Create server.py
Browse files- detector/server.py +155 -0
detector/server.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
from http.server import HTTPServer, SimpleHTTPRequestHandler
|
4 |
+
from multiprocessing import Process
|
5 |
+
import subprocess
|
6 |
+
from transformers import RobertaForSequenceClassification, RobertaTokenizer
|
7 |
+
import json
|
8 |
+
import fire
|
9 |
+
import torch
|
10 |
+
import re
|
11 |
+
from urllib.parse import urlparse, unquote, parse_qs, urlencode
|
12 |
+
|
13 |
+
model: RobertaForSequenceClassification = None
|
14 |
+
tokenizer: RobertaTokenizer = None
|
15 |
+
device: str = None
|
16 |
+
|
17 |
+
# Remove spaces query params from query
|
18 |
+
regex = r"__theme=(.+)"
|
19 |
+
|
20 |
+
|
21 |
+
def log(*args):
|
22 |
+
print(f"[{os.environ.get('RANK', '')}]", *args, file=sys.stderr)
|
23 |
+
|
24 |
+
|
25 |
+
class RequestHandler(SimpleHTTPRequestHandler):
|
26 |
+
|
27 |
+
def do_POST(self):
|
28 |
+
self.begin_content('application/json,charset=UTF-8')
|
29 |
+
|
30 |
+
content_length = int(self.headers['Content-Length'])
|
31 |
+
if content_length > 0:
|
32 |
+
post_data = self.rfile.read(content_length).decode('utf-8')
|
33 |
+
try:
|
34 |
+
post_data = json.loads(post_data)
|
35 |
+
|
36 |
+
if 'text' not in post_data:
|
37 |
+
self.wfile.write(json.dumps({"error": "missing key 'text'"}).encode('utf-8'))
|
38 |
+
else:
|
39 |
+
all_tokens, used_tokens, fake, real = self.infer(post_data['text'])
|
40 |
+
|
41 |
+
self.wfile.write(json.dumps(dict(
|
42 |
+
all_tokens=all_tokens,
|
43 |
+
used_tokens=used_tokens,
|
44 |
+
real_probability=real,
|
45 |
+
fake_probability=fake
|
46 |
+
)).encode('utf-8'))
|
47 |
+
|
48 |
+
except Exception as e:
|
49 |
+
self.wfile.write(json.dumps({"error": str(e)}).encode('utf-8'))
|
50 |
+
|
51 |
+
def do_GET(self):
|
52 |
+
query = urlparse(self.path).query
|
53 |
+
query = re.sub(regex, "", query, 0, re.MULTILINE)
|
54 |
+
query = unquote(query)
|
55 |
+
|
56 |
+
if not query:
|
57 |
+
self.begin_content('text/html')
|
58 |
+
|
59 |
+
html = os.path.join(os.path.dirname(__file__), 'index.html')
|
60 |
+
self.wfile.write(open(html).read().encode())
|
61 |
+
return
|
62 |
+
|
63 |
+
self.begin_content('application/json;charset=UTF-8')
|
64 |
+
|
65 |
+
all_tokens, used_tokens, fake, real = self.infer(query)
|
66 |
+
|
67 |
+
self.wfile.write(json.dumps(dict(
|
68 |
+
all_tokens=all_tokens,
|
69 |
+
used_tokens=used_tokens,
|
70 |
+
real_probability=real,
|
71 |
+
fake_probability=fake
|
72 |
+
)).encode())
|
73 |
+
|
74 |
+
def infer(self, query):
|
75 |
+
tokens = tokenizer.encode(query)
|
76 |
+
all_tokens = len(tokens)
|
77 |
+
tokens = tokens[:tokenizer.max_len - 2]
|
78 |
+
used_tokens = len(tokens)
|
79 |
+
tokens = torch.tensor([tokenizer.bos_token_id] + tokens + [tokenizer.eos_token_id]).unsqueeze(0)
|
80 |
+
mask = torch.ones_like(tokens)
|
81 |
+
|
82 |
+
with torch.no_grad():
|
83 |
+
logits = model(tokens.to(device), attention_mask=mask.to(device))[0]
|
84 |
+
probs = logits.softmax(dim=-1)
|
85 |
+
|
86 |
+
fake, real = probs.detach().cpu().flatten().numpy().tolist()
|
87 |
+
|
88 |
+
return all_tokens, used_tokens, fake, real
|
89 |
+
|
90 |
+
def begin_content(self, content_type):
|
91 |
+
self.send_response(200)
|
92 |
+
self.send_header('Content-Type', content_type)
|
93 |
+
self.send_header('Access-Control-Allow-Origin', '*')
|
94 |
+
self.end_headers()
|
95 |
+
|
96 |
+
def log_message(self, format, *args):
|
97 |
+
log(format % args)
|
98 |
+
|
99 |
+
|
100 |
+
def serve_forever(server, model, tokenizer, device):
|
101 |
+
log('Process has started; loading the model ...')
|
102 |
+
globals()['model'] = model.to(device)
|
103 |
+
globals()['tokenizer'] = tokenizer
|
104 |
+
globals()['device'] = device
|
105 |
+
|
106 |
+
log(f'Ready to serve at http://localhost:{server.server_address[1]}')
|
107 |
+
server.serve_forever()
|
108 |
+
|
109 |
+
|
110 |
+
def main(checkpoint, port=8080, device='cuda' if torch.cuda.is_available() else 'cpu'):
|
111 |
+
if checkpoint.startswith('gs://'):
|
112 |
+
print(f'Downloading {checkpoint}', file=sys.stderr)
|
113 |
+
subprocess.check_output(['gsutil', 'cp', checkpoint, '.'])
|
114 |
+
checkpoint = os.path.basename(checkpoint)
|
115 |
+
assert os.path.isfile(checkpoint)
|
116 |
+
|
117 |
+
print(f'Loading checkpoint from {checkpoint}')
|
118 |
+
data = torch.load(checkpoint, map_location='cpu')
|
119 |
+
|
120 |
+
model_name = 'roberta-large' if data['args']['large'] else 'roberta-base'
|
121 |
+
model = RobertaForSequenceClassification.from_pretrained(model_name)
|
122 |
+
tokenizer = RobertaTokenizer.from_pretrained(model_name)
|
123 |
+
|
124 |
+
model.load_state_dict(data['model_state_dict'])
|
125 |
+
model.eval()
|
126 |
+
|
127 |
+
print(f'Starting HTTP server on port {port}', file=sys.stderr)
|
128 |
+
server = HTTPServer(('0.0.0.0', port), RequestHandler)
|
129 |
+
|
130 |
+
# avoid calling CUDA API before forking; doing so in a subprocess is fine.
|
131 |
+
num_workers = int(subprocess.check_output([sys.executable, '-c', 'import torch; print(torch.cuda.device_count())']))
|
132 |
+
|
133 |
+
if num_workers <= 1:
|
134 |
+
serve_forever(server, model, tokenizer, device)
|
135 |
+
else:
|
136 |
+
print(f'Launching {num_workers} worker processes...')
|
137 |
+
|
138 |
+
subprocesses = []
|
139 |
+
|
140 |
+
for i in range(num_workers):
|
141 |
+
os.environ['RANK'] = f'{i}'
|
142 |
+
os.environ['CUDA_VISIBLE_DEVICES'] = f'{i}'
|
143 |
+
process = Process(target=serve_forever, args=(server, model, tokenizer, device))
|
144 |
+
process.start()
|
145 |
+
subprocesses.append(process)
|
146 |
+
|
147 |
+
del os.environ['RANK']
|
148 |
+
del os.environ['CUDA_VISIBLE_DEVICES']
|
149 |
+
|
150 |
+
for process in subprocesses:
|
151 |
+
process.join()
|
152 |
+
|
153 |
+
|
154 |
+
if __name__ == '__main__':
|
155 |
+
fire.Fire(main)
|