Running Inference on CPU

[suke18]

Hello,

I am trying to use run_infer.sh command on my own tiled images. However, my own MacBook (macOS) doesn't support GPU, is there a way to run built on the CPU only?

Best,
KS

[simongraham]

Hi, we have not tested this but have currently hard-coded it to work on the gpu. Please try changing the line here, to:

net = net.to("cpu")

and let us know how you get on.

[vqdang]

You also need to change this

hover_net/models/hovernet/run_desc.py

Line 176 in a0f80c7

patch_imgs_gpu = patch_imgs.to("cuda").type(torch.float32) # to NCHW

Also remove this because DataParallel is not applicable

hover_net/infer/base.py

Line 69 in a0f80c7

net = torch.nn.DataParallel(net)

There will be other places too, it will crop up once you try.

[suke18]

@vqdang @simongraham , Thanks for your reply. It works. It is slow so depending on the GPU is a better choice.

Best,
Kenong

[simongraham]

Keeping this issue open for easy tracking

[xiaoshuomin]

I would like to know that the modified model is trained on cpu, because I have encountered some problems and need to debug in cpu mode. I have commented on some statements as you said above, but I still encountered some problems, such as'Hovernet object has no attribute module'

[ajinkya-kulkarni]

Hello team, thanks for this repo! Is there an inference notebook/script available after I trained the model using run.py?

[KondratievaOlesya]

I also had to modify the following line:

hover_net/run_infer.py – Line 139

'batch_size' : int(args['batch_size']) * nr_gpus,

to:

'batch_size' : int(args['batch_size']),

because nr_gpus is 0 when no GPUs are detected at runtime, which results in a batch size of 0 and causes an error.