Commit
•
65062c0
1
Parent(s):
72a7d56
Create inference_c4.py
Browse files- inference_c4.py +155 -0
inference_c4.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# !pip install -q transformers datasets sentencepiece
|
2 |
+
import argparse
|
3 |
+
import gc
|
4 |
+
import json
|
5 |
+
import os
|
6 |
+
|
7 |
+
import datasets
|
8 |
+
import pandas as pd
|
9 |
+
import torch
|
10 |
+
from tqdm import tqdm
|
11 |
+
from transformers import AutoModelForSequenceClassification, AutoTokenizer
|
12 |
+
|
13 |
+
TOTAL_NUM_FILES_C4_TRAIN = 1024
|
14 |
+
|
15 |
+
|
16 |
+
def parse_args():
|
17 |
+
parser = argparse.ArgumentParser()
|
18 |
+
|
19 |
+
parser.add_argument(
|
20 |
+
"--start",
|
21 |
+
type=int,
|
22 |
+
required=True,
|
23 |
+
help="Starting file number to download. Valid values: 0 - 1023",
|
24 |
+
)
|
25 |
+
parser.add_argument(
|
26 |
+
"--end",
|
27 |
+
type=int,
|
28 |
+
required=True,
|
29 |
+
help="Ending file number to download. Valid values: 0 - 1023",
|
30 |
+
)
|
31 |
+
parser.add_argument("--batch_size", type=int, default=16, help="Batch size")
|
32 |
+
parser.add_argument(
|
33 |
+
"--model_name",
|
34 |
+
type=str,
|
35 |
+
default="taskydata/deberta-v3-base_10xp3nirstbbflanseuni_10xc4",
|
36 |
+
help="Model name",
|
37 |
+
)
|
38 |
+
parser.add_argument(
|
39 |
+
"--local_cache_location",
|
40 |
+
type=str,
|
41 |
+
default="c4_download",
|
42 |
+
help="local cache location from where the dataset will be loaded",
|
43 |
+
)
|
44 |
+
parser.add_argument(
|
45 |
+
"--use_local_cache_location",
|
46 |
+
type=bool,
|
47 |
+
default=True,
|
48 |
+
help="Set True if you want to load the dataset from local cache.",
|
49 |
+
)
|
50 |
+
parser.add_argument(
|
51 |
+
"--clear_dataset_cache",
|
52 |
+
type=bool,
|
53 |
+
default=False,
|
54 |
+
help="Set True if you want to delete the dataset files from the cache after inference.",
|
55 |
+
)
|
56 |
+
parser.add_argument(
|
57 |
+
"--release_memory",
|
58 |
+
type=bool,
|
59 |
+
default=True,
|
60 |
+
help="Set True if you want to release the memory of used variables.",
|
61 |
+
)
|
62 |
+
|
63 |
+
args = parser.parse_args()
|
64 |
+
return args
|
65 |
+
|
66 |
+
|
67 |
+
def chunks(l, n):
|
68 |
+
for i in range(0, len(l), n):
|
69 |
+
yield l[i : i + n]
|
70 |
+
|
71 |
+
|
72 |
+
def batch_tokenize(data, batch_size):
|
73 |
+
batches = list(chunks(data, batch_size))
|
74 |
+
tokenized_batches = []
|
75 |
+
for batch in batches:
|
76 |
+
# max_length will automatically be set to the max length of the model (512 for deberta)
|
77 |
+
tensor = tokenizer(
|
78 |
+
batch,
|
79 |
+
return_tensors="pt",
|
80 |
+
padding="max_length",
|
81 |
+
truncation=True,
|
82 |
+
max_length=512,
|
83 |
+
)
|
84 |
+
tokenized_batches.append(tensor)
|
85 |
+
return tokenized_batches, batches
|
86 |
+
|
87 |
+
|
88 |
+
def batch_inference(data, batch_size=16):
|
89 |
+
preds = []
|
90 |
+
tokenized_batches, batches = batch_tokenize(data, batch_size)
|
91 |
+
for i in tqdm(range(len(batches))):
|
92 |
+
with torch.no_grad():
|
93 |
+
logits = model(**tokenized_batches[i].to(device)).logits.cpu()
|
94 |
+
preds.extend(logits)
|
95 |
+
return preds
|
96 |
+
|
97 |
+
|
98 |
+
if __name__ == "__main__":
|
99 |
+
args = parse_args()
|
100 |
+
|
101 |
+
tokenizer = AutoTokenizer.from_pretrained(args.model_name)
|
102 |
+
model = AutoModelForSequenceClassification.from_pretrained(args.model_name)
|
103 |
+
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
|
104 |
+
model.to(device)
|
105 |
+
|
106 |
+
|
107 |
+
if args.use_local_cache_location:
|
108 |
+
file_name = f"c4-train.{global_id}.json.gz"
|
109 |
+
data_files = {"train": f"{args.local_cache_location}/{file_name}"}
|
110 |
+
c4 = datasets.load_dataset("json", data_files=data_files, split="train")
|
111 |
+
else:
|
112 |
+
file_name = f"en/c4-train.{global_id}.json.gz"
|
113 |
+
data_files = {"train": file_name}
|
114 |
+
c4 = datasets.load_dataset(
|
115 |
+
"allenai/c4", data_files=data_files, split="train"
|
116 |
+
)
|
117 |
+
df = pd.DataFrame(c4, index=None)
|
118 |
+
texts = df["text"].to_list()
|
119 |
+
preds = batch_inference(texts, batch_size=args.batch_size)
|
120 |
+
|
121 |
+
assert len(preds) == len(texts)
|
122 |
+
|
123 |
+
# Write two jsonl files:
|
124 |
+
# 1) Probas for all of C4
|
125 |
+
# 2) Probas + texts for samples predicted as tasky
|
126 |
+
df['timestamp'] = df['timestamp'].astype(str)
|
127 |
+
with open(c4taskyprobas_path, "w") as f, open(c4tasky_path, "w") as g:
|
128 |
+
for i in range(len(preds)):
|
129 |
+
predicted_class_id = preds[i].argmax().item()
|
130 |
+
pred = model.config.id2label[predicted_class_id]
|
131 |
+
tasky_proba = torch.softmax(preds[i], dim=-1)[-1].item()
|
132 |
+
f.write(json.dumps({"proba": tasky_proba}) + "\n")
|
133 |
+
# If it's tasky, save!
|
134 |
+
if int(predicted_class_id) == 1:
|
135 |
+
g.write(
|
136 |
+
json.dumps(
|
137 |
+
{
|
138 |
+
"proba": tasky_proba,
|
139 |
+
"text": texts[i],
|
140 |
+
"timestamp": df["timestamp"][i],
|
141 |
+
"url": df["url"][i],
|
142 |
+
}
|
143 |
+
)
|
144 |
+
+ "\n"
|
145 |
+
)
|
146 |
+
# release memory
|
147 |
+
if args.release_memory:
|
148 |
+
del preds
|
149 |
+
del texts
|
150 |
+
del df
|
151 |
+
gc.collect()
|
152 |
+
|
153 |
+
# Delete the processed dataset file from the cache
|
154 |
+
if args.clear_dataset_cache:
|
155 |
+
os.remove(f"{args.local_cache_location}/{file_name}")
|