PaddedConv1D runtime error

What commit version of aidge do you use

• aidge 0.2.2
• aidge_backend_cpu 0.2.3
• aidge_core 0.2.2

Hi, I need a PaddedConv1D for my application so I tried a minimalist inference, but get a bad_cast error.

• Did I do something wrong ?
• Is that Operator implemented for export to cpp or arm_cortexm ?

Thanks a million !

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)

---> 30 scheduler.forward(data=[input_tensor])

RuntimeError: std::bad_cast

Reproducible example code

import numpy as np
import aidge_core

# Create node PaddedConv1D
IN_CHANNELS, OUT_CHANNELS= 1, 32
graph_conv = aidge_core.GraphView()
graph_conv.add(aidge_core.PaddedConv1D(in_channels=IN_CHANNELS, out_channels=OUT_CHANNELS, kernel_dims=[3], name="padded_conv1d_0", padding_dims=[1,1], no_bias=True))

# add a zero bias
bias = aidge_core.Producer(aidge_core.Tensor(np.zeros([32], dtype=np.float32)), "bias")
bias.add_child(graph_conv.get_node("padded_conv1d_0"), other_in_id=2)
graph_conv.add(bias)

# visualize
#graph_conv.save("conv")
#visualize_mmd("conv.mmd")

# create input
batch_size = 64
signal_size = 16
x = np.random.rand(batch_size, IN_CHANNELS, signal_size).astype(np.float32) #(1 channel on first vector)
input_tensor=aidge_core.Tensor(x)
print(input_tensor.dims())

# compile
graph_conv.compile("cpu", aidge_core.dtype.float32,dims=[input_tensor.dims()])
## Create SCHEDULER
scheduler = aidge_core.SequentialScheduler(graph_conv)

# Run inference !
scheduler.forward(data=[input_tensor])

changed the description

changed the description

@olivierbichler

# add a zero bias
bias = aidge_core.Producer(aidge_core.Tensor(np.zeros([32], dtype=np.float32)), "bias")
bias.add_child(graph_conv.get_node("padded_conv1d_0"), other_in_id=2)
graph_conv.add(bias)

It looks like there is again a problem with no bias, if we don't add an empty bias, then the GraphView consider the bias as an input.

@idealbuq got the same problem.

Thanks a lot for your answer, indeed I had misunderstood the no_bias option.

However:

• if I set no_bias to True, I still see an extra input in my graph (the bias input from the conv node). That is why I need to provide the extra bias node (as shown above)
• If I set no_bias to False, I see an automatic bias node attached to the conv1d node (as shown bellow), so no need to create the bias node myself.
• But in both cases I still get the std::bad_cast error

Is there something I can do ?

Example with no_bias = False that still produces the error :

# Create node PaddedConv1D
IN_CHANNELS, OUT_CHANNELS= 1, 32
graph_conv = aidge_core.GraphView()
graph_conv.add(aidge_core.PaddedConv1D(in_channels=IN_CHANNELS, out_channels=OUT_CHANNELS, kernel_dims=[3]))

# create input
batch_size, signal_size = 64, 16
input_tensor=aidge_core.Tensor(np.random.rand(batch_size, IN_CHANNELS, signal_size).astype(np.float32))

# compile and scheduler
graph_conv.compile("cpu", aidge_core.dtype.float32,dims=[input_tensor.dims()])
scheduler = aidge_core.SequentialScheduler(graph_conv)

# Run inference
scheduler.forward(data=[input_tensor])

mentioned in commit aidge_backend_cpu@10ef1960

This was fixed in #172 (closed). It is not related to no bias.

closed

added Bug 🐛 label

@olivierbichler That is nice it is fixed but I can't compile aidge/dev, cf my issue #159 (closed)

What is the workaround to get to use the latest version ?

Thx, Louis