Torchexplorer

**Vega spec version 3d**

Release notes:
1. Add `attach` function for displaying intermediate values in `forward`
2. Add support for multiple `forward` invocations in training loop
3. Make input/output/attach nodes show up as Tensor nodes in Vega
4. Add documentation site
5. Fix double-hooking for non tracking hooks
6. Small visual tweaks

**Vega spec version 3c**

First stable release! Only a few small changes from 0.5.1.

Release notes:
1. Small tweaks to make logging happen on first epoch, fixing histogram order, better standalone reliability
2. Help button on the vega interface, help image in the README

**Vega spec version 3b**

Last "alpha" release before 1.0.0. Version 0.5.0 package is screwed up so this is the replacement.

Release notes:

1. Major code refactoring, both structural and line-by-line to improve clarity and modularity
2. Input/output nodes have tooltips showing tensor type and shape. They can be dragged into the right panel to visualize.
3. Small Vega improvements: magnifying glass locates parent node, histogram title and grad/grad norm text separated, left-align panel headers
4. Confirm desired functionality for nn.Embedding module, expand test suite
5. Documentation improvements

**Vega spec version 2a**

Release notes:
1. Add compatability for Python 3.8-3.11
2. Caching of graphviz calls (before graphviz was re-executed every log)
3. Nice tooltips in the Vega interface, showing input/output shapes and Module parameters
4. Small bugfixes and api default adjustments

**Vega spec version: 1c**

Release notes:
1. Fix histogram outlier rejection bug
2. Support for watching multiple modules in one training run
3. Better logging with wandb: immediate first-step structure logging and set `commit=False`
4. Add workaround for `ulimit` bug
5. Add "close panel" and "open module" buttons to the UI
5. Separate the single "Input" / "Output" node into several nodes in the explorer, if a Module has more than one input / output

**Vega spec version: 0a**

Release notes:
1. The code now typechecks cleanly via mypy
2. Small improvements to the hook `should_log_callable` function
3. Before, parameter gradients norms were being logged; it's now corrected to just the raw parameter gradients
4. Unnecessary development dependencies are separated from the `install_requires`