def build_engine_onnx(model_file):
#创建相应的实例
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
common.EXPLICIT_BATCH) as network, builder.create_builder_config(
) as config, trt.OnnxParser(network, TRT_LOGGER) as parser:
#设置相应的参数
config.max_workspace_size = common.GiB(1)
# Load the Onnx model and parse it in order to populate the TensorRT network.
#读取相应的模型文件
with open(model_file, 'rb') as model:
if not parser.parse(model.read()):
print('ERROR: Failed to parse the ONNX file.')
for error in range(parser.num_errors):
print(parser.get_error(error))
return None
#构建相应的引擎
return builder.build_engine(network, config)
def build_engine_caffe(model_file, deploy_file):
# You can set the logger severity higher to suppress messages (or lower to display more messages).
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, trt.CaffeParser() as parser:
# Workspace size is the maximum amount of memory available to the builder while building an engine.
# It should generally be set as high as possible.
builder.max_workspace_size = common.GiB(1)
# Load the Caffe model and parse it in order to populate the TensorRT network.
# This function returns an object that we can query to find tensors by name.
model_tensors = parser.parse(deploy=deploy_file,
model=model_file,
network=network,
dtype=ModelData.DTYPE)
# For Caffe, we need to manually mark the output of the network.
# Since we know the name of the output tensor, we can find it in model_tensors.
network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))
return builder.build_cuda_engine(network)
def build_int8_engine(deploy_file, model_file, calib, batch_size=32):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, trt.CaffeParser() as parser:
# We set the builder batch size to be the same as the calibrator's, as we use the same batches
# during inference. Note that this is not required in general, and inference batch size is
# independent of calibration batch size.
builder.max_batch_size = batch_size
builder.max_workspace_size = common.GiB(1)
builder.int8_mode = True
builder.int8_calibrator = calib
# Parse Caffe model
model_tensors = parser.parse(deploy=deploy_file,
model=model_file,
network=network,
dtype=ModelData.DTYPE)
network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))
# Build engine and do int8 calibration.
return builder.build_cuda_engine(network)
def build_engine_onnx_int8(TRT_LOGGER, model_file, calib):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(common.EXPLICIT_BATCH) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
builder.max_workspace_size = common.GiB(1)
# import pdb;pdb.set_trace()
builder.int8_mode = True
# calibration_cache = "res50_calibration.cache"
# calib = ResNetEntropyCalibrator(training_data='../datasets/ic15/train_list.txt', cache_file=calibration_cache)
builder.int8_calibrator = calib
# Load the Onnx model and parse it in order to populate the TensorRT network.
with open(model_file, 'rb') as model:
if not parser.parse(model.read()):
print ('ERROR: Failed to parse the ONNX file.')
for error in range(parser.num_errors):
print (parser.get_error(error))
return None
# import pdb;pdb.set_trace()
return builder.build_cuda_engine(network)
def build_engine_with_some_missing_weights(weights):
# For more information on TRT basics, refer to the introductory samples.
#创建builder和network实例
with trt.Builder(
TRT_LOGGER) as builder, builder.create_network() as network:
#设置最大工作空间的大小，GIB的具体实现参考common.py
builder.max_workspace_size = common.GiB(1)
# Set the refit flag in the builder
#在builder中设置refit标志位
builder.refittable = True
# Populate the network using weights from the PyTorch model.
#利用相应模型的权重填充tensorrt网络
#populate_network_with_some_dummy_weights的具体实现参考本文件的相应实现
populate_network_with_some_dummy_weights(network, weights)
# Build and return an engine.
#建立相应的引擎
#Builds an ICudaEngine from a INetworkDefinition
return builder.build_cuda_engine(network)
def build_engine(): with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, trt.CaffeParser() as parser: builder.fp16_mode = True builder.strict_type_constraints = True builder.max_batch_size = 16 # Workspace size is the maximum amount of memory available to the builder while building an engine. # It should generally be set as high as possible. builder.max_workspace_size = common.GiB(1) # Load the Caffe model and parse it in order to populate the TensorRT network. # This function returns an object that we can query to find tensors by name. model_tensors = parser.parse(deploy=deploy_file, model=model_file, network=network, dtype=ModelData.DTYPE) # For Caffe, we need to manually mark the output of the network. # Since we know the name of the output tensor, we can find it in model_tensors. network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME)) engine = builder.build_cuda_engine(network) with open(engine_file_path, "wb") as f: f.write(engine.serialize()) return engine
def build_int8_engine(deploy_file, model_file, calib, batch_size=32):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, builder.create_builder_config() as config, trt.CaffeParser(
) as parser, trt.Runtime(TRT_LOGGER) as runtime:
# We set the builder batch size to be the same as the calibrator's, as we use the same batches
# during inference. Note that this is not required in general, and inference batch size is
# independent of calibration batch size.
builder.max_batch_size = batch_size
config.max_workspace_size = common.GiB(1)
config.set_flag(trt.BuilderFlag.INT8)
config.int8_calibrator = calib
# Parse Caffe model
model_tensors = parser.parse(deploy=deploy_file,
model=model_file,
network=network,
dtype=ModelData.DTYPE)
network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))
# Build engine and do int8 calibration.
plan = builder.build_serialized_network(network, config)
return runtime.deserialize_cuda_engine(plan)
def build_engine_uff(model_file):
# You can set the logger severity higher to suppress messages (or lower to display more messages).
#创建相关的实例
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, builder.create_builder_config() as config, trt.UffParser(
) as parser:
# Workspace size is the maximum amount of memory available to the builder while building an engine.
# It should generally be set as high as possible.
#设定相应的参数
config.max_workspace_size = common.GiB(1)
# We need to manually register the input and output nodes for UFF.
#register_input用来注册一个uff网络的输入名称和相应的维度
#注册uff的输入和输出结点
parser.register_input(ModelData.INPUT_NAME, ModelData.INPUT_SHAPE)
parser.register_output(ModelData.OUTPUT_NAME)
# Load the UFF model and parse it in order to populate the TensorRT network.
#使用uff解析器解析模型
parser.parse(model_file, network)
# Build and return an engine.
#创建相关的引擎
return builder.build_engine(network, config)
def build_int8_engine(deploy_file, model_file, batch_size=32, trt_engine_datatype=trt.DataType.FLOAT):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, trt.CaffeParser() as parser:
# We set the builder batch size to be the same as the calibrator's, as we use the same batches
# during inference. Note that this is not required in general, and inference batch size is
# independent of calibration batch size.
builder.max_batch_size = batch_size
builder.max_workspace_size = common.GiB(1)
if trt_engine_datatype == trt.DataType.HALF:
builder.fp16_mode = True
elif trt_engine_datatype == trt.DataType.INT8:
# Now we create a calibrator and give it the location of our calibration data.
# We also allow it to cache calibration data for faster engine building.
_, [calib_data] = common.find_sample_data(description="Runs a Caffe MNIST network in Int8 mode", subfolder="mnist", find_files=["t10k-images-idx3-ubyte"])
calibration_cache = "mnist_calibration.cache"
builder.int8_mode = True
builder.int8_calibrator = MNISTEntropyCalibrator(calib_data, cache_file=calibration_cache)
# Parse Caffe model
model_tensors = parser.parse(deploy=deploy_file, model=model_file, network=network, dtype=ModelData.DTYPE)
network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))
# Build engine and do int8 calibration.
return builder.build_cuda_engine(network)
def build_engine_caffe(model_file, deploy_file, precision):
# precision: float, half, int8
# You can set the logger severity higher to suppress messages (or lower to display more messages).
with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, trt.CaffeParser() as parser:
# Workspace size is the maximum amount of memory available to the builder while building an engine.
# It should generally be set as high as possible.
builder.max_workspace_size = common.GiB(1)
# Load the Caffe model and parse it in order to populate the TensorRT network.
# This function returns an object that we can query to find tensors by name.
model_tensors = parser.parse(deploy=deploy_file, model=model_file, network=network, dtype=ModelData.DTYPE)
# For Caffe, we need to manually mark the output of the network.
# Since we know the name of the output tensor, we can find it in model_tensors.
print(model_tensors)
print(ModelData.OUTPUT_NAME)
print(model_tensors.find(ModelData.OUTPUT_NAME))
network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))
if precision == "half":
# enable fp16 (chenrong06)
builder.fp16_mode = True
builder.strict_type_constraints = True
print("pricision: half")
elif precision == "int8":
# enable int8 and set quantize (chenrong06)
# Incomplete version, please refer to workspace/tensorrt/samples/sampleINT8API/sampleINT8API.cpp
builder.int8_mode = True
builder.int8_calibrator = None
builder.strict_type_constraints = True
print(network.num_layers)
for i in range(network.num_layers):
layer = network[i]
tensor = layer.get_output(0)
tensor.set_dynamic_range(-1.0, 1.0)
tensor = layer.get_input(0)
tensor.set_dynamic_range(-1.0, 1.0)
print("pricision: int8")
else:
print("pricision: float")
return builder.build_cuda_engine(network)
def build_int8_engine(onnx_file_path, calib, batch_size=32):
# with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network, builder.create_builder_config() as config, trt.CaffeParser() as parser:
EXPLICIT_BATCH = 1 << (int)(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(common.EXPLICIT_BATCH) as network, \
builder.create_builder_config() as config, trt.OnnxParser(network, TRT_LOGGER) as parser:
# We set the builder batch size to be the same as the calibrator's, as we use the same batches
# during inference. Note that this is not required in general, and inference batch size is
# independent of calibration batch size.
builder.max_batch_size = batch_size
config.max_workspace_size = common.GiB(1)
config.set_flag(trt.BuilderFlag.INT8)
config.set_flag(trt.BuilderFlag.STRICT_TYPES)
config.int8_calibrator = calib
# Parse Onnx model
with open(onnx_file_path, 'rb') as model:
print('Beginning ONNX file parsing')
if not parser.parse(model.read()):
print('ERROR: Failed to parse the ONNX file.')
for error in range(parser.num_errors):
print(parser.get_error(error))
return None
network.get_input(0).shape = [batch_size, 3, 32, 32]
# Decide which layers fallback to FP32.
# If all layers fallback to FP32, you can use 'index>-1'
for index, layer in enumerate(network):
print('layer index', index, ':', layer.type)
if index < 10:
if layer.type == trt.LayerType.ACTIVATION or \
layer.type == trt.LayerType.CONVOLUTION or \
layer.type == trt.LayerType.FULLY_CONNECTED or \
layer.type == trt.LayerType.SCALE:
print('fallback to fp32!')
layer.precision = trt.float32
layer.set_output_type(0, trt.float32)
# Build engine and do int8 calibration.
return builder.build_engine(network, config)
def build_engine_onnx(model_file):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
builder.max_workspace_size = common.GiB(1)
builder.max_batch_size = args.batch_size
# Load the Onnx model and parse it in order to populate the TensorRT network.
with open(model_file, 'rb') as model:
ok = parser.parse(model.read())
if not ok:
print("Error: Parse onnx model \"{}\" failed.".format(
model_file))
error = parser.get_error(0)
print(" code: {}".format(error.code()))
print(" desc: {}".format(error.desc()))
print(" file: {}".format(error.file()))
print(" func: {}".format(error.func()))
print(" line: {}".format(error.line()))
print(" node: {}".format(error.node()))
exit(-1)
if args.q:
# enable int8 and set quantize (chenrong06)
# Incomplete version, please refer to workspace/tensorrt/samples/sampleINT8API/sampleINT8API.cpp
builder.int8_mode = True
builder.int8_calibrator = None
builder.strict_type_constraints = True
# print(network.num_layers)
for i in range(network.num_layers):
layer = network[i]
tensor = layer.get_output(0)
if tensor:
tensor.set_dynamic_range(-1.0, 1.0)
tensor = layer.get_input(0)
if tensor:
tensor.set_dynamic_range(-1.0, 1.0)
# print("pricision: int8")
return builder.build_cuda_engine(network)
def test_trt_export(model_name=ONNX_MODEL_NAME):
import tensorrt as trt
TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
trt.init_libnvinfer_plugins(TRT_LOGGER, '')
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
common.EXPLICIT_BATCH) as network, trt.OnnxParser(
network, TRT_LOGGER) as parser:
builder.max_workspace_size = common.GiB(1)
builder.fp16_mode = False
builder.max_batch_size = 1
with open(model_name, 'rb') as model:
if not parser.parse(model.read()):
print('ERROR: Failed to parse the ONNX file.')
for error in range(parser.num_errors):
print(parser.get_error(error))
return None
engine = builder.build_cuda_engine(network)
print("CUDA engine build successfully!")
return engine
def build_engine_onnx(model_file):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
common.EXPLICIT_BATCH) as network, trt.OnnxParser(
network, TRT_LOGGER) as parser:
builder.max_workspace_size = common.GiB(1)
builder.fp16_mode = True
builder.max_batch_size = 1 # always 1 for explicit batch
config = builder.create_builder_config()
# need to be set along with fp16_mode if config is specified.
config.set_flag(trt.BuilderFlag.FP16)
profile = builder.create_optimization_profile()
profile.set_shape('input', (1, 1, 4, 4), (2, 1, 4, 4), (4, 1, 4, 4))
profile.set_shape('grid', (1, 4, 4, 2), (2, 4, 4, 2), (4, 4, 4, 2))
config.add_optimization_profile(profile)
# Load the Onnx model and parse it in order to populate the TensorRT network.
with open(model_file, 'rb') as model:
if not parser.parse(model.read()):
print('ERROR: Failed to parse the ONNX file.')
for error in range(parser.num_errors):
print(parser.get_error(error))
return None
return builder.build_engine(network, config)
def build_engine_onnx(model_file, calibrator=None):
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, trt.OnnxParser(network, TRT_LOGGER) as parser:
builder.max_workspace_size = common.GiB(1)
builder.max_batch_size = 8
precision = "fp32"
if calibrator:
builder.int8_mode = True
builder.int8_calibrator = calibrator
precision = "int8"
else:
builder.fp16_mode = True
precision = "fp16"
# Load the Onnx model and parse it in order to populate the TensorRT network.
with open(model_file, 'rb') as model:
parser.parse(model.read())
engine = builder.build_cuda_engine(network)
serialized = engine.serialize()
with open(
"/work/models/flowers-152-b{}-{}.engine".format(
builder.max_batch_size, precision), "wb") as file:
file.write(serialized)
return engine
def build_int8_engine(deploy_file, model_file, calib, batch_size=32):
#创建相关的实例
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
) as network, builder.create_builder_config() as config, trt.CaffeParser(
) as parser:
# We set the builder batch size to be the same as the calibrator's, as we use the same batches
# during inference. Note that this is not required in general, and inference batch size is
# independent of calibration batch size.
#指定相关的参数
builder.max_batch_size = batch_size
config.max_workspace_size = common.GiB(1)
config.set_flag(trt.BuilderFlag.INT8)
config.int8_calibrator = calib
# Parse Caffe model
#使用caffe解析器解析模型，解析一个prototxt文件和一个binaryproto Caffe模型，分别提取网络定义和与网络相关的权值。
model_tensors = parser.parse(deploy=deploy_file,
model=model_file,
network=network,
dtype=ModelData.DTYPE)
#标记网络的输出
network.mark_output(model_tensors.find(ModelData.OUTPUT_NAME))
# Build engine and do int8 calibration.
#构建相应的引擎
return builder.build_engine(network, config)
def build_engine(model_dir):
"""Build TensorRT engine through the Python API.
Args:
model_dir: the trained TensorFlow PSENet model dir.
Returns:
engine: the build TensorRT engine.
"""
ckpt = tf.train.get_checkpoint_state(model_dir)
ckpt_path = ckpt.model_checkpoint_path
reader = pywrap_tensorflow.NewCheckpointReader(ckpt_path)
explicit_flag = 1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
explicit_flag) as network, builder.create_builder_config(
) as config:
data = network.add_input(INPUT_NAME, trt.float32, (-1, 3, -1, -1))
w = reader.get_tensor("resnet_v1_50/conv1/weights").transpose(
3, 2, 0, 1).reshape(-1)
b = np.zeros(64, dtype=np.float32)
conv1 = network.add_convolution(data, 64, (7, 7), trt.Weights(w),
trt.Weights(b))
conv1.stride = (2, 2)
conv1.padding = (3, 3)
bn1 = add_batchnorm(reader, network, conv1.get_output(0),
"resnet_v1_50/conv1/BatchNorm/", 1e-5)
relu1 = network.add_activation(bn1.get_output(0),
trt.ActivationType.RELU)
# C2
pool1 = network.add_pooling(relu1.get_output(0), trt.PoolingType.MAX,
(3, 3))
pool1.stride = (2, 2)
pool1.pre_padding = (0, 0)
pool1.post_padding = (1, 1)
x = bottleneck(reader, network, pool1.get_output(0), 64, 1,
"resnet_v1_50/block1/unit_1/bottleneck_v1/", 1)
x = bottleneck(reader, network, x.get_output(0), 64, 1,
"resnet_v1_50/block1/unit_2/bottleneck_v1/", 0)
# C3
block1 = bottleneck(reader, network, x.get_output(0), 64, 2,
"resnet_v1_50/block1/unit_3/bottleneck_v1/", 2)
x = bottleneck(reader, network, block1.get_output(0), 128, 1,
"resnet_v1_50/block2/unit_1/bottleneck_v1/", 1)
x = bottleneck(reader, network, x.get_output(0), 128, 1,
"resnet_v1_50/block2/unit_2/bottleneck_v1/", 0)
x = bottleneck(reader, network, x.get_output(0), 128, 1,
"resnet_v1_50/block2/unit_3/bottleneck_v1/", 0)
# C4
block2 = bottleneck(reader, network, x.get_output(0), 128, 2,
"resnet_v1_50/block2/unit_4/bottleneck_v1/", 2)
x = bottleneck(reader, network, block2.get_output(0), 256, 1,
"resnet_v1_50/block3/unit_1/bottleneck_v1/", 1)
x = bottleneck(reader, network, x.get_output(0), 256, 1,
"resnet_v1_50/block3/unit_2/bottleneck_v1/", 0)
x = bottleneck(reader, network, x.get_output(0), 256, 1,
"resnet_v1_50/block3/unit_3/bottleneck_v1/", 0)
x = bottleneck(reader, network, x.get_output(0), 256, 1,
"resnet_v1_50/block3/unit_4/bottleneck_v1/", 0)
x = bottleneck(reader, network, x.get_output(0), 256, 1,
"resnet_v1_50/block3/unit_5/bottleneck_v1/", 0)
block3 = bottleneck(reader, network, x.get_output(0), 256, 2,
"resnet_v1_50/block3/unit_6/bottleneck_v1/", 2)
x = bottleneck(reader, network, block3.get_output(0), 512, 1,
"resnet_v1_50/block4/unit_1/bottleneck_v1/", 1)
x = bottleneck(reader, network, x.get_output(0), 512, 1,
"resnet_v1_50/block4/unit_2/bottleneck_v1/", 0)
# C5
block4 = bottleneck(reader, network, x.get_output(0), 512, 1,
"resnet_v1_50/block4/unit_3/bottleneck_v1/", 0)
build_p5_r1 = add_conv_relu(reader, network, block4.get_output(0), 256,
1, 1, "build_feature_pyramid/build_P5/")
build_p4_r1 = add_conv_relu(
reader, network, block2.get_output(0), 256, 1, 1,
"build_feature_pyramid/build_P4/reduce_dimension/")
bfp_layer4_resize = network.add_resize(build_p5_r1.get_output(0))
build_p4_r1_shape = network.add_shape(
build_p4_r1.get_output(0)).get_output(0)
bfp_layer4_resize.set_input(1, build_p4_r1_shape)
bfp_layer4_resize.resize_mode = trt.ResizeMode.NEAREST
bfp_layer4_resize.align_corners = False
bfp_add = network.add_elementwise(build_p4_r1.get_output(0),
bfp_layer4_resize.get_output(0),
trt.ElementWiseOperation.SUM)
build_p4_r2 = add_conv_relu(
reader, network, bfp_add.get_output(0), 256, 3, 1,
"build_feature_pyramid/build_P4/avoid_aliasing/")
build_p3_r1 = add_conv_relu(
reader, network, block1.get_output(0), 256, 1, 1,
"build_feature_pyramid/build_P3/reduce_dimension/")
bfp_layer3_resize = network.add_resize(build_p4_r2.get_output(0))
bfp_layer3_resize.resize_mode = trt.ResizeMode.NEAREST
build_p3_r1_shape = network.add_shape(
build_p3_r1.get_output(0)).get_output(0)
bfp_layer3_resize.set_input(1, build_p3_r1_shape)
bfp_layer3_resize.align_corners = False
bfp_add1 = network.add_elementwise(bfp_layer3_resize.get_output(0),
build_p3_r1.get_output(0),
trt.ElementWiseOperation.SUM)
build_p3_r2 = add_conv_relu(
reader, network, bfp_add1.get_output(0), 256, 3, 1,
"build_feature_pyramid/build_P3/avoid_aliasing/")
build_p2_r1 = add_conv_relu(
reader, network, pool1.get_output(0), 256, 1, 1,
"build_feature_pyramid/build_P2/reduce_dimension/")
bfp_layer2_resize = network.add_resize(build_p3_r2.get_output(0))
bfp_layer2_resize.resize_mode = trt.ResizeMode.NEAREST
build_p2_r1_shape = network.add_shape(
build_p2_r1.get_output(0)).get_output(0)
bfp_layer2_resize.set_input(1, build_p2_r1_shape)
bfp_layer2_resize.align_corners = False
bfp_add2 = network.add_elementwise(bfp_layer2_resize.get_output(0),
build_p2_r1.get_output(0),
trt.ElementWiseOperation.SUM)
# P2
build_p2_r2 = add_conv_relu(
reader, network, bfp_add2.get_output(0), 256, 3, 1,
"build_feature_pyramid/build_P2/avoid_aliasing/")
build_p2_r2_shape = network.add_shape(
build_p2_r2.get_output(0)).get_output(0)
# P3 x2
layer1_resize = network.add_resize(build_p3_r2.get_output(0))
layer1_resize.resize_mode = trt.ResizeMode.LINEAR
layer1_resize.set_input(1, build_p2_r2_shape)
layer1_resize.align_corners = False
# P4 x4
layer2_resize = network.add_resize(build_p4_r2.get_output(0))
layer2_resize.resize_mode = trt.ResizeMode.LINEAR
layer2_resize.set_input(1, build_p2_r2_shape)
layer2_resize.align_corners = False
# p5 right
# P5 x8
layer3_resize = network.add_resize(build_p5_r1.get_output(0))
layer3_resize.resize_mode = trt.ResizeMode.LINEAR
layer3_resize.set_input(1, build_p2_r2_shape)
layer3_resize.align_corners = False
# C(P5,P4,P3,P2)
concat = network.add_concatenation([
layer3_resize.get_output(0),
layer2_resize.get_output(0),
layer1_resize.get_output(0),
build_p2_r2.get_output(0),
])
w = reader.get_tensor("feature_results/Conv/weights").transpose(
3, 2, 0, 1).reshape(-1)
b = np.zeros(256, dtype=np.float32)
feature_result_conv = network.add_convolution(concat.get_output(0),
256, (3, 3),
trt.Weights(w),
trt.Weights(b))
feature_result_conv.padding = (1, 1)
feature_result_bn = add_batchnorm(reader, network,
feature_result_conv.get_output(0),
"feature_results/Conv/BatchNorm/",
1e-5)
feature_result_relu = network.add_activation(
feature_result_bn.get_output(0), trt.ActivationType.RELU)
w = reader.get_tensor("feature_results/Conv_1/weights").transpose(
3, 2, 0, 1).reshape(-1)
b = reader.get_tensor("feature_results/Conv_1/biases")
feature_result_conv_1 = network.add_convolution(
feature_result_relu.get_output(0), 6, (1, 1), trt.Weights(w),
trt.Weights(b))
sigmoid = network.add_activation(feature_result_conv_1.get_output(0),
trt.ActivationType.SIGMOID)
sigmoid.get_output(0).name = OUTPUT_NAME
network.mark_output(sigmoid.get_output(0))
profile = builder.create_optimization_profile()
profile.set_shape("input",
min=(1, 3, 128, 128),
opt=(1, 3, 640, 640),
max=(4, 3, 1200, 1200))
config.add_optimization_profile(profile)
config.max_workspace_size = common.GiB(1)
if USE_FP16:
config_flags = 1 << int(trt.BuilderFlag.FP16)
config.flags = config_flags
engine = builder.build_engine(network, config)
return engine
