E3nn

Added
- Normalization testing with `assert_normalized`
- Optional logging for equivariance and normalization tests
- Public `e3nn.util.test.format_equivariance_error` method for printing equivariance test results
- Module `o3.SphericalHarmonicsAlphaBeta`

Changed
- Generated code (modules like `TensorProduct`, `Linear`, `Extract`) now pickled using TorchScript IR, rather than Python source code.
- e3nn now only requires PyTorch >= 1.8.0 rather than 1.8.1
- Changed `o3.legendre` into a module `o3.Legendre`

Removed
- Removed `e3nn.util.codegen.eval_code` in favor of `torch.fx`

Added
- `squared` option to `o3.Norm`
- `e3nn.nn.models.v2104.voxel_convolution.Convolution` made to be resolution agnostic
- `TensorProduct.visualize` keyword argument `aspect_ratio`

Changed
- `ReducedTensorProducts` is a (scriptable) `torch.nn.Module`
- e3nn now requires the latest stable PyTorch, >=1.8.1
- `TensorProduct.visualize`: color of paths based on `w.pow(2).mean()` instead of `w.sum().sign() * w.abs().sum()`

Fixed
- No more NaN gradients of `o3.Norm`/`nn.NormActivation` at zero when using `epsilon`
- Modules with `compile_mode('trace')` can now be compiled when their dtype and the current default dtype are different
- Fix errors in `ReducedTensorProducts` and add new tests

Added
- `uuu` connection mode in `o3.TensorProduct` now has specialized code

Fixed
- Fixed an issue with `Activation` (used by `Gate`). It was only applying the first activation function provided. `Activation('0e+0e', [act1, act2])` was equivalent to `Activation('2x0e', [act1])`. Solved by removing the `.simplify()` applied to `self.irreps_in`.
- `Gate` will not accept non-scalar `irreps_gates` or `irreps_scalars`

Added
- `e3nn.util.test.random_irreps` convenience function for writing tests

Changed
- `o3.Linear` now has more efficient specialized code

Fixed
- Fixed a problem with temporary files on windows

Added
- Added `e3nn.set_optimization_defaults()` and `e3nn.get_optimization_defaults()`
- Constructors for empty `Irreps`: `Irreps()` and `Irreps("")`
- Additional tests, docs, and refactoring for `Irrep` and `Irreps`.
- Added `TensorProduct.weight_views()` and `TensorProduct.weight_view_for_instruction()`
- Fix Docs for ExtractIr

Changed
- Renamed `o3.TensorProduct` arguments in `irreps_in1`, `irreps_in2` and `irreps_out`
- Renamed `o3.spherical_harmonics` arguement `xyz` into `x`
- Renamed `math.soft_one_hot_linspace` argument `endpoint` into `cutoff`, `cutoff = not endpoint`
- Variances are now provided to `o3.TensorProduct` through explicit `in1_var`, `in2_var`, `out_var` parameters
- Submodules define `__all__`; documentation uses shorter module names for the classes/methods.

Fixed
- Enabling/disabling einsum optimization no longer affects PyTorch RNG state.

Removed
- Variances can no longer be provided to `o3.TensorProduct` in the list-of-tuple format for `irreps_in1`, etc.

Added
- `basis='smooth_finite'` option to `math.soft_one_hot_linspace`
- `math.soft_unit_step` function
- `nn.model.v2103` generic message passing model + examples of networks using it.
- `o3.TensorProduct`: is jit scriptable
- `o3.TensorProduct`: also broadcast the `weight` argument
- simple e3nn models can be saved/loaded with `torch.save()`/`torch.load()`
- JITable `o3.SphericalHarmonics` module version of `o3.spherical_harmonics`
- `in_place` option for `e3nn.util.jit` compilation functions
- New `compile_mode("unsupported")` for modules that do not support TorchScript
- flake8 settings have been added to `setup.cfg` for improved code style
- `TensorProduct.visualize()` can now plot weights
- `basis='bessel'` option to `math.soft_one_hot_linspace`

Changed
- `o3.TensorProduct` now uses `torch.fx` to generate it's code
- e3nn now requires the latest stable PyTorch, >=1.8.0
- in `soft_one_hot_linspace` the argument `base` is renamed into `basis`
- `Irreps.slices()`, do `zip(irreps.slices(), irreps)` to retrieve the old behavior
- `math.soft_one_hot_linspace` very small change in the normalization of `fourier` basis
- `normalize2mom` is now a `torch.nn.Module`
- rename arguments `set_ir_...` into `filter_ir_...`
- Renamed `e3nn.nn.Gate` argument `irreps_nonscalars` to `irreps_gated`
- Renamed `e3nn.o3.TensorProduct` arguments `x1, x2` to `x, y`

Fixed
- `nn.Gate` was crashing when the number of scalars or gates was zero
- `device` edge cases for `Gate` and `SphericalHarmonics`