nelson-gon/cytounet

Deep Learning Based Cell Segmentation

cytounet's Issues

Wrong Image Order

• I have read the wiki and my issue is not
  handled (sufficiently) there or is not answered at all.

Describe the bug

Images shown by show_images are in the wrong order.

To Reproduce

Save images numbered 1 to 20.

Expected behavior

I expected to get images shown from 1 to 20(mathematically)

Unexpected behavior

I got images in the order 1, 10, 11, 12...20.

System Details

I am using this version.

Rethink reshaping

Current predictions are always black or random. This seems to stem from the adjustData part that may not be generalizable.

Improve unet_simple

Description

I would like to improve the initialization of unet_simple especially regarding making it similar to what the original paper proposed.

Similar Features

It is currently implemented in unet_simple

Feature Details

• Improve upsampling

• Review Downsampling

Proposed Implementation

None yet

Thank you

Fix issues with image shape

• Flexibility in show_images regarding ambiguous image shapes e.g 4 vs 3 vs 2

Provide Standalone Script

Description

I would like to run my training and/or fine tuning at the command line.

Similar Features

This is similar to installing the package with pip.
Feature Details

Often it is possible to run a model's pipeline by doing something like this:

python --args path_to_script --other-args

Proposed Implementation

I can think of two ways to do this.

• Use CI/CD or GitHub actions to run "tests"/training in the cloud.

• Provide a script under a folder named scripts that can be run as suggested above.

Remove calls to deprecated functions

Describe the bug
Model.fit_generator is deprecated

To Reproduce
Run via Colab

Expected behavior

Expected to run with no warnings

Unexpected behavior

Received a deprecation warning

System Details
Latest dev.

Using Custom Dataset

Hi,
Thanks very much for this wonderful project. I have a custom dataset formatted in form of the COCO dataset. Any suggestions on how I can plug that into this project?

Looking forward to your wonderful response

Extend Metrics

Description
I would like to have more metrics available.

Similar Features

"N/A"

Feature Details

Implement or at least add support for such metrics as Jaccard.

Proposed Implementation

Use native implementations or implement some from some papers.

Add Debugging Tips in README

Speed up show_augmented

Support non string input in show_images

Move to snake_case

Related to #12

Model fails to save when using callbacks

• I have read the wiki and my issue is not
  handled (sufficiently) there or is not answered at all.

Describe the bug

After training, I used a callback as shown in the example notebook and found that the model was not saved at the end of each epoch.

To Reproduce

Run the exact code in the above notebook.

Expected behavior

I expected to save the model at the end of each epoch based on the loss' improvement(or not).

Unexpected behavior

I got no error message. The model trained successfully but required that I manually save the model at the end of the entire training.

System Details

Using developer version with this hash.

Support XML and CSV based annotations

Image Distortion with save_predictions

Describe the bug

Images obtained after training on png input get distorted when saved with save_predictions

To Reproduce

This is reproducible with embroys.ipynb in the examples folder.

Expected behavior

I expected to save images to my directory as is.

Unexpected behavior

I got distorted noise like images instead.

System Details

Version 0.2.0, this version.

Indicate Name Change in Docs

Information lost after saving with saveImage

More Comprehensive Examples

Description

Current examples use very small data sets. Could it be possible to use a
Similar Features

N/A

Feature Details

Include examples that show the train-validate-test workflow in a more comprehensive manner.

Proposed Implementation

Use publicly available data sets.

Empty list when reading images of mixed formats in the same directory

• I have read the wiki and my issue is not
  handled (sufficiently) there or is not answered at all.

Describe the bug

read_images does not read images if they are of mixed formats but in the same folder.

To Reproduce

• Create a directory with images of mixed file formats

• Try to read the images with read_images

Expected behavior

I expected a list containing arrays of these images.

Unexpected behavior

I got an empty list instead.

System Details

I am using the developer version available here

Issues with test generation

Generating test data currently fails with the error:

OSError: Cannot understand given URI:

Add Training Helpers

Description

In training the model, I would like to be able to simply call some function that runs the training.

Similar Features

None.

Feature Details

For example instead of model.fit_generator, I would simply call something like train_model.

Proposed Implementation

Either write new methods or change to a class based API(this would need to be in some milestone).

Use Descriptive Names

Add Sanity Checks

• Tests
• Examples

Provide greater control over contours returned

Description

I would like to control what is drawn on the image when I use draw_contours.

Similar Features

This is related to draw_contours and find_contours.

Feature Details

It would be great if I could control font size and optionally provide an earlier list of images.

Proposed Implementation

Add arguments for the above.

Support HyperParameter Grid Search/Tuning

Show Images-Masks Side by Side

Description

In show_images, I would like to show images side by side.

Similar Features

show_images is close but not quite.

Feature Details

Show images either above or side by side with their predicted labels and/or show predicted, truth, image.

Proposed Implementation

Similar to what pyautocv does but see above.

Error on File Length Incompatibility

Description

When using a validation data set, I would like to be warned if lengths differ.

Similar Features

This would enhance the data generators e.g. generate_train_data or generate_validation_data.

Feature Details

N/A: This has been sufficiently described above.

Proposed Implementation

Raise an error on incompatible lengths and give a very specific error message.

Custom Model Depth

I would like to be able to reduce/increase model complexity preferably incrementally.