0

I want to train a LSTM network which starts at a patch in the image, outputs location where it should move next, and the input at next timestep is the patch around new cell.

Is there a way I could train this using Keras/PyTorch/Caffe/Tensorflow? Traditionally, networks have fixed input-output sequence for training. I need a network to have dynamic input-output sequence.

PS: I don't want to go for reinforcement learning, because I want to model path characteristics for future prediction.

Manish Sharma
  • 170
  • 2
  • 11
  • 1
    The only option I see is manually training a `stateful=True` layer step by step, creating the input steps manually between each batch. --- Details in "many to many" at the end: https://stackoverflow.com/questions/38714959/understanding-keras-lstms/50235563#50235563 – Daniel Möller Jun 04 '18 at 17:47
  • Or you could create a custom RNNCell based on LSTMCell where you select a slice of the total input. – Daniel Möller Jun 04 '18 at 17:48

0 Answers0