Sentiment

BertTextEncoder

Bases: Module

Source code in Agent/modules/emotion_detection/sentiment.py

class BertTextEncoder(nn.Module):
    def __init__(self, language="en", use_finetune=False):
        """
        language: en / cn
        """
        super(BertTextEncoder, self).__init__()

        assert language in ["en", "cn"]

        tokenizer_class = BertTokenizer
        model_class = BertModel
        # directory is fine
        # pretrained_weights = '/home/sharing/disk3/pretrained_embedding/Chinese/bert/pytorch'
        if language == "en":
            self.tokenizer = tokenizer_class.from_pretrained(
                f"{models_dir}/bert_en", do_lower_case=True
            )
            self.model = model_class.from_pretrained(f"{models_dir}/bert_en")
        elif language == "cn":
            self.tokenizer = tokenizer_class.from_pretrained(f"{models_dir}/bert_cn")
            self.model = model_class.from_pretrained(f"{models_dir}/bert_cn")

        self.use_finetune = use_finetune

    def get_tokenizer(self):
        return self.tokenizer

    def from_text(self, text):
        """
        text: raw data
        """
        input_ids = self.get_id(text)
        with torch.no_grad():
            last_hidden_states = self.model(input_ids)[
                0
            ]  # Models outputs are now tuples
        return last_hidden_states.squeeze()

    def forward(self, text):
        """
        text: (batch_size, 3, seq_len)
        3: input_ids, input_mask, segment_ids
        input_ids: input_ids,
        input_mask: attention_mask,
        segment_ids: token_type_ids
        """
        text = self.tokenizer(text)
        input_ids, input_mask, segment_ids = (
            torch.tensor(text["input_ids"]).long().unsqueeze(0),
            torch.tensor(text["token_type_ids"]).unsqueeze(0).float(),
            torch.tensor(text["attention_mask"]).unsqueeze(0).long(),
        )
        if self.use_finetune:
            last_hidden_states = self.model(
                input_ids=input_ids,
                attention_mask=input_mask,
                token_type_ids=segment_ids,
            )[
                0
            ]  # Models outputs are now tuples
        else:
            with torch.no_grad():
                last_hidden_states = self.model(
                    input_ids=input_ids,
                    attention_mask=input_mask,
                    token_type_ids=segment_ids,
                )[
                    0
                ]  # Models outputs are now tuples
        return last_hidden_states

__init__(language='en', use_finetune=False)

language: en / cn

Source code in Agent/modules/emotion_detection/sentiment.py

def __init__(self, language="en", use_finetune=False):
    """
    language: en / cn
    """
    super(BertTextEncoder, self).__init__()

    assert language in ["en", "cn"]

    tokenizer_class = BertTokenizer
    model_class = BertModel
    # directory is fine
    # pretrained_weights = '/home/sharing/disk3/pretrained_embedding/Chinese/bert/pytorch'
    if language == "en":
        self.tokenizer = tokenizer_class.from_pretrained(
            f"{models_dir}/bert_en", do_lower_case=True
        )
        self.model = model_class.from_pretrained(f"{models_dir}/bert_en")
    elif language == "cn":
        self.tokenizer = tokenizer_class.from_pretrained(f"{models_dir}/bert_cn")
        self.model = model_class.from_pretrained(f"{models_dir}/bert_cn")

    self.use_finetune = use_finetune

forward(text)

text: (batch_size, 3, seq_len) 3: input_ids, input_mask, segment_ids input_ids: input_ids, input_mask: attention_mask, segment_ids: token_type_ids

Source code in Agent/modules/emotion_detection/sentiment.py

def forward(self, text):
    """
    text: (batch_size, 3, seq_len)
    3: input_ids, input_mask, segment_ids
    input_ids: input_ids,
    input_mask: attention_mask,
    segment_ids: token_type_ids
    """
    text = self.tokenizer(text)
    input_ids, input_mask, segment_ids = (
        torch.tensor(text["input_ids"]).long().unsqueeze(0),
        torch.tensor(text["token_type_ids"]).unsqueeze(0).float(),
        torch.tensor(text["attention_mask"]).unsqueeze(0).long(),
    )
    if self.use_finetune:
        last_hidden_states = self.model(
            input_ids=input_ids,
            attention_mask=input_mask,
            token_type_ids=segment_ids,
        )[
            0
        ]  # Models outputs are now tuples
    else:
        with torch.no_grad():
            last_hidden_states = self.model(
                input_ids=input_ids,
                attention_mask=input_mask,
                token_type_ids=segment_ids,
            )[
                0
            ]  # Models outputs are now tuples
    return last_hidden_states

from_text(text)

text: raw data

Source code in Agent/modules/emotion_detection/sentiment.py

def from_text(self, text):
    """
    text: raw data
    """
    input_ids = self.get_id(text)
    with torch.no_grad():
        last_hidden_states = self.model(input_ids)[
            0
        ]  # Models outputs are now tuples
    return last_hidden_states.squeeze()