Optimization for tflite loops (#1289)

TomWildenhain-Microsoft · web-flow · commit bb7df354da64 · 2021-01-22T18:59:30.000-05:00
Signed-off-by: Tom Wildenhain &lt;tomwi@microsoft.com&gt;
diff --git a/tf2onnx/graph.py b/tf2onnx/graph.py
@@ -465,6 +465,9 @@ def __init__(self, nodes, output_shapes=None, dtypes=None, target=None, opset=No
         self.graph_name = graph_name or utils.make_name("tf2onnx")
         self._is_subgraph = is_subgraph
         self.ta_reads = []
+        # A list of index, output tuples of potential scan outputs in this graph
+        # Used by the tflite while loop handler
+        self.scan_outputs = []
         self.func_inputs = []
 
         self._target = set(target)
diff --git a/tf2onnx/tflite_handlers/tfl_controlflow.py b/tf2onnx/tflite_handlers/tfl_controlflow.py
@@ -12,6 +12,7 @@
 from tf2onnx.handler import tfl_op
 from tf2onnx import utils
 from tf2onnx.tf_loader import find_function
+from tf2onnx.graph_builder import GraphBuilder
 from tf2onnx.onnx_opset.controlflow import parameter_binding, inline_subgraph
 
 
@@ -40,6 +41,19 @@ def version_7(cls, ctx, node, **kwargs):
         cond_binding = parameter_binding(cond_graph, tfl_while_inputs)
         cond_outputs = inline_subgraph(ctx, cond_graph, cond_name, cond_binding)
 
+        # Potential scan output candidates are identified in the body subgraph using tfl_scan_output_rewriter.
+        # They can then be optimized in this tfl loop handler provided they are not used in the cond subgraph.
+        scan_outputs = sorted(body.scan_outputs, reverse=True)
+        def input_is_unused(g, index):
+            return len(g.find_output_consumers(g.func_inputs[index])) == 0
+        scan_outputs = [(i, out) for i, out in scan_outputs if input_is_unused(cond_graph, i)]
+
+        for idx, _ in scan_outputs:
+            del tfl_while_inputs[idx]
+            output_shapes.append(output_shapes.pop(idx))
+            output_dtypes.append(output_dtypes.pop(idx))
+            output_names.append(output_names.pop(idx))
+
         max_iterations = ctx.make_const(utils.make_name("max_iterations"), np.array(np.iinfo(np.int64).max))
 
         loop_node = ctx.make_node("Loop", [max_iterations.output[0], cond_outputs[0]] + tfl_while_inputs,
@@ -52,15 +66,21 @@ def version_7(cls, ctx, node, **kwargs):
         for k, v in output_map.items():
             ctx.replace_all_inputs(k, v)  # ops=ctx.get_nodes()
 
-        body = wire_tfl_while_body(body, loop_node.inputs, output_shapes, output_dtypes, cond_graph)
+        body = wire_tfl_while_body(body, loop_node.inputs, output_shapes, output_dtypes, cond_graph, scan_outputs)
+
+        for i in range(len(scan_outputs)):
+            squeeze_node = GraphBuilder(body).make_squeeze(
+                {'data': body.outputs[-1-i], "axes": [0]}, return_node=True)
+            body.outputs[-1-i] = squeeze_node.output[0]
 
         loop_node.set_body_graph_as_attr("body", body)
 
 def wire_tfl_while_body(g, loop_node_inputs, output_shapes,
-                        output_dtypes, cond_graph):
+                        output_dtypes, cond_graph, scan_outputs):
     """Wire subgraph graph into main."""
 
     g = copy.deepcopy(g)
+    graph_inputs = g.func_inputs.copy()
 
     # onnx will pass in cond as argument
     iter_node = g.make_node("Placeholder", [], name=utils.make_name("iteration_num"),
@@ -69,6 +89,28 @@ def wire_tfl_while_body(g, loop_node_inputs, output_shapes,
                             output_count=1, dtypes=[TensorProto.BOOL], shapes=[[]])
     cond_binding = parameter_binding(cond_graph, g.outputs)
 
+    to_remove = set()
+    for idx, scan_output in scan_outputs:
+        inp = g.get_node_by_output(graph_inputs[idx])
+
+        # Remove consumers of scan input
+        stack = [inp]
+        while stack:
+            node = stack.pop()
+            if node not in to_remove:
+                to_remove.add(node)
+                for out in node.output:
+                    stack += g.find_output_consumers(out)
+
+        # Remove scan input from cond graph
+        cond_binding = {k: "@@ALLOC" if v == g.outputs[idx] else v for k, v in cond_binding.items()}
+        del g.func_inputs[idx]
+        del g.outputs[idx]
+        g.outputs.append(scan_output)
+
+    for node in to_remove:
+        g.remove_node(node.name)
+
     # in onnx the body inputs are: index, cond, [loop_vars]
     g.func_inputs = [iter_node.output[0], cond_node.output[0]] + g.func_inputs
     # tell graph lib to keep inputs in order
diff --git a/tf2onnx/tflite_rewriters/__init__.py b/tf2onnx/tflite_rewriters/__init__.py
@@ -0,0 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
+
+"""tf2onnx.tflite_rewriters module"""
+
+from . import (
+    tfl_scan_output_rewriter
+)
diff --git a/tf2onnx/tflite_rewriters/tfl_scan_output_rewriter.py b/tf2onnx/tflite_rewriters/tfl_scan_output_rewriter.py
@@ -0,0 +1,156 @@
+# SPDX-License-Identifier: Apache-2.0
+
+
+"""
+tf2onnx.tflite_rewriters.tfl_scan_output_rewriter - Identify a common slice/concat pattern in tflite subgraphs
+Effectively replace A = A[:i] + [B] + A[i+1:] with A[i] = B
+"""
+import numpy as np
+
+from tf2onnx.graph_matcher import OpTypePattern, GraphMatcher
+
+
+# pylint: disable=missing-docstring
+
+def rewrite_slice_concat_to_scatter(g, ops):
+    pattern0 = \
+        OpTypePattern('TFL_CONCATENATION', name='concat', inputs=[
+            OpTypePattern('TFL_SLICE', name='begin_slice'),
+            OpTypePattern('*', name='middle'),
+            OpTypePattern('TFL_SLICE', name='end_slice')
+        ])
+
+    matcher = GraphMatcher(pattern0, allow_reorder=False)
+    match_results = list(matcher.match_ops(ops))
+    if match_results:
+        for match in match_results:
+            concat = match.get_op("concat")
+            begin_slice = match.get_op("begin_slice")
+            middle = match.get_op("middle")
+            end_slice = match.get_op("end_slice")
+            middle_shape = g.get_shape(middle.output[0])
+
+            # Both slices must be slicing the same tensor
+            if begin_slice.input[0] != end_slice.input[0]:
+                continue
+            original_tensor = begin_slice.input[0]
+            if concat.get_attr_int("axis") != 0:
+                continue
+            # The inserted slice must have length 1 (to be a single index)
+            if middle_shape is None or len(middle_shape) == 0 or middle_shape[0] != 1:
+                continue
+            rank = len(middle_shape)
+            scan_output = middle.output[0]
+            if not begin_slice.inputs[1].is_const() or not end_slice.inputs[2].is_const():
+                continue
+            # The first slice must start from the beginning (0) for all dims
+            if not all(v == 0 for v in begin_slice.inputs[1].get_tensor_value()):
+                continue
+            # The second slice must slice to the end (-1) for all dims
+            if not all(v == -1 for v in end_slice.inputs[2].get_tensor_value()):
+                continue
+            # The other slice dims are assembled by concatenation if rank > 1
+            if rank > 1:
+                begin_concat = begin_slice.inputs[2]
+                end_concat = end_slice.inputs[1]
+                if not begin_concat.type == "TFL_CONCATENATION":
+                    continue
+                if not end_concat.type == "TFL_CONCATENATION":
+                    continue
+                # Except for dim 0, slice from beginning to end
+                if not all(get_uniform_const_val(inp) == -1 for inp in begin_concat.inputs[1:]):
+                    continue
+                if not all(get_uniform_const_val(inp) == 0 for inp in end_concat.inputs[1:]):
+                    continue
+                begin_idx = begin_concat.inputs[0]
+                end_idx = end_concat.inputs[0]
+            else:
+                begin_idx = begin_slice.inputs[2]
+                end_idx = end_slice.inputs[1]
+            # For dim 0, slice to i for first part and from i+1 for second
+            if not node_is_one_plus_node(begin_idx, end_idx):
+                continue
+            out1, _ = get_out_and_offset(begin_idx)
+            graph_inps = [n.output[0] for n in g.inputs]
+            # To be a scan output, i must be a graph input
+            if out1 not in graph_inps:
+                continue
+            # The array being sliced must be a graph input
+            if original_tensor not in graph_inps:
+                continue
+            # The input/output index of i
+            idx = graph_inps.index(out1)
+            # The input/output index of the array
+            scan_output_idx = graph_inps.index(original_tensor)
+            # For a scan output, i must be assigned to i+1 with each iteration
+            if not node_is_one_plus_node(g.get_node_by_output(out1), g.get_node_by_output(g.outputs[idx])):
+                continue
+            if len(g.find_output_consumers(concat.output[0])) > 1:
+                continue
+
+            if g.opset < 10 and len(g.find_output_consumers(concat.output[0])) <= 1:
+                # If opset is < 10, conversion of the subgraph will fail unless we remove the slice nodes
+                # We add a tmp node to replace them.
+                shape = g.get_shape(concat.output[0])
+                dtype = g.get_dtype(concat.output[0])
+                tmp_node = g.make_node("TMP_SCAN_OUTPUT", [original_tensor, scan_output],
+                                       shapes=[shape], dtypes=[dtype])
+                g.replace_all_inputs(concat.output[0], tmp_node.output[0])
+
+            to_remove = []
+            out = g.outputs[scan_output_idx]
+            node = g.get_node_by_output(out)
+            to_remove.append(node)
+
+            while len(node.input) > 0 and node != concat:
+                out = node.input[0]
+                node = g.get_node_by_output(out)
+                to_remove.append(node)
+
+            to_remove += [begin_slice, end_slice, concat]
+
+            out = original_tensor
+            node = g.get_node_by_output(out)
+            to_remove.append(node)
+
+            while len(node.input) > 0:
+                out = node.input[0]
+                node = g.get_node_by_output(out)
+                to_remove.append(node)
+
+            if not g.is_safe_to_remove_nodes(to_remove):
+                continue
+
+            g.scan_outputs.append((scan_output_idx, scan_output))
+    return ops
+
+def get_uniform_const_val(n):
+    if not n.is_const():
+        return None
+    v = n.get_tensor_value(as_list=False).flatten()
+    if len(v) == 0:
+        return None
+    if np.all(v == v[0]):
+        return v[0]
+    return None
+
+def get_out_and_offset(n):
+    if n.type in ['TFL_RESHAPE', 'TFL_IDENTITY', 'Identity']:
+        return get_out_and_offset(n.inputs[0])
+    if n.type == 'TFL_ADD':
+        v1 = get_uniform_const_val(n.inputs[0])
+        v2 = get_uniform_const_val(n.inputs[1])
+        if v1 is not None and v2 is not None:
+            return '', v1 + v2
+        if v1 is not None:
+            inp2, o2 = get_out_and_offset(n.inputs[1])
+            return inp2, v1 + o2
+        if v2 is not None:
+            inp1, o1 = get_out_and_offset(n.inputs[0])
+            return inp1, v2 + o1
+    return n.output[0], 0
+
+def node_is_one_plus_node(node, one_plus_node):
+    n1, o1 = get_out_and_offset(node)
+    n2, o2 = get_out_and_offset(one_plus_node)
+    return n1 == n2 and o1 + 1 == o2
diff --git a/tf2onnx/tflite_utils.py b/tf2onnx/tflite_utils.py
@@ -180,7 +180,7 @@ def parse_tflite_graph(tflite_g, opcodes_map, model, input_prefix=''):
             output_shapes[name] = tensor.ShapeSignatureAsNumpy().tolist()
         buf = model.Buffers(tensor.Buffer())
         dtypes[name] = map_tflite_dtype_to_onnx(tensor.Type())
-        if not buf.DataIsNone():
+        if not buf.DataIsNone() and tensor.Buffer() > 0:
             # For const values we use TF to decode the binary data from the buffer
             t = tensor_pb2.TensorProto()
             t.tensor_content = buf.DataAsNumpy().tobytes()
