#!/usr/bin/env python3 # This file is part of the OpenMV project. # # Copyright (c) 2013-2022 Ibrahim Abdelkader # Copyright (c) 2013-2022 Kwabena W. Agyeman # # This work is licensed under the MIT license, see the file LICENSE for details. # # This script converts tflite models and labels to C structs. import sys import os import csv import glob import argparse import binascii import subprocess def run_vela(model_path, model_name, args): vela_dir = f'{args.build_dir}/{model_name}' vela_ini = os.path.dirname(os.path.abspath(__file__)) vela_args = args.vela_args.split() # Construct the command command = [ 'vela', *vela_args, '--output-dir', vela_dir, '--config', f'{vela_ini}/vela.ini', model_path ] # Call the command and capture the output try: result = subprocess.run(command, check=True, text=True, capture_output=True) except subprocess.CalledProcessError as e: print(e.stderr, file=sys.stderr) print(args.vela_args, file=sys.stderr) csv_file_path = glob.glob(os.path.join(vela_dir, "*.csv"))[0] with open(csv_file_path, mode='r') as file: row = next(csv.DictReader(file)) summary = { "Network:": row["network"], "Accelerator Configuration:": row["accelerator_configuration"], "System Configuration:": row["system_config"], "Memory Mode:": row["memory_mode"], "Compiler Mode: ": vela_args[-1], "Accelerator Clock:": str(int(float(row["core_clock"]) / 1000000))+" MHz", "Arena Size:": row["arena_cache_size"].split(".")[0], "Inference Time:": "%.2f ms, %.2f inferences/s"% (float(row["inference_time"]) * 1000, float(row["inferences_per_second"])), } print("", file=sys.stderr) for key, value in summary.items(): print(f"{key:<{30}} {value:<{50}}", file=sys.stderr) return f'{vela_dir}/{model_name}_vela.tflite' def main(): parser = argparse.ArgumentParser(description='Converts TFLite models to C file.') parser.add_argument('--input', action = 'store', help = 'Input tflite models directory.', required=True) parser.add_argument('--header', action = 'store_true', help = 'Generate header file.', required=False, default=False) parser.add_argument('--build-dir', action = 'store', help='Build directory', default='build', ) parser.add_argument('--vela-args', action = 'store', help='Vela compiler args', default='') args = parser.parse_args() tflm_builtin_models = [] tflm_builtin_models_index = {} print('/* NOTE: This file is auto-generated. */\n') index_headers = ['model', 'optimise'] # Open the file and parse it using DictReader with open(os.path.join(args.input, "index.csv"), 'r') as file: for row in csv.reader((line for line in file if not line.startswith('#'))): model = os.path.splitext(row[0])[0] tflm_builtin_models_index[model] = dict(zip(index_headers[1:], row[1:])) models_list = glob.glob(os.path.join(args.input, "*tflite")) if (args.header): # Generate the header file print('// Built-in TFLite Models.') print('typedef struct {') print(' const char *name;') print(' const unsigned int n_labels;') print(' const char **labels;') print(' const unsigned int size;') print(' const unsigned char *data;') print('}tflm_builtin_model_t;\n') print('extern const tflm_builtin_model_t tflm_builtin_models[];') else: # Generate the C file print('#include "imlib_config.h"') print('#include "tflm_builtin_models.h"') for model_path in models_list: model_size = os.path.getsize(model_path) model_name = os.path.basename(os.path.splitext(model_path)[0]) labels_file = os.path.splitext(model_path)[0]+'.txt' if (args.vela_args): args.vela_args += " --optimise %s"%tflm_builtin_models_index[model_name]["optimise"] # Compile the model using Vela and switch path to the new model. model_path = run_vela(model_path, model_name, args) model_size = os.path.getsize(model_path) # Generate model labels. labels = [] n_labels = 0 if os.path.exists(labels_file): with open(labels_file, 'r') as f: labels = ['"{:s}"'.format(l.strip()) for l in f.readlines()] n_labels = len(labels) print('static const char *tflm_{:s}_labels[] = {{{:s}}};'.format(model_name, ', '.join(labels))) # Generate model data. print('static const unsigned char tflm_{:s}_data[] __attribute__((aligned(16))) = {{'.format(model_name)) with open(model_path, 'rb') as f: for chunk in iter(lambda: f.read(12), b''): print(' ', end='') print(' '.join(['0x{:02x},'.format(x) for x in chunk])) print('};') # Store model info in builtin models table. tflm_builtin_models.append([ model_name, n_labels, 'tflm_{:s}_labels'.format(model_name), model_size, 'tflm_{:s}_data'.format(model_name)] ) # Generate built-in models table. print('const tflm_builtin_model_t tflm_builtin_models[] = {') for model in tflm_builtin_models: if model[0] in tflm_builtin_models_index: print(' #if defined(IMLIB_ENABLE_TFLM_BUILTIN_{:s})'.format(model[0].upper())) print(' {{ "{:s}", {:d}, {:s}, {:d}, {:s} }},'.format(*model)) if model[0] in tflm_builtin_models_index: print(' #endif') print(' {0, 0, 0, 0, 0}') print('};') if __name__ == '__main__': main()