The model.loadPyTorchDump() function does not handle nested module within Sequential layers, and is unable to track the layer names correctly to import the files.
Ex: If we have a model from pytorch:
class Dummy(nn.Module):
def __init__(self):
super(Dummy, self).__init__()
self.model = nn.Sequential(
nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False),
)and we dump it using chai_dump
dummy = Dummy()
dummy.chai_dump('models/dummy', 'dummy')The files we get are:
models/dummy
| - model.0.weight.chdata
| - model.0.weight.json
| - specification.json
And the loadPyTorchDump function does not correctly handle these file names since it does not keep track of the "nested" nature of the Conv2d layer since it's inside of the Sequential layer.
A more complicated example that expands on the above:
class DummyTwo(nn.Module):
def __init__(self, input_model):
super(DummyTwo, self).__init__()
self.model = nn.Sequential(
input_model,
)
dummy_two = DummyTwo(dummy)
dummy_two.chai_dump('models/dummy_two', 'dummy_two')Gives us the files:
models/dummy_two
| - model.0.model.0.weight.chdata
| - model.0.model.0.weight.json
| - specification.json
Again, the loadPyTorchDump cannot handle this nested nature of models.
The
model.loadPyTorchDump()function does not handle nested module withinSequentiallayers, and is unable to track the layer names correctly to import the files.Ex: If we have a model from pytorch:
and we dump it using
chai_dumpThe files we get are:
And the
loadPyTorchDumpfunction does not correctly handle these file names since it does not keep track of the "nested" nature of theConv2dlayer since it's inside of theSequentiallayer.A more complicated example that expands on the above:
Gives us the files:
Again, the
loadPyTorchDumpcannot handle this nested nature of models.