Files
pyTrainSeg/filter_functions.py
T

464 lines
18 KiB
Python

# -*- coding: utf-8 -*-
"""
Created on Mon Aug 15 16:35:22 2022
@author: fische_r
class to create feature stack on 4D tomo data
transient dimension, e.g. time, should be 4th dimension of input data
TODO: add GPU support (CUDA, cupy, cucim)
TODO: store git commit sha
"""
# necessary packages for feature stack
import numpy as np
from scipy import ndimage
from skimage import filters
from skimage.morphology import ball
import dask
import dask.array
# from dask.distributed import Client, LocalCluster
from itertools import combinations_with_replacement, combinations
import xarray as xr
# functions take chunked dask-array as input
# start-up cluster, TODO: option to connect to exisitng cluster
# TODO: class/function to boot up cluster with custom options, e.g. workers/threads
# esp. use SSD for memory spilling
# cluster = LocalCluster()
# client = Client(cluster)
# print('Dashboard at '+cluster.dashboard_link)
default_feature_dict = {'Gaussian': True,
# 'Sobel': True,
'Hessian': True,
'Diff of Gaussians': True,
'maximum': True,
'minimum': True,
'median': True,
'extra_time_ranks': True,
}
def ball_4d(sig):
bnd = np.zeros((sig*2+1,sig*2+1,sig*2+1,sig*2+1), dtype = bool)
bnd[sig,sig,sig,sig] = True
ecd = ndimage.distance_transform_edt(~bnd)
bnd = (ecd<sig+0.01).astype(int)
return bnd
class image_filter:
def __init__(self,
data_path = None,
outpath = None,
sigmas = [0,2, 4],
feature_dict = default_feature_dict,
mod_feat_dict = None,
chunksize = '20 MiB', #try to align chunks to extend far in time --> should be useful for most filters, esp. the dynamic rank filters
outchunks = '300 MiB',
ranks = ['maximum', 'minimum', 'median'], #, 'mean'
sigma_t = 40
):
if mod_feat_dict is not None:
for key in mod_feat_dict:
feature_dict[key] = mod_feat_dict[key]
self.data_path = data_path
self.outpath = outpath
self.sigmas = sigmas
self.feature_dict = feature_dict
# TODO: allow option of custom shaped chunks
self.chunks = chunksize
self.outchunks = outchunks
#wheter considering means for first and last time step
self.take_means = True
self.num_means = 7
# not sure if this is clever, does dask understand that this data is reused?
self.Gaussian_4D_dict = {}
self.Gaussian_space_dict = {}
self.Gaussian_time_dict = {}
self.Gradient_dict = {}
self.calculated_features = []
self.feature_names = []
self.considered_ranks = ranks
self.sigma_t = sigma_t
self.prepared = False
self.computed = False
# TODO: currently loads full dataset into memory, consider aligning desired chunk already for original dataset to avoid rechunking
# .rechunk() causes problems downstream: "Assertion error" , WTF?!
# if original data soe not fit in RAM, rechunk, store to disk and load again?
def open_raw_data(self):
data = xr.open_dataset(self.data_path)
da = dask.array.from_array(data.tomo.data, chunks = self.chunks)
self.original_dataset = data
self.data = da
def open_lazy_data(self, chunks=None):
if chunks is None:
chunks = self.chunks
data = xr.open_dataset(self.data_path, chunks = chunks)
da = dask.array.from_array(data.tomo)
# print('maybe re-introducing rechunking, but for large datasets auto might be ok')
# print('smaller chunks might be better for slicewise training')
# print('currently provided chunks are ignored')
self.original_dataset = data#.rechunk(self.chunks)
self.data = da
def load_raw_data(self):
data = xr.load_dataset(self.data_path)
# da = dask.array.from_array(data.tomo).rechunk(chunks = self.chunks)
da = dask.array.from_array(data.tomo.data, chunks = self.chunks)
self.original_dataset = data
self.data = da
def Gaussian_Blur_4D(self, sigma):
# TODO: check on boundary mode
deptharray = np.ones(self.data.ndim)+4*sigma
deptharray = tuple(np.min([deptharray, self.data.shape], axis=0))
G = self.data.map_overlap(filters.gaussian, depth=deptharray, boundary='nearest', sigma = sigma)
self.feature_names.append('Gaussian_4D_Blur_'+f'{sigma:.1f}')
self.calculated_features.append(G)
self.Gaussian_4D_dict[f'{sigma:.1f}'] = G
def Gaussian_Blur_space(self, sigma):
deptharray = np.ones(self.data.ndim)+4*sigma
deptharray[-1] = 0
sigmas = np.ones(deptharray.shape)*sigma
deptharray = tuple(np.min([deptharray, self.data.shape], axis=0))
sigmas[-1] = 0
G = self.data.map_overlap(filters.gaussian, depth=deptharray, boundary='nearest', sigma = sigmas)
self.feature_names.append('Gaussian_space_'+f'{sigma:.1f}')
self.calculated_features.append(G)
self.Gaussian_space_dict[f'{sigma:.1f}'] = G
def Gaussian_Blur_time(self, sigma):
deptharray = np.ones(self.data.ndim)+4*sigma
deptharray[:-1] = 0
sigmas = np.ones(deptharray.shape)*sigma
deptharray = tuple(np.min([deptharray, self.data.shape], axis=0))
sigmas[:-1] = 0
G = self.data.map_overlap(filters.gaussian, depth=deptharray, boundary='nearest', sigma = sigmas)
self.feature_names.append('Gaussian_time_'+f'{sigma:.1f}')
self.calculated_features.append(G)
self.Gaussian_time_dict[f'{sigma:.1f}'] = G
def Gaussian_4D_stack(self):
flag = True
for sigma in self.sigmas:
if np.abs(sigma-0)<0.1:
if flag:
flag = False
# self.Gaussian_4D_dict['original'] = self.data
# self.calculated_features.append(self.data)
# self.feature_names.append('original')
sig = 0
self.Gaussian_Blur_4D(sig)
else:
self.Gaussian_Blur_4D(sigma)
def Gaussian_space_stack(self):
flag = True
for sigma in self.sigmas:
if np.abs(sigma-0)<0.1:
if flag:
flag = False
# self.Gaussian_space_dict['original'] = self.data
sig = 0
self.Gaussian_Blur_space(sig)
else:
self.Gaussian_Blur_space(sigma)
def Gaussian_time_stack(self):
flag = True
for sigma in self.sigmas:
if np.abs(sigma-0)<0.1:
if flag:
flag = False
# self.Gaussian_time_dict['original'] = self.data
sig = 0
self.Gaussian_Blur_time(sig)
else:
self.Gaussian_Blur_time(sigma)
def diff_Gaussian(self, mode):
if mode == '4D':
lookup_dict = self.Gaussian_4D_dict
elif mode == 'space':
lookup_dict = self.Gaussian_space_dict
elif mode == 'time':
lookup_dict = self.Gaussian_time_dict
for comb in combinations(lookup_dict.keys(),2):
G1 = lookup_dict[comb[1]]
G0 = lookup_dict[comb[0]]
# DG = lookup_dict[comb[1]] - lookup_dict[comb[0]]
# DG = G1-G0
DG = dask.array.subtract(G1,G0)
name = ''.join(['diff_of_gauss_',mode,'_',comb[1],'_',comb[0]])
self.calculated_features.append(DG)
self.feature_names.append(name)
def diff_to_first_and_last(self, take_mean, means):
# TODO: take temporal mean/median for first and last
DA = self.data
if take_mean:
first = DA[...,:means].mean(axis=-1)
last = DA[...,-means:].mean(axis=-1)
else:
first = DA[...,0]
last = DA[...,-1]
if type(first) is not np.ndarray:
first = first.compute()
last = last.compute()
# ones = dask.array.ones(DA.shape, chunks=self.chunks)
if type(first) is np.ndarray:
firsts = dask.array.stack([first]*DA.shape[-1], axis=-1)
lasts = dask.array.stack([last]*DA.shape[-1], axis=-1)
firsts = firsts.rechunk(DA.chunksize)
lasts = lasts.rechunk(DA.chunksize)
DF = DA - firsts
DL = DA - lasts
self.calculated_features.append(DF)
self.feature_names.append('diff_to_first_')
self.calculated_features.append(DL)
self.feature_names.append('diff_to_last_')
self.feature_names.append('first_')
self.calculated_features.append(firsts)
self.feature_names.append('last_')
self.calculated_features.append(lasts)
else:
print('Diff first and last is an unexplainable pain in the ass, solve this at one point')
def time_mean(self):
DA = self.data
mean = DA.mean(axis=-1)
means = dask.array.stack([mean]*DA.shape[-1], axis=-1)
self.calculated_features.append(means)
self.feature_names.append('full_temporal_mean_')
def Gradients(self):
for key in self.Gaussian_4D_dict:
G = self.Gaussian_4D_dict[key]
gradients = dask.array.gradient(G)
self.Gradient_dict[key] = gradients
def Hessian(self):
# TODO: add max of all dimensions
for key in self.Gradient_dict.keys():
axes = range(self.data.ndim)
gradients = self.Gradient_dict[key]
H_elems = [dask.array.gradient(gradients[ax0], axis=ax1) for ax0, ax1 in combinations_with_replacement(axes, 2)]
gradnames = ['Gradient_sigma_'+key+'_'+str(ax0) for ax0 in axes]
elems = [(ax0,ax1) for ax0, ax1 in combinations_with_replacement(axes, 2)]
hessnames = [''.join(['hessian_sigma_',key,'_',str(elm[0]),str(elm[1])]) for elm in elems ]
self.feature_names = self.feature_names + gradnames + hessnames
self.calculated_features = self.calculated_features+gradients+H_elems
# def rank_filter(self, option, sigma):
# # note: rank filters not yet available in CUCIM and don't work with dask --> figure out why
# da = self.data
# if not np.abs(sigma-0)<1:
# if option == 'minimum':
# fun = filters.rank.minimum
# elif option == 'maximum':
# fun = filters.rank.maximum
# elif option == 'median':
# fun = filters.rank.median
# elif option == 'mean':
# fun = filters.rank.mean
# fp = ball_4d(sigma)
# deptharray = np.ones(da.ndim)+sigma
# deptharray = tuple(np.min([deptharray, da.shape], axis=0))
# R = da.map_overlap(fun, depth=deptharray, footprint=fp)
# name = ''.join([option,'_',f'{sigma:.1f}'])
# self.calculated_features.append(R)
# self.feature_names.append(name)
# def dynamic_rank_filter(self, option, sigma):
# # TODO: add custom dynamic model, eg. sigmoid
# da = self.data
# if option == 'minimum':
# fun = filters.rank.minimum
# elif option == 'maximum':
# fun = filters.rank.maximum
# elif option == 'median':
# fun = filters.rank.median
# elif option == 'mean':
# fun = filters.rank.mean
# fp_3D = ball(sigma)
# fp_4D = np.zeros(list(fp_3D.shape)+[2*self.sigma_t], dtype=int)
# fp_4D[fp_3D>0,:] = 1
# deptharray = np.ones(da.ndim)+sigma
# deptharray[-1] = self.sigma_t
# deptharray = tuple(np.min([deptharray, da.shape], axis=0))
# R = da.map_overlap(fun, depth=deptharray, footprint=fp_4D)
# name = ''.join([option,'_dynamic_',f'{sigma:.1f}'])
# self.calculated_features.append(R)
# self.feature_names.append(name)
def rank_like_filter(self, option, sigma):
# note: rank filters not yet available in CUCIM
da = self.data
if not np.abs(sigma-0)<1:
if option == 'minimum':
fun = ndimage.minimum_filter
elif option == 'maximum':
fun = ndimage.maximum_filter
elif option == 'median':
fun = ndimage.median_filter
# elif option == 'mean':
# fun = filters.rank.mean
fp = ball_4d(sigma)
deptharray = np.ones(da.ndim)+sigma
deptharray = tuple(np.min([deptharray, da.shape], axis=0))
R = da.map_overlap(fun, depth=deptharray, footprint=fp)
name = ''.join([option,'_',f'{sigma:.1f}'])
self.calculated_features.append(R)
self.feature_names.append(name)
def dynamic_rank_like_filter(self, option, sigma):
# TODO: add custom dynamic model, eg. sigmoid
da = self.data
if option == 'minimum':
fun = ndimage.minimum_filter
elif option == 'maximum':
fun = ndimage.maximum_filter
elif option == 'median':
fun = ndimage.median_filter
# elif option == 'mean':
# fun = filters.rank.mean
fp_3D = ball(sigma)
fp_4D = np.zeros(list(fp_3D.shape)+[2*self.sigma_t], dtype=int)
fp_4D[fp_3D>0,:] = 1
deptharray = np.ones(da.ndim)+sigma
deptharray[-1] = self.sigma_t
deptharray = tuple(np.min([deptharray, da.shape], axis=0))
R = da.map_overlap(fun, depth=deptharray, footprint=fp_4D)
name = ''.join([option,'_dynamic_',f'{sigma:.1f}'])
self.calculated_features.append(R)
self.feature_names.append(name)
def rank_filter_stack(self):
for option in self.considered_ranks:
for sigma in self.sigmas:
self.rank_like_filter(option, sigma)
self.dynamic_rank_like_filter(option, sigma)
def pixel_coordinates(self):
#create 3 arrays with the pixel coordinates
da = self.data
# coords = dask.array.where(da)
# for i in range(3):
# loc = coords[i].compute().reshape(da.shape) #compute() to know shape of coords, maybe find another way
# self.calculated_features.append(loc)
# self.feature_names.append('loc_'+'xyz'[i])
# the following looks less elegant, but seems more compatible with dask
# TODO: check performance
loc_x = dask.array.ones(da.shape)*dask.array.arange(da.shape[0])[:,None, None, None]
self.calculated_features.append(loc_x)
self.feature_names.append('loc_'+'x')
loc_y = dask.array.ones(da.shape)*dask.array.arange(da.shape[1])[None,:, None, None]
self.calculated_features.append(loc_y)
self.feature_names.append('loc_'+'y')
loc_z = dask.array.ones(da.shape)*dask.array.arange(da.shape[2])[None, None,:, None]
self.calculated_features.append(loc_z)
self.feature_names.append('loc_'+'z')
# TODO: include feature selection either in compute (better) or save
# TODO: maybe add purge function
# TODO: maybe add iterative segmentation results, i.e. median filter of segmentation
def prepare(self):
self.Gaussian_4D_dict = {}
self.Gaussian_space_dict = {}
self.Gaussian_time_dict = {}
self.Gradient_dict = {}
self.calculated_features = []
self.feature_names = []
self.diff_to_first_and_last(self.take_means, self.num_means)
self.Gaussian_4D_stack()
self.diff_Gaussian('4D')
self.Gradients()
self.Hessian()
self.Gaussian_time_stack()
self.diff_Gaussian('time')
self.Gaussian_space_stack()
self.diff_Gaussian('space')
self.pixel_coordinates()
# self.time_mean()
# self.rank_filter_stack() #you have to load the entire raw data set for this filter --> not so good for many time steps
self.prepared = True
def stack_features(self):
if not self.prepared:
print('prepare first')
else:
self.feature_stack = dask.array.stack(self.calculated_features, axis = 4)
def compute(self):
# self.feature_stack = self.feature_stack.compute()
self.feature_stack = self.feature_stack.persist() #not sure, but persist should be preferred
self.computed = True
def make_xarray_nc(self, outpath = None, store=False):
if outpath is None:
outpath = self.outpath
shp = self.feature_stack.shape
coords = {'x': np.arange(shp[0]), 'y': np.arange(shp[1]), 'z': np.arange(shp[2]), 'time': np.arange(shp[3]), 'feature': self.feature_names}
if store:
if self.computed:
if not type(self.feature_stack) is np.ndarray:
self.feature_stack.rechunk(self.outchunks)
#TODO avoid this explcit conversion. however seems necessary ?...
# if type(self.feature_stack) is not np.ndarray:
# self.feature_stack = self.feature_stack.compute()
self.result = xr.Dataset({'feature_stack': (['x','y','z','time', 'feature'], self.feature_stack)},
coords = coords
)
self.result.to_netcdf(outpath)
else:
print('maybe you have to compute the stack first ... ?!')
else:
self.result = xr.Dataset({'feature_stack': (['x','y','z','time', 'feature'], self.feature_stack)},
coords = coords
)