How set EncoderDecoderModel.config?

🤗Transformers
1

Is it correct to set the encoder, decoder, and encoder-decoder in the EncoderDecoderModel.config settings as shown below? Or can I just set the encoder-decoder config? If it is correct to set the encoder, decoder, and encoder-decoder separately, can I set all three configs the same? Or can I just match the decoder and encoder-decoder config? And I don’t know why there is a decoder config setting in the encoder config. Someone please tell me.


from transformers import EncoderDecoderModel

model = EncoderDecoderModel.from_pretrained(model_path)

encoder_max_length = 1280
decoder_max_length = 384
    
### encoder-decoder
model.config.pad_token_id = tokenizer.pad_token_id
model.config.sep_token_id = tokenizer.sep_token_id
model.config.bos_token_id = tokenizer.bos_token_id

# decoder_start_token_id (int, optional) — If an encoder-decoder model starts decoding with a different token than bos, the id of that token.
model.config.decoder_start_token_id = tokenizer.cls_token_id
model.config.eos_token_id = tokenizer.sep_token_id

model.config.vocab_size = model.config.encoder.vocab_size
    
model.config.max_length = decoder_max_length
model.config.min_length = 10
    
model.config.no_repeat_ngram_size = 2
model.config.length_penalty = 2.0     

model.config.num_beams = 4
model.config.early_stopping = True
    
### encoder
model.config.encoder.pad_token_id = tokenizer.pad_token_id
model.config.encoder.sep_token_id = tokenizer.sep_token_id
model.config.encoder.bos_token_id = tokenizer.bos_token_id

# decoder_start_token_id (int, optional) — If an encoder-decoder model starts decoding with a different token than bos, the id of that token.
model.config.encoder.decoder_start_token_id = tokenizer.cls_token_id
model.config.encoder.eos_token_id = tokenizer.sep_token_id

model.config.encoder.vocab_size = model.config.encoder.vocab_size

model.config.encoder.max_length = encoder_max_length
model.config.encoder.min_length = 10

model.config.encoder.no_repeat_ngram_size = 2
model.config.encoder.length_penalty = 2.0      

model.config.encoder.num_beams = 4
model.config.encoder.early_stopping = True
    
### decoder
model.config.decoder.pad_token_id = tokenizer.pad_token_id
model.config.decoder.sep_token_id = tokenizer.sep_token_id
model.config.decoder.bos_token_id = tokenizer.bos_token_id

# decoder_start_token_id (int, optional) — If an encoder-decoder model starts decoding with a different token than bos, the id of that token.
model.config.decoder.decoder_start_token_id = tokenizer.cls_token_id
model.config.decoder.eos_token_id = tokenizer.sep_token_id

model.config.decoder.vocab_size = model.config.encoder.vocab_size

model.config.decoder.max_length = decoder_max_length
model.config.decoder.min_length = 10
    
model.config.decoder.no_repeat_ngram_size = 2
model.config.decoder.length_penalty = 2.0       

model.config.decoder.num_beams = 4
    
model.config.decoder.early_stopping = True

model full config

EncoderDecoderConfig {
  "bos_token_id": 5,
  "decoder": {
    "_name_or_path": "./saved_models/kopatent_bigbird_mecab_mlm/checkpoint-271192/",
    "add_cross_attention": true,
    "architectures": [
      "BigBirdForMaskedLM"
    ],
    "attention_probs_dropout_prob": 0.1,
    "attention_type": "block_sparse",
    "bad_words_ids": null,
    "begin_suppress_tokens": null,
    "block_size": 64,
    "bos_token_id": 5,
    "chunk_size_feed_forward": 0,
    "classifier_dropout": null,
    "cross_attention_hidden_size": null,
    "decoder_start_token_id": 2,
    "diversity_penalty": 0.0,
    "do_sample": false,
    "early_stopping": true,
    "encoder_no_repeat_ngram_size": 0,
    "eos_token_id": 3,
    "exponential_decay_length_penalty": null,
    "finetuning_task": null,
    "forced_bos_token_id": null,
    "forced_eos_token_id": null,
    "hidden_act": "gelu_new",
    "hidden_dropout_prob": 0.1,
    "hidden_size": 768,
    "id2label": {
      "0": "LABEL_0",
      "1": "LABEL_1"
    },
    "initializer_range": 0.02,
    "intermediate_size": 2048,
    "is_decoder": true,
    "is_encoder_decoder": false,
    "label2id": {
      "LABEL_0": 0,
      "LABEL_1": 1
    },
    "layer_norm_eps": 1e-12,
    "length_penalty": 2.0,
    "max_length": 384,
    "max_position_embeddings": 4096,
    "min_length": 10,
    "model_type": "big_bird",
    "no_repeat_ngram_size": 2,
    "num_attention_heads": 8,
    "num_beam_groups": 1,
    "num_beams": 4,
    "num_hidden_layers": 4,
    "num_random_blocks": 3,
    "num_return_sequences": 1,
    "output_attentions": false,
    "output_hidden_states": false,
    "output_scores": false,
    "pad_token_id": 0,
    "prefix": null,
    "problem_type": null,
    "pruned_heads": {},
    "remove_invalid_values": false,
    "repetition_penalty": 1.0,
    "rescale_embeddings": false,
    "return_dict": true,
    "return_dict_in_generate": false,
    "sep_token_id": 3,
    "suppress_tokens": null,
    "task_specific_params": null,
    "temperature": 1.0,
    "tf_legacy_loss": false,
    "tie_encoder_decoder": false,
    "tie_word_embeddings": true,
    "tokenizer_class": null,
    "top_k": 50,
    "top_p": 1.0,
    "torch_dtype": "float32",
    "torchscript": false,
    "type_vocab_size": 2,
    "typical_p": 1.0,
    "use_bfloat16": false,
    "use_bias": true,
    "use_cache": true,
    "vocab_size": 32000
  },
  "decoder_start_token_id": 2,
  "early_stopping": true,
  "encoder": {
    "_name_or_path": "./saved_models/kopatent_bigbird_mecab_mlm/checkpoint-271192/",
    "add_cross_attention": false,
    "architectures": [
      "BigBirdForMaskedLM"
    ],
    "attention_probs_dropout_prob": 0.1,
    "attention_type": "block_sparse",
    "bad_words_ids": null,
    "begin_suppress_tokens": null,
    "block_size": 64,
    "bos_token_id": 5,
    "chunk_size_feed_forward": 0,
    "classifier_dropout": null,
    "cross_attention_hidden_size": null,
    "decoder_start_token_id": 2,
    "diversity_penalty": 0.0,
    "do_sample": false,
    "early_stopping": true,
    "encoder_no_repeat_ngram_size": 0,
    "eos_token_id": 3,
    "exponential_decay_length_penalty": null,
    "finetuning_task": null,
    "forced_bos_token_id": null,
    "forced_eos_token_id": null,
    "hidden_act": "gelu_new",
    "hidden_dropout_prob": 0.1,
    "hidden_size": 768,
    "id2label": {
      "0": "LABEL_0",
      "1": "LABEL_1"
    },
    "initializer_range": 0.02,
    "intermediate_size": 2048,
    "is_decoder": false,
    "is_encoder_decoder": false,
    "label2id": {
      "LABEL_0": 0,
      "LABEL_1": 1
    },
    "layer_norm_eps": 1e-12,
    "length_penalty": 2.0,
    "max_length": 1280,
    "max_position_embeddings": 4096,
    "min_length": 10,
    "model_type": "big_bird",
    "no_repeat_ngram_size": 2,
    "num_attention_heads": 8,
    "num_beam_groups": 1,
    "num_beams": 4,
    "num_hidden_layers": 4,
    "num_random_blocks": 3,
    "num_return_sequences": 1,
    "output_attentions": false,
    "output_hidden_states": false,
    "output_scores": false,
    "pad_token_id": 0,
    "prefix": null,
    "problem_type": null,
    "pruned_heads": {},
    "remove_invalid_values": false,
    "repetition_penalty": 1.0,
    "rescale_embeddings": false,
    "return_dict": true,
    "return_dict_in_generate": false,
    "sep_token_id": 3,
    "suppress_tokens": null,
    "task_specific_params": null,
    "temperature": 1.0,
    "tf_legacy_loss": false,
    "tie_encoder_decoder": false,
    "tie_word_embeddings": true,
    "tokenizer_class": null,
    "top_k": 50,
    "top_p": 1.0,
    "torch_dtype": "float32",
    "torchscript": false,
    "type_vocab_size": 2,
    "typical_p": 1.0,
    "use_bfloat16": false,
    "use_bias": true,
    "use_cache": true,
    "vocab_size": 32000
  },
  "eos_token_id": 3,
  "is_encoder_decoder": true,
  "length_penalty": 2.0,
  "max_length": 512,
  "min_length": 10,
  "model_type": "encoder-decoder",
  "no_repeat_ngram_size": 2,
  "num_beams": 4,
  "pad_token_id": 0,
  "sep_token_id": 3,
  "transformers_version": "4.37.0",
  "vocab_size": 32000
}
2

I think that the max length of the encoder and decoder should be the same for the training. Afterward, it is still possible to set the config.max_length for inference generation.

What do you mean by:

there is a decoder config setting in the encoder config

?