ChoiceNet
TensorFlow Implementation of ChoiceNet on regression tasks.
Summarized result:
Paper: arxiv
Classification (MNIST) Result
Error type: [Permutation]
| name | Result |
|---|---|
| Outlier Rate: 25.0% |  |
| Outlier Rate: 45.0% |  |
| Outlier Rate: 47.5% |  |
Error type: [Random Shuffle]
| name | Result |
|---|---|
| Outlier Rate: 50.0% |  |
| Outlier Rate: 90.0% |  |
| Outlier Rate: 95.0% |  |
Error type: [Label Bias]
| name | Result |
|---|---|
| Outlier Rate: 25.0% |  |
| Outlier Rate: 45.0% |  |
| Outlier Rate: 47.5% |  |
Regression Result
Reference Function: [cosexp]
| name | Training Data | Multi-Layer Perceptron | Mixture Density Network | ChoiceNet |
|---|---|---|---|---|
| oRate: 0.0% |  |  |  |  |
| oRate: 10.0% |  |  |  |  |
| oRate: 30.0% |  |  |  |  |
| oRate: 50.0% |  |  |  |  |
| oRate: 60.0% |  |  |  |  |
| oRate: 70.0% |  |  |  |  |
Reference Function: [linear]
| name | Training Data | Multi-Layer Perceptron | Mixture Density Network | ChoiceNet |
|---|---|---|---|---|
| oRate: 0.0% |  |  |  |  |
| oRate: 10.0% |  |  |  |  |
| oRate: 30.0% |  |  |  |  |
| oRate: 50.0% |  |  |  |  |
| oRate: 60.0% |  |  |  |  |
| oRate: 70.0% |  |  |  |  |
Reference Function: [step]
| name | Training Data | Multi-Layer Perceptron | Mixture Density Network | ChoiceNet |
|---|---|---|---|---|
| oRate: 0.0% |  |  |  |  |
| oRate: 10.0% |  |  |  |  |
| oRate: 30.0% |  |  |  |  |
| oRate: 50.0% |  |  |  |  |
| oRate: 60.0% |  |  |  |  |
| oRate: 70.0% |  |  |  |  |
HowTo?
- run code/main_reg_run.ipynb
- Properly modify followings based on the working environment:
nWorker = 16
maxGPU = 8- (I was using 16 CPUs / 8 TESLA P40s / 96GB RAM.)
Requirements
- Python3
- TF 1.4>=
Contact
This work was done in Kakao Brain.