def test_forward(self):
data = layers.data(name='X', shape=[1], dtype='float32')
data.stop_gradient = False
cond = layers.ConditionalBlock(inputs=[data])
out = layers.create_tensor(dtype='float32')
with cond.block():
hidden = layers.fc(input=data, size=10)
layers.assign(hidden, out)
cpu = core.CPUPlace()
exe = Executor(cpu)
exe.run(default_startup_program())
x = numpy.random.random(size=(10, 1)).astype('float32')
outs = exe.run(feed={'X': x}, fetch_list=[out])[0]
print outs
loss = layers.mean(x=out)
append_backward_ops(loss=loss)
outs = exe.run(feed={'X': x},
fetch_list=[
default_main_program().block(0).var(data.name +
"@GRAD")
])[0]
print outs
def test_grad(self):
place = core.CPUPlace()
program = Program()
x = layers.data(name='x',
shape=[1],
dtype='float32',
main_program=program,
stop_gradient=False)
table = layers.lod_rank_table(x, level=0, main_program=program)
array = layers.lod_tensor_to_array(x, table, main_program=program)
result = layers.array_to_lod_tensor(array, table, main_program=program)
mean = layers.mean(x=result, main_program=program)
append_backward_ops(mean)
tensor = core.LoDTensor()
tensor.set(numpy.arange(10).reshape(10, 1).astype('float32'), place)
tensor.set_lod([[0, 3, 9, 10]])
g_vars = program.global_block().var(x.name + "@GRAD")
exe = Executor(place)
g_out = [
numpy.array(item).sum() for item in exe.run(program,
feed={'x': tensor},
fetch_list=[g_vars],
return_numpy=False)
]
g_out_sum = numpy.array(g_out).sum()
self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
def test_grad(self):
place = core.CPUPlace()
program = Program()
x = layers.data(
name='x',
shape=[1],
dtype='float32',
main_program=program,
stop_gradient=False)
y = layers.data(
name='y',
shape=[1],
dtype='bool',
main_program=program,
stop_gradient=False)
level = 0
out_true, out_false = layers.split_lod_tensor(
input=x, mask=y, level=level, main_program=program)
out = layers.merge_lod_tensor(
in_true=out_true,
in_false=out_false,
mask=y,
x=x,
level=level,
main_program=program)
mean = layers.mean(x=out, main_program=program)
append_backward_ops(mean)
tensor = core.LoDTensor()
tensor.set(np.arange(10).reshape(10, 1).astype('float32'), place)
tensor.set_lod([[0, 3, 9, 10]])
mask_np = np.array([0, 1, 0]).astype('bool')
mask_np = np.expand_dims(mask_np, axis=1)
mask = core.LoDTensor()
mask.set(mask_np, place)
exe = Executor(place)
scope = core.Scope()
g_vars = program.global_block().var(x.name + "@GRAD")
g_out = [
item.sum()
for item in map(np.array,
exe.run(program,
feed={'x': tensor,
'y': mask},
fetch_list=[g_vars],
scope=scope,
return_numpy=False))
]
g_out_sum = np.array(g_out).sum()
self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
def test_forward(self):
data = layers.data(name='X', shape=[1], dtype='float32')
data.stop_gradient = False
cond = layers.ConditionalBlock(inputs=[data])
out = layers.create_tensor(dtype='float32')
with cond.block():
hidden = layers.fc(input=data, size=10)
layers.assign(hidden, out)
cpu = core.CPUPlace()
exe = Executor(cpu)
exe.run(default_startup_program())
x = numpy.random.random(size=(10, 1)).astype('float32')
outs = exe.run(feed={'X': x}, fetch_list=[out])[0]
print outs
loss = layers.mean(x=out)
append_backward_ops(loss=loss)
outs = exe.run(
feed={'X': x},
fetch_list=[
default_main_program().block(0).var(data.name + "@GRAD")
])[0]
print outs
def test_grad(self):
place = core.CPUPlace()
program = Program()
x = layers.data(
name='x',
shape=[1],
dtype='float32',
main_program=program,
stop_gradient=False)
table = layers.lod_rank_table(x, level=0, main_program=program)
array = layers.lod_tensor_to_array(x, table, main_program=program)
result = layers.array_to_lod_tensor(array, table, main_program=program)
mean = layers.mean(x=result, main_program=program)
append_backward_ops(mean)
tensor = core.LoDTensor()
tensor.set(numpy.arange(10).reshape(10, 1).astype('float32'), place)
tensor.set_lod([[0, 3, 9, 10]])
g_vars = program.global_block().var(x.name + "@GRAD")
exe = Executor(place)
g_out = [
numpy.array(item).sum()
for item in exe.run(program,
feed={'x': tensor},
fetch_list=[g_vars],
return_numpy=False)
]
g_out_sum = numpy.array(g_out).sum()
self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
def test_shrink_rnn_memory(self):
x = layers.data('x', shape=[100], dtype='float32')
x.stop_gradient = False
table = layers.lod_rank_table(x=x)
i = layers.zeros(dtype='int64', shape=[1])
mem1 = layers.shrink_memory(x=x, i=i, table=table)
i = layers.increment(x=i)
i.stop_gradient = True
mem2 = layers.shrink_memory(x=mem1, i=i, table=table)
i = layers.increment(x=i)
i.stop_gradient = True
mem3 = layers.shrink_memory(x=mem2, i=i, table=table)
cpu = core.CPUPlace()
tensor = core.LoDTensor()
tensor.set_lod([[0, 2, 5, 6]])
tensor_np = numpy.random.random(size=(3, 100)).astype('float32')
tensor.set(tensor_np, cpu)
exe = Executor(cpu)
outs = exe.run(feed={'x': tensor}, fetch_list=[mem1, mem2, mem3])
self.assertTrue(numpy.allclose(tensor_np[0:3], outs[0]))
self.assertTrue(numpy.allclose(tensor_np[0:2], outs[1]))
self.assertTrue(numpy.allclose(tensor_np[0:1], outs[2]))
mem3_mean = layers.mean(x=mem3)
append_backward_ops(loss=mem3_mean)
x_grad = exe.run(
feed={'x': tensor},
fetch_list=[main_program.global_block().var('x@GRAD')])[0]
self.assertAlmostEqual(1.0, x_grad.sum(), delta=0.1)
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(x=sum_result)
append_backward_ops(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_backward(self):
self.check_forward()
append_backward_ops(self.output)
ana_grad = [np.array(x) for x in self.backward()]
num_grad = self.get_numerical_gradient()
for idx, name in enumerate(self.data_field):
self.assertEqual(num_grad[idx].shape, ana_grad[idx].shape)
self.assertTrue(
np.isclose(num_grad[idx], ana_grad[idx], rtol=0.1).all())
def test_backward(self):
self.check_forward()
append_backward_ops(self.output)
ana_grad = [np.array(x) for x in self.backward()]
num_grad = self.get_numerical_gradient()
for idx, name in enumerate(self.data_field):
self.assertEqual(num_grad[idx].shape, ana_grad[idx].shape)
self.assertTrue(
np.isclose(
num_grad[idx], ana_grad[idx], rtol=0.1).all())
def test_l2decay_regularizer(self):
program = framework.Program()
block = program.global_block()
mul_x = block.create_parameter(
dtype="float32",
shape=[5, 10],
lod_level=0,
name="mul.x",
regularizer=regularizer.L2DecayRegularizer(0.5))
self.assertTrue(mul_x.regularizer is not None)
self.assertTrue(
isinstance(mul_x.regularizer, regularizer.L2DecayRegularizer))
mul_y = block.create_var(
dtype="float32", shape=[10, 8], lod_level=0, name="mul.y")
mul_out = block.create_var(
dtype="float32", shape=[5, 8], lod_level=0, name="mul.out")
block.append_op(
type="mul",
inputs={"X": mul_x,
"Y": mul_y},
outputs={"Out": mul_out},
attrs={"x_num_col_dims": 1})
mean_out = block.create_var(
dtype="float32", shape=[1], lod_level=0, name="mean.out")
block.append_op(
type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
params_grads = append_backward_ops(mean_out)
self.assertEqual(len(params_grads), 1)
count_ops = len(block.ops)
params_grads = optimizer.append_regularization_ops(params_grads)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(block.ops), count_ops + 2)
self.assertEqual(block.ops[-1].type, 'elementwise_add')
self.assertEqual(block.ops[-2].type, 'scale')
def test_adamax_optimizer(self):
init_program = framework.Program()
program = framework.Program()
block = program.global_block()
mul_x = block.create_parameter(
dtype="float32", shape=[5, 10], lod_level=0, name="mul.x")
mul_y = block.create_var(
dtype="float32", shape=[10, 8], lod_level=0, name="mul.y")
mul_out = block.create_var(
dtype="float32", shape=[5, 8], lod_level=0, name="mul.out")
block.append_op(
type="mul",
inputs={"X": mul_x,
"Y": mul_y},
outputs={"Out": mul_out},
attrs={"x_num_col_dims": 1})
mean_out = block.create_var(
dtype="float32", shape=[1], lod_level=0, name="mean.out")
block.append_op(
type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
learning_rate = 0.01
adamax_optimizer = self.MockAdamax(
learning_rate=learning_rate, beta1=0.9, beta2=0.999)
params_grads = append_backward_ops(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(adamax_optimizer.get_accumulators()), 0)
opts = adamax_optimizer.create_optimization_pass(params_grads, mul_out,
init_program)
self.assertEqual(len(opts), 2)
adam_op = opts[0]
self.assertEqual(adam_op.type, "adamax")
# Check accumulators
accumulators = adamax_optimizer.get_accumulators()
self.assertEqual(len(accumulators), 2)
self.assertTrue(adamax_optimizer.get_moment_str() in accumulators)
self.assertTrue(adamax_optimizer.get_inf_norm_str() in accumulators)
moment_acc = accumulators[adamax_optimizer.get_moment_str()]
inf_norm_acc = accumulators[adamax_optimizer.get_inf_norm_str()]
self.assertEqual(len(moment_acc), 1)
self.assertEqual(len(inf_norm_acc), 1)
self.assertTrue(mul_x.name in moment_acc)
self.assertTrue(mul_x.name in inf_norm_acc)
# Check init_program
init_ops = init_program.global_block().ops
self.assertEqual(len(init_ops), 4)
self.assertEqual(init_ops[0].type, "fill_constant")
self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
def test_vanilla_momentum_optimizer(self):
init_program = framework.Program()
program = framework.Program()
block = program.global_block()
mul_x = block.create_parameter(
dtype="float32", shape=[5, 10], lod_level=0, name="mul.x")
mul_y = block.create_var(
dtype="float32", shape=[10, 8], lod_level=0, name="mul.y")
mul_out = block.create_var(
dtype="float32", shape=[5, 8], lod_level=0, name="mul.out")
block.append_op(
type="mul",
inputs={"X": mul_x,
"Y": mul_y},
outputs={"Out": mul_out},
attrs={"x_num_col_dims": 1})
learning_rate = 0.01
momentum_optimizer = self.MockMomentum(
learning_rate=learning_rate, momentum=0.2)
mean_out = block.create_var(
dtype="float32", shape=[1], lod_level=0, name="mean.out")
block.append_op(
type="mean", inputs={"X": mul_out}, outputs={"Out": mean_out})
params_grads = append_backward_ops(mean_out)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(momentum_optimizer.get_accumulators()), 0)
opts = momentum_optimizer.create_optimization_pass(
params_grads, mul_out, init_program)
self.assertEqual(len(opts), 1)
sgd_op = opts[0]
self.assertEqual(sgd_op.type, "momentum")
self.assertFalse(sgd_op.attr('use_nesterov'))
# Check accumulators
accumulators = momentum_optimizer.get_accumulators()
self.assertEqual(len(accumulators), 1)
self.assertTrue(momentum_optimizer.get_velocity_str() in accumulators)
velocity_acc = accumulators[momentum_optimizer.get_velocity_str()]
self.assertEqual(len(velocity_acc), 1)
self.assertTrue(mul_x.name in velocity_acc)
# Check init_program
init_ops = init_program.global_block().ops
self.assertEqual(len(init_ops), 2)
self.assertEqual(init_ops[0].type, "fill_constant")
self.assertAlmostEqual(init_ops[0].attr('value'), learning_rate)
self.assertEqual(init_ops[1].type, "fill_constant")
self.assertAlmostEqual(init_ops[1].attr('value'), 0.0)
def test_l2decay_regularizer(self):
program = framework.Program()
block = program.global_block()
mul_x = block.create_parameter(
dtype="float32",
shape=[5, 10],
lod_level=0,
name="mul.x",
regularizer=regularizer.L1DecayRegularizer(0.5))
self.assertTrue(mul_x.regularizer is not None)
self.assertTrue(
isinstance(mul_x.regularizer, regularizer.L1DecayRegularizer))
mul_y = block.create_var(dtype="float32",
shape=[10, 8],
lod_level=0,
name="mul.y")
mul_out = block.create_var(dtype="float32",
shape=[5, 8],
lod_level=0,
name="mul.out")
block.append_op(type="mul",
inputs={
"X": mul_x,
"Y": mul_y
},
outputs={"Out": mul_out},
attrs={"x_num_col_dims": 1})
mean_out = block.create_var(dtype="float32",
shape=[1],
lod_level=0,
name="mean.out")
block.append_op(type="mean",
inputs={"X": mul_out},
outputs={"Out": mean_out})
params_grads = append_backward_ops(mean_out)
self.assertEqual(len(params_grads), 1)
count_ops = len(block.ops)
params_grads = optimizer.append_regularization_ops(params_grads)
self.assertEqual(len(params_grads), 1)
self.assertEqual(len(block.ops), count_ops + 3)
self.assertEqual(block.ops[-1].type, 'elementwise_add')
self.assertEqual(block.ops[-2].type, 'scale')
self.assertEqual(block.ops[-3].type, 'sign')
def test_read_write(self):
x = [
layers.data(name='x0', shape=[100]),
layers.data(name='x1', shape=[100]),
layers.data(name='x2', shape=[100])
]
for each_x in x:
each_x.stop_gradient = False
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = False
arr = layers.array_write(x=x[0], i=i)
i = layers.increment(x=i)
arr = layers.array_write(x=x[1], i=i, array=arr)
i = layers.increment(x=i)
arr = layers.array_write(x=x[2], i=i, array=arr)
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = False
a0 = layers.array_read(array=arr, i=i)
i = layers.increment(x=i)
a1 = layers.array_read(array=arr, i=i)
i = layers.increment(x=i)
a2 = layers.array_read(array=arr, i=i)
mean_a0 = layers.mean(x=a0)
mean_a1 = layers.mean(x=a1)
mean_a2 = layers.mean(x=a2)
a_sum = layers.sums(input=[mean_a0, mean_a1, mean_a2])
mean_x0 = layers.mean(x=x[0])
mean_x1 = layers.mean(x=x[1])
mean_x2 = layers.mean(x=x[2])
x_sum = layers.sums(input=[mean_x0, mean_x1, mean_x2])
scope = core.Scope()
cpu = core.CPUPlace()
exe = Executor(cpu)
tensor = numpy.random.random(size=(100, 100)).astype('float32')
outs = exe.run(feed={
'x0': tensor,
'x1': tensor,
'x2': tensor
},
fetch_list=[a_sum, x_sum],
scope=scope)
self.assertEqual(outs[0], outs[1])
total_sum = layers.sums(input=[a_sum, x_sum])
total_sum_scaled = layers.scale(x=total_sum, scale=1 / 6.0)
append_backward_ops(total_sum_scaled)
g_vars = map(default_main_program().global_block().var,
[each_x.name + "@GRAD" for each_x in x])
g_out = [
item.sum() for item in exe.run(feed={
'x0': tensor,
'x1': tensor,
'x2': tensor
},
fetch_list=g_vars)
]
g_out_sum = numpy.array(g_out).sum()
# since our final gradient is 1 and the neural network are all linear
# with mean_op.
# the input gradient should also be 1
self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
def _get_gradient(self, input_to_check, place, output_names, no_grad_set):
prog = Program()
block = prog.global_block()
inputs_with_np = {
key: value
for (key, value) in OpTest._create_var_descs_(
block, getattr(self, 'inputs', {}))
}
outputs_with_np = {
key: val
for (key, val) in OpTest._create_var_descs_(
block, getattr(self, 'outputs', {}))
}
inputs = {
k: [item[0] for item in inputs_with_np[k]]
for k in inputs_with_np
}
outputs = {
k: [item[0] for item in outputs_with_np[k]]
for k in outputs_with_np
}
op = block.append_op(
type=self.op_type,
inputs=inputs,
outputs=outputs,
attrs=getattr(self, 'attrs', {}))
# infer variable type and infer shape in compile-time
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
mean_inputs = map(block.var, output_names)
if len(mean_inputs) == 1:
loss = block.create_var(dtype=mean_inputs[0].dtype, shape=[1])
op = block.append_op(
inputs={"X": mean_inputs}, outputs={"Out": loss}, type='mean')
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
else:
avg_sum = []
for cur_loss in mean_inputs:
cur_avg_loss = block.create_var(dtype=cur_loss.dtype, shape=[1])
op = block.append_op(
inputs={"X": [cur_loss]},
outputs={"Out": [cur_avg_loss]},
type="mean")
op.desc.infer_var_type(block.desc)
op.desc.infer_shape(block.desc)
avg_sum.append(cur_avg_loss)
loss_sum = block.create_var(dtype=avg_sum[0].dtype, shape=[1])
op_sum = block.append_op(
inputs={"X": avg_sum}, outputs={"Out": loss_sum}, type='sum')
op_sum.desc.infer_var_type(block.desc)
op_sum.desc.infer_shape(block.desc)
loss = block.create_var(dtype=loss_sum.dtype, shape=[1])
op_loss = block.append_op(
inputs={"X": loss_sum},
outputs={"Out": loss},
type='scale',
attrs={'scale': 1.0 / float(len(avg_sum))})
op_loss.desc.infer_var_type(block.desc)
op_loss.desc.infer_shape(block.desc)
param_grad_list = append_backward_ops(
loss=loss, parameter_list=input_to_check, no_grad_set=no_grad_set)
feed_dict = {
item[0].name: OpTest._numpy_to_lod_tensor(item[1], item[2], place)
for p_name in inputs_with_np for item in inputs_with_np[p_name]
}
fetch_list = [g for p, g in param_grad_list]
executor = Executor(place)
return map(
np.array,
executor.run(prog, feed_dict, fetch_list, return_numpy=False))
def test_read_write(self):
x = [
layers.data(
name='x0', shape=[100]), layers.data(
name='x1', shape=[100]), layers.data(
name='x2', shape=[100])
]
for each_x in x:
each_x.stop_gradient = False
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = False
arr = layers.array_write(x=x[0], i=i)
i = layers.increment(x=i)
arr = layers.array_write(x=x[1], i=i, array=arr)
i = layers.increment(x=i)
arr = layers.array_write(x=x[2], i=i, array=arr)
i = layers.zeros(shape=[1], dtype='int64')
i.stop_gradient = False
a0 = layers.array_read(array=arr, i=i)
i = layers.increment(x=i)
a1 = layers.array_read(array=arr, i=i)
i = layers.increment(x=i)
a2 = layers.array_read(array=arr, i=i)
mean_a0 = layers.mean(x=a0)
mean_a1 = layers.mean(x=a1)
mean_a2 = layers.mean(x=a2)
a_sum = layers.sums(input=[mean_a0, mean_a1, mean_a2])
mean_x0 = layers.mean(x=x[0])
mean_x1 = layers.mean(x=x[1])
mean_x2 = layers.mean(x=x[2])
x_sum = layers.sums(input=[mean_x0, mean_x1, mean_x2])
scope = core.Scope()
cpu = core.CPUPlace()
exe = Executor(cpu)
tensor = numpy.random.random(size=(100, 100)).astype('float32')
outs = exe.run(feed={'x0': tensor,
'x1': tensor,
'x2': tensor},
fetch_list=[a_sum, x_sum],
scope=scope)
self.assertEqual(outs[0], outs[1])
total_sum = layers.sums(input=[a_sum, x_sum])
total_sum_scaled = layers.scale(x=total_sum, scale=1 / 6.0)
append_backward_ops(total_sum_scaled)
g_vars = map(default_main_program().global_block().var,
[each_x.name + "@GRAD" for each_x in x])
g_out = [
item.sum()
for item in exe.run(
feed={'x0': tensor,
'x1': tensor,
'x2': tensor},
fetch_list=g_vars)
]
g_out_sum = numpy.array(g_out).sum()
# since our final gradient is 1 and the neural network are all linear
# with mean_op.
# the input gradient should also be 1
self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)
def test_grad(self):
place = core.CPUPlace()
program = Program()
x = layers.data(name='x',
shape=[1],
dtype='float32',
main_program=program,
stop_gradient=False)
y = layers.data(name='y',
shape=[1],
dtype='bool',
main_program=program,
stop_gradient=False)
level = 0
out_true, out_false = layers.split_lod_tensor(input=x,
mask=y,
level=level,
main_program=program)
out = layers.merge_lod_tensor(in_true=out_true,
in_false=out_false,
mask=y,
x=x,
level=level,
main_program=program)
mean = layers.mean(x=out, main_program=program)
append_backward_ops(mean)
tensor = core.LoDTensor()
tensor.set(np.arange(10).reshape(10, 1).astype('float32'), place)
tensor.set_lod([[0, 3, 9, 10]])
mask_np = np.array([0, 1, 0]).astype('bool')
mask_np = np.expand_dims(mask_np, axis=1)
mask = core.LoDTensor()
mask.set(mask_np, place)
exe = Executor(place)
scope = core.Scope()
g_vars = program.global_block().var(x.name + "@GRAD")
g_out = [
item.sum() for item in map(
np.array,
exe.run(program,
feed={
'x': tensor,
'y': mask
},
fetch_list=[g_vars],
scope=scope,
return_numpy=False))
]
g_out_sum = np.array(g_out).sum()
self.assertAlmostEqual(1.0, g_out_sum, delta=0.1)