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. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 16 | # Description: |
| 17 | # Functions used to write to a TensorFlow Lite format file. Supports adding in file identifiers. |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 18 | import flatbuffers |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 19 | import flatbuffers.number_types as N |
| 20 | import numpy as np |
| 21 | from flatbuffers import encode |
Diego Russo | ea6111a | 2020-04-14 18:41:58 +0100 | [diff] [blame] | 22 | from flatbuffers.builder import UOffsetTFlags |
| 23 | |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 24 | from .nn_graph import PassPlacement |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 25 | from .tensor import MemType |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 26 | from .tensor import TensorPurpose |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 27 | from .tflite import Buffer |
| 28 | from .tflite import Metadata |
Diego Russo | e8a1045 | 2020-04-21 17:39:10 +0100 | [diff] [blame] | 29 | from .tflite import Model |
| 30 | from .tflite import Operator |
| 31 | from .tflite import OperatorCode |
| 32 | from .tflite import QuantizationParameters |
| 33 | from .tflite import SubGraph |
| 34 | from .tflite import Tensor |
| 35 | from .tflite_mapping import builtin_operator_inv_map |
| 36 | from .tflite_mapping import BuiltinOperator |
| 37 | from .tflite_mapping import custom_prefix |
| 38 | from .tflite_mapping import datatype_inv_map |
| 39 | |
| 40 | # ugh, the python flatbuffer interface is missing a method to add in file identifier. patching it in here: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 41 | |
| 42 | tflite_version = 3 |
| 43 | tflite_file_identifier = "TFL" + str(tflite_version) |
| 44 | |
| 45 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 46 | def FinishWithFileIdentifier(self, rootTable, fid): |
| 47 | if fid is None or len(fid) != 4: |
| 48 | raise Exception("fid must be 4 chars") |
| 49 | |
| 50 | flags = N.Uint8Flags |
| 51 | prepSize = 4 |
| 52 | self.Prep(self.minalign, prepSize + len(fid)) |
| 53 | for i in range(3, -1, -1): |
| 54 | self.head = self.head - flags.bytewidth |
| 55 | encode.Write(flags.packer_type, self.Bytes, self.Head(), ord(fid[i])) |
| 56 | |
| 57 | return self.Finish(rootTable) |
| 58 | |
| 59 | |
| 60 | flatbuffers.Builder.FinishWithFileIdentifier = FinishWithFileIdentifier |
| 61 | |
| 62 | |
| 63 | def make_vector(v): |
| 64 | try: |
| 65 | len(v) |
| 66 | return v |
| 67 | except TypeError: |
| 68 | return [v] |
| 69 | |
| 70 | |
| 71 | class TFLiteSerialiser: |
| 72 | def __init__(self, nng): |
| 73 | self.builder = flatbuffers.Builder(0) |
| 74 | self.nng = nng |
| 75 | |
| 76 | self.scratch_buf_id = 0 # Always assign scratch to buffer 0 |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 77 | self.scratch_fast_buf_id = 1 # Always assign scratch_fast to buffer 1 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 78 | self.buffer_offsets_map = {} |
| 79 | self.buffers_to_write = [] # have an empty array there |
| 80 | |
| 81 | self.input_tensors = [] |
| 82 | self.ops_to_ignore = set(("Const", "Placeholder", "SubgraphInput")) |
| 83 | |
| 84 | self.tensors_to_reshape = {} |
| 85 | |
| 86 | self.subgraphs_to_write = [sg for sg in self.nng.subgraphs if sg.placement == PassPlacement.Cpu] |
| 87 | |
| 88 | all_ops = [] |
| 89 | for sg in self.subgraphs_to_write: |
| 90 | for ps in sg.passes: |
| 91 | for op in ps.ops: |
| 92 | if op.type not in self.ops_to_ignore: |
| 93 | all_ops.append(op) |
| 94 | if op.type.startswith("Conv2D") or op.type.startswith("DepthwiseConv2d"): |
| 95 | self.tensors_to_reshape[op.inputs[1]] = (3, 0, 1, 2) |
| 96 | if op.type.startswith("FullyConnected"): |
| 97 | self.tensors_to_reshape[op.inputs[1]] = (1, 0) |
| 98 | |
| 99 | self.operator_codes = list(sorted(set(op.type for op in all_ops))) |
| 100 | self.operator_code_map = {} |
| 101 | |
| 102 | def write_byte_vector(self, v, alignment=1): |
| 103 | builder = self.builder |
| 104 | builder.StartVector(1, len(v), alignment) |
| 105 | for e in v[::-1]: |
| 106 | builder.PrependByte(e) |
| 107 | return builder.EndVector(len(v)) |
| 108 | |
| 109 | def write_int_vector(self, v): |
| 110 | builder = self.builder |
| 111 | builder.StartVector(4, len(v), 4) |
| 112 | for e in v[::-1]: |
| 113 | builder.PrependInt32(e) |
| 114 | return builder.EndVector(len(v)) |
| 115 | |
| 116 | def write_long_vector(self, v): |
| 117 | builder = self.builder |
| 118 | builder.StartVector(8, len(v), 8) |
| 119 | for e in v[::-1]: |
| 120 | builder.PrependInt64(e) |
| 121 | return builder.EndVector(len(v)) |
| 122 | |
| 123 | def write_float_vector(self, v): |
| 124 | builder = self.builder |
| 125 | builder.StartVector(4, len(v), 4) |
| 126 | for e in v[::-1]: |
| 127 | builder.PrependFloat32(e) |
| 128 | return builder.EndVector(len(v)) |
| 129 | |
| 130 | def write_offset_vector(self, v): |
| 131 | builder = self.builder |
| 132 | builder.StartVector(4, len(v), 4) |
| 133 | for e in v[::-1]: |
| 134 | builder.PrependUOffsetTRelative(e) |
| 135 | return builder.EndVector(len(v)) |
| 136 | |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 137 | def assign_buffers_to_tensors(self, tensors, scratch_tensor): |
| 138 | if scratch_tensor is not None: |
| 139 | scratch_tensor_mem_area = scratch_tensor.mem_area |
Tim Hall | 25f605c | 2020-05-18 18:04:26 +0100 | [diff] [blame] | 140 | else: |
| 141 | scratch_tensor_mem_area = None # all tensors are initialised to MemArea.Unknown |
| 142 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 143 | buffer_map = {} |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 144 | |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 145 | buf_idx = 2 |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 146 | |
| 147 | for tens in tensors: |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 148 | # Set buffer ids depending on allocation |
| 149 | if tens.is_allocated_in_tensor_arena(scratch_tensor_mem_area): |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 150 | buffer_map[tens] = self.scratch_buf_id |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 151 | elif tens.mem_type == MemType.Scratch_fast: |
| 152 | # For Scratch_fast when not co-allocated with scratch in the TensorArena: |
| 153 | buffer_map[tens] = self.scratch_fast_buf_id |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 154 | else: |
| 155 | buffer_map[tens] = buf_idx |
| 156 | buf_idx += 1 |
| 157 | |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 158 | # Initialize buffers_to_write to a length equal to number of buffers so |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 159 | # they can be appended at the correct index during tensor serialization |
| 160 | self.buffers_to_write = [None] * (buf_idx) |
| 161 | |
| 162 | return buffer_map |
| 163 | |
| 164 | def serialise_operator_code(self, idx, code): |
| 165 | builder = self.builder |
| 166 | custom_code_offset = None |
| 167 | if code.startswith(custom_prefix): |
| 168 | tf_code, opt_serializer = builtin_operator_inv_map[custom_prefix] |
| 169 | custom_code_offset = builder.CreateString(code[len(custom_prefix) :]) |
| 170 | else: |
Tim Hall | e9194df | 2020-08-04 20:37:01 +0100 | [diff] [blame] | 171 | assert ( |
| 172 | code in builtin_operator_inv_map |
| 173 | ), "Vela does not contain a mapping to serialise {} operator to a TensorFlow Lite operator".format(code) |
| 174 | tf_code, opt_serializer = builtin_operator_inv_map[code] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 175 | |
| 176 | if tf_code == BuiltinOperator.CUSTOM: |
Tim Hall | e9194df | 2020-08-04 20:37:01 +0100 | [diff] [blame] | 177 | assert ( |
| 178 | code == "NpuOp" |
| 179 | ), "Vela only supports serialising NpuOp operators as TensorFlow Lite Custom operators" |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 180 | custom_code_offset = builder.CreateString("ethos-u") |
| 181 | |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 182 | self.operator_code_map[code] = (idx, tf_code, opt_serializer) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 183 | |
| 184 | OperatorCode.OperatorCodeStart(builder) |
| 185 | OperatorCode.OperatorCodeAddBuiltinCode(builder, tf_code) |
| 186 | if custom_code_offset is not None: |
| 187 | OperatorCode.OperatorCodeAddCustomCode(builder, custom_code_offset) |
| 188 | |
| 189 | return OperatorCode.OperatorCodeEnd(builder) |
| 190 | |
| 191 | def serialise_quantization_parameters(self, quant): |
| 192 | builder = self.builder |
| 193 | |
| 194 | min = None |
| 195 | max = None |
| 196 | scale = None |
| 197 | zero_point = None |
| 198 | if quant is not None: |
| 199 | if quant.min is not None: |
| 200 | min = self.write_float_vector(make_vector(quant.min)) |
| 201 | if quant.max is not None: |
| 202 | max = self.write_float_vector(make_vector(quant.max)) |
| 203 | if quant.scale_f32 is not None: |
| 204 | scale = self.write_float_vector(make_vector(quant.scale_f32)) |
| 205 | if quant.zero_point is not None: |
| 206 | zero_point = self.write_long_vector(make_vector(quant.zero_point)) |
| 207 | |
| 208 | QuantizationParameters.QuantizationParametersStart(builder) |
| 209 | if min is not None: |
| 210 | QuantizationParameters.QuantizationParametersAddMin(builder, min) |
| 211 | if max is not None: |
| 212 | QuantizationParameters.QuantizationParametersAddMax(builder, max) |
| 213 | if scale is not None: |
| 214 | QuantizationParameters.QuantizationParametersAddScale(builder, scale) |
| 215 | if zero_point is not None: |
| 216 | QuantizationParameters.QuantizationParametersAddZeroPoint(builder, zero_point) |
| 217 | return QuantizationParameters.QuantizationParametersEnd(builder) |
| 218 | |
| 219 | def serialise_tensor(self, tens): |
| 220 | builder = self.builder |
| 221 | tens_shape = tens.shape |
| 222 | values = tens.quant_values |
| 223 | if values is None: |
| 224 | values = tens.values |
| 225 | |
| 226 | if values is None: |
| 227 | values = np.empty(shape=(0), dtype=np.uint8) |
| 228 | |
| 229 | if tens in self.tensors_to_reshape: |
| 230 | reorder = self.tensors_to_reshape[tens] |
| 231 | tens_shape = [tens_shape[idx] for idx in reorder] |
| 232 | values = values.transpose(reorder) |
| 233 | |
| 234 | if tens.purpose == TensorPurpose.Scratch: |
| 235 | tens_shape = [0] |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 236 | |
| 237 | buf_id = self.buffer_map[tens] |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 238 | self.buffers_to_write[buf_id] = values.flatten().view(np.uint8) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 239 | |
| 240 | shape = self.write_int_vector(tens_shape) |
| 241 | |
| 242 | name = builder.CreateString(tens.name) |
| 243 | quant = self.serialise_quantization_parameters(tens.quantization) |
| 244 | |
| 245 | Tensor.TensorStart(builder) |
| 246 | Tensor.TensorAddShape(builder, shape) |
| 247 | Tensor.TensorAddType(builder, datatype_inv_map[tens.dtype]) |
| 248 | # All tensors must have a valid backing buffer, even if it is empty. |
| 249 | # Empty buffers should be kept unique for TensorFlow Lite Micro |
| 250 | Tensor.TensorAddBuffer(builder, buf_id) |
| 251 | Tensor.TensorAddName(builder, name) |
| 252 | Tensor.TensorAddQuantization(builder, quant) |
| 253 | |
| 254 | res = Tensor.TensorEnd(builder) |
| 255 | return res |
| 256 | |
| 257 | def serialise_operator(self, op): |
| 258 | builder = self.builder |
| 259 | |
Michael McGeagh | bb1b09e | 2020-08-19 11:24:17 +0100 | [diff] [blame] | 260 | inputs_offset = self.write_int_vector([self.tensor_map[tens] for tens in op.inputs if tens in self.tensor_map]) |
| 261 | outputs_offset = self.write_int_vector( |
| 262 | [self.tensor_map[tens] for tens in op.outputs if tens in self.tensor_map] |
| 263 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 264 | |
| 265 | op_idx, tflop, opt_serializer = self.operator_code_map[op.type] |
| 266 | |
| 267 | builtin_opt_offset = None |
| 268 | custom_opt_offset = None |
| 269 | if opt_serializer is not None: |
| 270 | attrs = dict(op.attrs) |
| 271 | if "strides" in attrs: |
| 272 | attrs["stride_h"] = attrs["strides"][1] |
| 273 | attrs["stride_w"] = attrs["strides"][2] |
| 274 | if "ksize" in attrs: |
| 275 | attrs["filter_height"] = attrs["ksize"][1] |
| 276 | attrs["filter_width"] = attrs["ksize"][2] |
| 277 | if "dilation" in attrs: |
| 278 | attrs["dilation_h_factor"] = attrs["dilation"][1] |
| 279 | attrs["dilation_w_factor"] = attrs["dilation"][2] |
| 280 | if "channel_multiplier" in attrs: |
| 281 | attrs["depth_multiplier"] = attrs["channel_multiplier"] |
| 282 | |
| 283 | builtin_opt_offset, custom_opt_offset = opt_serializer.serialize(builder, attrs) |
| 284 | |
| 285 | mutating_variable_inputs_offset = self.write_byte_vector([]) |
| 286 | Operator.OperatorStart(builder) |
| 287 | Operator.OperatorAddOpcodeIndex(builder, op_idx) |
| 288 | Operator.OperatorAddInputs(builder, inputs_offset) |
| 289 | Operator.OperatorAddOutputs(builder, outputs_offset) |
| 290 | |
| 291 | if builtin_opt_offset is not None: |
| 292 | Operator.OperatorAddBuiltinOptionsType(builder, opt_serializer.builtin_opt_type) |
| 293 | Operator.OperatorAddBuiltinOptions(builder, builtin_opt_offset) |
| 294 | if custom_opt_offset is not None: |
| 295 | Operator.OperatorAddCustomOptions(builder, custom_opt_offset) |
| 296 | Operator.OperatorAddCustomOptionsFormat(builder, opt_serializer.custom_opt_format) |
| 297 | |
| 298 | Operator.OperatorAddMutatingVariableInputs(builder, mutating_variable_inputs_offset) |
| 299 | return Operator.OperatorEnd(builder) |
| 300 | |
| 301 | def serialise_subgraph(self, sg): |
| 302 | builder = self.builder |
| 303 | tensor_set = set() |
| 304 | |
| 305 | all_ops = [] |
| 306 | for ps in sg.passes: |
| 307 | for op in ps.ops: |
| 308 | if op.type not in self.ops_to_ignore: |
| 309 | all_ops.append(op) |
| 310 | |
| 311 | for op in all_ops: |
| 312 | for tens in op.inputs + op.outputs: |
| 313 | tensor_set.add(tens) |
| 314 | |
| 315 | all_tensors = [tens for nm, idx, tens in sorted((tens.name, idx, tens) for idx, tens in enumerate(tensor_set))] |
| 316 | |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 317 | scratch_tensors = [tens for tens in all_tensors if tens.name.endswith("scratch")] |
| 318 | |
Jacob Bohlin | 68a04b1 | 2020-07-13 11:39:36 +0200 | [diff] [blame] | 319 | scratch_fast_tensor = None |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 320 | for tens in all_tensors: |
| 321 | if tens.name.endswith("scratch_fast"): |
| 322 | scratch_fast_tensor = tens |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 323 | |
| 324 | if len(scratch_tensors) == 0: |
| 325 | scratch_tensor = None |
| 326 | else: |
| 327 | assert len(scratch_tensors) == 1, "Multiple scratch tensors" |
| 328 | scratch_tensor = scratch_tensors[0] |
| 329 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 330 | self.tensor_map = {tens: idx for idx, tens in enumerate(all_tensors)} |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 331 | self.buffer_map = self.assign_buffers_to_tensors(all_tensors, scratch_tensor) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 332 | |
| 333 | tensors_offset = self.write_offset_vector([self.serialise_tensor(tens) for tens in all_tensors]) |
| 334 | |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 335 | # Make sure the input_tensors haven't been modified |
| 336 | assert all(inp in sg.original_inputs for inp in sg.input_tensors) |
Michael McGeagh | bb1b09e | 2020-08-19 11:24:17 +0100 | [diff] [blame] | 337 | inputs = [self.tensor_map[tens] for tens in sg.original_inputs if tens in self.tensor_map] |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 338 | |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 339 | # Add the Scratch Tensors as input to the NPU subgraph to get them allocated by TensorFlow Lite Micro |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 340 | scratch_tensor_idx = self.tensor_map.get(scratch_tensor, None) |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 341 | scratch_fast_tensor_idx = self.tensor_map.get(scratch_fast_tensor, None) |
| 342 | |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 343 | if scratch_tensor_idx is not None and scratch_tensor_idx not in inputs: |
| 344 | inputs.append(scratch_tensor_idx) |
| 345 | |
Patrik Gustavsson | 3ab9452 | 2020-06-29 17:36:55 +0200 | [diff] [blame] | 346 | if scratch_fast_tensor_idx is not None and scratch_fast_tensor_idx not in inputs: |
| 347 | inputs.append(scratch_fast_tensor_idx) |
| 348 | |
Tim Hall | c8310b1 | 2020-06-17 14:53:11 +0100 | [diff] [blame] | 349 | inputs_offset = self.write_int_vector(inputs) |
Michael McGeagh | bb1b09e | 2020-08-19 11:24:17 +0100 | [diff] [blame] | 350 | outputs_offset = self.write_int_vector( |
| 351 | [self.tensor_map[tens] for tens in sg.output_tensors if tens in self.tensor_map] |
| 352 | ) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 353 | |
| 354 | operators_offset = self.write_offset_vector([self.serialise_operator(op) for op in all_ops]) |
| 355 | |
| 356 | SubGraph.SubGraphStart(builder) |
| 357 | SubGraph.SubGraphAddTensors(builder, tensors_offset) |
| 358 | SubGraph.SubGraphAddInputs(builder, inputs_offset) |
| 359 | SubGraph.SubGraphAddOutputs(builder, outputs_offset) |
| 360 | |
| 361 | SubGraph.SubGraphAddOperators(builder, operators_offset) |
| 362 | |
| 363 | return SubGraph.SubGraphEnd(builder) |
| 364 | |
| 365 | def write_aligned_bytes(self, buf): |
| 366 | builder = self.builder |
| 367 | builder.nested = True |
| 368 | data = bytes(buf) |
| 369 | length_bytes = UOffsetTFlags.py_type(len(data)) |
| 370 | builder.Prep(16, length_bytes) # Reserve aligned storage |
| 371 | builder.head = UOffsetTFlags.py_type(builder.Head() - length_bytes) # Update FlatBuffer internal pointer |
| 372 | builder.Bytes[builder.Head() : builder.Head() + length_bytes] = data # Assign bytes to aligned area |
| 373 | return builder.EndVector(length_bytes) |
| 374 | |
| 375 | def serialise_buffer(self, buf): |
| 376 | builder = self.builder |
| 377 | data = None |
| 378 | if buf is not None: |
| 379 | data = self.write_aligned_bytes(buf) |
| 380 | Buffer.BufferStart(builder) |
| 381 | if data is not None: |
| 382 | Buffer.BufferAddData(builder, data) |
| 383 | return Buffer.BufferEnd(builder) |
| 384 | |
| 385 | def serialise_metadata(self, metadata): |
| 386 | builder = self.builder |
| 387 | name = builder.CreateString(metadata[0]) |
| 388 | |
| 389 | Metadata.MetadataStart(builder) |
| 390 | Metadata.MetadataAddName(builder, name) |
| 391 | Metadata.MetadataAddBuffer(builder, metadata[1]) |
| 392 | |
| 393 | return Metadata.MetadataEnd(builder) |
| 394 | |
| 395 | def serialise_model(self): |
| 396 | builder = self.builder |
| 397 | operator_code_offset = self.write_offset_vector( |
| 398 | [self.serialise_operator_code(idx, code) for idx, code in enumerate(self.operator_codes)] |
| 399 | ) |
| 400 | |
| 401 | description = builder.CreateString("Vela Optimised") |
| 402 | |
| 403 | subgraph_offset = self.write_offset_vector([self.serialise_subgraph(sg) for sg in self.subgraphs_to_write]) |
| 404 | |
| 405 | # Fill the metadata buffer |
| 406 | version = np.int32(0) |
| 407 | subgraph_idx = np.int32(len(self.subgraphs_to_write)) # Only 1 supported currently |
| 408 | nbr_tensors = np.int32(len(self.tensor_map)) |
| 409 | |
| 410 | # An offset of -1 indicates that the tensor will be allocated online by Tensorflow Lite Micro |
| 411 | offsets = [np.int32(-1)] * nbr_tensors |
| 412 | |
| 413 | # Ensure that the order of the offsets match the order of the tensors |
| 414 | for tens, idx in self.tensor_map.items(): |
Patrik Gustavsson | eca2e95 | 2020-05-27 09:15:11 +0200 | [diff] [blame] | 415 | # Set offsets for tensor allocated in Tensor Arena or in the scratch_fast area |
Charles Xu | 04ce34c | 2020-06-23 12:42:28 +0200 | [diff] [blame] | 416 | if tens.mem_type in set((MemType.Scratch, MemType.Scratch_fast)) and tens.address is not None: |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 417 | offsets[idx] = np.int32(tens.address) |
| 418 | |
Michael McGeagh | 22f74e1 | 2020-08-07 16:21:03 +0100 | [diff] [blame] | 419 | self.nng.metadata.append(("OfflineMemoryAllocation", np.array([version, subgraph_idx, nbr_tensors] + offsets))) |
| 420 | |
| 421 | metadata_list = [] |
| 422 | for name, buffer in self.nng.metadata: |
| 423 | self.buffers_to_write.append(buffer) |
| 424 | metadata_list.append((name, len(self.buffers_to_write) - 1)) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 425 | |
| 426 | buffers_offset = self.write_offset_vector([self.serialise_buffer(buf) for buf in self.buffers_to_write]) |
Tim Hall | 79d07d2 | 2020-04-27 18:20:16 +0100 | [diff] [blame] | 427 | metadata_offset = self.write_offset_vector([self.serialise_metadata(metadata) for metadata in metadata_list]) |
| 428 | |
| 429 | Model.ModelStart(builder) |
| 430 | Model.ModelAddVersion(builder, tflite_version) |
| 431 | Model.ModelAddOperatorCodes(builder, operator_code_offset) |
| 432 | Model.ModelAddSubgraphs(builder, subgraph_offset) |
| 433 | Model.ModelAddDescription(builder, description) |
| 434 | Model.ModelAddBuffers(builder, buffers_offset) |
| 435 | Model.ModelAddMetadata(builder, metadata_offset) |
| 436 | return Model.ModelEnd(builder) |
| 437 | |
| 438 | def serialise(self): |
| 439 | |
| 440 | model = self.serialise_model() |
| 441 | |
| 442 | self.builder.FinishWithFileIdentifier(model, tflite_file_identifier) |
| 443 | |
| 444 | return self.builder.Output() |
| 445 | |
| 446 | def write(self, filename): |
| 447 | with open(self.filename, "wb") as f: |
| 448 | f.write(self.serialised_buf) |
| 449 | |
| 450 | |
| 451 | def write_tflite(nng, filename): |
| 452 | writer = TFLiteSerialiser(nng) |
| 453 | buf = writer.serialise() |
| 454 | |
| 455 | with open(filename, "wb") as f: |
| 456 | f.write(buf) |