as2528/neural-dashboard repository overview

⁠Neural Dashboard

An interactive neural network visualizer and training playground.
Designed for fast prototyping and educational insights into how neural networks learn in real time.

This container runs a fully self-contained dashboard where you can:

• Build custom neural networks visually (CNNs, FCNNs, etc.)
• Train on image and tabular datasets
• Visualize 2D loss curves and 3D loss landscapes (PCA-reduced)
• Export trained models
• Use CPU or GPU (CUDA supported)
• Supports CSV, ZIP image folders (ImageFolder), and custom parameters

⁠How to Run

⁠Option 1 — From Docker Hub

docker pull as2528/neural-dashboard:latest
docker run --gpus all -p 7860:7860 as2528/neural-dashboard:latest

Open in your browser:

http://localhost:7860

⁠Option 2 — From Local .tar File

If you have the exported Docker image .tar:

docker load -i neural-dashboard.tar
docker run --gpus all -p 7860:7860 neural-dashboard:latest

Open in your browser:

http://localhost:7860

⁠Neural Network Dashboard Manual

This program is used to quickly prototype and deploy CNN and FCNN architectures.
It automatically uses the CUDA-capable GPU when available; if no GPU is found, it falls back to CPU.

The backend uses PyTorch to build and train neural networks.
The program is fully editable and should be easily extendable; you can easily add more layers if you want to extend it.

⁠Table of Contents

• Supported Layers & Parameters⁠
• Architecture Validation⁠
• CUDA Errors⁠
• File Handling System⁠
  • CSV Format⁠
  • ZIP Format (Image Dataset)⁠
• Returns⁠
• Training Tab: Image Resize, Channels, Loss Function, Optimizer⁠
• Extendability⁠
• Final Note⁠

⁠Supported Layers & Parameters

Notes:

• in_dim and out_dim are mandatory for layers like Conv2d and Linear.
• For Dropout, enter the probability into the in_dim field. (Example: 0.5 for 50% dropout.)
• Layers like ReLU, Tanh, Sigmoid, Flatten, GELU, and Softmax have no user-configurable parameters (except Softmax, which uses dim=1 by default).
• Pooling layers accept optional parameters but work well with defaults.

⁠Architecture Validation

If you build an invalid architecture, the program will stop before training begins.

You’ll see a ⚠️ icon next to the problematic layer in the UI.
You can then edit or delete the layer using the Edit Layer tab.

⁠CUDA Errors

If the program encounters a CUDA error:

• The error will appear in the GPU section at the bottom of the UI.
• OR, a warning will appear advising you to restart the app or kernel.

Note:

• CUDA errors require a full kernel or app restart to clear.

⁠File Handling System

The app accepts CSV files for tabular data and ZIP files for image datasets.

⁠CSV Format

Your CSV should look like this (headers must match exactly):

dataset.csv

• x1, x2, x3, ..., y

• 0.5, 1.2, 3.4, ..., 0

• 0.7, 0.8, 2.1, ..., 1

• ...

• x1, x2, ..., xn are your features.

• y is the label.

• Headers are case-sensitive.

• Make sure to include all headers, or training will not start.

⁠ZIP Format (Image Dataset)

Organize your .zip file like this:

dataset.zip

• class1

  • image1.jpg
  • image2.jpg
  • ...

• class2

  • image1.jpg
  • image2.jpg
  • ...

• ...

• Each folder represents a class label.

• Inside folders, use image formats like .jpg, .png.

• Non-image files inside folders will be ignored.

• The app automatically extracts and deletes temporary files after training.

• The program only accepts zip files. Make sure the file is zipped!

• If any corrupted file is skipped the program outputs an info message on the UI.

⁠Returns

The program returns the following:

• A loss plot showing the losses through the epochs (downloadable).
• A video showing the path the model followed to its minima if the checkbox is selected, otherwise a black video is generated (downloadable).
• The complete trained model and architecture (downloadable).
• Losses on every Epoch written out at the bottom of the screen (readable but not downloadable).

⁠Training Tab: Image Resize, Channels, Loss Function, Optimizer

⁠Image Resize

• All input images are resized before training.

• Set value in: Image Resize (size_box)

• Example: 28 → Resizes images to 28×28.

• Purpose: Ensure consistent input size to avoid shape mismatches.

⁠Input Channels

• Control color mode of input images.

  • Options:

    1 → Grayscale

    3 → RGB color

  • Set in: Input Channels (channel_dropdown)

  Note: For tabular data (CSV), this is ignored.

⁠Loss Function

⁠Define loss calculation for backpropagation.

Options:

• CrossEntropyLoss: For classification (e.g., cat/dog). Use when labels are class indices (0, 1, 2, ...). Softmax is internal.

• MSELoss: For regression or if you want probability output with manual softmax.

• Set in: Loss Function (loss_dropdown) Behavior:

   - If MSELoss is selected and classification data is used, one-hot encoding + softmax is applied internally.
  
    - If CrossEntropyLoss, ensure target labels are integers (not one-hot).
  

  Copy

⁠Optimizer

• Controls weight updates during training.

  Options:

    - Adam: Adaptive learning rate, recommended default.
  
    - SGD: Standard stochastic gradient descent.
  

  Copy

  Set in: Optimizer in the dropdown

Corrupted Image Handling

- Corrupted images are skipped automatically.

- Warnings are shown in the Gradio Info bubble.

Workflow Summary

• Images resized to target size.

• Images converted to target channel mode (RGB or grayscale).

• Dataset checked for corruptions, skipped if needed.

• Loss and optimizer selected based on UI.

• Training starts, logs are shown live.

Notes

• Tabular data (CSV): Resize and channels are ignored.
• Animation: If enabled, generates 3D loss descent animation.

⁠Extendability

The program is built to be modular:

• You can easily add more layers to the backend.
• All layers follow a consistent internal config for easy expansion.

⁠Final Note

If you think of improvements or want to extend the features, feel free to hack away!

⁠Features

• Drag-and-drop neural network builder (Gradio web interface)
• Training progress log and loss curve plotting
• Dynamic PCA-based 3D animation of loss descent
• Downloadable model artifacts after training
• Auto GPU detection (NVIDIA CUDA)

⁠Requirements

• Docker 20.10+
• Optional: NVIDIA GPU with CUDA drivers (for acceleration)

⁠Links

• GitHub: https://github.com/as2528/Dashboard/tree/main⁠
• Huggingface: https://huggingface.co/spaces/as2528/Dashboard⁠

⁠License

MIT License — free for personal and commercial use.

⁠Author

Created by as2528.