# Copyright 2021 Ross Wightman . All rights reserved.
#
# 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.
# This code is modified by Zilliz.
"""PyTorch implementations of DropBlock and DropPath (Stochastic Depth) regularization layers.
PyTorch implementations of DropBlock and DropPath (Stochastic Depth) regularization layers.
Papers:
DropBlock: A regularization method for convolutional networks (https://arxiv.org/abs/1810.12890)
Copyright 2020 Ross Wightman
"""
import torch
from torch import nn
import torch.nn.functional as F
[docs]def drop_block_2d(
x, drop_prob: float = 0.1, block_size: int = 7, gamma_scale: float = 1.0,
with_noise: bool = False, inplace: bool = False, batchwise: bool = False):
""" DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
DropBlock with an experimental gaussian noise option. This layer has been tested on a few training
runs with success, but needs further validation and possibly optimization for lower runtime impact.
"""
_, c, h, w = x.shape
total_size = w * h
clipped_block_size = min(block_size, min(w, h))
# seed_drop_rate, the gamma parameter
gamma = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
(w - block_size + 1) * (h - block_size + 1))
# Forces the block to be inside the feature map.
w_i, h_i = torch.meshgrid(torch.arange(w).to(x.device), torch.arange(h).to(x.device))
valid_block = ((w_i >= clipped_block_size // 2) & (w_i < w - (clipped_block_size - 1) // 2)) & \
((h_i >= clipped_block_size // 2) & (h_i < h - (clipped_block_size - 1) // 2))
valid_block = torch.reshape(valid_block, (1, 1, h, w)).to(dtype=x.dtype)
if batchwise:
# one mask for whole batch, quite a bit faster
uniform_noise = torch.rand((1, c, h, w), dtype=x.dtype, device=x.device)
else:
uniform_noise = torch.rand_like(x)
block_mask = ((2 - gamma - valid_block + uniform_noise) >= 1).to(dtype=x.dtype)
block_mask = -F.max_pool2d(
-block_mask,
kernel_size=clipped_block_size, # block_size,
stride=1,
padding=clipped_block_size // 2)
if with_noise:
normal_noise = torch.randn((1, c, h, w), dtype=x.dtype, device=x.device) if batchwise else torch.randn_like(x)
if inplace:
x.mul_(block_mask).add_(normal_noise * (1 - block_mask))
else:
x = x * block_mask + normal_noise * (1 - block_mask)
else:
normalize_scale = (block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)).to(x.dtype)
if inplace:
x.mul_(block_mask * normalize_scale)
else:
x = x * block_mask * normalize_scale
return x
[docs]def drop_block_fast_2d(
x: torch.Tensor, drop_prob: float = 0.1, block_size: int = 7,
gamma_scale: float = 1.0, with_noise: bool = False, inplace: bool = False, batchwise: bool = False):
""" DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
DropBlock with an experimental gaussian noise option. Simplied from above without concern for valid
block mask at edges.
"""
_, c, h, w = x.shape
total_size = w * h
clipped_block_size = min(block_size, min(w, h))
gamma = gamma_scale * drop_prob * total_size / clipped_block_size ** 2 / (
(w - block_size + 1) * (h - block_size + 1))
if batchwise:
# one mask for whole batch, quite a bit faster
block_mask = torch.rand((1, c, h, w), dtype=x.dtype, device=x.device) < gamma
else:
# mask per batch element
block_mask = torch.rand_like(x) < gamma
block_mask = F.max_pool2d(
block_mask.to(x.dtype), kernel_size=clipped_block_size, stride=1, padding=clipped_block_size // 2)
if with_noise:
normal_noise = torch.randn((1, c, h, w), dtype=x.dtype, device=x.device) if batchwise else torch.randn_like(x)
if inplace:
x.mul_(1. - block_mask).add_(normal_noise * block_mask)
else:
x = x * (1. - block_mask) + normal_noise * block_mask
else:
block_mask = 1 - block_mask
normalize_scale = (block_mask.numel() / block_mask.to(dtype=torch.float32).sum().add(1e-7)).to(dtype=x.dtype)
if inplace:
x.mul_(block_mask * normalize_scale)
else:
x = x * block_mask * normalize_scale
return x
[docs]class DropBlock2d(nn.Module):
""" DropBlock. See https://arxiv.org/pdf/1810.12890.pdf
"""
[docs] def __init__(self,
drop_prob=0.1,
block_size=7,
gamma_scale=1.0,
with_noise=False,
inplace=False,
batchwise=False,
fast=True):
super().__init__()
self.drop_prob = drop_prob
self.gamma_scale = gamma_scale
self.block_size = block_size
self.with_noise = with_noise
self.inplace = inplace
self.batchwise = batchwise
self.fast = fast # FIXME finish comparisons of fast vs not
[docs] def forward(self, x):
if not self.training or not self.drop_prob:
return x
if self.fast:
return drop_block_fast_2d(
x, self.drop_prob, self.block_size, self.gamma_scale, self.with_noise, self.inplace, self.batchwise)
else:
return drop_block_2d(
x, self.drop_prob, self.block_size, self.gamma_scale, self.with_noise, self.inplace, self.batchwise)