import os, sys, time, shutil, tempfile, datetime, pathlib, subprocess
from pathlib import Path
import numpy as np
# from tqdm import trange, tqdm
from tqdm.auto import tqdm, trange
from urllib.parse import urlparse
import torch
from torch import nn, distributed, multiprocessing, optim
# from torch.nn.parallel import DistributedDataParallel as DDP
from scipy.ndimage import gaussian_filter, zoom
import logging
models_logger = logging.getLogger(__name__)
from . import transforms, dynamics, utils, plot
from .core import UnetModel, assign_device, check_mkl, MXNET_ENABLED, parse_model_string
from .io import OMNI_INSTALLED
from omnipose.gpu import empty_cache, ARM #, custom_nonzero_cuda
from omnipose.utils import hysteresis_threshold
from torchvf.numerics import interp_vf, ivp_solver
# from torchvf.utils import cluster
# _MODEL_URL = 'https://www.cellpose.org/models'
_MODEL_URL = 'https://raw.githubusercontent.com/kevinjohncutler/omnipose-models/main'
_MODEL_DIR_ENV = os.environ.get("CELLPOSE_LOCAL_MODELS_PATH")
_MODEL_DIR_DEFAULT = pathlib.Path.home().joinpath('.cellpose', 'models')
MODEL_DIR = pathlib.Path(_MODEL_DIR_ENV) if _MODEL_DIR_ENV else _MODEL_DIR_DEFAULT
if OMNI_INSTALLED:
import omnipose
from omnipose.core import C2_BD_MODELS, C1_BD_MODELS, C2_MODELS, C1_MODELS
else:
C2_BD_MODELS, C1_BD_MODELS, C2_MODELS, C1_MODELS = [],[],[],[]
print('omnipose not installed, omni models will not be available')
CP_MODELS = ['cyto','nuclei','cyto2']
C2_MODEL_NAMES = C2_BD_MODELS + C2_MODELS + CP_MODELS
BD_MODEL_NAMES = C2_BD_MODELS + C1_BD_MODELS
MODEL_NAMES = C1_MODELS + C2_BD_MODELS + C1_BD_MODELS + C2_MODELS + CP_MODELS
[docs]def model_path(model_type, model_index, use_torch):
torch_str = 'torch' if use_torch else ''
basename = '%s%s_%d' % (model_type, torch_str, model_index)
return cache_model_path(basename)
[docs]def size_model_path(model_type, use_torch):
torch_str = 'torch' if use_torch else ''
basename = 'size_%s%s_0.npy' % (model_type, torch_str)
return cache_model_path(basename)
[docs]def cache_model_path(basename):
MODEL_DIR.mkdir(parents=True, exist_ok=True)
url = f'{_MODEL_URL}/{basename}'
cached_file = os.fspath(MODEL_DIR.joinpath(basename))
if not os.path.exists(cached_file):
models_logger.info('Downloading: "{}" to {}\n'.format(url, cached_file))
print(url,cached_file)
utils.download_url_to_file(url, cached_file, progress=True)
return cached_file
[docs]def deprecation_warning_cellprob_dist_threshold(cellprob_threshold, dist_threshold):
models_logger.warning('cellprob_threshold and dist_threshold are being deprecated in a future release, use mask_threshold instead')
return cellprob_threshold if cellprob_threshold is not None else dist_threshold
[docs]class Cellpose():
""" main model which combines SizeModel and CellposeModel
Parameters
----------
gpu: bool (optional, default False)
whether or not to use GPU, will check if GPU available
model_type: str (optional, default 'cyto')
'cyto'=cytoplasm model; 'nuclei'=nucleus model
net_avg: bool (optional, default True)
loads the 4 built-in networks and averages them if True, loads one network if False
device: gpu device (optional, default None)
where model is saved (e.g. mx.gpu() or mx.cpu()), overrides gpu input,
recommended if you want to use a specific GPU (e.g. mx.gpu(4) or torch.cuda.device(4))
torch: bool (optional, default True)
run model using torch if available
"""
def __init__(self, gpu=False, model_type='cyto', net_avg=True, device=None,
use_torch=True, model_dir=None, dim=2, omni=None):
super(Cellpose, self).__init__()
if not torch:
if not MXNET_ENABLED:
use_torch = True
self.torch = use_torch
# assign device (GPU or CPU)
sdevice, gpu = assign_device(self.torch, gpu)
self.device = device if device is not None else sdevice
self.gpu = gpu
# set defaults and catch if cyto2 is being used without torch
model_type = 'cyto' if model_type is None else model_type
if model_type=='cyto2' and not self.torch:
model_type='cyto'
# default omni on if model specifies this, but also override by user setting
self.omni = ('omni' in model_type) if omni is None else omni
self.dim = dim # 2D vs 3D
# for now, omni models cannot do net_avg
if self.omni:
net_avg = False
model_range = range(4) if net_avg else range(1)
self.pretrained_model = [model_path(model_type, j, torch) for j in model_range]
self.diam_mean = 30. #default for any cyto model
nuclear = 'nuclei' in model_type
bacterial = ('bact' in model_type) or ('worm' in model_type)
plant = 'plant' in model_type
if nuclear:
self.diam_mean = 17.
elif bacterial or plant:
#self.diam_mean = 0.
net_avg = False # No bacterial, plant, or omni models have additional models
if not net_avg:
self.pretrained_model = self.pretrained_model[0]
self.cp = CellposeModel(device=self.device, gpu=self.gpu,
pretrained_model=self.pretrained_model,
diam_mean=self.diam_mean, use_torch=self.torch,
dim=self.dim, omni=self.omni)
self.cp.model_type = model_type
# size model not used for bacterial model
if not bacterial:
self.pretrained_size = size_model_path(model_type, torch)
self.sz = SizeModel(device=self.device, pretrained_size=self.pretrained_size,
cp_model=self.cp)
self.sz.model_type = model_type
else:
self.pretrained_size = None
[docs] def eval(self, x, batch_size=8, channels=None, channel_axis=None, z_axis=None,
invert=False, normalize=True, diameter=30., do_3D=False, anisotropy=None,
net_avg=True, augment=False, tile=True, tile_overlap=0.1, resample=True,
interp=True, cluster=False, boundary_seg=False, affinity_seg=False, despur=True,
flow_threshold=0.4, mask_threshold=0.0,
cellprob_threshold=None, dist_threshold=None, diam_threshold=12., min_size=15, max_size=None,
stitch_threshold=0.0, rescale=None, progress=None, omni=False, verbose=False,
transparency=False, model_loaded=False):
""" run cellpose and get masks
Parameters
----------
x: list or array of images
can be list of 2D/3D images, or array of 2D/3D images, or 4D image array
batch_size: int (optional, default 8)
number of 224x224 patches to run simultaneously on the GPU
(can make smaller or bigger depending on GPU memory usage)
channels: list (optional, default None)
list of channels, either of length 2 or of length number of images by 2.
First element of list is the channel to segment (0=grayscale, 1=red, 2=green, 3=blue).
Second element of list is the optional nuclear channel (0=none, 1=red, 2=green, 3=blue).
For instance, to segment grayscale images, input [0,0]. To segment images with cells
in green and nuclei in blue, input [2,3]. To segment one grayscale image and one
image with cells in green and nuclei in blue, input [[0,0], [2,3]].
channel_axis: int (optional, default None)
if None, channels dimension is attempted to be automatically determined
z_axis: int (optional, default None)
if None, z dimension is attempted to be automatically determined
invert: bool (optional, default False)
invert image pixel intensity before running network (if True, image is also normalized)
normalize: bool (optional, default True)
normalize data so 0.0=1st percentile and 1.0=99th percentile of image intensities in each channel
diameter: float (optional, default 30.)
if set to None, then diameter is automatically estimated if size model is loaded
do_3D: bool (optional, default False)
set to True to run 3D segmentation on 4D image input
anisotropy: float (optional, default None)
for 3D segmentation, optional rescaling factor (e.g. set to 2.0 if Z is sampled half as dense as X or Y)
net_avg: bool (optional, default True)
runs the 4 built-in networks and averages them if True, runs one network if False
augment: bool (optional, default False)
tiles image with overlapping tiles and flips overlapped regions to augment
tile: bool (optional, default True)
tiles image to ensure GPU/CPU memory usage limited (recommended)
tile_overlap: float (optional, default 0.1)
fraction of overlap of tiles when computing flows
resample: bool (optional, default True)
run dynamics at original image size (will be slower but create more accurate boundaries)
interp: bool (optional, default True)
interpolate during 2D dynamics (not available in 3D)
(in previous versions it was False)
flow_threshold: float (optional, default 0.4)
flow error threshold (all cells with errors below threshold are kept) (not used for 3D)
mask_threshold: float (optional, default 0.0)
all pixels with value above threshold kept for masks, decrease to find more and larger masks
dist_threshold: float (optional, default None) DEPRECATED
use mask_threshold instead
cellprob_threshold: float (optional, default None) DEPRECATED
use mask_threshold instead
min_size: int (optional, default 15)
minimum number of pixels per mask, can turn off with -1
stitch_threshold: float (optional, default 0.0)
if stitch_threshold>0.0 and not do_3D and equal image sizes, masks are stitched in 3D to return volume segmentation
rescale: float (optional, default None)
if diameter is set to None, and rescale is not None, then rescale is used instead of diameter for resizing image
progress: pyqt progress bar (optional, default None)
to return progress bar status to GUI
omni: bool (optional, default False)
use omnipose mask recontruction features
calc_trace: bool (optional, default False)
calculate pixel traces and return as part of the flow
verbose: bool (optional, default False)
turn on additional output to logs for debugging
verbose: bool (optional, default False)
turn on additional output to logs for debugging
transparency: bool (optional, default False)
modulate flow opacity by magnitude instead of brightness (can use flows on any color background)
model_loaded: bool (optional, default False)
internal variable for determining if model has been loaded, used in __main__.py
Returns
-------
masks: list of 2D arrays, or single 3D array (if do_3D=True)
labelled image, where 0=no masks; 1,2,...=mask labels
flows: list of lists 2D arrays, or list of 3D arrays (if do_3D=True)
flows[k][0] = XY flow in HSV 0-255
flows[k][1] = flows at each pixel
flows[k][2] = scalar cell probability (Cellpose) or distance transform (Omnipose)
flows[k][3] = final pixel locations after Euler integration
flows[k][4] = boundary output (nonempty for Omnipose)
flows[k][5] = pixel traces (nonempty for calc_trace=True)
styles: list of 1D arrays of length 256, or single 1D array (if do_3D=True)
style vector summarizing each image, also used to estimate size of objects in image
diams: list of diameters, or float (if do_3D=True)
"""
if cellprob_threshold is not None or dist_threshold is not None:
mask_threshold = deprecation_warning_cellprob_dist_threshold(cellprob_threshold, dist_threshold)
tic0 = time.time()
channels = [0,0] if channels is None else channels # why not just make this a default in the function header?
estimate_size = True if (diameter is None or diameter==0) else False
if estimate_size and self.pretrained_size is not None and not do_3D and x[0].ndim < 4:
tic = time.time()
models_logger.info('~~~ ESTIMATING CELL DIAMETER(S) ~~~')
diams, _ = self.sz.eval(x, channels=channels, channel_axis=channel_axis,
invert=invert, batch_size=batch_size,
augment=augment, tile=tile, normalize=normalize)
rescale = self.diam_mean / np.array(diams)
diameter = None
models_logger.info('estimated cell diameter(s) in %0.2f sec'%(time.time()-tic))
models_logger.info('>>> diameter(s) = ')
if isinstance(diams, list) or isinstance(diams, np.ndarray):
diam_string = '[' + ''.join(['%0.2f, '%d for d in diams]) + ']'
else:
diam_string = '[ %0.2f ]'%diams
models_logger.info(diam_string)
elif estimate_size:
if self.pretrained_size is None:
reason = 'no pretrained size model specified in model Cellpose'
else:
reason = 'does not work on non-2D images'
models_logger.warning(f'could not estimate diameter, {reason}')
diams = self.diam_mean
else:
diams = diameter
tic = time.time()
models_logger.info('~~~ FINDING MASKS ~~~')
masks, flows, styles = self.cp.eval(x,
batch_size=batch_size,
invert=invert,
normalize=normalize,
diameter=diameter,
rescale=rescale,
anisotropy=anisotropy,
channels=channels,
channel_axis=channel_axis,
z_axis=z_axis,
augment=augment,
tile=tile,
do_3D=do_3D,
net_avg=net_avg,
progress=progress,
tile_overlap=tile_overlap,
resample=resample,
interp=interp,
cluster=cluster,
flow_threshold=flow_threshold,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
boundary_seg=boundary_seg,
affinity_seg=affinity_seg,
despur=despur,
min_size=min_size,
max_size=max_size,
stitch_threshold=stitch_threshold,
omni=omni,
verbose=verbose,
transparency=transparency,
model_loaded=model_loaded)
models_logger.info('>>>> TOTAL TIME %0.2f sec'%(time.time()-tic0))
return masks, flows, styles, diams
# there is a bunch of repetiton in cellpose(), cellposemodel(), __main__ with nuclear, bacterial checks
# I need to figure out a way to facotr all that out, probably by making a function in models and calling it
# in all three contexts. Also I should just check for model existence for the 4-model averaging instead of
# requiring it for models based on name.
# I want to make a 1-class model (cellprob only). nclasses=1 is already an option, but much of the code requires 3+ classes.
[docs]class CellposeModel(UnetModel):
"""
Parameters
-------------------
gpu: bool (optional, default False)
whether or not to save model to GPU, will check if GPU available
pretrained_model: str or list of strings (optional, default False)
path to pretrained cellpose model(s), if None or False, no model loaded
model_type: str (optional, default None)
'cyto'=cytoplasm model; 'nuclei'=nucleus model; if None, pretrained_model used
net_avg: bool (optional, default True)
loads the 4 built-in networks and averages them if True, loads one network if False
torch: bool (optional, default True)
use torch nn rather than mxnet
diam_mean: float (optional, default 27.)
mean 'diameter', 27. is built in value for 'cyto' model
device: mxnet device (optional, default None)
where model is saved (mx.gpu() or mx.cpu()), overrides gpu input,
recommended if you want to use a specific GPU (e.g. mx.gpu(4))
model_dir: str (optional, default None)
overwrite the built in model directory where cellpose looks for models
omni: use omnipose model (optional, default False)
"""
def __init__(self, gpu=False, pretrained_model=False,
model_type=None, net_avg=True, use_torch=True,
diam_mean=30., device=None,
residual_on=True, style_on=True, concatenation=False,
nchan=1, nclasses=2, dim=2, omni=True,
checkpoint=False, dropout=False, kernel_size=2):
if not torch:
if not MXNET_ENABLED:
use_torch = True
self.torch = use_torch
# print('torch is', torch) # duplicated in unetmodel claass
if isinstance(pretrained_model, np.ndarray):
pretrained_model = list(pretrained_model)
elif isinstance(pretrained_model, str):
pretrained_model = [pretrained_model]
# initialize according to arguments
# these are overwritten if a model requires it (bact_omni the most restrictive)
self.omni = omni
self.nclasses = nclasses
self.diam_mean = diam_mean
self.dim = dim # 2D vs 3D
self.nchan = nchan
self.checkpoint = checkpoint
self.dropout = dropout
self.kernel_size = kernel_size
# channel axis might be useful here
pretrained_model_string = None
if model_type is not None or (pretrained_model and not os.path.exists(pretrained_model[0])):
pretrained_model_string = model_type
if not np.any([pretrained_model_string == s for s in MODEL_NAMES]): #also covers None case
pretrained_model_string = 'cyto'
if (pretrained_model and not os.path.exists(pretrained_model[0])):
models_logger.warning('pretrained model has incorrect path')
models_logger.info(f'>>{pretrained_model_string}<< model set to be used')
nuclear = 'nuclei' in model_type
bacterial = ('bact' in model_type) or ('worm' in model_type)
plant = 'plant' in model_type
if nuclear:
self.diam_mean = 17.
elif bacterial or plant:
#self.diam_mean = 0.
net_avg = False # No bacterial, plant, or omni models have additional models
# original omni models had the boundary field
if model_type in BD_MODEL_NAMES:
self.nclasses = 3
else:
self.nclasses = 2
# most original cellpose/omnipose models also were trained with 2 channels
# (even though most or all images were single-channel)
if model_type in C2_MODEL_NAMES:
self.nchan = 2
# for now, omni models cannot do net_avg
if self.omni:
net_avg = False
#changed to only look for multiple files if net_avg is selected
model_range = range(4) if net_avg else range(1)
pretrained_model = [model_path(pretrained_model_string, j, torch) for j in model_range]
residual_on, style_on, concatenation = True, True, False
else:
if pretrained_model:
pretrained_model_string = pretrained_model[0]
params = parse_model_string(pretrained_model_string)
if params is not None:
residual_on, style_on, concatenation = params
# set omni flag to true if the name contains it
if pretrained_model_string is not None:
self.omni = 'omni' in os.path.splitext(Path(pretrained_model_string).name)[0] if self.omni is None else self.omni
# convert abstract prediction classes number to actual count
# flow field components increase this by dim-1
self.nclasses = self.nclasses + (self.dim-1)
# initialize network
super().__init__(gpu=gpu, pretrained_model=False,
diam_mean=self.diam_mean, net_avg=net_avg, device=device,
residual_on=residual_on, style_on=style_on, concatenation=concatenation,
nclasses=self.nclasses, use_torch=self.torch, nchan=self.nchan,
dim=self.dim, checkpoint=self.checkpoint, dropout=self.dropout,
kernel_size=self.kernel_size)
self.unet = False
self.pretrained_model = pretrained_model
if self.pretrained_model and len(self.pretrained_model)==1:
# dataparallel A1
# if self.torch and self.gpu:
# net = self.net.module
# else:
# net = self.net
# if self.torch and gpu:
# self.net = nn.DataParallel(self.net)
self.net.load_model(self.pretrained_model[0], cpu=(not self.gpu))
if not self.torch:
self.net.collect_params().grad_req = 'null'
ostr = ['off', 'on']
omnistr = ['','_omni'] #toggle by containing omni phrase
self.net_type = 'cellpose_residual_{}_style_{}_concatenation_{}{}_abstract_nclasses_{}_nchan_{}_dim_{}'.format(ostr[residual_on],
ostr[style_on],
ostr[concatenation],
omnistr[omni],
self.nclasses-(self.dim-1), # "abstract"
self.nchan,
self.dim)
if self.torch and gpu:
self.net = nn.DataParallel(self.net)
#A1
# distributed.init_process_group()
# distributed.launch
# distributed.init_process_group(backend='nccl',rank=0,world_size=2)
# self.net = nn.parallel.DistributedDataParallel(self.net)
# rank = 0 # one computer
# world_size = 1
# setup(rank, world_size)
# # distributed.init_process_group('nccl',
# # init_method='env://')
# self.net = DDP(self.net,
# device_ids=[rank],
# output_device=rank)
# eval contains most of the tricky code handling all the cases for nclasses
# to get eval to efficiently run on an entire image set, we could pass a torch dataset
# this dataset could either parse out images loaded from memory or from storage
[docs] def eval(self, x, batch_size=8, indices=None, channels=None, channel_axis=None,
z_axis=None, normalize=True, invert=False,
rescale=None, diameter=None, do_3D=False, anisotropy=None, net_avg=True,
augment=False, tile=True, tile_overlap=0.1, bsize=224, num_workers=8,
resample=True, interp=True, cluster=False, suppress=None,
boundary_seg=False, affinity_seg=False, despur=True,
flow_threshold=0.4, mask_threshold=0.0, diam_threshold=12., niter=None,
cellprob_threshold=None, dist_threshold=None, flow_factor=5.0,
compute_masks=True, min_size=15, max_size=None, stitch_threshold=0.0,
progress=None, show_progress=True,
omni=False, calc_trace=False, verbose=False, transparency=False,
loop_run=False, model_loaded=False,hysteresis=True):
"""
Evaluation for CellposeModel. Segment list of images x, or 4D array - Z x nchan x Y x X
Parameters
----------
x: list or array of images
can be list of 2D/3D/4D images, or array of 2D/3D/4D images
batch_size: int (optional, default 8)
number of 224x224 patches to run simultaneously on the GPU
(can make smaller or bigger depending on GPU memory usage)
channels: list (optional, default None)
list of channels, either of length 2 or of length number of images by 2.
First element of list is the channel to segment (0=grayscale, 1=red, 2=green, 3=blue).
Second element of list is the optional nuclear channel (0=none, 1=red, 2=green, 3=blue).
For instance, to segment grayscale images, input [0,0]. To segment images with cells
in green and nuclei in blue, input [2,3]. To segment one grayscale image and one
image with cells in green and nuclei in blue, input [[0,0], [2,3]].
channel_axis: int (optional, default None)
if None, channels dimension is attempted to be automatically determined
z_axis: int (optional, default None)
if None, z dimension is attempted to be automatically determined
normalize: bool (default, True)
normalize data so 0.0=1st percentile and 1.0=99th percentile of image intensities in each channel
invert: bool (optional, default False)
invert image pixel intensity before running network
rescale: float (optional, default None)
resize factor for each image, if None, set to 1.0
diameter: float (optional, default None)
diameter for each image (only used if rescale is None),
if diameter is None, set to diam_mean
do_3D: bool (optional, default False)
set to True to run 3D segmentation on 4D image input
anisotropy: float (optional, default None)
for 3D segmentation, optional rescaling factor (e.g. set to 2.0 if Z is sampled half as dense as X or Y)
net_avg: bool (optional, default True)
runs the 4 built-in networks and averages them if True, runs one network if False
augment: bool (optional, default False)
tiles image with overlapping tiles and flips overlapped regions to augment
tile: bool (optional, default True)
tiles image to ensure GPU/CPU memory usage limited (recommended)
tile_overlap: float (optional, default 0.1)
fraction of overlap of tiles when computing flows
resample: bool (optional, default True)
run dynamics at original image size (will be slower but create more accurate boundaries)
interp: bool (optional, default True)
interpolate during 2D dynamics (not available in 3D)
(in previous versions it was False)
flow_threshold: float (optional, default 0.4)
flow error threshold (all cells with errors below threshold are kept) (not used for 3D)
mask_threshold: float (optional, default 0.0)
all pixels with value above threshold kept for masks, decrease to find more and larger masks
dist_threshold: float (optional, default None) DEPRECATED
use mask_threshold instead
cellprob_threshold: float (optional, default None) DEPRECATED
use mask_threshold instead
compute_masks: bool (optional, default True)
Whether or not to compute dynamics and return masks.
This is set to False when retrieving the styles for the size model.
min_size: int (optional, default 15)
minimum number of pixels per mask, can turn off with -1
stitch_threshold: float (optional, default 0.0)
if stitch_threshold>0.0 and not do_3D, masks are stitched in 3D to return volume segmentation
progress: pyqt progress bar (optional, default None)
to return progress bar status to GUI
omni: bool (optional, default False)
use omnipose mask reconstruction features
calc_trace: bool (optional, default False)
calculate pixel traces and return as part of the flow
verbose: bool (optional, default False)
turn on additional output to logs for debugging
transparency: bool (optional, default False)
modulate flow opacity by magnitude instead of brightness (can use flows on any color background)
loop_run: bool (optional, default False)
internal variable for determining if model has been loaded, stops model loading in loop over images
model_loaded: bool (optional, default False)
internal variable for determining if model has been loaded, used in __main__.py
Returns
-------
masks: list of 2D arrays, or single 3D array (if do_3D=True)
labelled image, where 0=no masks; 1,2,...=mask labels
flows: list of lists 2D arrays, or list of 3D arrays (if do_3D=True)
flows[k][0] = 8-bit RGb phase plot of flow field
flows[k][1] = flows at each pixel
flows[k][2] = scalar cell probability (Cellpose) or distance transform (Omnipose)
flows[k][3] = boundary output (nonempty for Omnipose)
flows[k][4] = final pixel locations after Euler integration
flows[k][5] = pixel traces (nonempty for calc_trace=True)
styles: list of 1D arrays of length 64, or single 1D array (if do_3D=True)
style vector summarizing each image, also used to estimate size of objects in image
"""
if cellprob_threshold is not None or dist_threshold is not None:
mask_threshold = deprecation_warning_cellprob_dist_threshold(cellprob_threshold, dist_threshold)
# images are given has a list, especially when heterogeneous in shape
is_grey = np.sum(channels)==0
slice_ndim = self.dim+(self.nchan>1 and not is_grey)+(channel_axis is not None)
# the logic here needs to be updated to account for the fact that images may not alreay match the expected dims
# and channels, namely mono channel might have a 2-channel model. I should just check for if the number of channels could
# possibly match, and warn that intenral conversion will happen or may break...
is_list = isinstance(x, list)
is_stack = is_image = False
if isinstance(x, np.ndarray):
# [0,0] is a special instance where we want to run the model on a single channel
dim_diff = x.ndim-slice_ndim
opt = np.array([0,1])#-is_grey
is_image, is_stack = [dim_diff==i for i in opt]
correct_shape = dim_diff in opt
# print('a1',interp,hysteresis,calc_trace)
# allow for a dataset to be passed so that we can do batches
# will be defined in omnipose.data.train_set
is_dataset = isinstance(x,torch.utils.data.Dataset)
if is_dataset:
correct_shape = True # assume the dataset has the right shape
if not (is_list or is_stack or is_dataset or is_image or loop_run):
models_logger.warning('input images must be a list of images, array of images, or dataloader')
else:
if is_list:
correct_shape = np.all([x[i].squeeze().ndim==slice_ndim] for i in range(len(x)))
if not correct_shape:
print(slice_ndim,x.ndim,is_list,is_stack)
models_logger.warning('input images do not match the expected number of dimensions ({}) and channels ({}) of model.'.format(self.dim,self.nchan))
if verbose and (is_dataset or not (is_list or is_stack)):
models_logger.info('Evaluating with flow_threshold %0.2f, mask_threshold %0.2f'%(flow_threshold, mask_threshold))
if omni:
models_logger.info(f'using omni model, cluster {cluster}')
# Note: dataset is finetuned for basic omnipose usage. No styles are returned, some options may not be supported.
if is_dataset:
if verbose:
models_logger.warning('Using dataset evaluation branch. Some options not yet supported.')
# set the tile parameter in dataset
x.tile = tile
# set the rescale parameter in dataset
x.rescale = rescale
# sample indices to evaluate
indices = list(range(len(x))) if indices is None else indices
# the sequential batch sampler gives us a set of indices in sequence, like 0-5, 6-11, etc.
sampler = torch.utils.data.sampler.BatchSampler(omnipose.data.sampler(indices),
batch_size=batch_size,
drop_last=False)
params = {'batch_size': 1, # this batch size is more like how many worker batches to aggregate
# 'shuffle': False, # use sampler instead
'collate_fn': x.collate_fn,
'pin_memory': False, # only useful for CPU tensors
'num_workers': num_workers,
'sampler': sampler,# iterabledataset does not need this
'persistent_workers': True if num_workers>0 else False,
'multiprocessing_context': 'spawn' if num_workers>0 else None,
'prefetch_factor': batch_size if num_workers>0 else None
}
loader = torch.utils.data.DataLoader(x, **params)
dist, dP, bd, masks, bounds, p, tr, affinity, flow_RGB = [], [], [], [], [], [], [], [], []
# I think the loader can at least do all the preprocessing work it will take to figure out
# padding and stitching and slicing
progress_bar = tqdm(total=len(indices),disable=not show_progress)
for batch,inds,subs in loader:
shape = batch.shape
nimg = batch.shape[0]
nchan = batch.shape[1]
shape = batch.shape[-(self.dim+1):] # nclasses, Y, X
resize = shape[-self.dim:] if not resample else None
# define the slice needed to get rid of padding required for net downsamples
slc = [slice(0, s+1) for s in shape]
slc[-(self.dim+1)] = slice(0, self.nclasses + 1)
for k in range(1,self.dim+1):
slc[-k] = slice(subs[-k][0], subs[-k][-1]+1)
slc = tuple(slc)
# catch cases where the images are 1-channel
# but the model is 2 channel
# if self.nchan-nchan:
# print('padding with extra chan dd',batch)
# batch = torch.cat([batch,torch.zeros_like(batch)],dim=1)#.permute(0,2,3,1)
# print('now',batch)
# batch = torch.cat([batch,batch],dim=1)
# batch = torch.cat([torch.zeros_like(batch),batch],dim=1)
# run the network on the batch
# yf, style = self.network(batch)
with torch.no_grad():
self.net.eval() # was missing this - some layers behave differently without it
if tile:
yf = x._run_tiled(batch,self,
batch_size=batch_size,
bsize=bsize,
augment=augment,
tile_overlap=tile_overlap).unsqueeze(0)
else:
yf = self.net(batch)[0]
del batch
# print('need to add normalization / invert /rescale options in dataloader')
# slice out padding
yf = yf[(Ellipsis,)+slc]
# rescale and resample
if resample and rescale!=1.0:
yf = omnipose.data.torch_zoom(yf, 1/rescale)
# compared to the usual per-image pipeline, this one will not support cellpose or u-net
flow_pred = yf[:,:self.dim]
dist_pred = yf[:,self.dim] #scalar field always after the vector field output
if self.nclasses>=self.dim+2:
bd_pred = yf[:,self.dim+1]
else:
bd_pred = torch.empty(nimg)
# clear from memory
del yf
# I made a vastly faster implementation using pytorch
rgb = omnipose.plot.rgb_flow(flow_pred,transparency=transparency)
# I implemented hysteresis with just pytorch
# it is faster than skimage with larger batches, but not by much
# it does better in thin sections, however (though might be broken skeleton fragments)
# I might just replace the main branch code with this
if hysteresis:
foreground = hysteresis_threshold(dist_pred.unsqueeze(1),mask_threshold-1, mask_threshold).squeeze(dim=1)
else:
foreground = dist_pred >= mask_threshold
# print('add flag')
# print('fg_here',torch.sum(foreground))
# vf = interp_vf(flow_pred/5., mode = "nearest_batched")
# initial_points = init_values_semantic(foreground, device=self.device)
shape = flow_pred.shape
B = shape[0]
dims = shape[-self.dim:]
coords = [torch.arange(0, l, device = self.device) for l in dims]
mesh = torch.meshgrid(coords, indexing = "ij")
init_shape = [B, 1] + ([1] * len(dims))
initial_points = torch.stack(mesh, dim = 0) # torchvf flips with mesh[::-1]
initial_points = initial_points.repeat(init_shape).float()
# final_points = ivp_solver(vf,initial_points,
# dx = 1,
# n_steps = 8,
# solver = "euler")[-1]
# these are equivalent
coords = torch.nonzero(foreground,as_tuple=True)
# coords = custom_nonzero_cuda(foreground.squeeze())
# coords = torch.where(foreground.squeeze())
# this block works
# # Assuming foreground is a boolean tensor of shape (B, D1, D2, ..., DN)
# fg = foreground.squeeze() # Now fg has shape (B, D1, D2, ..., DN)
# # Create a grid of indices
# grids = torch.meshgrid([torch.arange(size, device=fg.device) for size in fg.shape])
# # Stack the grids to create an index mesh
# index_mesh = torch.stack(grids, dim=0) # Now index_mesh has shape (N+1, B, D1, D2, ..., DN)
# # Move index_mesh to the same device as foreground
# index_mesh = index_mesh.to(fg.device)
# # Use the boolean tensor to index into the index mesh
# selected_indices = index_mesh[:, fg]
# coords = tuple(selected_indices)
# fg = foreground.squeeze() # Now fg has shape (B, D1, D2, ..., DN)
# # Create a grid of indices
# grids = torch.meshgrid([torch.arange(size, device=fg.device) for size in fg.shape])
# # Reshape each grid to have shape (-1)
# reshaped_grids = [grid.reshape(-1) for grid in grids]
# # Convert the reshaped grids to a tuple of indices
# selected_indices = tuple(reshaped_grids)
# # print(len(reshaped_grids),reshaped_grids[0].shape,reshaped_grids)
# coords = tuple(selected_indices)
# add to output lists
dP.extend(self._from_device(flow_pred))
dist.extend(self._from_device(dist_pred))
bd.extend(self._from_device(bd_pred))
flow_RGB.extend(self._from_device(rgb))
if torch.any(foreground):
cell_px = (Ellipsis,)+coords[-self.dim:]
if niter is None:
# niter = omnipose.core.get_niter(dist_pred).cpu()
# int(diameters(foreground,dist_pred)/(1+affinity_seg))
niter = int(2*(self.dim+1)*torch.mean(dist_pred[(Ellipsis,)+coords]) / (1+affinity_seg))
if verbose:
models_logger.info('niter set to %d'%niter)
final_points = initial_points.clone()
final_p, traced_p = omnipose.core.steps_batch(initial_points[cell_px],
flow_pred/5., #<<<<<<<<<<< add support for other options here
niter=niter,
omni=omni,
suppress=suppress,
interp=interp,
verbose=verbose,
calc_trace=calc_trace)
final_points[cell_px] = final_p.squeeze()
if affinity_seg:
steps, inds, idx, fact, sign = omnipose.utils.kernel_setup(self.dim)
affinity_graph = omnipose.core._get_affinity_torch(initial_points,
final_points,
flow_pred/5., #<<<<<<<<<<< add support for other options here
dist_pred,
foreground,
steps,
fact,
niter,
)
# cast to CPU for compute_masks
final_points = self._from_device(final_points)
traced_p = self._from_device(traced_p) if traced_p is not None else [None]*B
affinity_graph = self._from_device(affinity_graph).swapaxes(0,1) if affinity_seg else [None]*B
foreground = self._from_device(foreground)
dist_pred = self._from_device(dist_pred)
flow_pred = self._from_device(flow_pred)
bd_pred = self._from_device(bd_pred)
rgb = self._from_device(rgb)
# can loop through batch and run compute_masks
for iscell, disti, dPi, bdi, agi, pts, trp in zip(foreground, dist_pred, flow_pred, bd_pred, affinity_graph, final_points, traced_p):
parallel = 1
coords = np.nonzero(iscell)
# print('agi 33',agi.shape, affinity_graph.shape, np.sum(iscell), np.stack(coords).shape)
# agi = None
# print('torch computed affinity not quite ready yet ')
# print('PARALLEL', parallel)
#NOW THAT THE trajectories are "WORKING", I need to add the parallel affinity here
outputs = omnipose.core.compute_masks(dPi, disti,
affinity_graph=agi[(Ellipsis,)+coords] if agi is not None else agi,
bd = bdi,
p=pts.squeeze() if parallel else None,
coords = np.stack(coords),
iscell=iscell if parallel else None,
niter=niter,
rescale=rescale,
resize=resize,
min_size=min_size,
max_size=max_size,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
flow_threshold=flow_threshold,
flow_factor=flow_factor,
interp=interp,
cluster=cluster,
boundary_seg=boundary_seg,
affinity_seg=affinity_seg,
despur=despur,
calc_trace=calc_trace,
verbose=verbose,
use_gpu=self.gpu,
device=self.device,
nclasses=self.nclasses,
dim=self.dim)
masks.append(outputs[0])
p.append(outputs[1])
# tr.append(outputs[2])
tr.append(trp)
bounds.append(outputs[3])
affinity.append(outputs[4])
progress_bar.update()
empty_cache()
else:
progress_bar.update()
empty_cache()
models_logger.info('no cell pixels found')
masks = np.zeros((B,)+dims)
bounds = np.zeros((B,)+dims)
affinity = [None for _ in range(B)]
tr = [None for _ in range(B)]
p = [None for _ in range(B)]
masks = np.array(masks)
bounds = np.array(bounds)
p = np.array(p)
tr = np.array(tr)
ret = [masks, dP, dist, p, bd, tr, affinity, bounds, flow_RGB]
progress_bar.close()
for r in ret:
r.squeeze() if isinstance(r,np.ndarray) else r
# the flow list stores:
# (1) RGB representation of flows
# (2) flow components
# (3) cellprob (cp) or distance field (op)
# (4) pixel coordinates after Euler integration
# (5) boundary output (nclasses=4)
# (6) pixel trajectories during Euler integation (trace=True)
# (7) nstep_by_npix affinity graph
# (8) binary boundary map
# 5-8 were added in Omnipose, hence the unusual placement in the list.
# flows = [[o for o in out] for out in zip(rgb, dP, cellprob, p, bd, tr, affinity, bounds)]
flows = [list(item) for item in zip(flow_RGB, dP, dist, p, bd, tr, affinity, bounds)] # not sure which is faster of these yet
return masks, flows, []
elif (is_list or is_stack) and correct_shape:
masks, styles, flows = [], [], []
tqdm_out = utils.TqdmToLogger(models_logger, level=logging.INFO)
nimg = len(x)
iterator = trange(nimg, file=tqdm_out,disable=not show_progress) if nimg>1 else range(nimg)
# note: ~ is bitwise flip, overloaded to act as elementwise not for numpy arrays
# but for boolean variables, must use "not" operator isstead
for i in iterator:
dia = diameter[i] if isinstance(diameter, list) or isinstance(diameter, np.ndarray) else diameter
rsc = rescale[i] if isinstance(rescale, list) or isinstance(rescale, np.ndarray) else rescale
chn = channels if channels is None else channels[i] if (len(channels)==len(x) and
(isinstance(channels[i], list)
or isinstance(channels[i], np.ndarray)) and
len(channels[i])==2) else channels
maski, stylei, flowi = self.eval(x[i],
batch_size=batch_size,
channels = chn,
channel_axis=channel_axis,
z_axis=z_axis,
normalize=normalize,
invert=invert,
rescale=rsc,
diameter=dia,
do_3D=do_3D,
anisotropy=anisotropy,
net_avg=net_avg,
augment=augment,
tile=tile,
tile_overlap=tile_overlap,
bsize=bsize,
resample=resample,
interp=interp,
cluster=cluster,
suppress=suppress,
boundary_seg=boundary_seg,
affinity_seg=affinity_seg,
despur=despur,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
flow_threshold=flow_threshold,
niter=niter,
flow_factor=flow_factor,
compute_masks=compute_masks,
min_size=min_size,
max_size=max_size,
stitch_threshold=stitch_threshold,
progress=progress,
show_progress=show_progress,
omni=omni,
calc_trace=calc_trace,
verbose=verbose,
transparency=transparency,
loop_run=(i>0),
model_loaded=model_loaded)
masks.append(maski)
flows.append(flowi)
styles.append(stylei)
return masks, styles, flows
else:
if not model_loaded and (isinstance(self.pretrained_model, list) and not net_avg and not loop_run):
# whether or not we are using dataparallel
if self.torch and self.gpu:
models_logger.info('using dataparallel')
net = self.net.module
else:
net = self.net
models_logger.info('not using dataparallel')
net.load_model(self.pretrained_model[0], cpu=(not self.gpu))
if not self.torch:
net.collect_params().grad_req = 'null'
if verbose: models_logger.info('shape before transforms.convert_image(): {}'.format(x.shape))
# This takes care of the special case of grasycale, padding with zeros if the model was trained like that
x = transforms.convert_image(x, channels, channel_axis=channel_axis, z_axis=z_axis,
do_3D=(do_3D or stitch_threshold>0), normalize=False,
invert=False, nchan=self.nchan, dim=self.dim, omni=omni)
if verbose: models_logger.info('shape after transforms.convert_image(): {}'.format(x.shape))
if x.ndim < self.dim+2: # we need (nimg, *dims, nchan), so 2D has 4, 3D has 5, etc.
x = x[np.newaxis]
if verbose: models_logger.info('shape now {}'.format(x.shape))
self.batch_size = batch_size
rescale = self.diam_mean / diameter if (rescale is None and (diameter is not None and diameter>0)) else rescale
rescale = 1.0 if rescale is None else rescale
masks, styles, dP, cellprob, p, bd, tr, affinity, bounds = self._run_cp(x,
compute_masks=compute_masks,
normalize=normalize,
invert=invert,
rescale=rescale,
net_avg=net_avg,
resample=resample,
augment=augment,
tile=tile,
tile_overlap=tile_overlap,
bsize=bsize,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
flow_threshold=flow_threshold,
niter=niter,
flow_factor=flow_factor,
interp=interp,
cluster=cluster,
suppress=suppress,
boundary_seg=boundary_seg,
affinity_seg=affinity_seg,
despur=despur,
min_size=min_size,
max_size=max_size,
do_3D=do_3D,
anisotropy=anisotropy,
stitch_threshold=stitch_threshold,
omni=omni,
calc_trace=calc_trace,
verbose=verbose)
# the flow list stores:
# (1) RGB representation of flows
# (2) flow components
# (3) cellprob (cp) or distance field (op)
# (4) pixel coordinates after Euler integration
# (5) boundary output (nclasses=4)
# (6) pixel trajectories during Euler integation (trace=True)
# (7) augmented affinity graph (coords+affinity) of shape (dim,nstep,npix)
# (8) binary boundary map
# 5-8 were added in Omnipose, hence the unusual placement in the list.
flows = [plot.dx_to_circ(dP,transparency=transparency)
if self.nclasses>1 else np.zeros(cellprob.shape+(3+transparency,),np.uint8),
dP, cellprob, p, bd, tr, affinity, bounds]
return masks, flows, styles
def _run_cp(self, x, compute_masks=True, normalize=True, invert=False,
rescale=1.0, net_avg=True, resample=True,
augment=False, tile=True, tile_overlap=0.1, bsize=224,
mask_threshold=0.0, diam_threshold=12., flow_threshold=0.4, niter=None, flow_factor=5.0,
min_size=15, max_size=None,
interp=True, cluster=False, suppress=None, boundary_seg=False, affinity_seg=False, despur=True,
anisotropy=1.0, do_3D=False, stitch_threshold=0.0,
omni=False, calc_trace=False, verbose=False, pad=0):
# by this point, the image(s) will already have been formatted with channels, batch, etc
tic = time.time()
shape = x.shape
nimg = shape[0]
bd, tr, affinity = None, None, None
# set up image padding for prediction - set to 0 not as it actually doesn't really help
# note that this is not the same padding as what you need for the network to run
pad_seq = [(pad,)*2]*self.dim + [(0,)*2] # do not pad channel axis
unpad = tuple([slice(pad,-pad) if pad else slice(None,None)]*self.dim) # works in case pad is zero
if do_3D:
img = np.asarray(x)
if normalize or invert: # possibly make normalize a vector of upper-lower values
img = transforms.normalize_img(img, invert=invert, omni=omni)
# have not tested padding in do_3d yet
# img = np.pad(img,pad_seq,'reflect')
yf, styles = self._run_3D(img, rsz=rescale, anisotropy=anisotropy,
net_avg=net_avg, augment=augment, tile=tile,
tile_overlap=tile_overlap)
# unpadding
# yf = yf[unpad+(Ellipsis,)]
cellprob = np.sum([yf[k][2] for k in range(3)],axis=0)/3 if omni else np.sum([yf[k][2] for k in range(3)],axis=0)
bd = np.sum([yf[k][3] for k in range(3)],axis=0)/3 if self.nclasses==(self.dim+2) else np.zeros_like(cellprob)
dP = np.stack((yf[1][0] + yf[2][0], yf[0][0] + yf[2][1], yf[0][1] + yf[1][1]), axis=0) # (dZ, dY, dX)
if omni:
dP = np.stack([gaussian_filter(dP[a],sigma=1.5) for a in range(3)]) # remove some artifacts
bd = gaussian_filter(bd,sigma=1.5)
cellprob = gaussian_filter(cellprob,sigma=1.5)
dP = dP/2 #should be averaging components
del yf
else:
tqdm_out = utils.TqdmToLogger(models_logger, level=logging.INFO,)
iterator = trange(nimg, file=tqdm_out, disable=not show_progress) if nimg>1 else range(nimg)
styles = np.zeros((nimg, self.nbase[-1]), np.float32)
#indexing a little weird here due to channels being last now
if resample:
s = tuple(shape[-(self.dim+1):-1])
else:
s = tuple(np.round(np.array(shape[-(self.dim+1):-1])*rescale).astype(int))
dP = np.zeros((self.dim, nimg,)+s, np.float32)
cellprob = np.zeros((nimg,)+s, np.float32)
bounds = np.zeros((nimg,)+s, bool)
for i in iterator:
img = np.asarray(x[i])
# at this point, img should be (*DIMS,C)
if normalize or invert:
img = transforms.normalize_img(img, invert=invert, omni=omni)
# at this point, img should still be (*DIMS,C)
# pad the image to get cleaner output at the edges
# padding with edge values seems to work the best
# but actually, not really useful in the end...
if pad>0:
img = np.pad(img,pad_seq,'edge')
if rescale != 1.0:
# if self.dim>2:
# print('WARNING, resample not updated for ND')
# img = transforms.resize_image(img, rsz=rescale)
if img.ndim>self.dim: # then there is a channel axis, assume it is last here
img = np.stack([zoom(img[...,k],rescale,order=3) for k in range(img.shape[-1])],axis=-1)
else:
img = zoom(img,rescale,order=1)
yf, style = self._run_nets(img, net_avg=net_avg,
augment=augment, tile=tile,
tile_overlap=tile_overlap,
bsize=bsize)
# unpadding
yf = yf[unpad+(Ellipsis,)]
# resample interpolates the network output to native resolution prior to running Euler integration
# this means the masks will have no scaling artifacts. We could *upsample* by some factor to make
# the clustering etc. work even better, but that is not implemented yet
if resample and rescale!=1.0:
# for k in range(yf.shape[-1]):
# print('a',shape[1:1+self.dim]/np.array(yf.shape[:-1]))
# ND version actually gives better results than CV2 in some places.
yf = np.stack([zoom(yf[...,k], shape[1:1+self.dim]/np.array(yf.shape[:-1]), order=1)
for k in range(yf.shape[-1])],axis=-1)
# scipy.ndimage.affine_transform(A, np.linalg.inv(M), output_shape=tyx,
if self.nclasses>1:
cellprob[i] = yf[...,self.dim] #scalar field always after the vector field output
order = (self.dim,)+tuple([k for k in range(self.dim)]) #(2,0,1)
dP[:, i] = yf[...,:self.dim].transpose(order)
else:
cellprob[i] = yf[...,0]
# dP[i] = np.zeros(cellprob)
if self.nclasses>=self.dim+2:
if i==0:
bd = np.zeros_like(cellprob)
bd[i] = yf[...,self.dim+1]
styles[i] = style
del yf, style
styles = styles.squeeze()
net_time = time.time() - tic
if nimg > 1:
models_logger.info('network run in %2.2fs'%(net_time))
if compute_masks:
tic = time.time()
# allow user to specify niter
# Cellpose default is 200
# Omnipose default is None (dynamically adjusts based on distance field)
if niter is None and not omni:
niter = 200 if (do_3D and not resample) else (1 / rescale * 200)
if do_3D:
if not (omni and OMNI_INSTALLED):
# run cellpose compute_masks
masks, bounds, p, tr = dynamics.compute_masks(dP, cellprob,
bd=bd,
niter=niter,
resize=None,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
flow_threshold=flow_threshold,
interp=interp,
do_3D=do_3D,
min_size=min_size,
verbose=verbose,
use_gpu=self.gpu,
device=self.device,
nclasses=self.nclasses,
calc_trace=calc_trace)
affinity = []
else:
# run omnipose compute_masks
masks, bounds, p, tr, affinity = omnipose.core.compute_masks(dP, cellprob,
bd=bd,
do_3D=do_3D,
niter=niter,
resize=None,
min_size=min_size,
max_size=max_size,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
flow_threshold=flow_threshold,
flow_factor=flow_factor,
interp=interp,
cluster=cluster,
suppress=suppress,
boundary_seg=boundary_seg,
affinity_seg=affinity_seg,
despur=despur,
calc_trace=calc_trace,
verbose=verbose,
use_gpu=self.gpu,
device=self.device,
nclasses=self.nclasses,
dim=self.dim)
else:
masks, bounds, p, tr, affinity = [], [], [], [], []
resize = shape[-(self.dim+1):-1] if not resample else None
# print('compute masks 2',resize,shape,resample)
for i in iterator:
if not (omni and OMNI_INSTALLED):
# run cellpose compute_masks
outputs = dynamics.compute_masks(dP[:,i], cellprob[i], niter=niter,
mask_threshold=mask_threshold,
flow_threshold=flow_threshold,
interp=interp,
resize=resize,
verbose=verbose,
use_gpu=self.gpu,
device=self.device,
nclasses=self.nclasses,
calc_trace=calc_trace)
outputs = outputs + ([],) # affinity placeholder
else:
# run omnipose compute_masks
# important: resampling means that pixels need to go farther to cluser together;
# niter should be determined by dist, first of all; it currently is already scaled for resampling, good!
# dP needs to be scaled for magnitude to get pixels to move the same relative distance
# eps probably should be left the same if the above are changed
# if resample:
# print('rescale is',rescale,resize)
# dP[:,i] /= rescale this does nothign here since I normalize the flow anyway, have to pass in
bdi = bd[i] if bd is not None else None
outputs = omnipose.core.compute_masks(dP[:,i], cellprob[i],
bd=bdi,
niter=niter,
rescale=rescale,
resize=resize,
min_size=min_size,
max_size=max_size,
mask_threshold=mask_threshold,
diam_threshold=diam_threshold,
flow_threshold=flow_threshold,
flow_factor=flow_factor,
interp=interp,
cluster=cluster,
suppress=suppress,
boundary_seg=boundary_seg,
affinity_seg=affinity_seg,
despur=despur,
calc_trace=calc_trace,
verbose=verbose,
use_gpu=self.gpu,
device=self.device,
nclasses=self.nclasses,
dim=self.dim)
masks.append(outputs[0])
p.append(outputs[1])
tr.append(outputs[2])
bounds.append(outputs[3])
affinity.append(outputs[4])
masks = np.array(masks)
bounds = np.array(bounds)
p = np.array(p)
tr = np.array(tr)
affinity = np.array(affinity)
if stitch_threshold > 0 and nimg > 1:
models_logger.info(f'stitching {nimg} planes using stitch_threshold={stitch_threshold:0.3f} to make 3D masks')
masks = utils.stitch3D(masks, stitch_threshold=stitch_threshold)
flow_time = time.time() - tic
if nimg > 1:
models_logger.info('masks created in %2.2fs'%(flow_time))
ret = [masks, styles, dP, cellprob, p, bd, tr, affinity, bounds]
ret = [r.squeeze() if isinstance(r,np.ndarray) else r for r in ret]
else:
#pass back zeros for masks and p if not compute_masks
ret = [[], styles, dP, cellprob, [], bd, tr, affinity, bounds]
empty_cache()
return (*ret,)
[docs] def loss_fn(self, lbl, y):
"""
loss function between true labels lbl and prediction y
This is the one used to train the instance segmentation network.
"""
loss = omnipose.core.loss(self, lbl, y)
# if self.omni and OMNI_INSTALLED: #loss function for omnipose fields
# loss = omnipose.core.loss(self, lbl, y)
# else: # original loss function
# veci = 5. * self._to_device(lbl[:,1:])
# lbl = self._to_device(lbl[:,0]>.5)
# loss = self.criterion(y[:,:2] , veci)
# if self.torch:
# loss /= 2.
# loss2 = self.criterion2(y[:,2] , lbl)
# loss = loss + loss2
return loss
[docs] def train(self, train_data, train_labels, train_links=None, train_files=None,
test_data=None, test_labels=None, test_links=None, test_files=None,
channels=None, channel_axis=0, normalize=True,
save_path=None, save_every=100, save_each=False,
learning_rate=0.2, n_epochs=500, momentum=0.9, SGD=True,
weight_decay=0.00001, batch_size=8, dataloader=False, num_workers=0, nimg_per_epoch=None,
rescale=True, min_train_masks=5, netstr=None, tyx=None, timing=False, do_autocast=False,
affinity_field=False):
""" train network with images train_data
Parameters
------------------
train_data: list of arrays (2D or 3D)
images for training
train_labels: list of arrays (2D or 3D)
labels for train_data, where 0=no masks; 1,2,...=mask labels
can include flows as additional images
train_links: list of label links
These lists of label pairs define which labels are "linked",
i.e. should be treated as part of the same object. This is how
Omnipose handles internal/self-contact boundaries during training.
train_files: list of strings
file names for images in train_data (to save flows for future runs)
test_data: list of arrays (2D or 3D)
images for testing
test_labels: list of arrays (2D or 3D)
See train_labels.
test_links: list of label links
See train_links.
test_files: list of strings
file names for images in test_data (to save flows for future runs)
channels: list of ints (default, None)
channels to use for training
normalize: bool (default, True)
normalize data so 0.0=1st percentile and 1.0=99th percentile of image intensities in each channel
save_path: string (default, None)
where to save trained model, if None it is not saved
save_every: int (default, 100)
save network every [save_every] epochs
learning_rate: float or list/np.ndarray (default, 0.2)
learning rate for training, if list, must be same length as n_epochs
n_epochs: int (default, 500)
how many times to go through whole training set during training
weight_decay: float (default, 0.00001)
SGD: bool (default, True)
use SGD as optimization instead of RAdam
batch_size: int (optional, default 8)
number of tyx-sized patches to run simultaneously on the GPU
(can make smaller or bigger depending on GPU memory usage)
nimg_per_epoch: int (optional, default None)
minimum number of images to train on per epoch,
with a small training set (< 8 images) it may help to set to 8
rescale: bool (default, True)
whether or not to rescale images to diam_mean during training,
if True it assumes you will fit a size model after training or resize your images accordingly,
if False it will try to train the model to be scale-invariant (works worse)
min_train_masks: int (default, 5)
minimum number of masks an image must have to use in training set
netstr: str (default, None)
name of network, otherwise saved with name as params + training start time
tyx: int, tuple (default, 224x224 in 2D)
size of image patches used for training
"""
# torch.backends.cudnn.benchmark = True
# rank = args.nr * args.gpus + gpu
# distributed.init_process_group(backend='nccl',
# init_method='env://',
# world_size=args.world_size,
# rank=rank
# )
if rescale:
models_logger.info(f'Training with rescale = {rescale:.2f}')
# images may need some dimension shuffling to conform to standard, this is link-independent
train_data, train_labels, test_data, test_labels, run_test = transforms.reshape_train_test(train_data, train_labels,
test_data, test_labels,
channels, channel_axis,
normalize,
self.dim, self.omni)
# print('shape', train_data[0].shape, channels, train_labels[0].shape)
# check if train_labels have flows
# if not, flows computed, returned with labels as train_flows[i][0]
labels_to_flows = dynamics.labels_to_flows if not (self.omni and OMNI_INSTALLED) else omnipose.core.labels_to_flows
# Omnipose needs to recompute labels on-the-fly after image warping
models_logger.info('No precomuting flows with Omnipose. Computed during training.')
# We assume that if links are given, labels are properly formatted as 0,1,2,...,N
# might be worth implementing a remapping for the links just in case...
# for now, just skip this for any labels that come with a link file
for i,(labels,links) in enumerate(zip(train_labels,train_links)):
if links is None:
train_labels[i] = omnipose.utils.format_labels(labels)
# nmasks is inflated when using multi-label objects, so keep that in mind if you care about min_train_masks
nmasks = np.array([label.max() for label in train_labels])
# if self.omni and OMNI_INSTALLED:
# models_logger.info('No precomuting flows with Omnipose. Computed during training.')
# # We assume that if links are given, labels are properly formatted as 0,1,2,...,N
# # might be worth implementing a remapping for the links just in case...
# # for now, just skip this for any labels that come with a link file
# for i,(labels,links) in enumerate(zip(train_labels,train_links)):
# if links is None:
# train_labels[i] = omnipose.utils.format_labels(labels)
# # nmasks is inflated when using multi-label objects, so keep that in mind if you care about min_train_masks
# nmasks = np.array([label.max() for label in train_labels])
# else:
# train_labels = labels_to_flows(labels=train_labels, links=train_links, files=train_files,
# use_gpu=self.gpu, device=self.device, dim=self.dim)
# nmasks = np.array([label[0].max() for label in train_labels])
if run_test:
test_labels = labels_to_flows(test_labels, test_links, files=test_files,
use_gpu=self.gpu, device=self.device, dim=self.dim)
else:
test_labels = None
nremove = (nmasks < min_train_masks).sum()
if nremove > 0:
models_logger.warning(f'{nremove} train images with number of masks less than min_train_masks ({min_train_masks}), removing from train set')
ikeep = np.nonzero(nmasks >= min_train_masks)[0]
train_data = [train_data[i] for i in ikeep]
train_labels = [train_labels[i] for i in ikeep]
train_links = [train_links[i] for i in ikeep]
if channels is None:
models_logger.warning('channels is set to None, input must therefore have nchan channels (default is 2)')
model_path = self._train_net(train_data, train_labels, train_links,
test_data=test_data, test_labels=test_labels, test_links=test_links,
save_path=save_path, save_every=save_every, save_each=save_each,
learning_rate=learning_rate, n_epochs=n_epochs,
momentum=momentum, weight_decay=weight_decay,
SGD=SGD, batch_size=batch_size, dataloader=dataloader, num_workers=num_workers,
nimg_per_epoch=nimg_per_epoch, do_autocast=do_autocast, affinity_field=affinity_field,
rescale=rescale, netstr=netstr, tyx=tyx, timing=timing)
self.pretrained_model = model_path
return model_path
[docs]class SizeModel():
""" linear regression model for determining the size of objects in image
used to rescale before input to cp_model
uses styles from cp_model
Parameters
-------------------
cp_model: UnetModel or CellposeModel
model from which to get styles
device: mxnet device (optional, default mx.cpu())
where cellpose model is saved (mx.gpu() or mx.cpu())
pretrained_size: str
path to pretrained size model
omni: bool
whether or not to use distance-based size metrics
corresponding to 'omni' model
"""
def __init__(self, cp_model, device=None, pretrained_size=None, **kwargs):
super(SizeModel, self).__init__(**kwargs)
self.pretrained_size = pretrained_size
self.cp = cp_model
self.device = self.cp.device
self.diam_mean = self.cp.diam_mean
self.torch = self.cp.torch
if pretrained_size is not None:
self.params = np.load(self.pretrained_size, allow_pickle=True).item()
self.diam_mean = self.params['diam_mean']
if not hasattr(self.cp, 'pretrained_model'):
error_message = 'no pretrained cellpose model specified, cannot compute size'
models_logger.critical(error_message)
raise ValueError(error_message)
[docs] def eval(self, x, channels=None, channel_axis=None,
normalize=True, invert=False, augment=False, tile=True,
batch_size=8, progress=None, interp=True, omni=False):
""" Evaluation for SizeModel. Use images x to produce style or use style input to predict size of objects in image.
Object size estimation is done in two steps:
1. use a linear regression model to predict size from style in image
2. resize image to predicted size and run CellposeModel to get output masks.
Take the median object size of the predicted masks as the final predicted size.
Parameters
-------------------
x: list or array of images
can be list of 2D/3D images, or array of 2D/3D images
channels: list (optional, default None)
list of channels, either of length 2 or of length number of images by 2.
First element of list is the channel to segment (0=grayscale, 1=red, 2=green, 3=blue).
Second element of list is the optional nuclear channel (0=none, 1=red, 2=green, 3=blue).
For instance, to segment grayscale images, input [0,0]. To segment images with cells
in green and nuclei in blue, input [2,3]. To segment one grayscale image and one
image with cells in green and nuclei in blue, input [[0,0], [2,3]].
channel_axis: int (optional, default None)
if None, channels dimension is attempted to be automatically determined
normalize: bool (default, True)
normalize data so 0.0=1st percentile and 1.0=99th percentile of image intensities in each channel
invert: bool (optional, default False)
invert image pixel intensity before running network
augment: bool (optional, default False)
tiles image with overlapping tiles and flips overlapped regions to augment
tile: bool (optional, default True)
tiles image to ensure GPU/CPU memory usage limited (recommended)
progress: pyqt progress bar (optional, default None)
to return progress bar status to GUI
Returns
-------
diam: array, float
final estimated diameters from images x or styles style after running both steps
diam_style: array, float
estimated diameters from style alone
"""
if isinstance(x, list):
diams, diams_style = [], []
nimg = len(x)
tqdm_out = utils.TqdmToLogger(models_logger, level=logging.INFO)
iterator = trange(nimg, file=tqdm_out) if nimg>1 else range(nimg)
for i in iterator:
diam, diam_style = self.eval(x[i],
channels=channels[i] if (len(channels)==len(x) and
(isinstance(channels[i], list) or isinstance(channels[i], np.ndarray)) and
len(channels[i])==2) else channels,
channel_axis=channel_axis,
normalize=normalize,
invert=invert,
augment=augment,
tile=tile,
batch_size=batch_size,
progress=progress,
omni=omni)
diams.append(diam)
diams_style.append(diam_style)
return diams, diams_style
if x.squeeze().ndim > 3:
models_logger.warning('image is not 2D cannot compute diameter')
return self.diam_mean, self.diam_mean
styles = self.cp.eval(x,
channels=channels,
channel_axis=channel_axis,
normalize=normalize,
invert=invert,
augment=augment,
tile=tile,
batch_size=batch_size,
net_avg=False,
resample=False,
compute_masks=False)[-1]
diam_style = self._size_estimation(np.array(styles))
diam_style = self.diam_mean if (diam_style==0 or np.isnan(diam_style)) else diam_style
masks = self.cp.eval(x,
channels=channels,
channel_axis=channel_axis,
normalize=normalize,
invert=invert,
augment=augment,
tile=tile,
batch_size=batch_size,
net_avg=False,
resample=False,
rescale = self.diam_mean / diam_style if self.diam_mean>0 else 1,
#flow_threshold=0,
diameter=None,
interp=False,
omni=omni)[0]
# allow backwards compatibility to older scale metric
diam = utils.diameters(masks,omni=omni)[0]
if hasattr(self, 'model_type') and (self.model_type=='nuclei' or self.model_type=='cyto') and not self.torch and not omni:
diam_style /= (np.pi**0.5)/2
diam = self.diam_mean / ((np.pi**0.5)/2) if (diam==0 or np.isnan(diam)) else diam
else:
diam = self.diam_mean if (diam==0 or np.isnan(diam)) else diam
return diam, diam_style
def _size_estimation(self, style):
""" linear regression from style to size
sizes were estimated using "diameters" from square estimates not circles;
therefore a conversion factor is included (to be removed)
"""
szest = np.exp(self.params['A'] @ (style - self.params['smean']).T +
np.log(self.diam_mean) + self.params['ymean'])
szest = np.maximum(5., szest)
return szest
## Probably need channel axis here too
[docs] def train(self, train_data, train_labels,
test_data=None, test_labels=None,
channels=None, normalize=True,
learning_rate=0.2, n_epochs=10,
l2_regularization=1.0, batch_size=8): # dim and omni set by the model
""" train size model with images train_data to estimate linear model from styles to diameters
Parameters
------------------
train_data: list of arrays (2D or 3D)
images for training
train_labels: list of arrays (2D or 3D)
labels for train_data, where 0=no masks; 1,2,...=mask labels
can include flows as additional images
channels: list of ints (default, None)
channels to use for training
normalize: bool (default, True)
normalize data so 0.0=1st percentile and 1.0=99th percentile of image intensities in each channel
n_epochs: int (default, 10)
how many times to go through whole training set (taking random patches) for styles for diameter estimation
l2_regularization: float (default, 1.0)
regularize linear model from styles to diameters
batch_size: int (optional, default 8)
number of 224x224 patches to run simultaneously on the GPU
(can make smaller or bigger depending on GPU memory usage)
"""
batch_size /= 2 # reduce batch_size by factor of 2 to use larger tiles
batch_size = int(max(1, batch_size))
self.cp.batch_size = batch_size
print('sizemodel',self.cp.dim,self.cp.omni)
train_data, train_labels, test_data, test_labels, run_test = transforms.reshape_train_test(train_data, train_labels,
test_data, test_labels,
channels, normalize, self.cp.dim, self.cp.omni)
if isinstance(self.cp.pretrained_model, list):
cp_model_path = self.cp.pretrained_model[0]
self.cp.net.load_model(cp_model_path, cpu=(not self.cp.gpu))
if not self.torch:
self.cp.net.collect_params().grad_req = 'null'
else:
cp_model_path = self.cp.pretrained_model
diam_train = np.array([utils.diameters(lbl,omni=self.cp.omni)[0] for lbl in train_labels])
if run_test:
diam_test = np.array([utils.diameters(lbl,omni=self.cp.omni)[0] for lbl in test_labels])
# remove images with no masks
for i in range(len(diam_train)):
if diam_train[i]==0.0:
del train_data[i]
del train_labels[i]
if run_test:
for i in range(len(diam_test)):
if diam_test[i]==0.0:
del test_data[i]
del test_labels[i]
nimg = len(train_data)
styles = np.zeros((n_epochs*nimg, 256), np.float32)
diams = np.zeros((n_epochs*nimg,), np.float32)
tic = time.time()
for iepoch in range(n_epochs):
iall = np.arange(0,nimg,1,int)
for ibatch in range(0,nimg,batch_size):
inds = iall[ibatch:ibatch+batch_size]
#using the orignal rotate and resize just because mine apparently broke the feature that
# you could either pass in flows or masks... will eventually fix and streamline
imgi,lbl,scale = transforms.original_random_rotate_and_resize([train_data[i] for i in inds],
Y=[train_labels[i].astype(np.int16) for i in inds],
scale_range=1, xy=(512,512))
feat = self.cp.network(imgi)[1]
styles[inds+nimg*iepoch] = feat
diams[inds+nimg*iepoch] = np.log(diam_train[inds]) - np.log(self.diam_mean) + np.log(scale)
del feat
if (iepoch+1)%2==0:
models_logger.info('ran %d epochs in %0.3f sec'%(iepoch+1, time.time()-tic))
# create model
smean = styles.mean(axis=0)
X = ((styles - smean).T).copy()
ymean = diams.mean()
y = diams - ymean
A = np.linalg.solve(X@X.T + l2_regularization*np.eye(X.shape[0]), X @ y)
ypred = A @ X
models_logger.info('train correlation: %0.4f'%np.corrcoef(y, ypred)[0,1])
if run_test:
nimg_test = len(test_data)
styles_test = np.zeros((nimg_test, 256), np.float32)
for i in range(nimg_test):
styles_test[i] = self.cp._run_net(test_data[i].transpose((1,2,0)))[1]
diam_test_pred = np.exp(A @ (styles_test - smean).T + np.log(self.diam_mean) + ymean)
diam_test_pred = np.maximum(5., diam_test_pred)
models_logger.info('test correlation: %0.4f'%np.corrcoef(diam_test, diam_test_pred)[0,1])
self.pretrained_size = cp_model_path+'_size.npy'
self.params = {'A': A, 'smean': smean, 'diam_mean': self.diam_mean, 'ymean': ymean}
np.save(self.pretrained_size, self.params)
models_logger.info('model saved to '+self.pretrained_size)
return self.params