zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-08-21 12:06:42 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

updated calls to use loc &scale (#643)

Browse Source
This commit is contained in:
LeonEricsson 2024-02-08 18:01:59 +01:00 committed by GitHub
parent 1b97b2958b
commit 7dccd42133
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 4 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
python/mlx/nn/init.py
View File

@ -60,7 +60,7 @@ def normal(
""" """
def initializer(a: mx.array) -> mx.array: def initializer(a: mx.array) -> mx.array:
return std * mx.random.normal(shape=a.shape, dtype=dtype) + mean return mx.random.normal(shape=a.shape, scale=std, loc=mean, dtype=dtype)
return initializer return initializer
@ -184,7 +184,7 @@ def glorot_normal(
def initializer(a: mx.array, gain: float = 1.0) -> mx.array: def initializer(a: mx.array, gain: float = 1.0) -> mx.array:
fan_in, fan_out = _calculate_fan_in_fan_out(a) fan_in, fan_out = _calculate_fan_in_fan_out(a)
std = gain * math.sqrt(2.0 / (fan_in + fan_out)) std = gain * math.sqrt(2.0 / (fan_in + fan_out))
return mx.random.normal(shape=a.shape, dtype=dtype) * std return mx.random.normal(shape=a.shape, scale=std, dtype=dtype)
return initializer return initializer
@ -285,7 +285,7 @@ def he_normal(
raise ValueError(f"Invalid mode: {mode}. Valid modes are: fan_in, fan_out") raise ValueError(f"Invalid mode: {mode}. Valid modes are: fan_in, fan_out")
std = gain / math.sqrt(fan) std = gain / math.sqrt(fan)
return mx.random.normal(shape=a.shape, dtype=dtype) * std return mx.random.normal(shape=a.shape, scale=std, dtype=dtype)
return initializer return initializer

2
python/mlx/nn/layers/embedding.py
View File

@ -21,7 +21,7 @@ class Embedding(Module):
def __init__(self, num_embeddings: int, dims: int): def __init__(self, num_embeddings: int, dims: int):
super().__init__() super().__init__()
scale = math.sqrt(1 / dims) scale = math.sqrt(1 / dims)
self.weight = mx.random.normal((num_embeddings, dims)) * scale self.weight = mx.random.normal(shape=(num_embeddings, dims), scale=scale)
def _extra_repr(self): def _extra_repr(self):
return f"{self.weight.shape[0]}, {self.weight.shape[1]}" return f"{self.weight.shape[0]}, {self.weight.shape[1]}"