transformers/auto_transformers.py

# Copyright 2021 Zilliz. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

import numpy
import os
import torch
import shutil
from pathlib import Path

from transformers import AutoTokenizer, AutoModel, AutoModelForMaskedLM, AutoModelForCausalLM

from towhee.operator import NNOperator
from towhee import register

import warnings
import logging

from train_mlm_with_hf_trainer import train_mlm_with_hf_trainer
from train_clm_with_hf_trainer import train_clm_with_hf_trainer

log = logging.getLogger('run_op')

warnings.filterwarnings('ignore')


@register(output_schema=['vec'])
class AutoTransformers(NNOperator):
    """
    NLP embedding operator that uses the pretrained transformers model gathered by huggingface.
    Args:
        model_name (`str`):
            Which model to use for the embeddings.
    """

    def __init__(self, model_name: str = None, device: str = None, pretrain_weights_path=None,
                 load_pretrain_f=None, tokenizer=None) -> None:
        super().__init__()
        if device is None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        self.device = device
        self.model_name = model_name

        if self.model_name:
            try:
                self.model = AutoModel.from_pretrained(model_name).to(self.device)
                self.configs = self.model.config
            except Exception as e:
                model_list = self.supported_model_names()
                if model_name not in model_list:
                    log.error(f"Invalid model name: {model_name}. Supported model names: {model_list}")
                else:
                    log.error(f"Fail to load model by name: {self.model_name}")
                raise e
            if pretrain_weights_path is not None:
                if load_pretrain_f is None:
                    state_dict = torch.load(pretrain_weights_path, map_location='cpu')
                    self.model.load_state_dict(state_dict)
                else:
                    self.model = load_pretrain_f(self.model, pretrain_weights_path)
            self.model.eval()
            if tokenizer is None:
                try:
                    self.tokenizer = AutoTokenizer.from_pretrained(model_name)
                except Exception as e:
                    log.error(f'Fail to load tokenizer by name: {self.model_name}')
                    raise e
        else:
            log.warning('The operator is initialized without specified model.')
            pass

    def __call__(self, txt: str) -> numpy.ndarray:
        try:
            inputs = self.tokenizer(txt, return_tensors="pt").to(self.device)
        except Exception as e:
            log.error(f'Invalid input for the tokenizer: {self.model_name}')
            raise e
        try:
            outs = self.model(**inputs)
        except Exception as e:
            log.error(f'Invalid input for the model: {self.model_name}')
            raise e
        try:
            features = outs['last_hidden_state'].squeeze(0)
        except Exception as e:
            log.error(f'Fail to extract features by model: {self.model_name}')
            raise e
        vec = features.cpu().detach().numpy()
        return vec

    def save_model(self, format: str = 'pytorch', path: str = 'default'):
        if path == 'default':
            path = str(Path(__file__).parent)
            path = os.path.join(path, 'saved', format)
            os.makedirs(path, exist_ok=True)
            name = self.model_name.replace('/', '-')
            path = os.path.join(path, name)

        dummy_input = '[CLS]'
        inputs = self.tokenizer(dummy_input, return_tensors='pt')  # a dictionary
        if format == 'pytorch':
            torch.save(self.model, path + '.pt')
        elif format == 'torchscript':
            inputs = list(inputs.values())
            try:
                try:
                    jit_model = torch.jit.script(self.model)
                except Exception:
                    jit_model = torch.jit.trace(self.model, inputs, strict=False)
                torch.jit.save(jit_model, path + '.pt')
            except Exception as e:
                log.error(f'Fail to save as torchscript: {e}.')
                raise RuntimeError(f'Fail to save as torchscript: {e}.')
        elif format == 'onnx':
            from transformers.onnx.features import FeaturesManager
            from transformers.onnx import export
            model_kind, model_onnx_config = FeaturesManager.check_supported_model_or_raise(
                self.model, feature='default')
            onnx_config = model_onnx_config(self.model.config)
            if os.path.isdir(path):
                shutil.rmtree(path)
            onnx_inputs, onnx_outputs = export(
                self.tokenizer,
                self.model,
                config=onnx_config,
                opset=13,
                output=Path(path+'.onnx')
            )
        # todo: elif format == 'tensorrt':
        else:
            log.error(f'Unsupported format "{format}".')

    @staticmethod
    def supported_model_names(format: str = None):
        full_list = [
            "bert-large-uncased",
            "bert-base-cased",
            "bert-large-cased",
            "bert-base-multilingual-uncased",
            "bert-base-multilingual-cased",
            "bert-base-chinese",
            "bert-base-german-cased",
            "bert-large-uncased-whole-word-masking",
            "bert-large-cased-whole-word-masking",
            "bert-large-uncased-whole-word-masking-finetuned-squad",
            "bert-large-cased-whole-word-masking-finetuned-squad",
            "bert-base-cased-finetuned-mrpc",
            "bert-base-german-dbmdz-cased",
            "bert-base-german-dbmdz-uncased",
            "cl-tohoku/bert-base-japanese-whole-word-masking",
            "cl-tohoku/bert-base-japanese-char",
            "cl-tohoku/bert-base-japanese-char-whole-word-masking",
            "TurkuNLP/bert-base-finnish-cased-v1",
            "TurkuNLP/bert-base-finnish-uncased-v1",
            "wietsedv/bert-base-dutch-cased",
            "google/bigbird-roberta-base",
            "google/bigbird-roberta-large",
            "google/bigbird-base-trivia-itc",
            "albert-base-v1",
            "albert-large-v1",
            "albert-xlarge-v1",
            "albert-xxlarge-v1",
            "albert-base-v2",
            "albert-large-v2",
            "albert-xlarge-v2",
            "albert-xxlarge-v2",
            "facebook/bart-large",
            "google/bert_for_seq_generation_L-24_bbc_encoder",
            "google/bigbird-pegasus-large-arxiv",
            "google/bigbird-pegasus-large-pubmed",
            "google/bigbird-pegasus-large-bigpatent",
            "google/canine-s",
            "google/canine-c",
            "YituTech/conv-bert-base",
            "YituTech/conv-bert-medium-small",
            "YituTech/conv-bert-small",
            "ctrl",
            "microsoft/deberta-base",
            "microsoft/deberta-large",
            "microsoft/deberta-xlarge",
            "microsoft/deberta-base-mnli",
            "microsoft/deberta-large-mnli",
            "microsoft/deberta-xlarge-mnli",
            "distilbert-base-uncased",
            "distilbert-base-uncased-distilled-squad",
            "distilbert-base-cased",
            "distilbert-base-cased-distilled-squad",
            "distilbert-base-german-cased",
            "distilbert-base-multilingual-cased",
            "distilbert-base-uncased-finetuned-sst-2-english",
            "google/electra-small-generator",
            "google/electra-base-generator",
            "google/electra-large-generator",
            "google/electra-small-discriminator",
            "google/electra-base-discriminator",
            "google/electra-large-discriminator",
            "google/fnet-base",
            "google/fnet-large",
            "facebook/wmt19-ru-en",
            "funnel-transformer/small",
            "funnel-transformer/small-base",
            "funnel-transformer/medium",
            "funnel-transformer/medium-base",
            "funnel-transformer/intermediate",
            "funnel-transformer/intermediate-base",
            "funnel-transformer/large",
            "funnel-transformer/large-base",
            "funnel-transformer/xlarge-base",
            "funnel-transformer/xlarge",
            "gpt2",
            "gpt2-medium",
            "gpt2-large",
            "gpt2-xl",
            "distilgpt2",
            "EleutherAI/gpt-neo-1.3B",
            "EleutherAI/gpt-j-6B",
            "kssteven/ibert-roberta-base",
            "allenai/led-base-16384",
            "google/mobilebert-uncased",
            "microsoft/mpnet-base",
            "uw-madison/nystromformer-512",
            "openai-gpt",
            "google/reformer-crime-and-punishment",
            "tau/splinter-base",
            "tau/splinter-base-qass",
            "tau/splinter-large",
            "tau/splinter-large-qass",
            "squeezebert/squeezebert-uncased",
            "squeezebert/squeezebert-mnli",
            "squeezebert/squeezebert-mnli-headless",
            "transfo-xl-wt103",
            "xlm-mlm-en-2048",
            "xlm-mlm-ende-1024",
            "xlm-mlm-enfr-1024",
            "xlm-mlm-enro-1024",
            "xlm-mlm-tlm-xnli15-1024",
            "xlm-mlm-xnli15-1024",
            "xlm-clm-enfr-1024",
            "xlm-clm-ende-1024",
            "xlm-mlm-17-1280",
            "xlm-mlm-100-1280",
            "xlm-roberta-base",
            "xlm-roberta-large",
            "xlm-roberta-large-finetuned-conll02-dutch",
            "xlm-roberta-large-finetuned-conll02-spanish",
            "xlm-roberta-large-finetuned-conll03-english",
            "xlm-roberta-large-finetuned-conll03-german",
            "xlnet-base-cased",
            "xlnet-large-cased",
            "uw-madison/yoso-4096",
            "microsoft/deberta-v2-xlarge",
            "microsoft/deberta-v2-xxlarge",
            "microsoft/deberta-v2-xlarge-mnli",
            "microsoft/deberta-v2-xxlarge-mnli",
            "flaubert/flaubert_small_cased",
            "flaubert/flaubert_base_uncased",
            "flaubert/flaubert_base_cased",
            "flaubert/flaubert_large_cased",
            "camembert-base",
            "Musixmatch/umberto-commoncrawl-cased-v1",
            "Musixmatch/umberto-wikipedia-uncased-v1",
        ]
        full_list.sort()
        if format is None:
            model_list = full_list
        elif format == 'pytorch':
            to_remove = []
            assert set(to_remove).issubset(set(full_list))
            model_list = list(set(full_list) - set(to_remove))
        elif format == 'torchscript':
            to_remove = [
                'EleutherAI/gpt-j-6B',
                'EleutherAI/gpt-neo-1.3B',
                'allenai/led-base-16384',
                'ctrl',
                'distilgpt2',
                'facebook/bart-large',
                'google/bigbird-pegasus-large-arxiv',
                'google/bigbird-pegasus-large-bigpatent',
                'google/bigbird-pegasus-large-pubmed',
                'google/canine-c',
                'google/canine-s',
                'google/reformer-crime-and-punishment',
                'gpt2',
                'gpt2-large',
                'gpt2-medium',
                'gpt2-xl',
                'microsoft/deberta-base',
                'microsoft/deberta-base-mnli',
                'microsoft/deberta-large',
                'microsoft/deberta-large-mnli',
                'microsoft/deberta-xlarge',
                'microsoft/deberta-xlarge-mnli',
                'openai-gpt',
                'transfo-xl-wt103',
                'uw-madison/yoso-4096',
                'xlm-clm-ende-1024',
                'xlm-clm-enfr-1024',
                'xlm-mlm-100-1280',
                'xlm-mlm-17-1280',
                'xlm-mlm-en-2048',
                'xlm-mlm-ende-1024',
                'xlm-mlm-enfr-1024',
                'xlm-mlm-enro-1024',
                'xlm-mlm-tlm-xnli15-1024',
                'xlm-mlm-xnli15-1024',
                'xlnet-base-cased',
                'xlnet-large-cased',
                'microsoft/deberta-v2-xlarge',
                'microsoft/deberta-v2-xxlarge',
                'microsoft/deberta-v2-xlarge-mnli',
                'microsoft/deberta-v2-xxlarge-mnli',
                'flaubert/flaubert_small_cased',
                'flaubert/flaubert_base_uncased',
                'flaubert/flaubert_base_cased',
                'flaubert/flaubert_large_cased'
            ]
            assert set(to_remove).issubset(set(full_list))
            model_list = list(set(full_list) - set(to_remove))
        elif format == 'onnx':
            to_remove = [
                'ctrl',
                'funnel-transformer/intermediate',
                'funnel-transformer/large',
                'funnel-transformer/medium',
                'funnel-transformer/small',
                'funnel-transformer/xlarge',
                'google/canine-c',
                'google/canine-s',
                'google/fnet-base',
                'google/fnet-large',
                'google/reformer-crime-and-punishment',
                'transfo-xl-wt103',
                'uw-madison/yoso-4096',
                'xlnet-base-cased',
                'xlnet-large-cased'
            ]
            assert set(to_remove).issubset(set(full_list))
            model_list = list(set(full_list) - set(to_remove))
        # todo: elif format == 'tensorrt':
        else:
            log.error(f'Invalid format "{format}". Currently supported formats: "pytorch", "torchscript".')
        return model_list

    def train(self, training_config=None,
              train_dataset=None,
              eval_dataset=None,
              resume_checkpoint_path=None, **kwargs):
        task = kwargs.pop('task', None)
        data_args = kwargs.pop('data_args', None)
        training_args = kwargs.pop('training_args', None)
        prepare_model_weights_f = kwargs.pop('prepare_model_weights_f', None)

        if task == 'mlm' or task is None:
            model_with_head = AutoModelForMaskedLM.from_pretrained(self.model_name)
            if prepare_model_weights_f is not None:
                model_with_head = prepare_model_weights_f(self.model, model_with_head, **kwargs)

            train_mlm_with_hf_trainer(
                model_with_head,
                self.tokenizer,
                data_args,
                training_args,
                **kwargs
            )
        elif task == 'clm':
            model_with_head = AutoModelForCausalLM.from_pretrained(self.model_name)
            if prepare_model_weights_f is not None:
                model_with_head = prepare_model_weights_f(self.model, model_with_head, **kwargs)

            train_clm_with_hf_trainer(
                model_with_head,
                self.tokenizer,
                data_args,
                training_args,
                **kwargs
            )