Merge pull request #44 from rajewsky-lab/fix_imaging_preprocessing

Fix imaging preprocessing

docs/computational/generate_expression_matrix.md

@@ -13,7 +13,7 @@ To create such a spatial cell-by-gene ($M\times G$) expression matrix, you will
 
 We efficiently segment cells (or nuclei) from staining images using [cellpose](https://github.com/MouseLand/cellpose).
 We provide a model that we fine-tuned for segmentation of fresh-frozen, H&E-stained tissue,
-[here](https://github.com/danilexn/openst/blob/main/models/HE_cellpose_rajewsky).
+[here](http://bimsbstatic.mdc-berlin.de/rajewsky/openst-public-data/models/HE_cellpose_rajewsky).
 You can specify any other model that works best for your data -
 refer to the [cellpose](https://cellpose.readthedocs.io/en/latest/index.html) documentation.
 
@@ -31,7 +31,7 @@ By default, segmentation is extended radially 10 pixels. This can be changed wit
 Make sure to populate the arguments with the values specific to your dataset. Here, we provide `--h5-in` consistent
 with the previous steps, `--image-in` and `--mask-out` will read and write the staining and mask inside the Open-ST h5 object,
 and `--model` is `HE_cellpose_rajewsky`, the default used in our manuscript. This is the model we recommend for H&E images, and
-weights are automatically downloaded. It is also [provided in our repo](https://github.com/rajewsky-lab/openst/blob/main/models/HE_cellpose_rajewsky).
+weights are automatically downloaded. It is also [provided in our repo](http://bimsbstatic.mdc-berlin.de/rajewsky/openst-public-data/models/HE_cellpose_rajewsky).
 The rest of parameters can be checked with `openst segment --help`.
 
 !!! tip

docs/computational/preprocessing_imaging.md

@@ -63,16 +63,33 @@ tile-scan. You can run this by running the following command on the stitched ima
 
 ```bash
 openst image_preprocess \
-     --input=<path_to_input_image> \
-     --CUT \
-     --CUT-model=<path_to_model> \
-     --output=<path_to_output>
+     --image-in Image_Stitched_Composite.tif \
+     --image-out Image_Stitched_Composite_Restored.tif
+     # --device cuda # in case you have a CUDA-compatible GPU
 ```
 
-Make sure to replace the placeholders (`<...>`). For instance,
-`<path_to_input_image>` is the full path and file name of the previously stitched image; `<path_to_model>`
-is filename our pre-trained [CUT model](https://github.com/rajewsky-lab/openst/models/CUT.pth), and `<output_image>` 
-is the path to a folder (writeable) and desired filename for the output image.
+If you ran `openst merge_modalities`, then imaging data will be contained inside the Open-ST h5 object, and the command
+can be adapted:
+
+```bash
+openst image_preprocess \
+     --h5-in multimodal/spots_stitched.h5ad # just a placeholder, adapt
+     # --device cuda # in case you have a CUDA-compatible GPU
+```
+
+By default, the image will be loaded from the key `uns/spatial/staining_image`, and the CUT-restored image will be saved
+to `uns/spatial/staining_image_restored`. You can preview the image restoration results using:
+
+```bash
+openst preview \
+     --h5-in multimodal/spots_stitched.h5ad
+     --image-key uns/spatial/staining_image uns/spatial/staining_image_restored
+```
+
+This will load the two images and visualize it using `napari`. Later, you can run segmentation and pairwise alignment
+using either the default merged image (`uns/spatial/staining_image`), or the restored image (`uns/spatial/staining_image_restored`).
+Always assess these preprocessing choices (quantitatively and qualitatively) to decide whether these make sense for your data.
+
 
 ## Expected output
 After running the stitching (and optionally correction algorithms), you will have a single image file per sample. This, together with

docs/examples/adult_mouse/generate_expression_matrix.md

@@ -12,7 +12,7 @@ into a cell-by-genes matrix, where cells are defined from the segmentation mask.
 To create such a spatial cell-by-gene ($M\times G$) expression matrix, you will first need a segmentation mask.
 
 We efficiently segment cells (or nuclei) from staining images using [cellpose](https://github.com/MouseLand/cellpose).
-For the H&E-stained tissue provided in this example, we used our [fine-tuned model](https://github.com/rajewsky-lab/openst/blob/main/models/HE_cellpose_rajewsky).
+For the H&E-stained tissue provided in this example, we used our [fine-tuned model](http://bimsbstatic.mdc-berlin.de/rajewsky/openst-public-data/models/HE_cellpose_rajewsky).
 Make sure to download it and save it into a new `models` folder that you need to create under the `openst_adult_demo` main folder.
 
 You can run the segmentation on the previously created `openst_demo_adult_mouse_spatial_beads_puck_collection_aligned.h5ad` file, which

docs/examples/e13_mouse/generate_expression_matrix.md

@@ -12,7 +12,7 @@ into a cell-by-genes matrix, where cells are defined from the segmentation mask.
 To create such a spatial cell-by-gene ($M\times G$) expression matrix, you will first need a segmentation mask.
 
 We efficiently segment cells (or nuclei) from staining images using [cellpose](https://github.com/MouseLand/cellpose).
-For the H&E-stained tissue provided in this example, we used our [fine-tuned model](https://github.com/rajewsky-lab/openst/blob/main/models/HE_cellpose_rajewsky).
+For the H&E-stained tissue provided in this example, we used our [fine-tuned model](http://bimsbstatic.mdc-berlin.de/rajewsky/openst-public-data/models/HE_cellpose_rajewsky).
 Make sure to download it and save it into a new `models` folder that you need to create under the `openst_e13_demo` main folder.
 
 You can run the segmentation on the previously created `openst_demo_e13_mouse_head_spatial_beads_puck_collection_aligned.h5ad` file, which

openst/__init__.py

@@ -1 +1 @@
-__version__ = '0.0.2'
+__version__ = '0.0.11'

openst/cli.py

@@ -680,17 +680,46 @@ def get_image_preprocess_parser():
         allow_abbrev=False,
         add_help=False,
     )
-    parser.add_argument("--input_img", type=str, required=True, help="path to input image")
-    parser.add_argument("--cut_dir", type=str, required=True, help="path to CUT directory (to save patched images)")
-    parser.add_argument("--tile_size_px", type=int, required=True, help="size of the tile in pixels")
+    parser.add_argument(
+        "--h5-in",
+        type=str,
+        default="",
+        help="""If set, image is loaded from the Open-ST h5 object (key in --image-in),
+             and retored image is saved there (to the key --image-out)""",
+    )
+    parser.add_argument(
+        "--image-in",
+        type=str,
+        default="uns/spatial/staining_image",
+        help="Key or path to the input image",
+    )
+    parser.add_argument(
+        "--image-out",
+        type=str,
+        default="uns/spatial/staining_image_restored",
+        help="Key or path where the restored image will be written into",
+    )
+    parser.add_argument(
+        "--tile-size-px",
+        type=int,
+        default=512,
+        help="The input image is split into squared tiles of side `--tile-size-px`, for inference."+ 
+        "Larger values avoid boundary effects, but require more memory.",
+    )
+    parser.add_argument("--model", type=str, default="HE_CUT_rajewsky", help="CUT model used for image restoration")
     parser.add_argument(
         "--device",
         type=str,
         default="cpu",
         choices=["cpu", "cuda"],
-        help="Device used to run feature matching model. Can be ['cpu', 'cuda']",
+        help="Device used to run CUT restoration model. Can be ['cpu', 'cuda']",
+    )
+    parser.add_argument(
+        "--num-workers",
+        type=int,
+        default=-1,
+        help="Number of CPU workers for parallel processing",
     )
-    parser.add_argument("--checkpoints_dir", type=str, default="./checkpoints", help="models are saved here")
 
     return parser
 

openst/preprocessing/CUT/models/__init__.py

@@ -19,8 +19,9 @@
 """
 
 import importlib
-from models.base_model import BaseModel
+import logging
 
+from openst.preprocessing.CUT.models.base_model import BaseModel
 
 def find_model_using_name(model_name):
     """Import the module "models/[model_name]_model.py".
@@ -29,7 +30,7 @@ def find_model_using_name(model_name):
     be instantiated. It has to be a subclass of BaseModel,
     and it is case-insensitive.
     """
-    model_filename = "models." + model_name + "_model"
+    model_filename = "openst.preprocessing.CUT.models." + model_name + "_model"
     modellib = importlib.import_module(model_filename)
     model = None
     target_model_name = model_name.replace('_', '') + 'model'
@@ -63,5 +64,5 @@ def create_model(opt):
     """
     model = find_model_using_name(opt.model)
     instance = model(opt)
-    print("model [%s] was created" % type(instance).__name__)
+    logging.info("Model architecture `%s` was created" % type(instance).__name__)
     return instance

openst/preprocessing/CUT/models/base_model.py

@@ -1,8 +1,10 @@
 import os
 import torch
+import logging
 from collections import OrderedDict
 from abc import ABC, abstractmethod
-from . import networks
+
+from openst.preprocessing.CUT.models import networks
 
 
 class BaseModel(ABC):
@@ -211,7 +213,7 @@ def load_networks(self, epoch):
                 net = getattr(self, 'net' + name)
                 if isinstance(net, torch.nn.DataParallel):
                     net = net.module
-                print('loading the model from %s' % load_path)
+                logging.info(f'Loading model weights from {load_path}')
                 # if you are using PyTorch newer than 0.4 (e.g., built from
                 # GitHub source), you can remove str() on self.device
                 state_dict = torch.load(load_path, map_location=str(self.device))
@@ -229,17 +231,18 @@ def print_networks(self, verbose):
         Parameters:
             verbose (bool) -- if verbose: print the network architecture
         """
-        print('---------- Networks initialized -------------')
+        message = '---------- (Start) Networks initialized -------------\n'
         for name in self.model_names:
             if isinstance(name, str):
                 net = getattr(self, 'net' + name)
                 num_params = 0
                 for param in net.parameters():
                     num_params += param.numel()
                 if verbose:
-                    print(net)
-                print('[Network %s] Total number of parameters : %.3f M' % (name, num_params / 1e6))
-        print('-----------------------------------------------')
+                    message += f"{net}\n"
+                message += '[Network %s] Total number of parameters : %.3f M\n' % (name, num_params / 1e6)
+        message += '---------- (End) Networks initialized -------------'
+        logging.debug(message)
 
     def set_requires_grad(self, nets, requires_grad=False):
         """Set requies_grad=Fasle for all the networks to avoid unnecessary computations

openst/preprocessing/CUT/models/cut_model.py

@@ -1,10 +1,10 @@
 import numpy as np
 import torch
-from .base_model import BaseModel
-from . import networks
-from .patchnce import PatchNCELoss
-import util.util as util
 
+from openst.preprocessing.CUT.models.base_model import BaseModel
+from openst.preprocessing.CUT.models import networks
+from openst.preprocessing.CUT.models.patchnce import PatchNCELoss
+from openst.preprocessing.CUT.util import util
 
 class CUTModel(BaseModel):
     """ This class implements CUT and FastCUT model, described in the paper
@@ -98,10 +98,10 @@ def data_dependent_initialize(self, data):
         initialized at the first feedforward pass with some input images.
         Please also see PatchSampleF.create_mlp(), which is called at the first forward() call.
         """
-        bs_per_gpu = data["A"].size(0) // max(len(self.opt.gpu_ids), 1)
+        bs_per_gpu = self.opt.batch_size
         self.set_input(data)
         self.real_A = self.real_A[:bs_per_gpu]
-        self.real_B = self.real_B[:bs_per_gpu]
+        self.real_B = None
         self.forward()                     # compute fake images: G(A)
         if self.opt.isTrain:
             self.compute_D_loss().backward()                  # calculate gradients for D
@@ -138,14 +138,11 @@ def set_input(self, input):
             input (dict): include the data itself and its metadata information.
         The option 'direction' can be used to swap domain A and domain B.
         """
-        AtoB = self.opt.direction == 'AtoB'
-        self.real_A = input['A' if AtoB else 'B'].to(self.device)
-        self.real_B = input['B' if AtoB else 'A'].to(self.device)
-        self.image_paths = input['A_paths' if AtoB else 'B_paths']
+        self.real_A = input['A'].to(self.device)
 
     def forward(self):
         """Run forward pass; called by both functions <optimize_parameters> and <test>."""
-        self.real = torch.cat((self.real_A, self.real_B), dim=0) if self.opt.nce_idt and self.opt.isTrain else self.real_A
+        self.real = self.real_A
         if self.opt.flip_equivariance:
             self.flipped_for_equivariance = self.opt.isTrain and (np.random.random() < 0.5)
             if self.flipped_for_equivariance:

openst/preprocessing/CUT/models/networks.py

@@ -5,12 +5,6 @@
 import functools
 from torch.optim import lr_scheduler
 import numpy as np
-from .stylegan_networks import StyleGAN2Discriminator, StyleGAN2Generator, TileStyleGAN2Discriminator
-
-###############################################################################
-# Helper Functions
-###############################################################################
-
 
 def get_filter(filt_size=3):
     if(filt_size == 1):
@@ -256,10 +250,6 @@ def define_G(input_nc, output_nc, ngf, netG, norm='batch', use_dropout=False, in
         net = UnetGenerator(input_nc, output_nc, 7, ngf, norm_layer=norm_layer, use_dropout=use_dropout)
     elif netG == 'unet_256':
         net = UnetGenerator(input_nc, output_nc, 8, ngf, norm_layer=norm_layer, use_dropout=use_dropout)
-    elif netG == 'stylegan2':
-        net = StyleGAN2Generator(input_nc, output_nc, ngf, use_dropout=use_dropout, opt=opt)
-    elif netG == 'smallstylegan2':
-        net = StyleGAN2Generator(input_nc, output_nc, ngf, use_dropout=use_dropout, n_blocks=2, opt=opt)
     elif netG == 'resnet_cat':
         n_blocks = 8
         net = G_Resnet(input_nc, output_nc, opt.nz, num_downs=2, n_res=n_blocks - 4, ngf=ngf, norm='inst', nl_layer='relu')
@@ -323,8 +313,6 @@ def define_D(input_nc, ndf, netD, n_layers_D=3, norm='batch', init_type='normal'
         net = NLayerDiscriminator(input_nc, ndf, n_layers_D, norm_layer=norm_layer, no_antialias=no_antialias,)
     elif netD == 'pixel':     # classify if each pixel is real or fake
         net = PixelDiscriminator(input_nc, ndf, norm_layer=norm_layer)
-    elif 'stylegan2' in netD:
-        net = StyleGAN2Discriminator(input_nc, ndf, n_layers_D, no_antialias=no_antialias, opt=opt)
     else:
         raise NotImplementedError('Discriminator model name [%s] is not recognized' % netD)
     return init_net(net, init_type, init_gain, gpu_ids,
@@ -1214,8 +1202,6 @@ def forward(self, input):
 
 class UnetSkipConnectionBlock(nn.Module):
     """Defines the Unet submodule with skip connection.
-        X -------------------identity----------------------
-        |-- downsampling -- |submodule| -- upsampling --|
     """
 
     def __init__(self, outer_nc, inner_nc, input_nc=None,

openst/preprocessing/CUT/models/patchnce.py

@@ -0,0 +1,54 @@
+from packaging import version
+import torch
+from torch import nn
+
+class PatchNCELoss(nn.Module):
+    def __init__(self, opt):
+        super().__init__()
+        self.opt = opt
+        self.cross_entropy_loss = torch.nn.CrossEntropyLoss(reduction='none')
+        self.mask_dtype = torch.uint8 if version.parse(torch.__version__) < version.parse('1.2.0') else torch.bool
+
+    def forward(self, feat_q, feat_k):
+        num_patches = feat_q.shape[0]
+        dim = feat_q.shape[1]
+        feat_k = feat_k.detach()
+
+        # pos logit
+        l_pos = torch.bmm(
+            feat_q.view(num_patches, 1, -1), feat_k.view(num_patches, -1, 1))
+        l_pos = l_pos.view(num_patches, 1)
+
+        # neg logit
+
+        # Should the negatives from the other samples of a minibatch be utilized?
+        # In CUT and FastCUT, we found that it's best to only include negatives
+        # from the same image. Therefore, we set
+        # --nce_includes_all_negatives_from_minibatch as False
+        # However, for single-image translation, the minibatch consists of
+        # crops from the "same" high-resolution image.
+        # Therefore, we will include the negatives from the entire minibatch.
+        if self.opt.nce_includes_all_negatives_from_minibatch:
+            # reshape features as if they are all negatives of minibatch of size 1.
+            batch_dim_for_bmm = 1
+        else:
+            batch_dim_for_bmm = self.opt.batch_size
+
+        # reshape features to batch size
+        feat_q = feat_q.view(batch_dim_for_bmm, -1, dim)
+        feat_k = feat_k.view(batch_dim_for_bmm, -1, dim)
+        npatches = feat_q.size(1)
+        l_neg_curbatch = torch.bmm(feat_q, feat_k.transpose(2, 1))
+
+        # diagonal entries are similarity between same features, and hence meaningless.
+        # just fill the diagonal with very small number, which is exp(-10) and almost zero
+        diagonal = torch.eye(npatches, device=feat_q.device, dtype=self.mask_dtype)[None, :, :]
+        l_neg_curbatch.masked_fill_(diagonal, -10.0)
+        l_neg = l_neg_curbatch.view(-1, npatches)
+
+        out = torch.cat((l_pos, l_neg), dim=1) / self.opt.nce_T
+
+        loss = self.cross_entropy_loss(out, torch.zeros(out.size(0), dtype=torch.long,
+                                                        device=feat_q.device))
+
+        return loss

openst/preprocessing/CUT/models/template_model.py

@@ -9,15 +9,16 @@
 Given input-output pairs (data_A, data_B), it learns a network netG that can minimize the following L1 loss:
     min_<netG> ||netG(data_A) - data_B||_1
 You need to implement the following functions:
-    <modify_commandline_options>:　Add model-specific options and rewrite default values for existing options.
+    <modify_commandline_options>: Add model-specific options and rewrite default values for existing options.
     <__init__>: Initialize this model class.
     <set_input>: Unpack input data and perform data pre-processing.
     <forward>: Run forward pass. This will be called by both <optimize_parameters> and <test>.
     <optimize_parameters>: Update network weights; it will be called in every training iteration.
 """
 import torch
-from .base_model import BaseModel
-from . import networks
+
+from openst.preprocessing.CUT.models.base_model import BaseModel
+from openst.preprocessing.CUT.models import networks
 
 
 class TemplateModel(BaseModel):