search

NVIDIA / DIGITS

Deep Learning GPU Training System
https://developer.nvidia.com/digits
BSD 3-Clause "New" or "Revised" License
4.12k stars 1.38k forks source link

Issues with home page angular tables #741

Closed lukeyeager closed 8 years ago

lukeyeager commented 8 years ago

Following up on https://github.com/NVIDIA/DIGITS/pull/728, here are some bug reports / feature requests:

lukeyeager commented 8 years ago

BUG: When you click+drag over several entries, you can select them all at once. But if you drag too far, and your "mouse-up" action happens outside of the table, then Angular doesn't catch it. Then it stays stuck in "selection" mode.

Fixed in #742

lukeyeager commented 8 years ago

REQUEST: When you select a dataset (or several), can you filter to only show the models that were trained on that dataset[s]?

lukeyeager commented 8 years ago

REQUEST: When a job finishes, move it from the top section down into the appropriate table.

lukeyeager commented 8 years ago

REQUEST: update count of jobs based on filters (next to "Datasets" or "Models").

Fixed in https://github.com/NVIDIA/DIGITS/pull/750

TimZaman commented 8 years ago

Oh man we really cannot pull this in in the current state. But looks to have great potential. There are only around 30 characters available in which one can describe his model, right? How does one distinguish between jobs that have run. At least the datasets's name should be in there, probably some more stuff regarding the settings. This is how it then looks in practice (see screenshot) - also do notice that we are already using abberations: TLsq/TLf; 'ResNet' instead of 'Residual(Neural)Networks' .. This whole thing makes the multiselect useless since one in practice does not know what it's selecting. I can understand that for the minor few that use digits for only a few jobs this is fine, but as a power user we are using hundreds and need a descriptive naming convention.

Indeed i could shorten the dataset name, but then again, there is too little space for naming and annotating. My suggestion is to do away with the two column approach; maybe use tabs?

edit: wait a second.. i cannot filter for things like accuracy and loss anymore in the bigger overview table.. What's going on here . If we make a full-width table we can fit in user selectable columns i guess, right?

image

enoonIT commented 8 years ago

I agree with @TimZaman , in this current state it's impossible to distinguish between jobs. At the very least I'd make the column width resizable on the fly by mouse dragging

TimZaman commented 8 years ago

Suggestion with using tabs: image

and

image

In the wide way there is a lot of room to add more lovely column-metrics

jmancewicz commented 8 years ago

The side by side tables are very tight especially with the width restrictions from bootstrap. The non-name fields are about as small as they can be while still being readable.

I had initially had a tabbed view, but moved to the side by side to attempt to keep close to the origianal look. I can put out a tabbed version to see how that works.

lukeyeager commented 8 years ago

BTW, the tooltips added with https://github.com/NVIDIA/DIGITS/pull/748 help with these issues:

There are only around 30 characters available in which one can describe his model, right? How does one distinguish between jobs that have run.

in this current state it's impossible to distinguish between jobs

TimZaman commented 8 years ago

Definately the new homepage is an amazing improvement, it looks a lot cleaner and more professional. I think the tab view i just shown is not the ultimate solution, but it is a fair one. I tried putting the datasets below the models, but that made less sense and the scroll longer. But I do feel like we should try to use more of the width, so i can include more metrics and really compare different outcomes, especially given the latest improvements in the sweep that was implemented. So lets say i do a sweep of 10 different settings; i will have no way at all of comparing them in one view. If we would have more space in the job-row table we could do that.

enoonIT commented 8 years ago

I'd like to suggest a couple of things: 1) the new tabbed view is nice, but I think people tend create to new models (and check them) more often than datasets - so I'd make the model tab the default one. 2) Model names for running networks are still hidden when too long. screenshot from 2016-05-20 18 22 28 It seems some of the horizontal space is wasted (after Status and Progress). In the meantime it would be nice to have tooltips, but they don't seem to work on the running jobs.

lukeyeager commented 8 years ago

Thanks for all the fixes and updates @jmancewicz! Closing.