CoordConv
CoordConv copied to clipboard
I get zero accuracies by running the instructed command: ``` # coordconv version python train.py --arch coordconv_rendering -mb 16 -E 100 -L 0.005 --opt adam --l2 0.001 -mul 1 --use_mse_loss...
Hi, in your code you normalize xx_channel and yy_channel as following: https://github.com/uber-research/CoordConv/blob/27fab8b86efac87c262c7c596a0c384b83c9d806/CoordConv.py#L70 https://github.com/uber-research/CoordConv/blob/27fab8b86efac87c262c7c596a0c384b83c9d806/CoordConv.py#L71 I think it should be: `xx_channel = tf.cast(xx_channel, 'float32') / (self.y_dim - 1)` `yy_channel = tf.cast(yy_channel, 'float32')...
…s is reversed. As a result, the coordinates doesn't belong to interval [-1,1] when x_dym != y_dim.
In Python 3, dict.keys doesn't return a list, but something like a set that doesn't support indexing. To make indexing possible, it is necessary to convert that to a list.
Hi, This is a very minor issue--nevertheless, the rest of the repository working, I really wanted to have this fixed. This repository code at generating data is not python3 compatible.
Hi, I'm wondering where exactly did you put a coordconv layer for a2c algorithm. Is it before the input? The paper states 'Adding a CoordConv layer to an actor network...
I came across this issue, trying to apply CoordConv to a non squared window. Seems like the dimensions are swapped.