Spaces:
Running on Zero
Running on Zero
File size: 4,598 Bytes
4a9b62b | 1 2 3 4 5 6 7 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 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 | # Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
"""
Layer norm done in fp32 (for fp16 training)
"""
import torch.nn as nn
import torch.nn.functional as F
import torch
from math import ceil, floor
import torch
def torch_quantile(
input: torch.Tensor,
q: float | torch.Tensor,
dim: int | None = None,
keepdim: bool = False,
*,
interpolation: str = "nearest",
out: torch.Tensor | None = None,
) -> torch.Tensor:
"""Better torch.quantile for one SCALAR quantile.
Using torch.kthvalue. Better than torch.quantile because:
- No 2**24 input size limit (pytorch/issues/67592),
- Much faster, at least on big input sizes.
Arguments:
input (torch.Tensor): See torch.quantile.
q (float): See torch.quantile. Supports only scalar input
currently.
dim (int | None): See torch.quantile.
keepdim (bool): See torch.quantile. Supports only False
currently.
interpolation: {"nearest", "lower", "higher"}
See torch.quantile.
out (torch.Tensor | None): See torch.quantile. Supports only
None currently.
"""
# Sanitization: q
try:
q = float(q)
assert 0 <= q <= 1
except Exception:
raise ValueError(f"Only scalar input 0<=q<=1 is currently supported (got {q})!")
# Sanitization: dim
# Because one cannot pass `dim=None` to `squeeze()` or `kthvalue()`
if dim_was_none := dim is None:
dim = 0
input = input.reshape((-1,) + (1,) * (input.ndim - 1))
# Sanitization: inteporlation
if interpolation == "nearest":
inter = round
elif interpolation == "lower":
inter = floor
elif interpolation == "higher":
inter = ceil
else:
raise ValueError(
"Supported interpolations currently are {'nearest', 'lower', 'higher'} "
f"(got '{interpolation}')!"
)
# Sanitization: out
if out is not None:
raise ValueError(f"Only None value is currently supported for out (got {out})!")
# Logic
k = inter(q * (input.shape[dim] - 1)) + 1
out = torch.kthvalue(input, k, dim, keepdim=True, out=out)[0]
# Rectification: keepdim
if keepdim:
return out
if dim_was_none:
return out.squeeze()
else:
return out.squeeze(dim)
return out
class Fp32GroupNorm(nn.GroupNorm):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
def forward(self, input):
# print("In FP32 GroupNorm", input[0].mean(), input[0].std(), input[0].min(), input[0].max())
output = F.group_norm(
input.float(),
self.num_groups,
self.weight.float() if self.weight is not None else None,
self.bias.float() if self.bias is not None else None,
self.eps,
)
# NOTE: dirty hack
# if input.max()<1000:
# output = F.group_norm(
# input.float(),
# self.num_groups,
# self.weight.float() if self.weight is not None else None,
# self.bias.float() if self.bias is not None else None,
# self.eps,
# )
# else:
# print("Will remove outliers in a dirty way in FP32GroupNorm")
# q1 = torch_quantile(input.view(-1), 0.25)
# q3 = torch_quantile(input.view(-1), 0.75)
# IQR = q3-q1
# upper_bound = q3 + 1.5 * IQR
# lower_bound = q1 - 1.5 * IQR
# # correct_mean = input[(lower_bound<input) & (input<upper_bound)].mean()
# # correct_std = input[(lower_bound<input) & (input<upper_bound)].std()
# # print(f"After removing outliers, the correct mean and std are {correct_mean}, {correct_std}")
# # output = (input-correct_mean)/(correct_std+self.eps) * self.weight.float().unsqueeze(0).unsqueeze(-1) + self.bias.unsqueeze(0).unsqueeze(-1)
# input[(lower_bound<input) & (input<upper_bound)] = 0.0
# output = F.group_norm(
# input.float(),
# self.num_groups,
# self.weight.float() if self.weight is not None else None,
# self.bias.float() if self.bias is not None else None,
# self.eps,
# )
return output.type_as(input)
|