To avoid CPU -> GPU data transport

[Nanco-L]

1. Add map_location='cuda'. You can set 'cpu' instead for CPU only mode
   

   SIMPLE-NN_v2/simple_nn_v2/models/data_handler.py

   Line 60 in 28007d6

   return torch.load(self.filelist[idx])

2. Remove all .to(device) in training code

3. Set pin_memory=False in
   

   SIMPLE-NN_v2/simple_nn_v2/models/data_handler.py

   Line 223 in 28007d6

   num_workers=inputs['neural_network']['workers'], pin_memory=True)

   
   pin_memory only works CPU->GPU data transport.

4. Add .to(device) for scale and pca (Since all data is uploaded in the target device and collate_fn also works on that device).

5. Add .to(device) for all tensors generated in the code.

In my case, training speed increase by 5~10 times.