BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension 바트 모델(Bart model)
논문 : https://arxiv.org/pdf/1910.13461.pdf
github pytorch/fairseq/BART : https://github.com/pytorch/fairseq/tree/master/examples/bart
github pytorch/fairseq : https://github.com/pytorch/fairseq
Facebook AI Research
0. 소개
먼저 Github에는 이렇게 나와있습니다.
BART는 프리 트레이닝 목표로 노이즈 제거 기능을 갖춘 시퀀스-시퀀스 모델입니다. 우리는이 사전 훈련 목표가보다 일반적이며 SQuAD 및 GLUE 에서 RoBERTa 결과를 일치 시키고 요약 (XSum, CNN 데이터 세트), 긴 형식의 생성 질문 답변 (ELI5) 및 대화에 대한 최첨단 결과를 얻을 수 있음을 보여줍니다 반응 생성 (ConvAI2). 자세한 내용은 관련 용지를 참조하십시오.
Pre-trained models
Model | Description | params | Download |
bart.large | BART model with 12 encoder and decoder layers | 400M | bart.large.tar.gz |
bart.large.mnli | bart.large finetuned on MNLI | 400M | bart.large.mnli.tar.gz |
bart.large.cnn | bart.large finetuned on CNN-DM | 400M | bart.large.cnn.tar.gz |
bart.large.xsum | bart.large finetuned on Xsum | 400M | bart.large.xsum.tar.gz |
그 중에서 이번에 소개드릴 모델은 bart.large 모델입니다.
1. 모델의 구성
Transformer의 Encoder와 Decoder를 모두 이용한 모델입니다.
- Encoder : 12개
- Decoder : 12개
- Embedding : 50265개
- padding_idx = 1
- ...
BARTModel(
(encoder): TransformerEncoder(
(embed_tokens): Embedding(50265, 1024, padding_idx=1)
(embed_positions): LearnedPositionalEmbedding(1026, 1024, padding_idx=1)
(layers): ModuleList(
(0): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(1): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(2): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(3): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(4): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(5): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(6): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(7): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(8): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(9): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(10): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(11): TransformerEncoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
)
(layernorm_embedding): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(decoder): TransformerDecoder(
(embed_tokens): Embedding(50265, 1024, padding_idx=1)
(embed_positions): LearnedPositionalEmbedding(1026, 1024, padding_idx=1)
(layers): ModuleList(
(0): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(1): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(2): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(3): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(4): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(5): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(6): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(7): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(8): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(9): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(10): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(11): TransformerDecoderLayer(
(self_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(self_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(encoder_attn): MultiheadAttention(
(k_proj): Linear(in_features=1024, out_features=1024, bias=True)
(v_proj): Linear(in_features=1024, out_features=1024, bias=True)
(q_proj): Linear(in_features=1024, out_features=1024, bias=True)
(out_proj): Linear(in_features=1024, out_features=1024, bias=True)
)
(encoder_attn_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
(fc1): Linear(in_features=1024, out_features=4096, bias=True)
(fc2): Linear(in_features=4096, out_features=1024, bias=True)
(final_layer_norm): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
)
(layernorm_embedding): LayerNorm((1024,), eps=1e-05, elementwise_affine=True)
)
(classification_heads): ModuleDict()
)
2. 사용 예제
방법 1 : Load BART from torch.hub (PyTorch >= 1.1) : 파이토치 버전이 1.1 이상이라면 다음과 같이 실행할 수 있습니다.
import torch
bart = torch.hub.load('pytorch/fairseq', 'bart.large')
#bart.cuda() # GPU를 사용하고 싶다면 다음을 추가해줍니다.
bart.eval() # disable dropout (or leave in train mode to finetune)
이 방법은 .cache를 사용하는 방법입니다. 사용자의 컴퓨터가 .cache를 지원하지 않는다면 다음 혹은 다른 오류가 발생하게 됩니다. 필자의 경우에는 98%에서 계속 멈추었습니다. 이것을 해결할 방법이 있습니다.
컴퓨터 혹은 노트북이 2대 이상일 때 사용할 수 있는 이 방법은 다른 컴퓨터에서 .cache를 이용하여 파일을 옮기는 방법입니다.
아래의 사진은 .cache를 통해 bart 모델을 다운 받는 것입니다. 따라서 서버용 컴퓨터의 오류로 인해 노트북에서 다운을 받고 .cache 폴더에 들어가 파일을 압축하여 보낸 후 똑같은 .cache에서 압축을 풀어줍니다.
제 노트북은 삼성 윈10을 쓰고있는데 .cache의 폴더 위치는 C: > 사용자 > 사용자명 > .cache에 다운받아지고 이를 옮겨
서버용 컴퓨터 우분투 18.04 기준 home/사용자명/.cache 에 저장을 통해 문제를 해결하였습니다.
방법 2 : Load BART (for PyTorch 1.0 or custom models) : 또 다른 방법입니다. 이 방법은 custom model 또는 Pytorch 1.0에서 사용하실 수 있습니다. (추가로 1.1버전 이상에서도 사용하실 수 있습니다.)
# Download bart.large model
wget https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz
tar -xzvf bart.large.tar.gz
# Load the model in fairseq
from fairseq.models.bart import BARTModel
bart = BARTModel.from_pretrained('/path/to/bart.large', checkpoint_file='model.pt')
bart.eval() # disable dropout (or leave in train mode to finetune)
BART를 이용하려는 폴더에서 bart.large 다운 받은 후 압축을 풀고 그 폴더 내의 모델을 이용하는 것입니다.
저의 경우에는 이 방법 2번을 선호합니다. 통합된 환경에서 사용하는 것도 좋지만 무엇보다 제가 느끼기에는 코드 반응 체감이 더 빠릅니다.
다음 코드를 실행하시면 아래와 같이 결과가 생성됩니다.
이후에 다음을 테스트 할 수 있습니다.
tokens = bart.encode('Hello world!')
assert tokens.tolist() == [0, 31414, 232, 328, 2]
bart.decode(tokens) # 'Hello world!'
- bart.encode의 경우에는 인풋을 토큰으로 바꿔줍니다.
- bart.decode의 경우에는 토큰을 출력으로 바꿔줍니다.
모델에는 다양한 코드들이 존재하는데 이 부분은 모델 혹은 fairseq bart에서 사용하고자 하는 내용들을 찾아보시면 될 것 같습니다.
일반적인 모델 흐름에 대해 간단하게 말씀드리자면 다음과 같습니다.
- 입력 데이터 'English paragraph'가 입력으로 주어집니다.
- 입력 데이터가 encode를 통해 토큰화 됩니다.
- 토큰화 된 데이터가 Encoder에서 가공됩니다.
- 가공된 데이터가 Decoder에 들어갑니다.
- Decoder에서 나온 결과를 보고싶은 결과에 맞게 output 결과를 수정합니다.
사용하고 싶은 모델이나 모델에 대해서 자세하게 알고 싶으신 분들은 아래에서와 같이 Find File을 사용하여 함수들의 의미를 파악하실 수 있습니다.