Getting dimensions wrong when creating a feed-forward auto-encoder in Theano/Lasagne

Question

I want to create a simple autoencoder with 3000 input, 2 hidden and 3000 output neurons:

def build_autoencoder(input_var=None):
    l_in = InputLayer(shape=(None,3000), input_var=input_var)

    l_hid = DenseLayer(
            l_in, num_units=2,
            nonlinearity=rectify,
            W=lasagne.init.GlorotUniform())

    l_out = DenseLayer(
            l_hid, num_units=3000,
            nonlinearity=softmax)

    return l_out 

The shape of the training data is as follows:

train.shape = (3000,3)

This is input, target and loss function definition:

import sys
import os
import time
import numpy as np
import theano
import theano.tensor as T
import lasagne
from lasagne.updates import rmsprop
from lasagne.layers import DenseLayer, DropoutLayer, InputLayer
from lasagne.nonlinearities import rectify, softmax
from lasagne.objectives import categorical_crossentropy
# Creating the Theano variables
input_var = T.dmatrix('inputs')
target_var = T.dmatrix('targets')

# Building the Theano expressions on these variables
network = build_autoencoder(input_var)

prediction = lasagne.layers.get_output(network)
loss = categorical_crossentropy(prediction, target_var)
loss = loss.mean()

test_prediction = lasagne.layers.get_output(network,
                                                    deterministic=True)
test_loss = categorical_crossentropy(test_prediction, target_var)
test_loss = test_loss.mean()
test_acc = T.mean(T.eq(T.argmax(test_prediction, axis=1), target_var),
                          dtype=theano.config.floatX)

I'm just running one epoch but get an error:

params = lasagne.layers.get_all_params(network, trainable=True)
updates = rmsprop(loss, params, learning_rate=0.001)

# Compiling the graph by declaring the Theano functions

train_fn = theano.function([input_var, target_var],
                                   loss, updates=updates)
val_fn = theano.function([input_var, target_var],
                                 [test_loss, test_acc])

# For loop that goes each time through the hole training
# and validation data
print("Starting training...")
for epoch in range(1):

    # Going over the training data
    train_err = 0
    train_batches = 0
    start_time = time.time()
    print 'test1'
    train_err += train_fn(train, train)
    train_batches += 1

    # Going over the validation data
    val_err = 0
    val_acc = 0
    val_batches = 0
    err, acc = val_fn(train, train)
    val_err += err
    val_acc += acc
    val_batches += 1

    # Then we print the results for this epoch:
    print("Epoch {} of {} took {:.3f}s".format(epoch + 1, num_epochs, time.time() - start_time))
    print("training loss:\t\t{:.6f}".format(train_err / train_batches))
    print("validation loss:\t\t{:.6f}".format(val_err / val_batches))
    print("validation accuracy:\t\t{:.2f} %".format(val_acc / val_batches * 100))

This is the error:

ValueError: ('shapes (3000,3) and (3000,2) not aligned: 3 (dim 1) != 3000 (dim 0)', (3000, 3), (3000, 2)) Apply node that caused the error: Dot22(inputs, W) Toposort index: 3 Inputs types: [TensorType(float64, matrix), TensorType(float64, matrix)] Inputs shapes: [(3000, 3), (3000, 2)] Inputs strides: [(24, 8), (16, 8)] Inputs values: ['not shown', 'not shown'] Outputs clients: [[Elemwise{add,no_inplace}(Dot22.0, InplaceDimShuffle{x,0}.0), Elemwise{Composite{(i0 * (Abs(i1) + i2 + i3))}}[(0, 2)](TensorConstant{(1, 1) of 0.5}, Elemwise{add,no_inplace}.0, Dot22.0, InplaceDimShuffle{x,0}.0)]]

To me it seems that the bottleneck of the auto encoder is the problem. Any ideas?

Answer 1

Just got some help from my IBM college (Erwan), I've posted the working solution to this GIST, the relevant sections are these ones:

First, get the shape of the training data correct:

train.shape = (3, 3000)

Then use the same shape on the InputLayer:

def build_autoencoder(input_var=None):
    l_in = InputLayer(shape=(3, 3000), input_var=input_var)

    l_hid = DenseLayer(
            l_in, num_units=2,
            nonlinearity=rectify,
            W=lasagne.init.GlorotUniform())

    l_out = DenseLayer(
            l_hid, num_units=3000,
            nonlinearity=softmax)

return l_out

So this is solved, next problem is getting a descending cost during training, but this is another topic :)