pt --> torchscript --> pnnx --> ncnn
import os
import torch
# 0. pt模型下载及初始化
model = torch.hub.load('pytorch/vision:v0.10.0', 'deeplabv3_resnet50', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.10.0', 'deeplabv3_resnet101', pretrained=True)
# model = torch.hub.load('pytorch/vision:v0.10.0', 'deeplabv3_mobilenet_v3_large', pretrained=True)
model.eval()
x = torch.rand(1, 3, 224, 224)
# 方法1 ：pnnx
# 1. pt --> torchscript
traced_script_module = torch.jit.trace(model, x, strict=False)
traced_script_module.save("ts.pt")
# 2. ts --> pnnx --> ncnn
os.system("pnnx ts.pt inputshape=[1,3,224,224]") # # 2022年5月25日起，pnnx默认自动量化，不需要再次optmize
# ## 方法2： onnx
# # 1. pt ---> onnx
# torch_out = torch.onnx._export(model, x, "deeplabv3_resnet50.onnx", export_params=True)
# # 2. onnx --> onnxsim
# os.system("python3 -m onnxsim deeplabv3_resnet50.onnx sim.onnx")
# # 3. onnx --> ncnn
# os.system("onnx2ncnn sim.onnx ncnn.param ncnn.bin")
# # 4. ncnn --> optmize ---> ncnn
# os.system("ncnnoptimize ncnn.param ncnn.bin opt.param opt.bin 1") # 数字0 代表fp32 ；1代表fp16
# 两种转换都成功