oneflow
oneflow copied to clipboard
Published
20 hours ago •
Oneflow-Inc
Oneflow-Inc
Add gelu fast activation
这个和 torch.nn.GELU(approximate='tanh') 是重复的吗 https://pytorch.org/docs/stable/generated/torch.nn.GELU.html
相关讨论：https://github.com/huggingface/transformers/issues/15397
这个和 torch.nn.GELU(approximate='tanh') 是重复的吗 https://pytorch.org/docs/stable/generated/torch.nn.GELU.html
大老师，这个和 torch.nn.GELU(approximate='tanh')不是重复的，但是都是gelu的近似，gelu的变种有点多：
class NewGELUActivation(nn.Module):
"""
Implementation of the GELU activation function currently in Google BERT repo (identical to OpenAI GPT). Also see
the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
"""
def forward(self, input: Tensor) -> Tensor:
return 0.5 * input * (1.0 + torch.tanh(math.sqrt(2.0 / math.pi) * (input + 0.044715 * torch.pow(input, 3.0))))
class GELUActivation(nn.Module):
"""
Original Implementation of the GELU activation function in Google BERT repo when initially created. For
information: OpenAI GPT's GELU is slightly different (and gives slightly different results): 0.5 * x * (1 +
torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))) This is now written in C in nn.functional
Also see the Gaussian Error Linear Units paper: https://arxiv.org/abs/1606.08415
"""
def __init__(self, use_gelu_python: bool = False):
super().__init__()
if version.parse(version.parse(torch.__version__).base_version) < version.parse("1.4") or use_gelu_python:
self.act = self._gelu_python
else:
self.act = nn.functional.gelu
def _gelu_python(self, input: Tensor) -> Tensor:
return input * 0.5 * (1.0 + torch.erf(input / math.sqrt(2.0)))
def forward(self, input: Tensor) -> Tensor:
return self.act(input)
class FastGELUActivation(nn.Module):
"""
Applies GELU approximation that is slower than QuickGELU but more accurate. See: https://github.com/hendrycks/GELUs
"""
def forward(self, input: Tensor) -> Tensor:
return 0.5 * input * (1.0 + torch.tanh(input * 0.7978845608 * (1.0 + 0.044715 * input * input)))
class QuickGELUActivation(nn.Module):
"""
Applies GELU approximation that is fast but somewhat inaccurate. See: https://github.com/hendrycks/GELUs
"""
def forward(self, input: Tensor) -> Tensor:
return input * torch.sigmoid(1.702 * input)
class ClippedGELUActivation(nn.Module):
"""
Clip the range of possible GeLU outputs between [min, max]. This is especially useful for quantization purpose, as
it allows mapping negatives values in the GeLU spectrum. For more information on this trick, please refer to
https://arxiv.org/abs/2004.09602.
Gaussian Error Linear Unit. Original Implementation of the gelu activation function in Google Bert repo when
initially created.
For information: OpenAI GPT's gelu is slightly different (and gives slightly different results): 0.5 * x * (1 +
torch.tanh(math.sqrt(2 / math.pi) * (x + 0.044715 * torch.pow(x, 3)))). See https://arxiv.org/abs/1606.08415
"""
def __init__(self, min: float, max: float):
if min > max:
raise ValueError(f"min should be < max (got min: {min}, max: {max})")
super().__init__()
self.min = min
self.max = max
def forward(self, x: Tensor) -> Tensor:
return torch.clip(gelu(x), self.min, self.max)
从文档来看确实是和 torch.nn.GELU(approximate='tanh') 公式一样的。
oneflow 文档（渲染好像不太正确，以及混用了 input 和 x）：
pytorch 文档（pytorch 文档也没正确渲染 lol）：
oneflow 文档里 input 和 x 应该是同一个东西？把 input 换成 x 并放进括号里就和 pytorch 公式一样了。这么来看上面贴的 NewGELUActivation 和 FastGELUActivation 也是一样的。