Current Version: V1.0
Demo of Neural Network Sandbox
- Introduction
- Installation
- How to run
- Features
- Readme
- Help
- Optimization Technique
- Loss Function
- Add Layer
- Remove Layer
- Weights Intialization
- Learning rate
- Epochs
- Start
- Reset
- Batch Normalization
- Use GPU
- Shuffle Data
- Training Stats
- Layer Control
- Network Drawings
- Save Current Model
- Visualize Training
- Browse training Data
- Load Exsisting Model
- Testing Stats
- Browse testining Data
- Report a Bug
- Label at Start
- Smart PreProcess Data
- Validation Split
- Known Issues
- Acknowledgment
- Credits
Neural Network Sandbox is a GUI based application which makes making and training basic feed forward neural networks easy.
Before starting: Assume default value of parameters which are not listed here.(For Ex: batch_size=32)
Currently we are working on one executable file and pip package for this application.
Updates about package and executable will be posted here
While there is no executable available you can try Neural Network Sandbox with source code.
Before running Make Sure you have graphviz installed on your PC. Also make sure you have graphviz in environment variables.See #1 For More info on this.
- Clone this repo. (Stable : V1.0.You can also clone current code but it maybe unstable)
- Fulfill
requirements.txt(pip install -r requirements.txt). - Run with command
python main.py.
Above is the screenshot of Application.We will see each section one by one
This button will simply redirect you to README.md (Which contains documentation)of Github repo.
This button will simply redirect you to HELP.md(Which is useful for new users) of Github repo.
You can read HELP.md if you are beginner in Neural networks.All the buzzwords are explained there.
This option gives user number of choices for optimizer for their network.Click on SGD to see options and select one of them.
Options are:
- SGD
- RMSprop
- Adagrad
- Adam
- Adamax
- Nadam
- Adadelta
This option gives user number of choices for loss function for their network.Click on sparse_categorical_crossentropy to see options and select one of them.
Options are:
- sparse_categorical_crossentropy
- binary_crossentropy
- categorical_crossentropy
- mean_squared_error
- mean_absolute_error
- huber_loss
- cosine_proximity
- poisson
Clicking on this button will add new Layer in Layer Control.
Before Adding Layer:
After Adding Layer:
Clicking on this button will add remove Layer from Layer Control.Note that minimum of 2 Layer is required so it can only remove layer when number of layers are >=3.
Before Removing Layer:
After Removing Layer:
This option gives user number of choices for weights intialization (kernal intialization) for their network.Click on he_normal to see options and select one of them.
Note that picked intializer will be used for the whole network.
Options are:
- he_normal
- he_uniform
- lecun_normal
- lecun_uniform
- glorot_normal
- glorot_uniform
- RandomNormal
- RandomUniform
- Orthogonal
Input your choice of learning rate here.Generally it is between 0 and1.Here,default value is 0.01.
Buttons + and - can be used for increasing or decreasing number of epochs.Note that minimum value is 1 however there is no limit on maximum value.
Clicking this button will first check for errors (Error will be reported as popup if any) in selected options and than it will start training.
Clicking this button will reset app to it's initial state(i.e. mainscreen shown in Features).
Checking this will add BatchNormalize layer after each layer in your network except ouput layer.
If you check this app will try it's best to utilize GPU for training.
| Tensorflow Version | Effects of this option |
|---|---|
| Tensorflow <2.1.0 (CPU) | Checking or Unchecking will not make any difference |
| Tensorflow-gpu <=2.0.0 | Checking will use GPU (won't work if CUDA is not configured) unchecking will use CPU. |
| Tensorflow == 2.1.0 (Which supports both CPU and GPU) | Checking will use GPU (will work on CPU if CUDA is not configured) unchecking will use CPU. |
Checking this will shuffle training data before training.
After training if finished all stats (like accuracy,loss etc..) will be shown here.
Note that it only shows following metrics:
- Loss
- accuracy
- mse
- val_loss (if validation split >0)
- val_accuracy (if validation split >0)
- val_mse (if validation split >0)
You can give number of neurons and activation for each layer in layer control.Minimum number of neurons is 1 however there is no limit on maximum value.
Click on sigmoid to select other options for activation.
Options are:
- sigmoid
- relu
- elu
- selu
- tanh
- softmax
- linear
Note that You can't select activation for first layer.
Network Drawing will be shown in middle of application.Default Drawing is Input -> Ouput as shown in main screen.Drawing will be updated after training (given training is successful).
Note that it is not practical to include all neurons in drawing so for performance reasons number of neurons in drawing are limited to 13.
After Training you can save your model by clicking this.It will give dynamic name to your file and save it with .h5 extension.Popup is shown upon successful saving.
After training you can see graphical history of training by clicking this.It will open new window containing image.
You can select any training file with .csv extension.Before selecting note that app only supports single label in training data.
You can load any file (Even model which is trained outside this application) with .h5 extension.Layer control and Network Drawing will be set according to loaded model.
Note that as app only supports simple feed forward networks, loading file of other networks may not give appropiate output.
After training and testing (if testing data is selected) you can click this button to view testing statistics.Popup will be shown with this statistics.
This Contains:
- loss
- accuracy
- mse
You can select any testing file with .csv extension.Before selecting note that app only supports single label in testing data.
Clicking this button will simply redirect you to issues page of github repo.You can raise new issue and we'll try our best to fix it.
Checking it tells app that label is in first column of training/testing file. UnChecking it tells app that label is in last column of training/testing file.
This is an experimental feature.
Checking this will preprocess selected training/testing before starting.It does mainly following:
- Remove Unnecessary columns(Based on number of unique values).
- Fill Missing data with median.
- Replace text data with number representation.
Slide it to select percentage of validation data from training data.0% means no validation.
| Issue | Fixed |
|---|---|
| Sometimes Visualize training doesn't work | No |
We would like to thank Dr. Devang Pandya, Adani Institute of Infrastructure Engineering for his continuous guidance and support.
Contributors 💻 :
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent
