lihaoran

> Dear all, > > Contrary to previous comments (here + #54), I was able to reproduce all results from @tkipf's original paper using @philipjackson's implementation (see #25) of the...

Thank you buddy, @GuillaumeSalha! I can reproduce the results now.

I encounter the same issue using ds stage 3. Did you resolve this issue?

> My solution is to save checkpoints by myself or you can use [zero_to_fp32](https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/utils/zero_to_fp32.py) @mynewstart I found my converted ckpt `global_step_xxx` only contains meaningful `*optim_states.pt` but only empty `*model_states.pt`. Any...

> It's a bug in DGL implementation. Following the original paper, adding `nodes.data['iou']` is correct. Could you explain more about the bug? Is it related to the leaf nodes and...

You can sample from a standard normal distribution directly after training to generate new graphs

Sampling from a standard normal distribution : `self.z = tf.random_normal([1, FLAGS.hidden2])`. You don't need encoder to generate new graphs, only the decoder is enough.

Same question ...

Same problem when testing with py3.8, py3.9

@ChiYeungLaw May I ask why replacing 4 space chars with a tab before generation?