Merge branch 'master' of https://github.com/onnx/tensorflow-onnx into bench2

Author: sdpython

README.md

@@ -18,7 +18,7 @@ The common issues we run into we try to document here [Troubleshooting Guide](Tr
 
 | Build Type | OS | Python | Tensorflow | ONNX opset | Status |
 | ---        | ---    | ---    | ---        | ---        | ---    |
-| Unit Test - Basic | Linux, MacOS<sup>\*</sup>, Windows<sup>\*</sup> | 3.6-3.9 | 1.12-1.15, 2.1-2.5 | 8-15 | [![Build Status](https://dev.azure.com/tensorflow-onnx/tensorflow-onnx/_apis/build/status/unit_test?branchName=master)](https://dev.azure.com/tensorflow-onnx/tensorflow-onnx/_build/latest?definitionId=16&branchName=master) |
+| Unit Test - Basic | Linux, MacOS<sup>\*</sup>, Windows<sup>\*</sup> | 3.6-3.9 | 1.12-1.15, 2.1-2.5 | 8-14 | [![Build Status](https://dev.azure.com/tensorflow-onnx/tensorflow-onnx/_apis/build/status/unit_test?branchName=master)](https://dev.azure.com/tensorflow-onnx/tensorflow-onnx/_build/latest?definitionId=16&branchName=master) |
 | Unit Test - Full | Linux, MacOS, Windows | 3.6-3.9 | 1.12-1.15, 2.1-2.5 | 8-14 | [![Build Status](https://dev.azure.com/tensorflow-onnx/tensorflow-onnx/_apis/build/status/unit_test-matrix?branchName=master)](https://dev.azure.com/tensorflow-onnx/tensorflow-onnx/_build/latest?definitionId=18&branchName=master) | |
 <br/>
 

ci_build/azure_pipelines/keras2onnx_unit_test.yml

@@ -119,7 +119,7 @@ jobs:
     vmImage: 'vs2017-win2016'
   strategy:
     matrix:
-      # No python 2.x since no available ONNX package for Windows
+      ############ TF Keras Unit Tests ############
       Python36-onnx1.2:
         python.version: '3.6'
         ONNX_PATH: onnx==1.2.3
@@ -156,7 +156,27 @@ jobs:
         TENSORFLOW_PATH: tensorflow-cpu==2.5.0
         INSTALL_ORT: pip install onnxruntime==1.8.0
 
-    maxParallel: 3
+      ############ Pure Keras Unit Tests ############
+      Keras-Py36-tf1.15.0:
+        python.version: '3.6'
+        ONNX_PATH: onnx==1.5.0
+        KERAS: keras==2.2.5
+        TENSORFLOW_PATH: tensorflow==1.15.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
+
+      Keras-Py37-tf2.0.0:
+        python.version: '3.7'
+        ONNX_PATH: onnx==1.7.0
+        KERAS: keras==2.3.1
+        TENSORFLOW_PATH: tensorflow==2.0.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
+
+      Keras-Py37-tf2.2.0:
+        python.version: '3.7'
+        ONNX_PATH: onnx==1.9.0
+        KERAS: keras==2.4.3
+        TENSORFLOW_PATH: tensorflow==2.2.0
+        INSTALL_ORT: pip install onnxruntime==1.8.0
 
   steps:
   - task: UsePythonVersion@0
@@ -179,6 +199,7 @@ jobs:
       pip install protobuf
       pip install h5py==2.9.0
       pip install %TENSORFLOW_PATH%
+      IF NOT "%KERAS%"=="" (pip install %KERAS%)
       pip install git+https://github.com/microsoft/onnxconverter-common
       pip install pytest pytest-cov pytest-runner
       %INSTALL_ORT%
@@ -189,6 +210,7 @@ jobs:
       pip install -e .
       echo Test onnxruntime installation... && python -c "import onnxruntime"
       python -c "import onnxconverter_common"
+      IF NOT "%KERAS%"=="" (set TF_KERAS=0)
       pytest keras2onnx_tests --doctest-modules --junitxml=junit/test-results.xml
     displayName: 'pytest'
 

tests/test_backend.py

@@ -3791,6 +3791,18 @@ def func(input_x, boxes, box_ind, corp_size):
                             {_INPUT: input_x_val, _INPUT1: boxes_val, _INPUT2: box_ind_val, _INPUT3: corp_size_val},
                             rtol=1e-04, atol=1e-03)
 
+    @check_opset_min_version(11, "CropAndResize")
+    def test_crop_and_resize_empty_tensor(self):
+        def func(input_x, boxes, box_ind, corp_size):
+            return tf.image.crop_and_resize(input_x, boxes, box_ind, corp_size, name=_TFOUTPUT, extrapolation_value=1.0)
+        input_x_val = np.random.randint(low=0, high=256, size=[0, 36, 36, 3]).astype(np.float32)  # NHWC
+        boxes_val = np.array([]).astype(np.float32).reshape([0, 4])
+        box_ind_val = np.array([]).astype(np.int32)
+        corp_size_val = np.array([40, 40]).astype(np.int32)
+        self._run_test_case(func, [_OUTPUT],
+                            {_INPUT: input_x_val, _INPUT1: boxes_val, _INPUT2: box_ind_val, _INPUT3: corp_size_val},
+                            rtol=1e-04, atol=1e-03)
+
     def test_batch_to_space3d(self):
         block_size = [2, 2]
         crop = [[0, 1], [2, 1]]
@@ -4337,6 +4349,14 @@ def func(x):
             return tf.identity(x_, name=_TFOUTPUT)
         self._run_test_case(func, [_OUTPUT], {_INPUT: x_val})
 
+    def test_round_approx(self):
+        # In lower opsets there is no Round, but we can approximate it forgoing nearest even
+        x_val = np.array([-0.7, -0.5, -0.0, 0.0, +0.0, 0.3, 1.5, 0.7, float('nan')], dtype=np.float32)
+        def func(x):
+            x_ = tf.round(x)
+            return tf.identity(x_, name=_TFOUTPUT)
+        self._run_test_case(func, [_OUTPUT], {_INPUT: x_val})
+
     @check_opset_min_version(11, "Det")
     @unittest.skip("unclear how this is called in tf-2, fix later")
     def test_determinant(self):

tests/test_lstm.py

@@ -9,7 +9,7 @@
 from tensorflow.python.ops import init_ops
 from tensorflow.python.ops import variable_scope
 from backend_test_base import Tf2OnnxBackendTestBase
-from common import unittest_main, check_opset_after_tf_version, skip_tf2, skip_tf_versions
+from common import unittest_main, check_opset_after_tf_version, skip_tf2, skip_tf_versions, check_op_count
 
 from tf2onnx.tf_loader import is_tf2
 
@@ -36,12 +36,22 @@
 
 class LSTMTests(Tf2OnnxBackendTestBase):
 
-    def run_test_case(self, *args, **kwargs):  #pylint: disable=arguments-differ
+    def run_test_case(self, *args, require_lstm_count=1, **kwargs):  #pylint: disable=arguments-differ
         # TF LSTM has an unknown dim
         tmp = self.config.allow_missing_shapes
         self.config.allow_missing_shapes = True
+        def graph_validator(g):
+            good = True
+            if "graph_validator" in kwargs:
+                good = good and kwargs["graph_validator"](g)
+            if require_lstm_count is None or ":" not in g.outputs[0]:
+                # Skip checks for tflite graphs (no ":" in outputs)
+                return good
+            good = good and check_op_count(g, "LSTM", require_lstm_count, disabled=False)
+            good = good and check_op_count(g, "Loop", 0, disabled=False)
+            return good
         try:
-            super().run_test_case(*args, **kwargs)
+            super().run_test_case(*args, graph_validator=graph_validator, **kwargs)
         finally:
             self.config.allow_missing_shapes = tmp
 
@@ -385,7 +395,8 @@ def func(x):
         feed_dict = {"input_1:0": x_val}
         input_names_with_port = ["input_1:0"]
         output_names_with_port = ["output:0", "cell_state:0"]
-        self.run_test_case(func, feed_dict, input_names_with_port, output_names_with_port, rtol=1e-06)
+        self.run_test_case(func, feed_dict, input_names_with_port, output_names_with_port, rtol=1e-06,
+                           require_lstm_count=2)
 
     @check_opset_after_tf_version("1.15", 8, "might need Scan")
     @skip_tf2()     # Still failing likely due to inconsistent random number initialization

tf2onnx/graph.py

@@ -464,6 +464,8 @@ def __init__(self, nodes, output_shapes=None, dtypes=None, target=None, opset=No
         # A list of index, output tuples of potential scan outputs in this graph
         # Used by the tflite while loop handler
         self.scan_outputs = []
+        # Used by lstm_tf2_rewriter to indicate this subgraph is an LSTM cell
+        self.lstm_rewriter_context = None
         self.func_inputs = []
         self.ragged_variant_list_reads = []
         self.ragged_variant_list_writes = []

tf2onnx/onnx_opset/math.py

@@ -535,6 +535,15 @@ def version_11(cls, ctx, node, **kwargs):
 
 @tf_op("Round")
 class Round:
+    @classmethod
+    def version_1(cls, ctx, node, **kwargs):
+        # Not exactly nearest even but close enough
+        np_dtype = utils.map_onnx_to_numpy_type(ctx.get_dtype(node.input[0]))
+        const_half = ctx.make_const(utils.make_name("const_half"), np.array(0.5, np_dtype)).output[0]
+        add_node = ctx.make_node("Add", [node.input[0], const_half], op_name_scope=node.name).output[0]
+        node.type = "Floor"
+        ctx.replace_inputs(node, [add_node])
+
     @classmethod
     def version_11(cls, ctx, node, **kwargs):
         pass

tf2onnx/onnx_opset/nn.py

@@ -318,9 +318,10 @@ def build_dynamic_target_size(ctx, transposed_intput, target_hw):
     shape_of_transposed_input = ctx.make_node("Shape", [transposed_intput])
     first_half_of_shape = GraphBuilder(ctx).make_slice(
         {"data": shape_of_transposed_input.output[0], "ends": [2], "starts": [0]})
-    target_size_int64 = ctx.make_node("Cast", [target_hw], attr={'to': TensorProto.INT64})
+    if ctx.get_dtype(target_hw) != TensorProto.INT64:
+        target_hw = ctx.make_node("Cast", [target_hw], attr={'to': TensorProto.INT64}).output[0]
     # We build a tensor containing [n c nh nw]
-    final_target_size = ctx.make_node("Concat", [first_half_of_shape, target_size_int64.output[0]], {'axis': 0})
+    final_target_size = ctx.make_node("Concat", [first_half_of_shape, target_hw], {'axis': 0})
     return final_target_size
 
 
@@ -1183,9 +1184,13 @@ def any_version_after11(cls, opset, ctx, node, **kwargs):
             "method").s == b"nearest" else "linear"
         extrapolation_value = float(node.get_attr("extrapolation_value", "0").f)
         input_x = node.input[0]
+        x_shape = ctx.make_node("Shape", [input_x]).output[0]
+        num_channels = GraphBuilder(ctx).make_slice({"data": x_shape, "starts": [3], "ends": [4], "axes": [0]})
         boxes = node.input[1]
         box_ind = node.input[2]
         crop_size = node.input[3]
+        if ctx.get_dtype(crop_size) != TensorProto.INT64:
+            crop_size = ctx.make_node("Cast", [crop_size], attr={'to': TensorProto.INT64}).output[0]
         trip_name = utils.make_name(node.name + "_i")
         cond_name = utils.make_name(node.name + "_cond")
         cond_out_name = utils.make_name(node.name + "cond_out")
@@ -1233,6 +1238,10 @@ def any_version_after11(cls, opset, ctx, node, **kwargs):
         branches = {"body": g}
         inner_loop = ctx.make_node("Loop", [trip_node.output[0], cond_const.output[0]], name=node.name,
                                    outputs=node.output, branches=branches)
+        const_neg_one = ctx.make_const(utils.make_name("const_neg_one"), np.array([-1], np.int64)).output[0]
+        final_shape = ctx.make_node("Concat", [const_neg_one, crop_size, num_channels], attr={'axis': 0}).output[0]
+        # This reshape fixes the case when there are no iterations and the scan output is empty.
+        ctx.insert_new_node_on_output("Reshape", inner_loop.output[0], inputs=[inner_loop.output[0], final_shape])
 
     @classmethod
     def version_11(cls, ctx, node, **kwargs):

tf2onnx/rewriter/__init__.py

@@ -21,6 +21,7 @@
 from tf2onnx.rewriter.quantization_ops_rewriter import rewrite_quantize_and_dequantize
 from tf2onnx.rewriter.layer_normalization_rewriter import rewrite_layer_normalization
 from tf2onnx.rewriter.ragged_variant_shape_rewriter import rewrite_ragged_variant_shape
+from tf2onnx.rewriter.lstm_tf2_rewriter import rewriter_lstm_tf2
 
 
 __all__ = [
@@ -46,5 +47,6 @@
     "rewrite_quantize_and_dequantize",
     "rewrite_layer_normalization",
     "rewrite_conv_dilations",
-    "rewrite_ragged_variant_shape"
+    "rewrite_ragged_variant_shape",
+    "rewriter_lstm_tf2"
 ]