如何正确处理tensorflow RNN中的one-hot-encoding和多维数据

Question

I'm creating a binary classifier that classifiers letter sequences eg 'BA'.

Each sequence is made up of 2 letters encoded as one-hot vectors. For example, the sequence 'BA' is [[0, 1, 0, 0], [1, 0, 0, 0]] .

(The sequences are longer in my original code but I want to keep my question simple)

I'm reasonably new to machine learning, I've done some stuff before but only with flat inputs. I'm struggling to understand how I can pass a multi-dimensional input into my network. Do I need to flatten it somehow?

I've created a minimally reproducible example here - https://www.dropbox.com/s/dhg3huw6bh7dfjd/Example.py?dl=0 , Sequences only contains 3 examples of training data to keep it simple. I just need help working out how to pass 3d inputs into my model.

In the code I've provided, the entire dataset is just the following:

Sequences = [  [[0, 1, 0, 0], [1, 0, 0, 0]]   ,  [[0, 1, 0, 0], [1, 0, 0, 0]]  ,   [[0, 0, 0, 1], [0, 1, 0, 0]]  ]

It is a binary classifier, Targets for these sequences are:

Targets = [1.0, 1.0, 0.0]

If you run the code I've provided you should get this error:

ValueError: Input 0 of layer lstm is incompatible with the layer: expected ndim=3, found ndim=4. Full shape received: [None, 2, 2, 4]

If someone could help figure out how to correctly pass the sequences into my network that would be great. Thank you :)

Answer 1

This line is the culprit. LSTM expects 3 dimensional input. In your case you are trying to pass a 4 dimensional input (remember that Keras add an additional batch dimension to the input shape). So, getting rid of the first 2 will solve your issue.

model.add(LSTM(128,
                   input_shape=(2, 2, 4), return_sequences=True))

Simply change this to,

model.add(LSTM(128,
                   input_shape=(2, 4), return_sequences=True))