onnx动态分辨率实战

onnx导出动态分辨率

pytorch1.4.0版本，不挑cuda版本，mxnet tf挑cuda版本

人脸检测可以用

   output_onnx = 'abcd.onnx'

    input_names = ["images"]
    output_names = ["detects"]
    inputs = torch.randn(1, 3, 360, 640)


    dynamic_axes = {'images': {0: 'batch',2:'batch',3:'batch'}, 'detects': {0: 'batch'}}
    # torch.onnx.export(net, inputs, output_onnx, input_names=inputs, output_names=outputs,
    #                   dynamic_axes=dynamic_axes)

    torch_out = torch.onnx._export(net, inputs, output_onnx, export_params=True, verbose=False,
                                   input_names=input_names, output_names=output_names,opset_version=11,dynamic_axes=dynamic_axes)
    print("==> Exporting model to ONNX format at '{}'".format(output_onnx))

以下代码出自：

MNN和ONNX支持动态图像尺寸输入 - 知乎

def predict(self, image):
    img, scale, resize = self.pre_process(image)
    im_height, im_width = img.shape[2:]
        
    try:
        self.interpreter.resizeTensor(self.input_tensor, (1, 3, im_height, im_width))
        self.interpreter.resizeSession(self.session)
        tmp_input = MNN.Tensor((1, 3, im_height, im_width), MNN.Halide_Type_Float, img,
                                MNN.Tensor_DimensionType_Caffe)
        self.input_tensor.copyFrom(tmp_input)
        self.interpreter.runSession(self.session)
    except Exception as e:
        print(e)

    conf = np.array(self.interpreter.getSessionOutput(self.session, self.conf_name).getData())
    loc = np.array(self.interpreter.getSessionOutput(self.session, self.loc_name).getData())
    dets = self.post_process(resize, im_height, im_width, scale,  loc.reshape(-1, 4), conf.reshape(-1, 2))
    return dets

在输入Tensor维度不确定或需要修改时，需要调用resizeTensor来更新维度信息。这种情况一般发生在未设置输入维度和输入维度信息可变的情况。更新完所有Tensor的维度信息之后，需要再调用resizeSession来进行预推理，进行内存分配及复用。

需要注意的是，如果MNN模型要支持动态图像尺寸输入，导出ONNX模型时，不要使用torch.onnx.export函数中的dynamic_axes参数设置。即下段导出onnx代码中 dynamic_image = False。

ONNX

先给出代码：

output_onnx = 'faceDetector.onnx'
print("==> Exporting model to ONNX format at '{}'".format(output_onnx))
input_names = ["input0"]
output_names = ["output0", "output1", "output2"]
dynamic_image = True
# 转ONNX 动态输入 -> ONNX转换代码
if dynamic_image:
   dynamic_axes= {'input0':[0, 2, 3], 'output0':[0, 1], 'output1':[0, 1], 'output2':[0, 1]} #数字0，1等是指张量的维度，表示哪个维度需要动态输入
   inputs = torch.randn(1, 3, 120, args.long_side).to(device)
   torch_out = torch.onnx._export(net, inputs, output_onnx, export_params=True, verbose=False,
                                    input_names=input_names, output_names=output_names, dynamic_axes=dynamic_axes)
# MNN动态输入 -> ONNX转换代码
else:
   inputs = torch.randn(1, 3, 120, args.long_side).to(device)
   torch_out = torch.onnx.export(net, inputs, output_onnx, export_params=True, verbose=False,
                                    input_names=input_names, output_names=output_names)

不难发现，动态尺寸图像输入和固定尺寸图像输入的区别在于torch.onnx.export函数中dynamic_axes参数的设置，示例中的'input0':[0, 2, 3]中数字0，2，3是指张量的维度，表示0，2，3维度可以动态尺寸输入。

如果我们的输入模型参数是不断变化的呢，例如第三个维度的长度不是300，又该如何导出？我们只需要指定一下 第二个维度为动态变化的即可 ，如下：

torch.onnx.export(model,  # model being run
                  x,  # model input (or a tuple for multiple inputs)
                  "model2.onnx",  # where to save the model (can be a file or file-like object)
                  export_params=True,  # store the trained parameter weights inside the model file
                  opset_version=10,  # the ONNX version to export the model to
                  do_constant_folding=True,  # whether to execute constant folding for optimization
                  input_names=['input'],  # the model's input names
                  output_names=['output'],  # the model's output names
                  dynamic_axes={'input': {2: 'int_height'}})

那如果有多个变化的，也是写上相应维度即可，并且输出也可以是变化的，如下：

torch.onnx.export(model,  # model being run
                  x,  # model input (or a tuple for multiple inputs)
                  "model2.onnx",  # where to save the model (can be a file or file-like object)
                  export_params=True,  # store the trained parameter weights inside the model file
                  opset_version=10,  # the ONNX version to export the model to
                  do_constant_folding=True,  # whether to execute constant folding for optimization
                  input_names=['input'],  # the model's input names
                  output_names=['output'],  # the model's output names
                  dynamic_axes= {
                       input_name: {0: 'batch_size', 2 : 'in_width', 3: 'int_height'},
                       output_name: {0: 'batch_size', 2: 'out_width', 3:'out_height'}})

原文链接：https://blog.csdn.net/weixin_43198122/article/details/124452172

yolov5-face调整失败：

import argparse
import sys
import time

sys.path.append('./')  # to run '$ python *.py' files in subdirectories

import torch
import torch.nn as nn

import models
from models.experimental import attempt_load
from utils.activations import Hardswish, SiLU
from utils.general import set_logging, check_img_size
import onnx

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--weights', type=str, default='../yolov5n-face.pt', help='weights path')  # from yolov5/models/
    parser.add_argument('--img-size', nargs='+', type=int, default=[320, 320], help='image size')  # height, width
    parser.add_argument('--batch-size', type=int, default=1, help='batch size')
    opt = parser.parse_args()
    opt.img_size *= 2 if len(opt.img_size) == 1 else 1  # expand
    print(opt)
    set_logging()
    t = time.time()

    # Load PyTorch model
    model = attempt_load(opt.weights, map_location=torch.device('cpu'))  # load FP32 model
    model.eval()
    labels = model.names

    # Checks
    gs = int(max(model.stride))  # grid size (max stride)
    opt.img_size = [check_img_size(x, gs) for x in opt.img_size]  # verify img_size are gs-multiples

    # Input
    img = torch.zeros(opt.batch_size, 3, *opt.img_size)  # image size(1,3,320,192) iDetection

    # Update model
    for k, m in model.named_modules():
        m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
        if isinstance(m, models.common.Conv):  # assign export-friendly activations
            if isinstance(m.act, nn.Hardswish):
                m.act = Hardswish()
            elif isinstance(m.act, nn.SiLU):
                m.act = SiLU()
        # elif isinstance(m, models.yolo.Detect):
        #     m.forward = m.forward_export  # assign forward (optional)
        if isinstance(m, models.common.ShuffleV2Block):#shufflenet block nn.SiLU
            for i in range(len(m.branch1)):
                if isinstance(m.branch1[i], nn.SiLU):
                    m.branch1[i] = SiLU()
            for i in range(len(m.branch2)):
                if isinstance(m.branch2[i], nn.SiLU):
                    m.branch2[i] = SiLU()
    model.model[-1].export = True  # set Detect() layer export=True
    y = model(img)  # dry run

    # ONNX export
    print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
    output_onnx= opt.weights.replace('.pt', '.onnx')  # filename
    model.fuse()  # only for ONNX
    torch.onnx.export(model, img, output_onnx, verbose=False, opset_version=12, input_names=['data'],
                      output_names=['stride_' + str(int(x)) for x in model.stride])
    #
    # input_names = ["data"]
    # # output_names = ['stride_' + str(int(x)) for x in model.stride]
    # output_names = ['detects']
    # inputs = torch.randn(1, 3, 640, 640)
    #
    # dynamic_axes = {'data': {0: 'batch', 2: 'batch', 3: 'batch'}, 'detects': {0: 'batch'}}


    # torch_out = torch.onnx._export(model, img, output_onnx, export_params=True, verbose=False, input_names=input_names,
    #                                output_names=output_names, opset_version=12, dynamic_axes=dynamic_axes)

    # Checks
    onnx_model = onnx.load(output_onnx)  # load onnx model
    onnx.checker.check_model(onnx_model)  # check onnx model
    # print(onnx.helper.printable_graph(onnx_model.graph))  # print a human readable model
    print('ONNX export success, saved as %s' % output_onnx)
    # Finish
    print('\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.' % (time.time() - t))