Unofficial Pytorch implementation of the paper 'An Empirical Study of Example Forgetting during Deep Neural Network Learning' experiment on ImageNet-1K
Jupyter Notebook 96.95%Python 2.87%Shell 0.18%
example_forgetting's Introduction
Example Forgetting Implementation with Pytorch
Unofficial implementation of the paper An Empirical Study of Example Forgetting during Deep Neural Network Learning
0. Brief Explanation
Check data with no learning events with various epochs - 15, 40, 60
ResNet 50 mixed precision training on ImageNet-1K
All the experiments are done with 1 RTX 3080 GPU
1. Implementation Details
_eda : related with data pre-processing and eda
_not_learned_json : images with no learning events
_train_logs : train logs of ResNet 50 with mixed precision training
data.py : data augmentations, dataset
ImageNet_class_index.json : ImageNet-1K label information
main_single.py : train and evaluate model and save forgetting event logs
not_learned_train_filtered.json : pre-process not_learned_train.json using epochs
not_learned_train.json : data - not learned epochs output
run.sh : run python per gpu
utils.py : utils such as scheduler, logger
2. Results
2.1. Base Configuration
config
value
optimizer
SGD
base learning rate
0.05
weight decay
1e-4
optimizer momentum
0.9
batch size
128
learning rate schedule
cosine decay
warmup epochs
5
augmentation
Resize and Random Crop
loss
CrossEntropy
2.2. Train Results
epoch
acc
15
71.512
40
73.624
60
73.718
3. Reference
An Empirical Study of Example Forgetting during Deep Neural Network Learning [paper] [official code]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent