add pt2tf tool

pt2tf/onnx-tensorflow/onnx_tf/handlers/backend/arg_min.py

@@ -0,0 +1,45 @@
+import tensorflow as tf
+
+from onnx_tf.handlers.backend_handler import BackendHandler
+from onnx_tf.handlers.handler import onnx_op
+from onnx_tf.handlers.handler import tf_func
+from onnx_tf.common.tf_helper import tf_shape
+
+
+@onnx_op("ArgMin")
+@tf_func(tf.argmin)
+class ArgMin(BackendHandler):
+
+  @classmethod
+  def get_attrs_processor_param(cls):
+    return {"default": {"axis": 0}}
+
+  @classmethod
+  def _common(cls, node, **kwargs):
+    axis = node.attrs.get("axis", 0)
+    keepdims = node.attrs.get("keepdims", 1)
+    select_last_index = node.attrs.get("select_last_index", 0)
+    if select_last_index == 0:
+      arg_min = cls.make_tensor_from_onnx_node(node, **kwargs)
+    else:
+      # reverse the input and apply argmax on that to get last occurrence of max
+      x = kwargs["tensor_dict"][node.inputs[0]]
+      x = tf.reverse(x, axis=[axis])
+      arg_min = cls.make_tensor_from_onnx_node(node, inputs=[x], **kwargs)
+      # adjust indices to account for the reverse
+      arg_min = tf_shape(x)[axis] - arg_min - 1
+    if keepdims == 1:
+      return [tf.expand_dims(arg_min, axis=axis)]
+    return [arg_min]
+
+  @classmethod
+  def version_1(cls, node, **kwargs):
+    return cls._common(node, **kwargs)
+
+  @classmethod
+  def version_11(cls, node, **kwargs):
+    return cls._common(node, **kwargs)
+
+  @classmethod
+  def version_12(cls, node, **kwargs):
+    return cls._common(node, **kwargs)