New old error with force_range_in in custom yolo11 model

General Discussion
1

I am trying to quantize a custom model.

My script:
model_scripts_command = f"""normalization1 = normalization([0.0, 0.0, 0.0], [255.0, 255.0, 255.0])

model_optimization_config(calibration, calibset_size={calibset_size}, batch_size={batch_size})

performance_param(compiler_optimization_level=max)

model_optimization_flavor(optimization_level=2)

pre_quantization_optimization(activation_clipping, layers={{*}}, mode=percentile, clipping_values=[0.01, 99.99])

pre_quantization_optimization(weights_clipping, layers={{*}}, mode=percentile, clipping_values=[0.01, 99.99])

quantization_param([‘fastnas_flops40_1024x1280_har/matmul2’],force_range_in=[0.0,0.947,-9.935,10.146],force_range_index=[0,1])“”"

runner.load_model_script(model_scripts_command)

At first, this line wasn’t there:

quantization_param([‘fastnas_flops40_1024x1280_har/matmul2’],force_range_in=[0.0,0.947,-9.935,10.146],force_range_index=[0,1])

But after getting that error, I added it:

File “/opt/conda/lib/python3.11/site-packages/hailo_model_optimization/acceleras/atomic_ops/matmul_op.py”, line 223, in create_hw_params

raise AccelerasUnsupportedError(

hailo_model_optimization.acceleras.utils.acceleras_exceptions.AccelerasUnsupportedError: layer fastnas_flops40_1024x1280_har/matmul2 does not support shift delta. To overcome this issue you should force larger range at the inputs of the layer using command quantization_param([layer_name], force_range_in=[range_min0, range_max0, range_min1, range_max1], force_range_index=[0, 1]) current range of input 0 is [0.001, 0.761] and input 1 is [-7.985, 8.155].You should increase the multiplication of these ranges by a factor of 1.548, e.g. you can apply factor of sqrt(1.548) to both inputs:

quantization_param([fastnas_flops40_1024x1280_har/matmul2], force_range_in=[0.001, 0.947, -9.935, 10.146], force_range_index=[0,1])

Now error:
File “/opt/conda/lib/python3.11/site-packages/pydantic/v1/class_validators.py”, line 302, in

return lambda cls, v, values, field, config: validator(cls, v, values=values, field=field, config=config)

^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/opt/conda/lib/python3.11/site-packages/hailo_model_optimization/acceleras/model_optimization_config/mo_config_layer.py”, line 794, in validate_force_range_index

if 2 * len(v) != len(values[“force_range_in”]):

~~~~~~^^^^^^^^^^^^^^^^^^

KeyError: ‘force_range_in’

I couldn’t find a description of `force_range_in` in the documentation—only `force_range_out`.

I tried adding `‘` and `’'` to the layer name and splitting it into two indices, but nothing works; I’ve seen several posts on the forum about this error, but none offered a solution. Has the information been updated? I examined this layer using dfc-studio. It exists, and the name is exactly as specified.

Python version: 3.11

Hailo DFC version: 3.33.0

2

SOLUTION:

Since you need to constrain the input data, you also need to constrain the output data for the layer before the specified layer. This is not the first time I’ve encountered this error, and the problematic layer is always matmul2, followed by softmax1 + activation1. There’s no point in constraining softmax, as it’s already constrained. Activation1 is a different story. You can add the following line to model_script:

quantization_param([‘<model_name>/activation1’], force_range_out=[<min>, <max>])

And everything will work!

By the way, if you add a softmax constraint, the parser won’t find it.