-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_classifiers.py
357 lines (306 loc) · 13.8 KB
/
eval_classifiers.py
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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
from openai import OpenAI
from together import Together
import torch
from transformers import AutoModelForSequenceClassification, XLMRobertaForTokenClassification, AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
from utils import *
from length_mapping import *
from datasets import load_dataset
import evaluate
import re
import ast
import os
import numpy as np
import pandas as pd
from tqdm import tqdm
import random
class GPT():
def __init__(self, task):
self.client = OpenAI(
organization='org-OWe4x5PWbYh2yljPFA9odbCD',
)
self.task = task
def forward(self, x, k, cot=False):
if cot:
messages = create_messages_cot_classifier(self.task)
else:
messages = create_messages_classifier(self.task, k)
x = str(x)
messages.append({"role": "user", "content": x})
completion = self.client.chat.completions.create(
model="gpt-4-turbo",
messages=messages,
response_format={"type": "json_object"}
)
return completion.choices[0].message.content
class LLaMA():
def __init__(self, task):
self.client = Together()
self.task = task
def forward(self, x):
messages = create_messages_classifier(self.task, 0)
x = str(x)
messages.append({"role": "user", "content": x})
response = self.client.chat.completions.create(
model="meta-llama/Llama-3-8b-chat-hf",
messages=messages,
temperature=0.6,
top_p=0.9,
)
return response.choices[0].message.content
class LLaMAFT():
def __init__(self, task, device):
self.task = task
self.device = device
self.bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16
)
self.tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")
self.model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct", quantization_config=self.bnb_config)
self.model = PeftModel.from_pretrained(self.model, f"ULRs/llama-3-8b-{task}-ur").to(self.device)
def format_input(self, x):
input = f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n\n{LLAMA_FT_PROMPTS[self.task]}<|eot_id|><|start_header_id|>user<|end_header_id|>\n\n{x}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
return input
def forward(self, x):
x = self.format_input(x)
input_ids = self.tokenizer(x, return_tensors="pt").to(self.device)
terminators = [
self.tokenizer.eos_token_id,
self.tokenizer.convert_tokens_to_ids("<|eot_id|>"),
]
outputs = self.model.generate(
**input_ids,
max_new_tokens=50,
temperature=0.6,
top_p=0.9,
pad_token_id=self.tokenizer.eos_token_id,
eos_token_id=terminators
)
output = self.tokenizer.decode(outputs[0])
return output.split("assistant<|end_header_id|>\n\n")[1].replace("<|eot_id|>", "")
class XLMR():
def __init__(self, task, device):
self.device = device
self.task = task
if task == "ner-tagging" or task == "pos-tagging":
self.model = XLMRobertaForTokenClassification.from_pretrained(f"ULRs/xlm-roberta-large-{self.task}-ur").to(self.device)
else:
self.model = AutoModelForSequenceClassification.from_pretrained(f"ULRs/xlm-roberta-large-{self.task}-ur").to(self.device)
self.tokenizer = AutoTokenizer.from_pretrained(f"ULRs/xlm-roberta-large-{self.task}-ur")
def forward(self, x):
if self.task == "ner-tagging" or self.task == "pos-tagging":
inputs = self.tokenizer(x, add_special_tokens=False, is_split_into_words=True, return_tensors="pt")
else:
inputs = self.tokenizer(x, max_length=512, truncation=True, return_tensors="pt")
with torch.no_grad():
logits = self.model(**inputs.to(self.device)).logits
output = np.argmax(logits.cpu().detach().numpy(), axis=-1)[0]
if (self.task == "ner-tagging" or self.task == "pos-tagging") and len(output) != len(x):
new_output = []
prev_token = None
for i, token_id in enumerate(inputs.input_ids[0]):
token = self.tokenizer.decode(token_id)
if token != x[0]:
if token == "االله" and x[0] == "اﷲ":
x = x[1:]
new_output.append(output[i])
elif prev_token is None:
prev_token = token
elif prev_token + token == x[0]:
x = x[1:]
new_output.append(output[i])
prev_token = None
elif prev_token + token == "صلىاللهعليهوسلم":
x = x[1:]
new_output.append(output[i])
prev_token = None
elif prev_token + token == "اسلامصلىاللهعليهوسلم":
x = x[1:]
new_output.append(output[i])
prev_token = None
elif prev_token + token == "عبداالله":
x = x[1:]
new_output.append(output[i])
prev_token = None
elif prev_token + token == "لايٴن":
x = x[1:]
new_output.append(output[i])
prev_token = None
else:
prev_token += token
else:
x = x[1:]
new_output.append(output[i])
return new_output
else:
return output
class Metric():
def __init__(self):
self.accuracy = evaluate.load("accuracy")
self.f1_score = evaluate.load("f1")
def calculate_accuracy(self, predictions, references):
return self.accuracy.compute(predictions=predictions, references=references)
def calculate_f1_score(self, predictions, references):
return self.f1_score.compute(predictions=predictions, references=references, average="macro")
class Evaluation():
def __init__(self, task, k_shot, cot, to_eval, device):
self.task = task
if not os.path.exists(f"predictions/{self.task}"):
os.mkdir(f"predictions/{self.task}")
self.k_shot = k_shot
self.cot = cot
self.to_eval = to_eval
self.dataset = load_dataset("csv", data_files={"test": f"datasets/{self.task}.csv"})
if self.to_eval == "gpt":
self.gpt = GPT(self.task)
elif self.to_eval == "llama":
self.llama = LLaMA(self.task)
elif self.to_eval == "llama-ft":
self.llama_ft = LLaMAFT(self.task, device)
elif self.to_eval == "xlm-roberta":
self.xlmr = XLMR(self.task, device)
self.metric = Metric()
def extract_label(self, output):
try:
output = ast.literal_eval(output)["label"]
except:
try:
temp = "{" + re.search(rf"['\"]label['\"]\s*:\s*['\"].*?['\"]}}", output).group(0)
output = ast.literal_eval(temp)["label"]
except:
if self.task == "abuse-detection":
output = check_abusive_status(output)
elif self.task == "sarcasm-detection":
output = check_sarcastic_status(output)
elif self.task == "fake-news":
output = check_fake_news_status(output)
else:
labels = label_mapping[self.task]
pred_label = None
for label in labels:
if label in output.lower():
if pred_label is None:
output = str({'label': label})
pred_label = label
else:
output = ""
break
try:
output = ast.literal_eval(output)["label"]
except:
output = ""
if output != "":
try:
prediction = label_mapping[self.task][output.lower()]
except:
prediction = -1
else:
prediction = -1
return prediction
def eval_gpt(self, input, i):
for k in [0, 3, 6]:
if k not in self.k_shot:
continue
if os.path.exists(f"predictions/{self.task}/gpt-{k}-shot.csv"):
df = pd.read_csv(f"predictions/{self.task}/gpt-{k}-shot.csv")
else:
df = pd.DataFrame(columns=["prediction"])
if len(df) > i and df.iloc[i]["prediction"] != -1:
continue
if self.task == "ner-tagging" or self.task == "pos-tagging":
output = self.gpt.forward(highlight_word(input, self.dataset["test"][i]["index"]), k)
else:
output = self.gpt.forward(input, k)
prediction = self.extract_label(output)
df.loc[i] = [prediction]
df.to_csv(f"predictions/{self.task}/gpt-{k}-shot.csv", index=False)
if self.cot:
if os.path.exists(f"predictions/{self.task}/gpt-cot.csv"):
df = pd.read_csv(f"predictions/{self.task}/gpt-cot.csv")
else:
df = pd.DataFrame(columns=["prediction"])
if len(df) > i and df.iloc[i]["prediction"] != -1:
pass
else:
if self.task == "ner-tagging" or self.task == "pos-tagging":
output = self.gpt.forward(highlight_word(input, self.dataset["test"][i]["index"]), k, cot=True)
else:
output = self.gpt.forward(input, k, cot=True)
prediction = self.extract_label(output)
df.loc[i] = [prediction]
df.to_csv(f"predictions/{self.task}/gpt-cot.csv", index=False)
def eval_llama(self, input, i):
if os.path.exists(f"predictions/{self.task}/llama.csv"):
df = pd.read_csv(f"predictions/{self.task}/llama.csv")
else:
df = pd.DataFrame(columns=["prediction"])
if len(df) > i and df.iloc[i]["prediction"] != -1:
pass
else:
if self.task == "ner-tagging" or self.task == "pos-tagging":
output = self.llama.forward(highlight_word(input, self.dataset["test"][i]["index"]))
else:
output = self.llama.forward(input)
prediction = self.extract_label(output)
df.loc[i] = [prediction]
df.to_csv(f"predictions/{self.task}/llama.csv", index=False)
def eval_llama_ft(self, input, i):
if os.path.exists(f"predictions/{self.task}/llama-ft.csv"):
df = pd.read_csv(f"predictions/{self.task}/llama-ft.csv")
else:
df = pd.DataFrame(columns=["prediction"])
if len(df) > i and df.iloc[i]["prediction"] != -1:
pass
else:
if self.task == "ner-tagging" or self.task == "pos-tagging":
output = self.llama_ft.forward(highlight_word(input, self.dataset["test"][i]["index"]))
else:
output = self.llama_ft.forward(input)
prediction = self.extract_label(output)
df.loc[i] = [prediction]
df.to_csv(f"predictions/{self.task}/llama-ft.csv", index=False)
def eval_xlmr(self, input, i):
if os.path.exists(f"predictions/{self.task}/xlm-roberta.csv"):
df = pd.read_csv(f"predictions/{self.task}/xlm-roberta.csv")
else:
df = pd.DataFrame(columns=["prediction"])
if len(df) > i and type(df.iloc[i]["prediction"]) == np.int64 and df.iloc[i]["prediction"] != -1:
pass
else:
output = self.xlmr.forward(input)
prediction = output
df.loc[i] = prediction[self.dataset["test"][i]]
df.to_csv(f"predictions/{self.task}/xlm-roberta.csv", index=False)
def eval(self):
for i in tqdm(range(len(self.dataset["test"]))):
if self.task == "ner-tagging" or self.task == "pos-tagging":
input = string_to_list(self.dataset["test"][i]["input"])
else:
input = self.dataset["test"][i]["input"]
if self.to_eval == "gpt":
self.eval_gpt(input, i)
if self.to_eval == "llama":
self.eval_llama(input, i)
if self.to_eval == "llama-ft":
self.eval_llama_ft(input, i)
if self.to_eval == "xlm-roberta":
self.eval_xlmr(input, i)
def calcuate_metric(self):
f1 = {}
accuracy = {}
references = self.dataset["test"]["label"]
for model in ["gpt-0-shot", "gpt-3-shot", "gpt-6-shot", "gpt-cot", "llama", "llama-ft", "xlm-roberta"]:
if os.path.exists(f"predictions/{self.task}/{model}.csv"):
df = pd.read_csv(f"predictions/{self.task}/{model}.csv")
predictions = df["prediction"].tolist()
f1[model] = self.metric.calculate_f1_score(predictions, references)
accuracy[model] = self.metric.calculate_accuracy(predictions, references)
return f1, accuracy
ev = Evaluation("sarcasm-detection", [], False, "llama", "mps")
# ev.eval()
f1, accuracy = ev.calcuate_metric()
print(f1)
print("=========================")
# print(accuracy)