Tensorflow object_detection API源码——mobileDet

object_detection>python model_main.py   --logtostderr   --model_dir=../image/   --pipeline_config_path=../image/ssdlite_mobiledet_cpu_320x320_coco_sync_4x4.config

import tensorflow.compat.v1 as tf
from absl import flags
from object_detection import model_lib

# ①===== 命令行参数设定
flags.DEFINE_string(
    ‘model_dir‘, None, ‘Path to output model directory ‘
                       ‘where event and checkpoint files will be written.‘)
flags.DEFINE_string(‘pipeline_config_path‘, None, ‘Path to pipeline config ‘)
FLAGS = flags.FLAGS


def main(unused_argv):
    flags.mark_flag_as_required(‘model_dir‘)
    flags.mark_flag_as_required(‘pipeline_config_path‘)

    # ②===== 模型部分： 由配置信息初始化，返回4个fn和一个estimator
    train_and_eval_dict = model_lib.create_estimator_and_inputs(  
        run_config=tf.estimator.RunConfig(model_dir=FLAGS.model_dir),   # "../image"
        pipeline_config_path=FLAGS.pipeline_config_path,  # .config
        train_steps=FLAGS.num_train_steps,  # None
        sample_1_of_n_eval_examples=FLAGS.sample_1_of_n_eval_examples,  # 1
        sample_1_of_n_eval_on_train_examples=FLAGS.sample_1_of_n_eval_on_train_examples)    # 5

    estimator = train_and_eval_dict[‘estimator‘]
    train_input_fn = train_and_eval_dict[‘train_input_fn‘]
    eval_input_fns = train_and_eval_dict[‘eval_input_fns‘]
    eval_on_train_input_fn = train_and_eval_dict[‘eval_on_train_input_fn‘]
    predict_input_fn = train_and_eval_dict[‘predict_input_fn‘]
    train_steps = train_and_eval_dict[‘train_steps‘]  # 4000

    # ③===== 数据部分： 由上一步返回的4个fn创建
    train_spec, eval_specs = model_lib.create_train_and_eval_specs(
        train_input_fn,
        eval_input_fns,
        eval_on_train_input_fn,
        predict_input_fn,
        train_steps,
        eval_on_train_data=False)

    # Currently only a single Eval Spec is allowed.
    # ④===== 训练
    tf.estimator.train_and_evaluate(estimator, train_spec, eval_specs[0])


if __name__ == ‘__main__‘:
    tf.app.run()

import functools
import copy
from object_detection.utils import config_util
from object_detection import inputs
def create_estimator_and_inputs(run_config,
                                pipeline_config_path=None,  # .config
                                sample_1_of_n_eval_examples=1,  # 1
                                sample_1_of_n_eval_on_train_examples=1,  # 5
                                model_fn_creator=create_model_fn,
                                **kwargs):

    configs = config_util.get_configs_from_pipeline_file(pipeline_config_path, config_override=None)  # None
    kwargs.update({
        ‘train_steps‘: None,  # because train_steps is None
        ‘use_bfloat16‘: False  # because configs[‘train_config‘].use_bfloat16 and use_tpu is False and False
    })
    kwargs.update({‘sample_1_of_n_eval_examples‘: sample_1_of_n_eval_examples})  # because (sample_1_of_n_eval_examples=1) >= 1
    kwargs.update({‘eval_num_epochs‘: 1})   # because override_eval_num_epochs is True
    tf.logging.warning(‘Forced number of epochs for all eval validations to be 1.‘)
    configs = config_util.merge_external_params_with_configs(configs, False, kwargs_dict=kwargs)  # ③合并configs和kwargs hparms=False

    model_config = configs[‘model‘]  # type is <class ‘object_detection.protos.model_pb2.DetectionModel‘>
    train_config = configs[‘train_config‘]
    train_input_config = configs[‘train_input_config‘]
    eval_config = configs[‘eval_config‘]
    eval_input_configs = configs[‘eval_input_configs‘]  # 是一个len=1的列表

    eval_on_train_input_config = copy.deepcopy(train_input_config)
    eval_on_train_input_config.sample_1_of_n_examples = (sample_1_of_n_eval_on_train_examples)
    eval_on_train_input_config.num_epochs = 1  # because override_eval_num_epochs=True

    # Create the input functions for TRAIN/EVAL/PREDICT.
    train_input_fn = input.train_input_fn(train_config=train_config, train_input_config=train_input_config, model_config=model_config)
    eval_input_fns = []
    eval_input_fns[0] = input.create_eval_input_fn(eval_config, eval_input_configs[0], model_config)
    eval_on_train_input_fn = input.create_eval_input_fn(eval_config=eval_config, eval_input_config=eval_on_train_input_config, model_config=model_config)
    predict_input_fn = input.create_predict_input_fn(model_config=model_config, predict_input_config=eval_input_configs[0])

    detection_model_fn = functools.partial(model_builder.build, model_config=model_config)
    model_fn = model_fn_creator(detection_model_fn, configs, False, False, False)  # use_tpu = False hparms=False postprocess_on_cpu =False
    estimator = tf.estimator.Estimator(model_fn=model_fn, config=run_config)

    # Write the as-run pipeline config to disk.  
    if run_config.is_chief and save_final_config: 
        pipeline_config_final = config_util.create_pipeline_proto_from_configs(configs)
        config_util.save_pipeline_config(pipeline_config_final, estimator.model_dir)

    return dict(
        estimator=estimator,
        train_input_fn=train_input_fn,
        eval_input_fns=eval_input_fns,
        eval_on_train_input_fn=eval_on_train_input_fn,
        predict_input_fn=predict_input_fn,
        train_steps=train_config.num_steps)  # 训练次数

feature_extractor = _build_ssd_feature_extractor(    # 同文件下的函数
        feature_extractor_config=ssd_config.feature_extractor,
        freeze_batchnorm=ssd_config.freeze_batchnorm,  # False
        is_training=is_training)