Protein Image Multi-label Classification

Project code for CS385 Machine Learning in SJTU

Project Requirements

• python >= 3.6
• pytorch = 1.5.1
• tensorflow = 2.0.0
• opencv-python = 4.2.0.34

Project structure

• main.py: main training file, please run python main.py. And notice the file paths in main.py.
• predict.py: a predict class and 3 member functions
  1. random_predict(), which can pick an image randomly from the test dataset and predict it. The star sign(*) means the groundtruth class.
  2. specific_predict(img), which reads an image and gives the result of the prediction.
  3. overall_predict(), which can predict all of the images in the test set and give the roc_auc, macro_f1, micro_f1 score of the prediction.
• manage_data.py: file of spliting datasets.
• history.py: file of drawing history graph by using TensorBoard or matplotlib.
• cmd: file of command line instructions.
• logs/: log files of training.
• models/: trained models.

Results

We split the dataset by train:test:valid = 8:1:1. And the current result comparing to the baseline offered by TA:

	macro_f1	micro_f1	roc_auc
baseline	57.18%	73.37%	93.82%
my	69.23%	72.21%	94.23%
final result	74.22%	79.89%	96.19%