Add NNCFProfiler implementation

[avolkov-intel]

Changes

Added implementation of profiler to collect and compare activations value of openVino model to investigate cause of accuracy degradation.

Reason for changes

Related tickets

162317

Tests

[ljaljushkin]

Thank you for the contribution, Anatoly!
Please consider adding visualization functionality and enhancing the scalability

[SearchSavior]

Wonderful addition!!!!

@avolkov-intel once merged I will test on

Hermes-14B
Wayfarer-2-12B

Thanks you for your work.

[daniil-lyakhov]

Statistic collection API

Suggested change

	# Extract and convert collected statistics to numpy arrays
	result: ActivationData = {}
	for layer_name, statistic_points_list in statistics_aggregator.statistic_points.items():
	# Extract input activations (index 1 in statistic_points_list)
	in_container = list(
	statistic_points_list[1].algorithm_to_tensor_collectors["collect"][0].aggregators.values()
	)[0]._container
	in_vals = [np.array(elem.data) for elem in in_container]
	# Extract output activations (index 0 in statistic_points_list)
	out_container = list(
	statistic_points_list[0].algorithm_to_tensor_collectors["collect"][0].aggregators.values()
	)[0]._container
	out_vals = [np.array(elem.data) for elem in out_container]
	result[layer_name] = {"in": in_vals, "out": out_vals}
	# Extract and convert collected statistics to numpy arrays
	result: ActivationData = defaultdict(dict)
	target_type_to_str = {
	TargetType.PRE_LAYER_OPERATION: "in",
	TargetType.POST_LAYER_OPERATION: "out",
	}
	for _, statistic_point, tensor_collector in statistic_points.get_tensor_collectors():
	if statistic_point.target_point.type not in target_type_to_str:
	msg = f"Unsupported target type: {statistic_point.target_point.type}"
	raise RuntimeError(msg)
	insert_type = target_type_to_str[statistic_point.target_point.type]
	layer_name = statistic_point.target_point.target_node_name
	stats = tensor_collector.get_statistics().values
	result[layer_name][insert_type] = [np.array(elem.data) for elem in stats]

[daniil-lyakhov]

Sort nodes in topographical order. Currently nodes are sorted by lexicographical order

Suggested change

	target_ops = [graph.get_node_by_key(name) for name in target_names]
	target_ops = []
	for node in graph.topological_sort():
	if len(target_ops) == len(target_names):
	break
	if node.node_key in target_names:
	target_ops.append(node)

[daniil-lyakhov]

I'm suggesting to use following model-agnostic sniped there

from optimum.gptq.data import get_dataset
from optimum.gptq.data import prepare_dataset

dataset = "wikitext2"
seqlen = 50
nsamples = 2
calibration_dataset = get_dataset(dataset, tokenizer, seqlen=seqlen, nsamples=nsamples)
calibration_dataset = prepare_dataset(calibration_dataset)
quantization_dataset = nncf.Dataset(calibration_dataset, lambda x: model.prepare_inputs(**x))