Logging PyTorch scalars

[dkumor]

With latest tensorwatch and pytorch 1.4, I would expect the following code, based on the example in readme to create a plot of loss:

import tensorwatch as tw
import time
import torch

w = tw.Watcher(filename='test.log')
s = w.create_stream(name='metric1')
w.make_notebook()

# loss is a pytorch scalar with value 2
loss = torch.tensor(2)
for i in range(1000):
    # Without the next line, plotting does not work
    # loss = loss.reshape((1,1))
    print("loss",loss)
    s.write(loss)
    loss +=1
    time.sleep(1)

However, the plot remains empty when logging the pytorch scalar. Only when I reshape it to a 1x1 matrix with loss.reshape((1,1)) does the plot update in jupyter notebook.

I personally found this confusing - I think it would be very useful to have direct support for scalars here.