예제 #1
0
def train_once(
        sess,
        step,
        ops,
        names=None,
        gen_feed_dict_fn=None,
        deal_results_fn=melt.print_results,
        interval_steps=100,
        eval_ops=None,
        eval_names=None,
        gen_eval_feed_dict_fn=None,
        deal_eval_results_fn=melt.print_results,
        eval_interval_steps=100,
        print_time=True,
        print_avg_loss=True,
        model_dir=None,
        log_dir=None,
        is_start=False,
        num_steps_per_epoch=None,
        metric_eval_fn=None,
        metric_eval_interval_steps=0,
        summary_excls=None,
        fixed_step=None  #for epoch only, incase you change batch size
):

    timer = gezi.Timer()
    if print_time:
        if not hasattr(train_once, 'timer'):
            train_once.timer = Timer()
            train_once.eval_timer = Timer()
            train_once.metric_eval_timer = Timer()

    melt.set_global('step', step)
    epoch = (fixed_step
             or step) / num_steps_per_epoch if num_steps_per_epoch else -1
    epoch_str = 'epoch:%.4f' % (epoch) if num_steps_per_epoch else ''
    melt.set_global('epoch', '%.4f' % (epoch))

    info = BytesIO()
    stop = False

    if is_start or eval_interval_steps and step % eval_interval_steps == 0:
        if eval_ops is not None:
            if deal_eval_results_fn is None and eval_names is not None:
                deal_eval_results_fn = lambda x: melt.print_results(
                    x, eval_names)

            eval_feed_dict = {} if gen_eval_feed_dict_fn is None else gen_eval_feed_dict_fn(
            )
            #eval_feed_dict.update(feed_dict)

            #------show how to perf debug
            ##timer_ = gezi.Timer('sess run generate')
            ##sess.run(eval_ops[-2], feed_dict=None)
            ##timer_.print()

            timer_ = gezi.Timer('sess run eval_ops')
            eval_results = sess.run(eval_ops, feed_dict=eval_feed_dict)
            eval_loss = gezi.get_singles(eval_results)
            timer_.print()
            if deal_eval_results_fn is not None:
                #@TODO user print should also use logging as a must ?
                #print(gezi.now_time(), epoch_str, 'eval_step: %d'%step, 'eval_metrics:', end='')
                logging.info2('{} eval_step: {} eval_metrics:{}'.format(
                    epoch_str, step, eval_loss))
                eval_stop = deal_eval_results_fn(eval_results)

            assert len(eval_loss) > 0
            if eval_stop is True: stop = True
            eval_names_ = melt.adjust_names(eval_loss, eval_names)

            melt.set_global('eval_loss',
                            melt.parse_results(eval_loss, eval_names_))
        elif interval_steps != eval_interval_steps:
            #print()
            pass

    metric_evaluate = False
    # if metric_eval_function is not None \
    #   and ( (is_start and (step or ops is None))\
    #     or (step and ((num_steps_per_epoch and step % num_steps_per_epoch == 0) \
    #            or (metric_eval_interval_steps \
    #                and step % metric_eval_interval_steps == 0)))):
    #     metric_evaluate = True
    if metric_eval_fn is not None \
      and (is_start \
        or (num_steps_per_epoch and step % num_steps_per_epoch == 0) \
             or (metric_eval_interval_steps \
                 and step % metric_eval_interval_steps == 0)):
        metric_evaluate = True

    if metric_evaluate:
        evaluate_results, evaluate_names = metric_eval_fn()

    if ops is not None:
        if deal_results_fn is None and names is not None:
            deal_results_fn = lambda x: melt.print_results(x, names)

        if eval_names is None:
            eval_names = names

        feed_dict = {} if gen_feed_dict_fn is None else gen_feed_dict_fn()

        results = sess.run(ops, feed_dict=feed_dict)
        # #--------trace debug
        # if step == 210:
        #   run_metadata = tf.RunMetadata()
        #   results = sess.run(
        #         ops,
        #         feed_dict=feed_dict,
        #         options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE),
        #         run_metadata=run_metadata)
        #   from tensorflow.python.client import timeline
        #   trace = timeline.Timeline(step_stats=run_metadata.step_stats)

        #   trace_file = open('timeline.ctf.json', 'w')
        #   trace_file.write(trace.generate_chrome_trace_format())

        #reults[0] assume to be train_op
        results = results[1:]

        #@TODO should support aver loss and other avg evaluations like test..
        if print_avg_loss:
            if not hasattr(train_once, 'avg_loss'):
                train_once.avg_loss = AvgScore()
                if interval_steps != eval_interval_steps:
                    train_once.avg_loss2 = AvgScore()
            #assume results[0] as train_op return, results[1] as loss
            loss = gezi.get_singles(results)
            train_once.avg_loss.add(loss)
            if interval_steps != eval_interval_steps:
                train_once.avg_loss2.add(loss)

        if is_start or interval_steps and step % interval_steps == 0:
            train_average_loss = train_once.avg_loss.avg_score()
            if print_time:
                duration = timer.elapsed()
                duration_str = 'duration:{:.3f} '.format(duration)
                melt.set_global('duration', '%.3f' % duration)
                info.write(duration_str)
                elapsed = train_once.timer.elapsed()
                steps_per_second = interval_steps / elapsed
                batch_size = melt.batch_size()
                num_gpus = melt.num_gpus()
                instances_per_second = interval_steps * batch_size * num_gpus / elapsed
                gpu_info = '' if num_gpus <= 1 else ' gpus:[{}]'.format(
                    num_gpus)
                if num_steps_per_epoch is None:
                    epoch_time_info = ''
                else:
                    epoch_time_info = ' 1epoch:[{:.2f}h]'.format(
                        num_steps_per_epoch / interval_steps * elapsed / 3600)
                info.write(
                    'elapsed:[{:.3f}] batch_size:[{}]{} batches/s:[{:.2f}] insts/s:[{:.2f}] {} '
                    .format(elapsed, batch_size, gpu_info, steps_per_second,
                            instances_per_second, epoch_time_info))

            if print_avg_loss:
                #info.write('train_avg_metrics:{} '.format(melt.value_name_list_str(train_average_loss, names)))
                names_ = melt.adjust_names(train_average_loss, names)
                info.write('train_avg_metric:{} '.format(
                    melt.parse_results(train_average_loss, names_)))
                #info.write('train_avg_loss: {} '.format(train_average_loss))

            #print(gezi.now_time(), epoch_str, 'train_step:%d'%step, info.getvalue(), end=' ')
            logging.info2('{} {} {}'.format(epoch_str, 'train_step:%d' % step,
                                            info.getvalue()))

            if deal_results_fn is not None:
                stop = deal_results_fn(results)

    if is_start or eval_interval_steps and step % eval_interval_steps == 0:
        if ops is not None:
            if interval_steps != eval_interval_steps:
                train_average_loss = train_once.avg_loss2.avg_score()

            info = BytesIO()

            names_ = melt.adjust_names(results, names)

            train_average_loss_str = ''
            if print_avg_loss and interval_steps != eval_interval_steps:
                train_average_loss_str = melt.value_name_list_str(
                    train_average_loss, names_)
                melt.set_global('train_loss', train_average_loss_str)
                train_average_loss_str = 'train_avg_metric:{} '.format(
                    train_average_loss_str)

            if interval_steps != eval_interval_steps:
                #end = '' if eval_ops is None else '\n'
                #print(gezi.now_time(), epoch_str, 'eval_step: %d'%step, train_average_loss_str, end=end)
                logging.info2('{} eval_step: {} {}'.format(
                    epoch_str, step, train_average_loss_str))

        if log_dir:
            if not hasattr(train_once, 'summary_op'):
                #melt.print_summary_ops()
                if summary_excls is None:
                    train_once.summary_op = tf.summary.merge_all()
                else:
                    summary_ops = []
                    for op in tf.get_collection(tf.GraphKeys.SUMMARIES):
                        for summary_excl in summary_excls:
                            if not summary_excl in op.name:
                                summary_ops.append(op)
                    print('filtered summary_ops:')
                    for op in summary_ops:
                        print(op)
                    train_once.summary_op = tf.summary.merge(summary_ops)

                train_once.summary_train_op = tf.summary.merge_all(
                    key=melt.MonitorKeys.TRAIN)
                train_once.summary_writer = tf.summary.FileWriter(
                    log_dir, sess.graph)

                tf.contrib.tensorboard.plugins.projector.visualize_embeddings(
                    train_once.summary_writer, projector_config)

            summary = tf.Summary()
            #so the strategy is on eval_interval_steps, if has eval dataset, then tensorboard evluate on eval dataset
            #if not have eval dataset, will evaluate on trainset, but if has eval dataset we will also monitor train loss
            if train_once.summary_train_op is not None:
                summary_str = sess.run(train_once.summary_train_op,
                                       feed_dict=feed_dict)
                train_once.summary_writer.add_summary(summary_str, step)

            if eval_ops is None:
                #get train loss, for every batch train
                if train_once.summary_op is not None:
                    #timer2 = gezi.Timer('sess run')
                    try:
                        summary_str = sess.run(train_once.summary_op,
                                               feed_dict=feed_dict)
                    except Exception:
                        summary_str = ''
                    #timer2.print()
                    train_once.summary_writer.add_summary(summary_str, step)
            else:
                #get eval loss for every batch eval, then add train loss for eval step average loss
                try:
                    summary_str = sess.run(
                        train_once.summary_op, feed_dict=eval_feed_dict
                    ) if train_once.summary_op is not None else ''
                except Exception:
                    logging.warning(
                        'summary_str = sess.run(train_once.summary_op, feed_dict=eval_feed_dict) fail'
                    )
                    #logging.warning(traceback.format_exc())
                    summary_str = ''
                #all single value results will be add to summary here not using tf.scalar_summary..
                summary.ParseFromString(summary_str)
                suffix = 'eval' if not eval_names else ''
                melt.add_summarys(summary,
                                  eval_results,
                                  eval_names_,
                                  suffix=suffix)

            melt.add_summarys(summary,
                              train_average_loss,
                              names_,
                              suffix='train_avg%dsteps' % eval_interval_steps)

            if metric_evaluate:
                melt.add_summarys(summary,
                                  evaluate_results,
                                  evaluate_names,
                                  prefix='eval')

            train_once.summary_writer.add_summary(summary, step)
            train_once.summary_writer.flush()

            #timer_.print()

        if print_time:
            full_duration = train_once.eval_timer.elapsed()
            if metric_evaluate:
                metric_full_duration = train_once.metric_eval_timer.elapsed()
            full_duration_str = 'elapsed:{:.3f} '.format(full_duration)
            #info.write('duration:{:.3f} '.format(timer.elapsed()))
            duration = timer.elapsed()
            info.write('duration:{:.3f} '.format(duration))
            info.write(full_duration_str)
            info.write('eval_time_ratio:{:.3f} '.format(duration /
                                                        full_duration))
            if metric_evaluate:
                info.write('metric_time_ratio:{:.3f} '.format(
                    duration / metric_full_duration))
        #print(gezi.now_time(), epoch_str, 'eval_step: %d'%step, info.getvalue())
        logging.info2('{} {} {}'.format(epoch_str, 'eval_step: %d' % step,
                                        info.getvalue()))

        return stop
예제 #2
0
def tf_train_flow(
        train_once_fn,
        model_dir='./model',
        max_models_keep=1,
        save_interval_seconds=600,
        save_interval_steps=1000,
        num_epochs=None,
        num_steps=None,
        save_model=True,
        save_interval_epochs=1,
        num_steps_per_epoch=0,
        restore_from_latest=True,
        metric_eval_fn=None,
        init_fn=None,
        restore_fn=None,
        restore_scope=None,
        save_all_scope=False,  #TODO save load from restore scope only but svae all
        variables_to_restore=None,
        variables_to_save=None,  #by default will be the same as variables_to_restore
        sess=None):
    """
  similary flow as tf_flow, but add model try reload and save
  """
    if sess is None:
        #TODO melt.get_session is global session but may cause non close at last
        sess = melt.get_session()
    logging.info('tf_train_flow start')
    print('max_models_keep:', max_models_keep, file=sys.stderr)
    print('save_interval_seconds:', save_interval_seconds, file=sys.stderr)

    #this is usefull for you use another model with another scope, and just load and restore/save initalize your scope vars!
    #this is not for finetune but mainly for like using another model as in predict like this introducing graph other model scope and ignore here

    var_list = None if not restore_scope else tf.get_collection(
        tf.GraphKeys.GLOBAL_VARIABLES, scope=restore_scope)
    if not variables_to_restore:
        variables_to_restore = var_list
    if not variables_to_save:
        variables_to_save = variables_to_restore
    if save_all_scope:
        variables_to_save = None

    if variables_to_restore is None:
        #load all var in checkpoint try to save all var(might more then original checkpoint) if not specifiy variables_to_save
        varnames_in_checkpoint = melt.get_checkpoint_varnames(model_dir)
        #print(varnames_in_checkpoint)
        variables_to_restore = slim.get_variables_to_restore(
            include=varnames_in_checkpoint)

    #logging.info('variables_to_restore:{}'.format(variables_to_restore))
    loader = tf.train.Saver(var_list=variables_to_restore)

    saver = tf.train.Saver(
        max_to_keep=max_models_keep,
        keep_checkpoint_every_n_hours=save_interval_seconds / 3600.0,
        var_list=variables_to_save)
    epoch_saver = tf.train.Saver(var_list=variables_to_save, max_to_keep=1000)
    best_epoch_saver = tf.train.Saver(var_list=variables_to_save)

    ##TODO for safe restore all init will be ok ?
    #if variables_to_restore is None:
    init_op = tf.group(
        tf.global_variables_initializer(
        ),  #variables_initializer(global_variables())
        tf.local_variables_initializer()
    )  #variables_initializer(local_variables())
    # else:
    #   init_op = tf.group(tf.variables_initializer(variables_to_restore),
    #                      tf.local_variables_initializer())

    ##--mostly this will be fine except for using assistant predictor, initialize again! will make assistant predictor wrong
    ##so assume to all run init op! if using assistant predictor, make sure it use another session

    sess.run(init_op)

    #melt.init_uninitialized_variables(sess)

    #pre_step means the step last saved, train without pretrained,then -1
    pre_step = -1
    fixed_pre_step = -1  #fixed pre step is for epoch num to be correct if yu change batch size
    model_path = _get_model_path(model_dir, save_model)
    model_dir = gezi.get_dir(
        model_dir)  #incase you pass ./model/model-ckpt1000 -> ./model
    if model_path is not None:
        if not restore_from_latest:
            print('using recent but not latest model', file=sys.stderr)
            model_path = melt.recent_checkpoint(model_dir)
        model_name = os.path.basename(model_path)
        timer = gezi.Timer('Loading and training from existing model [%s]' %
                           model_path)
        if restore_fn is not None:
            restore_fn(sess)
        loader.restore(sess, model_path)
        timer.print()
        pre_step = melt.get_model_step(model_path)
        pre_epoch = melt.get_model_epoch(model_path)
        fixed_pre_step = pre_step
        if pre_epoch is not None:
            #like using batch size 32, then reload train using batch size 64
            if abs(pre_step / num_steps_per_epoch - pre_epoch) > 0.1:
                fixed_pre_step = int(pre_epoch * num_steps_per_epoch)
                logging.info('Warning, epoch is diff with pre_step / num_steps_per_epoch:{}, pre_epoch:{},maybe you change batch size and we will adjust to set pre_step as {}'\
                  .format(pre_step / num_steps_per_epoch, pre_epoch, fixed_pre_step))
    else:
        print('Train all start step 0', file=sys.stderr)
        #https://stackoverflow.com/questions/40220201/tensorflow-tf-initialize-all-variables-vs-tf-initialize-local-variables
        #tf.initialize_all_variables() is a shortcut to tf.initialize_variables(tf.all_variables()),
        #tf.initialize_local_variables() is a shortcut to tf.initialize_variables(tf.local_variables()),
        #which initializes variables in GraphKeys.VARIABLES and GraphKeys.LOCAL_VARIABLE collections, respectively.
        #init_op = tf.group(tf.global_variables_initializer(),
        #                   tf.local_variables_initializer())
        #[var for var in tf.all_variables() if var.op.name.startswith(restore_scope)] will be the same as tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=restore_scope)

        #sess.run(init_op)

        #like use image model, build image graph, reload first train, and then will go to same checkpoint all varaible just restore will ok
        #for finetune from loading other model init
        if init_fn is not None:
            init_fn(sess)

    if save_interval_epochs and num_steps_per_epoch:
        epoch_dir = os.path.join(model_dir, 'epoch')
        gezi.try_mkdir(epoch_dir)

    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    checkpoint_path = os.path.join(model_dir, 'model.ckpt')

    tf.train.write_graph(sess.graph_def, model_dir, 'train.pbtxt')
    only_one_step = False
    try:
        step = start = pre_step + 1
        fixed_step = fixed_pre_step + 1
        #hack just for save one model after load
        if num_steps < 0 or (num_steps and num_steps < step):
            print('just load and resave then exit', file=sys.stderr)
            saver.save(sess,
                       _get_checkpoint_path(checkpoint_path, step,
                                            num_steps_per_epoch),
                       global_step=step)
            sess.close()
            exit(0)

        if num_epochs < 0:
            only_one_step = True
            print('just run one step', file=sys.stderr)

        early_stop = True  #TODO allow config
        num_bad_epochs = 0
        pre_epoch_eval_loss = 1e20
        best_epoch_eval_loss = 1e20
        num_allowed_bad_epochs = 4  #allow 5 non decrease eval loss epochs  before stop
        while not coord.should_stop():
            stop = train_once_fn(sess,
                                 step,
                                 is_start=(step == start),
                                 fixed_step=fixed_step)
            if only_one_step:
                stop = True
            if save_model and step:
                #step 0 is also saved! actually train one step and save
                if step % save_interval_steps == 0:
                    timer = gezi.Timer('save model step %d to %s' %
                                       (step, checkpoint_path))
                    saver.save(sess,
                               _get_checkpoint_path(checkpoint_path,
                                                    fixed_step,
                                                    num_steps_per_epoch),
                               global_step=step)
                    timer.print()
                #if save_interval_epochs and num_steps_per_epoch and step % (num_steps_per_epoch * save_interval_epochs) == 0:
                #if save_interval_epochs and num_steps_per_epoch and step % num_steps_per_epoch == 0:
                if save_interval_epochs and num_steps_per_epoch and fixed_step % num_steps_per_epoch == 0:
                    #epoch = step // num_steps_per_epoch
                    epoch = fixed_step // num_steps_per_epoch
                    eval_loss = melt.eval_loss()
                    if eval_loss:
                        #['eval_loss:3.2','eal_accuracy:4.3']
                        eval_loss = float(
                            eval_loss.strip('[]').split(',')[0].strip(
                                "'").split(':')[-1])
                        if os.path.exists(
                                os.path.join(epoch_dir, 'best_eval_loss.txt')):
                            with open(
                                    os.path.join(epoch_dir,
                                                 'best_eval_loss.txt')) as f:
                                best_epoch_eval_loss = float(
                                    f.readline().split()[-1].strip())
                        if eval_loss < best_epoch_eval_loss:
                            best_epoch_eval_loss = eval_loss
                            logging.info(
                                'Now best eval loss is epoch %d eval_loss:%f' %
                                (epoch, eval_loss))
                            with open(
                                    os.path.join(epoch_dir,
                                                 'best_eval_loss.txt'),
                                    'w') as f:
                                f.write('%d %d %f\n' %
                                        (epoch, step, best_epoch_eval_loss))
                            best_epoch_saver.save(
                                sess, os.path.join(epoch_dir,
                                                   'model.ckpt-best'))

                        with open(os.path.join(epoch_dir, 'eval_loss.txt'),
                                  'a') as f:
                            f.write('%d %d %f\n' % (epoch, step, eval_loss))
                        if eval_loss >= pre_epoch_eval_loss:
                            num_bad_epochs += 1
                            if num_bad_epochs > num_allowed_bad_epochs:
                                logging.warning(
                                    'Evaluate loss not decrease for last %d epochs'
                                    % (num_allowed_bad_epochs + 1))
                                if not os.path.exists(
                                        os.path.join(epoch_dir,
                                                     'model.ckpt-noimprove')):
                                    best_epoch_saver.save(
                                        sess,
                                        os.path.join(epoch_dir,
                                                     'model.ckpt-noimprove'))
                                ##-------well remove it since
                                #if early_stop:
                                #  stop = True
                        else:
                            num_bad_epochs = 0
                        pre_epoch_eval_loss = eval_loss
                    if step % (num_steps_per_epoch *
                               save_interval_epochs) == 0:
                        epoch_saver.save(sess,
                                         os.path.join(epoch_dir,
                                                      'model.ckpt-%d' % epoch),
                                         global_step=step)
                    #--------do not add step
                    # epoch_saver.save(sess,
                    #        os.path.join(epoch_dir,'model.ckpt-%d'%epoch))
            if stop is True:
                print('Early stop running %d stpes' % (step), file=sys.stderr)
                raise tf.errors.OutOfRangeError(
                    None, None, 'Early stop running %d stpes' % (step))
            if num_steps and (step + 1) == start + num_steps:
                raise tf.errors.OutOfRangeError(None, None,
                                                'Reached max num steps')
            #max_num_epochs = 1000
            max_num_epochs = num_epochs
            if max_num_epochs and num_steps_per_epoch and step // num_steps_per_epoch >= max_num_epochs:
                raise tf.errors.OutOfRangeError(
                    None, None,
                    'Reached max num epochs of %d' % max_num_epochs)
            step += 1
            fixed_step += 1
    except tf.errors.OutOfRangeError, e:
        if not (step
                == start) and save_model and step % save_interval_steps != 0:
            saver.save(sess,
                       _get_checkpoint_path(checkpoint_path, step,
                                            num_steps_per_epoch),
                       global_step=step)
        if only_one_step:
            print('Done one step', file=sys.stderr)
            exit(0)
        if metric_eval_fn is not None:
            metric_eval_fn()
        if (num_epochs and step / num_steps_per_epoch >= num_epochs) or (
                num_steps and (step + 1) == start + num_steps):
            print('Done training for %.3f epochs, %d steps.' %
                  (step / num_steps_per_epoch, step + 1),
                  file=sys.stderr)
            #FIXME becase coord.join seems not work,  RuntimeError: Coordinator stopped with threads still running: Thread-9
            exit(0)
        else:
            print('Should not stop, but stopped at epoch: %.3f' %
                  (step / num_steps_per_epoch),
                  file=sys.stderr)
            print(traceback.format_exc(), file=sys.stderr)
            raise e
예제 #3
0
파일: flow.py 프로젝트: tangqiqi123/hasky
def tf_train_flow(train_once_fn, 
                  model_dir='./model', 
                  max_models_keep=1, 
                  save_interval_seconds=600, 
                  save_interval_steps=1000, 
                  num_epochs=None,
                  num_steps=None, 
                  save_model=True,
                  save_interval_epochs=1, 
                  num_steps_per_epoch=0,
                  restore_from_latest=True,
                  metric_eval_fn=None,
                  init_fn=None,
                  sess=None):
  """
  similary flow as tf_flow, but add model try reload and save
  """
  if sess is None:
    #TODO melt.get_session is global session but may cause
    sess = melt.get_session()
  logging.info('tf_train_flow start')
  print('max_models_keep:', max_models_keep)
  print('save_interval_seconds:', save_interval_seconds)
  
  saver = tf.train.Saver(
    max_to_keep=max_models_keep, 
    keep_checkpoint_every_n_hours=save_interval_seconds / 3600.0)
  
  epoch_saver = tf.train.Saver()
  best_epoch_saver = tf.train.Saver() 
  
  #pre_step means the step last saved, train without pretrained,then -1
  pre_step = -1;
  model_path = _get_model_path(model_dir, save_model)
  model_dir = gezi.get_dir(model_dir) #incase you pass ./model/model-ckpt1000 -> ./model
  if model_path is not None:
    if not restore_from_latest:
      print('using recent but not latest model', file=sys.stderr)
      model_path = melt.recent_checkpoint(model_dir)
    model_name = os.path.basename(model_path)
    timer = gezi.Timer('Loading and training from existing model [%s]'%model_path)
    saver.restore(sess, model_path)
    timer.print()
    pre_step = melt.get_model_step(model_path)
    if 'epoch' in model_name:
      pre_step *= num_steps_per_epoch
    #for non 0 eopochs  without this will be
    #Attempting to use uninitialized value input/input_producer/limit_epochs/epochs
    sess.run(tf.local_variables_initializer())
  else:
    print('Train all start step 0', file=sys.stderr)
    init_op = tf.group(tf.global_variables_initializer(),
                       tf.local_variables_initializer())
    sess.run(init_op)

    if init_fn is not None:
      init_fn(sess)
  
  if save_interval_epochs and num_steps_per_epoch:
    epoch_dir = os.path.join(model_dir, 'epoch')
    gezi.try_mkdir(epoch_dir)
  
  coord = tf.train.Coordinator()
  threads = tf.train.start_queue_runners(sess=sess, coord=coord)
  checkpoint_path = os.path.join(model_dir, 'model.ckpt')

  tf.train.write_graph(sess.graph_def, model_dir, 'train.pbtxt')
  try:
    step = start = pre_step +  1
    #hack just for save one model after load
    if num_steps and num_steps < step:
      print('just load and resave then exit', file=sys.stderr)
      saver.save(sess, 
                 _get_checkpoint_path(checkpoint_path, step, num_steps_per_epoch), 
                 global_step=step)
      sess.close()
      exit(0)

    early_stop = True #TODO allow config
    num_bad_epochs = 0
    pre_epoch_eval_loss = 1e20
    best_epoch_eval_loss = 1e20
    num_allowed_bad_epochs = 4 #allow 5 non decrease eval loss epochs  before stop
    while not coord.should_stop():
      stop = train_once_fn(sess, step, is_start=(step==start))
      if save_model and step:
        #step 0 is also saved! actually train one step and save
        if step % save_interval_steps == 0:
          timer = gezi.Timer('save model step %d to %s'%(step, checkpoint_path))
          saver.save(sess, 
                     _get_checkpoint_path(checkpoint_path, step, num_steps_per_epoch), 
                     global_step=step)
          timer.print()
        #if save_interval_epochs and num_steps_per_epoch and step % (num_steps_per_epoch * save_interval_epochs) == 0:
        if save_interval_epochs and num_steps_per_epoch and step % num_steps_per_epoch == 0:
          epoch = step // num_steps_per_epoch
          eval_loss = melt.eval_loss()
          if eval_loss:
            #['eval_loss:3.2','eal_accuracy:4.3']
            eval_loss = float(eval_loss.strip('[]').split(',')[0].strip("'").split(':')[-1])
            if os.path.exists(os.path.join(epoch_dir, 'best_eval_loss.txt')):
              with open(os.path.join(epoch_dir, 'best_eval_loss.txt')) as f:
                best_epoch_eval_loss = float(f.readline().split()[-1].strip())
            if eval_loss < best_epoch_eval_loss:
              best_epoch_eval_loss = eval_loss
              logging.info('Now best eval loss is epoch %d eval_loss:%f' % (epoch, eval_loss))
              with open(os.path.join(epoch_dir, 'best_eval_loss.txt'), 'w') as f:
                f.write('%d %d %f\n'%(epoch, step, best_epoch_eval_loss))
              best_epoch_saver.save(sess, 
                                    os.path.join(epoch_dir,'model.cpkt-best'))

            with open(os.path.join(epoch_dir, 'eval_loss.txt'), 'a') as f:
               f.write('%d %d %f\n'%(epoch, step, eval_loss))
            if eval_loss >= pre_epoch_eval_loss:
              num_bad_epochs += 1
              if num_bad_epochs > num_allowed_bad_epochs:
                logging.warning('Evaluate loss not decrease for last %d epochs'% (num_allowed_bad_epochs + 1))
                if not os.path.exists(os.path.join(epoch_dir,'model.cpkt-noimprove')):
                  best_epoch_saver.save(sess, os.path.join(epoch_dir,'model.cpkt-noimprove'))
                ##-------well remove it since 
                #if early_stop:
                #  stop = True 
            else:
              num_bad_epochs = 0
            pre_epoch_eval_loss = eval_loss
          if step % (num_steps_per_epoch * save_interval_epochs) == 0:
            epoch_saver.save(sess, 
                            os.path.join(epoch_dir,'model.cpkt-%d'%epoch), 
                            global_step=step)
          #--------do not add step
          # epoch_saver.save(sess, 
          #        os.path.join(epoch_dir,'model.cpkt-%d'%epoch))
      if stop is True:
        print('Early stop running %d stpes'%(step), file=sys.stderr)
        raise tf.errors.OutOfRangeError(None, None,'Early stop running %d stpes'%(step))
      if num_steps and (step + 1) == start + num_steps:
        raise tf.errors.OutOfRangeError(None, None,'Reached max num steps')
      #max_num_epochs = 1000
      max_num_epochs = num_epochs
      if num_steps_per_epoch and step // num_steps_per_epoch >= max_num_epochs:
        raise tf.errors.OutOfRangeError(None, None,'Reached max num epochs of %d'%max_num_epochs)
      step += 1
  except tf.errors.OutOfRangeError, e:
    if not (step==start) and save_model and step % save_interval_steps != 0:
      saver.save(sess, 
                 _get_checkpoint_path(checkpoint_path, step, num_steps_per_epoch), 
                 global_step=step)
    if metric_eval_fn is not None:
      metric_eval_fn()
    if (num_epochs and step / num_steps_per_epoch >= num_epochs) or (num_steps and (step + 1) == start + num_steps) :
      print('Done training for %.3f epochs, %d steps.' % (step / num_steps_per_epoch, step + 1), file=sys.stderr)
      #FIXME becase coord.join seems not work,  RuntimeError: Coordinator stopped with threads still running: Thread-9
      exit(0)
    else:
      print('Should not stop, but stopped at epoch: %.3f'%(step / num_steps_per_epoch), file=sys.stderr)
      print(traceback.format_exc(), file=sys.stderr)
      raise e