The Evaluation.ipynb notebook had its content duplicated in my last PR by mistake, so I removed the extra copy at the end and also added an outdir parameter to save the reconstructions into (useful for tracking multiple versions of CLIP image, text, image+text, etc).
For the train_voxel2clip.py script, I updated it to work with text CLIP by bringing in the change made to the e2e script. In the validation loop, I think the mixco parameters were mistakenly being passed in to the utils.mixco_nce loss, so I took those out since it was throwing exceptions.
The Evaluation.ipynb notebook had its content duplicated in my last PR by mistake, so I removed the extra copy at the end and also added an outdir parameter to save the reconstructions into (useful for tracking multiple versions of CLIP image, text, image+text, etc).
For the train_voxel2clip.py script, I updated it to work with text CLIP by bringing in the change made to the e2e script. In the validation loop, I think the mixco parameters were mistakenly being passed in to the utils.mixco_nce loss, so I took those out since it was throwing exceptions.