TensorFlow 训练模型保存四个文件
if mAP > best_mAP:
best_mAP = mAP
saver_best.save(sess, args.save_dir + 'best_model_Epoch_{}_step_{}_mAP_{:.4f}_loss_{:.4f}_lr_{:.7g}'.format(
epoch, int(__global_step), best_mAP, val_loss_total.average, __lr)) 如上部分代码，当训练结果比预设的best_mAP好，则保存此时的TensorFlow 训练模型,此时，在第几个epoch，第几个global_step,最好的best_mAP值，平均验证损失val_loss_total.average，lr。
tf.train.Saver().save(sess, 'ckpts/')在ckpts/ 路径下主要保存四个文件checkpoint：
checkpoint:model.ckpt.data-00000-of-00001: 某个ckpt的数据文件，保存每个变量的取值，保存的是网络的权值，偏置，操作等等。
model.ckpt.index ：某个ckpt的index文件 二进制 或者其他格式 不可直接查看 。是一个不可变得字符串表，每一个键都是张量的名称，它的值是一个序列化的BundleEntryProto。 每个BundleEntryProto描述张量的元数据：“数据”文件中的哪个文件包含张量的内容，该文件的偏移量，校验和，一些辅助数据等等。
model.ckpt.meta：某个ckpt的meta数据 二进制 或者其他格式 不可直接查看，保存了TensorFlow计算图的结构信息。model.ckpt-200.meta文件保存的是图结构，通俗地讲就是神经网络的网络结构。一般而言网络结构是不会发生改变，所以可以只保存一个就行了。我们可以使用下面的代码只在第一次保存meta文件。
checkpoint：记录训练较好的几次训练结果
本人训练后的checkpoint文件内容，如下（Epoch 26,30,34,94,98,）：
model_checkpoint_path: "best_model_Epoch_98_step_50291_mAP_0.7358_loss_3.6063_lr_1e-05"
all_model_checkpoint_paths: "best_model_Epoch_26_step_13715_mAP_0.7163_loss_3.3344_lr_0.0001"
all_model_checkpoint_paths: "best_model_Epoch_30_step_15747_mAP_0.7211_loss_3.3941_lr_0.0001"
all_model_checkpoint_paths: "best_model_Epoch_34_step_17779_mAP_0.7317_loss_3.3543_lr_3e-05"
all_model_checkpoint_paths: "best_model_Epoch_94_step_48259_mAP_0.7328_loss_3.5932_lr_1e-05"
all_model_checkpoint_paths: "best_model_Epoch_98_step_50291_mAP_0.7358_loss_3.6063_lr_1e-05"模型加载需要利用Saver.restore方法。可以加载固定参数，也可以加在所有参数。
tf.train.Saver.restore(sess,model_path)训练过程保存了大量tensorflow模型 :
-4.2$ ls
best_model_Epoch_10_step_5664_mAP_0.0729_loss_8.1562_lr_0.0001.data-00000-of-00001
best_model_Epoch_10_step_5664_mAP_0.0729_loss_8.1562_lr_0.0001.index
best_model_Epoch_10_step_5664_mAP_0.0729_loss_8.1562_lr_0.0001.meta
best_model_Epoch_12_step_6694_mAP_0.0748_loss_7.4896_lr_0.0001.data-00000-of-00001
best_model_Epoch_12_step_6694_mAP_0.0748_loss_7.4896_lr_0.0001.index
best_model_Epoch_12_step_6694_mAP_0.0748_loss_7.4896_lr_0.0001.meta
best_model_Epoch_16_step_8754_mAP_0.0775_loss_7.6105_lr_0.0001.data-00000-of-00001
best_model_Epoch_16_step_8754_mAP_0.0775_loss_7.6105_lr_0.0001.index
best_model_Epoch_16_step_8754_mAP_0.0775_loss_7.6105_lr_0.0001.meta
best_model_Epoch_20_step_10814_mAP_0.0715_loss_8.2555_lr_0.0001.data-00000-of-00001
best_model_Epoch_20_step_10814_mAP_0.0715_loss_8.2555_lr_0.0001.index
best_model_Epoch_20_step_10814_mAP_0.0715_loss_8.2555_lr_0.0001.meta
best_model_Epoch_24_step_12874_mAP_0.0717_loss_8.5067_lr_0.0001.data-00000-of-00001
best_model_Epoch_24_step_12874_mAP_0.0717_loss_8.5067_lr_0.0001.index
best_model_Epoch_24_step_12874_mAP_0.0717_loss_8.5067_lr_0.0001.meta
best_model_Epoch_26_step_13715_mAP_0.7163_loss_3.3344_lr_0.0001.data-00000-of-00001
best_model_Epoch_26_step_13715_mAP_0.7163_loss_3.3344_lr_0.0001.index
best_model_Epoch_26_step_13715_mAP_0.7163_loss_3.3344_lr_0.0001.meta
best_model_Epoch_26_step_13904_mAP_0.0726_loss_8.5469_lr_0.0001.data-00000-of-00001
best_model_Epoch_26_step_13904_mAP_0.0726_loss_8.5469_lr_0.0001.index
best_model_Epoch_26_step_13904_mAP_0.0726_loss_8.5469_lr_0.0001.meta
best_model_Epoch_26_step_13904_mAP_0.0788_loss_7.9227_lr_0.0001.data-00000-of-00001
best_model_Epoch_26_step_13904_mAP_0.0788_loss_7.9227_lr_0.0001.index
best_model_Epoch_26_step_13904_mAP_0.0788_loss_7.9227_lr_0.0001.meta
best_model_Epoch_28_step_14731_mAP_0.7189_loss_3.2645_lr_0.0001.data-00000-of-00001
best_model_Epoch_28_step_14731_mAP_0.7189_loss_3.2645_lr_0.0001.index
best_model_Epoch_28_step_14731_mAP_0.7189_loss_3.2645_lr_0.0001.meta
best_model_Epoch_30_step_15747_mAP_0.7211_loss_3.3941_lr_0.0001.data-00000-of-00001
best_model_Epoch_30_step_15747_mAP_0.7211_loss_3.3941_lr_0.0001.index
best_model_Epoch_30_step_15747_mAP_0.7211_loss_3.3941_lr_0.0001.meta
best_model_Epoch_30_step_15747_mAP_0.7310_loss_3.4142_lr_0.0001.data-00000-of-00001
best_model_Epoch_30_step_15747_mAP_0.7310_loss_3.4142_lr_0.0001.index
best_model_Epoch_30_step_15747_mAP_0.7310_loss_3.4142_lr_0.0001.meta
best_model_Epoch_30_step_15964_mAP_0.0797_loss_8.0055_lr_0.0001.data-00000-of-00001
best_model_Epoch_30_step_15964_mAP_0.0797_loss_8.0055_lr_0.0001.index
best_model_Epoch_30_step_15964_mAP_0.0797_loss_8.0055_lr_0.0001.meta
best_model_Epoch_32_step_16763_mAP_0.7343_loss_3.3788_lr_0.0001.data-00000-of-00001
best_model_Epoch_32_step_16763_mAP_0.7343_loss_3.3788_lr_0.0001.index
best_model_Epoch_32_step_16763_mAP_0.7343_loss_3.3788_lr_0.0001.meta
best_model_Epoch_34_step_17779_mAP_0.7317_loss_3.3543_lr_3e-05.data-00000-of-00001
best_model_Epoch_34_step_17779_mAP_0.7317_loss_3.3543_lr_3e-05.index
best_model_Epoch_34_step_17779_mAP_0.7317_loss_3.3543_lr_3e-05.meta
best_model_Epoch_34_step_17779_mAP_0.7419_loss_3.2561_lr_3e-05.data-00000-of-00001
best_model_Epoch_34_step_17779_mAP_0.7419_loss_3.2561_lr_3e-05.index
best_model_Epoch_34_step_17779_mAP_0.7419_loss_3.2561_lr_3e-05.meta
best_model_Epoch_36_step_18795_mAP_0.7390_loss_3.3124_lr_3e-05.data-00000-of-00001
best_model_Epoch_36_step_18795_mAP_0.7390_loss_3.3124_lr_3e-05.index
best_model_Epoch_36_step_18795_mAP_0.7390_loss_3.3124_lr_3e-05.meta
best_model_Epoch_38_step_19811_mAP_0.7427_loss_3.3455_lr_3e-05.data-00000-of-00001
best_model_Epoch_38_step_19811_mAP_0.7427_loss_3.3455_lr_3e-05.index
best_model_Epoch_38_step_19811_mAP_0.7427_loss_3.3455_lr_3e-05.meta
best_model_Epoch_38_step_19811_mAP_0.7428_loss_3.3389_lr_3e-05.data-00000-of-00001
best_model_Epoch_38_step_19811_mAP_0.7428_loss_3.3389_lr_3e-05.index
best_model_Epoch_38_step_19811_mAP_0.7428_loss_3.3389_lr_3e-05.meta
best_model_Epoch_48_step_24891_mAP_0.7435_loss_3.4464_lr_3e-05.data-00000-of-00001
best_model_Epoch_48_step_24891_mAP_0.7435_loss_3.4464_lr_3e-05.index
best_model_Epoch_48_step_24891_mAP_0.7435_loss_3.4464_lr_3e-05.meta
best_model_Epoch_4_step_2574_mAP_0.0042_loss_9.6720_lr_0.0001.data-00000-of-00001
best_model_Epoch_4_step_2574_mAP_0.0042_loss_9.6720_lr_0.0001.index
best_model_Epoch_4_step_2574_mAP_0.0042_loss_9.6720_lr_0.0001.meta
best_model_Epoch_4_step_2574_mAP_0.0428_loss_7.8114_lr_0.0001.data-00000-of-00001
best_model_Epoch_4_step_2574_mAP_0.0428_loss_7.8114_lr_0.0001.index
best_model_Epoch_4_step_2574_mAP_0.0428_loss_7.8114_lr_0.0001.meta
best_model_Epoch_4_step_2574_mAP_0.0484_loss_7.5300_lr_0.0001.data-00000-of-00001
best_model_Epoch_4_step_2574_mAP_0.0484_loss_7.5300_lr_0.0001.index
best_model_Epoch_4_step_2574_mAP_0.0484_loss_7.5300_lr_0.0001.meta
best_model_Epoch_4_step_2574_mAP_0.0492_loss_7.8293_lr_0.0001.data-00000-of-00001
best_model_Epoch_4_step_2574_mAP_0.0492_loss_7.8293_lr_0.0001.index
best_model_Epoch_4_step_2574_mAP_0.0492_loss_7.8293_lr_0.0001.meta
best_model_Epoch_52_step_26923_mAP_0.7448_loss_3.5138_lr_3e-05.data-00000-of-00001
best_model_Epoch_52_step_26923_mAP_0.7448_loss_3.5138_lr_3e-05.index
best_model_Epoch_52_step_26923_mAP_0.7448_loss_3.5138_lr_3e-05.meta
best_model_Epoch_58_step_30384_mAP_0.0728_loss_9.0220_lr_1e-05.data-00000-of-00001
best_model_Epoch_58_step_30384_mAP_0.0728_loss_9.0220_lr_1e-05.index
best_model_Epoch_58_step_30384_mAP_0.0728_loss_9.0220_lr_1e-05.meta
best_model_Epoch_6_step_3604_mAP_0.0531_loss_8.0416_lr_0.0001.data-00000-of-00001
best_model_Epoch_6_step_3604_mAP_0.0531_loss_8.0416_lr_0.0001.index
best_model_Epoch_6_step_3604_mAP_0.0531_loss_8.0416_lr_0.0001.meta
best_model_Epoch_6_step_3604_mAP_0.0633_loss_7.4694_lr_0.0001.data-00000-of-00001
best_model_Epoch_6_step_3604_mAP_0.0633_loss_7.4694_lr_0.0001.index
best_model_Epoch_6_step_3604_mAP_0.0633_loss_7.4694_lr_0.0001.meta
best_model_Epoch_74_step_38624_mAP_0.0731_loss_9.1056_lr_1e-05.data-00000-of-00001
best_model_Epoch_74_step_38624_mAP_0.0731_loss_9.1056_lr_1e-05.index
best_model_Epoch_74_step_38624_mAP_0.0731_loss_9.1056_lr_1e-05.meta
best_model_Epoch_80_step_41147_mAP_0.7455_loss_3.5324_lr_1e-05.data-00000-of-00001
best_model_Epoch_80_step_41147_mAP_0.7455_loss_3.5324_lr_1e-05.index
best_model_Epoch_80_step_41147_mAP_0.7455_loss_3.5324_lr_1e-05.meta
best_model_Epoch_8_step_4634_mAP_0.0639_loss_7.5254_lr_0.0001.data-00000-of-00001
best_model_Epoch_8_step_4634_mAP_0.0639_loss_7.5254_lr_0.0001.index
best_model_Epoch_8_step_4634_mAP_0.0639_loss_7.5254_lr_0.0001.meta
best_model_Epoch_8_step_4634_mAP_0.0710_loss_7.8850_lr_0.0001.data-00000-of-00001
best_model_Epoch_8_step_4634_mAP_0.0710_loss_7.8850_lr_0.0001.index
best_model_Epoch_8_step_4634_mAP_0.0710_loss_7.8850_lr_0.0001.meta
best_model_Epoch_8_step_4634_mAP_0.0726_loss_7.0773_lr_0.0001.data-00000-of-00001
best_model_Epoch_8_step_4634_mAP_0.0726_loss_7.0773_lr_0.0001.index
best_model_Epoch_8_step_4634_mAP_0.0726_loss_7.0773_lr_0.0001.meta
best_model_Epoch_94_step_48259_mAP_0.7328_loss_3.5932_lr_1e-05.data-00000-of-00001
best_model_Epoch_94_step_48259_mAP_0.7328_loss_3.5932_lr_1e-05.index
best_model_Epoch_94_step_48259_mAP_0.7328_loss_3.5932_lr_1e-05.meta
best_model_Epoch_98_step_50291_mAP_0.7358_loss_3.6063_lr_1e-05.data-00000-of-00001
best_model_Epoch_98_step_50291_mAP_0.7358_loss_3.6063_lr_1e-05.index
best_model_Epoch_98_step_50291_mAP_0.7358_loss_3.6063_lr_1e-05.meta
checkpoint
model-epoch_10_step_5664_loss_1.1807_lr_0.0001.data-00000-of-00001
model-epoch_10_step_5664_loss_1.1807_lr_0.0001.index
model-epoch_10_step_5664_loss_1.1807_lr_0.0001.meta
model-epoch_20_step_10667_loss_0.8032_lr_0.0001.data-00000-of-00001
model-epoch_20_step_10667_loss_0.8032_lr_0.0001.index
model-epoch_20_step_10667_loss_0.8032_lr_0.0001.meta
model-epoch_30_step_15747_loss_0.5327_lr_0.0001.data-00000-of-00001
model-epoch_30_step_15747_loss_0.5327_lr_0.0001.index
model-epoch_30_step_15747_loss_0.5327_lr_0.0001.meta
model-epoch_40_step_20827_loss_0.3800_lr_3e-05.data-00000-of-00001
model-epoch_40_step_20827_loss_0.3800_lr_3e-05.index
model-epoch_40_step_20827_loss_0.3800_lr_3e-05.meta
model-epoch_50_step_25907_loss_0.3512_lr_3e-05.data-00000-of-00001
model-epoch_50_step_25907_loss_0.3512_lr_3e-05.index
model-epoch_50_step_25907_loss_0.3512_lr_3e-05.meta
model-epoch_50_step_25907_loss_0.3513_lr_3e-05.data-00000-of-00001
model-epoch_50_step_25907_loss_0.3513_lr_3e-05.index
model-epoch_50_step_25907_loss_0.3513_lr_3e-05.meta
model-epoch_50_step_25907_loss_0.3590_lr_3e-05.data-00000-of-00001
model-epoch_50_step_25907_loss_0.3590_lr_3e-05.index
model-epoch_50_step_25907_loss_0.3590_lr_3e-05.meta
model-epoch_50_step_26264_loss_0.3299_lr_3e-05.data-00000-of-00001
model-epoch_50_step_26264_loss_0.3299_lr_3e-05.index
model-epoch_50_step_26264_loss_0.3299_lr_3e-05.meta
model-epoch_50_step_26264_loss_0.3480_lr_3e-05.data-00000-of-00001
model-epoch_50_step_26264_loss_0.3480_lr_3e-05.index
model-epoch_50_step_26264_loss_0.3480_lr_3e-05.meta
model-epoch_60_step_30987_loss_0.3373_lr_1e-05.data-00000-of-00001
model-epoch_60_step_30987_loss_0.3373_lr_1e-05.index
model-epoch_60_step_30987_loss_0.3373_lr_1e-05.meta
model-epoch_60_step_30987_loss_0.3422_lr_1e-05.data-00000-of-00001
model-epoch_60_step_30987_loss_0.3422_lr_1e-05.index
model-epoch_60_step_30987_loss_0.3422_lr_1e-05.meta
model-epoch_60_step_30987_loss_0.3430_lr_1e-05.data-00000-of-00001
model-epoch_60_step_30987_loss_0.3430_lr_1e-05.index
model-epoch_60_step_30987_loss_0.3430_lr_1e-05.meta
model-epoch_60_step_31414_loss_0.3199_lr_1e-05.data-00000-of-00001
model-epoch_60_step_31414_loss_0.3199_lr_1e-05.index
model-epoch_60_step_31414_loss_0.3199_lr_1e-05.meta
model-epoch_60_step_31414_loss_0.3371_lr_1e-05.data-00000-of-00001
model-epoch_60_step_31414_loss_0.3371_lr_1e-05.index
model-epoch_60_step_31414_loss_0.3371_lr_1e-05.meta
model-epoch_70_step_36067_loss_0.3288_lr_1e-05.data-00000-of-00001
model-epoch_70_step_36067_loss_0.3288_lr_1e-05.index
model-epoch_70_step_36067_loss_0.3288_lr_1e-05.meta
model-epoch_70_step_36067_loss_0.3350_lr_1e-05.data-00000-of-00001
model-epoch_70_step_36067_loss_0.3350_lr_1e-05.index
model-epoch_70_step_36067_loss_0.3350_lr_1e-05.meta
model-epoch_70_step_36564_loss_0.3164_lr_1e-05.data-00000-of-00001
model-epoch_70_step_36564_loss_0.3164_lr_1e-05.index
model-epoch_70_step_36564_loss_0.3164_lr_1e-05.meta
model-epoch_70_step_36564_loss_0.3291_lr_1e-05.data-00000-of-00001
model-epoch_70_step_36564_loss_0.3291_lr_1e-05.index
model-epoch_70_step_36564_loss_0.3291_lr_1e-05.meta
model-epoch_80_step_41147_loss_0.3309_lr_1e-05.data-00000-of-00001
model-epoch_80_step_41147_loss_0.3309_lr_1e-05.index
model-epoch_80_step_41147_loss_0.3309_lr_1e-05.meta
model-epoch_80_step_41147_loss_0.3321_lr_1e-05.data-00000-of-00001
model-epoch_80_step_41147_loss_0.3321_lr_1e-05.index
model-epoch_80_step_41147_loss_0.3321_lr_1e-05.meta
model-epoch_80_step_41714_loss_0.3133_lr_1e-05.data-00000-of-00001
model-epoch_80_step_41714_loss_0.3133_lr_1e-05.index
model-epoch_80_step_41714_loss_0.3133_lr_1e-05.meta
model-epoch_80_step_41714_loss_0.3244_lr_1e-05.data-00000-of-00001
model-epoch_80_step_41714_loss_0.3244_lr_1e-05.index
model-epoch_80_step_41714_loss_0.3244_lr_1e-05.meta
model-epoch_90_step_46227_loss_0.3235_lr_1e-05.data-00000-of-00001
model-epoch_90_step_46227_loss_0.3235_lr_1e-05.index
model-epoch_90_step_46227_loss_0.3235_lr_1e-05.meta
model-epoch_90_step_46227_loss_0.3270_lr_1e-05.data-00000-of-00001
model-epoch_90_step_46227_loss_0.3270_lr_1e-05.index
model-epoch_90_step_46227_loss_0.3270_lr_1e-05.meta
model-epoch_90_step_46864_loss_0.3098_lr_1e-05.data-00000-of-00001
model-epoch_90_step_46864_loss_0.3098_lr_1e-05.index
model-epoch_90_step_46864_loss_0.3098_lr_1e-05.meta
model-epoch_90_step_46864_loss_0.3239_lr_1e-05.data-00000-of-00001
model-epoch_90_step_46864_loss_0.3239_lr_1e-05.index
model-epoch_90_step_46864_loss_0.3239_lr_1e-05.meta
sh-4.2$如下为 yolov3 train.py:
# coding: utf-8
from __future__ import division, print_function
import tensorflow as tf
import numpy as np
import logging
from tqdm import trange
import args
from utils.data_utils import get_batch_data
from utils.misc_utils import shuffle_and_overwrite, make_summary, config_learning_rate, config_optimizer, AverageMeter
from utils.eval_utils import evaluate_on_cpu, evaluate_on_gpu, get_preds_gpu, voc_eval, parse_gt_rec
from utils.nms_utils import gpu_nms
from model import yolov3
# setting loggers
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s %(levelname)s %(message)s',
datefmt='%a, %d %b %Y %H:%M:%S', filename=args.progress_log_path, filemode='w')
# setting placeholders
is_training = tf.placeholder(tf.bool, name="phase_train")
handle_flag = tf.placeholder(tf.string, [], name='iterator_handle_flag')
# register the gpu nms operation here for the following evaluation scheme
pred_boxes_flag = tf.placeholder(tf.float32, [1, None, None])
pred_scores_flag = tf.placeholder(tf.float32, [1, None, None])
gpu_nms_op = gpu_nms(pred_boxes_flag, pred_scores_flag, args.class_num, args.nms_topk, args.score_threshold, args.nms_threshold)
##################
# tf.data pipeline
##################
train_dataset = tf.data.TextLineDataset(args.train_file)
train_dataset = train_dataset.shuffle(args.train_img_cnt)
train_dataset = train_dataset.batch(args.batch_size)
train_dataset = train_dataset.map(
lambda x: tf.py_func(get_batch_data,
inp=[x, args.class_num, args.img_size, args.anchors, 'train', args.multi_scale_train, args.use_mix_up, args.letterbox_resize],
Tout=[tf.int64, tf.float32, tf.float32, tf.float32, tf.float32]),
num_parallel_calls=args.num_threads
)
train_dataset = train_dataset.prefetch(args.prefetech_buffer)
val_dataset = tf.data.TextLineDataset(args.val_file)
val_dataset = val_dataset.batch(1)
val_dataset = val_dataset.map(
lambda x: tf.py_func(get_batch_data,
inp=[x, args.class_num, args.img_size, args.anchors, 'val', False, False, args.letterbox_resize],
Tout=[tf.int64, tf.float32, tf.float32, tf.float32, tf.float32]),
num_parallel_calls=args.num_threads
)
val_dataset.prefetch(args.prefetech_buffer)
iterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)
train_init_op = iterator.make_initializer(train_dataset)
val_init_op = iterator.make_initializer(val_dataset)
# get an element from the chosen dataset iterator
image_ids, image, y_true_13, y_true_26, y_true_52 = iterator.get_next()
y_true = [y_true_13, y_true_26, y_true_52]
# tf.data pipeline will lose the data `static` shape, so we need to set it manually
image_ids.set_shape([None])
image.set_shape([None, None, None, 3])
for y in y_true:
y.set_shape([None, None, None, None, None])
##################
# Model definition
##################
yolo_model = yolov3(args.class_num, args.anchors, args.use_label_smooth, args.use_focal_loss, args.batch_norm_decay, args.weight_decay, use_static_shape=False)
with tf.variable_scope('yolov3'):
pred_feature_maps = yolo_model.forward(image, is_training=is_training)
loss = yolo_model.compute_loss(pred_feature_maps, y_true)
y_pred = yolo_model.predict(pred_feature_maps)
l2_loss = tf.losses.get_regularization_loss()
# setting restore parts and vars to update
saver_to_restore = tf.train.Saver(var_list=tf.contrib.framework.get_variables_to_restore(include=args.restore_include, exclude=args.restore_exclude))
update_vars = tf.contrib.framework.get_variables_to_restore(include=args.update_part)
tf.summary.scalar('train_batch_statistics/total_loss', loss[0])
tf.summary.scalar('train_batch_statistics/loss_xy', loss[1])
tf.summary.scalar('train_batch_statistics/loss_wh', loss[2])
tf.summary.scalar('train_batch_statistics/loss_conf', loss[3])
tf.summary.scalar('train_batch_statistics/loss_class', loss[4])
tf.summary.scalar('train_batch_statistics/loss_l2', l2_loss)
tf.summary.scalar('train_batch_statistics/loss_ratio', l2_loss / loss[0])
global_step = tf.Variable(float(args.global_step), trainable=False, collections=[tf.GraphKeys.LOCAL_VARIABLES])
if args.use_warm_up:
learning_rate = tf.cond(tf.less(global_step, args.train_batch_num * args.warm_up_epoch),
lambda: args.learning_rate_init * global_step / (args.train_batch_num * args.warm_up_epoch),
lambda: config_learning_rate(args, global_step - args.train_batch_num * args.warm_up_epoch))
else:
learning_rate = config_learning_rate(args, global_step)
tf.summary.scalar('learning_rate', learning_rate)
if not args.save_optimizer:
saver_to_save = tf.train.Saver()
saver_best = tf.train.Saver()
optimizer = config_optimizer(args.optimizer_name, learning_rate)
# set dependencies for BN ops
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
# train_op = optimizer.minimize(loss[0] + l2_loss, var_list=update_vars, global_step=global_step)
# apply gradient clip to avoid gradient exploding
gvs = optimizer.compute_gradients(loss[0] + l2_loss, var_list=update_vars)
clip_grad_var = [gv if gv[0] is None else [
tf.clip_by_norm(gv[0], 100.), gv[1]] for gv in gvs]
train_op = optimizer.apply_gradients(clip_grad_var, global_step=global_step)
if args.save_optimizer:
print('Saving optimizer parameters to checkpoint! Remember to restore the global_step in the fine-tuning afterwards.')
saver_to_save = tf.train.Saver()
saver_best = tf.train.Saver()
with tf.Session() as sess:
sess.run([tf.global_variables_initializer(), tf.local_variables_initializer()])
saver_to_restore.restore(sess, args.restore_path)
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter(args.log_dir, sess.graph)
print('\n----------- start to train -----------\n')
best_mAP = -np.Inf
for epoch in range(args.total_epoches):
sess.run(train_init_op)
loss_total, loss_xy, loss_wh, loss_conf, loss_class = AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
for i in trange(args.train_batch_num):
_, summary, __y_pred, __y_true, __loss, __global_step, __lr = sess.run(
[train_op, merged, y_pred, y_true, loss, global_step, learning_rate],
feed_dict={is_training: True})
writer.add_summary(summary, global_step=__global_step)
loss_total.update(__loss[0], len(__y_pred[0]))
loss_xy.update(__loss[1], len(__y_pred[0]))
loss_wh.update(__loss[2], len(__y_pred[0]))
loss_conf.update(__loss[3], len(__y_pred[0]))
loss_class.update(__loss[4], len(__y_pred[0]))
if __global_step % args.train_evaluation_step == 0 and __global_step > 0:
# recall, precision = evaluate_on_cpu(__y_pred, __y_true, args.class_num, args.nms_topk, args.score_threshold, args.nms_threshold)
recall, precision = evaluate_on_gpu(sess, gpu_nms_op, pred_boxes_flag, pred_scores_flag, __y_pred, __y_true, args.class_num, args.nms_threshold)
info = "Epoch: {}, global_step: {} | loss: total: {:.2f}, xy: {:.2f}, wh: {:.2f}, conf: {:.2f}, class: {:.2f} | ".format(
epoch, int(__global_step), loss_total.average, loss_xy.average, loss_wh.average, loss_conf.average, loss_class.average)
info += 'Last batch: rec: {:.3f}, prec: {:.3f} | lr: {:.5g}'.format(recall, precision, __lr)
print(info)
logging.info(info)
writer.add_summary(make_summary('evaluation/train_batch_recall', recall), global_step=__global_step)
writer.add_summary(make_summary('evaluation/train_batch_precision', precision), global_step=__global_step)
if np.isnan(loss_total.average):
print('****' * 10)
raise ArithmeticError(
'Gradient exploded! Please train again and you may need modify some parameters.')
# NOTE: this is just demo. You can set the conditions when to save the weights.
if epoch % args.save_epoch == 0 and epoch > 0:
if loss_total.average <= 2.:
saver_to_save.save(sess, args.save_dir + 'model-epoch_{}_step_{}_loss_{:.4f}_lr_{:.5g}'.format(epoch, int(__global_step), loss_total.average, __lr))
# switch to validation dataset for evaluation
if epoch % args.val_evaluation_epoch == 0 and epoch >= args.warm_up_epoch:
sess.run(val_init_op)
val_loss_total, val_loss_xy, val_loss_wh, val_loss_conf, val_loss_class = \
AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
val_preds = []
for j in trange(args.val_img_cnt):
__image_ids, __y_pred, __loss = sess.run([image_ids, y_pred, loss],
feed_dict={is_training: False})
pred_content = get_preds_gpu(sess, gpu_nms_op, pred_boxes_flag, pred_scores_flag, __image_ids, __y_pred)
val_preds.extend(pred_content)
val_loss_total.update(__loss[0])
val_loss_xy.update(__loss[1])
val_loss_wh.update(__loss[2])
val_loss_conf.update(__loss[3])
val_loss_class.update(__loss[4])
# calc mAP
rec_total, prec_total, ap_total = AverageMeter(), AverageMeter(), AverageMeter()
gt_dict = parse_gt_rec(args.val_file, args.img_size, args.letterbox_resize)
info = '======> Epoch: {}, global_step: {}, lr: {:.6g} <======\n'.format(epoch, __global_step, __lr)
for ii in range(args.class_num):
npos, nd, rec, prec, ap = voc_eval(gt_dict, val_preds, ii, iou_thres=args.eval_threshold, use_07_metric=args.use_voc_07_metric)
info += 'EVAL: Class {}: Recall: {:.4f}, Precision: {:.4f}, AP: {:.4f}\n'.format(ii, rec, prec, ap)
rec_total.update(rec, npos)
prec_total.update(prec, nd)
ap_total.update(ap, 1)
mAP = ap_total.average
info += 'EVAL: Recall: {:.4f}, Precison: {:.4f}, mAP: {:.4f}\n'.format(rec_total.average, prec_total.average, mAP)
info += 'EVAL: loss: total: {:.2f}, xy: {:.2f}, wh: {:.2f}, conf: {:.2f}, class: {:.2f}\n'.format(
val_loss_total.average, val_loss_xy.average, val_loss_wh.average, val_loss_conf.average, val_loss_class.average)
print(info)
logging.info(info)
if mAP > best_mAP:
best_mAP = mAP
saver_best.save(sess, args.save_dir + 'best_model_Epoch_{}_step_{}_mAP_{:.4f}_loss_{:.4f}_lr_{:.7g}'.format(
epoch, int(__global_step), best_mAP, val_loss_total.average, __lr))
writer.add_summary(make_summary('evaluation/val_mAP', mAP), global_step=epoch)
writer.add_summary(make_summary('evaluation/val_recall', rec_total.average), global_step=epoch)
writer.add_summary(make_summary('evaluation/val_precision', prec_total.average), global_step=epoch)
writer.add_summary(make_summary('validation_statistics/total_loss', val_loss_total.average), global_step=epoch)
writer.add_summary(make_summary('validation_statistics/loss_xy', val_loss_xy.average), global_step=epoch)
writer.add_summary(make_summary('validation_statistics/loss_wh', val_loss_wh.average), global_step=epoch)
writer.add_summary(make_summary('validation_statistics/loss_conf', val_loss_conf.average), global_step=epoch)
writer.add_summary(make_summary('validation_statistics/loss_class', val_loss_class.average), global_step=epoch)