Centralized all deep learning segmentation tasks with new function "sct_deepseg" and refactored sct_download_data #2639

jcohenadad · 2020-03-19T18:39:51Z

With the need to more custom deep learning segmentation models and architectures, as well as new tasks (e.g. multi-class segmentation, CSF segmentation, etc.), it becomes more convenient to have a single CLI function, called sct_deepseg, that calls a specific model.

One notable feature of this new function is that it will rely on the actively-maintained ivadomed package. Another notable change is that we will now rely on the Pytorch framework (vs. Tensorflow).

For now, this new function will not replace sct_deepseg_X, but will be complementary to the existing functions, so we can gather feedback from users before making these functions deprecated. The "big" change will likely be associated with a major release (e.g. 5.0.0).

Changes include:

Introduced new function sct_deepseg that can point to a specific model. Fixes Centralize deep-learning segmentation tools into a single entry point and module #2626
Centralize hard-coded access to path for models and data. Fixes Centralize hard-coded access to path for models and data #2160
Deep-learning models are installed under $SCT_DIR/models
- New SCT variable: sct.__models_dir__
Added useful functions under sct.utils: add_suffix(), extract_fname(), tmp_create()
Implemented feature to set threshold for output segmentation
Installing default models during SCT installation by adding: sct_deepseg -download-default-models
Added ivadomed/master in requirements.txt
refactored sct_download_data and move some functions under spinalcordtoolbox.download

Todo in subsequent PRs:

Add a few default models in OSF with the appropriate json fields: threshold: Populate deepseg models and fill metadata fields #2695
Document how to create new models: Documentation for creating new models #2696
Add testing for online models: Add testing for sct_deepseg #2714

Test this branch on a development ivadomed version

# Checkout
cd $SCT_DIR
git fetch origin
git checkout -b jca/2626-deepseg origin/jca/2626-deepseg
# Activate SCT's venv
source ${SCT_DIR}/python/etc/profile.d/conda.sh
conda activate venv_sct
# Create alias to new sct_deepseg executable
pip install -e .
cp python/envs/venv_sct/bin/sct_deepseg bin/
# Clone ivadomed repos and pip install it inside SCT venv
cd ~  # or wherever you'd like to download that
git clone https://github.com/neuropoly/ivado-medical-imaging.git
cd ivado-medical-imaging
pip install -e .

kousu · 2020-03-31T20:01:49Z

scripts/sct_deepseg.py

+# This command-line tool is the interface for the deepseg API that performs segmentation using deep learning from the
+# ivadomed package.


You can write

Suggested change

# This command-line tool is the interface for the deepseg API that performs segmentation using deep learning from the

# ivadomed package.

"""

This command-line tool is the interface for the deepseg API that performs segmentation using deep learning from the

ivadomed package.

"""

To get pretty module-level docstrings. (python treats leading comments as module docstrings as well, but retains the #s, so I think using a triple-string is more idiomatic.)

indeed-- fixed here 021f205

scripts/sct_deepseg.py

jcohenadad · 2020-03-31T21:50:09Z

spinalcordtoolbox/deepseg/core.py

+    :param param_deepseg: class ParamDeepseg: Segmentation parameters.
+    :return: fname_out: str: Output filename.
+    """
+    nii_seg = segment_volume(folder_model, fname_image)


note: import ivadomed as imed

charleygros · 2020-04-02T03:27:12Z

spinalcordtoolbox/deepseg/models.py

+
+
+MODELS = {
+    'cord-t2star': {'url': 'https://osf.io/8nsa6/',


The other day we said tha the version is "encoded" in the url --> is it correct? (just checking if we do not need a version or date key)

fixed in c03c647

charleygros · 2020-04-02T03:29:38Z

spinalcordtoolbox/deepseg/models.py

+        """
+        Check if model (.pt file) is installed under SCT directory
+        """
+        if os.path.exists(os.path.join(self.folder, self.name + '.pt')):


Should we test if the .json is also present? or could we reasonably assume that?

i was thinking about that. My reasoning here was that the json file is not "mandatory" to execute a segmentation. Only the model is, which is why I didn't check for it. Open for discussion.

My reasoning here was that the json file is not "mandatory" to execute a segmentation.

ah! actually i am wrong. It is mandatory! See: https://github.com/neuropoly/ivado-medical-imaging/blob/master/ivadomed/utils.py#L624.

Oh well, I will check for it then ;-)

Done in b469daa

spinalcordtoolbox/deepseg/models.py

kousu · 2020-04-07T21:18:02Z

spinalcordtoolbox/deepseg/models.py

+    'uqueensland_mice_sc':
+        {'url': 'https://osf.io/nu3ma/download?version=2',
+         'description': 'Cord segmentation on mouse MRI. Data from University of Queensland.'},
+    'uqueensland_mice_gm':


According to the convention this should be

Suggested change

'uqueensland_mice_gm':

'mice_uqueensland_gm':

fixed in 57b99e3

charleygros · 2020-04-13T03:08:16Z

# Clone ivadomed repos and pip install it inside SCT venv
cd ~  # or wherever you'd like to download that
git pull https://github.com/neuropoly/ivado-medical-imaging.git
cd ivado-medical-imaging
pip install -e .

Minor comment:
For tests run by external users --> we should recommend to do a git clone instead of a git pull

charleygros · 2020-04-13T04:12:04Z

I think the field threshold is missing from ../spinalcordtoolbox/models/t2star_sc/t2star_sc.json.

charleygros · 2020-04-13T04:23:45Z

@jcohenadad : Please find herebelow the terminal output I got.

I followed the installation process you described in the first post of this PR.

manually added a threshold to the model metadata.
pull last version of the ivadomed PR (last commit, related to a typo in a nibabel function).

(venv_sct) chgroc@rosenberg:~/sct_ivado$ sct_deepseg -i sct_testing_data/t2s/t2s.nii.gz -m t2star_sc

--
Spinal Cord Toolbox (git-jca/2626-deepseg-53f3f5f03c9931f02e5f5c84b0a4d5cf0ba4e431)


Loaded 9 axial slices..
/home/GRAMES.POLYMTL.CA/chgroc/spinalcordtoolbox/python/envs/venv_sct/lib/python3.6/site-packages/torch/serialization.py:453: SourceChangeWarning: source code of class 'ivadomed.models.Decoder' has changed. you can retrieve the original source code by accessing the object's source attribute or set `torch.nn.Module.dump_patches = True` and use the patch tool to revert the changes.
  warnings.warn(msg, SourceChangeWarning)
/home/GRAMES.POLYMTL.CA/chgroc/spinalcordtoolbox/python/envs/venv_sct/lib/python3.6/site-packages/torch/serialization.py:453: SourceChangeWarning: source code of class 'ivadomed.models.UpConv' has changed. you can retrieve the original source code by accessing the object's source attribute or set `torch.nn.Module.dump_patches = True` and use the patch tool to revert the changes.
  warnings.warn(msg, SourceChangeWarning)
Traceback (most recent call last):
  File "/home/GRAMES.POLYMTL.CA/chgroc/spinalcordtoolbox/scripts/sct_deepseg.py", line 141, in <module>
    main()
  File "/home/GRAMES.POLYMTL.CA/chgroc/spinalcordtoolbox/scripts/sct_deepseg.py", line 136, in main
    sct.deepseg.core.segment_nifti(args.i, path_model, param)
  File "/home/GRAMES.POLYMTL.CA/chgroc/spinalcordtoolbox/spinalcordtoolbox/deepseg/core.py", line 50, in segment_nifti
    fname_out = sct.utils.add_suffix(fname_image, '_seg')
  File "/home/GRAMES.POLYMTL.CA/chgroc/spinalcordtoolbox/spinalcordtoolbox/utils.py", line 119, in add_suffix
    parent, stem, ext = extract_fname(fname)
NameError: name 'extract_fname' is not defined

jcohenadad · 2020-04-13T11:57:36Z

I think the field threshold is missing from ../spinalcordtoolbox/models/t2star_sc/t2star_sc.json

indeed-- i added it manually-- we will need to add it to all models (along with the other postproc relevant fields)

jcohenadad · 2020-04-14T20:02:09Z

spinalcordtoolbox/deepseg/core.py

+
+    # Postprocessing
+    metadata = sct.deepseg.models.get_metadata(folder_model)
+    postproc = PostProcessing(param, metadata)


@kousu is that an overkill to create a Class for this or should we simply define each postprocessing function with nii, param and metadata as input parameter?

I think a class doesn't offer anything here. A class is useful if you need to keep some state around for a while, but here you want to use the data immediately and get out.

I think what I would do is write it as

def postprocess(nii_seg, param, metadata): options = {**DEFAULTS, **metadata, **{k: v for k,v in param.__dict__.items() if v is not None}} if options['threshold']: nii_seg = threshold(nii_sig, options['threshold']) if options['keep_largest_object']: nii_seg = keep_largest_object(nii_sig) if options['fill_holes']: nii_seg = fill_holes(nii_sig) return nii_seg

or, if you take my suggestion about dropping ParamDeepseg, then

def postprocess(nii_seg, param, metadata): options = {**DEFAULTS, **metadata, **param} if options['threshold']: nii_seg = threshold(nii_sig, options['threshold']) if options['keep_largest_object']: nii_seg = keep_largest_object(nii_sig) if options['fill_holes']: nii_seg = fill_holes(nii_sig) return nii_seg

the "merge" approach is brilliant. Currently trying to implement it.

implemented in bea7a3c

spinalcordtoolbox/deepseg/core.py

kousu · 2020-04-16T05:23:30Z

spinalcordtoolbox/deepseg/core.py

+                nii_seg = imed.postprocessing.fill_holes(nii_seg)
+            else:
+                logger.warning("Algorithm 'fill holes' can only be run on binary segmentation. Skipping.")
+            return nii_seg


This return is tabbed in one level too deep.

fixed in bea7a3c

kousu · 2020-04-16T05:24:50Z

spinalcordtoolbox/deepseg/core.py

+                        "Neither I nor S is present in code: {}, for affine matrix: {}".format(code, affine))
+                nii_seg = imed.postprocessing.keep_largest_object_per_slice(nii_seg, axis=axis_infsup)
+            else:
+                logger.warning("Algorithm 'keep largest object' can only be run on binary segmentation. Skipping.")


Could this warning be an exception? "Errors should never pass silently. Unless explicitly silenced." advocates the zen of python.

well, technically it is not an error but just a warning that this step will be skipped. If we stop the program here, i anticipate that all users who try the flag "-thr 0" will see a crash, and since most users don't read their crash report, they will wrongly think that "-thr 0" doesn't work... so i'd rather give them an output and informing the small percentage of users who actually read reports that some algo have not run because of -thr 0.
@charleygros do you have any opinion?

I would agree that this is more a warning than an error, and would then agree to skip it.
Another approach would be to call the thresholding with a default value (eg 10e-3) in case the object is not binary (ie thr=0) when entering in keep_largest_object and then use the mask_prediction function to remove the small objects from the soft pred based on the binary output of keep_largest.

Another approach would be to call the thresholding with a default value (eg 10e-3) in case the object is not binary (ie thr=0) when entering in keep_largest_object and then use the mask_prediction function to remove the small objects from the soft pred based on the binary output of keep_largest.

yup i like this idea!

This should be implemented in ivadomed though, issue opened: ivadomed/ivadomed#205

kousu · 2020-04-16T06:25:58Z

spinalcordtoolbox/deepseg/core.py

+        # TODO: This if/elif below is ugly. Cannot think of something better for now...
+        do_process = DEFAULT_KEEP_LARGEST_OBJECT
+        if self.param.keep_largest_object is True:
+            do_process = True
+        elif self.param.keep_largest_object is None:
+            if 'keep_largest_object' in self.metadata:
+                do_process = self.metadata.keep_largest_object
+        if do_process:


I thought about what possible values everything could take here. It looks like you want self.param taking precedence, then self.metadata, then falling back to DEFAULT_, with None or not in ... meaning "undefined, so try the next option".

I would say you want do_process = param or metadata or default, except that will misinterpret False as "continue to fallback" instead of "stop here".

I rewrote this expression as a function I could test exhaustively with all the possible values; I made one simplification that I don't think would change the meaning -- to represent 'keep_largest_object' not in self.metadata as metadata is None.

from pprint import pprint def do_process(param, metadata, default): if param is True: return True elif param is None: if metadata is not None: return metadata return default pprint( [(param, metadata, default, "=>", do_process(param, metadata, default)) for param in [True, False, None] for metadata in [True, False, None] for default in [True, False]])

this gave me this truth table:

[(True, True, True, '=>', True), (True, True, False, '=>', True), (True, False, True, '=>', True), (True, False, False, '=>', True), (True, None, True, '=>', True), (True, None, False, '=>', True), (False, True, True, '=>', True), (False, True, False, '=>', False), (False, False, True, '=>', True), (False, False, False, '=>', False), (False, None, True, '=>', True), (False, None, False, '=>', False), (None, True, True, '=>', True), (None, True, False, '=>', True), (None, False, True, '=>', False), (None, False, False, '=>', False), (None, None, True, '=>', True), (None, None, False, '=>', False)]

In all cases where param is True, it's True (of course, that's written right into the logic). But when it's False it skips metadata and uses default instead.

Is that what you really meant? param=False and param=None both mean "continue", but metadata=False means "stop" yet metadata=None means "continue"? I think that's weird. I think the actual logic you want is

def do_process(param, metadata, default): if param is True: return True elif param is False: return False elif param is None: if metadata is not None: return metadata return default

which can be written shorter as

def do_process(param, metadata, default): if param is not None: return param if metadata is not None: return metadata return default

The truth table for this is

[(True, True, True, '=>', True), (True, True, False, '=>', True), (True, False, True, '=>', True), (True, False, False, '=>', True), (True, None, True, '=>', True), (True, None, False, '=>', True), (False, True, True, '=>', False), (False, True, False, '=>', False), (False, False, True, '=>', False), (False, False, False, '=>', False), (False, None, True, '=>', False), (False, None, False, '=>', False), (None, True, True, '=>', True), (None, True, False, '=>', True), (None, False, True, '=>', False), (None, False, False, '=>', False), (None, None, True, '=>', True), (None, None, False, '=>', False)]

The other tricky thing I see is that 'undefined' is represented in two different ways: for self.param, it's represented by None, for self.metadata it's represented by an actual missing value. What should happen if someone set self.metadata['keep_largest_object'] = None? Or if someone did del self.param.keep_largest_object?

If you could change the representation this would be easier: delete ParamDeepseg and just tell users to pass a dictionary -- you could even write def __init__(self, metadata, **params). Without that, we're going to have to accept some baseline ugliness.

Anyway, if you accept my changing of your logic above and the need to deal with different representations, you can rewrite this with the ternary hook:

Suggested change

# TODO: This if/elif below is ugly. Cannot think of something better for now...

do_process = DEFAULT_KEEP_LARGEST_OBJECT

if self.param.keep_largest_object is True:

do_process = True

elif self.param.keep_largest_object is None:

if 'keep_largest_object' in self.metadata:

do_process = self.metadata.keep_largest_object

if do_process:

if (self.param.keep_largest_object

if self.param.keep_largest_object is not None

else self.metadata['keep_largest_object']

if 'keep_largest_object' in self.metadata

else DEFAULT_KEEP_LARGEST_OBJECT):

Python has a shortcut for the if .. in .. else part, dict.get(...[, default]):

Suggested change

# TODO: This if/elif below is ugly. Cannot think of something better for now...

do_process = DEFAULT_KEEP_LARGEST_OBJECT

if self.param.keep_largest_object is True:

do_process = True

elif self.param.keep_largest_object is None:

if 'keep_largest_object' in self.metadata:

do_process = self.metadata.keep_largest_object

if do_process:

if (self.param.keep_largest_object

if self.param.keep_largest_object is not None

else self.metadata.get('keep_largest_object',

DEFAULT_KEEP_LARGEST_OBJECT)):

You could also map None => missing directly, and then make use of dict.get():

Suggested change

# TODO: This if/elif below is ugly. Cannot think of something better for now...

do_process = DEFAULT_KEEP_LARGEST_OBJECT

if self.param.keep_largest_object is True:

do_process = True

elif self.param.keep_largest_object is None:

if 'keep_largest_object' in self.metadata:

do_process = self.metadata.keep_largest_object

if do_process:

if ({k: v for k,v in self.param.__dict__.items() if v is not None}).get('keep_largest_object', self.metadata.get('keep_largest_object', DEFAULT_KEEP_LARGEST_OBJECT)):

Or what about viewing it explicitly as a merge? Apparently in Python3.9 you will be able to write defaults | metadata | param, but in the meantime we have this {**defaults, **metadata, **param} syntax:

DEFAULTS = { 'threshold': 0.9, 'keep_largest_object': True, 'fill_holes': True, } class PostProcessing: """ Deals with post-processing of the segmentation. Consider param (i.e. user's flag) with more priority than metadata (i.e. from model's json file). """ def __init__(self, param, metadata): """ :param param: dict: Defined by user's parameter :param metadata: dict: From model's json metadata """ # param overrides metadata overrides defaults self.param = {**DEFAULTS, **metadata, **param} [...] def keep_largest_object(self, nii_seg): """ Only keep largest object """ if self.param['keep_largest_object']: [...]

merged strategy implemented in bea7a3c

kousu · 2020-04-16T07:40:21Z

spinalcordtoolbox/deepseg/core.py

+
+    def threshold(self, nii_seg):
+        """
+        Threshold the prediction. For no prediction, set 'threshold' to 0.


Is this a typo?

Suggested change

Threshold the prediction. For no prediction, set 'threshold' to 0.

Threshold the prediction. For no threshold, set 'threshold' to 0.

Yes-- Fixed in ce8a184

kousu · 2020-04-16T07:42:18Z

spinalcordtoolbox/deepseg/core.py

+        Threshold the prediction. For no prediction, set 'threshold' to 0.
+        """
+        if self.param.threshold:


You can use threshold = None to mean "no threshold" instead of letting any falsey value -- namely 0.0 -- trigger this. That would be more uniform with the other inputs.

Editorializing: It's a remnant of C's limited type system that we have this instinct to repurpose the meanings of specific values.

no more relevant with current implementation (ce8a184)

kousu · 2020-04-16T17:52:32Z

spinalcordtoolbox/download.py

+    # if urls is not a list, make it one
+    if not isinstance(urls, (list, tuple)):
+        urls = [urls]


This should be a bit more reliable:

Suggested change

# if urls is not a list, make it one

if not isinstance(urls, (list, tuple)):

urls = [urls]

# if urls is not a list, make it one

if isinstance(urls, str):

urls = [urls]

indeed-- fixed in 88a04b5

kousu · 2020-04-16T22:45:28Z

scripts/sct_deepseg.py

+    args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])
+    # TODO: instead of assigning each args param, we could pass args while instanciating ParamDeepseg(args), and the
+    #  class would deal with assigning arguments to each field.


Instead of using ParamDeepseg, you could do something like:

Suggested change

args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])

# TODO: instead of assigning each args param, we could pass args while instanciating ParamDeepseg(args), and the

# class would deal with assigning arguments to each field.

args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])

args ={k: v for k in vars(args).items() if v is not None}

# separate out the segmentation `param` args from the top level args

input = args.pop('i')

list_models = args.pop('list_models')

install_model = args.pop('install_model')

install_default_models = args.pop('install_default_models')

...

You'd have to remove all the default values/make sure they're set to None; default=param_default.keep_largest_object is an extra layer of defaults which might accidentally override a model's metadata even when the the user doesn't ask for it.

actually it fails because v is not defined

You have to watch out because the argparse names don't match the postprocess() names. You have args.thr vs param.threshold and args.keep_largest vs param.keep_largest_object. You can use dest to fixup the naming, or just change the names of the flags in the CLI.

You could be explicit about which subset are destined for algorithm parameters instead of being using .pop() after the fact to pull off the other args:

params = ['keep_largest_object', 'fill_holes', 'threshold'] params = {k: vars(args)[k] for k in params if vars(args)[k] is not None}

Oh sorry!! Let me try that again:

Suggested change

args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])

# TODO: instead of assigning each args param, we could pass args while instanciating ParamDeepseg(args), and the

# class would deal with assigning arguments to each field.

args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])

args ={k: v for k, v in vars(args).items() if v is not None}

# separate out the segmentation `param` args from the top level args

input = args.pop('i')

list_models = args.pop('list_models')

install_model = args.pop('install_model')

install_default_models = args.pop('install_default_models')

...

@kousu how about not "popping" fields from args, and just passing args inside deepseg.core(args), because the two first arguments (fname_input and path_model) are already included in args.

Are you worried that we would be passing a bunch of irrelevant fields inside deepseg.core?

jcohenadad · 2020-04-18T19:59:14Z

Travis is crashes for two reasons:

kousu · 2020-04-18T21:31:14Z

An important subtlety: disables pip's progress bar on Travis. This is a small change but it does mean Travis logs are not quite comparable after this point.

kousu · 2020-04-18T21:44:45Z

The build is failing at sct_check_dependencies:

Check if ivadomed@git+https://github.com/neuropoly/ivado-medical-imaging@master is installed[FAIL]

No module named 'ivadomed@git+https://github'

This is because

https://github.com/neuropoly/spinalcordtoolbox/blob/72bc861f02e2075b12a2e4fd04d6767e2086f3d8/scripts/sct_check_dependencies.py#L171-L178

does not parse pip VCS URLs.

My best suggestion is to remove sct_check_dependencies entirely and instead rely on set -e + pip install -r requirements.txt. Then even if there somehow is a missing dependency it will fail at runtime but I don't think that's a big loss, since the error will be obvious. Trying to import all the dependencies is all sct_check_dependencies is doing anyway, just up front:

https://github.com/neuropoly/spinalcordtoolbox/blob/72bc861f02e2075b12a2e4fd04d6767e2086f3d8/scripts/sct_check_dependencies.py#L290-L293

My second best suggestion is to edit that code to parse VCS URLs; but that seems tricky because the docs don't even document the pkg@vcs+protocol://... format ivadomed is using here; they say you have to use -e vcs+protocol://....#egg=pkg

kousu

This is good now. Let it in!

EDIT ..no it's not. CI is failing now.

kousu

We need to figure out why CI is broken.

Fixes recent crashes in CI due to new ivadomed 1.1 version

jcohenadad · 2020-05-22T12:11:15Z

yess! CI is now passing. Issue was caused by ivadomed being recently bumped to 1.1, and this changed created compatibility issues with the current models (good thing our CI tests are sensitive 🙏). Issue was fixed by updating the URLs, which are now pointing to up-to-date models.

jcohenadad added the feature category: new functionality label Mar 19, 2020

jcohenadad added this to the 4.2.3 milestone Mar 19, 2020

This was referenced Mar 19, 2020

Segment volume -- Callable function from external CLI ivadomed/ivadomed#155

Merged

Silence CPU mode upon import of segment_volume() ivadomed/ivadomed#163

Open

jcohenadad added the sct_deepseg context: Global entry point for all deep learning segmentation methods label Mar 30, 2020

kousu reviewed Mar 31, 2020

View reviewed changes

scripts/sct_deepseg.py Outdated Show resolved Hide resolved

jcohenadad commented Mar 31, 2020

View reviewed changes

charleygros reviewed Apr 2, 2020

View reviewed changes

kousu reviewed Apr 3, 2020

View reviewed changes

spinalcordtoolbox/deepseg/models.py Outdated Show resolved Hide resolved

kousu reviewed Apr 3, 2020

View reviewed changes

spinalcordtoolbox/deepseg/models.py Outdated Show resolved Hide resolved

kousu reviewed Apr 7, 2020

View reviewed changes

jcohenadad commented Apr 14, 2020

View reviewed changes

spinalcordtoolbox/deepseg/core.py Outdated Show resolved Hide resolved

kousu reviewed Apr 16, 2020

View reviewed changes

kousu mentioned this pull request Apr 18, 2020

rmsct_check_dependencies #2662

Closed

jcohenadad added 6 commits May 18, 2020 13:06

Added test for deepseg models download

85bdc45

Create output folder if does not exist

10628f8

Removed mpath and now using m for both model name and path

3588ad3

Updated docstrings

26f0180

Changed keep-largest-object --> largest

14a748d

Clarified usage

490578c

jcohenadad mentioned this pull request May 18, 2020

Reorganize the way deepseg_sc works #2717

Closed

Replaced 'm' by 'model'

d0c7945

jcohenadad mentioned this pull request May 18, 2020

Long imports ivadomed/ivadomed#194

Open

jcohenadad added 8 commits May 18, 2020 16:40

Removed TODOs

293944c

Merge branch 'master' into jca/2626-deepseg

cb62ca9

Fetched default values for CLI args

32224f8

Added verbose

39ee443

Fixed logic for output file name

908cec5

Cosmetic on verbose

247f45a

Updated is_installed->is_valid in unit test

46421f9

Merge branch 'master' into jca/2626-deepseg

d696153

kousu approved these changes May 22, 2020

View reviewed changes

kousu suggested changes May 22, 2020

View reviewed changes

jcohenadad added 2 commits May 22, 2020 07:52

Replaced is_installed -> is_valid

dcb830d

Bumped deepseg models version

bb48deb

Fixes recent crashes in CI due to new ivadomed 1.1 version

Merge branch 'master' into jca/2626-deepseg

10ed5a1

kousu approved these changes May 22, 2020

View reviewed changes

jcohenadad added 2 commits May 22, 2020 16:27

Merge branch 'master' into jca/2626-deepseg

f2b5039

Merge branch 'master' into jca/2626-deepseg

fd3c8bf

jcohenadad merged commit dd004c7 into master May 22, 2020

jcohenadad deleted the jca/2626-deepseg branch May 22, 2020 22:03

joshuacwnewton mentioned this pull request Nov 15, 2020

Deprecate Tensorflow-based methods (sct_deepseg_<x>) and use only PyTorch (sct_deepseg) #3046

Closed

joshuacwnewton mentioned this pull request Dec 5, 2022

AttributeError: module 'spinalcordtoolbox' has no attribute 'image' #3972

Closed

		# This command-line tool is the interface for the deepseg API that performs segmentation using deep learning from the
		# ivadomed package.

	Threshold the prediction. For no prediction, set 'threshold' to 0.
	Threshold the prediction. For no threshold, set 'threshold' to 0.

-    args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])
-    # TODO: instead of assigning each args param, we could pass args while instanciating ParamDeepseg(args), and the
-    #  class would deal with assigning arguments to each field.
+    args = parser.parse_args(args=None if sys.argv[1:] else ['--help'])
+    args ={k: v for k in vars(args).items() if v is not None}
+    # separate out the segmentation `param` args from the top level args
+    input = args.pop('i')
+    list_models = args.pop('list_models')
+    install_model = args.pop('install_model')
+    install_default_models = args.pop('install_default_models')
+    ...

Centralized all deep learning segmentation tasks with new function "sct_deepseg" and refactored sct_download_data #2639

Centralized all deep learning segmentation tasks with new function "sct_deepseg" and refactored sct_download_data #2639

Conversation

jcohenadad commented Mar 19, 2020 • edited

Test this branch on a development ivadomed version

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

charleygros commented Apr 13, 2020

charleygros commented Apr 13, 2020

charleygros commented Apr 13, 2020

jcohenadad commented Apr 13, 2020

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jcohenadad Apr 16, 2020 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kousu Apr 16, 2020 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jcohenadad commented Apr 18, 2020 • edited

kousu commented Apr 18, 2020

kousu commented Apr 18, 2020 • edited

kousu left a comment • edited

Choose a reason for hiding this comment

kousu left a comment

Choose a reason for hiding this comment

jcohenadad commented May 22, 2020

jcohenadad commented Mar 19, 2020 •

edited

kousu Apr 16, 2020 •

edited

jcohenadad Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

kousu Apr 16, 2020 •

edited

jcohenadad commented Apr 18, 2020 •

edited

kousu commented Apr 18, 2020 •

edited

kousu left a comment •

edited