[POT] Skip per channel quantizers allining before concats (#12483)

* Skip per channel quantizers allining before concats

* Test coverage

tools/pot/openvino/tools/pot/algorithms/quantization/fake_quantize_configuration.py

@@ -290,6 +290,15 @@ def find_fqs_to_unify(model, config):
                 unified_scales_ops_.append(hw_op)
         return unified_scales_ops_
 
+    def _get_quantizable_per_ch_ops(hw_cfg_):
+        _quantizable_per_ch_ops = []
+        for hw_layer_cfg in hw_cfg_:
+            if 'quantization' in hw_layer_cfg and \
+               all(config['granularity'] == 'perchannel'
+                   for config in hw_layer_cfg['quantization']['activations']):
+                _quantizable_per_ch_ops.append(hw_layer_cfg['type'])
+        return _quantizable_per_ch_ops
+
     def _is_special_unify_conditions(node):
         check_map = {
             'Concat': _is_concat_unify_condition
@@ -310,6 +319,18 @@ def find_fqs_to_unify(model, config):
                 return True
             return False
 
+        def _is_followed_by_quantizable_per_channel(input_node):
+            if _is_quantize_agnostic_op(input_node):
+                concat_stack.extend(get_all_node_outputs(input_node))
+            else:
+                hw_op_type = get_hardware_config_operation_type(input_node, [hw_op['type'] for hw_op in hw_ops])
+                if hw_op_type in per_channel_quantizable:
+                    concat_stack.clear()
+                    logger.debug('Found per channel quantizible op %s %s as Concat %s output',
+                                 input_node.type, input_node.fullname, node.fullname)
+                    return True
+            return False
+
         res = False
         concat_inputs = get_node_inputs(node)
         for concat_input in concat_inputs:
@@ -317,10 +338,17 @@ def find_fqs_to_unify(model, config):
                 logger.debug('Concat %s without FQ or Concat as input will not unified',
                              node.fullname)
                 return res
+
         concat_stack = [node]
         while concat_stack:
             node_to_check = concat_stack.pop()
             res = _is_followed_by_conv(node_to_check)
+        if res:
+            concat_stack = [node]
+            while concat_stack:
+                node_to_check = concat_stack.pop()
+                res = not _is_followed_by_quantizable_per_channel(node_to_check)
+
         return res
 
     def _is_agnostic_branching_op(node_):
@@ -360,6 +388,7 @@ def find_fqs_to_unify(model, config):
                     stack_.append(parent)
 
     hardware_config = load_hardware_config(config)
+    per_channel_quantizable = _get_quantizable_per_ch_ops(hardware_config)
     hw_ops = get_operation_list(hardware_config)
     quantize_agnostic_ops = [op[1] for op in find_operation_matches(QUANTIZE_AGNOSTIC_OPERATIONS, hw_ops)]
     unified_scales_ops = _get_unified_scales_ops(hw_ops)

tools/pot/tests/data/engine_configs/concat_depthwise_model.json

@@ -0,0 +1,47 @@
+{
+    "launchers": [
+        {
+            "framework": "openvino",
+            "adapter": "classification"
+        }
+    ],
+    "datasets": [
+        {
+            "name": "imagenet_1001_classes",
+            "annotation_conversion": {
+                "converter": "imagenet",
+                "annotation_file": "val.txt",
+                "has_background": true
+            },
+            "annotation": "imagenet1001.pickle",
+            "data_source": "ILSVRC2012_img_val",
+            "metrics": [
+                {
+                    "name": "accuracy@top1",
+                    "type": "accuracy",
+                    "top_k": 1,
+                    "reference": 0.7185
+                },
+                {
+                    "name": "accuracy@top5",
+                    "type": "accuracy",
+                    "top_k": 5,
+                    "reference": 0.9069
+                }
+            ],
+            "preprocessing": [
+                {
+                    "type": "crop",
+                    "central_fraction": 0.875
+                },
+                {
+                    "type": "resize",
+                    "size": 224
+                }
+            ],
+            "_command_line_mapping": {
+                "annotation_file": null
+            }
+        }
+    ]
+}

tools/pot/tests/data/models/concat_depthwise_model/concat_depthwise_model.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e771874a16658b9899576e3c877b54a2277272c176edcd0c897906a658981c03
+size 168

tools/pot/tests/data/models/concat_depthwise_model/concat_depthwise_model.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6d6eed00903c057a4770b28645b549201f143c8eab91e756e5d19121f3a7d5a
+size 1119

tools/pot/tests/data/reference_models/concat_depthwise_model_MinMaxQuantization_pytorch.xml

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5c5f76b57e5972208a4ee862208daa1b56938f77996774c9d8d76c6134bbfea
+size 30782

tools/pot/tests/data/reference_unify/concat_depthwise_model_to_unify.json

@@ -0,0 +1 @@
+[]

tools/pot/tests/test_graph_with_stats.py

@@ -17,7 +17,8 @@ from .utils.config import merge_configs
 TEST_MODELS = [
     ('resnet_example', 'pytorch', 'DefaultQuantization', 'CPU'),
     ('resnet_example', 'pytorch', 'RangeSupervision', 'ANY'),
-    ('act_act_example', 'pytorch', 'DefaultQuantization', 'GNA')
+    ('act_act_example', 'pytorch', 'DefaultQuantization', 'GNA'),
+    ('concat_depthwise_model', 'pytorch', 'MinMaxQuantization', 'CPU'),
 ]
 
 

tools/pot/tests/test_unify_scales.py

@@ -17,9 +17,10 @@ from .utils.config import get_engine_config, merge_configs
 from .utils.path import TEST_ROOT
 
 TEST_MODELS = [
-    ('mobilenet-v2-pytorch', 'pytorch', 'MinMaxQuantization', 'performance'),
-    ('resnet-50-tf', 'tf', 'DefaultQuantization', 'performance'),
-    ('octave-resnet-26-0.25', 'mxnet', 'DefaultQuantization', 'accuracy'),
+    ('mobilenet-v2-pytorch', 'pytorch', 'MinMaxQuantization', 'performance', 'VPU'),
+    ('resnet-50-tf', 'tf', 'DefaultQuantization', 'performance', 'VPU'),
+    ('octave-resnet-26-0.25', 'mxnet', 'DefaultQuantization', 'accuracy', 'VPU'),
+    ('concat_depthwise_model', 'pytorch', 'MinMaxQuantization', 'accuracy', 'CPU'),
 ]
 
 REFERENCES_PATH = TEST_ROOT / 'data/reference_unify'
@@ -32,13 +33,13 @@ def _params(request):
 
 
 def test_unify_scales(_params, tmp_path, models):
-    model_name, model_framework, algorithm, preset = _params
+    model_name, model_framework, algorithm, preset, device = _params
 
     algorithm_config = Dict({
         'algorithms': [{
             'name': algorithm,
             'params': {
-                'target_device': 'VPU',
+                'target_device': device,
                 'preset': preset,
                 'stat_subset_size': 2
             }