未验证 提交 1ee5ba1c 编写于 作者: A Aurelius84 提交者: GitHub

Enhance checking on sub branch of backward (#21582)

* add backward unittest test=develop

* add sub-branch in test_backward.py test=develop

* refine code, add comment test=develop

* reconstruct TestBackward Class test=develop

* fix typo of comment test=develop
上级 34dc7106
...@@ -16,72 +16,205 @@ from __future__ import print_function ...@@ -16,72 +16,205 @@ from __future__ import print_function
import unittest import unittest
import paddle.fluid as fluid import paddle.fluid as fluid
from simple_nets import init_data import numpy as np
def case1_fill_grad_vars(): class BackwardNet(object):
x = fluid.layers.data(name='image', shape=[784], dtype='float32') """
label = fluid.layers.data(name='label', shape=[1], dtype='int64') Abstract Base Class.
feature = fluid.layers.fc(input=x, size=20, act=None) All Net inherited this Class should implement two functions:
part1, part2 = fluid.layers.split(feature, num_or_sections=[10, 10], dim=1) build_model: build net to test the logic of backward
# Note that: part2 is not used. init_data: fake input data to test all programs.
loss = fluid.layers.cross_entropy(input=part1, label=label) """
loss = fluid.layers.mean(loss)
return loss def __init__(self):
self.stop_gradient_grad_vars = set()
self.no_grad_vars = set()
def case2_prune_no_grad_branch(): self.params_names = set()
x = fluid.layers.data(name='image', shape=[784], dtype='float32') self.op_path = []
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
feature = fluid.layers.fc(input=x, size=10, act=None) def build_model(self):
label = fluid.layers.cast(label, dtype="float32") """
label = fluid.layers.cast(label, dtype='int64') Build net to test the logic of backward.
# Note that the label is not persistable in fluid.layers.cross_entropy. :return: loss
loss = fluid.layers.cross_entropy(input=feature, label=label) """
loss = fluid.layers.mean(loss) raise NotImplementedError
return loss
def init_data(self):
"""
def case3_prune_no_grad_branch2(): Fake input data to test all programs.
label = fluid.layers.data(name='label', shape=[1], dtype='int64') :return: dict, {'var_name': var_data}
label = fluid.layers.cast(label, dtype="float32") """
label = fluid.layers.cast(label, dtype='int64') raise NotImplementedError
out = fluid.layers.one_hot(input=label, depth=100)
loss = fluid.layers.mean(out)
return loss
def case4_with_no_grad_op_maker():
out = fluid.layers.gaussian_random(shape=[20, 30])
loss = fluid.layers.mean(out)
return loss
class TestBackward(unittest.TestCase): class TestBackward(unittest.TestCase):
def check_backward(self, model, feed_dict): """
place = fluid.CPUPlace() All related TestClass should inherit this class,
and only implement test_backward function.
"""
def _check_all(self, net):
place = fluid.CUDAPlace(0) if fluid.core.is_compiled_with_cuda(
) else fluid.CPUPlace()
exe = fluid.Executor(place) exe = fluid.Executor(place)
main = fluid.Program() main = fluid.Program()
startup = fluid.Program() startup = fluid.Program()
with fluid.program_guard(main, startup): with fluid.program_guard(main, startup):
loss = model() loss = net.build_model()
self._check_backward(loss, main)
optimizer = fluid.optimizer.SGD(learning_rate=0.1) optimizer = fluid.optimizer.SGD(learning_rate=0.1)
optimizer.minimize(loss) optimizer.minimize(loss)
exe.run(startup)
exe.run(fluid.default_startup_program()) exe.run(feed=net.init_data())
exe.run(feed=feed_dict)
def _check_backward(self, loss, main_program):
global_block_idx = self.global_block_idx
params_grads = self._check_params_grad(loss)
# 1.1 get_stop_gradients
no_grad_dict = self._check_stop_gradient(main_program)
# 1.2 find_op_path
op_path, block_no_grad_set = self._check_op_path(
main_program.block(global_block_idx), [loss], [], no_grad_dict)
# 1.3 _find_no_grad_vars
no_grad_vars = self._check_find_no_grad_vars(
main_program.block(global_block_idx), op_path, [loss],
block_no_grad_set)
# update no_grad_dict
block_no_grad_set.update(no_grad_vars)
no_grad_dict[global_block_idx].update(
list(map(fluid.backward._append_grad_suffix_, block_no_grad_set)))
def _check_params_grad(self, loss, parameter_list=None, no_grad_set=None):
params_grads = fluid.backward.append_backward(loss, parameter_list,
no_grad_set)
params_names = set(
[param_var.name for (param_var, grad_var) in params_grads])
self.assertSetEqual(params_names, self.net.params_names)
return params_grads
def _check_stop_gradient(self, program):
no_grad_dict = fluid.backward._get_stop_gradients_(program)
if no_grad_dict is not None and isinstance(no_grad_dict, dict):
self.assertSetEqual(no_grad_dict[self.global_block_idx],
self.net.stop_gradient_grad_vars)
return no_grad_dict
def _check_op_path(self, root_block, outputs, inputs=[], no_grad_dict=None):
if no_grad_dict is None or not isinstance(no_grad_dict, dict):
block_no_grad_set = None
else:
block_no_grad_set = set(
map(fluid.backward._strip_grad_suffix_, no_grad_dict[
self.global_block_idx]))
op_path = fluid.backward._find_op_path_(root_block, outputs, inputs,
block_no_grad_set)
op_types = [op.type for op in op_path]
self.assertListEqual(op_types, self.net.op_path)
return op_path, block_no_grad_set
def _check_find_no_grad_vars(self, root_block, op_path, targets,
block_no_grad_set):
no_grad_vars = fluid.backward._find_no_grad_vars(
root_block, op_path, targets, block_no_grad_set)
self.assertSetEqual(no_grad_vars, self.net.no_grad_vars)
return no_grad_vars
class SimpleNet(BackwardNet):
def __init__(self):
super(BackwardNet, self).__init__()
self.stop_gradient_grad_vars = set([
u'x_no_grad@GRAD', u'x2_no_grad@GRAD', u'x3_no_grad@GRAD',
u'label_no_grad@GRAD'
])
self.no_grad_vars = set()
self.params_names = set([u'w2v', u'fc_predict.b_0', u'fc_w'])
self.op_path = [
u'lookup_table_v2',
u'lookup_table_v2', # embedding
u'elementwise_add', # merge
u'mul',
u'elementwise_add',
u'softmax', # fc
u'elementwise_sub',
u'square',
u'mean'
] # loss
self.shape = [16, 50]
def init_data(self):
assert len(self.shape) == 2
x = np.random.randint(0, 90, self.shape).astype('int64')
x2 = np.random.randint(0, 90, self.shape).astype('int64')
x3 = np.random.randint(0, 90, self.shape).astype('int64')
label = np.random.random([self.shape[0], 1]).astype('float32')
return {
'x_no_grad': x,
'x2_no_grad': x2,
'x3_no_grad': x3,
'label_no_grad': label
}
def build_model(self):
# stop_gradient = True in input
x = fluid.data(name='x_no_grad', shape=self.shape, dtype='int64')
x2 = fluid.data(name='x2_no_grad', shape=self.shape, dtype='int64')
x3 = fluid.data(name='x3_no_grad', shape=self.shape, dtype='int64')
label = fluid.data(
name='label_no_grad', shape=[self.shape[0], 1], dtype='float32')
# shared layer, the grad of 'w2v' will be summed and renamed.
# To test _addup_repetitive_outputs_
x_emb = fluid.embedding(
x, size=[100, 64], param_attr=fluid.ParamAttr(name='w2v'))
x2_emb = fluid.embedding(
x2, size=[100, 64], param_attr=fluid.ParamAttr(name='w2v'))
x3_emb = fluid.embedding(
x3, size=[100, 64], param_attr=fluid.ParamAttr(name='w2v'))
# merge layers
x_merge = fluid.layers.elementwise_add(x_emb, x2_emb, name='x_add_x2')
x2_merge = fluid.layers.elementwise_add(
x2_emb, x3_emb, name='x2_add_x3')
# shared fc_w
predict = fluid.layers.fc(input=x_merge,
size=1,
act='softmax',
param_attr=fluid.ParamAttr(name='fc_w'),
name='fc_predict')
# useless layer for calculating loss
fc_no_use = fluid.layers.fc(input=x2_merge,
size=1,
act='sigmoid',
param_attr=fluid.ParamAttr(name='fc_w'),
name='fc_no_use')
# loss
cost = fluid.layers.square_error_cost(input=predict, label=label)
loss = fluid.layers.mean(cost, name='mean_loss')
return loss
class TestSimpleNet(TestBackward):
def test_backward(self): def test_backward(self):
batch_size = 2 """
img, label = init_data(batch_size, img_shape=[784], label_range=9) Instantiate each NetClass to test backward.
feed_dict = {'image': img, 'label': label} """
self.check_backward(case1_fill_grad_vars, feed_dict) self.global_block_idx = 0
self.check_backward(case2_prune_no_grad_branch, feed_dict) self.net = SimpleNet()
self.check_backward(case3_prune_no_grad_branch2, {'label': label}) self._check_all(self.net)
self.check_backward(case4_with_no_grad_op_maker, {})
# TODO(Aurelius84): add conditional network test
class ConditionalNet(BackwardNet):
def __init__(self):
super(BackwardNet, self).__init__()
if __name__ == '__main__': if __name__ == '__main__':
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册