Example adding various images

[gar1t]

This is inspired by #21, which wants a way to easily add a PNG as an image summary. The examples demonstrate how to do this.

FWIW I think this example is a very reasonable feature addition:

writer.add_image('sample', 'sample.png')

The rationale for closing #21 was that this sort of encoding is best left to the user. I disagree based on the following:

1. tensorboardX already uses a PIL.Image instance to encode a numpy array to PNG bytes

2. If given a PIL.Image instance, tensorboardX can skip the step of generating one from the numpy array (this is shown in the example)

3. If given a string value as the second argument to writer.add_image, tensorboardX can trivially call PIL.Image.open and use that image instance to encode the PNG bytes for the summary

This PR is valid as a stand-alone example, but I also hope it makes the case to support the following enhancements to writer.add_image:

• Support a string arg for image, which is used in a call to PIL.Image.open
• Support a PIL Image instance directly

I'm happy to implement those changes in another PR but I wanted to start this discussion with an example (useful in its own right).

[codecov-io]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.12%. Comparing base (366bc8f) to head (18e2690).
⚠️ Report is 155 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #490      +/-   ##
==========================================
+ Coverage   89.07%   89.12%   +0.05%     
==========================================
  Files          37       37              
  Lines        2562     2566       +4     
==========================================
+ Hits         2282     2287       +5     
+ Misses        280      279       -1     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

[lanpa]

Sorry for missing this PR. It would be good to minimize the overhead to convert the data back and forth. As of your suggestion, I think passing "string arg for image" might conflict with the usage of caffe2 blob name (use string to index tensor). I would suggest using writer.add_image_from_file to separate the code section instead of adding logics inside the add_image function. This function will be much more easy to support other formats in the future.

def add_image_from_file(self, tag, filename_or_PILobject, global_step=None, walltime=None):
  # if("PIL onj"):
  # if("png"):
  # if("jpg"):

What do you think?