CODEWITHSUNDEEP
pythontensorflow
tastyminerals
tastyminerals

Reputation: 6538

How to apply linear transform on a 3D feature vector in Tensorflow?

Imagine there is a tensor with the following dimensions (32, 20, 3) where batch_size = 32, num_steps = 20 and features = 3. The features are taken from a .csv file that has the following format:

feat1, feat2, feat3
200, 100, 0
5.5, 200, 0.5
23.2, 1, 9.3

Each row is transformed into 3-dim vector (numpy array): [200, 100, 0], [5.5, 200, 0.5], [23.2, 1, 9.3].

We want to use these features in a recurrent neural network but directly feeding them into rnn won't do, we'd like to process these feature vectors first by applying linear transformation to each 3-dim vector inside the batch sample and reshape the input tensor into (32, 20, 100).

This is easily done in Torch for example via: nn.MapTable():add(nn.Linear(3, 100)) which is applied on the input batch tensor of size 20 x 32 x 3 (num_steps and batch_size are switched in Torch). We split it into 20 arrays each 32x3 in size 

  1 : DoubleTensor - size: 32x3
  2 : DoubleTensor - size: 32x3
  3 : DoubleTensor - size: 32x3
  ...

and use nn.Linear(3, 100) to transform them into 32x100 vectors. We then pack them up back into 20 x 32 x 100 tensor. How can we implement the same operation in Tensorflow?

Upvotes: 2

Views: 1297

Answers (1)

Burton2000
Burton2000

Reputation: 2072

Could reshape into [batchsize*num_steps, features] use a Tensorflow linear layer with 100 outputs and then reshape back would that work?

reshaped_tensor = tf.reshape(your_input, [batchsize*num_steps, features])
linear_out = tf.layers.dense(inputs=reshaped_tensor, units=100)
reshaped_back = tf.reshape(linear_out, [batchsize, num_steps, features]

Upvotes: 2

Related Questions