Accuracy of MobileNetV3-minimalistic model is different from given checkpoint

System information

• What is the top-level directory of the model you are using: /research/slim/nets/mobilenet
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 18.04
• TensorFlow installed from (source or binary): source
• TensorFlow version (use command below): 1.15.0
• Bazel version (if compiling from source): 0.24.1
• CUDA/cuDNN version: CUDA 10.0 CuDNN 7
• GPU model and memory: 5 Titan XP / 12G
• Exact command to reproduce:

python research/slim/train_image_classifier.py \
    --train_dir={TRAIN_DIR} \
    --dataset_name=imagenet \
    --dataset_split_name=train \
    --dataset_dir={DATASET_DIR} \
    --model_name=v3_large_min \
    --learning_rate=0.1 \
    --preprocessing_name="inception_v2" \
    --label_smoothing=0.1 \
    --moving_average_decay=0.9999 \
    --batch_size=64 \
    --num_readers=16 \
    --num_preprocessing_threads=16 \
    --num_clones=5 \
    --learning_rate_decay_factor=0.01 \
    --num_epoches_per_decay=3.0 \
    --weight_decay=0.00001

Custom code

1. nets_factory.py
   Since there is no keys for MobilenetV3 in mapping dictionary in nets_factory.py, I added 'v3_large_min' in networks_map and arg_scopes_map (I think it should be added in official repo)

2. mobilenet.py
   According to paper, I change default value of bn_decay in training_scope to 0.99

Describe the problem

Since I want to train MobileNetV3 Large minimalistic (float) model, I measured accuracy from pre-trained checkpoint (it was step 342500). It performs 72.292% in Top 1 Accuracy. (I use eval_image_classifier.py)

However, when I train and evaluate this model from scratch, it seems to have 61% peak accuracy.
(even more than 342500 steps)

As you can see in command, I use hyperparameter according to MobileNetV3 paper.
Did I miss something for training (e.g. other hyperparameter)? If then, can you share them?