Dominic Jack
Dominic Jack
This is a significantly simplified version of [this PR](https://github.com/google/gin-config/pull/26) I made ages back and was too busy to finish off, inspired by `keras`s [CustomObjectScope](https://www.tensorflow.org/api_docs/python/tf/keras/utils/CustomObjectScope). I'm not sold on the class...
Love this package. One thing I've been using is relative includes, and this is potentially buggy without doing dirty hacks to prioritize relative paths. Also added variable/path expansion, because I...
[`LayerNorm`'s uses the final dimension of inputs to define `param_shape`](https://github.com/deepmind/dm-haiku/blob/main/haiku/_src/layer_norm.py#L126). This is inconsistent with both [tensorflow's implementation](https://github.com/tensorflow/tensorflow/blob/v2.5.0/tensorflow/python/keras/layers/normalization.py#L1186), and my understanding of the original paper. haiku's implementation looks closer to [sonnet's](https://github.com/deepmind/sonnet/blob/v2/sonnet/src/axis_norm.py#L193)....
Compared to the [original repository](https://github.com/PetarV-/GAT/blob/master/utils/layers.py#L38), the [GATConv](https://github.com/danielegrattarola/spektral/blob/master/spektral/layers/convolutional/gat_conv.py) implementation has some minor differences. In particular, for the original implementation: - a different dropout is applied to each head before dense transformations;...
I've recently started looking at GNNs and wanted to build myself a decent benchmarking suite. Having found this repo (which is awesome - nice work!) I'm thinking it might be...
I found the supplementary material for the paper a while back, but can't seem to find it any more. Is it still available somewhere?
Is there a prefered way of citing this repository in academic articles? Could a .bib entry be provided either here or in the READMEmd?
See e.g. [this failed CI test](https://github.com/tensorflow/graphics/pull/286/checks?check_run_id=1608017643) (the PR does not change anything to do with the rasterizer). Based on a very quick google I suspect it may be a bazel...
* add `tf.keras.Model.fit` scipt (equivalent to `train`, for demo purposes) * specifies `axis` of a `squeeze` in model that was only running successfully because of a bug in preprocessing Test...
Not sure if these are wanted, where or how to name them, but I saw the `py_function` in the data pipeline and couldn't do nothing. Functionality is -almost- identical, though...