fix einsum

onnx2kerastl/customonnxlayer/onnxeinsum.py

@@ -0,0 +1,40 @@
+from keras.layers import Layer, TFOpLambda
+import tensorflow as tf
+import numpy as np
+
+
+# this custom layer needed because of a tensorflow bug on einsum serielization
+class OnnxEinsumLayer(Layer):
+    """
+    Layer wrapping a single tf.einsum operation.
+
+    Usage:
+    x = EinsumLayer("bmhwf,bmoh->bmowf")((x1, x2))
+    """
+
+    def __init__(self, equation: str, constant_input, constant_place):
+        super().__init__()
+        self.equation = equation
+        if constant_input is not None:
+            if hasattr(constant_input, 'numpy'):
+                constant_input = constant_input.numpy()
+            if not isinstance(constant_input, np.ndarray):
+                constant_input = np.array(constant_input)
+            self.constant_input = constant_input
+        self.constant_place = constant_place
+
+    def call(self, inputs, *args, **kwargs):
+        if self.constant_input is not None:
+            if self.constant_place == 1:
+                inputs = [inputs, self.constant_input]
+            else:
+                inputs = [self.constant_input, inputs]
+
+        return tf.einsum(self.equation, *inputs)
+
+    def get_config(self):
+        return {
+            "equation": self.equation,
+            "constant_input": self.constant_input,
+            "constant_place": self.constant_place
+        }

onnx2kerastl/operation_layers.py

@@ -2,9 +2,10 @@
 
 import keras
 import numpy as np
-import tensorflow as tf
 from keras import backend as K
+import tensorflow as tf
 
+from .customonnxlayer.onnxeinsum import OnnxEinsumLayer
 from .exceptions import UnsupportedLayer
 from .utils import is_numpy, ensure_tf_type, ensure_float
 from .tfops_funcs import tf_math_abs, tf_clip_by_value, tf_math_negative, K_mean, tf_math_reduce_prod,\
@@ -889,10 +890,19 @@ def get_empty_array(x, dtype=new_dtype, keras_name=keras_name):
 
 
 def convert_einsum(node, params, layers, lambda_func, node_name, keras_name):
-    # input_0 = layers[node.input[0]]
-    # input_1 = layers[node.input[1]]
-    # equation = params['equation'].decode('utf-8')
-    # layers[node_name] = tf.einsum(equation, *[input_0, input_1], name=keras_name)
-    input_0 = tf_expand_dims(layers[node.input[0]], axis=2, tf_name=f"{params['cleaned_name']}_einsum_0")
-    input_1 = tf_expand_dims(layers[node.input[1]], axis=0, tf_name=f"{params['cleaned_name']}_einsum_1")
-    layers[node_name] = tf_multiply(input_0, input_1, tf_name=f"{params['cleaned_name']}_einsum_mult")
+    input_0 = layers[node.input[0]]
+    input_1 = layers[node.input[1]]
+    equation = params['equation'].decode('utf-8')
+
+    is_input_0_constant = isinstance(input_0, tf.Tensor)
+    is_input_1_constant = isinstance(input_1, tf.Tensor)
+    if is_input_0_constant and is_input_1_constant:
+        layers[node_name] = tf.einsum(equation, *[input_0, input_1], name=keras_name)
+    elif is_input_0_constant:
+        layers[node_name] = OnnxEinsumLayer(equation, input_0, 0)(input_1, name=keras_name)
+    elif is_input_1_constant:
+        layers[node_name] = OnnxEinsumLayer(equation, input_1, 1)(input_0, name=keras_name)
+    else:
+        layers[node_name] = OnnxEinsumLayer(equation, None, None)([input_0, input_1], name=keras_name)
+
+

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "onnx2kerastl"
-version = "0.0.142"
+version = "0.0.145"
 description = ""
 authors = ["dorhar <[email protected]>"]
 license = "MIT"