Electra-base returns always same output

🤗Transformers
1

I pretrained Electra-base model using Google-electra

In hparams i had only set “model_type:base”

Then i converted checkpoints using this script

with setting:

{
  "vocab_size": 30522,
  "embedding_size": 768,
  "hidden_size": 768,
  "num_hidden_layers": 12,
  "num_attention_heads": 12,
  "intermediate_size": 3072,
  "generator_size": "1/4",
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "attention_probs_dropout_prob": 0.1,
  "max_position_embeddings": 512,
  "type_vocab_size": 2,
  "initializer_range": 0.02,
  "model_type": "electra"
}

I’m loading model with:

discriminator = AutoModel.from_pretrained("./electraBase")

and my output is always constant vector.

tensor([[ 2.4182, -0.0906, -1.5302,  ...,  6.7738,  0.4463,  0.0859],
        [ 2.4182, -0.0906, -1.5302,  ...,  6.7738,  0.4463,  0.0859],
        [ 2.4182, -0.0906, -1.5302,  ...,  6.7738,  0.4463,  0.0859],
        [ 2.4182, -0.0906, -1.5302,  ...,  6.7738,  0.4463,  0.0859]],
       grad_fn=<SliceBackward0>)

Is something wrong in loading process?