[TF FE] Use regular Convolution in case dynamic input channels (#19253)

* [TF FE] Use regular Convolution in case dynamic input channels

This solution is aligned with the legacy frontend but it has limitation.
This is a temporal solution until the core obtains ShapeOf evaluator.

Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>

* Remove unused variable from the test

Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>

* Fix unit-test

* Update mo unit-test

---------

Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>

src/frontends/tensorflow/tests/convert_tricky_models.cpp

@@ -749,3 +749,29 @@ TEST_F(FrontEndConversionWithReferenceTestsF, TF1IfWithNonExistentOpInBranch) {
         model_ref = make_shared<Model>(OutputVector{mul}, ParameterVector{y, ind});
     }
 }
+
+TEST_F(FrontEndConversionWithReferenceTestsF, ConvolutionWithDynamicInputChannel) {
+    // This test aims to check conversion of a model with convolution of dynamic input channel
+    // Namely, the resulted model must contain the regular convolution, not grouped convolution
+    { model = convert_model("conv_with_dynamic_input_channel"); }
+    {
+        auto input = make_shared<Parameter>(f32, PartialShape{Dimension::dynamic(), 10, 10, Dimension::dynamic()});
+
+        auto transpose_order = make_shared<Constant>(i64, Shape{4}, vector<int32_t>{0, 3, 1, 2});
+        auto transpose = make_shared<Transpose>(input, transpose_order);
+
+        auto filter = make_shared<Constant>(element::f32, Shape{6, 6, 3, 3}, vector<float>(6 * 6 * 3 * 3, 0.0f));
+        auto conv = make_shared<Convolution>(transpose,
+                                             filter,
+                                             Strides{1, 1},
+                                             CoordinateDiff{0, 0},
+                                             CoordinateDiff{0, 0},
+                                             Strides{1, 1},
+                                             op::PadType::SAME_UPPER);
+
+        auto transpose_order_back = make_shared<Constant>(i64, Shape{4}, vector<int32_t>{0, 2, 3, 1});
+        auto transpose_back = make_shared<Transpose>(conv, transpose_order_back);
+
+        model_ref = make_shared<Model>(OutputVector{transpose_back}, ParameterVector{input});
+    }
+}

src/frontends/tensorflow/tests/test_models/gen_scripts/generate_conv_with_dynamic_input_channel.py

@@ -0,0 +1,19 @@
+# Copyright (C) 2023 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+import sys
+
+import tensorflow as tf
+
+# Create the graph and model
+tf.compat.v1.reset_default_graph()
+with tf.compat.v1.Session() as sess:
+    filter = tf.constant(value=0, shape=[3, 3, 6, 6], dtype=tf.float32)
+    input = tf.compat.v1.placeholder(dtype=tf.float32, shape=[None, 10, 10, None], name='input')
+    conv = tf.raw_ops.Conv2D(input=input,
+                             filter=filter,
+                             strides=[1, 1, 1, 1],
+                             padding='SAME')
+    tf.compat.v1.saved_model.simple_save(sess, os.path.join(sys.argv[1], "conv_with_dynamic_input_channel"),
+                                         inputs={'input': input}, outputs={'conv': conv})

src/frontends/tensorflow_common/src/utils.cpp

@@ -208,11 +208,7 @@ OutputVector translate_convolution_op(const frontend::NodeContext& node, size_t
     }
 
     Output<Node> conv;
-    if (input_channels_static && num_groups == 1) {
+    if (input_channels_static && num_groups > 1) {
-        // regular convolutional operation
-        // we assume that input channel size will not be changed if they are already static
-        conv = make_shared<Convolution>(input, filter, strides, pads_begin, pads_end, dilations, auto_pad);
-    } else {
         // grouped convolutional operation
         // compute input channels given from the input and the filter
         // and number of groups required to split the filter
@@ -233,6 +229,12 @@ OutputVector translate_convolution_op(const frontend::NodeContext& node, size_t
         auto filter_new_shape = make_shared<Concat>(OutputVector{num_groups, filter_new_cout, shape_cin_xy}, 0);
         auto new_filter = make_shared<Reshape>(filter, filter_new_shape, false);
         conv = make_shared<GroupConvolution>(input, new_filter, strides, pads_begin, pads_end, dilations, auto_pad);
+    } else {
+        // assumption to use regular convolution for all other cases is taken from the legacy frontend
+        // this solution is sufficient for all observed models in the validation
+        // in general, it has limitation and it needs to use grouped convolution when num_groups is not static
+        // 118107: remove this assumtpion when it obtains complete shape propagation in the core
+        conv = make_shared<Convolution>(input, filter, strides, pads_begin, pads_end, dilations, auto_pad);
     }
 
     convert_nchw_to_nhwc(is_nhwc, conv, Rank(spatial_dims_num + 2));

tools/mo/unit_tests/moc_tf_fe/conversion_with_layout_test.py

@@ -68,7 +68,7 @@ class TestConversionWithBatchAndLayout(unittest.TestCase):
         *[
             (
                     "model_with_convolution_dynamic_rank.pbtxt", 7, "x(n???),kernel(????)",
-                    {"x": PartialShape([7, Dimension.dynamic(), Dimension.dynamic(), Dimension.dynamic()]),
+                    {"x": PartialShape([7, Dimension.dynamic(), Dimension.dynamic(), 3]),
                      "kernel": PartialShape([2, 2, 3, 1])},
             ),
             (