Andrew Luo
commited on
Commit
•
c98f9e5
1
Parent(s):
07655e0
customer handler
Browse files- handler.py +35 -0
- requirements.txt +0 -0
handler.py
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
+
from transformers import AutoModelForMaskedLM, AutoTokenizer
|
3 |
+
import torch
|
4 |
+
|
5 |
+
|
6 |
+
class EndpointHandler():
|
7 |
+
def __init__(self, path=""):
|
8 |
+
tokenizer = AutoTokenizer.from_pretrained(path)
|
9 |
+
model = AutoModelForMaskedLM.from_pretrained(path)
|
10 |
+
self.tokenizer = tokenizer
|
11 |
+
self.model = model
|
12 |
+
|
13 |
+
def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
|
14 |
+
"""
|
15 |
+
data args:
|
16 |
+
inputs (:obj: `str`)
|
17 |
+
date (:obj: `str`)
|
18 |
+
Return:
|
19 |
+
A :obj:`list` | `dict`: will be serialized and returned
|
20 |
+
"""
|
21 |
+
# get inputs
|
22 |
+
tokens = self.tokenizer(text, return_tensors='pt')
|
23 |
+
output = self.model(**tokens)
|
24 |
+
vec = torch.max(
|
25 |
+
torch.log(
|
26 |
+
1 + torch.relu(output.logits)
|
27 |
+
) * tokens.attention_mask.unsqueeze(-1),
|
28 |
+
dim=1)[0].squeeze()
|
29 |
+
instruction = data.pop("instruction", data)
|
30 |
+
cols = vec.nonzero().squeeze().cpu().tolist()
|
31 |
+
# extract the non-zero values
|
32 |
+
weights = vec[cols].cpu().tolist()
|
33 |
+
# use to create a dictionary of token ID to weight
|
34 |
+
sparse_dict = dict(zip(cols, weights))
|
35 |
+
return sparse_dict
|
requirements.txt
ADDED
File without changes
|