from typing import List, Tuple
import pickle
import numpy as np
import sentencepiece as spm
from ..utils import check_and_download
from ._spu_context_bigru_utils import (
create_spucbigru_sentiment_model,
process_text_input,
)
# Resolving parent dependencies
from inspect import getsourcefile
import os
import sys
current_path = os.path.abspath(getsourcefile(lambda: 0))
current_dir = os.path.dirname(current_path)
parent_dir = current_dir[: current_dir.rfind(os.path.sep)]
sys.path.insert(0, parent_dir)
RESOURCES_PATH = os.path.join(os.path.dirname(__file__), "resources/")
PROD_WEIGHTS_LOC = RESOURCES_PATH + "Sentiment_SPUCBiGRU_prod.weights"
EVAL_WEIGHTS_LOC = RESOURCES_PATH + "Sentiment_SPUCBiGRU_eval.weights"
WORD_EMBEDDING_MATRIX_LOC = os.path.abspath(
os.path.join(
os.path.dirname(__file__),
"..",
"resources/SPUTokenized_word_embedding_16k.matrix",
)
)
PROD_WEIGHTS_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/Sentiment_SPUCBiGRU_prod.weights"
EVAL_WEIGHTS_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/Sentiment_SPUCBiGRU_eval.weights"
WORD_EMBEDDING_MATRIX_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/SPUTokenized_word_embedding_16k.matrix"
SPU_TOKENIZER_WORD_LOC = os.path.abspath(
os.path.join(
os.path.dirname(__file__),
"..",
"resources/SPU_word_tokenizer_16k.model",
)
)
# Data Preprocessing Config
TEXT_MAX_LEN = 256
# Loading Tokenizer
spu_tokenizer_word = spm.SentencePieceProcessor(SPU_TOKENIZER_WORD_LOC)
sp_key_to_index = {
spu_tokenizer_word.id_to_piece(id): id
for id in range(spu_tokenizer_word.get_piece_size())
}
sp_index_to_key = {
id: spu_tokenizer_word.id_to_piece(id)
for id in range(spu_tokenizer_word.get_piece_size())
}
WORD_EMBEDDING_VOCAB_SIZE = len(sp_key_to_index)
WORD_EMBEDDING_VECTOR_SIZE = 128
WORD_EMBEDDING_MATRIX = np.zeros(
(WORD_EMBEDDING_VOCAB_SIZE, WORD_EMBEDDING_VECTOR_SIZE)
)
NUM_RNN_STACKS = 3
NUM_RNN_UNITS = 128
DROPOUT = 0.2
[docs]class SPUCBiGRUSentimentAnalyzer:
"""
SentencePiece Unigram Context Bidirectional GRU Sentiment Analyzer class.
- This is a Bidirectional `GRU <https://arxiv.org/abs/1412.3555>`_ based Sentiment Analyzer that uses `SentencePiece Unigram <https://arxiv.org/abs/1804.10959>`_ tokenizer and pre-trained Word2Vec embeddings.
- It achieves 0.9469 Accuracy, 0.9380 F1 macro score and 0.9147 F1 score (treating class 0 as minority).
- For more details about the training procedure, dataset and evaluation metrics, see `ReadMe <https://github.com/vngrs-ai/VNLP/blob/main/vnlp/sentiment_analyzer/ReadMe.md>`_.
"""
def __init__(self, evaluate):
self.model = create_spucbigru_sentiment_model(
TEXT_MAX_LEN,
WORD_EMBEDDING_VOCAB_SIZE,
WORD_EMBEDDING_VECTOR_SIZE,
WORD_EMBEDDING_MATRIX,
NUM_RNN_UNITS,
NUM_RNN_STACKS,
DROPOUT,
)
# Check and download word embedding matrix and model weights
check_and_download(
WORD_EMBEDDING_MATRIX_LOC, WORD_EMBEDDING_MATRIX_LINK
)
if evaluate:
MODEL_WEIGHTS_LOC = EVAL_WEIGHTS_LOC
MODEL_WEIGHTS_LINK = EVAL_WEIGHTS_LINK
else:
MODEL_WEIGHTS_LOC = PROD_WEIGHTS_LOC
MODEL_WEIGHTS_LINK = PROD_WEIGHTS_LINK
check_and_download(MODEL_WEIGHTS_LOC, MODEL_WEIGHTS_LINK)
# Load Word embedding matrix
word_embedding_matrix = np.load(WORD_EMBEDDING_MATRIX_LOC)
# Load Model weights
with open(MODEL_WEIGHTS_LOC, "rb") as fp:
model_weights = pickle.load(fp)
# Insert word embedding weights to correct position (0 for SPUBiGRUSentimentAnalyzer model)
model_weights.insert(0, word_embedding_matrix)
# Set model weights
self.model.set_weights(model_weights)
self.spu_tokenizer_word = spu_tokenizer_word
[docs] def predict(self, text: str) -> List[Tuple[str, str]]:
"""
Args:
text:
Input text.
Returns:
Sentiment label of input text.
"""
prob = self.predict_proba(text)
return 1 if prob > 0.5 else 0
[docs] def predict_proba(self, text: str) -> float:
"""
Args:
text:
Input text.
Returns:
Probability that the input text has positive sentiment.
"""
tokenized_text = process_text_input(
text, self.spu_tokenizer_word, TEXT_MAX_LEN
)
num_int_tokens = len(tokenized_text[0])
num_str_tokens = len(text.split())
# if the text is longer than the length the model is trained on
if num_int_tokens > TEXT_MAX_LEN:
first_half_of_preprocessed_text = " ".join(
text.split()[: (num_str_tokens // 2)]
)
second_half_of_preprocessed_text = " ".join(
text.split()[(num_str_tokens // 2) :]
)
prob = (
self.predict_proba(first_half_of_preprocessed_text)
+ self.predict_proba(second_half_of_preprocessed_text)
) / 2
else:
prob = self.model(tokenized_text).numpy()[0][0]
return prob