AIDeveloper Versions Save

This is an update for AIDeveloper 0.2.0. Please first install AIDeveloper 0.2.0. Within AIDeveloper 0.2.0 go to the Menu 'Help'->'Check for updates' and use the dropdown-menu on the right to select an update file. After clicking 'Download + Install' restart AIDeveloper and the voila, the update is done.

Bugfixes

• icons (art) missing after update
• Major improvements on Plotting tab (tools for showing 0,1,2 dimensional data)

This is an update for AIDeveloper 0.2.0. Please first install AIDeveloper 0.2.0. Within AIDeveloper 0.2.0 go to the Menu 'Help'->'Check for updates' and use the dropdown-menu on the right to select an update file. After clicking 'Download + Install' restart AIDeveloper and the voila, the update is done.

Bugfixes

• Importing jpeg/png did not work (when using cropping) #46

This is an update for AIDeveloper 0.2.0. Please first install AIDeveloper 0.2.0. Within AIDeveloper 0.2.0 go to the Menu 'Help'->'Check for updates' and use the dropdown-menu on the right to select an update file. After clicking 'Download + Install' restart AIDeveloper and the voila, the update is done.

New features

• New padding options: 'delete', 'alternate' 36404435b7da68de0513431af56fba5b6dcd32b5

Bugfixes

• Altering padding mode had no effect 2b92a45c281535f21d3132e578e65d17e512d4cc
• Clearer names for padding options 032c08fbc9b8b9a0cacbae7fed58311a1b0757f7 0e07346d0173e25103993a88060cacf7012d1238 ba963140cafd0f16ecf20dec3863522101a5a58d
• Remove scipy.ndimge.zoom (use cv2.resize instead) c9f358ff0d207a7ba86b4bb48df1fd55d1911a6a

New rapid update feature! Within AIDeveloper go to 'Help'->'Check for updates' to download and install updates. After that you only need to restart AIDeveloper.

Limitation of the GPU Version: Only NVIDIA GPUs with a compute capability>3.7 are supported. What is my compute capability? AID was tested on multiple Systems with different GPUs. On most systems, AID only worked after the 2nd start. If AID crashes or does not detect your GPU, you may need to update your GPU driver. If you still have issues with the GPU based version, you may go for the CPU based version. Mac Compatibility: Only MacOS>=10.15.3

New features

• #43 #42 Went from NumPy to OpenCV for padding and zooming operations (https://github.com/maikherbig/AIDeveloper/tree/master/Tutorial%20Deploy%20to%20OpenCV%20dnn)
• #44 Reviewed image processing pipeline: images are no longer dropped if object is too close to the edge of the image (applies only for RT-DC data)
• #40 transfer learning: when loaded model has less classes than required for new data, nodes are added correspondingly

Bugfixes

• #38 Precision, Recall were not shown correctly
• #39 Allow saving meta files without specifying the model

Limitation of the GPU Version: Only NVIDIA GPUs with a compute capability>3.7 are supported. What is my compute capability? AID was tested on multiple Systems with different GPUs. On most systems, AID only worked after the 2nd start. If AID crashes or does not detect your GPU, you may need to update your GPU driver. If you still have issues with the GPU based version, you may go for the CPU based version. Mac Compatibility: Only MacOS>=10.15.3

LR screening

• more efficient data management
• plot (val.) loss, 1st derivative of (val.) loss, (val.) accuracy, 1st deriv. of (val.) accuracy

Model training

• more efficient data management
• when directory becomes un-available (drive unmounted, server connection lost...), continue saving to a temp. folder

Loading and saving sessions

• Load session: all hyper parameters can be restored automatically
• Save session all hyper parameters are saved

New pre-trained models

• ResNeXt 50
• ResNeXt 101

Bug fixes

• Adamax optimizer
• ResNet v2
• Checking for updates of AID

Limitation of the GPU Version: Only NVIDIA GPUs with a compute capability>3.7 are supported. What is my compute capability? AID was tested on multiple Systems with different GPUs. On most systems, AID only worked after the 2nd start. If AID crashes or does not detect your GPU, you may need to update your GPU driver. If you still have issues with the GPU based version, you may go for the CPU based version. Mac Compatibility: Only MacOS>=10.15.3

Additions to the "Expert" menu:

• Learning rate screening (to find optimal learning rates)
• Learning rate schedules
  • Exponential decrease
  • Cyclical learning rates (https://arxiv.org/abs/1506.01186)
• Advanced settings for optimizer
  • beta, momentum, etc. for all optimizers (SGD, Adam, RMSprop,...)

As usual, those settings are tracked in the meta-file and can be changed during training.

Limitation of the GPU Version: Only NVIDIA GPUs with a compute capability>3.7 are supported. What is my compute capability? AID was tested on multiple Systems with different GPUs. On most systems, AID only worked after the 2nd start. If AID crashes or does not detect your GPU, you may need to update your GPU driver. If you still have issues with the GPU based version, you may go for the CPU based version. Mac Compatibility: Only MacOS>=10.15.3

• Display model attention as heatmap (Grad-CAM) -> accessible via the interactive confusion matrix
• Display model architecture in TensorBoard (Windows only) -> accessible via the interactive confusion matrix
• Multiple bug-fixes
• SqueezeNet (pretrained on Imagenet) now available in model zoo

Limitation of the GPU Version: Only NVIDIA GPUs with a compute capability>3.7 are supported. What is my compute capability? AID was tested on multiple Systems with different GPUs. On some systems, AID only worked after the 2nd start. If AID crashes or does not detect your GPU, you may need to update your GPU driver. If you still have issues with the GPU based version, you may go for the CPU based version. Mac Compatibility: Only MacOS>=10.15.3

• GPU Support out of the box! AIDeveloper now detects your NVIDIA GPU automatically (currently only for the Windows version)
• Train a model on multiple GPUs
• Select if CPU, GPU, or Multi-GPU (if available) should be used for training
• Train multiple models on CPU and/or GPU in parallel
• Multi-GPU models are saved such that they can be loaded on any other system
• Any trained model from AIDeveloper can be loaded on another PC (e.g. to to continue training or to predict)
• AIDeveloper works off-line. There is no "home-calling" (Data-protection)
• More pretrained models (ResNet, DenseNet, Inception,...)

Limitation of the GPU Version: Only NVIDIA GPUs with a compute capability>3.7 are supported. What is my compute capability? AID was tested on multiple Systems with different GPUs. On some systems, AID only worked after the 2nd start. If AID crashes or does not detect your GPU, you may need to update your GPU driver. If you still have issues with the GPU based version, you may go for the CPU based version.

Special thanks to @alfrem for running hundreds of tests for the Multi-GPU options!

• More conversion methods (available on History-Tab): models created by AIDeveloper (Keras models) can now be converted to: TensorFlow (.pb), PyTorch, Caffe, CNTK, MXNet, ONNX, and CoreML.
• New Expert option: choose the padding_mode: besides zero padding, AIDeveloper now offers plenty other options including reflection and symmetric padding. The padding_mode can be changed in-between epochs and is tracked in the meta-file.
• Improved support for the normalization method "standard scaling".
• Bugfix related to pyqtgraph.