Weightslab
A tool to develop models between BlackBox and Whitebox
7 followers
A tool to develop models between BlackBox and Whitebox
7 followers
Build your models anywhere between the black box and the white box with full control, flexibility, and scale. People trade clarity for speed, but then repeat the iterations many times because of the lack of clarity. We believe that gaining greater clarity and more granular insights during the training of AI models is key to making development cheaper and more trustworthy.
How granular can I actually get with these insights during training, and does it slow things down noticeably on larger models?
@kadersusuzlu Right now, we cover the data part, so you trace and attach different signals (numbers) produced by the models to the individual samples, so you get better insights into the loss landscapes. Because of this, it does not slow down one forward pass by more than 5%.
How granular can I actually get with the insights during training, is it per-layer per-step or more of a higher-level summary?
@abdulsametvgoc At the moment, the supported granularity refers to per-sample (e.g., one image in a detection setup) or per-instance (one bounding box in a detection setup) stats and signals. Very soon, we will have per-layer and potentially per neuron/head stats as well.
How granular can the insights actually get during training, like per-layer behavior or something deeper, and does that come with a noticeable hit to training speed?
@keremkorni1924 You can trace individual samples, or even individual segmentation maps or bounding boxes, depending on the task. As we capture these signals and make them available for processing, the training overhead is not noticeable.