文章目录
1 准备工作
1.1 efficientnet网络介绍
详见参考链接EfficientNet网络结构及代码详解。
1.2 efficientnet训练分类数据集
详见参考链接EfficientNet训练自定义分类数据集。
2 为何要转
想部署在开发板上，通常需要先转成onnx形式。
3 安装相关依赖
其它的依赖在网络训练过程中通常都会遇到，用于转换的依赖安装：
pip install onnxruntime
pip install onnx
4 转换过程
需要准备的东西很少：
主要代码如下，运行后可得到转换后的onnx模型：
import onnxruntime
import torch
import numpy as np
import onnx
import os
from PIL import Image
import matplotlib.pyplot as plt
import json
# 导入网络结构
from model import efficientnet_b0 as create_model
def model_convert_onnx(model, input_shape, output_path):
dummy_input = torch.randn(1, 3, input_shape[0], input_shape[1])
input_names = ["input1"]
output_names = ["output1"]
torch.onnx.export(
model,
dummy_input,
output_path,
verbose=False, # 如果指定为True，在导出的onnx中会有详细的导出过程信息description
keep_initializers_as_inputs=False, # 若为True，会出现5.3节中的warning消除
opset_version=11, # 版本通常为10 or 11
input_names=input_names,
output_names=output_names,
)
if __name__ == '__main__':
# create model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = create_model(num_classes=5).to(device)
# load pth model weights
model_weight_path = "./output/model-25.pth"
model.load_state_dict(torch.load(model_weight_path, map_location=device))
# 据说因为BN、dropout的存在，所以这儿要转成eval()模式
model.eval()
# 导出onnx模型的输入尺寸，要和pytorch模型的输入尺寸一致
input_shape = (224, 224)
# onnx模型输出到哪里去
output_path = './output/efficientnet_b0.onnx'
## ------------------------------------------#
## pth模型转换为onnx模型，转换完成后，可注释掉
## ------------------------------------------#
# model_convert_onnx(model, input_shape, output_path)
# print("model convert onnx finsh.")
5 检验生成的onnx模型
5.1 onnx.checker检验
通常这一步都没啥问题，要是有问题，当我没说。
import onnx
onnx_model = onnx.load("efficientnet_b0.onnx")
onnx.checker.check_model(onnx_model)
5.2 np.testing.assert_allclose校验
直接看代码注释即可。
def check_onnx_2(model, ort_session, input_shape, device):
# -----------------------------------#
# 给个模型输入，分辨率要对
# -----------------------------------#
x = torch.randn(size=(1, 3, input_shape[0], input_shape[1]), dtype=torch.float32)
# -----------------------------------#
# torch模型推理
# -----------------------------------#
with torch.no_grad():
torch_out = model(x.to(device))
# print(torch_out) # tensor([[-0.5728, 0.1695, -0.3256, 1.1357, -0.4081]])
# print(type(torch_out)) # <class 'torch.Tensor'>
# -----------------------------------#
# onnx模型推理
# -----------------------------------#
ort_inputs = {ort_session.get_inputs()[0].name: x.numpy()} # 初始化数据，注意这儿的x是上面的输入数据x，后期应该是img
ort_outs = ort_session.run(None, ort_inputs) # 推理得到输出
# print(ort_outs) # [array([[-0.5727689 , 0.16947027, -0.32555276, 1.13574 , -0.40812433]], dtype=float32)]
# print(type(ort_outs)) # <class 'list'>，里面是个numpy矩阵
# print(type(ort_outs[0])) # <class 'numpy.ndarray'>
ort_outs = ort_outs[0] # 因此这儿需要把内部numpy矩阵取出来，这一步很有必要
# print(torch_out.numpy().shape) # (1, 5)，1张图片，该图片属于5个类别的概率
# print(ort_outs.shape) # (1, 5)
# ---------------------------------------------------------#
# 比较实际值与期望值的差异，通过没啥事，不通过引发AssertionError
# 这儿需要两个numpy输入
# ---------------------------------------------------------#
np.testing.assert_allclose(torch_out.numpy(), ort_outs, rtol=1e-03, atol=1e-05)
if __name__ == '__main__':
# create model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = create_model(num_classes=5).to(device)
# load model weights
model_weight_path = "./output/model-25.pth"
model.load_state_dict(torch.load(model_weight_path, map_location=device))
# 据说因为BN、dropout的存在，所以这儿要转成eval()模式
model.eval()
# 导出onnx模型的输入尺寸，要和pytorch模型的输入尺寸一致
input_shape = (224, 224)
# onnx模型输出到哪里去
output_path = './output/efficientnet_b0.onnx'
# -------------------------#
# 第二轮验证
# -------------------------#
# 初始化onnx模型
ort_session_1 = onnxruntime.InferenceSession(output_path)
check_onnx_2(model, ort_session_1, input_shape, device)
print("onnx model check_2 finsh.")
5.3 warning消除记录
[W:onnxruntime:, graph.cc:1237 onnxruntime::Graph::Graph] Initializer 893 appears in graph inputs and will not be treated as constant value/weight. This may prevent some of the graph optimizations, like const folding. Move it out of graph inputs if there is no need to override it, by either re-generating the model with latest exporter/converter or with the tool onnxruntime/tools/python/remove_initializer_from_input.py.
解决方案：
在得到的onnx模型文件夹下，新建remove_initializer_from_input.py，内容为：
import onnx
import argparse
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument("--input", required=True, help="input model")
parser.add_argument("--output", required=True, help="output model")
args = parser.parse_args()
return args
def remove_initializer_from_input():
args = get_args()
model = onnx.load(args.input)
if model.ir_version < 4:
print(
'Model with ir_version below 4 requires to include initilizer in graph input'
)
return
inputs = model.graph.input
name_to_input = {}
for input in inputs:
name_to_input[input.name] = input
for initializer in model.graph.initializer:
if initializer.name in name_to_input:
inputs.remove(name_to_input[initializer.name])
onnx.save(model, args.output)
if __name__ == '__main__':
remove_initializer_from_input()
打开终端，运行python remove_initializer_from_input.py --input efficientnet_b0.onnx --output efficientnet_b0.onnx即可。
5.4 测试一张图片校验
测试代码里面，不能涉及torch，torchvision这种包了。
import onnxruntime
import numpy as np
import onnx
import os
from PIL import Image
import matplotlib.pyplot as plt
import json
def softmax_2D(X):
"""
针对二维numpy矩阵每一行进行softmax操作
X: np.array. Probably should be floats.
return: 二维矩阵
"""
# looping through rows of X
# 循环遍历X的行
ps = np.empty(X.shape)
for i in range(X.shape[0]):
ps[i,:] = np.exp(X[i,:])
ps[i,:] /= np.sum(ps[i,:])
return ps
def check_onnx_3(ort_session, img, json_path, input_shape):
# ----------------------------------------------------------------#
# 图像预处理，包括resize，归一化，减均值，除方差，HWC变为CHW，添加batch维度
# ----------------------------------------------------------------#
img = img.convert('RGB')
img_resize = img.resize(input_shape, Image.BICUBIC) # PIL.Image类型
# PIL.Image类型无法直接除以255，需要先转成array
img_resize = np.array(img_resize, dtype='float32') / 255.0
img_resize -= [0.485, 0.456, 0.406]
img_resize /= [0.229, 0.224, 0.225]
img_CHW = np.transpose(img_resize, (2, 0, 1))
# ---------------------------------------------------------#
# 添加batch_size维度，缺少这个维度，网络没法预测
# ---------------------------------------------------------#
img = np.expand_dims(img_CHW, 0)
# -----------------------------------#
# class_indict用于可视化类别
# -----------------------------------#
with open(json_path, "r") as f:
class_indict = json.load(f)
# -----------------------------------#
# onnx模型推理
# 初始化数据，注意此时img是numpy格式
# -----------------------------------#
ort_inputs = {ort_session.get_inputs()[0].name: img}
ort_outs = ort_session.run(None, ort_inputs) # 推理得到输出
# print(ort_outs) # [array([[-4.290639 , -2.267056 , 7.666328 , -1.4162455 , 0.57391334]], dtype=float32)]
# -----------------------------------#
# 经过softmax转化为概率
# softmax_2D按行转化，一行一个样本
# -----------------------------------#
predict_probability = softmax_2D(ort_outs[0])
# print(predict_probability) # array([[0.1],[0.2],[0.3],[0.3],[0.1]])
# -----------------------------------#
# argmax得到最大概率索引，也就是类别对应索引
# -----------------------------------#
predict_cla = np.argmax(predict_probability, axis=-1)
# print(predict_cla) # array([2])
print_res = "class: {} prob: {:.3}".format(class_indict[str(predict_cla[0])],
predict_probability[0][predict_cla[0]])
plt.title(print_res)
for i in range(len(predict_probability[0])):
print("class: {:10} prob: {:.3}".format(class_indict[str(i)],
predict_probability[0][i]))
plt.show()
if __name__ == '__main__':
output_path = './output/efficientnet_b0.onnx'
# -------------------------#
# 第三轮验证
# -------------------------#
# load image
img_path = "./data/rose.jpg"
assert os.path.exists(img_path), "file: '{}' dose not exist.".format(img_path)
img = Image.open(img_path)
plt.imshow(img)
# read class_indict
json_path = './class_indices.json'
# 加载onnx模型
ort_session_2 = onnxruntime.InferenceSession(output_path)
check_onnx_3(ort_session_2, img, json_path, input_shape)
print("onnx model check_3 finsh.")
6 整合到一起的代码
model_convert_onnx.py代码如下：
import onnxruntime
import torch
import numpy as np
import onnx
import os
from PIL import Image
import matplotlib.pyplot as plt
import json
# 导入网络结构
from model import efficientnet_b0 as create_model
def softmax_2D(X):
"""
针对二维numpy矩阵每一行进行softmax操作
X: np.array. Probably should be floats.
return: 二维矩阵
"""
# looping through rows of X
# 循环遍历X的行
ps = np.empty(X.shape)
for i in range(X.shape[0]):
ps[i,:] = np.exp(X[i,:])
ps[i,:] /= np.sum(ps[i,:])
return ps
def model_convert_onnx(model, input_shape, output_path):
dummy_input = torch.randn(1, 3, input_shape[0], input_shape[1])
input_names = ["input1"]
output_names = ["output1"]
torch.onnx.export(
model,
dummy_input,
output_path,
verbose=False,
keep_initializers_as_inputs=False,
opset_version=11, # 版本通常为10 or 11
input_names=input_names,
output_names=output_names,
)
def check_onnx_2(model, ort_session, input_shape, device):
# -----------------------------------#
# 给个模型输入，分辨率要对
# -----------------------------------#
x = torch.randn(size=(1, 3, input_shape[0], input_shape[1]), dtype=torch.float32)
# -----------------------------------#
# torch模型推理
# -----------------------------------#
with torch.no_grad():
torch_out = model(x.to(device))
# print(torch_out) # tensor([[-0.5728, 0.1695, -0.3256, 1.1357, -0.4081]])
# print(type(torch_out)) # <class 'torch.Tensor'>
# -----------------------------------#
# onnx模型推理
# -----------------------------------#
ort_inputs = {ort_session.get_inputs()[0].name: x.numpy()} # 初始化数据，注意这儿的x是上面的输入数据x，后期应该是img
ort_outs = ort_session.run(None, ort_inputs) # 推理得到输出
# print(ort_outs) # [array([[-0.5727689 , 0.16947027, -0.32555276, 1.13574 , -0.40812433]], dtype=float32)]
# print(type(ort_outs)) # <class 'list'>，里面是个numpy矩阵
# print(type(ort_outs[0])) # <class 'numpy.ndarray'>
ort_outs = ort_outs[0] # 因此这儿需要把内部numpy矩阵取出来，这一步很有必要
# print(torch_out.numpy().shape) # (1, 5)，1张图片，该图片属于5个类别的概率
# print(ort_outs.shape) # (1, 5)
# ---------------------------------------------------------#
# 比较实际值与期望值的差异，通过没啥事，不通过引发AssertionError
# 这儿需要两个numpy输入
# ---------------------------------------------------------#
np.testing.assert_allclose(torch_out.numpy(), ort_outs, rtol=1e-03, atol=1e-05)
def check_onnx_3(ort_session, img, json_path):
# ----------------------------------------------------------------#
# 图像预处理，包括resize，归一化，减均值，除方差，HWC变为CHW，添加batch维度
# ----------------------------------------------------------------#
img = img.convert('RGB')
img_resize = img.resize(input_shape, Image.BICUBIC) # PIL.Image类型
# PIL.Image类型无法直接除以255，需要先转成array
img_resize = np.array(img_resize, dtype='float32') / 255.0
img_resize -= [0.485, 0.456, 0.406]
img_resize /= [0.229, 0.224, 0.225]
img_CHW = np.transpose(img_resize, (2, 0, 1))
# ---------------------------------------------------------#
# 添加batch_size维度，缺少这个维度，网络没法预测
# ---------------------------------------------------------#
img = np.expand_dims(img_CHW, 0)
# -----------------------------------#
# class_indict用于可视化类别
# -----------------------------------#
with open(json_path, "r") as f:
class_indict = json.load(f)
# -----------------------------------#
# onnx模型推理
# 初始化数据，注意此时img是numpy格式
# -----------------------------------#
ort_inputs = {ort_session.get_inputs()[0].name: img}
ort_outs = ort_session.run(None, ort_inputs) # 推理得到输出
# print(ort_outs) # [array([[-4.290639 , -2.267056 , 7.666328 , -1.4162455 , 0.57391334]], dtype=float32)]
# -----------------------------------#
# 经过softmax转化为概率
# softmax_2D按行转化，一行一个样本
# -----------------------------------#
predict_probability = softmax_2D(ort_outs[0])
# print(predict_probability) # array([[0.1],[0.2],[0.3],[0.3],[0.1]])
# -----------------------------------#
# argmax得到最大概率索引，也就是类别对应索引
# -----------------------------------#
predict_cla = np.argmax(predict_probability, axis=-1)
# print(predict_cla) # array([2])
print_res = "class: {} prob: {:.3}".format(class_indict[str(predict_cla[0])],
predict_probability[0][predict_cla[0]])
plt.title(print_res)
for i in range(len(predict_probability[0])):
print("class: {:10} prob: {:.3}".format(class_indict[str(i)],
predict_probability[0][i]))
plt.show()
if __name__ == '__main__':
# create model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = create_model(num_classes=5).to(device)
# load model weights
model_weight_path = "./output/model-25.pth"
model.load_state_dict(torch.load(model_weight_path, map_location=device))
# 据说因为BN、dropout的存在，所以这儿要转成eval()模式
model.eval()
# 导出onnx模型的输入尺寸，要和pytorch模型的输入尺寸一致
input_shape = (224, 224)
# onnx模型输出到哪里去
output_path = './output/efficientnet_b0.onnx'
## ------------------------------------------#
## pth模型转换为onnx模型，转换完成后，可注释掉
## ------------------------------------------#
# model_convert_onnx(model, input_shape, output_path)
# print("model convert onnx finsh.")
# -------------------------#
# 第一轮验证
# -------------------------#
onnx_model = onnx.load(output_path)
onnx.checker.check_model(onnx_model)
print("onnx model check_1 finsh.")
# -------------------------#
# 第二轮验证
# -------------------------#
# 初始化onnx模型
ort_session_1 = onnxruntime.InferenceSession(output_path)
check_onnx_2(model, ort_session_1, input_shape, device)
print("onnx model check_2 finsh.")
# -------------------------#
# 第三轮验证
# -------------------------#
# load image
img_path = "./data/rose.jpg"
assert os.path.exists(img_path), "file: '{}' dose not exist.".format(img_path)
img = Image.open(img_path)
plt.imshow(img)
# read class_indict
json_path = './class_indices.json'
# 加载onnx模型
ort_session_2 = onnxruntime.InferenceSession(output_path)
check_onnx_3(ort_session_2, img, json_path, input_shape)
print("onnx model check_3 finsh.")
1512
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
