Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame^] | 1 | # Copyright (C) 2020 Arm Limited or its affiliates. All rights reserved. |
| 2 | # |
| 3 | # SPDX-License-Identifier: Apache-2.0 |
| 4 | # |
| 5 | # Licensed under the Apache License, Version 2.0 (the License); you may |
| 6 | # not use this file except in compliance with the License. |
| 7 | # You may obtain a copy of the License at |
| 8 | # |
| 9 | # www.apache.org/licenses/LICENSE-2.0 |
| 10 | # |
| 11 | # Unless required by applicable law or agreed to in writing, software |
| 12 | # distributed under the License is distributed on an AS IS BASIS, WITHOUT |
| 13 | # WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | # See the License for the specific language governing permissions and |
| 15 | # limitations under the License. |
| 16 | |
| 17 | |
| 18 | # Description: |
| 19 | # Functions used to read from a TensorFlow Lite format file. |
| 20 | |
| 21 | from .tflite.Model import Model |
| 22 | from .tflite.BuiltinOperator import BuiltinOperator |
| 23 | |
| 24 | import numpy as np |
| 25 | import os.path |
| 26 | from .nn_graph import Graph, Operation, Subgraph |
| 27 | from .tensor import Tensor, QuantizationParameters |
| 28 | |
| 29 | from .tflite_mapping import builtin_operator_map, datatype_map, datatype_map_numpy, DataType |
| 30 | |
| 31 | |
| 32 | def decode_str(s): |
| 33 | if s is None: |
| 34 | return "" |
| 35 | return s.decode("utf-8") |
| 36 | |
| 37 | |
| 38 | def reshape_tensor_add_const_op(tens, reorder): |
| 39 | if not tens.reshaped: |
| 40 | original_shape = tens.shape |
| 41 | tens.name = tens.name + "_reshape" |
| 42 | tens.shape = [original_shape[idx] for idx in reorder] |
| 43 | tens.bandwidth_shape = tens.shape |
| 44 | tens.storage_shape = tens.shape |
| 45 | |
| 46 | if tens.values is not None: |
| 47 | tens.values = tens.values.transpose(reorder) |
| 48 | |
| 49 | if tens.quant_values is not None: |
| 50 | tens.quant_values = tens.quant_values.transpose(reorder) |
| 51 | |
| 52 | op = Operation("Const", tens.name) |
| 53 | op.outputs = [tens] |
| 54 | tens.ops = [op] |
| 55 | tens.reshaped = True |
| 56 | |
| 57 | |
| 58 | class TFLiteSubgraph: |
| 59 | def __init__(self, graph, subgraph): |
| 60 | self.graph = graph |
| 61 | self.name = decode_str(subgraph.Name()) |
| 62 | |
| 63 | self.tensors = [] |
| 64 | for idx in range(subgraph.TensorsLength()): |
| 65 | self.tensors.append(self.parse_tensor(subgraph.Tensors(idx))) |
| 66 | |
| 67 | for idx in range(subgraph.OperatorsLength()): |
| 68 | self.parse_operator(subgraph.Operators(idx)) |
| 69 | |
| 70 | self.outputs = [self.tensors[idx] for idx in subgraph.OutputsAsNumpy()] |
| 71 | self.inputs = [self.tensors[idx] for idx in subgraph.InputsAsNumpy()] |
| 72 | |
| 73 | # Fix up tensors without operations. Generate either Placeholder or Constant ops |
| 74 | for tens in self.inputs: |
| 75 | assert not tens.ops |
| 76 | op = Operation("Placeholder", tens.name) |
| 77 | op.outputs = [tens] |
| 78 | tens.ops = [op] |
| 79 | |
| 80 | for tens in self.tensors: |
| 81 | if not tens.ops: |
| 82 | op = Operation("Const", tens.name) |
| 83 | op.outputs = [tens] |
| 84 | tens.ops = [op] |
| 85 | |
| 86 | def parse_tensor(self, tens_data): |
| 87 | np_shape = tens_data.ShapeAsNumpy() |
| 88 | shape = list(np_shape) if type(np_shape) is np.ndarray else [] |
| 89 | name = decode_str(tens_data.Name()) |
| 90 | dtype = datatype_map[tens_data.Type()] |
| 91 | |
| 92 | tens = Tensor(shape, dtype, name) |
| 93 | |
| 94 | quant = tens_data.Quantization() |
| 95 | |
| 96 | def len1_array_to_scalar(arr): |
| 97 | # The following flatbuffer quantisation fields all return a scalar value of 0 if they are not definied in |
| 98 | # the input buffer. This is represented in Vela by using None. |
| 99 | # Otherwise, the fields returned are a single or multi-element array. In which case, single element arrays |
| 100 | # are converted to scalars |
| 101 | if isinstance(arr, int) and arr == 0: |
| 102 | return None |
| 103 | if len(arr) == 1: |
| 104 | return arr[0] |
| 105 | return arr |
| 106 | |
| 107 | tens.quantization = QuantizationParameters() |
| 108 | tens.quantization.min = len1_array_to_scalar(quant.MinAsNumpy()) |
| 109 | tens.quantization.max = len1_array_to_scalar(quant.MaxAsNumpy()) |
| 110 | tens.quantization.scale_f32 = len1_array_to_scalar(quant.ScaleAsNumpy()) |
| 111 | tens.quantization.zero_point = len1_array_to_scalar(quant.ZeroPointAsNumpy()) |
| 112 | |
| 113 | if dtype == DataType.uint8: |
| 114 | tens.quantization.quant_min = 0 |
| 115 | tens.quantization.quant_max = (1 << dtype.bits) - 1 |
| 116 | elif dtype in set((DataType.int8, DataType.int16, DataType.int32, DataType.int64)): |
| 117 | tens.quantization.quant_min = -(1 << (dtype.bits - 1)) |
| 118 | tens.quantization.quant_max = (1 << (dtype.bits - 1)) - 1 |
| 119 | else: |
| 120 | raise Exception("DataType '" + str(dtype) + "' is not supported for quantization.") |
| 121 | |
| 122 | if tens.quantization.scale_f32 is None and tens.quantization.zero_point is None: |
| 123 | tens.quantization = None |
| 124 | |
| 125 | tens.values = None |
| 126 | buf = self.graph.buffers[tens_data.Buffer()] |
| 127 | if buf is not None: |
| 128 | tens.values = np.array(buf.view(datatype_map_numpy[tens_data.Type()]).reshape(shape)) |
| 129 | if tens.quantization is not None: |
| 130 | tens.quant_values = tens.values |
| 131 | tens.values = tens.quantization.dequantize(tens.quant_values) |
| 132 | return tens |
| 133 | |
| 134 | def parse_operator(self, op_data): |
| 135 | op_type, opt_serializer = self.graph.operator_codes[op_data.OpcodeIndex()] |
| 136 | inputs = [self.tensors[idx] for idx in op_data.InputsAsNumpy()] |
| 137 | outputs = [self.tensors[idx] for idx in op_data.OutputsAsNumpy()] |
| 138 | name = "unknown_op_name" |
| 139 | if len(outputs): |
| 140 | name = outputs[0].name |
| 141 | op = Operation(op_type, name) |
| 142 | op.inputs = inputs |
| 143 | op.outputs = outputs |
| 144 | for out in op.outputs: |
| 145 | out.ops = [op] |
| 146 | |
| 147 | activation_function_to_split_out = None |
| 148 | |
| 149 | if op_type.startswith("DepthwiseConv2d") or op_type.startswith("Conv2D"): |
| 150 | reshape_tensor_add_const_op(inputs[1], (1, 2, 3, 0)) |
| 151 | |
| 152 | if op_type.startswith("FullyConnected"): |
| 153 | reshape_tensor_add_const_op(inputs[1], (1, 0)) |
| 154 | |
| 155 | if opt_serializer is not None: |
| 156 | op.attrs = opt_serializer.deserialize(op_data.BuiltinOptions(), op_data.CustomOptionsAsNumpy()) |
| 157 | |
| 158 | if "stride_w" in op.attrs: |
| 159 | op.attrs["strides"] = (1, op.attrs["stride_h"], op.attrs["stride_w"], 1) |
| 160 | if "filter_width" in op.attrs: |
| 161 | op.attrs["ksize"] = (1, op.attrs["filter_height"], op.attrs["filter_width"], 1) |
| 162 | if "dilation_w_factor" in op.attrs: |
| 163 | op.attrs["dilation"] = (1, op.attrs["dilation_h_factor"], op.attrs["dilation_w_factor"], 1) |
| 164 | if "depth_multiplier" in op.attrs: |
| 165 | op.attrs["channel_multiplier"] = op.attrs["depth_multiplier"] |
| 166 | |
| 167 | if "fused_activation_function" in op.attrs: |
| 168 | if op_type in set(("ConcatTFLite",)): |
| 169 | act = op.attrs["fused_activation_function"] |
| 170 | del op.attrs["fused_activation_function"] |
| 171 | if act is not None: |
| 172 | activation_function_to_split_out = act |
| 173 | |
| 174 | if activation_function_to_split_out is not None: |
| 175 | act_op = Operation(activation_function_to_split_out, name + activation_function_to_split_out) |
| 176 | out_tens = op.outputs[0] |
| 177 | intermediate_tens = out_tens.clone("_act_intermediate") |
| 178 | out_tens.ops = [act_op] |
| 179 | act_op.outputs = [out_tens] |
| 180 | intermediate_tens.ops = [op] |
| 181 | op.outputs[0] = intermediate_tens |
| 182 | act_op.inputs = [intermediate_tens] |
| 183 | |
| 184 | |
| 185 | class TFLiteGraph: |
| 186 | def __init__( |
| 187 | self, |
| 188 | filename, |
| 189 | batch_size=1, |
| 190 | feed_dict={}, |
| 191 | output_node_names=[], |
| 192 | initialisation_nodes=[], |
| 193 | ): |
| 194 | |
| 195 | self.op_times = {} |
| 196 | if batch_size is None: |
| 197 | batch_size = 1 |
| 198 | self.batch_size = batch_size |
| 199 | self.name = os.path.splitext(os.path.basename(filename))[0] |
| 200 | self.initialisation_nodes = initialisation_nodes |
| 201 | |
| 202 | with open(filename, "rb") as f: |
| 203 | buf = bytearray(f.read()) |
| 204 | |
| 205 | model = Model.GetRootAsModel(buf, 0) |
| 206 | |
| 207 | self.buffers = [] |
| 208 | for idx in range(model.BuffersLength()): |
| 209 | self.buffers.append(self.parse_buffer(model.Buffers(idx))) |
| 210 | |
| 211 | self.operator_codes = [] |
| 212 | for idx in range(model.OperatorCodesLength()): |
| 213 | self.operator_codes.append(self.parse_operator_code(model.OperatorCodes(idx))) |
| 214 | |
| 215 | self.subgraphs = [] |
| 216 | for idx in range(model.SubgraphsLength()): |
| 217 | self.subgraphs.append(TFLiteSubgraph(self, model.Subgraphs(idx))) |
| 218 | |
| 219 | self.nng = Graph(self.name, self.batch_size) |
| 220 | for tflite_sg in self.subgraphs: |
| 221 | sg = Subgraph(tflite_sg.name) |
| 222 | sg.original_inputs = tflite_sg.inputs # Preserve the original input order |
| 223 | sg.output_tensors = tflite_sg.outputs |
| 224 | self.nng.subgraphs.append(sg) |
| 225 | |
| 226 | def parse_buffer(self, buf_data): |
| 227 | if buf_data.DataLength() == 0: |
| 228 | return None |
| 229 | data = buf_data.DataAsNumpy() |
| 230 | return data |
| 231 | |
| 232 | def parse_operator_code(self, code): |
| 233 | c = code.BuiltinCode() |
| 234 | op_type, ser = builtin_operator_map[c] |
| 235 | if c == BuiltinOperator.CUSTOM: |
| 236 | op_type += decode_str(code.CustomCode()) |
| 237 | return op_type, ser |
| 238 | |
| 239 | |
| 240 | def read_tflite( |
| 241 | filename, |
| 242 | batch_size=1, |
| 243 | feed_dict={}, |
| 244 | output_node_names=[], |
| 245 | initialisation_nodes=[], |
| 246 | ): |
| 247 | tflite_graph = TFLiteGraph( |
| 248 | filename, batch_size, feed_dict, output_node_names, initialisation_nodes |
| 249 | ) |
| 250 | nng = tflite_graph.nng |
| 251 | nng.refresh_after_modification() |
| 252 | return nng |