[MO] Range output_type correction for FP16 (openvinotoolkit#6590)

* added ChangeRangeOutputType.py

* applied review comments

* corrected error message - warn user to use FP32

* renamed ChangeCastOutputType.py et ell.

* merged ChangeRangeOutputType.py, ChangeCastOutputType.py into a singe file

* corrections

* typo fix

* applied comments: faster find_and_replace loop, wording correction

model-optimizer/automation/package_BOM.txt

@@ -11,7 +11,7 @@ extensions/back/__init__.py
 extensions/back/AvgPool.py
 extensions/back/blob_normalizer.py
 extensions/back/CellNormalizer.py
-extensions/back/ChangeCastOutputType.py
+extensions/back/ChangeOutputTypeAttributes.py
 extensions/back/ClampNormalizer.py
 extensions/back/compress_quantized_weights.py
 extensions/back/ConvolutionNormalizer.py

model-optimizer/extensions/back/ChangeOutputTypeAttributes.py

@@ -0,0 +1,100 @@
+# Copyright (C) 2018-2021 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import logging as log
+
+import numpy as np
+
+from mo.back.replacement import BackReplacementPattern
+from mo.graph.graph import Graph
+from mo.graph.graph import Node
+from mo.middle.passes.convert_data_type import data_type_str_to_np
+from mo.utils.error import Error
+
+operations_with_data_type_attributes = {
+    'Cast': {'attr_name': 'dst_type', 'in_ports_to_check': (0,)},
+    'Range': {'attr_name': 'output_type', 'in_ports_to_check': (0, 1, 2)},
+}
+
+
+class ChangeOutputTypeAttributes(BackReplacementPattern):
+    """
+    The transformation changes output type for the specific operations defined in the
+    operations_with_data_type_attributes dictionary if one of the following conditions is met:
+    - The operation output type is fp64. Since not all plugins support fp64 data type it is converted to fp32.
+    - Changes output type from fp32 to fp16 (and ensure that this is possible) when generating fp16 IR.
+    - Keep operation output type equal to fp32 for operations located in the shape calculation sub-graphs to
+    avoid floating point overflow.
+    """
+    enabled = True
+    force_shape_inference = True
+
+    def run_after(self):
+        from extensions.back.MarkNodesWithShapeValues import MarkNodesWithShapeValues
+        return [MarkNodesWithShapeValues]
+
+    def run_before(self):
+        return []
+
+    def find_and_replace_pattern(self, graph: Graph):
+        ir_data_type = data_type_str_to_np(graph.graph['cmd_params'].data_type)
+
+        for node in graph.get_op_nodes():
+            if node.op in operations_with_data_type_attributes:
+                dst_type = operations_with_data_type_attributes[node.op]['attr_name']
+                node_name = node.soft_get('name', node.id)
+                assert node.has_valid(dst_type), '{} attribute is missing for node {}'.format(dst_type, node_name)
+
+                final_type = None
+                if node[dst_type] == np.float64:
+                    final_type = np.float32
+
+                if node[dst_type] in [np.float32, np.float64] and ir_data_type == np.float16 and \
+                        not node.has_and_set('returns_shape_value'):
+                    final_type = np.float16
+                elif node.has_and_set('returns_shape_value') and node.dst_type == np.float16:
+                    # return back FP32 for all nodes with shape values
+                    final_type = np.float32
+
+                if final_type is not None:
+                    log.warning('Change data type from {} to {} for node {}'.format(node[dst_type], final_type,
+                                                                                    node_name))
+                    node[dst_type] = final_type
+
+                if final_type == np.float16:
+                    assert_that_is_castable_to_fp16(node)
+
+
+def assert_that_is_castable_to_fp16(node: Node):
+    op_name = node.soft_get('op')
+    node_name = node.soft_get('name', node.id)
+
+    for i in operations_with_data_type_attributes[op_name]['in_ports_to_check']:
+        val = node.in_port(i).data.get_value()
+        if val is None:
+            return
+
+        if np.any(val > np.finfo(np.float16).max) or np.any(val < np.finfo(np.float16).min):
+            raise Error("Try to convert with --data_type=FP32 argument. "
+                        "This model can not be converted to FP16 precision, since "
+                        "'{}' node value {} exceeds FP16 allowed limits: [{}, {}]"
+                        .format(node_name, val, np.finfo(np.float16).min, np.finfo(np.float16).max))
+        # further this input values will be rewritten since force_shape_inference=True
+        node.in_port(i).data.set_value(val.astype(np.float16))
+
+    original_output = node.out_port(0).data.get_value()
+    node.infer(node)
+    casted_output = node.out_port(0).data.get_value()
+    original_output_len = len(original_output) if hasattr(original_output, '__len__') else None
+    casted_output_len = len(casted_output) if hasattr(casted_output, '__len__') else None
+
+    if original_output_len != casted_output_len:
+        raise Error("Try to convert with --data_type=FP32 argument. "
+                    "This model can not be converted to FP16 precision, since "
+                    "after conversion of '{}' node to FP16 output shape {} differs from the original {}."
+                    .format(node_name, casted_output_len, original_output_len))
+
+    diff_count = np.count_nonzero(np.subtract(original_output, casted_output) > 1.e-4)
+    if diff_count > 0:
+        log.warning("{} elements of {} of Range node '{}' output differ from the original values while "
+                    "converting network to FP16 precision".format(diff_count, len(original_output), node_name))

model-optimizer/extensions/back/MarkNodesWithShapeValues.py

@@ -23,7 +23,6 @@
     'Tile': [1],  # repeats input
     'TopK': [1],  # K input
     'Pad': [1, 2],  # pads_begin, pads_end
-    'Range': [0, 1, 2],  # start, stop, step inputs
     'OneHot': [1],  # depth input
 }
 

model-optimizer/extensions/ops/range.py

@@ -1,8 +1,6 @@
 # Copyright (C) 2018-2021 Intel Corporation
 # SPDX-License-Identifier: Apache-2.0
 
-import logging as log
-
 import numpy as np
 
 from mo.graph.graph import Node, Graph

model-optimizer/unit_tests/extensions/back/ChangeOutputTypeAttributes_test.py

@@ -0,0 +1,122 @@
+# Copyright (C) 2018-2021 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import unittest
+from copy import deepcopy
+
+import numpy as np
+
+from extensions.back.ChangeOutputTypeAttributes import ChangeOutputTypeAttributes
+from extensions.ops.Cast import Cast
+from extensions.ops.range import Range
+from mo.front.common.partial_infer.utils import float32_array
+from mo.middle.passes.convert_data_type import convert_blobs, data_type_str_to_np
+from mo.middle.passes.infer import partial_infer
+from mo.utils.error import Error
+from mo.utils.ir_engine.compare_graphs import compare_graphs
+from unit_tests.utils.graph import build_graph, result, regular_op_with_empty_data, connect
+from unit_tests.utils.graph import valued_const_with_data
+
+
+class ChangeOutputTypeAttributesTests(unittest.TestCase):
+
+    def test_range_correct_case(self):
+        graph, graph_ref = build_range_test_graphs(start=0, limit=10, delta=1, dst_type_str='FP16')
+        ChangeOutputTypeAttributes().find_and_replace_pattern(graph)
+        (flag, resp) = compare_graphs(graph, graph_ref, 'res', check_op_attrs=True)
+        self.assertTrue(flag, resp)
+
+    # starting from ~1000 FP16 absolute difference between neighbor values is more than 1
+    # fails because of shape inconsistency
+    def test_range_different_values(self):
+        graph, graph_ref = build_range_test_graphs(start=0, limit=50000, delta=1, dst_type_str='FP16')
+        self.assertRaises(Error, ChangeOutputTypeAttributes().find_and_replace_pattern, graph)
+
+    def test_range_out_of_fp16_max(self):
+        graph, graph_ref = build_range_test_graphs(start=0, limit=100000, delta=1, dst_type_str='FP16')
+        self.assertRaises(Error, ChangeOutputTypeAttributes().find_and_replace_pattern, graph)
+
+    def test_range_out_of_fp16_min(self):
+        graph, graph_ref = build_range_test_graphs(start=0, limit=-100000, delta=-1, dst_type_str='FP16')
+        self.assertRaises(Error, ChangeOutputTypeAttributes().find_and_replace_pattern, graph)
+
+    def test_cast_correct_case(self):
+        input_data = np.array([0, 1000, 4, 9, 0])
+        graph, graph_ref = build_cast_test_graphs(input_data, dst_type_str='FP16')
+        ChangeOutputTypeAttributes().find_and_replace_pattern(graph)
+        (flag, resp) = compare_graphs(graph, graph_ref, 'res', check_op_attrs=True)
+        self.assertTrue(flag, resp)
+
+    def test_cast_out_of_fp16_max(self):
+        input_data = np.array([0, 100000, 4, 9, 0])
+        graph, graph_ref = build_cast_test_graphs(input_data, dst_type_str='FP16')
+        self.assertRaises(Error, ChangeOutputTypeAttributes().find_and_replace_pattern, graph)
+
+    def test_cast_out_of_fp16_min(self):
+        input_data = np.array([0, -100000, 4, 9, 0])
+        graph, graph_ref = build_cast_test_graphs(input_data, dst_type_str='FP16')
+        self.assertRaises(Error, ChangeOutputTypeAttributes().find_and_replace_pattern, graph)
+
+
+def build_range_test_graphs(start=0, limit=10, delta=1, dst_type_str='FP16'):
+    nodes = {
+        **valued_const_with_data('start', float32_array(start)),
+        **valued_const_with_data('limit', float32_array(limit)),
+        **valued_const_with_data('delta', float32_array(delta)),
+        **regular_op_with_empty_data('range', {'type': 'Range', 'op': 'Range',
+                                               'output_type': np.float32,
+                                               'infer': Range.infer}),
+        **result('res'),
+    }
+
+    nodes_ref = deepcopy(nodes)
+    nodes_ref.update({
+        **regular_op_with_empty_data('range', {'type': 'Range', 'op': 'Range',
+                                               'output_type': data_type_str_to_np(dst_type_str),
+                                               'infer': Range.infer}),
+    })
+
+    edges = [
+        *connect('start', '0:range'),
+        *connect('limit', '1:range'),
+        *connect('delta', '2:range'),
+        *connect('range', 'res'),
+    ]
+    graph = build_graph(nodes, edges)
+    graph_ref = build_graph(nodes_ref, edges)
+
+    graph = partial_infer(graph)
+
+    graph.graph['cmd_params'].data_type = dst_type_str
+    convert_blobs(graph, dst_type_str)
+    return graph, graph_ref
+
+
+def build_cast_test_graphs(input_data, dst_type_str='FP16'):
+    nodes = {
+        **valued_const_with_data('input', float32_array(input_data)),
+        **regular_op_with_empty_data('cast', {'type': 'Convert', 'op': 'Cast',
+                                              'dst_type': np.float32,
+                                              'infer': Cast.infer}),
+        **result('res'),
+    }
+
+    nodes_ref = deepcopy(nodes)
+    nodes_ref.update({
+        **regular_op_with_empty_data('cast', {'type': 'Convert', 'op': 'Cast',
+                                              'dst_type': data_type_str_to_np(dst_type_str),
+                                              'infer': Cast.infer}),
+    })
+
+    edges = [
+        *connect('input', 'cast'),
+        *connect('cast', 'res'),
+    ]
+    graph = build_graph(nodes, edges)
+    graph_ref = build_graph(nodes_ref, edges)
+
+    graph = partial_infer(graph)
+
+    graph.graph['cmd_params'].data_type = dst_type_str
+    convert_blobs(graph, dst_type_str)
+    return graph, graph_ref