Gavia Gray

Hey, I wrote some notes replicating and loading the pretrained model here: https://github.com/gngdb/human_body_prior/tree/master/src/human_body_prior/train

@nghorbani could you tell me if the metrics I got running the validation set are correct? They were: ``` [V02_05] -- Epoch 0: val_loss:0.51, v2v:0.25, kl:3.52, geodesic_matrot:2.17, jtr:0.26 ```

Here's a workaround, it also deals with some other issues when trying to load using the SMPL body models that are on the website (some arrays are actually `chumpy` or...

Neither the EGL or OSMesa notebook exist anymore, [links in the documentation](https://pyrender.readthedocs.io/en/latest/examples/offscreen.html#google-colab-examples) lead nowhere. Edit: there's a link to the EGL notebook in the repository README, sorry, I guess the...

The README says 4 days on 8xA100s, which is enough info to estimate the MFU. Copying code out of other parts of the repo: ```python n_layer = 12 n_head =...

Honestly, the easiest thing to do is going to be add a step in your bash script to move the output dir after each iteration so you don't have to...

Same problem, thanks for the fix!

Oh yeah, and I could just be calculating the gradient variances wrong; am currently using: ``` python T.var(T.grad(expressions.loss_train,l_hidden.W)) ``` Should probably be: ``` python T.mean(T.var(T.grad(expressions.loss_train, l_hidden.W), axis=0)) ``` Repeating with...

Unsurprisingly, initial results suggest that's not going to fix it.

Comment from talk: we could be having problems with normalisation in this calculation; should normalise in some way.