CNN GAP

About

• This work is related to Physionet 2017 contest. (https://physionet.org/content/challenge-2017/1.0.0/)
• This work aim at classifying ECG signals into 4-types, including "Atrial Fibrillation", "Normal", "Other Diseases" and "Noisy".
• We reproduced works of the paper "Towards Understanding ECG Rhythm Classification Using Convolutional Neural Networks and Attention Mappings" by GoodFellow et. al.

Our Works

• Train on N/A/O 3 type and got F1 score 0.84. (same as the paper)
• Observe CAM on ECG of each type and find what the model have learned.
• Train N/A/O/W 4 type with a fine-tuning procedure on the model pretrained by 3 type, and got F1 score 0.81.
• Fine-tuning the previous model with F1 loss and got F1 score 0.82, which is higher than all of the work.
• Analyze convergence property of F1 loss and WCE.
• Analyze convergence property of Softmax and Log-Softmax activation.

Files discription

• Net/ : All models network structure
• utils/ : Preprocess function and loss function
• report/ : All .ppt of this work
• dataset/ : Dataset
• jupyter notebook/Plot confusion matrix.ipynb : Plot confusion matrix of a model
• jupyter notebook/plot.py : Plot accuracy, loss or F1 score of the training process
• jupyter notebook/analyze f1 loss.ipynb : Analysis of convergence speed of F1, recall and precision.
• jupyter notebook/Train* : Training models
• model_description.txt : Description of all models