Add stacked LSTM support

- added new base class exclusive for LSTM which supports stacked LSTMs
- this base class can be combined with unit_rnn_rewriter_base once we add support for stacked layers in other RNN types like GRU
- modified LSTM rewriter such that it allows multiple cell matches and for all the variable finder and weight-bias methods we pass the LSTM index
- added new test for the same

Author: buddhapuneeth

tests/test_stacked_lstm.py

@@ -0,0 +1,56 @@
+"""Unit Tests for layered lstm"""
+
+import numpy as np
+import tensorflow as tf
+
+from tensorflow.python.ops import init_ops
+from backend_test_base import Tf2OnnxBackendTestBase
+from common import unittest_main, check_lstm_count, skip_tf2
+
+from tf2onnx.tf_loader import is_tf2
+
+if is_tf2():
+    LSTMCell = tf.compat.v1.nn.rnn_cell.LSTMCell
+    MultiRNNCell = tf.compat.v1.nn.rnn_cell.MultiRNNCell
+    dynamic_rnn = tf.compat.v1.nn.dynamic_rnn
+else:
+    LSTMCell = tf.contrib.rnn.LSTMCell
+    LSTMBlockCell = tf.contrib.rnn.LSTMBlockCell
+    MultiRNNCell = tf.contrib.rnn.MultiRNNCell
+    dynamic_rnn = tf.nn.dynamic_rnn
+
+class LSTMLayeredTests(Tf2OnnxBackendTestBase):
+    def test_layered_lstm(self):
+        units = 5
+        batch_size = 6
+        x_val = np.array([[1., 1.], [2., 2.], [3., 3.], [4., 4.]], dtype=np.float32)
+        x_val = np.stack([x_val] * batch_size)
+
+        def func(x):
+            initializer = init_ops.constant_initializer(0.5)
+            num_layers = 2
+
+            # no scope
+            def lstm_cell():
+                return LSTMCell(
+                    units,
+                    initializer=initializer,
+                    state_is_tuple=True)
+
+            stacked_lstm = MultiRNNCell(
+                [lstm_cell() for _ in range(num_layers)])
+            outputs, cell_state = dynamic_rnn(
+                stacked_lstm,
+                x,
+                dtype=tf.float32)
+            return tf.identity(outputs, name="output"), tf.identity(cell_state, name="cell_state")
+
+        input_names_with_port = ["input_1:0"]
+        feed_dict = {"input_1:0": x_val}
+
+        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,
+                           graph_validator=lambda g: check_lstm_count(g, 2))
+
+if __name__ == '__main__':
+    unittest_main()