File size: 7,180 Bytes
9df4cc0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
import warnings
warnings.filterwarnings("ignore")

from sklearn.metrics import accuracy_score,f1_score
from datasets import load_dataset, load_from_disk, Dataset
from tqdm import tqdm
import datasets
import torch

from torch.utils.data import DataLoader
from functools import partial
from pathlib import Path

dic = {
        0:"negative",
        1:'neutral',
        2:'positive',
    }

with open(Path(__file__).parent / 'sentiment_templates.txt') as f:
    templates = [l.strip() for l in f.readlines()]
    

def format_example(example: dict) -> dict:
    context = f"Instruction: {example['instruction']}\n"
    if example.get("input"):
        context += f"Input: {example['input']}\n"
    context += "Answer: "
    target = example["output"]
    return {"context": context, "target": target}

def change_target(x):
    if 'positive' in x or 'Positive' in x:
        return 'positive'
    elif 'negative' in x or 'Negative' in x:
        return 'negative'
    else:
        return 'neutral'

       
def vote_output(x):
    output_dict = {'positive': 0, 'negative': 0, 'neutral': 0} 
    for i in range(len(templates)):
        pred = change_target(x[f'out_text_{i}'].lower())
        output_dict[pred] += 1
    if output_dict['positive'] > output_dict['negative']:
        return 'positive'
    elif output_dict['negative'] > output_dict['positive']:
        return 'negative'
    else:
        return 'neutral'
    
def test_fpb(args, model, tokenizer, prompt_fun=None):
    batch_size = args.batch_size
    # instructions = load_dataset("financial_phrasebank", "sentences_50agree")
    instructions = load_from_disk(Path(__file__).parent.parent / "data/financial_phrasebank-sentences_50agree/")
    instructions = instructions["train"]
    instructions = instructions.train_test_split(seed = 42)['test']
    instructions = instructions.to_pandas()
    instructions.columns = ["input", "output"]
    instructions["output"] = instructions["output"].apply(lambda x:dic[x])

    if prompt_fun is None:
        instructions["instruction"] = "What is the sentiment of this news? Please choose an answer from {negative/neutral/positive}."
    else:
        instructions["instruction"] = instructions.apply(prompt_fun, axis = 1)
    
    instructions[["context","target"]] = instructions.apply(format_example, axis = 1, result_type="expand")

    # print example
    print(f"\n\nPrompt example:\n{instructions['context'][0]}\n\n")


    context = instructions['context'].tolist()
    
    total_steps = instructions.shape[0]//batch_size + 1
    print(f"Total len: {len(context)}. Batchsize: {batch_size}. Total steps: {total_steps}")


    out_text_list = []
    for i in tqdm(range(total_steps)):
        tmp_context = context[i* batch_size:(i+1)* batch_size]
        tokens = tokenizer(tmp_context, return_tensors='pt', padding=True, max_length=512, return_token_type_ids=False)
        for k in tokens.keys():
            tokens[k] = tokens[k].cuda()
        res = model.generate(**tokens, max_length=512, eos_token_id=tokenizer.eos_token_id)
        res_sentences = [tokenizer.decode(i, skip_special_tokens=True) for i in res]
        # print(f'{i}: {res_sentences[0]}')
        out_text = [o.split("Answer: ")[1] for o in res_sentences]
        out_text_list += out_text
        torch.cuda.empty_cache()

    instructions["out_text"] = out_text_list
    instructions["new_target"] = instructions["target"].apply(change_target)
    instructions["new_out"] = instructions["out_text"].apply(change_target)

    acc = accuracy_score(instructions["new_target"], instructions["new_out"])
    f1_macro = f1_score(instructions["new_target"], instructions["new_out"], average = "macro")
    f1_micro = f1_score(instructions["new_target"], instructions["new_out"], average = "micro")
    f1_weighted = f1_score(instructions["new_target"], instructions["new_out"], average = "weighted")

    print(f"Acc: {acc}. F1 macro: {f1_macro}. F1 micro: {f1_micro}. F1 weighted (BloombergGPT): {f1_weighted}. ")

    return instructions


def test_fpb_mlt(args, model, tokenizer):
    batch_size = args.batch_size
    # dataset = load_dataset("financial_phrasebank", "sentences_50agree")
    dataset = load_from_disk(Path(__file__).parent.parent / 'data/financial_phrasebank-sentences_50agree/')
    dataset = dataset["train"]#.select(range(300))
    dataset = dataset.train_test_split(seed=42)['test']
    dataset = dataset.to_pandas()
    dataset.columns = ["input", "output"]
    dataset["output"] = dataset["output"].apply(lambda x: dic[x])
    dataset["text_type"] = dataset.apply(lambda x: 'news', axis=1)
    
    dataset["output"] = dataset["output"].apply(change_target)
    dataset = dataset[dataset["output"] != 'neutral']
    
    out_texts_list = [[] for _ in range(len(templates))]
    
    def collate_fn(batch):
        inputs = tokenizer(
            [f["context"] for f in batch], return_tensors='pt',
            padding=True, max_length=args.max_length,
            return_token_type_ids=False
        )
        return inputs
    
    for i, template in enumerate(templates):
        dataset = dataset[['input', 'output', "text_type"]]
        dataset["instruction"] = dataset['text_type'].apply(lambda x: template.format(type=x) + "\nOptions: positive, negative")
        # dataset["instruction"] = dataset['text_type'].apply(lambda x: template.format(type=x) + "\nOptions: negative, positive")
        dataset[["context", "target"]] = dataset.apply(format_example, axis=1, result_type="expand")
        
        dataloader = DataLoader(Dataset.from_pandas(dataset), batch_size=args.batch_size, collate_fn=collate_fn, shuffle=False)

        log_interval = len(dataloader) // 5

        for idx, inputs in enumerate(tqdm(dataloader)):
            inputs = {key: value.to(model.device) for key, value in inputs.items()}
            res = model.generate(**inputs, do_sample=False, max_length=args.max_length, eos_token_id=tokenizer.eos_token_id, max_new_tokens=10)
            res_sentences = [tokenizer.decode(i, skip_special_tokens=True) for i in res]
            tqdm.write(f'{idx}: {res_sentences[0]}')
            # if (idx + 1) % log_interval == 0:
            #     tqdm.write(f'{idx}: {res_sentences[0]}')
            out_text = [o.split("Answer: ")[1] for o in res_sentences]
            out_texts_list[i] += out_text
            torch.cuda.empty_cache()
            
    for i in range(len(templates)):
        dataset[f"out_text_{i}"] = out_texts_list[i]
        dataset[f"out_text_{i}"] = dataset[f"out_text_{i}"].apply(change_target)
    
    dataset["new_out"] = dataset.apply(vote_output, axis=1, result_type="expand")
    dataset.to_csv('tmp.csv')

    for k in [f"out_text_{i}" for i in range(len(templates))] + ["new_out"]:

        acc = accuracy_score(dataset["target"], dataset[k])
        f1_macro = f1_score(dataset["target"], dataset[k], average="macro")
        f1_micro = f1_score(dataset["target"], dataset[k], average="micro")
        f1_weighted = f1_score(dataset["target"], dataset[k], average="weighted")

        print(f"Acc: {acc}. F1 macro: {f1_macro}. F1 micro: {f1_micro}. F1 weighted (BloombergGPT): {f1_weighted}. ")

    return dataset