How to import model parameters without participating in training

[zengnanhua]

How to import model parameters without participating in training

[zengnanhua]

怎么导入模型参数不参与训练 如： model = models.vgg16(pretrained = True) for param in model.parameters(): param.requires_grad = False

[shaltielshmid]

In TorchSharp it would look the same:

var model = ...;
foreach (var param in model.parameters())
    param.requires_grad = False;

[zengnanhua]

model = models.vgg16(pretrained = True) How to import the trained parameters of python tourch model

[yueyinqiu]

model = models.vgg16(pretrained = True) How to import the trained parameters of python tourch model

I suppose it could be another issue. I've tried that, but it seems that currently there are no beautiful solutions available.

If you don't mind, maybe you could download the pre-trained model with pytorch, and then try TorchSharp.PyBridge or exportsd.py to load it into csharp.

Actually torch.hub.download_url_to_file could work in csharp. But I haven't found a simple enough way to resolve such a .pkl file. Perhaps some more features should be added into TorchSharp.PyBridge (to just load a dictionary of tensor, without a module or optimizer)?

[yueyinqiu]

#713

[yueyinqiu]

I have uploaded the converted model to hugging face. Could the following code work for you?

using TorchSharp;

var huggingFace = "https://huggingface.co/";
// var huggingFace = "https://hf-mirror.com/";

var file = new FileInfo("./vgg16.dat");
if (!file.Exists)
    torch.hub.download_url_to_file(
        $"{huggingFace}yueyinqiu/vision-TorchSharp/resolve/main/VGG16_Weights.IMAGENET1K_V1",
        file.FullName);

var model = torchvision.models.vgg16(weights_file: file.FullName);