numer.ai | Validation and prediction code for numer.ai | Machine Learning library

Based on the paper you shared, it looks like you need to change the weight arrays per each output neuron per each layer. Unfortunately, this means that the implementation of your optimization routine is going to depend on the layer type, since an "output neuron" for a convolution layer is quite different than a fully-connected layer. In other words, just looping over Flux.params(model) is not going to be sufficient, since this is just a set of all the weight arrays in the model and each weight array is treated differently depending on which layer it comes from.

Fortunately, Julia's multiple dispatch does make this easier to write if you use separate functions instead of a giant loop. I'll summarize the algorithm using the pseudo-code below:

What you could do in this situation is to iterate on the validation set(or on the test set for that matter) and manually create a list of y_true and y_pred.

This is more of a comment, but worth pointing out.

The reason in general is indeed what talonmies commented, but you are summing up the numbers incorrectly. Let's see what happens when tensors are moved to GPU (I tried this on my PC with RTX2060 with 5.8G usable GPU memory in total):

Let's run the following python commands interactively:

No, they aren't comparable.

Your baseline model used X_train to fit the model. Then you're using the fitted model to score the X_train sample. This is like cheating because the model is going to already perform the best since you're evaluating it based on data that it has already seen.

The grid searched model is at a disadvantage because:

1. It's working with less data since you have split the X_train sample.
2. Compound that with the fact that it's getting trained with even less data due to the 5 folds (it's training with only 4/5 of X_val per fold).

So your score for the grid search is going to be worse than your baseline.

Now you might ask, "so what's the point of best_model.best_score_? Well, that score is used to compare all the models used when searching for the optimal hyperparameters in your search space, but in no way should be used to compare against a model that was trained outside of the grid search context.

So how should one go about conducting a fair comparison?

1. Split your training data for both models.

You should try this Google Notebook trouble shooting section about 524 errors : https://cloud.google.com/notebooks/docs/troubleshooting?hl=ja#opening_a_notebook_results_in_a_524_a_timeout_occurred_error

Turns out its just documented incorrectly.

In reality the export from brain.js is this:

You're right. Just one thing to consider for choosing OrdinalEncoder or OneHotEncoder is that does the order of data matter?

Most ML algorithms will assume that two nearby values are more similar than two distant values. This may be fine in some cases e.g., for ordered categories such as:

• quality = ["bad", "average", "good", "excellent"] or
• shirt_size = ["large", "medium", "small"]

but it is obviously not the case for the:

• color = ["white","orange","black","green"]

column (except for the cases you need to consider a spectrum, say from white to black. Note that in this case, white category should be encoded as 0 and black should be encoded as the highest number in your categories), or if you have some cases for example, say, categories 0 and 4 may be more similar than categories 0 and 1. To fix this issue, a common solution is to create one binary attribute per category (One-Hot encoding)

Increasing the dimension of a trained model is not possible (without many difficulties and re-training the model). The model you are using was pre-trained with dimension 768, i.e., all weight matrices of the model have a corresponding number of trained parameters. Increasing the dimensionality would mean adding parameters which however need to be learned.

Also, the dimension of the model does not reflect the amount of semantic or context information in the sentence representation. The choice of the model dimension reflects more a trade-off between model capacity, the amount of training data, and reasonable inference speed.

If the model that you are using does not provide representation that is semantically rich enough, you might want to search for better models, such as RoBERTa or T5.

You could build a model like this.

x = features you have. y = true_lable

from that you can extract features importance. also, if you want to go the extra mile,you can do Bootstrapping, so that the features importance would be more stable (statistical).