Replace the GPU card number when pytorch loads the historical model, map_ Location settings

pytorch When saving the model after training , Will put the equipment number used in the training process （ for example gpu Card number cuda:0 ,cpu） Also preserved . When pytorch When reloading the historical model , The model defaults to the equipment card number during training , Load the weight onto the corresponding card number .

However , Sometimes the equipment conditions of the test process and the training process are inconsistent .
for instance ,A The main engine has four GPU card , And then we use cuda:3 Training models , And save the model .
During the test , We need to be in the customer's B Host running model , however B The host has only one piece gpu card ：cuda:0 .

If the model is loaded by default ,pytorch We can't find gpu equipment , Or some other mistake .

here , When loading, you need to make a transformation , by torch.load Appoint gpu How devices are mapped ：

according to pytorch Documents , When loading the model , You can specify the tensor Load to a specific target GPU On .
Loading methods are ：

>>> torch.load('tensors.pt')
# 1. Load all tensors onto the GPU 0
>>> torch.load('tensors.pt', map_location=torch.device('cuda:0'))
# 2. Load all tensors onto GPU 1
>>> torch.load('tensors.pt', map_location=lambda storage, loc: storage.cuda(1))
# 3. Map tensors from GPU 1 to GPU 0
>>> torch.load('tensors.pt', map_location={'cuda:1':'cuda:0'})
 
The measured found ：
Method 1 It is not loaded into the target card at all , What card did the model originally train on , Or load it onto the original old card number , Therefore, the designation failed .
Method 3, Errors are reported between codes ,location.startswith(‘cuda’): AttributeError: ‘NoneType’ object has no attribute ‘startswitch’, Analyze the code and find that this is torch Their own bug! Cheating father .
Method 2： You can handle it normally tensor All loaded to cuda1 On .

————————————————
Copyright notice ： This paper is about CSDN Blogger 「Icoding_F2014」 The original article of , follow CC 4.0 BY-SA Copyright agreement , For reprint, please attach the original source link and this statement .
Link to the original text ：https://blog.csdn.net/jmh1996/article/details/111041108