domain_mapping/cosines_with_augmentations/classify.py

# %%

# from datasets import load_from_disk
import os
import glob

os.environ['NCCL_P2P_DISABLE'] = '1'
os.environ['NCCL_IB_DISABLE'] = '1'
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"

import re
import torch
from torch.utils.data import DataLoader
import torch
import torch.nn as nn

from transformers import (
    AutoTokenizer,
    AutoModelForSequenceClassification,
    AutoModel,
    DataCollatorWithPadding,
)
import evaluate
import numpy as np
import pandas as pd
# import matplotlib.pyplot as plt
from datasets import Dataset, DatasetDict

from tqdm import tqdm

torch.set_float32_matmul_precision('high')


BATCH_SIZE = 256

# %%
# construct the target id list
# data_path = '../../../esAppMod_data_import/train.csv'
data_path = '../esAppMod_data_import/train.csv'
train_df = pd.read_csv(data_path, skipinitialspace=True)
# rather than use pattern, we use the real thing and property
entity_ids = train_df['entity_id'].to_list()
target_id_list = sorted(list(set(entity_ids)))


# %%
id2label = {}
label2id = {}
for idx, val in enumerate(target_id_list):
    id2label[idx] = val
    label2id[val] = idx


# introduce pre-processing functions
def preprocess_text(text):
    # 1. Make all uppercase
    text = text.lower()

    # Substitute digits with '#'
    # text = re.sub(r'\d+', '#', text)

    # standardize spacing
    text = re.sub(r'\s+', ' ', text).strip()

    return text




# outputs a list of dictionaries
# processes dataframe into lists of dictionaries
# each element maps input to output
# input: tag_description
# output: class label
def process_df_to_dict(df):
    output_list = []
    for _, row in df.iterrows():
        desc = row['mention']
        desc = preprocess_text(desc)
        index = row['entity_id']
        element = {
            'text' : desc,
            'label': label2id[index], # ensure labels starts from 0
        }
        output_list.append(element)

    return output_list


def create_dataset():
    # train 
    data_path = '../esAppMod_data_import/test.csv'
    test_df = pd.read_csv(data_path, skipinitialspace=True)


    # combined_data = DatasetDict({
    #     'train': Dataset.from_list(process_df_to_dict(train_df)),
    # })
    return Dataset.from_list(process_df_to_dict(test_df))



# %%

def test():

    test_dataset = create_dataset()

    # prepare tokenizer

    MODEL_NAME = 'prajjwal1/bert-small' # 'prajjwal1/bert-small' 'bert-base-cased'  'distilbert-base-cased' 
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, return_tensors="pt", clean_up_tokenization_spaces=True)
    # Define additional special tokens
    # additional_special_tokens = ["<THING_START>", "<THING_END>", "<PROPERTY_START>", "<PROPERTY_END>", "<NAME>", "<DESC>", "<SIG>", "<UNIT>", "<DATA_TYPE>"]
    # Add the additional special tokens to the tokenizer
    # tokenizer.add_special_tokens({"additional_special_tokens": additional_special_tokens})

    # %%
    # compute max token length
    max_length = 0
    for sample in test_dataset['text']:
        # Tokenize the sample and get the length
        input_ids = tokenizer(sample, truncation=False, add_special_tokens=True)["input_ids"]
        length = len(input_ids)
        
        # Update max_length if this sample is longer
        if length > max_length:
            max_length = length

    print(max_length)

    # %%

    max_length = 128

    # given a dataset entry, run it through the tokenizer
    def preprocess_function(example):
        input = example['text']
        # text_target sets the corresponding label to inputs
        # there is no need to create a separate 'labels'
        model_inputs = tokenizer(
            input,
            # max_length=max_length,
            # padding='max_length'
        )
        return model_inputs

    # map maps function to each "row" in the dataset
    # aka the data in the immediate nesting
    datasets = test_dataset.map(
        preprocess_function,
        batched=True,
        num_proc=8,
        remove_columns="text",
    )


    datasets.set_format(type='torch', columns=['input_ids', 'attention_mask', 'label'])

    data_collator = DataCollatorWithPadding(tokenizer=tokenizer)


    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    bert_model = AutoModel.from_pretrained(MODEL_NAME)

    class BertForClassificationAndTriplet(nn.Module):
        def __init__(self, bert_model, num_classes):
            super().__init__()
            self.bert = bert_model
            self.classifier = nn.Linear(bert_model.config.hidden_size, num_classes)

        def forward(self, input_ids, attention_mask=None):
            outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
            cls_embeddings = outputs.last_hidden_state[:, 0, :]  # CLS token
            logits = self.classifier(cls_embeddings)
            return cls_embeddings, logits

    model = BertForClassificationAndTriplet(bert_model, num_classes=len(label2id))
    state_dict = torch.load('./checkpoint/classification.pt')
    model.load_state_dict(state_dict)

    model = model.eval()

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model.to(device)

    pred_labels = []
    actual_labels = []


    dataloader = DataLoader(datasets, batch_size=BATCH_SIZE, shuffle=False, collate_fn=data_collator)
    for batch in tqdm(dataloader):
            # Inference in batches
            input_ids = batch['input_ids']
            attention_mask = batch['attention_mask']
            # save labels too
            actual_labels.extend(batch['labels'])
            

            # Move to GPU if available
            input_ids = input_ids.to(device)
            attention_mask = attention_mask.to(device)

            # Perform inference
            with torch.no_grad():
                cls, logits = model(
                    input_ids,
                    attention_mask)
                predicted_class_ids = logits.argmax(dim=1).to("cpu")
                pred_labels.extend(predicted_class_ids)

    pred_labels = [tensor.item() for tensor in pred_labels]


    # %%
    from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score, confusion_matrix
    y_true = actual_labels
    y_pred = pred_labels

    # Compute metrics
    accuracy = accuracy_score(y_true, y_pred)
    average_parameter = 'weighted'
    zero_division_parameter = 0
    f1 = f1_score(y_true, y_pred, average=average_parameter, zero_division=zero_division_parameter)
    precision = precision_score(y_true, y_pred, average=average_parameter, zero_division=zero_division_parameter)
    recall = recall_score(y_true, y_pred, average=average_parameter, zero_division=zero_division_parameter)

    with open("results/output.txt", "a") as f:

        print('*' * 80, file=f)
        # Print the results
        print(f'Accuracy: {accuracy:.5f}', file=f)
        print(f'F1 Score: {f1:.5f}', file=f)
        print(f'Precision: {precision:.5f}', file=f)
        print(f'Recall: {recall:.5f}', file=f)

    # export result
    label_list = [id2label[id] for id in pred_labels]
    df = pd.DataFrame({
        'class_prediction': pd.Series(label_list) 
    })

    # we can save the t5 generation output here
    df.to_csv(f"results/classify.csv", index=False)






# %%
# reset file before writing to it
with open("results/output.txt", "w") as f:
    print('', file=f)
    test()