This is even faster … (less than 5 minutes!)
Prerequisite: a Google account and thus an access to Google Drive (https://drive.google.com).This gives us cloud storage that can be used for the model.
Step 1 (2 min): Selecting an example in the Seedbank project.The example should be similar to what we want to do later “actually” (if we already know). For a change from the well-known MNIST (optical digit recognition) we could take, for example, the newer Fashion MNIST (optical recognition of garments).
Step 2 (20 sec): Click “Run Seed in Colab”.this opens the selected sample in our Colabarea (and saves it in our Google Drive at the same time).Colab is an environment with Jupyter notebooks and TPUs where you can train and use machine learning models in the cloud.
The final 4 steps that follow are all initiated by one click each in the Colab/Jupyter notebook, which opened in step 2.
Step 3 (1 min): Download the Fashion MNIST data
Step 4 (10 sec): Model definition (Convolutional Network in Keras, based on tensorflow)
Step 5 (1 min): Training on a TPU
Step 6 (20 sec): Checking the trained network using examples:
What have we achieved in those just under 5 minutes?
- We have “built” a working network model (through Copy & Paste) and trained and tested it with a non-trivial data set.
- Based on this, we can now experiment independently with this network model, e.g. by
- Changing the learning rate (default 1e-3)
- Replacing the optimizer (default: Adam)
- Changing the network architecture (default: CNN with 3 convolutional layers and 2 fully connected layers as well as dropout and batch normalization)
- Changing the data set (which usually requires an adjustment of the input and output layer of the network model to take into account, for example, a changed image size and/or a different number of classes)
- and remarkably, it all works with a three-year-old discount laptop, as we calculate and store everything in the cloud.
Access to deep learning was probably never easier.