Attempt at using Mobile Net v2.0 on DeepFashion2 dataset to detect top, bottom, and full body clothing.

There are possible bugs in this code as it is not the up-to-date version Pytorch Implementation of MobileNetv2.0 is from qfgaohao with slight adjustemnts to meet our needs. The dataset used can be found by going to deepfashion2.

Heirarchy of folders should be the following:

 Root: pytorch-ssd
    ->train:
        ->Annotations
        ->image
        ... 
    ->val:
        ->Annotations
        ->image
        ...
    ->deepfashion-classes.txt 
    ->generate_xml_annotations.py
    ->train_ssd_lite.py
    ->generate_deepfashion_data.py

Ensure deepfashion2_Dataset.py is within the vision/models folder

Extracting the data

#create the xml annotations to be read in by generate_deepfashion_data
python generate_xml_annotations.py 'train' 
python generate_xml_annotations.py 'val' 

Create the Image Sets for training

python generate_deepfashion_data.py 'train'
python generate_deepfashion_data.py 'val'

Train the model

python train_ssd_lite.py --dataset_type 'deep_fashion_2'  --datasets ./  --validation_dataset ./ --net mb2-ssd-lite --scheduler cosine lr 0.01 --t_max 200 --validation_epochs 5 --num_epochs 20