wehub-resource-sync
304 lines
9.3 KiB
Python
304 lines
9.3 KiB
Python
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
|
|
# Copyright (c) 2023 NVIDIA 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 numpy as np
|
|
from op_test import get_device_place, is_custom_device
|
|
|
|
import paddle
|
|
from paddle import nn
|
|
|
|
|
|
def skip_unit_test():
|
|
return (
|
|
not (paddle.is_compiled_with_cuda() or is_custom_device())
|
|
or paddle.device.cuda.get_device_capability()[0] < 8
|
|
or paddle.get_cudnn_version() < 8900
|
|
)
|
|
|
|
|
|
skip_msg = (
|
|
"only support with cuda and CUDNN 8.9 or later,"
|
|
" and only Ampere or later devices are supported"
|
|
)
|
|
|
|
|
|
def verify_node_count(graph, node_name, target_count):
|
|
count = 0
|
|
for node in graph.nodes():
|
|
if node.name() == node_name:
|
|
count += 1
|
|
return count == target_count
|
|
|
|
|
|
class ConvBNActLayer(paddle.nn.Layer):
|
|
def __init__(self, num_channels, num_filters, filter_size):
|
|
super().__init__()
|
|
self.act = nn.ReLU()
|
|
self.conv = nn.Conv2D(
|
|
in_channels=num_channels,
|
|
out_channels=num_filters,
|
|
kernel_size=filter_size,
|
|
stride=1,
|
|
padding=(filter_size - 1) // 2,
|
|
groups=1,
|
|
bias_attr=False,
|
|
data_format="NHWC",
|
|
)
|
|
self.bn = nn.BatchNorm(num_filters, data_layout="NHWC")
|
|
|
|
def forward(self, x):
|
|
x = self.conv(x)
|
|
x = self.bn(x)
|
|
return self.act(x)
|
|
|
|
|
|
class ResUnit(paddle.nn.Layer):
|
|
def __init__(self, hidden, is_shortcut):
|
|
super().__init__()
|
|
self.is_shortcut = is_shortcut
|
|
filter_size = 3
|
|
num_channels = hidden
|
|
num_filters = hidden
|
|
self.conv_bn1 = ConvBNActLayer(hidden, hidden, filter_size)
|
|
self.conv_bn2 = ConvBNActLayer(hidden, hidden, filter_size)
|
|
self.act = nn.ReLU()
|
|
self.conv1 = nn.Conv2D(
|
|
in_channels=num_channels,
|
|
out_channels=num_filters,
|
|
kernel_size=filter_size,
|
|
stride=1,
|
|
padding=(filter_size - 1) // 2,
|
|
groups=1,
|
|
bias_attr=False,
|
|
data_format="NHWC",
|
|
)
|
|
self.bn1 = nn.BatchNorm(num_filters, data_layout="NHWC")
|
|
if not self.is_shortcut:
|
|
self.conv2 = nn.Conv2D(
|
|
in_channels=num_channels,
|
|
out_channels=num_filters,
|
|
kernel_size=filter_size,
|
|
stride=1,
|
|
padding=(filter_size - 1) // 2,
|
|
groups=1,
|
|
bias_attr=False,
|
|
data_format="NHWC",
|
|
)
|
|
self.bn2 = nn.BatchNorm(num_filters, data_layout="NHWC")
|
|
|
|
def forward(self, input):
|
|
x1 = self.conv_bn1(input)
|
|
x1 = self.conv_bn2(x1)
|
|
x1 = self.conv1(x1)
|
|
x1 = self.bn1(x1)
|
|
if not self.is_shortcut:
|
|
x2 = self.conv2(input)
|
|
x2 = self.bn2(x2)
|
|
else:
|
|
x2 = input
|
|
output = self.act(paddle.add(x1, x2))
|
|
return output
|
|
|
|
|
|
@unittest.skipIf(skip_unit_test(), skip_msg)
|
|
class TestFuseResUnitBase(unittest.TestCase):
|
|
def setUp(self):
|
|
self.batch = 8
|
|
self.hidden = 16
|
|
self.width = 64
|
|
self.height = 64
|
|
self.iter = 5
|
|
self.set_attr()
|
|
|
|
paddle.enable_static()
|
|
np.random.seed(10)
|
|
paddle.seed(10)
|
|
paddle.framework.random._manual_program_seed(10)
|
|
|
|
self.place = get_device_place()
|
|
self.exe = paddle.static.Executor(self.place)
|
|
|
|
self.feeds = [
|
|
{
|
|
"input": np.random.random(
|
|
(self.batch, self.height, self.width, self.hidden)
|
|
).astype("float16")
|
|
+ i,
|
|
}
|
|
for i in range(self.iter)
|
|
]
|
|
|
|
def build_program(
|
|
self, main_program, startup_program, is_shortcut=True, is_training=False
|
|
):
|
|
with (
|
|
paddle.static.program_guard(main_program, startup_program),
|
|
paddle.utils.unique_name.guard(),
|
|
):
|
|
x1 = paddle.static.data(
|
|
name="input",
|
|
shape=[-1, self.height, self.width, self.hidden],
|
|
dtype='float16',
|
|
)
|
|
layer1 = ConvBNActLayer(self.hidden, self.hidden, 3)
|
|
resunit_layer1 = ResUnit(self.hidden, is_shortcut)
|
|
resunit_layer2 = ResUnit(self.hidden, is_shortcut)
|
|
layer2 = ConvBNActLayer(self.hidden, self.hidden, 3)
|
|
optimizer = None
|
|
with paddle.static.amp.fp16_guard():
|
|
out = layer1(x1)
|
|
out = resunit_layer1(out)
|
|
out = resunit_layer2(out)
|
|
out = layer2(out)
|
|
loss = paddle.mean(out)
|
|
if is_training:
|
|
optimizer = paddle.optimizer.SGD(learning_rate=0.001)
|
|
optimizer = paddle.static.amp.decorate(
|
|
optimizer=optimizer,
|
|
init_loss_scaling=128.0,
|
|
use_dynamic_loss_scaling=True,
|
|
use_pure_fp16=True,
|
|
use_fp16_guard=True,
|
|
)
|
|
optimizer.minimize(loss)
|
|
return loss.name, optimizer
|
|
|
|
def cal_output(self, enable_fusion):
|
|
main_prog = paddle.static.Program()
|
|
startup_prog = paddle.static.Program()
|
|
output_name, optimizer = self.build_program(
|
|
main_prog, startup_prog, self.is_shortcut, self.is_training
|
|
)
|
|
loss_list = []
|
|
scope = paddle.static.Scope()
|
|
with paddle.static.scope_guard(scope):
|
|
self.exe.run(startup_prog)
|
|
if self.is_training:
|
|
optimizer.amp_init(self.place, scope=scope)
|
|
else:
|
|
self._cast_model_to_fp16(main_prog)
|
|
build_strategy = paddle.static.BuildStrategy()
|
|
build_strategy.fuse_resunit = enable_fusion
|
|
program = paddle.static.CompiledProgram(
|
|
main_prog, build_strategy=build_strategy
|
|
)
|
|
|
|
for i in range(self.iter):
|
|
result = self.exe.run(
|
|
program, feed=self.feeds[i], fetch_list=[output_name]
|
|
)
|
|
loss_list.append(result)
|
|
|
|
if enable_fusion:
|
|
self.assertTrue(
|
|
verify_node_count(
|
|
program._graph, "fused_scale_bias_add_relu", 2
|
|
),
|
|
f"[{type(self).__name__}] The number of fused_scale_bias_add_relu is miss-matched in the computing graph.",
|
|
)
|
|
conv_bnstats_count = 6 if self.is_shortcut else 8
|
|
self.assertTrue(
|
|
verify_node_count(
|
|
program._graph,
|
|
"fused_scale_bias_relu_conv_bn",
|
|
conv_bnstats_count,
|
|
),
|
|
f"[{type(self).__name__}] The number of fused_scale_bias_relu_conv_bn is miss-matched in the computing graph.",
|
|
)
|
|
|
|
return np.array(loss_list)
|
|
|
|
def _test_output(self):
|
|
results_ref = self.cal_output(enable_fusion=False)
|
|
results_actual = self.cal_output(enable_fusion=True)
|
|
|
|
np.testing.assert_allclose(
|
|
results_ref,
|
|
results_actual,
|
|
rtol=self.rtol,
|
|
atol=self.atol,
|
|
err_msg=f"[{type(self).__name__}] outputs are miss-matched.",
|
|
)
|
|
|
|
def set_attr(self):
|
|
self.atol = 1e-4
|
|
self.rtol = 1e-4
|
|
|
|
self.is_shortcut = True
|
|
self.is_training = False
|
|
|
|
def _cast_model_to_fp16(self, prog):
|
|
fp16_var_list = paddle.static.amp.cast_model_to_fp16(prog)
|
|
paddle.static.amp.cast_parameters_to_fp16(
|
|
self.place, prog, to_fp16_var_names=fp16_var_list
|
|
)
|
|
|
|
|
|
@unittest.skipIf(skip_unit_test(), skip_msg)
|
|
class TestFuseResUnitShortcutFwd(TestFuseResUnitBase):
|
|
def set_attr(self):
|
|
self.atol = 1e-3
|
|
self.rtol = 1e-2
|
|
|
|
self.is_shortcut = True
|
|
self.is_training = False
|
|
|
|
def test_output(self):
|
|
self._test_output()
|
|
|
|
|
|
@unittest.skipIf(skip_unit_test(), skip_msg)
|
|
class TestFuseResUnitDualFwd(TestFuseResUnitBase):
|
|
def set_attr(self):
|
|
self.atol = 1e-3
|
|
self.rtol = 1e-2
|
|
|
|
self.is_shortcut = False
|
|
self.is_training = False
|
|
|
|
def test_output(self):
|
|
self._test_output()
|
|
|
|
|
|
@unittest.skipIf(skip_unit_test(), skip_msg)
|
|
class TestFuseResUnitBwd(TestFuseResUnitBase):
|
|
def set_attr(self):
|
|
self.atol = 1e-3
|
|
self.rtol = 1e-2
|
|
|
|
self.is_shortcut = True
|
|
self.is_training = True
|
|
|
|
def test_output(self):
|
|
self._test_output()
|
|
|
|
|
|
@unittest.skipIf(skip_unit_test(), skip_msg)
|
|
class TestFuseResUnitDualBwd(TestFuseResUnitBase):
|
|
def set_attr(self):
|
|
self.atol = 1e-3
|
|
self.rtol = 1e-2
|
|
|
|
self.is_shortcut = False
|
|
self.is_training = True
|
|
|
|
def test_output(self):
|
|
self._test_output()
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|