wehub-resource-sync
188 lines
7.1 KiB
Python
188 lines
7.1 KiB
Python
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import unittest
|
|
|
|
import paddle
|
|
from paddle.autograd.ir_backward import grad
|
|
from paddle.base.core import call_vjp, has_vjp
|
|
from paddle.base.libpaddle.pir import (
|
|
build_pipe_for_block,
|
|
get_used_external_value,
|
|
)
|
|
|
|
paddle.enable_static()
|
|
|
|
|
|
class TestConstructModuleWithPyLayerOp(unittest.TestCase):
|
|
def test_fwd_only_with_single_output(self):
|
|
"""
|
|
pseudocode:
|
|
|
|
y = 3 * x
|
|
"""
|
|
|
|
def forward_fn(x):
|
|
return 3 * x
|
|
|
|
with paddle.pir_utils.IrGuard():
|
|
main_program = paddle.static.Program()
|
|
startup_program = paddle.static.Program()
|
|
with paddle.static.program_guard(main_program, startup_program):
|
|
x = paddle.static.data(name="x", shape=[6, 1], dtype="float32")
|
|
x.stop_gradient = True
|
|
out = paddle.static.nn.static_pylayer(
|
|
forward_fn, [x], backward_fn=None
|
|
)
|
|
y = paddle.mean(out)
|
|
|
|
pylayer_op = main_program.global_block().ops[-3]
|
|
self.assertEqual(pylayer_op.name(), "pd_op.pylayer")
|
|
self.assertEqual(len(pylayer_op.results()), 1)
|
|
value_list = get_used_external_value(pylayer_op)
|
|
self.assertEqual(len(value_list), 1)
|
|
self.assertTrue(value_list[0].is_same(pylayer_op.operand_source(0)))
|
|
|
|
def test_fwd_only_with_multi_inputs_multi_outputs(self):
|
|
"""
|
|
pseudocode:
|
|
|
|
ret1 = x * y
|
|
ret2 = x - y
|
|
"""
|
|
|
|
def forward_fn(x, y):
|
|
z = 3 * x
|
|
return z * y, z - y
|
|
|
|
with paddle.pir_utils.IrGuard():
|
|
main_program = paddle.static.Program()
|
|
startup_program = paddle.static.Program()
|
|
with paddle.static.program_guard(main_program, startup_program):
|
|
x = paddle.static.data(name="x", shape=[6, 1], dtype="float32")
|
|
y = paddle.static.data(name="x", shape=[6, 1], dtype="float32")
|
|
x.stop_gradient = True
|
|
y.stop_gradient = True
|
|
ret1, ret2 = paddle.static.nn.static_pylayer(
|
|
forward_fn, [x, y], backward_fn=None
|
|
)
|
|
out = ret1 + ret2
|
|
|
|
pylayer_op = main_program.global_block().ops[-2]
|
|
self.assertEqual(pylayer_op.name(), "pd_op.pylayer")
|
|
self.assertEqual(len(pylayer_op.results()), 2)
|
|
value_list = get_used_external_value(pylayer_op)
|
|
self.assertEqual(len(value_list), 2)
|
|
self.assertTrue(value_list[0].is_same(pylayer_op.operand_source(0)))
|
|
self.assertTrue(value_list[1].is_same(pylayer_op.operand_source(1)))
|
|
|
|
# check build_pipe_for_block interface
|
|
fwd_block = pylayer_op.as_pylayer_op().forward_block()
|
|
build_pipe_for_block(fwd_block)
|
|
|
|
def test_pylayer_op_call_vjp_interface(self):
|
|
def forward_fn(x, y):
|
|
z = 3 * x
|
|
tmp = paddle.mean(x)
|
|
return z * y, z - y, tmp
|
|
|
|
def backward_fn(dx, dy, dtmp):
|
|
return dx - dy, paddle.broadcast_to(dtmp, dx.shape)
|
|
|
|
with paddle.pir_utils.IrGuard():
|
|
main_program = paddle.static.Program()
|
|
startup_program = paddle.static.Program()
|
|
with paddle.static.program_guard(main_program, startup_program):
|
|
x = paddle.static.data(name="x", shape=[6, 1], dtype="float32")
|
|
x.stop_gradient = False
|
|
y = paddle.static.data(name="y", shape=[6, 1], dtype="float32")
|
|
y.stop_gradient = False
|
|
|
|
out_0, out_1, out_2 = paddle.static.nn.static_pylayer(
|
|
forward_fn, [x, y], backward_fn=backward_fn
|
|
)
|
|
ret = paddle.mean(out_0 - 2 * out_1 + out_2)
|
|
|
|
dataop = main_program.global_block().ops[0]
|
|
pylayer_op = main_program.global_block().ops[-7]
|
|
self.assertEqual(pylayer_op.name(), "pd_op.pylayer")
|
|
self.assertEqual(len(pylayer_op.results()), 3)
|
|
|
|
out_grad_0 = paddle.full(
|
|
shape=[6, 1], dtype='float32', fill_value=3
|
|
)
|
|
out_grad_1 = paddle.full(
|
|
shape=[6, 1], dtype='float32', fill_value=2
|
|
)
|
|
out_grad_2 = paddle.full(
|
|
shape=[], dtype='float32', fill_value=0.1
|
|
)
|
|
pylayer_input = [
|
|
[input] for input in get_used_external_value(pylayer_op)
|
|
]
|
|
pylayer_input_stop_gradients = [[False], [False]]
|
|
pylayer_output = [pylayer_op.results()]
|
|
pylayer_output_grad = [[out_grad_0], [out_grad_1], [out_grad_2]]
|
|
self.assertEqual(has_vjp(pylayer_op), False)
|
|
|
|
grad_outs = call_vjp(
|
|
pylayer_op,
|
|
pylayer_input,
|
|
pylayer_output,
|
|
pylayer_output_grad,
|
|
pylayer_input_stop_gradients,
|
|
)
|
|
|
|
def test_pylayer_op_backward(self):
|
|
def forward_fn(x, y):
|
|
z = 3 * x
|
|
tmp = paddle.mean(x)
|
|
return z * y, z - y, tmp
|
|
|
|
def backward_fn(dx, dy, dtmp):
|
|
return dx - dy, dtmp * dy * dx
|
|
|
|
with paddle.pir_utils.IrGuard():
|
|
main_program = paddle.static.Program()
|
|
startup_program = paddle.static.Program()
|
|
with paddle.static.program_guard(main_program, startup_program):
|
|
x = paddle.static.data(name="x", shape=[6, 1], dtype="float32")
|
|
x.stop_gradient = False
|
|
y = paddle.static.data(name="y", shape=[6, 1], dtype="float32")
|
|
y.stop_gradient = False
|
|
|
|
out_0, out_1, out_2 = paddle.static.nn.static_pylayer(
|
|
forward_fn, [x, y], backward_fn=backward_fn
|
|
)
|
|
ret = paddle.mean(out_0 - 2 * out_1 + out_2)
|
|
|
|
dataop = main_program.global_block().ops[0]
|
|
pylayer_op = main_program.global_block().ops[-7]
|
|
self.assertEqual(pylayer_op.name(), "pd_op.pylayer")
|
|
self.assertEqual(len(pylayer_op.results()), 3)
|
|
|
|
self.assertEqual(has_vjp(pylayer_op), False)
|
|
dataop_0 = main_program.global_block().ops[0]
|
|
dataop_1 = main_program.global_block().ops[1]
|
|
|
|
# check vjp interface for if_op
|
|
grad_outs = grad(
|
|
pylayer_op.results(),
|
|
[dataop_0.result(0), dataop_1.result(0)],
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|