Pretrained model from torchvision perform bad compared with caffe or matconvnet

i use the pretrained vgg16_bn downloaded from torchvision for images classification.First, extracting the feature from the first full connection layers. Then, some of 4096-d features are trained for liblinear. After the SVM training, rest of the features are tested.
The same operation is implemented on the caffe and matconvnet. Comparing the three results, I find the model’s performance from torchvision is the worst.I want to know wether other people face the same problem, or my method is wrong. And i have one more question is that how can i use the pretrained model from other deep learning platform on the pytorch. Thank you in advance.

It seems this is also happen to me, if you can tell me how bad the performance was? such as top1 or top5, thx

Could you post a link to the repo of the Caffe and Matconvnet implementation of these models?
We’ve had this discussion in the past, but I can’t find the thread.
The conclusion was, as far as I remember, that the Caffe model was trained differently, i.e. an advanced model instead of the original architecture. But as I said, I try to find the thread where I compared the implementations. So take this statement with a grain of salt.