compile a tensorflow StandardLSTM graph for streaming

tools/compile_tf_standardlstm_graph_for_streaming.py

@@ -0,0 +1,117 @@
+#!/usr/bin/env python3
+
+"""
+We need to create a simple LSTM model to run in web browser with javascript and tensorflow js.
+Therefore, we cannot use NativeLSTM2 here and have to use StandardLSTM. At the same time, the
+inference will run in a frame-by-frame base. The goal of this script to create a compatible graph.
+network = {
+    "source": {
+        "class": "eval",
+        "eval": "self.network.get_config().typed_value('transform')(source(0, as_data=True),
+        network=self.network)"
+    },
+    "dropout1": {
+        "class": "dropout",
+        "dropout": 0.2,
+        "from": "source"
+    },
+    "lstm_fw": {
+        "class": "rec",
+        "unit": "StandardLSTM",
+        "n_out": 64,
+        "dropout": 0.2,
+        "direction": 1,
+        "from": "dropout1",
+        "unit_opts": {"forget_bias": 0.0}
+    },
+    "dropout2": {
+        "class": "dropout",
+        "dropout": 0.2,
+        "from": "lstm_fw"
+    },
+    "output": {
+        "class": "softmax",
+        "n_out": 2,
+        "target": "classes",
+        "loss": "ce",
+        "dropout": 0.2,
+        "from": "dropout2"
+    }
+}
+"""
+
+import os
+import sys
+import argparse
+import tensorflow as tf
+from tensorflow.python.framework import graph_io
+
+try:
+    from tensorflow.python.ops.nn import rnn_cell
+except ImportError:
+    from tensorflow.python.ops import rnn_cell
+from tensorflow.python.ops import rnn
+
+tf.compat.v1.disable_eager_execution()
+
+
+def main(argv):
+    """
+    Main entry.
+    """
+    # noinspection SpellCheckingInspection
+    argparser = argparse.ArgumentParser(description="Compile some StandardLSTM ops")
+    argparser.add_argument("--output_file", help="allowed extensions: meta")
+    argparser.add_argument("--hidden_size", default=64, type=int, help="LSTM hidden size")
+    argparser.add_argument("--input_size", default=75, type=int, help="input feature size")
+    argparser.add_argument("--output_classes", default=2, type=int, help="number of output classes")
+    args = argparser.parse_args(argv[1:])
+
+    tf.compat.v1.reset_default_graph()
+
+    inputs = tf.compat.v1.placeholder(tf.float32, [1, None, args.input_size], name="inputs")
+    initial_c = tf.compat.v1.placeholder(tf.float32, [1, args.hidden_size], name="initial_c")
+    initial_h = tf.compat.v1.placeholder(tf.float32, [1, args.hidden_size], name="initial_h")
+    lstm_cell = rnn_cell.LSTMCell(num_units=args.hidden_size, forget_bias=0.0, name="lstm_cell")
+    initial_state = tf.compat.v1.nn.rnn_cell.LSTMStateTuple(initial_c, initial_h)
+
+    with tf.compat.v1.variable_scope("lstm_fw"):
+        lstm_outputs, lstm_final_state = rnn.dynamic_rnn(
+            cell=lstm_cell, inputs=inputs, initial_state=initial_state, dtype=tf.float32, scope="rec"
+        )
+
+    with tf.compat.v1.variable_scope("output"):
+        # noinspection PyPep8Naming
+        W = tf.compat.v1.get_variable(
+            name="W",
+            shape=[args.hidden_size, args.output_classes],
+            initializer=tf.compat.v1.glorot_uniform_initializer(),
+        )
+        b = tf.compat.v1.get_variable(
+            name="b", shape=[args.output_classes], initializer=tf.compat.v1.glorot_uniform_initializer()
+        )
+        matmul_output = tf.einsum("bth,hk->btk", lstm_outputs, W)
+        dense_logits = tf.add(matmul_output, b)
+        tf.nn.softmax(dense_logits, axis=-1, name="softmax_output")
+
+    tf.identity(lstm_final_state.c, name="final_state_c")
+    tf.identity(lstm_final_state.h, name="final_state_h")
+
+    saver = tf.compat.v1.train.Saver()
+    graph_def = saver.export_meta_graph()
+
+    if args.output_file:
+        filename = args.output_file
+        _, ext = os.path.splitext(filename)
+        assert ext in [".meta"], "filename %r extension invalid" % filename
+        print("Write graph to file:", filename)
+        graph_io.write_graph(
+            graph_def,
+            logdir=os.path.dirname(filename),
+            name=os.path.basename(filename),
+            as_text=ext.endswith("txt"),
+        )
+
+
+if __name__ == "__main__":
+    main(sys.argv)