Comments (11)
So, not really sure what the problem is.
Single rgb works fine
Single depth works fine
Double stream does not work, is that correct?
from admd.
After you train rgb
and depth
, you should have the trained checkpoints saved in the model
folder.
When you train the double_stream
, the two branches of the network are initialized with those checkpoints. From your error, it look like you do not have the depth checkpoint there.
from admd.
I think my path was wrong. Also when in training it's written test acc. do you really check the accuracy on the test set at every epoch?
from admd.
Well at the beginning I was just testing at the end with the --mode test_xxx
options. Early stopping is done when loss is nearly stable since there is no validation set for this dataset. I inserted the test check during training just to verify everything was ok. But in no way test accuracy is used for validation.
from admd.
Ok it's working but I dont really undertstand I dont get performance like you there must be a bug, just to recap what I do for training the double stream, I first train RGB and depth single stream after I start the training for the double stream, at start the test. acc is 54% but after first epoch it decrease to 37% and at the end of the 2000 iterations its even lower, around 23% acc. I didnt change anything in your code.
from admd.
Hi, as we state in the paper, finetuning the two-stream does not help much and is often harmful.
Simply taking the two separate streams (trained independently) and average logits works.
For NYUD I think I only finetuned for 1000 iterations with a very low learning rate.
Please see
Line 11 in 754d5cc
Running this will finetune for a tiny bit and save the two stream in a single model for testing and further use.
Do you get a reasonable accuracy for the two separate RGB and Depth stream (cf. Table 5 in the paper)?
from admd.
I'm gonna give you numbers tomorrow or after tomorrow, but it was less than what you put in your paper to be frank.
from admd.
This is what I get, which is the number in the paper.
from admd.
What is your training strategy ? What do I have to train: single stream RGB, single stream Depth, double stream after that (?) and finally Hallucination network ? I followed your paper, but certainly I was wrong somewhere.
from admd.
Yes, that is the idea. In the file run.sh
you have everything, also some other baselines such as modDrop and rgb ensemble.
from admd.
Closing this since inactive for a week
from admd.
Related Issues (14)
- Feature for comparison ResNet-50 on NYU
- Resnet 50 v1 or v2? HOT 2
- How to run the code on NTU for Action Recignition HOT 2
- Theoretical question about your TPAMI paper HOT 3
- confusion about dataset in each training step HOT 1
- Northwestern-UCLA dataset HOT 3
- The depth images of N-UCLA HOT 1
- Your code don't work... HOT 15
- Help to change the first convolution HOT 12
- Cannot load pretrained weights into a modified network HOT 7
- NYU classification: What modification do you do to your network ? HOT 10
- train_hallucination vs train_hallucination_p2 for NYUD HOT 9
- Help with train_hallucination_p2 for NYU HOT 12
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from admd.