feed data to fitDataset() - tensorflow-datasets

Question

I'm trying to fit a model using fitDataset(). I can train using the "normal" approach, with a for loop and getting random batches of data (20000 data points).
I'd like to use the fitDataset() and be able to use the entire dataset and not rely on "randomness" of my getBatch function.
I'm getting closer, using the API docs and the example on tfjs-data but, i'm stuck on a probably dumb data manipulation...
So here's how i'm doing it:
const [trainX, trainY] = await bigData
const model = await cnnLSTM // gru performing well
const BATCH_SIZE = 32
const dataSet = flattenDataset(trainX.slice(200), trainY.slice(200))
model.compile({
loss: 'categoricalCrossentropy',
optimizer: tf.train.adam(0.001),
metrics: ['accuracy']
})
await model.fitDataset(dataSet.train.batch(32), {
epochs: C.trainSteps,
validationData: dataSet.validation,
callbacks: {
onBatchEnd: async (batch, logs) => (await tf.nextFrame()),
onEpochEnd: (epoch, logs) => {
let i = epoch + 1
lossValues.push({'epoch': i, 'loss': logs.loss, 'val_loss': logs.val_loss, 'set': 'train'})
accuracyValues.push({'epoch': i, 'accuracy': logs.acc, 'val_accuracy': logs.val_acc, 'set': 'train'})
// await md `${await plotLosses(train.lossValues)} ${await plotAccuracy(train.accuracyValues)}`
}
}
})
here's my interpretation of the dataset creation:
flattenDataset = (features, labels, split = 0.35) => {
return tf.tidy(() => {
let slice =features.length - Math.floor(features.length * split)
const featuresTrain = features.slice(0, slice)
const featuresVal = features.slice(slice)
const labelsTrain = labels.slice(0, slice)
const labelsVal = labels.slice(slice)
const data = {
train: tf.data.array(featuresTrain, labelsTrain),
validation: tf.data.array(featuresVal, labelsVal)
}
return data
})
}
I'm getting an error:
Error: Dataset iterator for fitDataset() is expected to generate an Array of length 2: `[xs, ys]`, but instead generates Tensor
[[0.4106583, 0.5408, 0.4885066, 0.9021732, 0.1278526],
[0.3711334, 0.5141, 0.4848816, 0.9021571, 0.2688071],
[0.4336613, 0.5747, 0.4822159, 0.9021728, 0.3694479],
...,
[0.4123166, 0.4553, 0.478438 , 0.9020132, 0.8797594],
[0.3963479, 0.3714, 0.4871198, 0.901996 , 0.7170534],
[0.4832076, 0.3557, 0.4892016, 0.9019232, 0.9999322]],Tensor
[[0.3711334, 0.5141, 0.4848816, 0.9021571, 0.2688071],
[0.4336613, 0.5747, 0.4822159, 0.9021728, 0.3694479],
[0.4140858, 0.5985, 0.4789927, 0.9022084, 0.1912155],
...,
The input data is 6 timesteps with 5 dimensions and the labels are just one-hot encoded classes [0,0,1], [0,1,0] and [1, 0, 0]. I guess the flattenDataset() is not sending the data in the correct way.
Does data.train needs to output for each data point [6 timesteps with 5 dims, label] ? I get this error when i tried that:
Error: The feature data generated by the dataset lacks the required input key 'conv1d_Conv1D5_input'.
Could really use some pro insight...
--------------------
Edit #1:
I feel i'm close to an answer.
const X = tf.data.array(trainX.slice(0, 100))//.map(x => x)
const Y = tf.data.array(trainY.slice(0, 100))//.map(x => x)
const zip = tf.data.zip([X, Y])
const dataSet = {
train: zip
}
dataSet.train.forEach(x => console.log(x))
With this i get on the console:
[Array(6), Array(3)]
[Array(6), Array(3)]
[Array(6), Array(3)]
...
[Array(6), Array(3)]
[Array(6), Array(3)]
but the fitDataset is giving me: Error: The feature data generated by the dataset lacks the required input key 'conv1d_Conv1D5_input'.
my model look like this:
const model = tf.sequential()
model.add(tf.layers.conv1d({
inputShape: [6, 5],
kernelSize: (3),
filters: 64,
strides: 1,
padding: 'same',
activation: 'elu',
kernelInitializer: 'varianceScaling',
}))
model.add(tf.layers.maxPooling1d({poolSize: (2)}))
model.add(tf.layers.conv1d({
kernelSize: (1),
filters: 64,
strides: 1,
padding: 'same',
activation: 'elu'
}))
model.add(tf.layers.maxPooling1d({poolSize: (2)}))
model.add(tf.layers.lstm({
units: 18,
activation: 'elu'
}))
model.add(tf.layers.dense({units: 3, activation: 'softmax'}))
model.compile({
loss: 'categoricalCrossentropy',
optimizer: tf.train.adam(0.001),
metrics: ['accuracy']
})
return model
What is wrong here?

Answer 1

What model.fitDataset expects are a Dataset, each element inside this dataset is a tuple of two items, [feature, label].
So in your case, you need to create featureDataset and labelDataset, then merge then with tf.data.zip to create trainDataset. Same for validation dataset.

Answer 2

Solved it
so after a lot of trial an error i found a way to make it work.
So, i had an input shape of [6, 5], meaning an array with 6 arrays of 5 floats each.
[[[0.3467378, 0.3737, 0.4781905, 0.90665, 0.68142351],
[0.44003019602788285, 0.3106, 0.4864576, 0.90193448, 0.5841830879700972],
[0.30672944860847245, 0.3404, 0.490295674, 0.90720676, 0.8331748581920732],
[0.37475716007758336, 0.265, 0.4847249, 0.902056932, 0.6611207914113887],
[0.5639427928616854, 0.2423002, 0.483168235, 0.9020202294447865, 0.82823],
[0.41581425627336555, 0.4086, 0.4721923, 0.902094287, 0.914699]], ... 20k more]
What i did was to flatten the array becoming an array of 5 dimensions arrays. Then applied the .batch(6) to it.
const BATCH_SIZE = 20 //batch size fed to the NN
const X = tf.data.array([].concat(...trainX)).batch(6).batch(BATCH_SIZE)
const Y = tf.data.array(trainY).batch(BATCH_SIZE)
const zip = tf.data.zip([X, Y])
const dataSet = {
train: zip
}
Hope it can help others on complex data!!