Introduction

Welcome to the DeepDetect platform!

The DD platform simplifies best practices for Deep Learning / Machine Learning. As applied researchers and engineneers, we use it at Jolibrain. As it works for us, we hope it works for you.

Best Practices

We see the platform as an help for using best practices when building and testing Deep Learning models. In more details this means:

A selection of fully tested neural network architectures that work extremely well in a large number of cases
Data pipelines for training with data augmentation that are well tested, bug-free and yield excellent results in many real-world applications
REST API and server to use deep models in production, with error handling, UI building and testing made easy.

Our goal is to transit the best results from academia to the industry, avoiding bugs and time losses, while preserving openess and usability.

Quickstart

User Setup

Use the New Folder button

New user folder

work directory in JupyterLab

Work jupyter directory

Inside work directory

Work inside jupyter

Make a folder under your name

Work username directory

User setup proceeds with two steps:

Setup of your User Data Directory that will store all your data files
Setup of your JupyterLab User Directory that will store all your Python notebooks

Setup your data directory

From the main UI:

Click on Data
Go the file manager tab
Use the ‘New Folder’ button to create a folder for your username, e.g. JeanDupont

This sets up your own Data directory with backups

Setup your JupyterLab Python directory

From the main UI:

Click on Jupyter
Go to the 'work' directory

Click the New Folder button in the left panel of the Jupyter Lab window
Right-click on the newly created folder Untitled Folder, choose Rename and set your username, e.g. JeanDupont

This sets up your own JupyterLab directory with backups

Available models

Available models

Available models

Example predict image labels

Example image prediction

The platform sets up a range of pre-trained models that are ready to be used within applications.

To use these models:

Click on Predict from the main UI
Select your model from the Current Predict Service list or from the left-hand tab.

These models are pre-loaded on the platform's GPUs.

For loading up your own private models see section Prediction from models

Deep Learning training examples

The platform presets a range of examples for training your own models via Jupyter calls and monitoring with the DeepDetect platform UI.

Examples cover training from CSV data, images for classification, object detection and segmentation, text classification, audio classification and generic data autoencoder.

To try one of the examples, follow the steps below:

Copy the work/examples/dd_examples.ipynb file to your JupyterLab user directory, e.g. work/JeanDupont/dd_examples.ipynb

Do not forget to replace JeanDupont with your user name!

Open a new notebook from JupyterLab, rename it for instance test1.ipynb

Copy the first cell from the dd_examples.ipynb notebook to the first cell of your new notebook, that is the way to start any training notebook.

# Copy the first cell from the `dd_examples.ipynb` notebook to the first cell of your new notebook
import sys
sys.path.append('/opt/platform/code/dd_widgets/')
from dd_widgets import Classification, CSV, Text, Segmentation, Detection

Find the example that is of interest to your task, and copy its cell to your test1.ipynb notebook
In the example cells you have copied, change the model_repo variable so that the model and files are stored within your user data directory
```
# Modify the model_repo
model_repo="/opt/platform/models/training/JeanDupont/covert",
```

Make sure you are setting up the right `model_repo` by using the path completion (press Tab from the `model_repo` path string)

Do not forget to replace JeanDupont with your user name!

Click on Run training.

The `Status` indicator should show `running`

Go to the platform UI, click on the Training tab where your job should appear
Click on Monitor from your training job to look at live metrics