Cannot generate scheduler after quantization

Required prerequisites

Aidge_quantization

What commit version of aidge do you use

aidge_core: 1e21aa51
aidge_quantization: b164c89a843fac5897f14a6e6bf59d15fe539a06 (no cuda version for aidge_quantization)

Problem description

When I fuse Conv and Scaling. The scheduler cannot compute the dims of the new metaOP. Before that, I also have an error who tell me that the input isn't connected, or when i save my model in .mmd i can see that an input is connected to my Conv.

aidge_core.fuse_to_metaops(model, "Conv2D->Scaling", "Cmsis_Nn_Conv")
model.save("model_af_fuse")

scheduler = aidge_core.SequentialScheduler(model)
scheduler.generate_scheduling()

Replaced 0 (out of 0) matching sub-graph with meta operators
Conv_Op::setBackend(): could not set backend for weight input, because input is not connected
Replaced 1 (out of 1) matching sub-graph with meta operators
Assertion failed: requiredSize.type == Elts_t::Data in /local2/is148265/wb274724/STM32_dev/dev/superpoint/aidge/aidge/aidge_core/src/scheduler/Scheduler.cpp:442
Cannot generate memory with token-based producer-consumer model for node  (of type Cmsis_Nn_Conv).
Traceback (most recent call last):

Reproducible example code

np.random.seed(123)

##########################################################
# Création du model qui contient qu'une seule Convolution
##########################################################

model = aidge_core.sequential([
    aidge_core.Conv2D(in_channels=64, out_channels=10, kernel_dims = [3,3], name='Conv')
    ], name = "aha")
model.set_datatype(aidge_core.dtype.float32)



# Paramètres aléatoire avec seed pour les poids et bias 
for node in model.get_nodes():
    if node.type() == "Producer":
        prod_op = node.get_operator()
        value = prod_op.get_output(0)
        tuple_out = node.output(0)[0]
        # No conv in current network
        if tuple_out[0].type() == "Conv" and tuple_out[1]==1:
            # Conv weight
            aidge_core.xavier_uniform_filler(value)
        elif tuple_out[0].type() == "Conv" and tuple_out[1]==2:
            # Conv bias
            aidge_core.constant_filler(value, 0.01)

scheduler = aidge_core.SequentialScheduler(model)
# Configuration for the model + forward dimensions
dims = [1,64,32,32]
input_tab = np.random.uniform(low=0.0, high=1.0, size=dims)
aidge_export_cpp.generate_input_file(array_name="simple_conv_export", array=input_tab.reshape(-1), export_folder="simple_conv_export")

aidge_core.remove_flatten(model)
aidge_core.fuse_batchnorm(model)

# input_node = aidge_core.Producer(dims, "input")
# input_node.add_child(model)
# model.add(input_node)

model.compile("cpu", aidge_core.dtype.float32,dims=[dims])


scheduler = aidge_core.SequentialScheduler(model)


##########################################################
# Quantification avec des samples aléatoires
##########################################################

def propagate(model, scheduler, sample):
    # Setup the input
    sample = np.reshape(sample,dims)
    input_tensor = aidge_core.Tensor(sample)
    input_tensor.set_backend("cpu")
    input_tensor.set_datatype(aidge_core.dtype.float32)
    # input_tensor.get_operator().set_output(0, input_tensor)
    # Run the inference
    scheduler.forward(True, [input_tensor])    
    # Gather the results
    output_node = model.get_output_nodes().pop()
    output_tensor = output_node.get_operator().get_output(0)
    return np.array(output_tensor)

# Create Samples
samples = []
for _ in range(10):
    samples.append(np.random.uniform(low=0.0, high=1.0, size=dims))
    
    
for i in range(10):
    output_array = propagate(model, scheduler, samples[i])

print('Propagate ok ')

NB_SAMPLES  = 10
NB_BITS     = 8

tensors = []
for sample in samples[0:NB_SAMPLES]:
    tensor = aidge_core.Tensor(sample)
    tensors.append(tensor)
    
aidge_quantization.quantize_network(model, NB_BITS, tensors, use_cuda = False, apply_rounding = True)

scheduler = aidge_core.SequentialScheduler(model)

model.save("model_af_quant")
print("PTQ done")

##########################################################
# Change dtype
##########################################################
aidge_core.expand_metaops(model)

aidge_core.fuse_to_metaops(model, "FC->Scaling->ReLU?", "Cmsis_Nn_FC")
aidge_core.fuse_to_metaops(model, "Conv2D->Scaling", "Cmsis_Nn_Conv")
model.save("model_af_fuse")

scheduler = aidge_core.SequentialScheduler(model)
scheduler.generate_scheduling()


aidge_export_arm_cortexm.export("simple_conv_export", model, scheduler, board="stm32h7")

Cannot generate scheduler after quantization

Required prerequisites

What commit version of aidge do you use

Problem description

Reproducible example code

Designs

Child items ...

Activity