This package provides the Dataset API for MNIST dataset. It is built using @tensorflow/tfjs-data package (which is now included in @tensorflow/tfjs union package) that provides a uniform and consistent way to access various datasets.
npm install tfjs-data-mnist
// get the dataset
const ds = await MNISTDataset.create();
// there are 2 properties in ds (testDataset and trainDataset)
// get the iterator for testDataset
const it = await ds.testDataset.iterator();
// iterate by invoking next
const dataElement = await it.next();
// dataElement.done === true => there are no more elements
// dataElement.value is **TensorContainer** of type [feature, label]
// where feature and label are of type Tensor1D
//
// feature is Tensor1D with shape [784]
// label is Tensor1D with shape [10]
//
//
// label is actually a one-hot encoded vector
// how to get the feature and label
const feature = dataElement.value[0] as tfjs.Tensor;
const label = dataElement.value[1] as tfjs.Tensor;
// The nice thing about dataset API is that you get
// lot of operations such as suffle, repeat, take etc
// for free
// Here is an example to first shuffle the dataset
// and then take only first 5 samples
const shuffled5 = await ds.testDataset.shuffle(10).take(5).iterator();
// You can also pass dataset to train the model
await model.fitDataset(ds.trainDataset.batch(32), {
epochs: 1,
callbacks: {
onBatchEnd: async (batch: number, logs?: tf.Logs) => {
batchProgressEl.innerText =
`${batch} - ${logs['loss']} - ${logs['acc']}`;
},
onEpochEnd: async (epoch: number, logs?: tf.Logs) => {
epochEndResultEl.innerText =
`${epoch} - ${logs['loss']} - ${logs['acc']}`;
}
}
});
# do npm install at the root of this directory
npm install
# install peer dependnencies
npm install @tensorflow/tfjs-core @tensorflow/tfjs-data --no-save
# change directory into example
cd examples
# do npm install in example
npm install
# Run a basic example that shows
# how to use the api of Dataset
npm run basic
# Another example is to train a model
# where I use fitDataset api that takes Dataset
# as an input
npm run train