Add patch_tensorboard

[daavoo]

• ❗ I have followed the Contributing to DVCLive guide.

• 📖 If this PR requires documentation updates, I have created a separate PR (or issue, at least) in dvc.org and linked it here.

---

Add patch_tensorboard function.

It can override or extend existing tensorboard logging calls by redirecting those to Live.log/Live.log_image calls.

This allows to patch both raw tensorboard calls but also existing callbacks using tensorboard:

from dvclive.tensorboard import patch_tensorboard
 
patch_tensorboard()

. . .
model.fit(
    x,
    y,
    epochs=2,
    callbacks=[tf.keras.callbacks.TensorBoard()],
)

patch_tensorboard(override=False) would generate DVCLive outputs while preserving the original behavior.

[daavoo]

Not sure if patch_tensorboard should instead optionally accept a Live instance

[codecov-commenter]

Codecov Report

Merging #204 (c6458e3) into main (8471869) will decrease coverage by 0.01%.
The diff coverage is 92.20%.

@@            Coverage Diff             @@
##             main     #204      +/-   ##
==========================================
- Coverage   91.07%   91.05%   -0.02%     
==========================================
  Files          19       19              
  Lines         605      671      +66     
==========================================
+ Hits          551      611      +60     
- Misses         54       60       +6     

Impacted Files	Coverage Δ
dvclive/tensorboard.py	82.85% <82.85%> (ø)
dvclive/data/base.py	88.40% <100.00%> (+1.96%)	⬆️
dvclive/data/image.py	95.55% <100.00%> (+1.11%)	⬆️
dvclive/data/scalar.py	98.18% <100.00%> (+0.62%)	⬆️
dvclive/__init__.py

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8471869...c6458e3. Read the comment docs.

[pared]

So in case of using dvclive with tensorboard we would need to revert the patches after the execution in order to not use modified tensorboard later in our project?

Shouldn't patch_tensorboard be a context manager that would revert the patch after exiting the context?

[daavoo]

I need to test it but I think that the patch doesn't persist across python executions.

[pared]

What about the same process? I think safest way would be to let the user control when the dvclive is patching the tensorboard.

[dberenbaum]

Discussed with @daavoo that we will hold off on merging this.