New answers tagged

2

Seems quite strange. I bet you've got a bug or test set size is small enough. Like if there're only 2 objects: 1 positive and 1 negative, you could have classified them correctly even with a random classifier. I assume your test set is not that small, which means there's a bug somewhere. Also, in theory, if both of your sets are large enough and are drawn ...


0

Less loss means low bias on training set. It always recommended to first aim for a model with low bias so you go and choose "loss_exp-resLayer10". It would have been better if we've loss for validation set because we can't assess the "overfitting". So in case, if your chosen model doesn't perform well on test data then use regularisation ...


1

You can't. If the model is trained with 6 features, it means that this model is like a function which requires 6 arguments. For instance the model might calculate the answer like this: answer = 0 * f1 + 1 * f2 + 0 * f3 + 5*f4 + 0.5*f5 +10*f6 Obviously there's no way to know the answer of this function without knowing all its arguments. Another way to look ...


1

The general rule of thumb is to run the number of epochs until validation error starts to increase. Sometimes fast initial learning will not lead to the best performance later.


1

Something similar to what you describe is frequently used in some domains and it is called gradient accumulation. In layman terms, it consists of computing the gradients for several batches without updating the weight and, after N batches, you aggregate the gradients and apply the weight update. This certainly allows using batch sizes greater than the size ...


0

The problem with having overlapping examples between Validation set and training set is the purpose of using a validation set is to ideally tune hyperparameters for your model and to have overlapping examples would mean that since your model has already been trained on this overlapping data, it would ideally have a greater probability of predicting the ...


3

You want to manually label some cases and then extend that "manual labeling" to the rest of the data. This is a supervised learning excercise with prior manual labeling by you. Let's suppose you have partitioned a random, suitably sized training data set. Now you need to model a classification algorithm via the classical modeling pipeline and use ...


2

The problem you are talking about is unsupervised sentiment analysis. You can try: VADER: It gives the polarity of the sentence based on which you can tag your training data. But this library has certain limitations - it can't sense sarcasm and sometimes the accuracy is not that great. But for initial understanding, you can check this library. Text Blob - ...


1

Have you looked into focal loss? The idea seems to be similar to what you are describing - If predictions (~0.8) is close to GT Label of 1, it does not add to the loss value.


1

In the simple example you’ve given, the outcome should be similar. Not the same because default Keras batch size is 32 and whereas in the first formulation that would mean 32 usable training examples, in the second the model could only benefit from the percentage of the 32 that you allow the loss function to see. I would prefer the first option where the ...


1

1.Technically they are the same.They will compute the same loss. Although there will be a subtle difference in performance. In the first case gradient will only be calculated for the subset but in the second case gradient will be calculated for the whole training set but only the gradient of subset will be used to compute cost and update. 2.No, the points ...


0

You have to use the same CountVectorizer instance on all data and have a method to handle out of training sample tokens.


1

When you are working with time-series data, the most recent data captures the most relevant information possible, so it is more prudent to include them in training data. So a more prudent decision would be to opt for Roll-Forward Partitioning. Roll-Forward Partitioning: We start with a short training period and we gradually increase it, at each iteration of ...


Top 50 recent answers are included