Tom/tf 2.5 (#1462)

* add tf-2.5 to ci

Signed-off-by: Guenther Schmuelling <[email protected]>

* theta must be >- 0

Signed-off-by: Guenther Schmuelling <[email protected]>

* Update tflite flatbuffer

Signed-off-by: Tom Wildenhain <[email protected]>

* Fixes for tf2.5

Signed-off-by: Tom Wildenhain <[email protected]>

Co-authored-by: Guenther Schmuelling <[email protected]>

Author: TomWildenhain-Microsoft

ci_build/azure_pipelines/onnxruntime_nightly_test.yml

@@ -39,6 +39,18 @@ stages:
           - template: 'unit_test.yml'
         report_coverage: 'True'
 
+    - template: 'templates/job_generator.yml'
+      parameters:
+        platforms: ['linux', 'windows']
+        python_versions: [3.8']
+        tf_versions: ['2.5.0rc1']
+        onnx_opsets: ['']
+        onnx_backends: {onnxruntime: ['nightly']}
+        job:
+          steps:
+          - template: 'unit_test.yml'
+        report_coverage: 'True'
+  
   - template: 'templates/combine_test_coverage.yml'
 
 schedules:

ci_build/azure_pipelines/pretrained_model_test-matrix.yml

@@ -27,4 +27,12 @@ jobs:
     job:
       steps:
       - template: 'pretrained_model_test.yml'
-    
+
+- template: 'templates/job_generator.yml'
+  parameters:
+    platforms: ['linux', 'windows']
+    python_versions: ['3.8']
+    tf_versions: ['2.5.0rc1']
+    job:
+      steps:
+      - template: 'pretrained_model_test.yml'

ci_build/azure_pipelines/unit_test-matrix.yml

@@ -7,7 +7,7 @@ stages:
       parameters:
         platforms: ['linux', 'windows']
         python_versions: ['3.6']
-        tf_versions: ['1.13.1', '1.12.3']
+        tf_versions: ['1.12.3']
         onnx_opsets: ['']
         job:
           steps:
@@ -35,5 +35,16 @@ stages:
           steps:
           - template: 'unit_test.yml'
         report_coverage: 'True'
+
+    - template: 'templates/job_generator.yml'
+      parameters:
+        platforms: ['linux', 'windows']
+        python_versions: ['3.8']
+        tf_versions: ['2.5.0rc1']
+        onnx_opsets: ['']
+        job:
+          steps:
+          - template: 'unit_test.yml'
+        report_coverage: 'True'
 
   - template: 'templates/combine_test_coverage.yml'

ci_build/azure_pipelines/unit_test.yml

@@ -6,7 +6,7 @@ stages:
     - template: 'templates/job_generator.yml'
       parameters:
         python_versions: ['3.8']
-        tf_versions: ['2.4.0']
+        tf_versions: ['2.5.0rc1']
         onnx_opsets: ['']
         skip_tflite_tests: 'False'
         skip_tf_tests: 'True'
@@ -30,7 +30,7 @@ stages:
     - template: 'templates/job_generator.yml'
       parameters:
         python_versions: ['3.8']
-        tf_versions: ['2.4.0']
+        tf_versions: ['2.5.0rc1']
         onnx_opsets: ['']
         job:
           steps:
@@ -40,7 +40,7 @@ stages:
     - template: 'templates/job_generator.yml'
       parameters:
         python_versions: ['3.7']
-        tf_versions: ['1.14.0','1.15.2','2.2.0','2.3.0']
+        tf_versions: ['1.14.0','1.15.2','2.3.0','2.4.1']
         onnx_opsets: ['']
         job:
           steps:
@@ -71,7 +71,7 @@ stages:
       parameters:
         python_versions: ['3.7']
         platforms: ['windows']
-        tf_versions: ['2.3.0']
+        tf_versions: ['2.4.1']
         onnx_opsets: ['']
         job:
           steps:

tests/backend_test_base.py

@@ -30,6 +30,7 @@
 from tf2onnx.tf_loader import tf_optimize, is_tf2, get_hash_table_info
 from tf2onnx.tf_utils import compress_graph_def
 from tf2onnx.graph import ExternalTensorStorage
+from tf2onnx.tflite.Model import Model
 
 
 if is_tf2():
@@ -217,6 +218,22 @@ def convert_to_tflite(self, graph_def, feed_dict, outputs):
             except ConverterError:
                 return None
 
+    def tflite_has_supported_types(self, tflite_path):
+        try:
+            with open(tflite_path, 'rb') as f:
+                buf = f.read()
+            buf = bytearray(buf)
+            model = Model.GetRootAsModel(buf, 0)
+            tensor_cnt = model.Subgraphs(0).TensorsLength()
+            interpreter = tf.lite.Interpreter(tflite_path)
+            for i in range(tensor_cnt):
+                dtype = interpreter._get_tensor_details(i)['dtype']   # pylint: disable=protected-access
+                if np.dtype(dtype).kind == 'O':
+                    return False
+            return True
+        except (RuntimeError, ValueError):
+            return False
+
     def run_tflite(self, tflite_path, feed_dict):
         try:
             interpreter = tf.lite.Interpreter(tflite_path)
@@ -293,7 +310,7 @@ def run_test_case(self, func, feed_dict, input_names_with_port, output_names_wit
 
         if test_tflite:
             tflite_path = self.convert_to_tflite(graph_def, feed_dict, output_names_with_port)
-            test_tflite = tflite_path is not None
+            test_tflite = tflite_path is not None and self.tflite_has_supported_types(tflite_path)
 
         if test_tf:
             tf_reset_default_graph()

tests/test_backend.py

@@ -3328,6 +3328,16 @@ def func(x, y):
 
         self._run_test_case(func, [_OUTPUT], {_INPUT: input_x, _INPUT1: input_y})
         self._run_test_case(func, [_OUTPUT], {_INPUT: input_x.astype(np.int32), _INPUT1: input_y})
+
+    @check_opset_min_version(8, "BroadcastTo")
+    def test_zeros_like_bool(self):
+        input_x = np.random.random_sample([10, 20]).astype(np.float32)
+        input_y = np.array([20, 10]).astype(np.int64)
+
+        def func(x, y):
+            z = tf.reshape(x, y)
+            return tf.zeros_like(z, name=_TFOUTPUT)
+
         self._run_test_case(func, [_OUTPUT], {_INPUT: input_x > 0.5, _INPUT1: input_y})
 
     @check_opset_min_version(9, "is_nan")
@@ -3746,7 +3756,7 @@ def test_conv1d_5(self):
     def test_thresholded_relu(self):
         # tf.keras.layers.ThresholdedReLU only supports `float32` for x
         x_val = np.array([0.0, 1.0, -1.0, 2.0, -2.0, 0.5, -0.5, 1.5, -1.5], dtype=np.float32).reshape((3, 3))
-        theta_vals = [-1.0, -0.5, 0.0, 0.5, 1.0]
+        theta_vals = [0.0, 0.5, 1.0, 2.0]
         for theta_val in theta_vals:
             def func(x):
                 t = tf.keras.layers.ThresholdedReLU(theta=theta_val)

tf2onnx/onnx_opset/controlflow.py

@@ -183,7 +183,15 @@ def version_9(cls, ctx, node, **kwargs):
         # T1 output = Where(bool condition, T1 x, T1 y)
         # NOTE: condition can be 1-dimension in tensorflow, while in onnx,
         # it should be broadcastable with other two inputs
-        if ctx.get_dtype(node.output[0]) != TensorProto.STRING:
+
+        # We can't use the mul/add trick if a NaN is involved. handles_nan is added earlier in the converter.
+        handles_nan = node.get_attr_value("handles_nan", False)
+        if ctx.get_dtype(node.output[0]) in [TensorProto.FLOAT, TensorProto.DOUBLE]:
+            for inp in node.inputs[1:]:
+                if inp.is_const() and np.any(np.isnan(inp.get_tensor_value(as_list=False))):
+                    handles_nan = True
+
+        if ctx.get_dtype(node.output[0]) != TensorProto.STRING and not handles_nan:
             # Due to bad ORT implementation, Mul/Add ops are faster than Where op
             cls.version_7(ctx, node, **kwargs)
             return

tf2onnx/onnx_opset/logical.py

@@ -74,6 +74,11 @@ def version_1(cls, ctx, node, **kwargs):
     def version_7(cls, ctx, node, **kwargs):
         # T2 output = Equal(T1, x, T1 y), T1 \in {bool, int32, int64}
         need_not = node.type == "NotEqual"
+        if need_not and node.input[0] == node.input[1]:
+            # The only value not equal to itself is NaN
+            node.type = "IsNaN"
+            ctx.replace_inputs(node, [node.input[0]])
+            return
         supported_dtypes = [
             TensorProto.BOOL,
             TensorProto.INT32,

tf2onnx/onnx_opset/tensor.py

@@ -2065,11 +2065,13 @@ def version_10(cls, ctx, node, **kwargs):
         batch_dim = node.get_attr_value("batch_dim", 0)
 
         ctx.remove_node(node.name)
+        shape = ctx.get_shape(node.input[0])
+        dtype = ctx.get_dtype(node.input[0])
         node = ctx.make_node(
             "ReverseSequence",
             node.input,
             outputs=node.output,
-            attr={"batch_axis": batch_dim, "time_axis": seq_dim})
+            attr={"batch_axis": batch_dim, "time_axis": seq_dim}, shapes=[shape], dtypes=[dtype])
 
         seq_len_dtype = ctx.get_dtype(node.input[1])
         utils.make_sure(seq_len_dtype is not None, "dtype of {} is None".format(node.input[1]))

tf2onnx/tflite/BuiltinOperator.py

@@ -137,4 +137,13 @@ class BuiltinOperator(object):
     CALL_ONCE = 129
     BROADCAST_TO = 130
     RFFT2D = 131
+    CONV_3D = 132
+    IMAG = 133
+    REAL = 134
+    COMPLEX_ABS = 135
+    HASHTABLE = 136
+    HASHTABLE_FIND = 137
+    HASHTABLE_IMPORT = 138
+    HASHTABLE_SIZE = 139
+    REDUCE_ALL = 140