Once constructing data like this:
[[[[-0.64189893]]
[[-0.5155563 ]]
[[-0.5729615 ]]
...
[[-0.00979282]]
[[-0.00870954]]
[[ 0.00463101]]]
...
[[[ 1.4152465 ]]
[[-0.77834904]]
[[-0.8964073 ]]
...
[[-0.00979282]]
[[-0.00870954]]
[[ 0.00463101]]]]
And the input shape is (1000, 1000, 1, 1). It is all fine to train the model using keras and output model in .h5 format. The model structure is as follows:
Conv2D->BN->Conv2D->BN->Flatten->Dense->BN->Dense->BN->Dense.
But when using tvm to compile this keras model.
model = keras.models.load_model(model_path)
shape_dict = {'input': (1, 1000, 1, 1)}
mod, params = relay.frontend.from_keras(model, shape_dict)
# compile the model
target = "llvm"
ctx = tvm.cpu(0)
with relay.build_config(opt_level=3):
....
)
I’m getting the error below:
TVMError: Check failed: checked_type.as<IncompleteTypeNode>() == nullptr: Cannot resolve type of Var(conv2d_1_input) at (nullptr)
I’m not sure if this is intended behaviour or not. And if this is intended behaviour, I’m curious what use case does it support. And how could I solve this problem?