wehub-resource-sync
307 lines
10 KiB
Python
307 lines
10 KiB
Python
# Copyright (c) 2024 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 os
|
|
import tempfile
|
|
import unittest
|
|
|
|
import numpy as np
|
|
from get_program import (
|
|
get_r50_program,
|
|
get_r50_refit_program,
|
|
)
|
|
|
|
import paddle
|
|
import paddle.inference as paddle_infer
|
|
from paddle.quantization import PTQ, QuantConfig
|
|
from paddle.quantization.observers import AbsmaxObserver
|
|
from paddle.tensorrt.export import (
|
|
Input,
|
|
PrecisionMode,
|
|
TensorRTConfig,
|
|
convert,
|
|
convert_to_trt,
|
|
)
|
|
from paddle.tensorrt.util import (
|
|
predict_program,
|
|
)
|
|
from paddle.vision.models import resnet18
|
|
|
|
# NOTE(Pan Zhaowu): using legacy linear to fulfill promise of tensorrt graph capturing
|
|
# and converting.
|
|
paddle.set_flags({"FLAGS_use_legacy_linear": True})
|
|
|
|
|
|
def standardize(array):
|
|
mean_val = np.mean(array)
|
|
std_val = np.std(array)
|
|
standardized_array = (array - mean_val) / std_val
|
|
return standardized_array
|
|
|
|
|
|
class TestConverterResNet50(unittest.TestCase):
|
|
def setUp(self):
|
|
self.temp_dir = tempfile.TemporaryDirectory()
|
|
self.path = os.path.join(self.temp_dir.name, 'pir-trt')
|
|
|
|
def test_paddle_to_tensorrt_conversion_r50(self):
|
|
# Step1: get program and init fake inputs
|
|
program, scope, param_dict = get_r50_program()
|
|
|
|
# Set input
|
|
input_config = Input(
|
|
min_input_shape=(1, 3, 224, 224),
|
|
optim_input_shape=(1, 3, 224, 224),
|
|
max_input_shape=(4, 3, 224, 224),
|
|
input_data_type='float32',
|
|
name='input',
|
|
)
|
|
_, input_optim_data, _ = input_config.generate_input_data()
|
|
|
|
# Create a TensorRTConfig with inputs as a required field.
|
|
trt_config = TensorRTConfig(inputs=[input_config])
|
|
trt_config.disable_passes = ['dead_code_elimination_pass']
|
|
|
|
output_var = program.list_vars()[-1]
|
|
|
|
# get original results(for tests only)
|
|
|
|
output_expected = predict_program(
|
|
program, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
program_with_trt = convert_to_trt(program, trt_config, scope)
|
|
output_var = program_with_trt.list_vars()[-1]
|
|
|
|
# Step6: run inference(converted_program)
|
|
output_converted = predict_program(
|
|
program_with_trt, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
output_expected = standardize(output_expected[0])
|
|
output_trt = standardize(output_converted[0])
|
|
|
|
# Check that the results are close to each other within a tolerance of 1e-3
|
|
np.testing.assert_allclose(
|
|
output_expected,
|
|
output_trt,
|
|
rtol=1e-3,
|
|
atol=1e-3,
|
|
err_msg="Outputs are not within the 1e-3 tolerance",
|
|
)
|
|
|
|
def test_refit(self):
|
|
# Step1: get program and init fake inputs
|
|
paddle.enable_static()
|
|
save_path = os.path.join(self.temp_dir.name, 'resnet50')
|
|
program, scope, param_dict = get_r50_refit_program(save_path)
|
|
|
|
# Set input
|
|
input_config = Input(
|
|
min_input_shape=(1, 3, 224, 224),
|
|
optim_input_shape=(1, 3, 224, 224),
|
|
max_input_shape=(4, 3, 224, 224),
|
|
input_data_type='float32',
|
|
)
|
|
_, input_optim_data, _ = input_config.generate_input_data()
|
|
|
|
# Create a TensorRTConfig with inputs as a required field.
|
|
trt_config = TensorRTConfig(inputs=[input_config])
|
|
|
|
output_var = program.list_vars()[-1]
|
|
|
|
# get original results(for tests only)
|
|
|
|
output_expected = predict_program(
|
|
program, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
trt_save_path = os.path.join(self.temp_dir.name, 'resnet50trt')
|
|
trt_config.save_model_dir = trt_save_path
|
|
trt_config.refit_params_path = save_path + '.pdiparams'
|
|
model_dir = save_path
|
|
|
|
program_with_trt = paddle.tensorrt.convert(model_dir, trt_config)
|
|
config = paddle_infer.Config(
|
|
trt_config.save_model_dir + '.json',
|
|
trt_config.save_model_dir + '.pdiparams',
|
|
)
|
|
config.switch_ir_debug(True)
|
|
if paddle.is_compiled_with_cuda():
|
|
config.enable_use_gpu(100, 0)
|
|
else:
|
|
config.disable_gpu()
|
|
predictor = paddle_infer.create_predictor(config)
|
|
|
|
paddle.disable_static()
|
|
for i, input_instance in enumerate(trt_config.inputs):
|
|
min_data, _, max_data = input_instance.generate_input_data()
|
|
model_inputs = paddle.to_tensor(min_data)
|
|
output_converted = predictor.run([model_inputs])
|
|
|
|
output_expected = standardize(output_expected[0])
|
|
output_trt = standardize(output_converted[0].numpy())
|
|
|
|
np.testing.assert_allclose(
|
|
output_expected,
|
|
output_trt,
|
|
rtol=1e-1,
|
|
atol=1e-1,
|
|
err_msg="Outputs are not within the 1e-1 tolerance",
|
|
)
|
|
|
|
def test_paddle_to_tensorrt_conversion_r50_collect_shape(self):
|
|
# Step1: get program and init fake inputs
|
|
program, scope, param_dict = get_r50_program()
|
|
|
|
# Set input
|
|
input_data = tuple(
|
|
np.random.rand(n, 3, 224, 224).astype(np.float32) for n in (1, 2, 4)
|
|
)
|
|
input_optim_data = input_data[1]
|
|
input_config = Input(warmup_data=input_data)
|
|
|
|
# Create a TensorRTConfig with inputs as a required field.
|
|
trt_config = TensorRTConfig(inputs=[input_config])
|
|
trt_config.disable_passes = ['dead_code_elimination_pass']
|
|
|
|
output_var = program.list_vars()[-1]
|
|
|
|
# get original results(for tests only)
|
|
|
|
output_expected = predict_program(
|
|
program, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
program_with_trt = convert_to_trt(program, trt_config, scope)
|
|
output_var = program_with_trt.list_vars()[-1]
|
|
|
|
# Step6: run inference(converted_program)
|
|
output_converted = predict_program(
|
|
program_with_trt, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
output_expected = standardize(output_expected[0])
|
|
output_trt = standardize(output_converted[0])
|
|
|
|
# Check that the results are close to each other within a tolerance of 1e-3
|
|
np.testing.assert_allclose(
|
|
output_expected,
|
|
output_trt,
|
|
rtol=1e-2,
|
|
atol=1e-2,
|
|
err_msg="Outputs are not within the 1e-2 tolerance",
|
|
)
|
|
|
|
def test_convert_quant_model(self):
|
|
paddle.disable_static()
|
|
image = paddle.ones([1, 3, 224, 224], dtype="float32")
|
|
model = resnet18()
|
|
model.eval()
|
|
output_fp32 = model(image)
|
|
|
|
observer = AbsmaxObserver(quant_bits=8)
|
|
q_config = QuantConfig(activation=observer, weight=observer)
|
|
ptq = PTQ(q_config)
|
|
quant_model = ptq.quantize(model)
|
|
out = quant_model(image)
|
|
converted_model = ptq.convert(quant_model)
|
|
save_path = os.path.join(self.temp_dir.name, 'int8_infer')
|
|
paddle.jit.save(converted_model, save_path, input_spec=[image])
|
|
|
|
paddle.enable_static()
|
|
trt_save_path = os.path.join(self.temp_dir.name, 'int8_trt_infer')
|
|
# Set input
|
|
input_config = Input(
|
|
min_input_shape=(1, 3, 224, 224),
|
|
optim_input_shape=(1, 3, 224, 224),
|
|
max_input_shape=(1, 3, 224, 224),
|
|
input_data_type='float32',
|
|
)
|
|
trt_config = TensorRTConfig(inputs=[input_config])
|
|
trt_config.disable_passes = ['dead_code_elimination_pass']
|
|
trt_config.save_model_dir = trt_save_path
|
|
trt_config.precision_mode = PrecisionMode.INT8
|
|
convert(save_path, trt_config)
|
|
|
|
config = paddle_infer.Config(
|
|
trt_config.save_model_dir + '.json',
|
|
trt_config.save_model_dir + '.pdiparams',
|
|
)
|
|
config.enable_use_gpu(100, 0)
|
|
predictor = paddle_infer.create_predictor(config)
|
|
output_trt_int8 = predictor.run([image])
|
|
|
|
# Check that the results are close to each other within a tolerance of 0.9
|
|
np.testing.assert_allclose(
|
|
output_fp32,
|
|
output_trt_int8[0],
|
|
rtol=0.9,
|
|
atol=0.9,
|
|
err_msg="Outputs are not within the 0.9 tolerance",
|
|
)
|
|
|
|
def test_paddle_to_tensorrt_conversion_r50_use_cuda_graph(self):
|
|
# Step1: get program and init fake inputs
|
|
program, scope, param_dict = get_r50_program()
|
|
|
|
# Set input
|
|
input_config = Input(
|
|
min_input_shape=(1, 3, 224, 224),
|
|
optim_input_shape=(1, 3, 224, 224),
|
|
max_input_shape=(4, 3, 224, 224),
|
|
input_data_type='float32',
|
|
name='input',
|
|
)
|
|
_, input_optim_data, _ = input_config.generate_input_data()
|
|
|
|
# Create a TensorRTConfig with inputs as a required field.
|
|
trt_config = TensorRTConfig(inputs=[input_config])
|
|
trt_config.disable_passes = ['dead_code_elimination_pass']
|
|
|
|
# use_cuda_graph: True
|
|
trt_config.use_cuda_graph = True
|
|
|
|
output_var = program.list_vars()[-1]
|
|
|
|
# get original results(for tests only)
|
|
|
|
output_expected = predict_program(
|
|
program, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
program_with_trt = convert_to_trt(program, trt_config, scope)
|
|
output_var = program_with_trt.list_vars()[-1]
|
|
|
|
# Step6: run inference(converted_program)
|
|
output_converted = predict_program(
|
|
program_with_trt, {"input": input_optim_data}, [output_var]
|
|
)
|
|
|
|
output_expected = standardize(output_expected[0])
|
|
output_trt = standardize(output_converted[0])
|
|
|
|
# Check that the results are close to each other within a tolerance of 1e-3
|
|
np.testing.assert_allclose(
|
|
output_expected,
|
|
output_trt,
|
|
rtol=1e-3,
|
|
atol=1e-3,
|
|
err_msg="Outputs are not within the 1e-3 tolerance",
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|