torch2paddle
用于torch模型到paddle模型权重的转换以及前反向对齐
安装
- 本地安装
python3 setup.py bdist_wheel
pip3 install dist/model_align-1.0.1-py3-none-any.whl --force-reinstall
- pip安装
使用方法
使用需要保证padle的模型和torch的模型结构完全一致且内部的网络名称命名一样
具体参数说明见 !pip uninstall model_ailgn这!。
使用resnet18的分类任务举例
import paddle
import torch
import paddle.nn as pn
import torch.nn as tn
from model_align import ModelAlign
from paddle.vision.models import resnet18 as paddle_resnet18
from torchvision.models import resnet18 as torch_resnet18
def paddle_loss(paddle_out, label=None):
return pn.CrossEntropyLoss()(paddle_out, label)
def torch_loss(torch_out, label=None):
return tn.CrossEntropyLoss()(torch_out, label)
paddle.set_device('cpu')
paddle_model = paddle_resnet18()
torch_model = torch_resnet18()
input_data = torch.randn((2,3,224,224))
align = ModelAlign(paddle_model,
torch_model,
paddle_loss_func=paddle_loss,
torch_loss_func=torch_loss,
input_data=input_data,
diff_threshold=100,
save_path="./output/resnet18",
iters=3,
feat_align=True)
align.convert_weight()
align.forward()
torch_input = torch.randint(0, 100, (2,))
paddle_input = paddle.to_tensor(torch_input.numpy())
align.backward(paddle_input={'label':paddle_input},
torch_input={'label':torch_input})
模型结果默认输出到 ./output中。
