def __init__(self, engine_file_path):
# Create a Context on this device,
self.cfx = cuda.Device(0).make_context()
stream = cuda.Stream()
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
runtime = trt.Runtime(TRT_LOGGER)
# Deserialize the engine from file
with open(engine_file_path, "rb") as f:
engine = runtime.deserialize_cuda_engine(f.read())
context = engine.create_execution_context()
host_inputs = []
cuda_inputs = []
host_outputs = []
cuda_outputs = []
bindings = []
for binding in engine:
size = trt.volume(
engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
cuda_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(cuda_mem))
# Append to the appropriate list.
if engine.binding_is_input(binding):
host_inputs.append(host_mem)
cuda_inputs.append(cuda_mem)
else:
host_outputs.append(host_mem)
cuda_outputs.append(cuda_mem)
# Store
self.stream = stream
self.context = context
self.engine = engine
self.host_inputs = host_inputs
self.cuda_inputs = cuda_inputs
self.host_outputs = host_outputs
self.cuda_outputs = cuda_outputs
self.bindings = bindings
def __init__(self, model):
print('setting up Yolov5s-simple.trt processor')
# load tensorrt engine
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
TRTbin = '{0}/models/{1}'.format(os.path.dirname(__file__), model)
with open(TRTbin, 'rb') as f, trt.Runtime(TRT_LOGGER) as runtime:
engine = runtime.deserialize_cuda_engine(f.read())
self.context = engine.create_execution_context()
# allocate memory
inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
for binding in engine:
size = trt.volume(engine.get_binding_shape(binding))
dtype = trt.nptype(engine.get_binding_dtype(binding))
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
bindings.append(int(device_mem))
if engine.binding_is_input(binding):
inputs.append({'host': host_mem, 'device': device_mem})
else:
outputs.append({'host': host_mem, 'device': device_mem})
# save to class
self.inputs = inputs
self.outputs = outputs
self.bindings = bindings
self.stream = stream
# post processing config
filters = (80 + 5) * 3
self.output_shapes = [(1, 3, 80, 80, 85), (1, 3, 40, 40, 85),
(1, 3, 20, 20, 85)]
self.strides = np.array([8., 16., 32.])
anchors = np.array([
[[10, 13], [16, 30], [33, 23]],
[[30, 61], [62, 45], [59, 119]],
[[116, 90], [156, 198], [373, 326]],
])
self.nl = len(anchors)
self.nc = 80 # classes
self.no = self.nc + 5 # outputs per anchor
self.na = len(anchors[0])
a = anchors.copy().astype(np.float32)
a = a.reshape(self.nl, -1, 2)
self.anchors = a.copy()
self.anchor_grid = a.copy().reshape(self.nl, 1, -1, 1, 1, 2)
def allocate_buffers(engine, profile_id):
"""Allocate device memory for I/O bindings of engine and return them."""
d_inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
if engine.has_implicit_batch_dimension:
max_batch_size = engine.max_batch_size
else:
shape = engine.get_binding_shape(0)
if -1 in list(shape):
batch_dim = list(shape).index(-1)
max_batch_size = engine.get_profile_shape(0, 0)[2][batch_dim]
else:
max_batch_size = shape[0]
nb_bindings_per_profile = engine.num_bindings // engine.num_optimization_profiles
bindings = [0 for i in range(engine.num_bindings)]
for binding in range(profile_id * nb_bindings_per_profile, (profile_id + 1) * nb_bindings_per_profile):
logging.info("Binding {:}".format(binding))
dtype = engine.get_binding_dtype(binding)
format = engine.get_binding_format(binding)
shape = engine.get_binding_shape(binding)
if format == trt.TensorFormat.CHW4:
shape[-3] = ((shape[-3] - 1) // 4 + 1) * 4
elif format == trt.TensorFormat.DHWC8:
shape[-4] = ((shape[-4] - 1) // 8 + 1) * 8
if not engine.has_implicit_batch_dimension:
if -1 in list(shape):
batch_dim = list(shape).index(-1)
shape[batch_dim] = max_batch_size
size = trt.volume(shape)
else:
size = trt.volume(shape) * max_batch_size
# Allocate device buffers
device_mem = cuda.mem_alloc(size * dtype.itemsize)
# Append device buffer to device bindings.
bindings[binding] = int(device_mem)
# Append to the appropriate list.
if engine.binding_is_input(binding):
d_inputs.append(device_mem)
else:
host_mem = cuda.pagelocked_empty(size, trt.nptype(dtype))
outputs.append(HostDeviceMem(host_mem, device_mem))
return d_inputs, outputs, bindings, stream
def __init__(self, model, anchor_nums, nc, anchors, output_shapes,
img_size):
# load tensorrt engine
self.cfx = cuda.Device(0).make_context()
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
TRTbin = model
# print('trtbin', TRTbin)
runtime = trt.Runtime(TRT_LOGGER)
with open(TRTbin, 'rb') as f:
engine = runtime.deserialize_cuda_engine(f.read())
self.context = engine.create_execution_context()
# allocate memory
inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
for binding in engine:
size = trt.volume(
engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
bindings.append(int(device_mem))
if engine.binding_is_input(binding):
inputs.append({'host': host_mem, 'device': device_mem})
else:
outputs.append({'host': host_mem, 'device': device_mem})
# save to class
self.inputs = inputs
self.outputs = outputs
self.bindings = bindings
self.stream = stream
self.anchor_nums = anchor_nums
self.nc = nc # classes
self.no = self.nc + 5 # outputs per anchor
# post processing config
self.output_shapes = output_shapes
self.strides = np.array([8., 16., 32.])
self.na = len(anchors[0])
self.nl = len(anchors)
self.img_size = img_size
a = anchors.copy().astype(np.float32)
a = a.reshape(self.nl, -1, 2)
self.anchors = a.copy()
self.anchor_grid = a.copy().reshape(self.nl, 1, -1, 1, 1, 2)
def allocate_buffers(cls, engine):
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
for binding in engine:
size = trt.volume(engine.get_binding_shape(binding)) \
* engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_memory = cuda.pagelocked_empty(size, dtype)
device_memory = cuda.mem_alloc(host_memory.nbytes)
bindings.append(int(device_memory))
if engine.binding_is_input(binding):
inputs.append(HostDeviceMemory(host_memory, device_memory))
else:
outputs.append(HostDeviceMemory(host_memory, device_memory))
return inputs, outputs, bindings, stream
def allocate_buffers(engine):
"""Allocates host and device buffer for TRT engine inference.
This function is similair to the one in ../../common.py, but
converts network outputs (which are np.float32) appropriately
before writing them to Python buffer. This is needed, since
uff plugins doesn't support output type description, and
in our particular case, we use NMS plugin as network output.
Args:
engine (trt.ICudaEngine): uff engine
Returns:
inputs [HostDeviceMem]: engine input memory
outputs [HostDeviceMem]: engine output memory
bindings [int]: buffer to device bindings
stream (cuda.Stream): cuda stream for engine inference synchronization
"""
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
# Current NMS implementation in TRT only supports DataType.FLOAT but
# it may change in the future, which could brake this sample here
# when using lower precision [e.g. NMS output would not be np.float32
# anymore, even though this is assumed in binding_to_type]
for binding in engine:
size = trt.volume(
engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
# Append to the appropriate list.
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def allocate_buffers(engine):
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
for binding in engine:
size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
# Append to the appropriate list.
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def build_engine():
"""Takes an ONNX file and creates a TensorRT engine to run inference with"""
builder = trt.Builder(TRT_LOGGER)
network = builder.create_network(common.EXPLICIT_BATCH)
parser = trt.OnnxParser(network, TRT_LOGGER)
runtime = trt.Runtime(TRT_LOGGER)
# Parse model file
print('Loading ONNX file from path {}...'.format(onnx_file_path))
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
print('Completed parsing of ONNX file')
# Print input info
print('Network inputs:')
for i in range(network.num_inputs):
tensor = network.get_input(i)
print(tensor.name, trt.nptype(tensor.dtype), tensor.shape)
network.get_input(0).shape = [10, 1]
network.get_input(1).shape = [10, 1, 1, 16]
network.get_input(2).shape = [6, 1]
network.get_input(3).shape = [6, 1, 1, 16]
config = builder.create_builder_config()
config.set_flag(trt.BuilderFlag.REFIT)
config.max_workspace_size = 1 << 28 # 256MiB
print(
'Building an engine from file {}; this may take a while...'.format(
onnx_file_path))
plan = builder.build_serialized_network(network, config)
engine = runtime.deserialize_cuda_engine(plan)
print("Completed creating Engine")
with open(engine_file_path, "wb") as f:
f.write(plan)
return engine
def allocate_buffers(self):
self.inputs = []
self.outputs = []
self.bindings = []
self.stream = cuda.Stream()
for binding in self.engine:
size = trt.volume(self.engine.get_binding_shape(binding)) * self.max_batch_size
dtype = trt.nptype(self.engine.get_binding_dtype(binding))
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
self.bindings.append(int(device_mem))
# print(bindings)
# Append to the appropriate list.
if self.engine.binding_is_input(binding):
self.inputs.append(HostDeviceMem(host_mem, device_mem))
else:
self.outputs.append(HostDeviceMem(host_mem, device_mem))
def get_layer_weights(self, layer_name):
"""
Get the weights of a constant layer as a numpy array
:param layer_name: string
:return:
"""
layer = self._ops[layer_name]
if isinstance(layer, trt.IConstantLayer):
return np.reshape(self._ops[layer_name].weights,
self._ops_output[layer_name].shape)
elif isinstance(layer, trt.IIdentityLayer):
cast_to_dtype = trt.nptype(layer.precision)
const_layer_name = layer.get_input(0).name
weights = self.get_layer_weights(const_layer_name)
return weights.astype(cast_to_dtype)
else:
return layer.get_output(0)
def __init__(self, trt_path):
# get model name
self._model_name = os.path.basename(trt_path)
self._model_name = self._model_name[:self._model_name.rfind(".")]
# create engine
self.trt_path = trt_path
self.logger = trt.Logger()
self.runtime = trt.Runtime(self.logger)
with open(trt_path, "rb") as f:
self.engine = self.runtime.deserialize_cuda_engine(f.read())
# create context and buffer
self.context = self.engine.create_execution_context()
self.stream = cuda.Stream()
bindings = []
host_input = device_input = host_output = device_output = None
for binding in self.engine:
binding_idx = self.engine.get_binding_index(binding)
print(f"binding name {binding}, idx {binding_idx}")
shape = trt.volume(self.context.get_binding_shape(binding_idx))
dtype = trt.nptype(self.engine.get_binding_dtype(binding))
if self.engine.binding_is_input(binding):
print(shape)
host_input = np.empty(shape, dtype=np.float32)
device_input = cuda.mem_alloc(host_input.nbytes)
bindings.append(int(device_input))
else:
host_output = cuda.pagelocked_empty(shape, dtype)
device_output = cuda.mem_alloc(host_output.nbytes)
bindings.append(int(device_output))
assert device_input is not None
assert device_output is not None
assert len(bindings) == 2
self.bindings = bindings
self.device_input = device_input
self.host_input = host_input
self.device_output = device_output
self.host_output = host_output
def build_engine(onnx_file_path, engine_file_path, calib=None):
EXPLICIT_BATCH = 1 << (int)(
trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
with trt.Builder(TRT_LOGGER) as builder, builder.create_network(
(EXPLICIT_BATCH)) as network, trt.OnnxParser(network,
TRT_LOGGER) as parser:
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
print('Completed parsing of ONNX file')
print('Network inputs:')
for i in range(network.num_inputs):
tensor = network.get_input(i)
print(tensor.name, trt.nptype(tensor.dtype), tensor.shape)
config = builder.create_builder_config()
config.max_workspace_size = 1 << 20
if calib:
config.set_flag(trt.BuilderFlag.INT8)
config.int8_calibrator = calib
else:
builder.fp16_mode = True
print(
'Building an engine from file {}; this may take a while...'.format(
onnx_file_path))
engine = builder.build_engine(network, config)
print("Completed creating Engine. Writing file to: {}".format(
engine_file_path))
with open(engine_file_path, "wb") as f:
f.write(engine.serialize())
return engine
def run(batchSize, nRow, nCol):
print("test", batchSize, nRow, nCol)
logger = trt.Logger(trt.Logger.ERROR)
trt.init_libnvinfer_plugins(logger, '')
ctypes.cdll.LoadLibrary(soFilePath)
engine = buildEngine(logger, nRow, nCol)
if engine == None:
print("Failed building engine!")
return None
print("Succeeded building engine!")
context = engine.create_execution_context()
stream = cuda.Stream()
condition = np.array(np.random.randint(0, 2, [batchSize, nRow, nCol]),
dtype=np.int32)
inputX = np.full([batchSize, nRow, nCol], 1, dtype=np.float32)
inputY = np.full([batchSize, nRow, nCol], -1, dtype=np.float32)
inputH0 = np.ascontiguousarray(condition.reshape(-1))
inputH1 = np.ascontiguousarray(inputX.reshape(-1))
inputH2 = np.ascontiguousarray(inputY.reshape(-1))
inputD0 = cuda.mem_alloc(inputH0.nbytes)
inputD1 = cuda.mem_alloc(inputH1.nbytes)
inputD2 = cuda.mem_alloc(inputH2.nbytes)
outputH0 = np.empty((batchSize, ) + tuple(engine.get_binding_shape(3)),
dtype=trt.nptype(engine.get_binding_dtype(3)))
outputD0 = cuda.mem_alloc(outputH0.nbytes)
cuda.memcpy_htod_async(inputD0, inputH0, stream)
cuda.memcpy_htod_async(inputD1, inputH1, stream)
cuda.memcpy_htod_async(inputD2, inputH2, stream)
context.execute_async(
batchSize,
[int(inputD0), int(inputD1),
int(inputD2), int(outputD0)], stream.handle)
cuda.memcpy_dtoh_async(outputH0, outputD0, stream)
stream.synchronize()
outputH0CPU = whereCPU(condition, inputX, inputY)
print("Check result:",
["True" if np.all(outputH0 == outputH0CPU) else "False"][0])
def _allocate_buffers(self):
""" nvidia function to allocate memory on the device """
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
for binding in self.engine:
size = (
trt.volume(self.engine.get_binding_shape(binding))
* self.engine.max_batch_size
)
dtype = trt.nptype(self.engine.get_binding_dtype(binding))
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
bindings.append(int(device_mem))
if self.engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def allocate_buffers(self):
d_inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
# FIXME: Clarify this with Po-han. What if we want a batch size smaller than max_batch_size from builder?
# max_batch_size = self.engine.max_batch_size if self.engine.has_implicit_batch_dimension else self.engine.get_profile_shape(0, 0)[2][0]
if self.engine.has_implicit_batch_dimension:
max_batch_size = min(self.engine.max_batch_size,
self.args.batch_size)
else:
max_batch_size = self.args.batch_size
for binding in self.engine:
logging.info("Binding {:}".format(binding))
desc = self.engine.get_binding_format_desc(
self.engine.get_binding_index(binding))
logging.info(" Binding info {:} with shape {:}".format(
desc,
self.engine.get_binding_shape(
self.engine.get_binding_index(binding))))
dtype = self.engine.get_binding_dtype(binding)
format = self.engine.get_binding_format(
self.engine.get_binding_index(binding))
shape = self.engine.get_binding_shape(binding)
if format == trt.TensorFormat.CHW4:
shape[-3] = ((shape[-3] - 1) // 4 + 1) * 4
if not self.engine.has_implicit_batch_dimension:
shape[0] = max_batch_size
size = trt.volume(shape)
else:
size = trt.volume(shape) * max_batch_size
# Allocate host and device buffers
device_mem = cuda.mem_alloc(size * dtype.itemsize)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
# Append to the appropriate list.
if self.engine.binding_is_input(binding):
d_inputs.append(device_mem)
else:
host_mem = cuda.pagelocked_empty(size, trt.nptype(dtype))
outputs.append(HostDeviceMem(host_mem, device_mem))
return d_inputs, outputs, bindings, stream
def allocate_buffers(engine):
print(engine.get_binding_shape(0))
print(engine.get_binding_shape(1))
print(engine.get_binding_shape(2))
#print(engine.get_binding_shape(3))
#bindings = []
# Determine dimensions and create page-locked memory buffers (i.e. won't be swapped to disk) to hold host inputs/outputs.
h_input = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(0)), dtype=trt.nptype(ModelData.DTYPE))
h_output_1 = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(1)), dtype=trt.nptype(ModelData.DTYPE))
h_output_2 = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(2)), dtype=trt.nptype(ModelData.DTYPE))
#h_output_3 = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(3)), dtype=trt.nptype(ModelData.DTYPE))
# Allocate device memory for inputs and outputs.
d_input = cuda.mem_alloc(h_input.nbytes)
d_output_1 = cuda.mem_alloc(h_output_1.nbytes)
d_output_2 = cuda.mem_alloc(h_output_2.nbytes)
#d_output_3 = cuda.mem_alloc(h_output_3.nbytes)
print('bbb')
# Create a stream in which to copy inputs/outputs and run inference.
stream = cuda.Stream()
return h_input, d_input, [h_output_1,h_output_2] , [d_output_1,d_output_2], stream
def _init_buffers(self):
self._stream = cuda.Stream()
self._input_buffers = []
self._output_buffers = []
self._bindings = []
for binding in self._engine:
dtype = np.dtype(trt.nptype(self._engine.get_binding_dtype(binding)))
shape = self._engine.get_binding_shape(binding)
size = trt.volume(shape) * self._engine.max_batch_size
device = cuda.mem_alloc(size * dtype.itemsize)
self._bindings.append(int(device))
if self._engine.binding_is_input(binding):
host = None # will be added during inference
buffer = MemoryBuffer(dtype, shape, host, device)
self._input_buffers.append(buffer)
else:
host = cuda.pagelocked_empty(size, dtype)
buffer = MemoryBuffer(dtype, shape, host, device)
self._output_buffers.append(buffer)
def allocate_buffers(engine, batch_size=None):
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
class HostDeviceMem(object):
def __init__(self, host_mem, device_mem):
self.host = host_mem
self.device = device_mem
def __str__(self):
return "Host:\n" + str(self.host) + "\nDevice:\n" + str(
self.device)
def __repr__(self):
return self.__str__()
for binding in engine:
dims = engine.get_binding_shape(binding)
# print('buff--dims:', dims)
if dims[0] == -1:
assert (batch_size is not None)
dims[0] = batch_size
size = trt.volume(
dims
) * engine.max_batch_size # The maximum batch size which can be used for inference.
# print("size:",size)
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
if engine.binding_is_input(
binding): # Determine whether a binding is an input binding.
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def allocate_buffers(engine):
class HostDeviceMem(object):
def __init__(self, host_mem, device_mem):
"""
host_mem: cpu memory
device_mem: gpu memory
"""
self.host = host_mem
self.device = device_mem
def __str__(self):
return "Host:\n" + str(self.host) + "\nDevice:\n" + str(
self.device)
def __repr__(self):
return self.__str__()
inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
for binding in engine:
# print(binding) # 绑定的输入输出
# print(engine.get_binding_shape(binding)) # get_binding_shape 是变量的大小
size = trt.volume(
engine.get_binding_shape(binding)) * engine.max_batch_size
# volume 计算可迭代变量的空间，指元素个数
# size = trt.volume(engine.get_binding_shape(binding)) # 如果采用固定bs的onnx，则采用该句
dtype = trt.nptype(engine.get_binding_dtype(binding))
# get_binding_dtype 获得binding的数据类型
# nptype等价于numpy中的dtype，即数据类型
# allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype) # 创建锁业内存
device_mem = cuda.mem_alloc(host_mem.nbytes) # cuda分配空间
# print(int(device_mem)) # binding在计算图中的缓冲地址
bindings.append(int(device_mem))
# append to the appropriate list
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def __init__(self, model_path):
# load model
with open(model_path, 'rb') as f:
with trt.Runtime(trt.Logger(trt.Logger.ERROR)) as runtime:
self.engine = runtime.deserialize_cuda_engine(f.read())
# allocate buffers
self.inputs = list()
self.outputs = list()
self.bindings = list()
self.stream = cuda.Stream()
for binding in self.engine:
shape = self.engine.get_binding_shape(binding)
size = trt.volume(shape) * self.engine.max_batch_size
dtype = trt.nptype(self.engine.get_binding_dtype(binding))
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
self.bindings.append(int(device_mem))
if self.engine.binding_is_input(binding):
self.inputs.append([host_mem, device_mem])
else:
self.outputs.append([host_mem, device_mem])
self.context = self.engine.create_execution_context()
def main():
model_path="model/trt/conv_3d_2.trt"
TRT_LOGGER = trt.Logger(trt.Logger.WARNING)
runtime=trt.Runtime(TRT_LOGGER)
f=open(model_path,"rb")
engine = runtime.deserialize_cuda_engine(f.read())
context=engine.create_execution_context()
f.close()
for binding in engine:
size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Append to the appropriate list.
if engine.binding_is_input(binding):
print("input_size: ", size, "dtype: ", dtype)
else:
print("output_size: ", size, "dtype: ", dtype)
inputs, outputs, bindings, stream = common.allocate_buffers(engine)
length=input_shape[0]*input_shape[1]*input_shape[2]*input_shape[3]
data=np.zeros(length,dtype=np.float32)
data[:]=1.0
inputs[0].host=data.reshape(input_shape)
print(inputs[0].host[0][0][0][:10])
outputs[0].host=np.zeros(output_shape, dtype=np.float32)
trt_outputs = common.do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream)
print(trt_outputs[0][0][0][0][:10])
print("starting...")
starttime = time.time()
for i in range(1000):
trt_outputs = common.do_inference(context, bindings=bindings, inputs=inputs, outputs=outputs, stream=stream)
endtime = time.time()
print (endtime - starttime)
print(trt_outputs[0][0][0][0][:10])
def allocate_buffer(self):
bindings = []
inputs = []
outputs = []
for binding in self.engine:
size = trt.volume(self.engine.get_binding_shape(
binding)) * self.engine.max_batch_size
dtype = trt.nptype(self.engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
if self.engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings
def get_input_metadata(self):
inputs = OrderedDict()
active_profile = self.context.active_optimization_profile
bindings_per_profile = len(
self.engine) // self.engine.num_optimization_profiles
logging.debug(
"Total # of Profiles: {:}, Bindings Per Profile: {:}, Active Profile: {:}"
.format(self.engine.num_optimization_profiles,
bindings_per_profile, active_profile))
start_binding = bindings_per_profile * active_profile
end_binding = start_binding + bindings_per_profile
logging.info("Start Binding: {:}, End Binding: {:}".format(
start_binding, end_binding))
for binding in range(start_binding, end_binding):
if self.engine.binding_is_input(binding):
inputs[self.engine[binding]] = (
trt.nptype(self.engine.get_binding_dtype(binding)),
list(self.engine.get_binding_shape(binding)),
)
return inputs
def allocate_buffers(engine):
"""
Allocates all buffers required for an engine, i.e. host/device inputs/outputs.
Parameters
----------
engine : tensorrt.ICudaEngine
An ICudaEngine for executing inference on a built network
Returns
-------
list
All input HostDeviceMem of an engine
list
All output HostDeviceMem of an engine
GPU bindings
Device bindings
GPU stream
A stream is a sequence of commands (possibly issued by different host threads) that execute in order
"""
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
for binding in engine:
size = trt.volume(engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
# Append to the appropriate list.
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def allocate_buffers(engine):
inputs = []
outputs = []
bindings = []
stream = cuda.Stream()
for binding in engine: # binding相当于一组字符串名称，包括'data','prob', engine可以切片， engine[0]='data', engine[1]='prob'
size = trt.volume(
engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size,
dtype) # (784,)是把img拉直的一个一维数组，作为主机的缓存
device_mem = cuda.mem_alloc(
host_mem.nbytes) # obj，可以int(obj)，是所占的GPU内存比特数
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
# Append to the appropriate list.
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem,
device_mem)) # (2,) 分别是data和prob
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def allocate_buffers(engine):
for binding in engine:
size = trt.volume(
engine.get_binding_shape(binding)) * engine.max_batch_size
dtype = trt.nptype(engine.get_binding_dtype(binding))
# Allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
# Append the device buffer to device bindings.
bindings.append(int(device_mem))
# Append to the appropriate list.
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
# input_output_path = "/home/nsathish/Efficient_object_detection/mmdetection/buffer/ssd_pickle.dat"
# buffers=[inputs,outputs,bindings]
# with open(input_output_path, "wb") as f:
# pickle.dump(buffers, f, pickle.HIGHEST_PROTOCOL)
return inputs, outputs, bindings, stream
def allocate_buffers(engine):
inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
for binding in engine:
# print(binding) # 绑定的输入输出
# print(engine.get_binding_shape(binding)) # get_binding_shape 是变量的大小
size = trt.volume(engine.get_binding_shape(binding))*engine.max_batch_size
# volume 计算可迭代变量的空间，指元素个数
# size = trt.volume(engine.get_binding_shape(binding)) # 如果采用固定bs的onnx，则采用该句
dtype = trt.nptype(engine.get_binding_dtype(binding))
# get_binding_dtype 获得binding的数据类型
# nptype等价于numpy中的dtype，即数据类型
# allocate host and device buffers
host_mem = cuda.pagelocked_empty(size, dtype) # 创建锁业内存
device_mem = cuda.mem_alloc(host_mem.nbytes) # cuda分配空间
# print(int(device_mem)) # binding在计算图中的缓冲地址
bindings.append(int(device_mem))
#append to the appropriate list
if engine.binding_is_input(binding):
inputs.append(HostDeviceMem(host_mem, device_mem))
else:
outputs.append(HostDeviceMem(host_mem, device_mem))
return inputs, outputs, bindings, stream
def __init__(self, model):
# load tensorrt engine
TRT_LOGGER = trt.Logger(trt.Logger.INFO)
TRTbin = model
print('trtbin', TRTbin)
with open(TRTbin, 'rb') as f, trt.Runtime(TRT_LOGGER) as runtime:
engine = runtime.deserialize_cuda_engine(f.read())
self.context = engine.create_execution_context()
# allocate memory
inputs, outputs, bindings = [], [], []
stream = cuda.Stream()
for binding in engine:
size = trt.volume(engine.get_binding_shape(binding))
dtype = trt.nptype(engine.get_binding_dtype(binding))
host_mem = cuda.pagelocked_empty(size, dtype)
device_mem = cuda.mem_alloc(host_mem.nbytes)
bindings.append(int(device_mem))
if engine.binding_is_input(binding):
inputs.append({'host': host_mem, 'device': device_mem})
else:
outputs.append({'host': host_mem, 'device': device_mem})
# save to class
self.inputs = inputs
self.outputs = outputs
self.bindings = bindings
self.stream = stream
self.no = 12
self.output_shapes = [(1, 3, 56, 56, self.no), (1, 3, 28, 28, self.no),
(1, 3, 14, 14, self.no)]
self.names = [
'angular_leafspot', 'anthracnose_fruit_rot', 'blossom_blight',
'gray_mold', 'leaf_spot', 'powdery_mildew_fruit',
'powdery_mildew_leaf'
]
self.img_size = 448
def build(self):
assert os.path.exists(self.engine_file), "Engine file doesn't exist"
runtime = trt.Runtime(TrtModel.TRT_LOGGER)
with open(self.engine_file, 'rb') as engine_file:
self.engine = runtime.deserialize_cuda_engine(engine_file.read())
if self.engine is None:
raise RuntimeError('Unable to load the engine file')
self.context = self.engine.create_execution_context()
self.stream = cp.cuda.Stream(non_blocking=True)
self.max_batch_size = self.engine.get_profile_shape(0, 0)[2][0]
for binding in self.engine:
shape = self.engine.get_binding_shape(binding)
if shape[0] == -1:
shape = (self.max_batch_size, ) + shape[1:]
size = trt.volume(shape)
dtype = trt.nptype(self.engine.get_binding_dtype(binding))
buffer = HostDeviceMem(size, dtype)
self.bindings.append(buffer.devptr)
if self.engine.binding_is_input(binding):
self.input = buffer
self.input_shapes.append(
self.engine.get_binding_shape(binding))
else:
self.outputs.append(buffer)
self.out_shapes.append(self.engine.get_binding_shape(binding))
self.out_names.append(binding)
assert self.input is not None
self.start = cp.cuda.Event()
self.end = cp.cuda.Event()
def load_input(img_path, host_buffer):
print('load input')
c, h, w = INPUT_SHAPE
img = cv2.imread(img_path)
mean = [123.829891747, 127.351147446, 110.256170154]
stdv = [0.016895854, 0.017222115, 0.014714524]
img = cv2.resize(img, (h, w))
img = np.swapaxes(img, 0, 2)
img = np.swapaxes(img, 1, 2)
img = np.array(img, dtype=float)
mean = np.array(mean, dtype=float)
stdv = np.array(stdv, dtype=float)
img[0, :, :] -= mean[0]
img[1, :, :] -= mean[1]
img[2, :, :] -= mean[2]
img[0, :, :] *= stdv[0]
img[1, :, :] *= stdv[1]
img[2, :, :] *= stdv[2]
dtype = trt.nptype(DTYPE)
img_array = np.asarray(img).astype(dtype).ravel()
np.copyto(host_buffer, img_array)
