low mAP, high classification loss in eval - object detection resnet 50

[qazalkz]

Hi guys. I have about 4000 images and Im using resnrt50 640*640 with 70000 steps. but after all Im getting low mAP. could you please give me some advice.

p.n: also I divide original learning rate to my batch size

pipeline_config.model.ssd.num_classes = 1
pipeline_config.train_config.batch_size = 64//8
pipeline_config.train_config.fine_tune_checkpoint = "checkpointmodels/retina101/ssd_/checkpoint/ckpt-0"
pipeline_config.train_config.fine_tune_checkpoint_type = "detection"
pipeline_config.train_input_reader.label_map_path= LABELMAP
pipeline_config.train_config.optimizer.momentum_optimizer.learning_rate.cosine_decay_learning_rate.warmup_learning_rate=0.013333000242710114/8
pipeline_config.train_config.optimizer.momentum_optimizer.learning_rate.cosine_decay_learning_rate.learning_rate_base=0.03999999910593033/8

on 20000 stept eval

on 70000 steps


also you can see that classification loss is high on eval output