Author: yoojoonl

This week I worked on:

• fine-tuned detector. It didn’t result in significant increases in accuracy.
• I ran into some overfitting problems with the classifier. The training accuracy would hover around 70% but the testing accuracy was only 30%. I added a batch regularization layer, dropout layer, and better preprocessing. Hopefully by next week I can get a model with good accuracy. If I still can’t get a good model trained, as risk mitigation I can find an existing model online, however it is not ideal as existing models use old architecture so I would not maximize accuracy.
• implemented clothing images and attributes storage. Uses pickle to store images and attributes in a dictionary.
• Implemented color attribution. Uses MMCQ algorithm with color thief API.

For next week:

• Retrain classifier with new model.
• train fine-grained clothing attribution model.

Although I am slightly behind schedule for the models, I am ahead of schedule for the clothing storage, which means I can focus more time onto the models.

The past two weeks I worked on:

• Finished the design report. I wrote most of the documentation on the visualizer component.
• formatted datasets for training. Detector required tfrecord files from DeepFashion2 dataset while classifier required images sorted into classes from DeepFashion dataset.
• Trained a clothing detector that detects bounding boxes and does simple classification of clothing articles.
  • tried yolov4 using pytorch, but API was really bad and had many problems I had to fix by looking at pull requests. Ultimately, didn’t use.
  • tried efficientDet0 with tensorflow2 object detection API and it was much better documented and had more community support. Ultimately, used this.
  • Using DeepFashion2 dataset, which doesn’t provide good classification labels, but has bounding boxes over every clothing item (DeepFashion only has 1 bounding box per image)
  • unfortunately, I didn’t measure accuracy during training, but here are the loss figures:
  • 
• Training a clothing classifier that provides more fine-grained classification.
  • Using efficientNetb0 from keras. Below is test accuracy:
  • 
  • Still requires more epochs until convergence.
• Created simple clothing recognition API with aforementioned models. Currently returns top 10 most likely bounding boxes and labels.

For Next Week:

• Finish training models and perform fine-tuning.
• Start training models for more fine-grained classification like shape and material.
• Implement color classification.
• Right now detector could predict differently from classifier. I’m not sure what to do when this happens, but I should figure it out.

I am slightly ahead of schedule.

This week I worked on:

• Presented the design review presentation. We got a lot of good feedback, especially for our hardware design and our ML algorithm.
• Worked with Fred on details of hardware design. Came up with a design that uses gears so the servo doesn’t need to be in the middle of the axis. This ensures there’s no weight on the servo and it also allows us to potentially add additional servos if we need more torque.
• Wrote code that can train any prebuilt models on pytorch. Uses fastai wrappers to make the job easier.
•  
• Created a resnet34 classification model with 66% top-1 test accuracy. Not a metric worth noting as I still need to have better data augmentation and hyperparameter tuning. We also need validation testing from internet sources and we’re looking at top-3 accuracy, but it’s a start.

For next week:

• Finish the visualizer section of the design report.
• Bootstrap training data with additional data from the web.
• Create validation set by manually labeling web-scraped images.
• If I have the time, try training an object detection model that can find top and bottom.

I am on schedule.