Spaces:
Runtime error
Runtime error
File size: 5,822 Bytes
455a40f |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
# coding=utf-8
# Copyright 2020-present, AllenAI Authors, University of Illinois Urbana-Champaign,
# Intel Nervana Systems and the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Binarizers take a (real value) matrix as input and produce a binary (values in {0,1}) mask of the same shape.
"""
import torch
from torch import autograd
class ThresholdBinarizer(autograd.Function):
"""
Thresholdd binarizer.
Computes a binary mask M from a real value matrix S such that `M_{i,j} = 1` if and only if `S_{i,j} > \tau`
where `\tau` is a real value threshold.
Implementation is inspired from:
https://github.com/arunmallya/piggyback
Piggyback: Adapting a Single Network to Multiple Tasks by Learning to Mask Weights
Arun Mallya, Dillon Davis, Svetlana Lazebnik
"""
@staticmethod
def forward(ctx, inputs: torch.tensor, threshold: float, sigmoid: bool):
"""
Args:
inputs (`torch.FloatTensor`)
The input matrix from which the binarizer computes the binary mask.
threshold (`float`)
The threshold value (in R).
sigmoid (`bool`)
If set to ``True``, we apply the sigmoid function to the `inputs` matrix before comparing to `threshold`.
In this case, `threshold` should be a value between 0 and 1.
Returns:
mask (`torch.FloatTensor`)
Binary matrix of the same size as `inputs` acting as a mask (1 - the associated weight is
retained, 0 - the associated weight is pruned).
"""
nb_elems = inputs.numel()
nb_min = int(0.005 * nb_elems) + 1
if sigmoid:
mask = (torch.sigmoid(inputs) > threshold).type(inputs.type())
else:
mask = (inputs > threshold).type(inputs.type())
if mask.sum() < nb_min:
# We limit the pruning so that at least 0.5% (half a percent) of the weights are remaining
k_threshold = inputs.flatten().kthvalue(max(nb_elems - nb_min, 1)).values
mask = (inputs > k_threshold).type(inputs.type())
return mask
@staticmethod
def backward(ctx, gradOutput):
return gradOutput, None, None
class TopKBinarizer(autograd.Function):
"""
Top-k Binarizer.
Computes a binary mask M from a real value matrix S such that `M_{i,j} = 1` if and only if `S_{i,j}`
is among the k% highest values of S.
Implementation is inspired from:
https://github.com/allenai/hidden-networks
What's hidden in a randomly weighted neural network?
Vivek Ramanujan*, Mitchell Wortsman*, Aniruddha Kembhavi, Ali Farhadi, Mohammad Rastegari
"""
@staticmethod
def forward(ctx, inputs: torch.tensor, threshold: float):
"""
Args:
inputs (`torch.FloatTensor`)
The input matrix from which the binarizer computes the binary mask.
threshold (`float`)
The percentage of weights to keep (the rest is pruned).
`threshold` is a float between 0 and 1.
Returns:
mask (`torch.FloatTensor`)
Binary matrix of the same size as `inputs` acting as a mask (1 - the associated weight is
retained, 0 - the associated weight is pruned).
"""
# Get the subnetwork by sorting the inputs and using the top threshold %
mask = inputs.clone()
_, idx = inputs.flatten().sort(descending=True)
j = int(threshold * inputs.numel())
# flat_out and mask access the same memory.
flat_out = mask.flatten()
flat_out[idx[j:]] = 0
flat_out[idx[:j]] = 1
return mask
@staticmethod
def backward(ctx, gradOutput):
return gradOutput, None
class MagnitudeBinarizer(object):
"""
Magnitude Binarizer.
Computes a binary mask M from a real value matrix S such that `M_{i,j} = 1` if and only if `S_{i,j}`
is among the k% highest values of |S| (absolute value).
Implementation is inspired from https://github.com/NervanaSystems/distiller/blob/2291fdcc2ea642a98d4e20629acb5a9e2e04b4e6/distiller/pruning/automated_gradual_pruner.py#L24
"""
@staticmethod
def apply(inputs: torch.tensor, threshold: float):
"""
Args:
inputs (`torch.FloatTensor`)
The input matrix from which the binarizer computes the binary mask.
This input marix is typically the weight matrix.
threshold (`float`)
The percentage of weights to keep (the rest is pruned).
`threshold` is a float between 0 and 1.
Returns:
mask (`torch.FloatTensor`)
Binary matrix of the same size as `inputs` acting as a mask (1 - the associated weight is
retained, 0 - the associated weight is pruned).
"""
# Get the subnetwork by sorting the inputs and using the top threshold %
mask = inputs.clone()
_, idx = inputs.abs().flatten().sort(descending=True)
j = int(threshold * inputs.numel())
# flat_out and mask access the same memory.
flat_out = mask.flatten()
flat_out[idx[j:]] = 0
flat_out[idx[:j]] = 1
return mask
|