatten_dims=2),
shape=[-1, trg_vocab_size],
act="softmax")
cost = layers.cross_entropy(input=predict, label=gold)
weighted_cost = cost * weights
return layers.reduce_sum(weighted_cost)
def train_main(use_cuda, is_sparse, is_local=True):
if use_cuda and not fluid.core.is_compiled_with_cuda():
return
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
context = encoder(is_sparse)
rnn_out = decoder_train(context, is_sparse)
label = pd.data(name="target_language_next_word",
shape=[1],
dtype='int64',
lod_level=1)
cost = pd.cross_entropy(input=rnn_out, label=label)
avg_cost = pd.mean(cost)
optimizer = fluid.optimizer.Adagrad(
learning_rate=1e-4,
regularization=fluid.regularizer.L2DecayRegularizer(
regularization_coeff=0.1))
optimize_ops, params_grads = optimizer.minimize(avg_cost)
train_data = paddle.batch(paddle.reader.shuffle(
paddle.dataset.wmt14.train(dict_size), buf_size=1000),
batch_size=batch_size)
exe = Executor(place)
def train_loop(main_program):
exe.run(framework.default_startup_program())
batch_id = 0
for pass_id in xrange(1):
for data in train_data():
word_data = to_lodtensor(map(lambda x: x[0], data), place)
trg_word = to_lodtensor(map(lambda x: x[1], data), place)
trg_word_next = to_lodtensor(map(lambda x: x[2], data), place)
outs = exe.run(main_program,
feed={
'src_word_id': word_data,
'target_language_word': trg_word,
'target_language_next_word': trg_word_next
},
fetch_list=[avg_cost])
avg_cost_val = np.array(outs[0])
print('pass_id=' + str(pass_id) + ' batch=' + str(batch_id) +
" avg_cost=" + str(avg_cost_val))
if batch_id > 3:
break
batch_id += 1
if is_local:
train_loop(framework.default_main_program())
else:
port = os.getenv("PADDLE_INIT_PORT", "6174")
pserver_ips = os.getenv("PADDLE_INIT_PSERVERS") # ip,ip...
eplist = []
for ip in pserver_ips.split(","):
eplist.append(':'.join([ip, port]))
pserver_endpoints = ",".join(eplist) # ip:port,ip:port...
trainers = int(os.getenv("TRAINERS"))
current_endpoint = os.getenv("POD_IP") + ":" + port
trainer_id = int(os.getenv("PADDLE_INIT_TRAINER_ID"))
training_role = os.getenv("TRAINING_ROLE", "TRAINER")
t = fluid.DistributeTranspiler()
t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers)
if training_role == "PSERVER":
pserver_prog = t.get_pserver_program(current_endpoint)
pserver_startup = t.get_startup_program(current_endpoint,
pserver_prog)
exe.run(pserver_startup)
exe.run(pserver_prog)
elif training_role == "TRAINER":
train_loop(t.get_trainer_program())
def transformer(
src_vocab_size,
trg_vocab_size,
max_length,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate,
src_pad_idx,
trg_pad_idx,
pos_pad_idx, ):
file_obj = fluid.layers.open_recordio_file(
filename='./wmt16.recordio',
shapes=[
[batch_size * max_length, 1],
[batch_size * max_length, 1],
[batch_size * max_length, 1],
[batch_size * max_length, 1],
[batch_size, n_head, max_length, max_length],
[batch_size, n_head, max_length, max_length],
[batch_size, n_head, max_length, max_length],
[batch_size * max_length, 1],
[batch_size * max_length, 1],
],
dtypes=[
'int64',
'int64',
'int64',
'int64',
'float32',
'float32',
'float32',
'int64',
'float32',
],
lod_levels=[0] * 9)
src_word, src_pos, trg_word, trg_pos, src_slf_attn_bias, trg_slf_attn_bias, trg_src_attn_bias, gold, weights = fluid.layers.read_file(
file_obj)
enc_input = prepare_encoder(
src_word,
src_pos,
src_vocab_size,
d_model,
src_pad_idx,
max_length,
dropout_rate, )
enc_output = encoder(
enc_input,
src_slf_attn_bias,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate, )
dec_input = prepare_decoder(
trg_word,
trg_pos,
trg_vocab_size,
d_model,
trg_pad_idx,
max_length,
dropout_rate, )
dec_output = decoder(
dec_input,
enc_output,
trg_slf_attn_bias,
trg_src_attn_bias,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate, )
# TODO(guosheng): Share the weight matrix between the embedding layers and
# the pre-softmax linear transformation.
predict = layers.reshape(
x=layers.fc(input=dec_output,
size=trg_vocab_size,
param_attr=fluid.initializer.Xavier(uniform=False),
bias_attr=False,
num_flatten_dims=2),
shape=[-1, trg_vocab_size],
act="softmax")
cost = layers.cross_entropy(input=predict, label=gold)
weighted_cost = cost * weights
return layers.reduce_sum(weighted_cost)
