Upload folder using huggingface_hub
Browse files- pipeline.py +25 -0
pipeline.py
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import pipeline, Pipeline, AutoTokenizer, AutoConfig, AutoModelForSequenceClassification
|
2 |
+
from transformers.pipelines import PIPELINE_REGISTRY
|
3 |
+
import torch
|
4 |
+
|
5 |
+
|
6 |
+
class SpanClassificationPipeline(Pipeline):
|
7 |
+
def __init__(self, model, tokenizer, device="cpu", **kwargs):
|
8 |
+
super().__init__(model=model, tokenizer=tokenizer, device=device, **kwargs)
|
9 |
+
self.model.to(self.device)
|
10 |
+
self.model.eval()
|
11 |
+
|
12 |
+
def _sanitize_parameters(self, **kwargs):
|
13 |
+
return {}, kwargs, {}
|
14 |
+
|
15 |
+
def preprocess(self, inputs):
|
16 |
+
return self.tokenizer(inputs, return_tensors="pt").to(self.device)
|
17 |
+
|
18 |
+
def _forward(self, model_inputs):
|
19 |
+
with torch.no_grad():
|
20 |
+
outputs = self.model(**model_inputs)
|
21 |
+
return outputs
|
22 |
+
|
23 |
+
def postprocess(self, model_outputs):
|
24 |
+
logits = model_outputs.logits
|
25 |
+
return int(torch.argmax(logits, dim=1).item())
|