potatoref / vision_ui Goto Github PK

Graphical User interface for fastai

currently only works with version 1

Visual UI adds a graphical interface to fastai allowing the user to quickly load, choose parameters, train and view results without the need to dig deep into the code.

• Can now be pip installed fast-gui

• Update for compatability with fastai2
• Files: Visual_UI2.ipyb and vision_ui2.py

Files: Visual_UI.ipyb and vision_ui.py

• Inclusion of ImageDataBunch.from_csv
• Additional augmentations included [cutout, jitter, contrast, brightness, rotate, symmetric warp, padding]
• Inclusion of ClassConfusion widget
• Addition of 'Code' tab to view code

• Under the 'Info' tab you can now easily upload some common datasets: Cats&Dogs, Imagenette, Imagewoof, Cifar and Mnist

• Under the 'Results' tab if there are more than 2 classes the confusion matrix upgrades will not work but you can now view the confusion matrix

• Works with Google Colab (https://github.com/asvcode/Colab_UI) - Results tab not currently available in Colab

• xresnet architectures now working (using xresnet2.py from fastai)

• Under the Results tab, the confusion matrix tab now includes enhanced viewing features:

Option to view images with heatmaps or not

Option to view images within each section of the matrix

If heatmap option is 'YES' you can choose colormap, interpolation and alpha parameters

Examples of using different parameters for viewing images

Also have the option to view the images without the heatmap feature. Images within each matrix class display Index, Actual_Class, Predicted_Class, Prediction value, Loss and Image location

Images are stored within the path folder under their respective confusion matrix tags

View saved image files from various sections of the confusion matrix and compare their heatmap images.

False Positive True Positive True Negative False Negative

• after a training run, the model is saved in the models folder with the following name: 'architecture' + 'pretrained' + batchsize + image size eg: resnet50_pretrained_True_batch_32_image_128.pth
• updated tkinter askdirectory code: now after choosing a file the tkinter dialogue box will be destroyed - previously the box would remain open

• results tab added where you can load your saved model and plot multi_plot_losses, top_losses and Confusion_matrix

• path and image_path (for augmentations) is now within vision_ui so no need to have a seperate cell to specify path
• included link to fastai docs and forum in 'info' tab

All tabs are provided within an accordion design using ipywidgets, this allows for all aspects of choosing and viewing parameters in one line of sight

The Augmentation tab utilizes fastai parameters so you can view what different image augmentations look like and compare

View batch information

Review model data and choose suitable metrics for training

Review parameters get learning rate and train using the one cycle policy

Can experiment with various learning rates and train

• fastai

I am using the developer version:

git clone https://github.com/fastai/fastai

cd fastai

tools/run-after-git-clone

pip install -e ".[dev]"

for installation instructions visit Fastai Installation Readme

• ipywidgets

pip install ipywidgets jupyter nbextension enable --py widgetsnbextension

or

conda install -c conda-forge ipywidgets

for installation instructions visit Installation docs

• psutil

psutil (process and system utilities) is a cross-platform library for retrieving information on running processes and system utilization (CPU, memory, disks, network, sensors) in Python

pip install psutil

git clone this repository

git clone https://github.com/asvcode/Vision_UI.git

run Visual_UI.ipynb and run display_ui()

• Colab - version 1 works with colab Colab_UI but is glitchy

• Integrate into fastai v2 - Compatability with fastai v2 done but not with the full functionality as the v1 version
• Create pip install verson - Done! fast-gui
• Include full functionality with v2 version
• Create a v2 that is compatible with Colab