- Ubuntu 16.04 LTS
- NVIDIA GTX 1080
- TensorFlow 1.0
- Python 3.5 (Lab 3 ~ Lab 6, Lab 9, Lab 10, Final Project)
- Python 2.7 (Lab 7)
- C++ (Lab 8)
Implement NIN, all convolutional NIN and train on CIFAR-10.
Method | Test Error |
---|---|
NIN + Dropout | 10.89% |
All Conv. NIN + Dropout + Data Augmentation | 10.31% |
NIN + Dropout + Data Augmentation | 8.88% |
Combine NIN with different activation functions, BN, He weight initialization and train on CIFAR-10.
Method | Test Error |
---|---|
ReLU NIN + Dropout + Data Augmentation + BN | 8.22% |
Maxout NIN (k=3) + Dropout + Data Augmentation + BN | 7.67% |
Use VGG-19 to build an object recognition system, and retrain VGG-19 on CIFAR-10.
Method | Test Error |
---|---|
Random initialization + BN | 7.97% |
Pretrained model + BN | 6.94% |
Build LSTM to perform the copy task.
Training length | 1~20 | Training length | 30 | ||
Test length | 10 | 99% | Test length | 20 | 85% |
20 | 99% | 30 | 99% | ||
30 | 30% | 50 | 7% |
Add a hard attention mechanism to this code.
Attention | BLEU-1 | BLEU-2 | BLEU-3 | BLEU-4 | METEOR |
---|---|---|---|---|---|
Hard | 63.6 | 42.0 | 28.2 | 19.3 | 19.8 |
Soft | 65.3 | 43.5 | 29.2 | 19.9 | 20.5 |
Build an AI to play 2048 through TD(0).
After 1000K training games, the winning rate is 0.974 (averaged over 10K test games).
Train a DQN to play Breakout.
During training, the game score obtained by the agent could achieve 65.
Apply DDPG to the pendulum problem.
Real-Time Partial Style Transfer.