semi-conductor-image-classification-first
semi-conductor-image-classification-first
The in-class image classification contest, in which you need to build a model and recognise bad chips and good chips.
Data
Semi-conductor images. Two classes: {"0": "good_0", "1": "bad_1"}.
Bad chips for example:
Good chips for example:
Approches
主要用了简单 CNN，ResNet20v2，和 ResNet56v2。v2 代表使用的是 ResNet version 2，20 和 56 是层数。
Loss function
Loss Function 使用 categorical_crossentropy, 随后又使用tensorflow imbalanced_data中方法进行 class weights 加权。
Results
Current best score is obtained by ResNet56v2:
| Epoch | AUC | Public Score | Date |
|---|---|---|---|
| 109 | 0.9832 | 0.97156 | 20200225 |
Engineering contributions
Confusion matrix
Modified TensorFlow official confusion matrix metrics codes in confusion_matrix_v2_1_0.py so that
- 2D tensor y_pred and y_true input can be accepted;
- class_id can be specified when collecting these metrics such as: false positive, false negative, Recall, AUC.
Test focal loss from Tensorflow Models
https://github.com/tensorflow/models/blob/master/official/vision/keras_cv/losses/focal_loss.py
License
MIT License