|
5 | 5 | ------
|
6 | 6 |
|
7 | 7 | (Add your change to a random empty line to avoid merge conflicts)
|
8 |
| - - |
9 |
| - - |
| 8 | +- |
| 9 | +- |
| 10 | +- |
| 11 | +- |
| 12 | +- |
| 13 | +- |
| 14 | +- |
| 15 | +- |
| 16 | +- |
| 17 | +- |
| 18 | +- |
| 19 | +- |
| 20 | +- |
| 21 | +- |
| 22 | +- |
| 23 | +- |
| 24 | +- |
| 25 | +- |
| 26 | +- |
| 27 | +- |
| 28 | +- |
| 29 | +- |
| 30 | +- |
| 31 | +- |
| 32 | + |
| 33 | + |
| 34 | + |
| 35 | +0.3.5 |
| 36 | +------ |
| 37 | + |
| 38 | +Breaking changes: |
| 39 | + - You can no longer pass an int as the `kernel_size` for a `flax.linen.Conv. |
| 40 | + Instead a type error is raised stating that |
| 41 | + a tuple/list should be provided. Stride and dilation arguments do support broadcasting a single int value now because this is not |
| 42 | + ambigious when the kernel rank is known. |
| 43 | + - `flax.linen.enable_named_call` and `flax.linen.disable_named_call` now work anywhere instead of only affecting Modules constructed after the enable/disable call. Additionally, there is now `flax.linen.override_named_call` that provided a context manager to locally disable/enable named_call. |
| 44 | + - NamedTuples are no longer converted to tuples on assignment to a `linen.Module`. |
| 45 | + |
| 46 | +New features: |
| 47 | + - Flax internal stack frames are now removed from exception state traces. |
| 48 | + - Added `flax.linen.nowrap` to decorate method that should not be transformed |
| 49 | + because they are stateful. |
10 | 50 | - Flax no longer uses implicit rank broadcasting. Thus, you can now use Flax with `--jax_numpy_rank_promotion=raise`.
|
11 |
| - - |
12 |
| - - |
13 |
| - - |
14 |
| - - NamedTuples are no longer converted to tuples on assingment to a `linen.Module`. |
15 |
| - - |
16 |
| - - |
| 51 | + |
| 52 | +Bugfixes: |
| 53 | + - linen Modules and dataclasses made with `flax.struct.dataclass` or `flax.struct.PyTreeNode` are now correctly recognized as dataclasses by static analysis tools like PyLance. Autocomplete of constructors has been verified to work with VSCode. |
17 | 54 | - Fixed a bug in FrozenDict which didn't allow copying dicts with reserved names.
|
18 |
| - - |
19 |
| - - |
20 |
| - - |
21 |
| - - `flax.linen.enable_named_call` and `flax.linen.disable_named_call` now work anywhere instead of only affecting Modules constructed after the enable/disable call. Additionally, there is now `flax.linen.override_named_call` that provided a context manager to locally disable/enable named_call. |
22 |
| - - |
23 |
| - - |
24 | 55 | - Fix the serialization of named tuples. Tuple fields are no longer stored in the state dict and the named tuple class is no longer recreated ([bug](https://github.com/google/flax/issues/1429)).
|
25 |
| - - |
26 |
| - - linen Modules and dataclasses made with `flax.struct.dataclass` or `flax.struct.PyTreeNode` are now correctly recognized as dataclasses by static analysis tools like PyLance. Autocomplete of constructors has been verified to work with VSCode. |
27 |
| - - |
28 |
| - - |
29 |
| - - `flax.linen.Conv` no longer interprets an int past as kernel_size as a 1d convolution. Instead a type error is raised stating that |
30 |
| - a tuple/list should be provided. Stride and dilation arguments do support broadcasting a single int value now because this is not |
31 |
| - ambigious when the kernel rank is known. |
32 |
| - - |
| 56 | + - Mixed precision training with float16 now works correctly with the attention layers. |
| 57 | + - auto-generated linen Module `__hash__`, `__eq__`, `__repr__` no longer fail by default on non-init attributes. |
| 58 | + |
| 59 | + |
33 | 60 |
|
34 | 61 | 0.3.4
|
35 | 62 | ------
|
|
0 commit comments