Created by: winston-zillow

Download https://ernie.bj.bcebos.com/ERNIE_Base_en_stable-2.0.0.tar.gz, uncompress and the ernie_config.json does not have a value for use_task_id.

{
  "attention_probs_dropout_prob": 0.1, 
  "hidden_act": "gelu", 
  "hidden_dropout_prob": 0.1, 
  "hidden_size": 768, 
  "initializer_range": 0.02, 
  "max_position_embeddings": 512, 
  "num_attention_heads": 12, 
  "num_hidden_layers": 12, 
  "sent_type_vocab_size": 4, 
  "task_type_vocab_size": 16, 
  "vocab_size": 30522
}

As a consequence, when attempting to save the model using save_inference_model (so that I can convert it to ONNX using the paddle2onnx tool), it would complain the task id input is missing after pruning the model for inference inside save_inference_model as shown in my debug outputs by modifying the save method locally:

cloned 2 -> read_file_0.tmp_0 VarType.LOD_TENSOR False
cloned 6 -> _input_reader_reader VarType.READER True
cloned 32 -> read_file_0.tmp_6 VarType.LOD_TENSOR False
cloned 41 -> read_file_0.tmp_5 VarType.LOD_TENSOR False
cloned 44 -> read_file_0.tmp_4 VarType.LOD_TENSOR False
cloned 63 -> read_file_0.tmp_1 VarType.LOD_TENSOR False
cloned 219 -> read_file_0.tmp_3 VarType.LOD_TENSOR False
cloned 374 -> read_file_0.tmp_2 VarType.LOD_TENSOR False
pruned 116 -> read_file_0.tmp_4 VarType.LOD_TENSOR False
pruned 117 -> read_file_0.tmp_0 VarType.LOD_TENSOR False
pruned 129 -> read_file_0.tmp_1 VarType.LOD_TENSOR False
pruned 718 -> read_file_0.tmp_2 VarType.LOD_TENSOR False

For your reference, my model saving codes look like

# cloned and modified from `ernie/finetune/classifier`
def create_model(args,
                 pyreader_name,
                 ernie_config,
                 is_prediction=False,
                 task_name="",
                 is_classify=False,
                 is_regression=False,
                 ernie_version="1.0"):
     ...
    (src_ids, sent_ids, pos_ids, task_ids, input_mask, labels,
     qids) = fluid.layers.read_file(pyreader)

    ernie = ErnieModel(
    ...
    cls_feats = ernie.get_pooled_output()
    inputs = (src_ids, sent_ids, pos_ids, task_ids, input_mask)
    return inputs, cls_feats, pyreader, (labels, qids)

# imitating `run_classifier.py`
startup_prog = fluid.Program()
train_program = fluid.Program()

with fluid.program_guard(train_program, startup_prog):
    with fluid.unique_name.guard():
        inputs, ernie_latent, train_pyreader, _ = create_model(
            args,
            pyreader_name='input_reader',
            ernie_config=ernie_config,
            is_classify=args.is_classify,
            is_regression=args.is_regression)

        # exe.run(startup_prog) . # i think there is no need to run this
        init_pretraining_params(
            exe,
            args.init_pretraining_params,
            main_program=startup_prog,
            use_fp16=args.use_fp16)
    fluid.io.save_inference_model(dirname="/ernie_onnx_path",
                                 feeded_var_names=[var.name for var in inputs],
                                 target_vars=[ernie_latent],
                                 main_program=train_program,
                                 executor=exe)

After I set use_task_id to true in the config file, saving the inference model will work. Please confirm if that's the right fix (and fix the pre-trained archive.)