Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

AEC_exp #397

Merged
merged 68 commits into from
Aug 27, 2024
Merged

AEC_exp #397

merged 68 commits into from
Aug 27, 2024

Conversation

GuyPerets106
Copy link
Collaborator

  • Added AEC capabilities
  • Added AEC stats
  • Fixed calculate_size function in nerl_tools
  • Added experiments json files
  • Modified hugging face json file datasets indexes (IMPORTANT FOR EXISTING NOTEBOOKS)

Tests:

  • Passed full flow test locally
  • Passed NIF test locally
  • Passed Distributed & Federated experiments (as presented Aug 20)

@@ -1,5 +1,7 @@
#include "nerlWorkerOpenNN.h"
#include "ae_red.h"
#include <fstream>
Copy link
Owner

@leondavi leondavi Aug 23, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We don't need fstream and iostream
I guess it was for debug
Please remove

@@ -28,7 +30,7 @@ namespace nerlnet
void NerlWorkerOpenNN::perform_training()
{
this->_training_strategy_ptr->set_data_set_pointer(this->_data_set.get());

this->_training_strategy_ptr->get_loss_index_pointer()->set_regularization_method(LossIndex::RegularizationMethod::L2); // ! ADDED NOW
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

No need. Please validate that this is rebased on top of master.
In master there is already implementation for regularization.

_training_strategy_ptr->get_loss_index_pointer()->set_regularization_method(parse_regularization_loss_args(_loss_args_str));

int num_of_samples = _aec_data_set->dimension(0);
loss_val_tensor = std::make_shared<fTensor2D>(1, 1);
(*loss_val_tensor)(0, 0) = static_cast<float>(_last_loss);
(*loss_val_tensor)(1, 0) = _ae_red_ptr->_ema_event; // Mask the following lines to get reduction in data tranfers sizes, or Unmask to enable AEC stats
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Add TODO here that we should envelope it with if statement and pass model arg that controls it.

// Add _aec_all_loss_values to loss_val_tensor
for (int i = 0; i < num_of_samples; i++)
{
(*loss_val_tensor)(3 + i, 0) = (*_aec_all_loss_values)(i, 0);
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It looks wrong to go through all samples and them one by one.
Add TODO optimization here.

break;
*loss_values_return = mse2D;
_aec_all_loss_values = loss_values_return;
// cout << "MSE Loss: " << mse_loss << endl;
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

remove commented cout

_aec_all_loss_values = loss_values_return;
// cout << "MSE Loss: " << mse_loss << endl;
_ae_red_ptr->update_batch(loss_values_mse);
// cout << "AE_RED RESULT VECTOR:" << endl << *res << endl;
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

remove commented cout

*loss_values_mse = mse2D;
result_ptr = _ae_red_ptr->update_batch(loss_values_mse);
// string filename = "/tmp/nerlnet/predict_errors.csv";
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

remove commented code block or move it as a function of ae_red.

@@ -569,6 +598,11 @@ namespace nerlnet
}
curr_layer = curr_layer->get_next_layer_ptr();
}
// Write the model parameters to file
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

add a dedicated function that saves model parameters.
We should add this capability to NIF and API-Server at some point.

But remove this comment from here anyway.

src_cpp/opennn Outdated
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is maybe the reason for the crash.
We should carefully merge your changes to opennn. I think that Ori did changes as well.

fTensor2D diff = (*calculate_res - *_aec_data_set);
fTensor2D squared_diff = diff.pow(2);
fTensor1D sum_squared_diff = squared_diff.sum(Eigen::array<int, 1>({1}));
fTensor1D mse1D = (1.0 / static_cast<float>(_aec_data_set->dimension(0))) * sum_squared_diff;
fTensor2D mse2D = mse1D.reshape(Eigen::array<int, 2>({num_of_samples, 1}));
fTensor2D mse2D = mse1D.reshape(Eigen::array<int, 2>({(int)num_of_samples, 1}));
// cout << "MSE2D: " << mse2D << endl;
*loss_values_mse = mse2D;
Copy link
Owner

@leondavi leondavi Aug 23, 2024

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This can cause a memory issue. You try to put a tensor that will be deleted when this scope ends into a shared pointer. The shared pointer cannot take control of this memory block. This is a wrong usage of shared pointers.

_ae_red_ptr->update_batch(loss_values_mse); // Update thresholds

break;
*loss_values_return = mse2D;
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This line again


break;
*loss_values_return = mse2D;
_aec_all_loss_values = loss_values_return;
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't get what the purpose of this line. It is a switch between pointers but at least needs an explanation here.

@leondavi leondavi merged commit 0916ba0 into master Aug 27, 2024
1 check passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants