This page provides a changelog that lists new releases and updates to the PerceptiLabs products and services.
To upgrade to the latest version of PerceptiLabs, run:
$ pip install --upgrade perceptilabs
To install or upgrade to the latest nightly build of PerceptiLabs, run:
$ pip install --upgrade pl-nightly
- Added more logging to make sure trickier issues can be traced.
- The right-side panel in the workspace now hides correctly.
- Components no longer cover the full space when previews are hidden.
- Changing one preview no longer changes all of them.
- 1D and 3D Conv layers are now properly supported again.
- Unsaved models are now properly visualized in the Overview.
- A back-button has been added to the public datasets modal.
- There's now an easy way to leave requests for datasets in case you don't find one you are interested in.
- The overview interaction has now been improved so that:
- Datasets can be both unregistered and deleted. You can reach these options through right-clicking the dataset.
- Pressing the trash bin icon will delete models and unregister datasets (so no accidents happen).
- The Delete and Unregister confirmation popup now better describes what happens and always gives you the option to confirm or cancel.
- All "Kernel Offline" errors have been removed and replaced with more descriptive errors.
- Export fail messages now have the full traceback in them to see what exactly went wrong.
- The Viewbox in the Statistics View now remembers which tab you were at when changing between components.
- The DataType recommendation will no longer fail because of a bad or blocked external internet connection.
- The DataSettings are no longer removed if the model can't be created (for example if the name already exists).
- The RNN component now uses
- Going backwards and forwards in history of a model now works as it should.
- Redundant code line containing the number of neurons has been removed from the Dense component.
- Changing settings now always makes the model unsaved.
- Long dataset names in the Overview will not get cut off as easily any longer.
- Images are now visible when choosing what model type you want to build.
- Refreshing the browser during training now works properly.
- You can now load a folder instead of a CSV file. Whenever you create a project, you will now see three options Image Classification, Image Segmentation, and Multi-Modal.
- In Image Classification you get to choose a folder of folders, where each sub-folder represents a class.
- In Image Segmentation you get to choose two folders, one containing images and one containing masks.
- Multi-Model projects are non-image classification or image segmentation projects.
- If choosing an Image Classification or Image Segmentation project, you will automatically get the Input and Targets recommended, and the recommended pre-processing settings have been improved since we can be more specific.
- Whenever you start training, you will now see
model.summary()in the terminal.
- The progress bar has been removed from the workspace because it was breaking the design a bit.
- When starting training, you will automatically see the Overview tab as the first thing you see in the Evaluate View.
- Fixed an issue where the end results were sometimes missing after training completes.
- Errors in the Components no longer throw
ValueError: max() arg is an empty sequencebut will instead highlight the correct issue.
- Models will no longer vanish from the Deploy View after restarting the browser.
- Fixed an issue where Windows would occasionally crash because of faulty paths.
- The Kernel will no longer throw an error if you stop training before the first epoch is complete.
- If there is an error while loading data, the Data Wizard will no longer open and the dataset will not be added to the Overview.
- The Upload .CSV button will now be disabled after pressing on it once, preventing accidentally opening the file dialog multiple times.
- Model errors are now only shown on the relevant model's workspace.
- You can now find the tool version next to the logo at the top left.
- You can now export a model as a "PL Package" under the Deploy View. This exports a zipped package that you easily can share and load. Currently it only contains the model.json file but will soon be enhanced to contain checkpoints, statistics, etc.
- The "Load model" option now only accepts .zip files exported from PerceptiLabs.
- The Pause button now works as intended, whereas it previously did not properly pause the training nor change icon correctly.
- The Stop button will now always change the model status to "stopped" and will use an intermediate status of "stopping".
- The tutorials link now points to the correct YouTube video.
- Previews are no longer stored in model.json files, ensuring that they can't randomly go corrupt if the previews are too large.
- Pressing + New Model will no longer take you to the Load Dataset window in cases where a backend service isn't running.
- Upgrading PerceptiLabs will no longer fail because of a corrupt model.
- You can now load models into an existing dataset by selecting a model.json file (NOTE: this is in experimental mode).
- "Public" has now been added to the Component categories to distinguish Deep Learning Components from public Components.
- The CSV table along with pre-processing settings are now accessible from the Data Settings again.
- Custom Components no longer prevent Data Settings from opening.
- You can now remove datasets from the Overview by right-clicking on them and selecting Delete. In the case of local models, this will just remove them from the Overview, whilst for public datasets, this will also delete the dataset from your storage.
- You can now unregister models in the Overview by right-clicking on them and selecting Unregister. This will remove them from the Overview without deleting the model.
- The Confusion Matrix in the Evaluate View has been changed back to use absolute values instead of relative, as using relative values hid some important information.
- Models will now be saved in the correct locations â this used to cause issues where models appeared as deleted or didn't use the correct checkpoints.
- If your model recommendation crashes, an error will be displayed rather than the model becoming broken.
- Components will no longer place themselves on top of the settings on the right side if you are using FireFox.
- The Run Test button should now always be clickable.
- Fixed the issue where some systems were unable to start the tool because of the error Unable to upgrade your PerceptiLabs database that was caused of a Celery version issue.
- The frontend memory usage has been optimized.
- In case Gradio gets blocked and does not open in a new tab, you now get a popup inside the tool with the URL it's running on.
- The Max and Avg Pooling settings inside MobileNetV2 now works properly.
- The file explorer will now properly open on all systems.
- Tooltips from charts (e.g., in the Training View) will no longer be cut off when the tooltip extends past the chart.
- The Mime type error which used to prevent visualizations on some systems has now been fixed.
- Fixed the issue where the map in the Training View would zoom in and no longer show the entire model when training was stopped or completed.
- Fixed previews which randomly didn't load (but you could still see the preview if you pressed the Component).
- The Evaluate View will no longer be randomly empty after running a test.
- There's now a warning in the Data Wizard if you pick the Mask datatype, guiding you to set the pre-process settings correctly.
- In the Training View during training, the tab which was previously named the same as your target component is now named Overview.
- The status Waiting has now been changed to the more descriptive status Untrained.
- The Data Wizard view now scales well with many inputs.
- Pre-0.13 models will now run properly with the Gradio deployment.
- Segmentation tests are no longer available for Image targets.
- Running multiple tests as once now displays properly.
- The default theme of the app is now dark theme rather than light.
- Error messages now have a max size to their popup windows.
- New UI - PerceptiLabs has a brand new UI with this release, making it look sleeker than ever!
- Public Datasets - The Public Datasets feature allows you to download a variety of different datasets to use straight in PerceptiLabs. We are going to keep adding to these datasets and are also creating a Garden on the PerceptiLabs website. If you want to help the Garden (and Public Datasets) grow, you can check out details on how to do that here.
- New data-centric Overview - The Model Hub is renamed to Overview and now also contains information about which dataset the models were created from. Every time you load a new dataset into PerceptiLabs it's going to register as a new dataset which then can have multiple models created under it. This makes the models easier to keep track of and promotes a data-centric approach.
- Multi-class segmentation - With the Mask datatype being added, you can now create and train multi-class segmentation models! Just make sure to select the Mask datatype when you are loading your data and change the Image and Mask Resize options to match each other.
- Native File Explorer - The custom built File Explorer that was previously used to browse your computers files have now been replaced with a native OS one.
- Improved training speed and responsiveness - The training speed has been increased by about 60% and the responsiveness in the tool has generally increased. Things are running faster than ever before!
- Precision, Recall, and F1 score have now been added to the Training Statistics if you are running with a Categorical target.
- The Deploy View (previously export view) has received a face lift and some more functionality added to it. You can now also deploy as:
- FastAPI - This will export a FastAPI server and client script for you that you can use straight away, making it easy to start serving your model outside of PerceptiLabs.
- Gradio - We have integrated Gradio into PerceptiLabs! With this, you can spin up a Gradio serving with just a button click, which will open the Gradio app in a new browser tab for you to run your model on new data or to demo to other people.
- The pre-processing progress is now a bit more descriptive than before, letting you see how far it has progressed so you can judge roughly when it will be done.
- You will now see a link to get help whenever something gives an error.
- We changed up the workflow for creating a model so that you no longer need to go through the Training Settings every time.
- The random Partition now has a Seed option that lets you control the randomness.
- The Confusion Matrix now contains relative values instead of absolute to make it easier to see the performance on specific classes.
- The Input and Target components are now renamed to be the same as the Column name.
- The Variable dropdown on the components - This was a feature that allowed you to select which variable you wanted a Component to output. Cool idea in theory but confusing in practice and better handled through custom code in a Custom component.
- Weights toggle - This feature allowed you to toggle if you wanted the components to show the trained or untrained versions of themselves on the workspace. This feature will come back in an improved version later on, likely in the Evaluate view (previously Test view).
- Automatically downloading tutorial datasets â Since Public Datasets now let you choose which datasets you want to download and use, there is no need to automatically download them on installation.
- Multi-input models are now correctly recommended.
- Numerical inputs can now be merged.
- Code in components can now be updated more than once.
- Statistics can now be seen after restarting PerceptiLabs.
- The maximize button now works as it should in the Statistics view.
- Updating the Learning Rate now works as it should.
- Text data can now be loaded properly.
- Many many more smaller ones...
- Removed a bug which blocked some preprocessing from finishing if the data took too long to preprocess
- The export list is now scrollable
- You will now always see the preview of the components when you enter a workspace
- If you get a crash when starting the training, you will get an error which sends you back into the workspace and tells you what went wrong
- The Jupyter Notebook export option has been temporarily hidden to remove confusion, until the export is working again
- Fixed a bug where models were not loaded correctly at login
- The pre-processing pipeline is now exported with the the model
- You are now required to enter a path before exporting a model (so it has somewhere to export to)
- Partially-trained models now show up in the export list
- âConvolution components without a connection will no longer make the model look like it still works
- Fixed the case where if there was an issue with the connection, the questionnaire couldn't be closed.
- Trained models will now appear in the export view.
- The kernel status icon in the bottom right of the workspace now updates appropriately.
- Changing settings of a component will now always take effect, which it didn't before in some cases.
- There is now only one Dice loss
- The Kernel status at the bottom right of the workspace will no longer be permanent red
- The Kernel no longer seemingly stops responding randomly
- Only trained models can now be seen in the testing view
- Segmentation tests now have the PerceptiLabs theme
- Tests are now using caching and will therefore be a lot faster for large datasets or datasets that use pre-processing
- The scrollbar in the settings will now always go down to the bottom, regardless of screen resolution
- Tests will no longer randomly disappear right after appearing, which used to happen if you ran a larger test
- Segmentation test image visualizations are no longer mirrored
- Quantized export has now been re-enabled. Note: It's restricted to models with a single input and single target
- The login questionnaire has been re-enabled
- Previews in the workspace are now calculated much faster
- Training will now start faster than before in case you are using any pre-processing
- UNet has been added, which includes:
- A UNet component
- Improved segmentation visualizations
- Two Segmentation tests: A Segmentation metrics table and an image slider that let's you see the five best and worst predictions
- One new dice option called "Keras Dice" which is calculated slightly differently
- You can now update the "Data Settings" on the workspace, which includes:
- Changing Data Types
- Changing pre-processing
- Changing partition
- Changing data path
- Image Target will no longer throw an error in the Data Wizard
- Stopping training will no longer cause a crash if stopped at the wrong time
- Labels are now shown as text during training, making it easier to see which class is which.
- The Softmax component has been removed because of redundancy.
- If you stop training and start it again, you will now see it continue from the previous statistics if you choose to start from the last checkpoint.
- The error messages in the Data Wizard have now been improved to give a hint as to what may be wrong.
- Random Partition now uses a fixed seed.
- Compressed models can now be exported again without crashing.
- Preview generation in the workspace is now faster.
- Pre-processing has been improved by pre-generating the samples rather than generating them on the fly.
- Backend requests have been cleaned up and should be more responsive in general.
- Crashes for very small (1x1) previews will no longer occur.
- Improved recommendations for Categorical Decoders.
- The questionnaire is now visible when you first start the tool.
- You can now select "Do not use" as an Input/Target type, which will ignore that column when creating the model.
- Added three new datasets that will automatically download when you start the tool the first time: "Human Activity", "Wildfire" and "Covid19".
- All CSV filetypes should now be readable in the Data Wizard, whereas before some of them caused strange symbols to appear which caused an error.
- A visual bug causing the entry boxes in the Training Settings to be too large has now been fixed.
- A new onboarding experience has been added.
- We have now upgraded to using TensorFlow 2.5! Make sure to upgrade your CUDA to 11.2 and cuDNN to 8.1 to be able to use GPUs with this version.
- The Accuracy, Loss and other line charts are no longer empty before the second epoch starts.
- You can no longer set height or width to 0 in the Reshape pre-processing in the Data Wizard.
- "Kernel Offline" should no longer happen, instead, error messages will be much more precise as to the cause of the issue.
- The error messages in the workspace and the Data Wizard have been improved.
- The numerical target statistics view (often what you use for regression) has now been improved.
- The pre-processing UI has been updated.
- Onboarding will now be visible when you first enter the tool, rather than only appearing after selecting the ModelHub in the left menu.
- The Binary data type has been temporarily disabled because Categorical is currently better in every way.
- Fixed a dependency issue in the frontend that was causing strange behaviours such as being unable to delete models, test popup not showing up, a project modal showing up, etc.
- Test loading message is now complete.
- Image Targets now have an updated look in the statics view during training.
- There is now a loading indicator after pressing Run Model or Customize in the last stage of the Data Wizard.
- A legend has been added to the Predictions for Each Class chart.
- The resource section in the top right has been improved.
- You can now access our community sites like YouTube, Forum, Slack and GitHub through the top-right icons.
- In the Training Settings, you can now choose to automatically save the weights of your model every epoch, rather than just at the end of the training.
- There is now a restriction on creating models with only a single Target column, until we implement better support for multi-modal problems.
- Text inputs are now possible to use; they will use bag-of-word as the preprocessing method.
- The UI for the test view modal has been updated.
- A Python 3.8 distribution is now included.
- PerceptiLabs will no longer install (or prompt you to install) gym['atari'].
- Ctrl+V now works properly in the custom code.
- Tests are no longer interrupted by changing views.
- The Categorical column selection in the Data Wizard will now always be an option, even if it's not recommended.
- A lot of unnecessary kernel calls have been removed.
- Extensions of your loaded data are now case insensitive.
- When refreshing the tool, you will now open the model you were viewing rather than the tool going through all of them.
Note: this version has been removed due a critical error.
- Loading messages for tests have been improved so you can now see their progression.
- The training statistics have now been updated for categorical targets during training; you can now see:
- Loss per epoch
- Accuracy per epoch
- Prediction versus ground truth for a single sample
- Predictions for each class summed over all samples in that epoch
- A deleted model will now stay deleted, rather than randomly coming back.
- It's now no longer possible to press the Open button in ModelHub without having any models selected to go to a bugged workspace.
- A few preprocessing options have been added: Random cropping. Random rotation.
- We have added a direct messenger to the tool, so you can easily get in contact with us. You can find it under the ? -> Send us a Message.
- The Kernel will no longer crash on certain learning rates.
- Fixed a backwards compatibility issue with versions earlier than 0.12.4.
- There is now a loading message when running a test.
- A few preprocessing options have been added:
- Flipping images
- Resizing images
- Min-Max and Standardization are now both Normalization options
- Data Wizard now loads faster after selecting a file.
- Data Wizard will no longer skip past steps from clicking on Continue too many times.
- Binary data now works with Yes/No and yes/no as well as True/False, true/false, 1/0.
- Multiple columns now work inside the tool, although some manual work is still required to put the model together.
- You can now pre-process your data inside the Data Wizard by pressing the little symbol on top of the column. The pre-processing available so far is Normalize for Image and Numerical data.
- The data types in the Data Wizard are now filtered on what is possible for each column. For example, if a column contains a path then it can only be Image type (and later on Text).
- Fixed a common error message "getBoundingClinetRect" which sometimes stopped the workspace from functioning as it should.
- The components in the map will no longer have eternal spinners if you start training straight from the Data Wizard.
- Binary data now works as intended. Fixed a bug where data types sometimes were not recommended.
- Fixed an issue where the Next button sometimes was disabled.
- Training progress will now not be accidentally shared between models.
- Reshape settings have now been fixed.
- Training results will no longer randomly show up after already having interacted with it.
- A data type will now be automatically recommended for each column.
- Tests for a model are now removed when that model is deleted.
- The Output component has now been properly renamed to Target.
- The questionnaire will no longer get stuck where you are unable to close it by pressing Confirm.
- The Reset component button has been removed from the Input and Target components as they have no settings to reset.
- The Create model UI has been updated to look a bit cleaner.
- The dropdown on the components in the workspace now works.
- Data settings can no longer be randomly copied over to another model (this used to cause the model to use the wrong input and target data).
- Training can now be started in incognito mode.
- Fixed progress bar during training; it now reaches 100% every time.
- Pre-trained components now automatically use their preprocessing function to ensure that data input to them is correctly formatted.
- Results are now seen in the Training Completed popup.
Note: this version has been removed due a critical error.
- The old Create Model popup together with the template has been removed. This means that models created pre-0.12 will no longer be usable after upgrading to 0.12.0.
- A new featured called Data Wizard has been introduced. The Data Wizard provides a more structured way to load your data, making the models data centric. It works by:
- allowing you to load a CSV file.
- allowing you to define the CSV file (specifying which columns are the input and target, and their types).
- providing you with a recommended model that can be run immediately.
- There are currently four supported column types in the Data Wizard: Numerical, Categorical, Binary, and Image, allowing you to mix them to build models such as: Image Classification, Regression, Segmentation, simple Image Generation, Autoencoder, etc. Read more about the Data Wizard here. Note that this link will become available once we document this functionality.
- Object Detection has been temporarily removed.
- GAN has been temporarily removed.
- Pre-trained models have been added as components under the Deep Learning, including VGG16, ResNet50, InceptionV3, and MobileNetV2.
- Data components have been removed and replaced by Input and Target components.
- Training components have been removed. The training engine is now running behind the scenes and has been generalized to work for all above mentioned cases.
- The Statistics View has received a slight update, where it now shows statistics based on the Inputs and Targets rather than the Training components.
- The Test View has been reworked. You can now run one (or many) tests on one (or many) of your models. The current tests available are: Confusion Matrix and Metrics Table.
- Changing the model path will no longer cause saving checkpoints to crash in Windows.
- Zooming by inputting a number will no longer cause the components placement to change.
- The correct TensorFlow version (2.4.1) is now shown in the bottom right.