Open eddiebergman opened 5 months ago
The first step with PyTorch integration is to make it work with a simple MLP with 1 hidden layer. This works quite trivially if you have a class MyNet
that implements it but that's not what the amltk pipelines are for. We'd rather define it as so:
pipeline = Sequential(
nn.Flatten(start_dim=1),
Component(nn.Linear, config={"in_features": 724, "out_features": 20}, name="fc1"),
nn.ReLU,
Component(nn.Linear, config={"in_features": 20, "out_features": 10}, name="fc2"),
Component(nn.LogSoftmax, config={"dim": 1}),
name="my-mlp-pipeline",
)
The first challenge is to somehow define the search space in the pipeline, where that number 20
can go between something like (10, 30)
. The main issue is:
request
functionality to make this work? Typically we've just defined the search space with the component its parameterize.The basic requirements of the previous features are mostly implemented aside from Join
and Split
which I will work on soon.
In the meantime, the next steps will be towards taking the ResNet models family from PyTorch and do the following:
This issue will serve as a log as to the PyTorch progress in AMLTK. Please feel free to chime in with any information/suggestions/solutions to problems.