Fahim Dalvi
Fahim Dalvi
CTranslate silently quits if the model being loaded was trained with the `-encoder_type dbrnn` option. I tried to run `cli/translate` under `gdb`, but it did not give additional information, other...
The original paper does not define a set threshold for the binary mask creation over neuron activation vectors, however, in our implementation is implicitly set using `average_precision_score`; we should clarify...
https://github.com/fdalvi/NeuroX/blob/f2314ccd964f744eb3209f1438fa53704a6665b2/neurox/interpretation/iou_probe.py#L37 assumes that y_train only has 0's and 1's. For multiclass, we should run IoU for each tag, and then combine into a global ranking.
Some tokenizers do not have `max_model_length` in their parameters, and truncation is turned off for these. However, the model still expects truncated inputs, so extraction fails in this case. Test...
Partial files (hdf5/json) as left as is when extraction of activations fails midway. An error is printed but might be missed by users and they may inadvertently use the files...
There are several hurdles when trying to use the latest source for libjingle and following the instructions in the README. - The node-gyp rebuild module looks for **libvoice_engine_core.a**, but this...
If we use the following code: > ``` > c.on( 'audio', function( evt ) { > // Audio from the local socket. > // Write it to a file. >...
Several BLOOM assets for news categorization results in null outputs
Post processing can be improved my mapping common names to ISO names