Article directory
-
- 2022-Job Description HW3
-
- learning target
- Dataset introduction
- baseline
- model selection
- data augmentation
- Test Time Augmentation (it needs to be done many times to show its advantages, the weight of the final test is better)
- Cross validation (need to make changes to the data set, so that you can train several models, if you don't want to do cross validation, just divide the validation set a little less, you don't need so much, just to verify whether there is overfitting)
- ensemble (average is relatively simple, and there may be a tie in voting that needs to be dealt with)
- No GPU resources to use Kaggle
- checkpoing to resume training (a trick to avoid interruptions for a long time)
- It is safer to train multiple models to do ensemble
- Experimental Tips
- Other tricks...
- Colab IO is very slow, it is recommended to use kaggle
- Q1. Data Augmentation
- Q2. Residual connection
- print dimension, don't calculate it yourself
- Training an early-stop strategy
2022-Job Description HW3
from：https://www.bilibili.com/video/BV1Wv411h7kN?p=35
learning target
Dataset introduction
baseline
model selection
data augmentation
mixup
What needs to be changed to use mixup
Test Time Augmentation (it needs to be done many times to show its advantages, the weight of the final test is better)
Cross validation (need to make changes to the data set, so that you can train several models, if you don't want to do cross validation, just divide the validation set a little less, you don't need so much, just to verify whether there is overfitting)
ensemble (average is relatively simple, and there may be a tie in voting that needs to be dealt with)
No GPU resources to use Kaggle
kaggle time management
checkpoing to resume training (a trick to avoid interruptions for a long time)
It is safer to train multiple models to do ensemble
Experimental Tips
Other tricks…
Colab IO is very slow, it is recommended to use kaggle
Q1. Data Augmentation
Q2. Residual connection
The provided code implements residual connections
print dimension, don't calculate it yourself
Training an early-stop strategy
