rg314 / autoballs Goto Github PK

The tests currently have poor coverage. Need to add tests for all parts of the package.

In some cases, there are multiple eyeballs in the FOV. This can cause complications for the sholl analysis for obvious reasons. Ensure that a warning message is created for these instances and is flagged in the output result.

The main problem with the network is thresholding the images after the bandpass filter has been applied.

this line of code (line 71 of pipeline_cnn.py as of commit 6f1968a):
best_model = torch.hub.load_state_dict_from_url('https://docs.google.com/uc?export=download&id=13CLZoNyvCt2K46UvAyHUqH7099FbnBh_')

errors on _cuda_deserialize.

Assuming that this model specifically is cuda-based and therefore cannot be deserialised for cpu computation.

Should the model be ensured to work across both CUDA and CPU compute?

Issues with people raising PRs without being colab

The folder structure is quite rigid and users may find it difficult to define the folder structure and metadata file. It could also be possible to set up a GUI that users can drag and drop images. If people start using the package I will implement this.

In some instances the sample only contains an eyeball with no axon growth. This needs to be classified as no growth and should be easy to integrate within the current architecture for example

aux_params=dict(
    pooling='avg',             # one of 'avg', 'max'
    dropout=0.5,               # dropout ratio, default is None
    activation='sigmoid',      # activation function, default is None
    classes=4,                 # define number of output labels
)

model = smp.Unet('resnet34', classes=4, aux_params=aux_params)
mask, label = model(x)

mask.shape, label.shape
# (N, 4, H, W), (N, 4)

See segmentation models for more details smp

1. Most functions are missing documentation
2. Usage section in the README.md need updating