add_embedding expected hierarchy

[lucabergamini]

Hello again
I'm currently experimenting with the embeddings' visualization, and i'm having some troubles with the expected folder hierarchy. As far as I have understand, the add_embedding function does not have any timestep parameters (although there is one in the function called with value None), so it does not provide a time distributed visualization. The current hierarchy of my loggings is the following:

-project
--runs
---August14_11:24:12
----event.out
---August14_11:28:12
----event.out

i.e. every running gets his own directory with the SummaryWriter events (file event.out), and i can run everything using tensorboard from the project dir (tensorboard --logdir runs).

If I use add_embedding for one of the experiments with the following notation

add_embedding(torch.FloatTensor(out_numpy),save_path='runs/August14_11:28:12',metadata=[k.data.numpy()[0]  for k in label_batch ])

I get one warning each time complaining about the fact the dir already exists, and the embedding log file get overwritten each time, so I can see only the last one written.

However if I log to a subdir with the following notation

add_embedding(torch.FloatTensor(out_numpy),save_path='runs/August14_11:28:12/emb{}'.format(unique_id),metadata=[k.data.numpy()[0]  for k in label_batch ])

I can't see any embedding in the visualization's page, and i have to run tensorboard directly in the embedding log directory (e.g. tensorboard --logdir runs/August14_11:28:12/emb0), which is a bit of a pain if I log every 25 batch.

Any solutions of best practices?

[lanpa]

This will be fixed together with #5. Maybe this weekend.

[lucabergamini]

I did not see it, my bad :) Keep up the good work :)

[lucabergamini]

I've written a class named EmbeddingWriter which may be useful as baseline to include embeddings in the SummaryWriter. It use a single projector_config file and a level of subfolders to hold the different timestep, each one including the metadata,tensors and sprite. The results seem pretty good for now as shown below , should i submit a PR?

[lanpa]

Of course!