forked from KEMT/zpwiki
Nahrát soubory do „pages/students/2016/lukas_pokryvka/dp2021/lungCancer/util“
This commit is contained in:
parent
526fe93ff9
commit
4dca468612
Binary file not shown.
@ -0,0 +1,331 @@
|
||||
import math
|
||||
import random
|
||||
import warnings
|
||||
|
||||
import numpy as np
|
||||
import scipy.ndimage
|
||||
|
||||
import torch
|
||||
from torch.autograd import Function
|
||||
from torch.autograd.function import once_differentiable
|
||||
import torch.backends.cudnn as cudnn
|
||||
|
||||
from util.logconf import logging
|
||||
log = logging.getLogger(__name__)
|
||||
# log.setLevel(logging.WARN)
|
||||
# log.setLevel(logging.INFO)
|
||||
log.setLevel(logging.DEBUG)
|
||||
|
||||
def cropToShape(image, new_shape, center_list=None, fill=0.0):
|
||||
# log.debug([image.shape, new_shape, center_list])
|
||||
# assert len(image.shape) == 3, repr(image.shape)
|
||||
|
||||
if center_list is None:
|
||||
center_list = [int(image.shape[i] / 2) for i in range(3)]
|
||||
|
||||
crop_list = []
|
||||
for i in range(0, 3):
|
||||
crop_int = center_list[i]
|
||||
if image.shape[i] > new_shape[i] and crop_int is not None:
|
||||
|
||||
# We can't just do crop_int +/- shape/2 since shape might be odd
|
||||
# and ints round down.
|
||||
start_int = crop_int - int(new_shape[i]/2)
|
||||
end_int = start_int + new_shape[i]
|
||||
crop_list.append(slice(max(0, start_int), end_int))
|
||||
else:
|
||||
crop_list.append(slice(0, image.shape[i]))
|
||||
|
||||
# log.debug([image.shape, crop_list])
|
||||
image = image[crop_list]
|
||||
|
||||
crop_list = []
|
||||
for i in range(0, 3):
|
||||
if image.shape[i] < new_shape[i]:
|
||||
crop_int = int((new_shape[i] - image.shape[i]) / 2)
|
||||
crop_list.append(slice(crop_int, crop_int + image.shape[i]))
|
||||
else:
|
||||
crop_list.append(slice(0, image.shape[i]))
|
||||
|
||||
# log.debug([image.shape, crop_list])
|
||||
new_image = np.zeros(new_shape, dtype=image.dtype)
|
||||
new_image[:] = fill
|
||||
new_image[crop_list] = image
|
||||
|
||||
return new_image
|
||||
|
||||
|
||||
def zoomToShape(image, new_shape, square=True):
|
||||
# assert image.shape[-1] in {1, 3, 4}, repr(image.shape)
|
||||
|
||||
if square and image.shape[0] != image.shape[1]:
|
||||
crop_int = min(image.shape[0], image.shape[1])
|
||||
new_shape = [crop_int, crop_int, image.shape[2]]
|
||||
image = cropToShape(image, new_shape)
|
||||
|
||||
zoom_shape = [new_shape[i] / image.shape[i] for i in range(3)]
|
||||
|
||||
with warnings.catch_warnings():
|
||||
warnings.simplefilter("ignore")
|
||||
image = scipy.ndimage.interpolation.zoom(
|
||||
image, zoom_shape,
|
||||
output=None, order=0, mode='nearest', cval=0.0, prefilter=True)
|
||||
|
||||
return image
|
||||
|
||||
def randomOffset(image_list, offset_rows=0.125, offset_cols=0.125):
|
||||
|
||||
center_list = [int(image_list[0].shape[i] / 2) for i in range(3)]
|
||||
center_list[0] += int(offset_rows * (random.random() - 0.5) * 2)
|
||||
center_list[1] += int(offset_cols * (random.random() - 0.5) * 2)
|
||||
center_list[2] = None
|
||||
|
||||
new_list = []
|
||||
for image in image_list:
|
||||
new_image = cropToShape(image, image.shape, center_list)
|
||||
new_list.append(new_image)
|
||||
|
||||
return new_list
|
||||
|
||||
|
||||
def randomZoom(image_list, scale=None, scale_min=0.8, scale_max=1.3):
|
||||
if scale is None:
|
||||
scale = scale_min + (scale_max - scale_min) * random.random()
|
||||
|
||||
new_list = []
|
||||
for image in image_list:
|
||||
# assert image.shape[-1] in {1, 3, 4}, repr(image.shape)
|
||||
|
||||
with warnings.catch_warnings():
|
||||
warnings.simplefilter("ignore")
|
||||
# log.info([image.shape])
|
||||
zimage = scipy.ndimage.interpolation.zoom(
|
||||
image, [scale, scale, 1.0],
|
||||
output=None, order=0, mode='nearest', cval=0.0, prefilter=True)
|
||||
image = cropToShape(zimage, image.shape)
|
||||
|
||||
new_list.append(image)
|
||||
|
||||
return new_list
|
||||
|
||||
|
||||
_randomFlip_transform_list = [
|
||||
# lambda a: np.rot90(a, axes=(0, 1)),
|
||||
# lambda a: np.flip(a, 0),
|
||||
lambda a: np.flip(a, 1),
|
||||
]
|
||||
|
||||
def randomFlip(image_list, transform_bits=None):
|
||||
if transform_bits is None:
|
||||
transform_bits = random.randrange(0, 2 ** len(_randomFlip_transform_list))
|
||||
|
||||
new_list = []
|
||||
for image in image_list:
|
||||
# assert image.shape[-1] in {1, 3, 4}, repr(image.shape)
|
||||
|
||||
for n in range(len(_randomFlip_transform_list)):
|
||||
if transform_bits & 2**n:
|
||||
# prhist(image, 'before')
|
||||
image = _randomFlip_transform_list[n](image)
|
||||
# prhist(image, 'after ')
|
||||
|
||||
new_list.append(image)
|
||||
|
||||
return new_list
|
||||
|
||||
|
||||
def randomSpin(image_list, angle=None, range_tup=None, axes=(0, 1)):
|
||||
if range_tup is None:
|
||||
range_tup = (0, 360)
|
||||
|
||||
if angle is None:
|
||||
angle = range_tup[0] + (range_tup[1] - range_tup[0]) * random.random()
|
||||
|
||||
new_list = []
|
||||
for image in image_list:
|
||||
# assert image.shape[-1] in {1, 3, 4}, repr(image.shape)
|
||||
|
||||
image = scipy.ndimage.interpolation.rotate(
|
||||
image, angle, axes=axes, reshape=False,
|
||||
output=None, order=0, mode='nearest', cval=0.0, prefilter=True)
|
||||
|
||||
new_list.append(image)
|
||||
|
||||
return new_list
|
||||
|
||||
|
||||
def randomNoise(image_list, noise_min=-0.1, noise_max=0.1):
|
||||
noise = np.zeros_like(image_list[0])
|
||||
noise += (noise_max - noise_min) * np.random.random_sample(image_list[0].shape) + noise_min
|
||||
noise *= 5
|
||||
noise = scipy.ndimage.filters.gaussian_filter(noise, 3)
|
||||
# noise += (noise_max - noise_min) * np.random.random_sample(image_hsv.shape) + noise_min
|
||||
|
||||
new_list = []
|
||||
for image_hsv in image_list:
|
||||
image_hsv = image_hsv + noise
|
||||
|
||||
new_list.append(image_hsv)
|
||||
|
||||
return new_list
|
||||
|
||||
|
||||
def randomHsvShift(image_list, h=None, s=None, v=None,
|
||||
h_min=-0.1, h_max=0.1,
|
||||
s_min=0.5, s_max=2.0,
|
||||
v_min=0.5, v_max=2.0):
|
||||
if h is None:
|
||||
h = h_min + (h_max - h_min) * random.random()
|
||||
if s is None:
|
||||
s = s_min + (s_max - s_min) * random.random()
|
||||
if v is None:
|
||||
v = v_min + (v_max - v_min) * random.random()
|
||||
|
||||
new_list = []
|
||||
for image_hsv in image_list:
|
||||
# assert image_hsv.shape[-1] == 3, repr(image_hsv.shape)
|
||||
|
||||
image_hsv[:,:,0::3] += h
|
||||
image_hsv[:,:,1::3] = image_hsv[:,:,1::3] ** s
|
||||
image_hsv[:,:,2::3] = image_hsv[:,:,2::3] ** v
|
||||
|
||||
new_list.append(image_hsv)
|
||||
|
||||
return clampHsv(new_list)
|
||||
|
||||
|
||||
def clampHsv(image_list):
|
||||
new_list = []
|
||||
for image_hsv in image_list:
|
||||
image_hsv = image_hsv.clone()
|
||||
|
||||
# Hue wraps around
|
||||
image_hsv[:,:,0][image_hsv[:,:,0] > 1] -= 1
|
||||
image_hsv[:,:,0][image_hsv[:,:,0] < 0] += 1
|
||||
|
||||
# Everything else clamps between 0 and 1
|
||||
image_hsv[image_hsv > 1] = 1
|
||||
image_hsv[image_hsv < 0] = 0
|
||||
|
||||
new_list.append(image_hsv)
|
||||
|
||||
return new_list
|
||||
|
||||
|
||||
# def torch_augment(input):
|
||||
# theta = random.random() * math.pi * 2
|
||||
# s = math.sin(theta)
|
||||
# c = math.cos(theta)
|
||||
# c1 = 1 - c
|
||||
# axis_vector = torch.rand(3, device='cpu', dtype=torch.float64)
|
||||
# axis_vector -= 0.5
|
||||
# axis_vector /= axis_vector.abs().sum()
|
||||
# l, m, n = axis_vector
|
||||
#
|
||||
# matrix = torch.tensor([
|
||||
# [l*l*c1 + c, m*l*c1 - n*s, n*l*c1 + m*s, 0],
|
||||
# [l*m*c1 + n*s, m*m*c1 + c, n*m*c1 - l*s, 0],
|
||||
# [l*n*c1 - m*s, m*n*c1 + l*s, n*n*c1 + c, 0],
|
||||
# [0, 0, 0, 1],
|
||||
# ], device=input.device, dtype=torch.float32)
|
||||
#
|
||||
# return th_affine3d(input, matrix)
|
||||
|
||||
|
||||
|
||||
|
||||
# following from https://github.com/ncullen93/torchsample/blob/master/torchsample/utils.py
|
||||
# MIT licensed
|
||||
|
||||
# def th_affine3d(input, matrix):
|
||||
# """
|
||||
# 3D Affine image transform on torch.Tensor
|
||||
# """
|
||||
# A = matrix[:3,:3]
|
||||
# b = matrix[:3,3]
|
||||
#
|
||||
# # make a meshgrid of normal coordinates
|
||||
# coords = th_iterproduct(input.size(-3), input.size(-2), input.size(-1), dtype=torch.float32)
|
||||
#
|
||||
# # shift the coordinates so center is the origin
|
||||
# coords[:,0] = coords[:,0] - (input.size(-3) / 2. - 0.5)
|
||||
# coords[:,1] = coords[:,1] - (input.size(-2) / 2. - 0.5)
|
||||
# coords[:,2] = coords[:,2] - (input.size(-1) / 2. - 0.5)
|
||||
#
|
||||
# # apply the coordinate transformation
|
||||
# new_coords = coords.mm(A.t().contiguous()) + b.expand_as(coords)
|
||||
#
|
||||
# # shift the coordinates back so origin is origin
|
||||
# new_coords[:,0] = new_coords[:,0] + (input.size(-3) / 2. - 0.5)
|
||||
# new_coords[:,1] = new_coords[:,1] + (input.size(-2) / 2. - 0.5)
|
||||
# new_coords[:,2] = new_coords[:,2] + (input.size(-1) / 2. - 0.5)
|
||||
#
|
||||
# # map new coordinates using bilinear interpolation
|
||||
# input_transformed = th_trilinear_interp3d(input, new_coords)
|
||||
#
|
||||
# return input_transformed
|
||||
#
|
||||
#
|
||||
# def th_trilinear_interp3d(input, coords):
|
||||
# """
|
||||
# trilinear interpolation of 3D torch.Tensor image
|
||||
# """
|
||||
# # take clamp then floor/ceil of x coords
|
||||
# x = torch.clamp(coords[:,0], 0, input.size(-3)-2)
|
||||
# x0 = x.floor()
|
||||
# x1 = x0 + 1
|
||||
# # take clamp then floor/ceil of y coords
|
||||
# y = torch.clamp(coords[:,1], 0, input.size(-2)-2)
|
||||
# y0 = y.floor()
|
||||
# y1 = y0 + 1
|
||||
# # take clamp then floor/ceil of z coords
|
||||
# z = torch.clamp(coords[:,2], 0, input.size(-1)-2)
|
||||
# z0 = z.floor()
|
||||
# z1 = z0 + 1
|
||||
#
|
||||
# stride = torch.tensor(input.stride()[-3:], dtype=torch.int64, device=input.device)
|
||||
# x0_ix = x0.mul(stride[0]).long()
|
||||
# x1_ix = x1.mul(stride[0]).long()
|
||||
# y0_ix = y0.mul(stride[1]).long()
|
||||
# y1_ix = y1.mul(stride[1]).long()
|
||||
# z0_ix = z0.mul(stride[2]).long()
|
||||
# z1_ix = z1.mul(stride[2]).long()
|
||||
#
|
||||
# # input_flat = th_flatten(input)
|
||||
# input_flat = x.contiguous().view(x[0], x[1], -1)
|
||||
#
|
||||
# vals_000 = input_flat[:, :, x0_ix+y0_ix+z0_ix]
|
||||
# vals_001 = input_flat[:, :, x0_ix+y0_ix+z1_ix]
|
||||
# vals_010 = input_flat[:, :, x0_ix+y1_ix+z0_ix]
|
||||
# vals_011 = input_flat[:, :, x0_ix+y1_ix+z1_ix]
|
||||
# vals_100 = input_flat[:, :, x1_ix+y0_ix+z0_ix]
|
||||
# vals_101 = input_flat[:, :, x1_ix+y0_ix+z1_ix]
|
||||
# vals_110 = input_flat[:, :, x1_ix+y1_ix+z0_ix]
|
||||
# vals_111 = input_flat[:, :, x1_ix+y1_ix+z1_ix]
|
||||
#
|
||||
# xd = x - x0
|
||||
# yd = y - y0
|
||||
# zd = z - z0
|
||||
# xm1 = 1 - xd
|
||||
# ym1 = 1 - yd
|
||||
# zm1 = 1 - zd
|
||||
#
|
||||
# x_mapped = (
|
||||
# vals_000.mul(xm1).mul(ym1).mul(zm1) +
|
||||
# vals_001.mul(xm1).mul(ym1).mul(zd) +
|
||||
# vals_010.mul(xm1).mul(yd).mul(zm1) +
|
||||
# vals_011.mul(xm1).mul(yd).mul(zd) +
|
||||
# vals_100.mul(xd).mul(ym1).mul(zm1) +
|
||||
# vals_101.mul(xd).mul(ym1).mul(zd) +
|
||||
# vals_110.mul(xd).mul(yd).mul(zm1) +
|
||||
# vals_111.mul(xd).mul(yd).mul(zd)
|
||||
# )
|
||||
#
|
||||
# return x_mapped.view_as(input)
|
||||
#
|
||||
# def th_iterproduct(*args, dtype=None):
|
||||
# return torch.from_numpy(np.indices(args).reshape((len(args),-1)).T)
|
||||
#
|
||||
# def th_flatten(x):
|
||||
# """Flatten tensor"""
|
||||
# return x.contiguous().view(x[0], x[1], -1)
|
||||
@ -0,0 +1,136 @@
|
||||
import gzip
|
||||
|
||||
from diskcache import FanoutCache, Disk
|
||||
from diskcache.core import BytesType, MODE_BINARY, BytesIO
|
||||
|
||||
from util.logconf import logging
|
||||
log = logging.getLogger(__name__)
|
||||
# log.setLevel(logging.WARN)
|
||||
log.setLevel(logging.INFO)
|
||||
# log.setLevel(logging.DEBUG)
|
||||
|
||||
|
||||
class GzipDisk(Disk):
|
||||
def store(self, value, read, key=None):
|
||||
"""
|
||||
Override from base class diskcache.Disk.
|
||||
|
||||
Chunking is due to needing to work on pythons < 2.7.13:
|
||||
- Issue #27130: In the "zlib" module, fix handling of large buffers
|
||||
(typically 2 or 4 GiB). Previously, inputs were limited to 2 GiB, and
|
||||
compression and decompression operations did not properly handle results of
|
||||
2 or 4 GiB.
|
||||
|
||||
:param value: value to convert
|
||||
:param bool read: True when value is file-like object
|
||||
:return: (size, mode, filename, value) tuple for Cache table
|
||||
"""
|
||||
# pylint: disable=unidiomatic-typecheck
|
||||
if type(value) is BytesType:
|
||||
if read:
|
||||
value = value.read()
|
||||
read = False
|
||||
|
||||
str_io = BytesIO()
|
||||
gz_file = gzip.GzipFile(mode='wb', compresslevel=1, fileobj=str_io)
|
||||
|
||||
for offset in range(0, len(value), 2**30):
|
||||
gz_file.write(value[offset:offset+2**30])
|
||||
gz_file.close()
|
||||
|
||||
value = str_io.getvalue()
|
||||
|
||||
return super(GzipDisk, self).store(value, read)
|
||||
|
||||
|
||||
def fetch(self, mode, filename, value, read):
|
||||
"""
|
||||
Override from base class diskcache.Disk.
|
||||
|
||||
Chunking is due to needing to work on pythons < 2.7.13:
|
||||
- Issue #27130: In the "zlib" module, fix handling of large buffers
|
||||
(typically 2 or 4 GiB). Previously, inputs were limited to 2 GiB, and
|
||||
compression and decompression operations did not properly handle results of
|
||||
2 or 4 GiB.
|
||||
|
||||
:param int mode: value mode raw, binary, text, or pickle
|
||||
:param str filename: filename of corresponding value
|
||||
:param value: database value
|
||||
:param bool read: when True, return an open file handle
|
||||
:return: corresponding Python value
|
||||
"""
|
||||
value = super(GzipDisk, self).fetch(mode, filename, value, read)
|
||||
|
||||
if mode == MODE_BINARY:
|
||||
str_io = BytesIO(value)
|
||||
gz_file = gzip.GzipFile(mode='rb', fileobj=str_io)
|
||||
read_csio = BytesIO()
|
||||
|
||||
while True:
|
||||
uncompressed_data = gz_file.read(2**30)
|
||||
if uncompressed_data:
|
||||
read_csio.write(uncompressed_data)
|
||||
else:
|
||||
break
|
||||
|
||||
value = read_csio.getvalue()
|
||||
|
||||
return value
|
||||
|
||||
def getCache(scope_str):
|
||||
return FanoutCache('data-unversioned/cache/' + scope_str,
|
||||
disk=GzipDisk,
|
||||
shards=64,
|
||||
timeout=1,
|
||||
size_limit=3e11,
|
||||
# disk_min_file_size=2**20,
|
||||
)
|
||||
|
||||
# def disk_cache(base_path, memsize=2):
|
||||
# def disk_cache_decorator(f):
|
||||
# @functools.wraps(f)
|
||||
# def wrapper(*args, **kwargs):
|
||||
# args_str = repr(args) + repr(sorted(kwargs.items()))
|
||||
# file_str = hashlib.md5(args_str.encode('utf8')).hexdigest()
|
||||
#
|
||||
# cache_path = os.path.join(base_path, f.__name__, file_str + '.pkl.gz')
|
||||
#
|
||||
# if not os.path.exists(os.path.dirname(cache_path)):
|
||||
# os.makedirs(os.path.dirname(cache_path), exist_ok=True)
|
||||
#
|
||||
# if os.path.exists(cache_path):
|
||||
# return pickle_loadgz(cache_path)
|
||||
# else:
|
||||
# ret = f(*args, **kwargs)
|
||||
# pickle_dumpgz(cache_path, ret)
|
||||
# return ret
|
||||
#
|
||||
# return wrapper
|
||||
#
|
||||
# return disk_cache_decorator
|
||||
#
|
||||
#
|
||||
# def pickle_dumpgz(file_path, obj):
|
||||
# log.debug("Writing {}".format(file_path))
|
||||
# with open(file_path, 'wb') as file_obj:
|
||||
# with gzip.GzipFile(mode='wb', compresslevel=1, fileobj=file_obj) as gz_file:
|
||||
# pickle.dump(obj, gz_file, pickle.HIGHEST_PROTOCOL)
|
||||
#
|
||||
#
|
||||
# def pickle_loadgz(file_path):
|
||||
# log.debug("Reading {}".format(file_path))
|
||||
# with open(file_path, 'rb') as file_obj:
|
||||
# with gzip.GzipFile(mode='rb', fileobj=file_obj) as gz_file:
|
||||
# return pickle.load(gz_file)
|
||||
#
|
||||
#
|
||||
# def dtpath(dt=None):
|
||||
# if dt is None:
|
||||
# dt = datetime.datetime.now()
|
||||
#
|
||||
# return str(dt).rsplit('.', 1)[0].replace(' ', '--').replace(':', '.')
|
||||
#
|
||||
#
|
||||
# def safepath(s):
|
||||
# s = s.replace(' ', '_')
|
||||
# return re.sub('[^A-Za-z0-9_.-]', '', s)
|
||||
@ -0,0 +1,19 @@
|
||||
import logging
|
||||
import logging.handlers
|
||||
|
||||
root_logger = logging.getLogger()
|
||||
root_logger.setLevel(logging.INFO)
|
||||
|
||||
# Some libraries attempt to add their own root logger handlers. This is
|
||||
# annoying and so we get rid of them.
|
||||
for handler in list(root_logger.handlers):
|
||||
root_logger.removeHandler(handler)
|
||||
|
||||
logfmt_str = "%(asctime)s %(levelname)-8s pid:%(process)d %(name)s:%(lineno)03d:%(funcName)s %(message)s"
|
||||
formatter = logging.Formatter(logfmt_str)
|
||||
|
||||
streamHandler = logging.StreamHandler()
|
||||
streamHandler.setFormatter(formatter)
|
||||
streamHandler.setLevel(logging.DEBUG)
|
||||
|
||||
root_logger.addHandler(streamHandler)
|
||||
@ -0,0 +1,143 @@
|
||||
# From https://github.com/jvanvugt/pytorch-unet
|
||||
# https://raw.githubusercontent.com/jvanvugt/pytorch-unet/master/unet.py
|
||||
|
||||
# MIT License
|
||||
#
|
||||
# Copyright (c) 2018 Joris
|
||||
#
|
||||
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
||||
# of this software and associated documentation files (the "Software"), to deal
|
||||
# in the Software without restriction, including without limitation the rights
|
||||
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
||||
# copies of the Software, and to permit persons to whom the Software is
|
||||
# furnished to do so, subject to the following conditions:
|
||||
#
|
||||
# The above copyright notice and this permission notice shall be included in all
|
||||
# copies or substantial portions of the Software.
|
||||
#
|
||||
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
||||
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
||||
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
||||
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
||||
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
||||
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
||||
# SOFTWARE.
|
||||
|
||||
# Adapted from https://discuss.pytorch.org/t/unet-implementation/426
|
||||
|
||||
import torch
|
||||
from torch import nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
|
||||
class UNet(nn.Module):
|
||||
def __init__(self, in_channels=1, n_classes=2, depth=5, wf=6, padding=False,
|
||||
batch_norm=False, up_mode='upconv'):
|
||||
"""
|
||||
Implementation of
|
||||
U-Net: Convolutional Networks for Biomedical Image Segmentation
|
||||
(Ronneberger et al., 2015)
|
||||
https://arxiv.org/abs/1505.04597
|
||||
|
||||
Using the default arguments will yield the exact version used
|
||||
in the original paper
|
||||
|
||||
Args:
|
||||
in_channels (int): number of input channels
|
||||
n_classes (int): number of output channels
|
||||
depth (int): depth of the network
|
||||
wf (int): number of filters in the first layer is 2**wf
|
||||
padding (bool): if True, apply padding such that the input shape
|
||||
is the same as the output.
|
||||
This may introduce artifacts
|
||||
batch_norm (bool): Use BatchNorm after layers with an
|
||||
activation function
|
||||
up_mode (str): one of 'upconv' or 'upsample'.
|
||||
'upconv' will use transposed convolutions for
|
||||
learned upsampling.
|
||||
'upsample' will use bilinear upsampling.
|
||||
"""
|
||||
super(UNet, self).__init__()
|
||||
assert up_mode in ('upconv', 'upsample')
|
||||
self.padding = padding
|
||||
self.depth = depth
|
||||
prev_channels = in_channels
|
||||
self.down_path = nn.ModuleList()
|
||||
for i in range(depth):
|
||||
self.down_path.append(UNetConvBlock(prev_channels, 2**(wf+i),
|
||||
padding, batch_norm))
|
||||
prev_channels = 2**(wf+i)
|
||||
|
||||
self.up_path = nn.ModuleList()
|
||||
for i in reversed(range(depth - 1)):
|
||||
self.up_path.append(UNetUpBlock(prev_channels, 2**(wf+i), up_mode,
|
||||
padding, batch_norm))
|
||||
prev_channels = 2**(wf+i)
|
||||
|
||||
self.last = nn.Conv2d(prev_channels, n_classes, kernel_size=1)
|
||||
|
||||
def forward(self, x):
|
||||
blocks = []
|
||||
for i, down in enumerate(self.down_path):
|
||||
x = down(x)
|
||||
if i != len(self.down_path)-1:
|
||||
blocks.append(x)
|
||||
x = F.avg_pool2d(x, 2)
|
||||
|
||||
for i, up in enumerate(self.up_path):
|
||||
x = up(x, blocks[-i-1])
|
||||
|
||||
return self.last(x)
|
||||
|
||||
|
||||
class UNetConvBlock(nn.Module):
|
||||
def __init__(self, in_size, out_size, padding, batch_norm):
|
||||
super(UNetConvBlock, self).__init__()
|
||||
block = []
|
||||
|
||||
block.append(nn.Conv2d(in_size, out_size, kernel_size=3,
|
||||
padding=int(padding)))
|
||||
block.append(nn.ReLU())
|
||||
# block.append(nn.LeakyReLU())
|
||||
if batch_norm:
|
||||
block.append(nn.BatchNorm2d(out_size))
|
||||
|
||||
block.append(nn.Conv2d(out_size, out_size, kernel_size=3,
|
||||
padding=int(padding)))
|
||||
block.append(nn.ReLU())
|
||||
# block.append(nn.LeakyReLU())
|
||||
if batch_norm:
|
||||
block.append(nn.BatchNorm2d(out_size))
|
||||
|
||||
self.block = nn.Sequential(*block)
|
||||
|
||||
def forward(self, x):
|
||||
out = self.block(x)
|
||||
return out
|
||||
|
||||
|
||||
class UNetUpBlock(nn.Module):
|
||||
def __init__(self, in_size, out_size, up_mode, padding, batch_norm):
|
||||
super(UNetUpBlock, self).__init__()
|
||||
if up_mode == 'upconv':
|
||||
self.up = nn.ConvTranspose2d(in_size, out_size, kernel_size=2,
|
||||
stride=2)
|
||||
elif up_mode == 'upsample':
|
||||
self.up = nn.Sequential(nn.Upsample(mode='bilinear', scale_factor=2),
|
||||
nn.Conv2d(in_size, out_size, kernel_size=1))
|
||||
|
||||
self.conv_block = UNetConvBlock(in_size, out_size, padding, batch_norm)
|
||||
|
||||
def center_crop(self, layer, target_size):
|
||||
_, _, layer_height, layer_width = layer.size()
|
||||
diff_y = (layer_height - target_size[0]) // 2
|
||||
diff_x = (layer_width - target_size[1]) // 2
|
||||
return layer[:, :, diff_y:(diff_y + target_size[0]), diff_x:(diff_x + target_size[1])]
|
||||
|
||||
def forward(self, x, bridge):
|
||||
up = self.up(x)
|
||||
crop1 = self.center_crop(bridge, up.shape[2:])
|
||||
out = torch.cat([up, crop1], 1)
|
||||
out = self.conv_block(out)
|
||||
|
||||
return out
|
||||
Loading…
Reference in New Issue
Block a user