fix len() of unsized object

[XinyuWuu]

Changes

Added new logic when hasattr(tensor_collector.get_statistics().mean_values, "__len__") == False.

Reason for changes

if tensor_collector.get_statistics().mean_values is a nncf.tensor with shape () of dim 0, extend function call will fail.

Tests

tests/onnx/quantization/test_bias_correction.py is tested.
I am not sure how to test it as I encountered this bug during quantization of a network from one of my robotics projects. I guess the bug will only be triggled by certain networks and certain calibration datasets, therefor no new tests are added to varify the fix.

[daniil-lyakhov]

Greetings, @XinyuWuu! Thank you for your interest in our project!

I afraid we need to find the root cause for this bug, looks like the effect you are describing is a consequence of the deeper error. MeanTensorStatistic, which unexpectedly returns a scalar, indicate a problems inside.
It is combined only with the MeanAggregator which should ALWAYS return a NNCFTensor (https://github.com/openvinotoolkit/nncf/blob/develop/nncf/torch/tensor_statistics/collectors.py#L293-L321). Combined with MeanPerChReducer and BatchMeanReducer, the aggregator should always receive NNCFTensors as an input. (For further details on reducers/aggregators, please see https://github.com/openvinotoolkit/nncf/blob/develop/nncf/experimental/common/tensor_statistics/collectors.py#L184-L191)

So please answer following question to clarify the problem:

• Which NNCF backend did you use?
• Could you please attach a minimal reproducer?

We indeed need to test it, I'll help you with that

[XinyuWuu]

Glad to spot something real.
I am using the OpenVino backend. I recorded some data and zipped them with the model for reproduce:
test.zip
test code:

import pickle

import nncf
import openvino as ov
import torch

ovcore = ov.Core()
ovmodel = ovcore.read_model("test.xml")
with open("test.pkl", "rb") as f:
    calibration_dataset = pickle.load(f)
calibration_dataset = nncf.Dataset(
    calibration_dataset,
    lambda x: {k: torch.tensor(v).to(torch.float32) for k, v in x.items()},
)
quantized_model = nncf.quantize(
    ovmodel,
    calibration_dataset,
    fast_bias_correction=False,
    target_device=nncf.TargetDevice.CPU,
    model_type=nncf.ModelType.TRANSFORMER,
)

The return type of MeanTensorStatistic is actully correct <class 'nncf.tensor.tensor.Tensor'> but the shape is ().

[MaximProshin]

@daniil-lyakhov , please clarify the next step here.