Error creating and exporting TFLite after fine-tuning (using run_squad_v1.py/run_squad_v2.py)

[manojpreveen]

python -m albert.run_squad_v1 \ --albert_config_file=albert/assets/albert_config.json \ --output_dir=albert/output \ --train_file=albert/train-v1.1.json \ --predict_file=albert/dev-v1.1.json \ --train_feature_file=albert/output/train.tfrecord \ --predict_feature_file=albert/output/dev.tfrecord \ --predict_feature_left_file=albert/output/pred_left_file_example.pkl \ --albert_hub_module_handle=https://tfhub.dev/google/albert_base/1 \ --spm_model_file=albert/assets/30k-clean.model \ --do_lower_case \ --max_seq_length=384 \ --doc_stride=128 \ --max_query_length=64 \ --do_train=true \ --do_predict=false \ --train_batch_size=16 \ --predict_batch_size=8 \ --learning_rate=5e-5 \ --num_train_epochs=2.0 \ --warmup_proportion=.1 \ --save_checkpoints_steps=5000 \ --n_best_size=20 \ --max_answer_length=128 \ --export_dir=albert/modelite

The above is the command I ran to finetune Albert-Base on SQuAD1.1 The finetune was successful and I got both tensorflow checkpoints and the model in saved_model format.

When the Conversion of saved_model to TFLite Model step starts, it's throwing out an error as below :

File "/usr/local/share/anaconda3/lib/python3.7/site-packages/tensorflow_core/lite/python/convert_saved_model.py", line 194, in freeze_saved_model raise ValueError("SavedModels with assets/ directory are not supported.") ValueError: SavedModels with assets/ directory are not supported.

Is it just happening for me or is it not supported by tensorflow itself yet. Is there any workaround to create TFLite model. Please Help.

Thanks.

[taylorshin]

I'm having a similar problem. We would greatly appreciate some help/guidance!

[PantherYan]

We also meet this issue, any solutions?