Getting the learned representation of the data fro

Question

We can train an autoencoder in pylearn2 using below YAML file (along with pylearn2/scripts/train.py)

!obj:pylearn2.train.Train {
    dataset: &train !obj:pylearn2.datasets.mnist.MNIST {
        which_set: 'train',
        start: 0,
        stop: 50000
    },
    model: !obj:pylearn2.models.autoencoder.DenoisingAutoencoder {
        nvis : 784,
        nhid : 500,
        irange : 0.05,
        corruptor: !obj:pylearn2.corruption.BinomialCorruptor {
            corruption_level: .2,
        },
        act_enc: "tanh",
        act_dec: null,    # Linear activation on the decoder side.
    },
    algorithm: !obj:pylearn2.training_algorithms.sgd.SGD {
        learning_rate : 1e-3,
        batch_size : 100,
        monitoring_batches : 5,
        monitoring_dataset : *train,
        cost : !obj:pylearn2.costs.autoencoder.MeanSquaredReconstructionError {},
        termination_criterion : !obj:pylearn2.termination_criteria.EpochCounter {
            max_epochs: 10,
        },
    },
    save_path: "./dae_l1.pkl",
    save_freq: 1
}

What we get is the learned autoencoder model as "dae_l1.pkl".

If I want to use this model for supervised training, I can use "dae_l1.pkl" to initialize the layer of an MLP. I can then train this model. I can even predict the output of the model using 'fprop' function.

But what if I dun want to use this pretrained model for supervised learning and I just want to save the new learned representation of my data with the autoencoder.

How can I do this?

Even more detailed question is put here

Answer 1

I think you can use the encode and decode functions of the autoencoder to get the hidden representation. E.g:

l1_path = 'dae_l1.pkl'
l1 = serial.load(l1_path)
"""encode"""
#layer 1
l1Input = l1.get_input_space().make_theano_batch()
l1Encode = l1.encode(l1Input)
l1Decode = l1.decode(l1Encode)
l1EncodeFunction = theano.function([l1Input], l1Encode)
l1DecodeFunction = theano.function([l1Encode], l1Decode)

Then, the representation will be:

l1encode = l1EncodeFunction(YourData)

Answer 2

The reconstruct method of the pickled model should do it - I believe usage is the same as fprop.