CODEWITHSUNDEEP
machine-learningtensorflowconvolution
Uylenburgh
Uylenburgh

Reputation: 1307

What does 4d tensor mean in Tensorflow?

I need to perform convolution along the text line of fixed size. So essentially, a training example is of a form: 1*N_FEATURES where N_FEATURES is equal to 3640 (140 characters encoded as one-hot, so 40*26=3640). I am trying to make sense of the example here, precisely:

def my_conv_model(X, y):
   X = tf.reshape(X, [-1, N_FEATURES, 1, 1])  # to form a 4d tensor of shape batch_size x n_features x 1 x 1
   features = skflow.ops.conv2d(X, N_FILTERS, [WINDOW_SIZE, 1], padding='VALID') # this will give you sliding window of WINDOW_SIZE x 1 convolution.
   pool = tf.squeeze(tf.reduce_max(features, 1), squeeze_dims=[1])
   return return skflow.models.logistic_regression(pool, y)

I do not understand why in this line:

 features = skflow.ops.conv2d(X, N_FILTERS, [WINDOW_SIZE, 1], padding='VALID')

we have : [WINDOW_SIZE, 1] and not [1, WINDOW_SIZE] ? Since as far as I understand, convolution should be performed as following:

training example: '001010101000100101'
sliding window:   |---|
                   |---|
                    |---|

and so on, each window of size [1, WINDOW_SIZE] since its height is 1, and width is 3. But why does the example given says " features = skflow.ops.conv2d(X, N_FILTERS, [WINDOW_SIZE, 1], padding='VALID')"?

Upvotes: 0

Views: 6548

Answers (1)

Prune
Prune

Reputation: 77857

A 4-D tensor is exactly what it indicates: a tensor with four dimensions (subscripts, features, ...).

The notation for a sliding window in one dimension is a notational convention. For convenience, we put the significant dimension first, and keep the 1 sizes for the end. Thus, a 3-wide slice (1-D window) is given as a 3x1 window.

Upvotes: 2

Related Questions