musdl's People
musdl's Issues
The normalization for AQA-7 dataset
Hi! I found the value range difference in the AQA-7 dataset.
So what is the way for converting the hard target(Regression) to soft target(USDL)? The output_dim is same as 101 for all kinds of actions?
I didn't find details about this in the paper and codes, could you give me some suggestions? Thanks a lot!
Unable to get the proper regression results
Hello, thank you for your wonderful work!
I can get wonderful results for USDL and MUSDL settings. However, the results are terrible when I turn the labels to the initial hard targets(Our-Regression in Table 3). All outputs are around 70. I tried MSE loss and SmoothL1Loss and they all did not work.
The modification for code is as follows:
1.I modified the output_dim for "USDL" in opts.py, from 101 to 1.
output_dim = {'USDL':1, 'MUSDL': 21}
2.I removed the softmax function on layer3 in evaluator.py
output = self.layer3(x)
3.I regarded the output from the evaluator as the final predicted scores and calculated the loss.
preds = evaluator(clip_feats.mean(1)).float() #preds = compute_score(args.type, probs, data) pred_scores.extend([i.item() for i in preds]) target=torch.tensor(data['final_score']).float().cuda() if split == 'train': loss = criterion(preds,target) optimizer.zero_grad() loss.backward() optimizer.step()
If possible, could you release your code for "Our-Regression" in the paper?
Problem with JIGSAW
When trying to run with Jigsaw data, the code gives error of No CUDA devices available and on line in main.py of JIGSAW
i3ds = nn.ModuleList([InceptionI3d() for _ in range(4)]).cuda()
Although Cuda is using Nvidia Tesla from Colab and is enabled. The same environment was able to run with other dataset.
about the pick of standard deviation in Gaussian distribution
Hi Zanlin Ni,
Thanks for your wonderful work!
I was wondering whether the selection of standard deviation has an effect on the loss(Here ฯ is a hyper-parameter which serves as the level of uncertainty for assessing an action) since if I pick a bigger one, the loss in the training phase is close to zero, and if I choose 1, the loss is about 0.01. Will the size of the loss value have an impact on training?
Looking forward to your reply!
Spearman Correlation is too low
Hi I executed the MTL-AQA training and got only 0.68665 Sp corr. I have kept all the config setting as it is except for num_worker 1, training batch size 1 and test 2.
I am using a 2048 Nvidia GPU with 10 GB memory
CUDA 11.4
Pytorch '1.10.2'
Could you tell why the drastic decrease in Sp corr. Can you please share your pretrained weights?
Unable to get stable results for multiple iterations (JIGSAW)
Hi there,
I am using your repository to reproduce the results on the JIGSAW dataset and I am using the default parameter values as defined in the config.py file except for the batch size. I was using the batch size of 1 on colab. But I am unable to get stable results for multiple runs.
For example, I ran USDL for Suturing and got 0.3 test correlation, then I ran it again and this time I got nan value for test correlation, and for the third time, I got .19 test correlation. This behavior is there in all 3 surgical videos.
What could be the reason for this unstability? What were the hyperparameter values that you have used?
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
๐ 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. ๐๐๐
-
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 โค๏ธ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.