def net_profiler(self, state, profile_path='/tmp/profile'):
enable_if_gpu = state == 'GPU' or state == "All"
if enable_if_gpu and not core.is_compiled_with_cuda():
return
startup_program = fluid.Program()
main_program = fluid.Program()
with fluid.program_guard(main_program, startup_program):
image = fluid.layers.data(name='x', shape=[784], dtype='float32')
hidden1 = fluid.layers.fc(input=image, size=64, act='relu')
i = layers.zeros(shape=[1], dtype='int64')
counter = fluid.layers.zeros(shape=[1],
dtype='int64',
force_cpu=True)
until = layers.fill_constant([1], dtype='int64', value=10)
data_arr = layers.array_write(hidden1, i)
cond = fluid.layers.less_than(x=counter, y=until)
while_op = fluid.layers.While(cond=cond)
with while_op.block():
hidden_n = fluid.layers.fc(input=hidden1, size=64, act='relu')
layers.array_write(hidden_n, i, data_arr)
fluid.layers.increment(x=counter, value=1, in_place=True)
layers.less_than(x=counter, y=until, cond=cond)
hidden_n = layers.array_read(data_arr, i)
hidden2 = fluid.layers.fc(input=hidden_n, size=64, act='relu')
predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
label = fluid.layers.data(name='y', shape=[1], dtype='int64')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(cost)
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=predict,
label=label,
total=batch_size)
optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
opts = optimizer.minimize(avg_cost, startup_program=startup_program)
place = fluid.CPUPlace() if state == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(startup_program)
pass_acc_calculator = fluid.average.WeightedAverage()
with profiler.profiler(state, 'total', profile_path) as prof:
for iter in range(10):
if iter == 2:
profiler.reset_profiler()
x = np.random.random((32, 784)).astype("float32")
y = np.random.randint(0, 10, (32, 1)).astype("int64")
outs = exe.run(main_program,
feed={
'x': x,
'y': y
},
fetch_list=[avg_cost, batch_acc, batch_size])
acc = np.array(outs[1])
b_size = np.array(outs[2])
pass_acc_calculator.add(value=acc, weight=b_size)
pass_acc = pass_acc_calculator.eval()
def check_switch(self, value):
x = layers.fill_constant(shape=[1], dtype='float32', value=value)
zero_var = layers.fill_constant(shape=[1], dtype='float32', value=0.0)
one_var = layers.fill_constant(shape=[1], dtype='float32', value=1.0)
two_var = layers.fill_constant(shape=[1], dtype='float32', value=2.0)
three_var = layers.fill_constant(shape=[1], dtype='float32', value=3.0)
result = layers.create_global_var(shape=[1],
value=-1.0,
dtype='float32',
persistable=True)
with layers.Switch() as switch:
with switch.case(layers.less_than(x, zero_var)):
layers.assign(zero_var, result)
with switch.case(layers.less_than(x, one_var)):
layers.assign(one_var, result)
with switch.case(layers.less_than(x, two_var)):
layers.assign(two_var, result)
with switch.default():
layers.assign(three_var, result)
cpu = core.CPUPlace()
exe = Executor(cpu)
exe.run(default_startup_program())
out = exe.run(feed={}, fetch_list=[result])[0][0]
return out
def is_finished(alive_log_prob, finished_scores,
finished_in_finished):
max_out_len = 200
max_length_penalty = layers.pow(
layers.fill_constant([1],
dtype='float32',
value=((5.0 + max_out_len) /
6.0)), alpha)
lower_bound_alive_score = layers.slice(
alive_log_prob, starts=[0], ends=[1],
axes=[0]) / max_length_penalty
lowest_score_of_fininshed_in_finished = finished_scores * finished_in_finished
lowest_score_of_fininshed_in_finished += (
1.0 - finished_in_finished) * -INF
lowest_score_of_fininshed_in_finished = layers.reduce_min(
lowest_score_of_fininshed_in_finished)
met = layers.less_than(
lower_bound_alive_score,
lowest_score_of_fininshed_in_finished)
met = layers.cast(met, 'float32')
bound_is_met = layers.reduce_sum(met)
finished_eos_num = layers.reduce_sum(finished_in_finished)
finish_cond = layers.less_than(
finished_eos_num,
layers.fill_constant([1],
dtype='float32',
value=beam_size))
return finish_cond
def check_switch(self, value):
x = layers.fill_constant(shape=[1], dtype='float32', value=value)
zero_var = layers.fill_constant(shape=[1], dtype='float32', value=0.0)
one_var = layers.fill_constant(shape=[1], dtype='float32', value=1.0)
two_var = layers.fill_constant(shape=[1], dtype='float32', value=2.0)
three_var = layers.fill_constant(shape=[1], dtype='float32', value=3.0)
result = layers.create_global_var(
shape=[1], value=-1.0, dtype='float32', persistable=True)
with layers.Switch() as switch:
with switch.case(layers.less_than(x, zero_var)):
layers.assign(zero_var, result)
with switch.case(layers.less_than(x, one_var)):
layers.assign(one_var, result)
with switch.case(layers.less_than(x, two_var)):
layers.assign(two_var, result)
with switch.default():
layers.assign(three_var, result)
cpu = core.CPUPlace()
exe = Executor(cpu)
exe.run(default_startup_program())
out = exe.run(feed={}, fetch_list=[result])[0][0]
return out
def simple_net(self):
d0 = layers.data(
"d0", shape=[10], append_batch_size=False, dtype='float32')
d1 = layers.data(
"d1", shape=[10], append_batch_size=False, dtype='float32')
d2 = layers.data(
"d2", shape=[10], append_batch_size=False, dtype='float32')
# fill_constant npu op doesn't support int64
i = layers.zeros(shape=[1], dtype='int32')
i = layers.cast(i, 'int64')
i.stop_gradient = True
init = layers.zeros(shape=[10], dtype='float32')
mem_array = layers.array_write(x=init, i=i)
data_array = layers.array_write(x=d0, i=i)
i = layers.increment(i)
layers.array_write(d1, i, array=data_array)
i = layers.increment(i)
layers.array_write(d2, i, array=data_array)
i = layers.zeros(shape=[1], dtype='int32')
i = layers.cast(i, 'int64')
i.stop_gradient = True
array_len = layers.fill_constant(shape=[1], dtype='int32', value=5)
array_len = layers.cast(array_len, 'int64')
array_len.stop_gradient = True
cond = layers.ones(shape=[1], dtype='int32')
cond = layers.cast(cond, 'bool')
j = layers.fill_constant(shape=[1], dtype='int32', value=1)
j = layers.cast(j, 'int64')
j.stop_gradient = True
array_len2 = layers.fill_constant(shape=[1], dtype='int32', value=3)
array_len2 = layers.cast(array_len2, 'int64')
array_len2.stop_gradient = True
cond2 = layers.logical_or(x=j, y=array_len2)
cond2 = layers.ones(shape=[1], dtype='int32')
cond2 = layers.cast(cond2, 'bool')
while_op = layers.While(cond=cond)
while_op2 = layers.While(cond=cond2)
with while_op.block():
d = layers.array_read(array=data_array, i=i)
prev = layers.array_read(array=mem_array, i=i)
result = layers.sums(input=[d, prev])
i = layers.increment(x=i, in_place=True)
layers.array_write(result, i=i, array=mem_array)
layers.less_than(x=i, y=array_len, cond=cond)
with while_op2.block():
d2 = layers.array_read(array=data_array, i=j)
prev2 = layers.array_read(array=mem_array, i=j)
result2 = layers.sums(input=[d2, prev2])
j = layers.increment(x=j, in_place=True)
layers.array_write(result2, i=j, array=mem_array)
layers.less_than(x=j, y=array_len2, cond=cond2)
sum_result = layers.array_read(array=mem_array, i=j)
loss = layers.mean(sum_result)
return loss, sum_result
def net_profiler(self, state, profile_path='/tmp/profile'):
enable_if_gpu = state == 'GPU' or state == "All"
if enable_if_gpu and not core.is_compiled_with_cuda():
return
startup_program = fluid.Program()
main_program = fluid.Program()
with fluid.program_guard(main_program, startup_program):
image = fluid.layers.data(name='x', shape=[784], dtype='float32')
hidden1 = fluid.layers.fc(input=image, size=64, act='relu')
i = layers.zeros(shape=[1], dtype='int64')
counter = fluid.layers.zeros(
shape=[1], dtype='int64', force_cpu=True)
until = layers.fill_constant([1], dtype='int64', value=10)
data_arr = layers.array_write(hidden1, i)
cond = fluid.layers.less_than(x=counter, y=until)
while_op = fluid.layers.While(cond=cond)
with while_op.block():
hidden_n = fluid.layers.fc(input=hidden1, size=64, act='relu')
layers.array_write(hidden_n, i, data_arr)
fluid.layers.increment(x=counter, value=1, in_place=True)
layers.less_than(x=counter, y=until, cond=cond)
hidden_n = layers.array_read(data_arr, i)
hidden2 = fluid.layers.fc(input=hidden_n, size=64, act='relu')
predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
label = fluid.layers.data(name='y', shape=[1], dtype='int64')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(cost)
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size)
optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
opts = optimizer.minimize(avg_cost, startup_program=startup_program)
place = fluid.CPUPlace() if state == 'CPU' else fluid.CUDAPlace(0)
exe = fluid.Executor(place)
exe.run(startup_program)
pass_acc_calculator = fluid.average.WeightedAverage()
with profiler.profiler(state, 'total', profile_path) as prof:
for iter in range(10):
if iter == 2:
profiler.reset_profiler()
x = np.random.random((32, 784)).astype("float32")
y = np.random.randint(0, 10, (32, 1)).astype("int64")
outs = exe.run(main_program,
feed={'x': x,
'y': y},
fetch_list=[avg_cost, batch_acc, batch_size])
acc = np.array(outs[1])
b_size = np.array(outs[2])
pass_acc_calculator.add(value=acc, weight=b_size)
pass_acc = pass_acc_calculator.eval()
def test_simple_forward(self):
d0 = layers.data(
"d0", shape=[10], append_batch_size=False, dtype='float32')
d1 = layers.data(
"d1", shape=[10], append_batch_size=False, dtype='float32')
d2 = layers.data(
"d2", shape=[10], append_batch_size=False, dtype='float32')
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = True
init = layers.zeros(shape=[10], dtype='float32')
mem_array = layers.array_write(x=init, i=i)
data_array = layers.array_write(x=d0, i=i)
i = layers.increment(i)
layers.array_write(d1, i, array=data_array)
i = layers.increment(i)
layers.array_write(d2, i, array=data_array)
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = True
array_len = layers.fill_constant(shape=[1], dtype='int64', value=3)
array_len.stop_gradient = True
cond = layers.less_than(x=i, y=array_len)
while_op = layers.While(cond=cond)
with while_op.block():
d = layers.array_read(array=data_array, i=i)
prev = layers.array_read(array=mem_array, i=i)
result = layers.sums(input=[d, prev])
i = layers.increment(x=i, in_place=True)
layers.array_write(result, i=i, array=mem_array)
layers.less_than(x=i, y=array_len, cond=cond)
sum_result = layers.array_read(array=mem_array, i=i)
loss = layers.mean(sum_result)
append_backward(loss)
cpu = core.CPUPlace()
exe = Executor(cpu)
d = []
for i in xrange(3):
d.append(numpy.random.random(size=[10]).astype('float32'))
outs = exe.run(feed={'d0': d[0],
'd1': d[1],
'd2': d[2]},
fetch_list=[sum_result])
self.assertAlmostEqual(numpy.sum(d), numpy.sum(outs[0]), delta=0.01)
def test_return_single_var(self):
def fn_1():
return layers.fill_constant(shape=[4, 2], dtype='int32', value=1)
def fn_2():
return layers.fill_constant(shape=[4, 2], dtype='int32', value=2)
def fn_3():
return layers.fill_constant(shape=[4, 3], dtype='int32', value=3)
main_program = Program()
startup_program = Program()
with program_guard(main_program, startup_program):
x = layers.fill_constant(shape=[1], dtype='float32', value=0.3)
y = layers.fill_constant(shape=[1], dtype='float32', value=0.1)
z = layers.fill_constant(shape=[1], dtype='float32', value=0.2)
pred_2 = layers.less_than(x, y) # false: 0.3 < 0.1
pred_1 = layers.less_than(z, x) # true: 0.2 < 0.3
# call fn_1
out_0 = layers.case(pred_fn_pairs=[(pred_1, fn_1), (pred_1, fn_2)],
default=fn_3)
# call fn_2
out_1 = layers.case(pred_fn_pairs=[(pred_2, fn_1), (pred_1, fn_2)],
default=fn_3)
# call default fn_3
out_2 = layers.case(pred_fn_pairs=((pred_2, fn_1), (pred_2, fn_2)),
default=fn_3)
# no default, call fn_2
out_3 = layers.case(pred_fn_pairs=[(pred_1, fn_2)])
# no default, call fn_2. but pred_2 is false
out_4 = layers.case(pred_fn_pairs=[(pred_2, fn_2)])
place = fluid.CUDAPlace(
0) if core.is_compiled_with_cuda() else fluid.CPUPlace()
exe = fluid.Executor(place)
res = exe.run(main_program,
fetch_list=[out_0, out_1, out_2, out_3, out_4])
self.assertTrue(np.allclose(res[0], 1))
self.assertTrue(np.allclose(res[1], 2))
self.assertTrue(np.allclose(res[2], 3))
self.assertTrue(np.allclose(res[3], 2))
self.assertTrue(np.allclose(res[4], 2))
def test_return_single_var(self):
"""
pseudocode:
if 0.23 < 0.1:
return 2
else:
return -1
"""
def true_func():
return layers.fill_constant(shape=[2, 3], dtype='int32', value=2)
def false_func():
return layers.fill_constant(shape=[3, 2], dtype='int32', value=-1)
main_program = Program()
startup_program = Program()
with program_guard(main_program, startup_program):
x = layers.fill_constant(shape=[1], dtype='float32', value=0.1)
y = layers.fill_constant(shape=[1], dtype='float32', value=0.23)
pred = layers.less_than(y, x)
out = layers.cond(pred, true_func, false_func)
# out is one tensor
place = fluid.CUDAPlace(
0) if core.is_compiled_with_cuda() else fluid.CPUPlace()
exe = fluid.Executor(place)
ret = exe.run(main_program, fetch_list=[out.name])
self.assertTrue(
np.allclose(np.asarray(ret), np.full((3, 2), -1, np.int32)))
def is_finished(self, step_idx, source_length, alive_log_probs, finished_scores, finished_in_finished):
"""
is_finished
"""
base_1 = layers.cast(source_length, 'float32') + 55.0
base_1 /= 6.0
max_length_penalty = layers.pow(base_1, self.alpha)
flat_alive_log_probs = layers.reshape(alive_log_probs, [-1])
lower_bound_alive_scores_1 = layers.gather(flat_alive_log_probs, [self.get_alive_index])
lower_bound_alive_scores = lower_bound_alive_scores_1 / max_length_penalty
lowest_score_of_finished_in_finish = layers.reduce_min(finished_scores * finished_in_finished, dim=1)
finished_in_finished = layers.cast(finished_in_finished, 'bool')
lowest_score_of_finished_in_finish += \
((1.0 - layers.cast(layers.reduce_any(finished_in_finished, 1), 'float32')) * -INF)
#print lowest_score_of_finished_in_finish
bound_is_met = layers.reduce_all(layers.greater_than(lowest_score_of_finished_in_finish,
lower_bound_alive_scores))
decode_length = source_length + 50
length_cond = layers.less_than(x=step_idx, y=decode_length)
return layers.logical_and(x=layers.logical_not(bound_is_met), y=length_cond)
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 compare_ifelse_op_and_numpy(self, place):
self.set_test_case()
prog = Program()
startup_prog = Program()
with program_guard(prog, startup_prog):
src = layers.data(name='data', shape=[1], dtype='float32')
cond = layers.fill_constant(
[1], dtype='float32', value=self.cond_value)
ifcond = layers.less_than(x=src, y=cond)
ie = layers.IfElse(ifcond)
with ie.true_block():
true_target = ie.input(src)
true_target = fluid.layers.exp(true_target)
ie.output(true_target)
with ie.false_block():
false_target = ie.input(src)
false_target = fluid.layers.tanh(false_target)
ie.output(false_target)
if_out = ie()
out = layers.reduce_sum(if_out)
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())
fetch_list = [out]
o1, = exe.run(fluid.default_main_program(),
feed={'data': self.data},
fetch_list=[out])
o2 = self.numpy_cal()
self.assertTrue(
np.allclose(
o1, o2, atol=1e-8),
"IfElse result : " + str(o1) + "\n Numpy result :" + str(o2))
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 build_program(self, compile_program=True):
startup_program = fluid.Program()
main_program = fluid.Program()
with fluid.program_guard(main_program, startup_program):
image = fluid.layers.data(name='x', shape=[784], dtype='float32')
hidden1 = fluid.layers.fc(input=image, size=64, act='relu')
i = layers.zeros(shape=[1], dtype='int64')
counter = fluid.layers.zeros(shape=[1],
dtype='int64',
force_cpu=True)
until = layers.fill_constant([1], dtype='int64', value=10)
data_arr = layers.array_write(hidden1, i)
cond = fluid.layers.less_than(x=counter, y=until)
while_op = fluid.layers.While(cond=cond)
with while_op.block():
hidden_n = fluid.layers.fc(input=hidden1, size=64, act='relu')
layers.array_write(hidden_n, i, data_arr)
fluid.layers.increment(x=counter, value=1, in_place=True)
layers.less_than(x=counter, y=until, cond=cond)
hidden_n = layers.array_read(data_arr, i)
hidden2 = fluid.layers.fc(input=hidden_n, size=64, act='relu')
predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
label = fluid.layers.data(name='y', shape=[1], dtype='int64')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(cost)
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(input=predict,
label=label,
total=batch_size)
optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
opts = optimizer.minimize(avg_cost, startup_program=startup_program)
if compile_program:
# TODO(luotao): profiler tool may have bug with multi-thread parallel executor.
# https://github.com/PaddlePaddle/Paddle/pull/25200#issuecomment-650483092
exec_strategy = fluid.ExecutionStrategy()
exec_strategy.num_threads = 1
train_program = fluid.compiler.CompiledProgram(
main_program).with_data_parallel(loss_name=avg_cost.name,
exec_strategy=exec_strategy)
else:
train_program = main_program
return train_program, startup_program, avg_cost, batch_size, batch_acc
def test_exceptions(self):
i = layers.zeros(shape=[2], dtype='int64')
array_len = layers.fill_constant(shape=[2], dtype='int64', value=1)
cond = layers.less_than(x=i, y=array_len)
with self.assertRaises(TypeError):
layers.While(cond=cond)
cond = layers.cast(cond, dtype='float64')
with self.assertRaises(TypeError):
layers.While(cond=cond)
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 build_and_run_program(place, batch_size, beam_size, stop_gradient=False):
fluid.default_startup_program().random_seed = 1
fluid.default_main_program().random_seed = 1
np.random.seed(2)
x = layers.assign(
np.random.rand(batch_size, beam_size, 32).astype("float32"))
indices = fluid.data(shape=[None, beam_size], dtype="int64", name="indices")
step_idx = layers.fill_constant(
shape=[1], dtype="int64", value=0, force_cpu=True)
max_len = layers.fill_constant(
shape=[1], dtype="int64", value=10, force_cpu=True)
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
scores = layers.array_write(x, step_idx)
with while_op.block():
bs = layers.cast(layers.shape(x)[0], "int64")
for _ in range(20):
bs = layers.cast(bs, 'int64')
bs.stop_gradient = stop_gradient
batch_pos = layers.expand(
layers.unsqueeze(
layers.range(
0, bs, 1, dtype=bs.dtype), [1]), [1, beam_size])
topk_coordinates = layers.stack([batch_pos, indices], axis=2)
topk_coordinates.stop_gradient = stop_gradient
score = layers.gather_nd(x, topk_coordinates)
layers.increment(x=step_idx, value=1.0, in_place=True)
layers.array_write(score, i=step_idx, array=scores)
length_cond = layers.less_than(x=step_idx, y=max_len)
layers.assign(length_cond, cond)
out = layers.tensor_array_to_tensor(scores, axis=0, use_stack=True)[0]
loss = layers.reduce_mean(out)
opt = fluid.optimizer.Adam(0.01)
opt.minimize(loss)
exe = fluid.Executor(place)
data = np.random.random_integers(
low=0, high=beam_size - 1, size=(batch_size, beam_size)).astype("int64")
loss_val, = exe.run(feed={"indices": data}, fetch_list=[loss])
return loss_val
def decrement(self):
new_scale = self.scale / self.factor
one = layers.fill_constant(shape=[1], dtype='float32', value=1.0)
less_than_one = layers.less_than(new_scale, one)
with layers.Switch() as switch:
with switch.case(less_than_one):
layers.assign(one, self.scale)
with switch.default():
layers.assign(new_scale, self.scale)
layers.assign(layers.zeros_like(self.good_steps), self.good_steps)
def append_cond_op(self, program):
def true_func():
return layers.fill_constant(shape=[2, 3], dtype='int32', value=2)
def false_func():
return layers.fill_constant(shape=[3, 2], dtype='int32', value=-1)
with fluid.program_guard(program):
x = layers.fill_constant(shape=[1], dtype='float32', value=0.1)
y = layers.fill_constant(shape=[1], dtype='float32', value=0.23)
pred = layers.less_than(y, x)
out = layers.cond(pred, true_func, false_func)
def build_program(self, compile_program=True):
startup_program = fluid.Program()
main_program = fluid.Program()
with fluid.program_guard(main_program, startup_program):
image = fluid.layers.data(name='x', shape=[784], dtype='float32')
hidden1 = fluid.layers.fc(input=image, size=64, act='relu')
i = layers.zeros(shape=[1], dtype='int64')
counter = fluid.layers.zeros(
shape=[1], dtype='int64', force_cpu=True)
until = layers.fill_constant([1], dtype='int64', value=10)
data_arr = layers.array_write(hidden1, i)
cond = fluid.layers.less_than(x=counter, y=until)
while_op = fluid.layers.While(cond=cond)
with while_op.block():
hidden_n = fluid.layers.fc(input=hidden1, size=64, act='relu')
layers.array_write(hidden_n, i, data_arr)
fluid.layers.increment(x=counter, value=1, in_place=True)
layers.less_than(x=counter, y=until, cond=cond)
hidden_n = layers.array_read(data_arr, i)
hidden2 = fluid.layers.fc(input=hidden_n, size=64, act='relu')
predict = fluid.layers.fc(input=hidden2, size=10, act='softmax')
label = fluid.layers.data(name='y', shape=[1], dtype='int64')
cost = fluid.layers.cross_entropy(input=predict, label=label)
avg_cost = fluid.layers.mean(cost)
batch_size = fluid.layers.create_tensor(dtype='int64')
batch_acc = fluid.layers.accuracy(
input=predict, label=label, total=batch_size)
optimizer = fluid.optimizer.Momentum(learning_rate=0.001, momentum=0.9)
opts = optimizer.minimize(avg_cost, startup_program=startup_program)
if compile_program:
train_program = fluid.compiler.CompiledProgram(
main_program).with_data_parallel(loss_name=avg_cost.name)
else:
train_program = main_program
return train_program, startup_program, avg_cost, batch_size, batch_acc
def increment(self):
enough_steps = layers.less_than(self.increment_every,
self.good_steps + 1)
with layers.Switch() as switch:
with switch.case(enough_steps):
new_scale = self.scale * self.factor
scale_valid = layers.isfinite(new_scale)
with layers.Switch() as switch2:
with switch2.case(scale_valid):
layers.assign(new_scale, self.scale)
layers.assign(layers.zeros_like(self.good_steps),
self.good_steps)
with switch2.default():
layers.increment(self.good_steps)
with switch.default():
layers.increment(self.good_steps)
def test_error(self):
def fn_1():
return layers.fill_constant(shape=[4, 2], dtype='int32', value=1)
main_program = Program()
startup_program = Program()
with program_guard(main_program, startup_program):
x = layers.fill_constant(shape=[1], dtype='float32', value=0.23)
z = layers.fill_constant(shape=[1], dtype='float32', value=0.2)
pred_1 = layers.less_than(z, x) # true
# The type of 'pred_fn_pairs' in case must be list or tuple
def type_error_pred_fn_pairs():
layers.case(pred_fn_pairs=1, default=fn_1)
self.assertRaises(TypeError, type_error_pred_fn_pairs)
# The elements' type of 'pred_fn_pairs' in Op(case) must be tuple
def type_error_pred_fn_1():
layers.case(pred_fn_pairs=[1], default=fn_1)
self.assertRaises(TypeError, type_error_pred_fn_1)
# The tuple's size of 'pred_fn_pairs' in Op(case) must be 2
def type_error_pred_fn_2():
layers.case(pred_fn_pairs=[(1, 2, 3)], default=fn_1)
self.assertRaises(TypeError, type_error_pred_fn_2)
# The pred's type of 'pred_fn_pairs' in Op(case) must be bool Variable
def type_error_pred():
layers.case(pred_fn_pairs=[(1, fn_1)], default=fn_1)
self.assertRaises(TypeError, type_error_pred)
# The function of pred_fn_pairs in case must be callable
def type_error_fn():
layers.case(pred_fn_pairs=[(pred_1, 2)], default=fn_1)
self.assertRaises(TypeError, type_error_fn)
# The default in Op(case) must be callable
def type_error_default():
layers.case(pred_fn_pairs=[(pred_1, fn_1)], default=fn_1())
self.assertRaises(TypeError, type_error_default)
def pairwise_hinge(self):
"""pairwise model"""
poi_repr = L.split(self.poi_repr, 2, dim=0)
pos_repr, neg_repr = poi_repr
pos_pred = L.cos_sim(self.query_repr, pos_repr)
neg_pred = L.cos_sim(self.query_repr, neg_repr)
mode = 'hinge_loss'
# log(1 + e-z), max(0, 1 - z)
if 'hinge_loss' == mode:
theta_z = L.relu(1 + neg_pred - pos_pred)
elif 'logistic_loss' == mode:
theta_z = L.log(1 + L.exp(neg_pred - pos_pred))
self.loss = L.reduce_mean(theta_z)
pos_cnt = L.reduce_sum(L.cast(L.greater_than(pos_pred, neg_pred), dtype="float32"))
neg_cnt = L.reduce_sum(L.cast(L.less_than(pos_pred, neg_pred), dtype="float32"))
self.order = pos_cnt / (1e-5 + neg_cnt)
self.metrics = [self.loss, self.order]
def test_input_type_error(self):
main_program = Program()
startup_program = Program()
with program_guard(main_program, startup_program):
src = layers.data(name='data', shape=[1], dtype='float32')
const_value = layers.fill_constant(
[1], dtype='float32', value=123.0)
ifcond = layers.less_than(x=src, y=const_value)
with self.assertRaises(TypeError):
ie = layers.IfElse(set())
with self.assertRaises(TypeError):
ie = layers.IfElse(ifcond, set())
with self.assertRaises(TypeError):
ie = layers.IfElse(ifcond)
with ie.true_block():
true_target = ie.input(src)
true_target = fluid.layers.exp(true_target)
ie.output([])
def increment(self):
enough_steps = layers.less_than(self.increment_every,
self.good_steps + 1)
def increment_step():
layers.increment(self.good_steps)
def maybe_update():
new_scale = self.scale * self.factor
scale_valid = layers.isfinite(new_scale)
def update_scale_and_step():
layers.assign(new_scale, self.scale)
layers.assign(
layers.zeros_like(self.good_steps), self.good_steps)
layers.cond(scale_valid, update_scale_and_step)
layers.cond(enough_steps, maybe_update, increment_step)
def pairwise_loss(self):
"""pairwise model"""
# TODO: for neg_num neg poi, split num should be (neg_num + 1) on dim 0
poi_repr = L.split(self.poi_repr, [1 * self.batch_size, self.neg_num * self.batch_size], dim=0)
pos_repr, neg_repr = poi_repr
# size [-1 x emb_size]
# size [-1*n x emb_size]
prefix_expand = L.reshape(L.expand(self.query_repr, [1, self.neg_num]), [-1, self.hidden_size])
# size [-1*n x 1]
neg_pred_n = self.safe_cosine_sim(neg_repr, prefix_expand)
# size [-1 x 1]
pos_pred = self.safe_cosine_sim(pos_repr, self.query_repr)
cost = self.loss_neg_log_of_pos(pos_pred, L.reshape(neg_pred_n, [-1, self.neg_num]), 15)
self.loss = L.mean(x=cost)
# size [-1 x 1]
neg_avg = L.reduce_mean(L.reshape(neg_pred_n, [-1, self.neg_num]), dim=1, keep_dim=True)
pos_cnt = L.reduce_sum(L.cast(L.greater_than(pos_pred, neg_avg), dtype="float32"))
neg_cnt = L.reduce_sum(L.cast(L.less_than(pos_pred, neg_avg), dtype="float32"))
# equal to positive and negative order
self.order = pos_cnt / (1e-5 + neg_cnt)
self.metrics = [self.loss, self.order]
def build_graph_with_sub_graph(self):
def linear_fc(num):
data = fluid.layers.data(name='image',
shape=[1, 32, 32],
dtype='float32')
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
hidden = data
for _ in six.moves.xrange(num):
hidden = fluid.layers.fc(hidden, size=128, act='relu')
loss = fluid.layers.cross_entropy(input=hidden, label=label)
loss = fluid.layers.mean(loss)
return loss
main_program = Program()
startup_program = Program()
def true_func():
return linear_fc(3)
def false_func():
return linear_fc(5)
with program_guard(main_program, startup_program):
x = layers.fill_constant(shape=[1], dtype='float32', value=0.1)
y = layers.fill_constant(shape=[1], dtype='float32', value=0.23)
pred = layers.less_than(y, x)
out = layers.cond(pred, true_func, false_func)
core_graph = core.Graph(main_program.desc)
# We should create graph for test, otherwise it will throw a
# error that it cannot find the node of "STEP_COUNTER"
graph = IrGraph(core_graph, for_test=True)
sub_graph = graph.get_sub_graph(0)
all_sub_graphs = graph.all_sub_graphs(
for_test=True) # same reason for subgraph
# Should return graph and sub_graphs at the same time. If only return sub_graph, the graph will
# be destructed and the sub_graphs will be empty.
return graph, all_sub_graphs
def build_program(self):
def true_func():
return layers.fill_constant(shape=[1, 2], dtype='int32',
value=1), layers.fill_constant(
shape=[2, 3],
dtype='bool',
value=True)
def false_func():
return layers.fill_constant(shape=[3, 4], dtype='float32',
value=3), layers.fill_constant(
shape=[4, 5],
dtype='int64',
value=2)
main_program = Program()
startup_program = Program()
with program_guard(main_program, startup_program):
x = layers.fill_constant(shape=[1], dtype='float32', value=0.1)
y = layers.fill_constant(shape=[1], dtype='float32', value=0.23)
pred = layers.less_than(x, y)
out = layers.cond(pred, true_func, false_func)
# out is a tuple containing 2 tensors
return main_program, startup_program, out
def decoder_decode(context, is_sparse):
init_state = context
array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
# fill the first element with init_state
state_array = pd.create_array('float32')
pd.array_write(init_state, array=state_array, i=counter)
# ids, scores as memory
ids_array = pd.create_array('int64')
scores_array = pd.create_array('float32')
init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2)
init_scores = pd.data(name="init_scores",
shape=[1],
dtype="float32",
lod_level=2)
pd.array_write(init_ids, array=ids_array, i=counter)
pd.array_write(init_scores, array=scores_array, i=counter)
cond = pd.less_than(x=counter, y=array_len)
while_op = pd.While(cond=cond)
with while_op.block():
pre_ids = pd.array_read(array=ids_array, i=counter)
pre_state = pd.array_read(array=state_array, i=counter)
pre_score = pd.array_read(array=scores_array, i=counter)
# expand the lod of pre_state to be the same with pre_score
pre_state_expanded = pd.sequence_expand(pre_state, pre_score)
pre_ids_emb = pd.embedding(input=pre_ids,
size=[dict_size, word_dim],
dtype='float32',
is_sparse=is_sparse)
# use rnn unit to update rnn
current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb],
size=decoder_size,
act='tanh')
current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score)
# use score to do beam search
current_score = pd.fc(input=current_state_with_lod,
size=target_dict_dim,
act='softmax')
topk_scores, topk_indices = pd.topk(current_score, k=50)
selected_ids, selected_scores = pd.beam_search(pre_ids,
topk_indices,
topk_scores,
beam_size,
end_id=10,
level=0)
pd.increment(x=counter, value=1, in_place=True)
# update the memories
pd.array_write(current_state, array=state_array, i=counter)
pd.array_write(selected_ids, array=ids_array, i=counter)
pd.array_write(selected_scores, array=scores_array, i=counter)
pd.less_than(x=counter, y=array_len, cond=cond)
translation_ids, translation_scores = pd.beam_search_decode(
ids=ids_array, scores=scores_array)
# return init_ids, init_scores
return translation_ids, translation_scores
def decode(context, is_sparse):
init_state = context
array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
# fill the first element with init_state
state_array = pd.create_array('float32')
pd.array_write(init_state, array=state_array, i=counter)
# ids, scores as memory
ids_array = pd.create_array('int64')
scores_array = pd.create_array('float32')
init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2)
init_scores = pd.data(
name="init_scores", shape=[1], dtype="float32", lod_level=2)
pd.array_write(init_ids, array=ids_array, i=counter)
pd.array_write(init_scores, array=scores_array, i=counter)
cond = pd.less_than(x=counter, y=array_len)
while_op = pd.While(cond=cond)
with while_op.block():
pre_ids = pd.array_read(array=ids_array, i=counter)
pre_state = pd.array_read(array=state_array, i=counter)
pre_score = pd.array_read(array=scores_array, i=counter)
# expand the lod of pre_state to be the same with pre_score
pre_state_expanded = pd.sequence_expand(pre_state, pre_score)
pre_ids_emb = pd.embedding(
input=pre_ids,
size=[dict_size, word_dim],
dtype='float32',
is_sparse=is_sparse)
# use rnn unit to update rnn
current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb],
size=decoder_size,
act='tanh')
current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score)
# use score to do beam search
current_score = pd.fc(input=current_state_with_lod,
size=target_dict_dim,
act='softmax')
topk_scores, topk_indices = pd.topk(current_score, k=topk_size)
selected_ids, selected_scores = pd.beam_search(
pre_ids, topk_indices, topk_scores, beam_size, end_id=10, level=0)
pd.increment(x=counter, value=1, in_place=True)
# update the memories
pd.array_write(current_state, array=state_array, i=counter)
pd.array_write(selected_ids, array=ids_array, i=counter)
pd.array_write(selected_scores, array=scores_array, i=counter)
pd.less_than(x=counter, y=array_len, cond=cond)
translation_ids, translation_scores = pd.beam_search_decode(
ids=ids_array, scores=scores_array)
# return init_ids, init_scores
return translation_ids, translation_scores
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 infilling_decode(self):
if self.task_type == "dialog":
emb_num = 4
else:
emb_num = 3
input_shapes = [[-1, self.max_seq_len, 1]] * emb_num + \
[[-1, self.max_seq_len, self.max_seq_len]]
input_dtypes = ['int64'] * emb_num + ['float32']
input_lod_levels = [0] * emb_num + [0]
shapes = input_shapes + [[-1, self.max_seq_len, 1],
[-1, self.max_seq_len, 1], [-1, 1], [-1],
[-1, 1, self.max_seq_len], [-1, 1]]
dtypes = input_dtypes + [
'int64', 'int64', 'float32', 'int32', 'float32', 'int64'
]
lod_levels = input_lod_levels + [2, 2, 2, 0, 0, 0]
inputs = self.to_ternsor(shapes, dtypes, lod_levels)
pyreader = fluid.io.DataLoader.from_generator(feed_list=inputs,
capacity=50,
iterable=False)
emb_ids = {}
for key, value in zip(self.emb_keys, inputs[:emb_num]):
emb_ids[key] = value
input_mask = inputs[emb_num]
tgt_ids, tgt_pos, init_scores, parent_idx, tgt_input_mask, data_ids = inputs[
-6:]
ernie = ErnieModel(emb_ids=emb_ids,
input_mask=input_mask,
config=self.ernie_config,
use_fp16=self.use_fp16,
task_type=self.task_type,
decoding=True,
gather_idx=parent_idx)
max_len = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=self.max_dec_len,
force_cpu=True)
step_idx = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=0,
force_cpu=True)
pos_idx = layers.fill_constant(shape=[1],
dtype=tgt_ids.dtype,
value=1,
force_cpu=True)
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
ids = layers.array_write(layers.reshape(tgt_ids, (-1, 1)), step_idx)
pos_biases = layers.array_write(layers.reshape(tgt_pos, (-1, 1)),
step_idx)
scores = layers.array_write(init_scores, step_idx)
tgt_masks = layers.array_write(tgt_input_mask, step_idx)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
pos_bias = layers.array_read(array=pos_biases, i=step_idx)
pos_bias = layers.gather(input=pos_bias, index=parent_idx)
tmp_mask = layers.array_read(tgt_masks, i=step_idx)
def gen_batch_like(value,
dtype="int64",
shape=[-1, 1, 1],
is_scalar=True):
if is_scalar:
return layers.fill_constant_batch_size_like(
input=parent_idx,
value=value,
shape=shape,
dtype=dtype)
else:
return layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=parent_idx,
value=1,
shape=shape,
dtype=dtype),
y=value,
axis=0)
tmp_mask = layers.gather(input=tmp_mask, index=parent_idx)
append_0_mask = gen_batch_like(0.0, dtype=tmp_mask.dtype)
append_1_mask = gen_batch_like(1.0, dtype=tmp_mask.dtype)
tmp_mask = layers.concat([tmp_mask, append_1_mask], axis=2)
pre_mask = layers.concat([tmp_mask, append_0_mask], axis=2)
cur_mask = layers.concat([tmp_mask, append_1_mask], axis=2)
cur_ids = gen_batch_like(self.attn_id)
pre_pos = gen_batch_like(step_idx, is_scalar=False)
cur_pos = gen_batch_like(pos_idx, is_scalar=False)
if self.continuous_position:
pre_pos = pre_pos + pos_bias
cur_pos = cur_pos + pos_bias
dec_emb_ids = {
"word_embedding": layers.concat([pre_ids, cur_ids], axis=1),
"pos_embedding": layers.concat([pre_pos, cur_pos], axis=1)
}
if self.task_type == "dialog":
role_ids = gen_batch_like(0)
turn_ids = gen_batch_like(0)
dec_emb_ids["role_embedding"] = layers.concat(
[role_ids, role_ids], axis=1)
dec_emb_ids["turn_embedding"] = layers.concat(
[turn_ids, turn_ids], axis=1)
else:
sent_ids = gen_batch_like(self.tgt_type_id)
dec_emb_ids["sent_embedding"] = layers.concat(
[sent_ids, sent_ids], axis=1)
dec_mask = layers.concat([pre_mask, cur_mask], axis=1)
dec_out = ernie.encode(dec_emb_ids,
dec_mask,
parent_idx,
remove_query=True)
fc_out = self.cal_logit(dec_out[:, 1:, :], None)
topk_scores, topk_indices = layers.topk(
input=layers.softmax(fc_out), k=self.beam_size)
pre_lenpen = layers.pow(
(5.0 + layers.cast(step_idx, pre_scores.dtype)) / 6.0,
self.length_penalty)
cur_lenpen = layers.pow(
(5.0 + layers.cast(pos_idx, pre_scores.dtype)) / 6.0,
self.length_penalty)
accu_scores = layers.elementwise_add(x=layers.log(topk_scores),
y=pre_scores * pre_lenpen,
axis=0) / cur_lenpen
topk_indices = layers.lod_reset(topk_indices, pre_ids)
accu_scores = layers.lod_reset(accu_scores, pre_ids)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=self.beam_size,
end_id=self.eos_idx,
return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
layers.increment(x=pos_idx, value=1.0, in_place=True)
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.array_write(tmp_mask, i=step_idx, array=tgt_masks)
layers.array_write(pos_bias, i=step_idx, array=pos_biases)
layers.assign(gather_idx, parent_idx)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=self.beam_size, end_id=self.eos_idx)
graph_vars = {
"finished_ids": finished_ids,
"finished_scores": finished_scores,
"data_ids": data_ids
}
for k, v in graph_vars.items():
v.persistable = True
return pyreader, graph_vars
def test_simple_forward(self):
d0 = layers.data("d0",
shape=[10],
append_batch_size=False,
dtype='float32')
d1 = layers.data("d1",
shape=[10],
append_batch_size=False,
dtype='float32')
d2 = layers.data("d2",
shape=[10],
append_batch_size=False,
dtype='float32')
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = True
init = layers.zeros(shape=[10], dtype='float32')
mem_array = layers.array_write(x=init, i=i)
data_array = layers.array_write(x=d0, i=i)
i = layers.increment(i)
layers.array_write(d1, i, array=data_array)
i = layers.increment(i)
layers.array_write(d2, i, array=data_array)
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = True
array_len = layers.fill_constant(shape=[1], dtype='int64', value=1)
array_len.stop_gradient = True
cond = layers.less_than(x=i, y=array_len)
j = layers.fill_constant(shape=[1], dtype='int64', value=1)
j.stop_gradient = True
array_len2 = layers.fill_constant(shape=[1], dtype='int64', value=3)
array_len2.stop_gradient = True
cond2 = layers.less_than(x=j, y=array_len2)
while_op = layers.While(cond=cond)
while_op2 = layers.While(cond=cond2)
with while_op.block():
d = layers.array_read(array=data_array, i=i)
prev = layers.array_read(array=mem_array, i=i)
result = layers.sums(input=[d, prev])
i = layers.increment(x=i, in_place=True)
layers.array_write(result, i=i, array=mem_array)
layers.less_than(x=i, y=array_len, cond=cond)
with while_op2.block():
d2 = layers.array_read(array=data_array, i=j)
prev2 = layers.array_read(array=mem_array, i=j)
result2 = layers.sums(input=[d2, prev2])
j = layers.increment(x=j, in_place=True)
layers.array_write(result2, i=j, array=mem_array)
layers.less_than(x=j, y=array_len2, cond=cond2)
sum_result = layers.array_read(array=mem_array, i=j)
loss = layers.mean(sum_result)
append_backward(loss)
cpu = core.CPUPlace()
exe = Executor(cpu)
d = []
for i in range(3):
d.append(numpy.random.random(size=[10]).astype('float32'))
outs = exe.run(feed={
'd0': d[0],
'd1': d[1],
'd2': d[2]
},
fetch_list=[sum_result])
self.assertAlmostEqual(numpy.sum(d), numpy.sum(outs[0]), delta=0.01)
def inference(self, model, inputs, outputs):
"""
Run inference.
Args:
inputs(dict): Its key is input name(str) and its value is a Variable.
model(object): A generate model. Need to implement `_generation_network` and `_calc_logits`.
Returns:
dict(str:Variable): Its key is output name(str) and its value is a Variable.
"""
# prepare while loop
max_len = layers.fill_constant(
shape=[1], dtype="int64", value=self.max_dec_len, force_cpu=True)
min_len = layers.fill_constant(
shape=[1], dtype="int64", value=self.min_dec_len, force_cpu=True)
step_idx = layers.fill_constant(
shape=[1], dtype="int64", value=0, force_cpu=True)
ids = layers.array_write(layers.reshape(inputs["tgt_ids"], (-1, 1)), step_idx)
pos_biases = layers.array_write(layers.reshape(inputs["tgt_pos"], (-1, 1)), step_idx)
scores = layers.array_write(inputs["init_score"], step_idx)
tgt_generation_mask = layers.array_write(inputs["tgt_generation_mask"], step_idx)
parent_idx = inputs["parent_idx"]
if self.decoding_strategy == "beam_search":
beam_size = self.beam_size
else:
beam_size = 1
eos_penalty = np.zeros(self.vocab_size, dtype="float32")
eos_penalty[self.eos_id] = -1e9
eos_penalty = layers.assign(eos_penalty)
token_penalty = np.zeros(self.vocab_size, dtype="float32")
token_penalty[self.unk_id] = -1e9
if self.mask_id >= 0:
token_penalty[self.mask_id] = -1e9
token_penalty = layers.assign(token_penalty)
# start while loop
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
pos_bias = layers.array_read(array=pos_biases, i=step_idx)
pos_bias = layers.gather(input=pos_bias, index=parent_idx)
tmp_tgt_generation_mask = layers.array_read(tgt_generation_mask, i=step_idx)
dtype = tmp_tgt_generation_mask.dtype
append_mask = layers.fill_constant_batch_size_like(
input=pre_ids,
value=1.0,
shape=[-1, 1, 1],
dtype=dtype)
tmp_tgt_generation_mask = layers.concat([tmp_tgt_generation_mask, append_mask], axis=2)
pre_mask = tmp_tgt_generation_mask = layers.gather(input=tmp_tgt_generation_mask, index=parent_idx)
pre_sent = layers.fill_constant_batch_size_like(
input=pre_mask,
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype)
if self.continuous_position:
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_mask,
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype), y=step_idx, axis=0) + pos_bias
else:
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_mask,
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype), y=step_idx, axis=0)
if self.use_role:
pre_role = layers.fill_constant_batch_size_like(
input=pre_mask,
value=0,
shape=[-1, 1, 1],
dtype=pre_ids.dtype)
else:
pre_role = None
dec_out, _ = model._generation_network(
token_ids=pre_ids,
type_ids=pre_sent,
pos_ids=pre_pos,
role_ids=pre_role,
generation_mask=tmp_tgt_generation_mask,
gather_idx=parent_idx)
logits = model._calc_logits(dec_out)
# ignore unk and mask token
if self.ignore_unk:
logits = layers.elementwise_add(logits, token_penalty, axis=1)
# min dec length
min_len_cond = layers.less_than(x=step_idx, y=min_len)
def min_len_penalty():
"""Plus minimum length penalty."""
return layers.elementwise_add(logits, eos_penalty, axis=1)
def no_penalty():
"""No penalty."""
return logits
logits = layers.case([(min_len_cond, min_len_penalty)], default=no_penalty)
# get probs
probs = layers.softmax(logits / self.temperature)
if self.decoding_strategy == "beam_search":
topk_scores, topk_indices = layers.topk(
input=probs, k=beam_size)
else:
if self.decoding_strategy.startswith("sampling"):
sampling_ids = layers.sampling_id(probs, dtype="int")
elif self.decoding_strategy.startswith("topk_sampling"):
topk_probs, _ = layers.topk(input=probs, k=self.topk)
ge_cond = layers.cast(
layers.greater_equal(
probs,
layers.unsqueeze(topk_probs[:, -1], [1])),
"float32")
old_probs = probs
probs = probs * ge_cond / layers.reduce_sum(topk_probs, dim=-1, keep_dim=True)
sampling_ids = layers.sampling_id(probs, dtype="int")
probs = old_probs
else:
raise ValueError(self.decoding_strategy)
sampling_scores = layers.one_hot(
layers.unsqueeze(sampling_ids, [1]), probs.shape[1]
)
sampling_scores = sampling_scores * probs - (1 - sampling_scores) * 1e3
topk_scores, topk_indices = layers.topk(
input=sampling_scores, k=1)
pre_len = layers.cast(step_idx, "float32")
layers.increment(x=step_idx, value=1.0, in_place=True)
cur_len = layers.cast(step_idx, "float32")
# update scores
if self.length_average:
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores), y=pre_scores * pre_len, axis=0) / cur_len
elif self.length_penalty > 0:
pre_lp = layers.pow((5 + pre_len) / 6, self.length_penalty)
cur_lp = layers.pow((5 + cur_len) / 6, self.length_penalty)
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores), y=pre_scores * pre_lp, axis=0) / cur_lp
else:
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores), y=pre_scores, axis=0)
topk_indices = layers.lod_reset(topk_indices, pre_ids)
accu_scores = layers.lod_reset(accu_scores, pre_ids)
selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
end_id=self.eos_id,
return_parent_idx=True)
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.array_write(pre_mask, i=step_idx, array=tgt_generation_mask)
layers.array_write(pos_bias, i=step_idx, array=pos_biases)
layers.assign(gather_idx, parent_idx)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=beam_size, end_id=self.eos_id)
predictions = {
"finished_ids": finished_ids,
"finished_scores": finished_scores,
"token_ids": inputs["token_ids"],
"data_id": inputs["data_id"]
}
return predictions
def decoder_decode(context, is_sparse):
init_state = context
array_len = pd.fill_constant(shape=[1], dtype='int64', value=max_length)
counter = pd.zeros(shape=[1], dtype='int64', force_cpu=True)
# fill the first element with init_state
state_array = pd.create_array('float32')
pd.array_write(init_state, array=state_array, i=counter)
# ids, scores as memory
ids_array = pd.create_array('int64')
scores_array = pd.create_array('float32')
init_ids = pd.data(name="init_ids", shape=[1], dtype="int64", lod_level=2)
init_scores = pd.data(
name="init_scores", shape=[1], dtype="float32", lod_level=2)
pd.array_write(init_ids, array=ids_array, i=counter)
pd.array_write(init_scores, array=scores_array, i=counter)
cond = pd.less_than(x=counter, y=array_len)
while_op = pd.While(cond=cond)
with while_op.block():
pre_ids = pd.array_read(array=ids_array, i=counter)
pre_state = pd.array_read(array=state_array, i=counter)
pre_score = pd.array_read(array=scores_array, i=counter)
# expand the recursive_sequence_lengths of pre_state to be the same with pre_score
pre_state_expanded = pd.sequence_expand(pre_state, pre_score)
pre_ids_emb = pd.embedding(
input=pre_ids,
size=[dict_size, word_dim],
dtype='float32',
is_sparse=is_sparse)
# use rnn unit to update rnn
current_state = pd.fc(input=[pre_state_expanded, pre_ids_emb],
size=decoder_size,
act='tanh')
current_state_with_lod = pd.lod_reset(x=current_state, y=pre_score)
# use score to do beam search
current_score = pd.fc(input=current_state_with_lod,
size=target_dict_dim,
act='softmax')
topk_scores, topk_indices = pd.topk(current_score, k=beam_size)
# calculate accumulated scores after topk to reduce computation cost
accu_scores = pd.elementwise_add(
x=pd.log(topk_scores), y=pd.reshape(
pre_score, shape=[-1]), axis=0)
selected_ids, selected_scores = pd.beam_search(
pre_ids,
pre_score,
topk_indices,
accu_scores,
beam_size,
end_id=10,
level=0)
pd.increment(x=counter, value=1, in_place=True)
# update the memories
pd.array_write(current_state, array=state_array, i=counter)
pd.array_write(selected_ids, array=ids_array, i=counter)
pd.array_write(selected_scores, array=scores_array, i=counter)
# update the break condition: up to the max length or all candidates of
# source sentences have ended.
length_cond = pd.less_than(x=counter, y=array_len)
finish_cond = pd.logical_not(pd.is_empty(x=selected_ids))
pd.logical_and(x=length_cond, y=finish_cond, out=cond)
translation_ids, translation_scores = pd.beam_search_decode(
ids=ids_array, scores=scores_array, beam_size=beam_size, end_id=10)
# return init_ids, init_scores
return translation_ids, translation_scores
def beam_search():
max_len = layers.fill_constant(
shape=[1], dtype=start_tokens.dtype, value=max_out_len)
step_idx = layers.fill_constant(
shape=[1], dtype=start_tokens.dtype, value=0)
cond = layers.less_than(x=step_idx, y=max_len)
while_op = layers.While(cond)
# array states will be stored for each step.
ids = layers.array_write(start_tokens, step_idx)
scores = layers.array_write(init_scores, step_idx)
# cell states will be overwrited at each step.
# caches contains states of history steps to reduce redundant
# computation in decoder.
caches = [{
"k": layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, 0, d_model],
dtype=enc_output.dtype,
value=0),
"v": layers.fill_constant_batch_size_like(
input=start_tokens,
shape=[-1, 0, d_model],
dtype=enc_output.dtype,
value=0)
} for i in range(n_layer)]
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
pre_scores = layers.array_read(array=scores, i=step_idx)
# sequence_expand can gather sequences according to lod thus can be
# used in beam search to sift states corresponding to selected ids.
pre_src_attn_bias = layers.sequence_expand(
x=trg_src_attn_bias, y=pre_scores)
pre_enc_output = layers.sequence_expand(x=enc_output, y=pre_scores)
pre_caches = [{
"k": layers.sequence_expand(
x=cache["k"], y=pre_scores),
"v": layers.sequence_expand(
x=cache["v"], y=pre_scores),
} for cache in caches]
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
input=pre_enc_output, # cann't use pre_ids here since it has lod
value=1,
shape=[-1, 1],
dtype=pre_ids.dtype),
y=layers.increment(
x=step_idx, value=1.0, in_place=False),
axis=0)
logits = wrap_decoder(
trg_vocab_size,
max_in_len,
n_layer,
n_head,
d_key,
d_value,
d_model,
d_inner_hid,
dropout_rate,
weight_sharing,
dec_inputs=(
pre_ids, pre_pos, None, pre_src_attn_bias, trg_data_shape,
slf_attn_pre_softmax_shape, slf_attn_post_softmax_shape,
src_attn_pre_softmax_shape, src_attn_post_softmax_shape),
enc_output=pre_enc_output,
caches=pre_caches)
topk_scores, topk_indices = layers.topk(
input=layers.softmax(logits), k=beam_size)
accu_scores = layers.elementwise_add(
x=layers.log(topk_scores),
y=layers.reshape(
pre_scores, shape=[-1]),
axis=0)
# beam_search op uses lod to distinguish branches.
topk_indices = layers.lod_reset(topk_indices, pre_ids)
selected_ids, selected_scores = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
end_id=eos_idx)
layers.increment(x=step_idx, value=1.0, in_place=True)
# update states
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
layers.assign(pre_src_attn_bias, trg_src_attn_bias)
layers.assign(pre_enc_output, enc_output)
for i in range(n_layer):
layers.assign(pre_caches[i]["k"], caches[i]["k"])
layers.assign(pre_caches[i]["v"], caches[i]["v"])
layers.assign(
layers.elementwise_add(
x=slf_attn_pre_softmax_shape,
y=attn_pre_softmax_shape_delta),
slf_attn_pre_softmax_shape)
layers.assign(
layers.elementwise_add(
x=slf_attn_post_softmax_shape,
y=attn_post_softmax_shape_delta),
slf_attn_post_softmax_shape)
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
finished_ids, finished_scores = layers.beam_search_decode(
ids, scores, beam_size=beam_size, end_id=eos_idx)
return finished_ids, finished_scores