Features Requests

At present, time series can only be provided as columnar data in an imported CSV. Please would you update so that each row can represent a sample, and provide suitable time-series support capabilities such as * Parsing time-series time/datestamps with ~ms/us resolution * Automatic re-sampling for uneven time series (average and interpolate) * Automatic differentiation w.r.t - since this can, sometimes, be an alternative to resampling since dx/dt is by definition per unit time Anyone else have specific TS needs?

Despite a certain degree of control over random number generation in various Python packages (see below) TF training on GPU is said not to be 100% reproducible because of uncontrolled operation order over parallel CUDA cores. Being able to force training to occur on CPU would increase reproducibility of results, albeit at the cost of longer training times; however, that is a choice for the user to make. def resetRandoms(rSeed = commonSeed): # 0. set PYTHONHASHSEED per https://stackoverflow.com/questions/58067359/is-there-a-way-to-set-pythonhashseed-for-a-jupyter-notebook-session/61953451#61953451 # This https://stackoverflow.com/questions/54865930/how-to-set-pythonhashseed-for-jupyter-notebook says setting the env from within a notebook is too late... # but we'll keep that code, just in case # 1. Set PYTHONHASHSEED environment variable os.environ['PYTHONHASHSEED'] = str(rSeed) # NB this key does not exist until this is executed # 2. Set python built-in pseudo-random generator pyrandom.seed(rSeed) # 3. Set numpy pseudo-random generator np.random.seed(rSeed) # 4. Set the tensorflow pseudo-random generator; see also https://www.tensorflow.org/api_docs/python/tf/random/set_seed tf.random.set_seed(rSeed) # for later versions: # tf.compat.v1.set_random_seed(rSeed)

ResNet18 - and many other models - do not use 'glorot_uniform' as the kernel initializer. Whilst it is possible to set alternatives, such as 'he_normal' via custom code, 1/ This is far more time consuming and error prone than choosing a value from a dropdown list 2/ The component is then "non-standard" - values previously set by the settings panel are invisible and inaccessible - this makes it much, much harder to review a model for correctness since one cannot do so by inspecting the settings 3/ Resetting the component then wipes out customisation whether set by the panel or not; resetting may be deliberate - but it can also be accidental as there is no confirmation dialog associated with component reset Resolution sought: add kernel_initializer to settings panel with the full range of TF presets available as options in a drop down menu. Bias initializers are very rarely used and are not a priority. There are however other arguments that should also be considered for inclusion, such as kernel_regularizer (e.g. for weight decay). Ideally,. a general purpose mechanism that allows the full range of TF arguments to be specified by the user without coding would be welcome. I note however that the TF source code for convolutions (see https://github.com/keras-team/keras/blob/v2.8.0/keras/layers/convolutional.py#L533-L691 ) shows that type hinting is not used, therefore it is acknowledged that a general purpose mechanism may not be possible at this time.

The dataset ive been developing defines things using the polygon tool inside vgg for image segmentation. Being able to load in multiple polygons per image is about all that keeps perceptilabs from being perfect for my use case (even if I have to do some data transformation into your preferred csv format)

Whilst I believe there is >1 request for more info to be displayable on components, it would also be rather handy to be able to find/select components by settings e.g. * Components with dropout on (or numeric relation? >0, >=0.3, <0.2?), maxpooling, ... * Convolution stride * Activation * Etc. As this would make implementing and reviewing larger models significantly easier. It might not be hard to do either: ?reuse settings panel for a "search/select" dialog?

PerceptiLabs makes good use of the browser's database/cache. However, when it is necessary to use PL with an "empty" or null/initialised cache (e.g. to establish a known PL state & initialisation actions for fault finding, or robustness checking, or...) it is necessary either to use a private/incognito browser window (with independent cache)or to clear the existing cache, which requires careful control if other sites, etc. are not to be unintentionally affected. It would therefore be very useful if PL offered a single button control so reset its data structures in the browser cache; the user should of course be required to confirm the action after being presented with a warning that resetting could take some time and that PL might not be usable until e.g. (current experience on V0.13.9) pre-processing of all data has been run again. If the PL internal architecture would permit selective clearing of browser data structures per model/project that would be even better - as long as the user also has an option to reset all.

It would be most helpful for larger models if one could have component and settings panels that could be undocked, resized, etc. so that key aspects were always visible, or always accessible (such as, as much of the component list as one can fit on screen) At present the component list is awkward with larger models, and the settings don't show all settings at once, requiring a lot if interaction to access information that is typically needed (but genuinely needed ) only momentarily, e.g. to check a value. Panels might be "torn off" or have a dock/undock control. Tabbed panels might also help make better use of the limited screen area available, eg. tab for components, component settings etc. (name of selected component on tab 1 being used for related tabs). Data settings could then become a tab, further unifying the UI.

One can of course see the effect of each feature on the component preview, with the aid of the < > navigation. However, there can be a lot of filters and that's not an efficient way to examine them. It would be very nice to be able to have a pop-up window (resizable, scrollable, non-modal) showing the effects of all features in a grid, and to be able to save the image. It would be doubly nice if the raw image were also included in that view, and if the previewed image sizes were also adjustable, from the default scaled down version used in standard previews to full-size (of the source image)

Provide sorts for longer component lists: * Alphanumeric * Datetime added to diagram * Distance from top left

It would be really neat to be able to annotate components directly without “modifying code” with a little comment indicator and text display on mouseover This could be used to explain settings, design, anything that the user feels is important and it would not clutter the model with additional boxes (as my current empty custom component does).