Lab 2: Convolutional Nets

Goal of the lab

• Use a simple convolutional network to recognize EMNIST characters.
• Construct a synthetic dataset of EMNIST lines.
• Move from reading single characters to reading lines.

Follow along

git pull
cd lab2

Using a convolutional network for recognizing EMNIST characters

We left off in Lab 1 having trained an MLP model on the EMNIST characters dataset.

Let's also train a CNN on the same task. We can start in the notebook notebooks/01b-cnn-for-emnist.ipynb.

We can also run the same experiment with

training/run_experiment.py '{"dataset": "EmnistDataset", "model": "CharacterModel", "network": "lenet", "train_args": {"epochs": 1}}'

Training the single epoch will take about 2 minutes (that's why we only do one epoch in this lab :)). Leave it running while we go on to the next part.

Subsampling data

It is very useful to be able to subsample the dataset for quick experiments. This is possibe by passing subsample_fraction=0.1 (or some other fraction) at dataset initialization, or in dataset_args in the run_experiment.py dictionary, for example:

training/run_experiment.py '{"dataset": "EmnistDataset", "dataset_args": {"subsample_fraction": 0.25}, "model": "CharacterModel", "network": "lenet"}'

Making a synthetic dataset of EMNIST Lines

• Synthetic dataset we built for this project
• Sample sentences from Brown corpus
• For each character, sample random EMNIST character and place on a line (with some random overlap)
• Look at: notebooks/02-look-at-emnist-lines.ipynb

Reading multiple characters at once

Now that we have a dataset of lines and not just single characters, we can apply our convolutional net to it.

Let's look at notebooks/02b-cnn-for-simple-emnist-lines.ipynb, where we generate a datset with at most 8 characters and no overlap.

The first network we try is simply the same LeNet network we used for single characters, applied to each character in sequence, using the TimeDistributed layer.

We can also express the same network using all convolutional layers, which we do next.

We can train this model with a command, too:

python training/run_experiment.py --save '{"train_args": {"epochs": 5}, "dataset": "EmnistLinesDataset", "dataset_args": {"max_length": 8, "max_overlap": 0}, "model": "LineModel", "network": "line_cnn_all_conv"}'