predict_deep_fashion_attributes

Garments in fashion domain are characterised by attributes like pattern, neck types, sleeve types, size, material. Ecommerce websites use this data to help users navigate through their catalog via filters and effective categorization. To predict these visual features using Deep Learning. Please train a ​Single Neural Network to predict all three attributes for each garment.

Data Representation

• Dataset can be downloaded form here
• All images are stored in images folder.
• Attributes for each image is stored in attributes.csv file with id as image filenames and "neck ", "sleeve_length" and "pattern" as their attributes
  • Neck attribute is distributed to 0-6 types
  • Sleeve length attribute is distributed to 0-4 types
  • Pattern attribute is distributed to 0-9 types

Convolutional Neural Network is trained :

• Download the trained model here and paste them in "models" folder
• To train the model run : python3 main_multi_label.py /path/to/data/folder
• To predict individual attributes for an image run : python3 main_multi_label.py /path/example/image /path/to/saved_model/
  • The current accuracy 90.11% on 10% validation set.
  • Output of a predicted image is a string of pattern : ("attribute1" + "attribute_code1"), ("attribute2 " + "attribute_code2"), ("attribute3" + "attribute_code3")

Fine-tuned pretrained vgg16 model and retrained last layers with our dataset

• Download the trained model here
• To train teh model run : python3 multi_label_vgg_nn.py
• To predict individual attributes for an image run : python3 main_multi_label_pretrained.py "path/to /example/image/to/predict /path/to/saved/weights/folder
  • The first trained model gave current accuracy 68.16% on 10% validation set.
  • The approach is to share the pretrained layers between all attributes and defining last layers for each attributes to predict inidividually.

Need for improvement

• Taking care of #NA values, will take much more development. One way is implemented to treat them as a separate class.
• A new approach is to train a multi-task network which performs better in terms of model generalization.
• The pretrained model is still in training and taking lots of time. Will update when model is trained again and upload to google drive. Training is done and took 996 minutes.
• Introducing k cross fold validation to make every batch_size a representable whole of complete dataset.
• Increasing no of data points via data augmentation.
• Changing the transformations performed and using more related normalizing with mean and standard variance values.