What parameters did you use for training on the imdb when running the run_classifier.py? Could you share your configuration? When I run it with your code (imdb branch) with default parameters, I get very low accuracy, like 50%.
python run_classifier.py --task_name=imdb --do_train=true --do_eval=true --data_dir=$BERT_DATA_DIR --vocab_file=$BERT_BASE_DIR/vocab.txt --bert_config_file=$BERT_BASE_DIR/bert_config.json --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt --max_seq_length=256 --train_batch_size=32 --learning_rate=2e-5 --num_train_epochs=5.0 --output_dir=$TMP_DIR