Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Failed to reproduce the results and Loss soar----pytorch #62

Open
catherinezll95 opened this issue Nov 29, 2019 · 1 comment
Open

Failed to reproduce the results and Loss soar----pytorch #62

catherinezll95 opened this issue Nov 29, 2019 · 1 comment

Comments

@catherinezll95
Copy link

catherinezll95 commented Nov 29, 2019

Hi, thank you for opening up the code. In the process of running the program implemented in pytorch, I encountered several problems, I hope to get your answer.

  1. Failed to reproduce the results.

For example:
1)Res. sup. (MA)
python src/translate.py --residual_velocities --learning_rate 0.005

the results I get:
global step:10000 learning rate: 0.0047

milliseconds | 80 | 160 | 320 | 400 | 560 | 1000 |
walking | 0.374 | 0.636 | 0.878 | 0.980 | n/a | n/a |
eating | 0.263 | 0.457 | 0.710 | 0.864 | n/a | n/a |
...

In your paper
milliseconds | 80 | 160 | 320 | 400 |
walking | 0.28 | 0.49 | 0.72 | 0.81 |
eating | 0.23 | 0.39 | 0.62 | 0.76 |
...

2)Sampling-based loss (SA)
python src/translate.py --action walking --seq_length_out 25

the results I get:
global step:10000 learning rate: 0.0047

milliseconds | 80 | 160 | 320 | 400 | 560 | 1000 |
walking | 0.60 | 0.607 | 0.808 | 0.889 | 1.029 | 1.186 |

In your paper
milliseconds | 80 | 160 | 320 | 400 |
walking | 0.92 | 0.98 | 1.02 | 1.20 |


2.When the global step>10000, the loss will soar.

Res. sup. (MA)
Global step = 10000

milliseconds | 80 | 160 | 320 | 400 | 560 | 1000 |
walking | 0.374 | 0.636 | 0.878 | 0.980 | n/a | n/a |
eating | 0.263 | 0.457 | 0.710 | 0.864 | n/a | n/a |
...

Global step = 10200

milliseconds | 80 | 160 | 320 | 400 | 560 | 1000 |
walking | 1.559 | 2.374 | 3.420 | 3.835 | n/a | n/a |
eating | 1.271 | 1.732 | 2.439 | 2.789 | n/a | n/a |
smoking | 1.464 | 2.132 | 3.167 | 3.508 | n/a | n/a |
discussion | 1.828 | 2.441 | 3.116 | 3.379 | n/a | n/a |
...

@yuyujunjun
Copy link

I tried pytorch version and I encountered the problem 2, too. At line 211 of translate.py, changing the optimiser from Adam to SGD will solve the problem 2. The loss exactly will grow up when iterations are too large but do not soar anymore. In #22, the author suggests the iterations might be 10000 or 20000 so it's not a big problem.
My experiment:
python src/translate.py --residual_velocities --learning_rate 0.005 --action walking

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants