Comments (6)
@x94carbone I worked this out but found that there was quite a bit of discrepancy in the way various sources said was the right way to handle it, and did not have time to test validity of the implementation I did properly so I left it out for now. There is also the case of how to really handle this beyond just the code implementation.
For example you might end up with 10-20 configurations with close to identical result. So it seems there is a need rather to add another layer of automation, where a configurable number of best results is picked, and then run again in a meaningful process that picks the best one. This leads to the question of "best" in terms of a well generalizing model vs. test performance.
The way I found the process tends to play out is something like...start with something pretty broad, and end up with some ideas on how to improve. Then narrow the boundaries, and end up with some ideas on how to improve even more. Maybe repeat this a few times, depending on the prediction task, and end up with a bunch of well performing configurations. So this will be another useful layer of process automation but involves more open questions than the one mentioned above. So I'm a little inclined to skip the "legacy" approach of picking the best model all together at least for now and go straight to figuring out an approach that appreciates the problem more.
In which case the parameter save_best_model is redundant for now.
from talos.
@mikkokotila Thanks for the response! I have some ways in mind that could be used address these concerns as well. They pertain closely to my current research. What it appears you're talking about is a more insightful way of picking the best hyperparameters (HP). I have two ideas about this depending on how far you're willing to take this.
First, the first step could be exactly what Talos already does: screen different HP configurations, but then pick the top N of them. Second, of those top N models, randomly permute the test/train/dev sets and retrain/eval for each permutation. Implement some measure of model robustness such that only models that are robust to the random permutations of the data are kept and return the best overall. This also might involve a more thoughtful split of the data initially so as to avoid bias.
Second, you (we?) should consider evolutionary algorithms. This will greatly improve performance overall and could be built on top of the first step.
But anyway regarding the actual issue here I understand what you're saying. Is there any way to completely remove the option and prior package (Hyperio) to avoid confusion?
from talos.
@x94carbone 👍
Regarding your first idea, I agree. Maybe there is an option to pick top N configurations / all configurations that meet a set metric threshold. It will probably be a good idea to always throw in some randomly or otherwise picked configurations for control purpose.
Regarding the split of the data, I think the starting point would be that a "large as possible" portion of the original dataset is reserved for the Val() stage and it never touch the Scan() stage. Then split further in some meaningful way when it gets to Val(). It seems that the splitting part will be trivial and not affecting the mainline of Val() so might be better to handle it separately later. The other thing is to have all configurations in Val() run n times and measure std, mean and max as evaluation metric as opposed to just the objective metric (e.g. f1 score).
Regarding evolutionary algorithms. I think that is a great idea. After having lightly reviewed the many approaches used for this problem, I thought the evolutionary algorithm approach was very interesting. Are you familiar with the python libraries for that purpose? Do you have suggestions for how to best approach adding it. I think it would be a good idea to prioritize this as it makes the current iterative approach redundant and allows shifting focus to the next layer of process automation. I'll create a separate issue on that with some related questions.
[EDIT] There is now #18 for the evolutionary algorithm topic.
from talos.
For the actual issue, I will remove all the redundant codes.
from talos.
@mikkokotila Sounds good. I absolutely agree on all points. I'll comment in #18 about evolutionary algorithms and elsewhere where necessary.
from talos.
The artifacts were cleaned from the current version.
The wider discussion continues in #40 so closing here.
from talos.
Related Issues (20)
- Key duplication results in exception: param dict vs. csv file stored (e.g., 'loss') HOT 3
- Heatmap always have missing parts HOT 2
- Failed to build scipy while installing talos HOT 1
- Using `Evaluate` with multiple inputs HOT 6
- ImportError: cannot import name 'get_config' HOT 4
- add arguments to the function syntax HOT 1
- Using train, val data HOT 1
- Cannot replicate typical example from documentation HOT 1
- Can't find description of layer shapes HOT 3
- Talos installation stuck at downloading statsmodel HOT 1
- [RFC00102] Increase Visibility and Access to Power Draw Data HOT 1
- Getting errors when handling custom object HOT 4
- Add comprehensive support for n-dimensional `x` or `y`
- AttributeError: module 'talos.utils' has no attribute 'ExperimentLogCallback' HOT 1
- Live is not in talos HOT 1
- Advanced hyperparameter configuration HOT 1
- Error with Autom8 import HOT 1
- pip install talos is failing due to sklearn in requirements.txt instead of scikit-learn HOT 4
- Will talos support Maximal Update Parametrization later? HOT 1
- batch_size not support list HOT 1
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 talos.