MobileNetV3

A tensorflow implementation of the paper "Searching for MobileNetV3" with a R-ASPP segmenter head and classification head.

This project is in active development and as such, has a lot of rough edges. The implementation should be true to the original paper, but your mileage may vary.

Installation

pip install -r requirements.txt

Results

The implementation was initially done to apply the MobileNet v3 architecture to skin lesion segmentation, as per the ISIC Challenge 2018. The 2018 edition was chosen as it is the last one with a segmentation task.

Metrics

For segmentation, two metrics are availables, namely Dice Coefficient and Jaccard Index. In both cases a matching loss is also provided.

Optimizers

The original paper used RMSProp, but we found out that training could be sped up by using Adam instead.

Train from scratch

COCO

You can download the files here.

Once that's done, create this hierarchy in your directories:

You should be able to run it with python3 train.py

ISIC 2018

You can download the dataset here.

TODO

Evaluation

Segmentation

To get predictions for one image

Run python3 eval.py --model-path out/ -i your_image.png to output the segmentation mask.

To get predictions for multiple images

TODO

Classification

To get predictions for one image

TODO

To get predictions for multiple images

Run python3 eval_classification.py --model-path out/ --input-dir your_test_dir/ --labels-file labels.csv to output the confusion matrix.