-
Notifications
You must be signed in to change notification settings - Fork 0
/
histogram.py
172 lines (162 loc) · 5.95 KB
/
histogram.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
from transformers import AutoTokenizer, AutoModelForCausalLM
from datasets import load_dataset
from accelerate import Accelerator
import numpy as np
import argparse
import glob
import os
import torch
import json
from tqdm import tqdm
import re
from torch.nn import CrossEntropyLoss
def remove_emojis(text):
# This regex pattern targets common emoji ranges in Unicode
emoji_pattern = re.compile(
"["
"\U0001F600-\U0001F64F" # Emoticons
"\U0001F300-\U0001F5FF" # Symbols & Pictographs
"\U0001F680-\U0001F6FF" # Transport & Map Symbols
"\U0001F700-\U0001F77F" # Alchemical Symbols
"\U0001F780-\U0001F7FF" # Geometric Shapes Extended
"\U0001F800-\U0001F8FF" # Supplemental Arrows-C
"\U0001F900-\U0001F9FF" # Supplemental Symbols and Pictographs
"\U0001FA00-\U0001FA6F" # Chess Symbols
"\U0001FA70-\U0001FAFF" # Symbols and Pictographs Extended-A
"\U00002702-\U000027B0" # Dingbats
"\U000024C2-\U0001F251"
"]+",
flags=re.UNICODE,
)
return emoji_pattern.sub(r"", text)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument(
"--dataset_name_or_path",
type=str,
default="facebook/flores",
help="Name or path to the HF dataset.",
)
parser.add_argument("--config", type=str, help="Config for HF datasets.")
parser.add_argument(
"--split",
type=str,
default="dev",
help="Name of the split to consider e.g. 'train'",
)
parser.add_argument(
"--model_name_or_path", type=str, default="google/gemma-2b", help=""
)
parser.add_argument(
"--column_name",
type=str,
default="sentence",
help="Name of the column of interest.",
)
parser.add_argument(
"--request_batch_size", type=int, default=32, help="Batch size."
)
parser.add_argument(
"--max_length",
type=int,
default=128,
help="Maximum sequence length (number of tokens)",
)
parser.add_argument(
"--max_samples",
type=int,
default=1000000,
help="Maximum number of samples to evaluate.",
)
parser.add_argument("--seed", type=int, default=122, help="seed")
parser.add_argument(
"--output_filename", type=str, help="Name or path to the output file."
)
parser.add_argument(
"--add_start_token",
action="store_true",
help="Whether to add a bos token during the perplexity computation.",
)
return parser.parse_args()
def _per_token_loss(inputs, logits):
"""
https://huggingface.co/learn/nlp-course/chapter7/6?fw=pt
return a loss of size (batch size, sequence length)
"""
# Shift so that tokens < n predict n
shift_labels = inputs[..., 1:].contiguous()
shift_logits = logits[..., :-1, :].contiguous()
# Calculate per-token loss
loss_fn = CrossEntropyLoss(reduction="none")
loss = loss_fn(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
loss = loss.view(shift_logits.size(0), shift_logits.size(1))
"""
loss = loss_fn(shift_logits.transpose(1, 2), shift_labels)
"""
return loss
if __name__ == "__main__":
args = parse_args()
rng = np.random.default_rng(args.seed)
dataset = load_dataset(args.dataset_name_or_path, args.config, split=args.split)
samples = [example[args.column_name] for example in dataset]
samples = [remove_emojis(sample) for sample in samples]
tokenizer = AutoTokenizer.from_pretrained(args.model_name_or_path)
model = AutoModelForCausalLM.from_pretrained(
args.model_name_or_path,
device_map={"": Accelerator().process_index},
torch_dtype=torch.float16,
)
if args.output_filename:
output_filename = args.output_filename
else:
output_filename = f"{args.dataset_name_or_path.split('/')[-1]}_{args.config}_{args.split}_seed_{args.seed}.jsonl"
start = 0
if os.path.exists(output_filename):
with open(output_filename, "r") as fin:
for line in fin:
start += 1
print("Start ...")
device = model.device
for i in tqdm(range(start, len(samples), args.request_batch_size)):
prompts = samples[i : i + args.request_batch_size]
# print("\n".join([f"{k}: {e}" for k, e in enumerate(prompts)]))
inputs = tokenizer(
prompts,
return_tensors="pt",
padding="max_length",
truncation=True,
max_length=args.max_length,
)
inputs = inputs.to(device)
if args.add_start_token:
bos_tokens_tensor = torch.tensor(
[[tokenizer.bos_token_id]] * inputs["input_ids"].size(dim=0)
).to(device)
inputs["input_ids"] = torch.cat(
[bos_tokens_tensor, inputs["input_ids"]], dim=1
)
inputs["attention_mask"] = torch.cat(
[
torch.ones(bos_tokens_tensor.size(), dtype=torch.int64).to(device),
inputs["attention_mask"],
],
dim=1,
)
# labels = inputs["input_ids"]
# outputs = model(**inputs, labels=labels)
with torch.no_grad():
outputs = model(**inputs)
per_token_loss = _per_token_loss(inputs["input_ids"], outputs.logits)
shift_attention_mask = inputs["attention_mask"][..., 1:].contiguous()
per_sample_loss = per_token_loss * shift_attention_mask
per_sample_loss = per_sample_loss.sum(-1) / shift_attention_mask.sum(-1)
perplexity = torch.exp(per_sample_loss)
# print(f"First sentence perplexity: {perplexity[0].item()}")
with open(output_filename, "a") as fout:
for j in range(len(prompts)):
fout.write(
json.dumps(
{"sample": prompts[j], "perplexity": perplexity[j].item()}
)
+ "\n"
)