You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Correct the predict in AutoEncoder to consider non-vector layer outputs #68
Fixed in 17e9096: the AutoEncoder now computes the cached encodings, predict and inverse_predict on the whole batch through the network's own (batched) predict, instead of looping over the layers and the rows, so inner layers can have outputs of any shape (e.g. ConvLayer or PoolingLayer). A regression test with convolutional and pooling layers in both the encoder and the decoder has been added in 7665bfa.
Note that the encoded data (the output of the last encoder layer) must still be one-dimensional for each record, as predict returns a (records × features) matrix: end the encoder with a ReshaperLayer if needed.
Activity
sylvaticus commentedon Nov 3, 2026
Fixed in 17e9096: the AutoEncoder now computes the cached encodings, predict and inverse_predict on the whole batch through the network's own (batched) predict, instead of looping over the layers and the rows, so inner layers can have outputs of any shape (e.g. ConvLayer or PoolingLayer). A regression test with convolutional and pooling layers in both the encoder and the decoder has been added in 7665bfa.
Note that the encoded data (the output of the last encoder layer) must still be one-dimensional for each record, as predict returns a (records × features) matrix: end the encoder with a ReshaperLayer if needed.
Minor doc edit to add explicitly GPU support