70 lines
2.1 KiB
Python
70 lines
2.1 KiB
Python
"""
|
|
Query the tensorflow_model_server's REST API.
|
|
"""
|
|
|
|
import logging
|
|
from typing import Optional, Union
|
|
|
|
import aiohttp
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
class TensorFlow:
|
|
"""
|
|
Fetch an embedding vector from the tensorflow model server.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
tf_config,
|
|
session: aiohttp.ClientSession,
|
|
timeout_sock_connect: Union[int, float] = 0.5,
|
|
timeout_sock_read: Union[int, float] = 10,
|
|
):
|
|
self.config = tf_config
|
|
self.session = session
|
|
self.timeout = aiohttp.ClientTimeout(
|
|
sock_connect=timeout_sock_connect, sock_read=timeout_sock_read
|
|
)
|
|
|
|
async def embed(
|
|
self, text: Union[str, list[str]]
|
|
) -> Optional[Union[list[float], list[list[float]]]]:
|
|
"""
|
|
Query the tensorflow_model_server's REST API for a prediction.
|
|
|
|
Take a string or a list of strings and return an embedding vector
|
|
or a list of embedding vectors.
|
|
|
|
If the request fails or times out, return None.
|
|
"""
|
|
text_ = text if isinstance(text, list) else [text]
|
|
data = {'signature_name': 'serving_default', 'instances': text_}
|
|
try:
|
|
async with self.session.post(
|
|
self.config['model_server_endpoint'],
|
|
json=data,
|
|
timeout=self.timeout,
|
|
) as resp:
|
|
try:
|
|
res = await resp.json()
|
|
if isinstance(text, list):
|
|
return res.get('predictions')
|
|
else:
|
|
return res.get('predictions')[0]
|
|
except:
|
|
msg = 'Got invalid response from tensorflow'
|
|
logger.error(msg)
|
|
return None
|
|
except Exception as err:
|
|
msg = 'Could not get embedding from tensorflow for '
|
|
if isinstance(text, str):
|
|
msg += f'string of length {len(text)}'
|
|
else:
|
|
msg += 'list of strings with lengths '
|
|
msg += ','.join([str(len(s)) for s in text])
|
|
msg += f', reason: {err}'
|
|
logger.error(msg)
|
|
return None
|