Added support for regression prediction, ie raw values instead of cla… #7
Conversation
…ssification, as a minor modification of fine_tune_ae().
|
I just used the auto_encoder code and instead of using the samples as the target output, used an additional set of iterators. This is the test code. #include <dll/network.hpp> #include int main() {
|
|
Hi, Thank you very much for this Pull Request. Sorry for the delay in answering, I'm not very active during the holidays. Everything seems in order. If I understand correctly: This enable to train regression models by using the same kind of generator as the auto encoder, but not enforcing auto-encoder model. Right ? Thanks Baptiste |
include/dll/dbn_impl.hpp
Outdated
|
|
||
| validate_generator(generator); | ||
|
|
||
| //cpp_assert(dll::input_size(layer_get<0>()) == dll::output_size(layer_get<layers - 1>()), "The network is not build //as an autoencoder"); |
There was a problem hiding this comment.
You can directly remove this instead of commenting
include/dll/dbn_impl.hpp
Outdated
| */ | ||
| template <typename InIterator, typename OutIterator> | ||
| weight fine_tune_reg(InIterator&& in_first, InIterator&& in_last, OutIterator&& out_first, OutIterator&& out_last, size_t max_epochs) { | ||
| //return ""; |
Removed commented code. The methods use a custom generator type reg_generator_t which currently matches categorical_generator_t. Only issue is the return value. return trainer.train(*this, generator, max_epochs); For regression, average loss seems like a good fit, but the trainer is treating it like a categorical generator, with classification error and nan loss. So the return value and the printed output is invalid, but as far as I can tell the training does work.
|
Correct. My only issue is the return value. |
|
Seems fine now 👍 What is you issue with the return value ? Seems correct to me. |
|
If no more issues, I'll merge this now. |
|
If you are interested, I added a few (dumb) tests for regressions: https://github.com/wichtounet/dll/blob/master/test/src/unit/reg.cpp You can also see the configuration of the error using MSE rather than CCE. |
…ssification, as a minor modification of fine_tune_ae().