Scott Sievert
Scott Sievert
**What does this PR implement?** This PR allows skipping the bracket that trains all models to completion if `bool(patience)`. This bracket is likely of marginal use because the hyperparameters will...
**What does this PR implement?** This PR prioritizes model training/scoring based on the most recent score in model selection.
When I came across this package, I was glad to see close integration with optimization algorithms. I wish there was support for `CDClassifier.partial_fit`. This would allow the user to have...
Shouldn't this code work? ``` python from sklearn.datasets import make_classification from scikeras.wrappers import KerasClassifier import tensorflow as tf def model(): model = tf.keras.Sequential() model.add(tf.keras.layers.Dense(8)) model.add(tf.keras.layers.Dense(1)) return model X, y =...
As discussed in #218, it's hard to choose the loss automatically (categorical cross entropy loss is best for independent labels, binary cross entropy is okay for multiple labels). Motivation for...
A recent example has confusion around the input arguments. This is the code in question: ``` python class Autoencoder(...): def score(self, X) -> float: reconstructed = self.predict(X) return binary_accuracy(X, reconstructed).numpy().mean()...
**What does this PR implement?** It adds `loss="categorical_crossentropy"` to KerasClassifier by default. It adds some protection in case the user passes multiple classes in `y` but the model is only...
Why does KerasClassifier not have a default loss? https://github.com/adriangb/scikeras/blob/d83bffaf8497289d9755f42654bdcc15584e68e6/scikeras/wrappers.py#L1229-L1231 Shouldn't there be a least a sensible default that works in most use cases? **Reference issues/PRs** https://github.com/dask/dask-ml/pull/794
Currently, the number of outputs of the Keras model is checked: https://github.com/adriangb/scikeras/blob/9f633f5c21a45b43986a8d779c3ef9e3005038bb/scikeras/wrappers.py#L418-L419 It'd be nice if the number of inputs could be checked too (and their shape/dtype). This should be...
Right now this library is tied to NumPy arrays pretty heavily. Will this library work with [CuPy][2] arrays? CuPy arrays are NumPy arrays for CUDA GPUs and are nearly a...