About SLAYER deConv feature

[alexggener]

About SLAYER deConv feature

[bamsumit]

@alexggener transposed convolution synapses are already available in lava-dl SLAYER. The block api for transposed convolution is not available though.

So there are two options for you:

1. Not use block api (block api is just for convenience) and initialize the neuron, synapse and quantization directly in your network definition until transposed convolution support in block api is available.
   e.g:

   neuron = slayer.neuron.cuba.Neuron(...)
   synapse = slayer.synapse.ConvTranspose(...)
   synapse.pre_hook_fx = neuron.quantize_8bit  # if you need quantization

   and call them in sequence in forward():

   spike_output = neuron(synapse(spike_input))

2. Implement block api and contribute to the repo. I can list out what needs to be done.

[alexggener]

@bamsumit Thanks for pointing us to the right direction. Implementing it on the block api seems doable so we'll try to contribute to it. Any guidance during the process would be appreciated.

[bamsumit]

@alexggener sounds good. Here are the steps

1. in slayer/base.py, implement AbstractConvT and AbstractUnpool. Follow the flow similar to AbstractConv and AbstractPool respectively, with different synapse of-course.
2. Derive ConvT and Unpool in all the other neuron specific blocks.