[人工智能]tensorRT教程——使用tensorRT OP 搭建自己的

????????如下提供一个可以运行的使用tensorRT的OP来搭建自己定义的层或者直接重写自己网络，使用OP的场景：

? ? ? ? 1、自己的网络无法通过paser来直接转换为TRT的网络。如果自己写cuda实现，那么量化的操作也得自己实现，这样难度其实很高，建议还是使用TRT的OP搭建，搭建完支持量化等操作。

? ? ? ? 2、学习测试TRT的OP。

? ? ? ? 关于OP的一些我遇到的疑惑解读见我的另一篇博客：https://blog.csdn.net/u012863603/article/details/119358339

import tensorrt as trt
import numpy
import pycuda.driver as cuda
import pycuda.autoinit


TRT_LOGGER = trt.Logger(trt.Logger.WARNING)


def generate_net(network, weights):
    input_tensor = network.add_input(name="input", dtype=trt.float32, shape=(4, 5, 5))
    print(input_tensor.shape)

#    scale_1 = network.add_scale(input=input_tensor, mode=trt.tensorrt.ScaleMode.UNIFORM, shift=np.zeros((1), dtype=np.float32), scale=np.array([4], dtype=np.float32), power=np.ones((1), dtype=np.float32))
 
#    unary_1 = network.add_unary(input=input_tensor, op=trt.tensorrt.UnaryOperation.EXP)
#    print(unary_1.get_output(0).shape)

    elemnet_1 = network.add_elementwise(input1=input_tensor, input2=input_tensor, op=trt.tensorrt.ElementWiseOperation.DIV)
    print(elemnet_1.get_output(0).shape)
#    reduce_1 = network.add_reduce(input=input_tensor, op=trt.tensorrt.ReduceOperation.MAX, axes=1, keep_dims=True)
#    print(reduce_1.get_output(0).shape)
#    network.mark_output(reduce_1.get_output(0))

#    div_1 = network.add_elementwise(input1=input_tensor, input2=reduce_1.get_output(0), op=trt.tensorrt.ElementWiseOperation.DIV)
#    network.mark_output(div_1.get_output(0))

#    const_1 = network.add_constant(shape=[], weights=numpy.zeros((2,), dtype=numpy.int32))
#    gather_1 = network.add_gather(input=input_tensor, indices=const_1.get_output(0), axis=2)
#    print(const_1.get_output(0).shape)
#    network.mark_output(gather_1.get_output(0))

#    shuffle_1 = network.add_shuffle(input=gather_1.get_output(0))
#    shuffle_1.reshape_dims=[1, 5, 5]
#    print(shuffle_1.get_output(0).shape)

#    concat_1 = network.add_concatenation(inputs=[input_tensor, gather_1.get_output(0)])
#    print(concat_1.get_output(0).shape)

    network.mark_output(elemnet_1.get_output(0))


def build_engine(weights):
    # For more information on TRT basics, refer to the introductory samples.
    with trt.Builder(TRT_LOGGER) as builder, builder.create_network() as network:
        builder.max_workspace_size = 1 << 30
        builder.max_batch_size = 8
        # Populate the network using weights from the PyTorch model.
        generate_net(network, weights)
        # Build and return an engine.
        return builder.build_cuda_engine(network)


def allocate_buffers(engine):
    # 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)) * engine.max_batch_size, dtype=trt.nptype(trt.float32))
    h_output = cuda.pagelocked_empty(trt.volume(engine.get_binding_shape(1)) * engine.max_batch_size, dtype=trt.nptype(trt.float32))

    # Allocate:device memory for inputs and outputs.
    d_input = cuda.mem_alloc(h_input.nbytes)
    d_output = cuda.mem_alloc(h_output.nbytes)

    # Create a stream in which to copy inputs/outputs and run inference.
    stream = cuda.Stream()

    return h_input, d_input, h_output, d_output, stream


def inference(batch_size_inference, h_input, d_input, h_output, d_output, stream, context):
    cuda.memcpy_htod_async(d_input, h_input, stream)

    context.execute_async(batch_size=batch_size_inference, bindings=[int(d_input), int(d_output)], stream_handle=stream.handle)

    cuda.memcpy_dtoh_async(h_output, d_output, stream)

    stream.synchronize()

    return h_output
    #result_trt = numpy.reshape(h_output, (8, -1))[:batch_size_inference, ]
    
    #return result_trt

def main():
    input_array = numpy.array(range(300), dtype=numpy.float32).reshape(3,4,5,5) 
    weights = None
    with build_engine(weights) as engine:
        h_input, d_input, h_output, d_output, stream = allocate_buffers(engine)
        with engine.create_execution_context() as context:
            h_input = input_array
            result = inference(3,h_input, d_input, h_output, d_output, stream, context)
            print(result)
if __name__ == '__main__':
    main()