Comments (2)
The loss prediction that the ProgressBar
is showing you is necessarily something of an estimate. If we want to compute an exact loss value, we need to actually compute predictions with the current state of the neural network on every point in in
---and that can be a large amount of computation! So, we don't do this, and instead use the loss just on each minibatch of training points as the progress bar estimate (all other toolkits do something like this too). This means that the first minibatch loss estimate has a lot of variance, and then as we add estimates for other minibatches, it slowly converges to the full-data loss (but even then it is still something of an estimate). This explains the "jumpy" behavior you see at the beginning of the epoch.
The optimizer isn't actually being reset between iterations, and the model isn't getting worse---it's just a matter of the approximation of how we can efficiently get some kind of loss estimate.
One experiment you could do to see this is use the PrintLoss()
callback instead of ProgressBar()
. If you run for 10 epochs, this will print 10 loss values. Then, compare this with what happens if you use an adaptation of your single-epoch training loop:
int epochs = 1000;
for (int e = 0; e < epochs; e++)
{
double loss = Train(in, out, opt);
std::cout << loss << std::endl;
}
and this should print very similar numbers (they may not be exactly the same due to randomness in shuffling the data and other small effects).
from mlpack.
@rcurtin Thank you for thorough explanation!!! I have tried something similar to the code you proposed but i couldn't understand the slight differences in values. Now everything is clear. Sooo Resolved!!!
Thanks again
Cheers!!!
from mlpack.
Related Issues (20)
- stb_image_write warning while compiling HOT 1
- Benchmark to replace the transform functions HOT 15
- Can't train a model having bias addition layer Add() HOT 8
- Reverse Convolution? HOT 6
- Documentation issue
- [R] - `verbose` argument has no effect HOT 1
- Get rid of `arma::fill::zeros` when we upgrade the minimum armadillo version HOT 5
- Document `internal_compact::` name space for `arma::fill` HOT 2
- [R] - Global option for 'verbose' argument HOT 5
- Add `.prepare` script to have r-universe automatically build new nightlies HOT 1
- bfd.h:35:2: error: #error config.h must be included before this header HOT 4
- Any ideas about Random Forest regressor? HOT 2
- Switch from `-j 2` to `-j ${nproc}` HOT 2
- dimensionality mismatch: Decision Tree CLI with both -t and -T specified
- [R] - Should the returning model object gain a class?
- NumPy 2.0 support
- [R] Switch `sprintf` to `snprintf` HOT 4
- Physics-Informed Neural Network possible with MLPack? HOT 1
- 1-D Convolution issues about time series data HOT 1
- Using Header-Only mlpack via CMake FetchContent and Automatic Dependency Download
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 mlpack.