Spaces:
Running
Running
Upload image_captioner.py
Browse files- image_captioner.py +75 -0
image_captioner.py
ADDED
@@ -0,0 +1,75 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Optional
|
2 |
+
import requests
|
3 |
+
|
4 |
+
import logging
|
5 |
+
|
6 |
+
from haystack import Document, component
|
7 |
+
from haystack.lazy_imports import LazyImport
|
8 |
+
from PIL import Image
|
9 |
+
|
10 |
+
logger = logging.getLogger(__name__)
|
11 |
+
|
12 |
+
with LazyImport(message="Run 'pip install transformers[torch,sentencepiece]'") as torch_and_transformers_import:
|
13 |
+
import torch
|
14 |
+
from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer, BlipProcessor, BlipForConditionalGeneration
|
15 |
+
from PIL import Image
|
16 |
+
|
17 |
+
@component
|
18 |
+
class ImageCaptioner:
|
19 |
+
def __init__(
|
20 |
+
self,
|
21 |
+
model_name: str = "Salesforce/blip-image-captioning-base",
|
22 |
+
):
|
23 |
+
torch_and_transformers_import.check()
|
24 |
+
self.model_name = model_name
|
25 |
+
|
26 |
+
if model_name == "nlpconnect/vit-gpt2-image-captioning":
|
27 |
+
self.model = VisionEncoderDecoderModel.from_pretrained(model_name)
|
28 |
+
self.feature_extractor = ViTImageProcessor.from_pretrained(model_name)
|
29 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
30 |
+
max_length = 16
|
31 |
+
num_beams = 4
|
32 |
+
self.gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
|
33 |
+
else:
|
34 |
+
self.processor = BlipProcessor.from_pretrained(model_name)
|
35 |
+
self.model = BlipForConditionalGeneration.from_pretrained(model_name)
|
36 |
+
|
37 |
+
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
38 |
+
self.model.to(self.device)
|
39 |
+
|
40 |
+
@component.output_types(captions=List[str])
|
41 |
+
def run(self, image_file_paths: List[str]) -> List[Document]:
|
42 |
+
|
43 |
+
images = []
|
44 |
+
for image_path in image_file_paths:
|
45 |
+
i_image = Image.open(image_path)
|
46 |
+
if i_image.mode != "RGB":
|
47 |
+
i_image = i_image.convert(mode="RGB")
|
48 |
+
|
49 |
+
images.append(i_image)
|
50 |
+
|
51 |
+
preds = []
|
52 |
+
if self.model_name == "nlpconnect/vit-gpt2-image-captioning":
|
53 |
+
pixel_values = self.feature_extractor(images=images, return_tensors="pt").pixel_values
|
54 |
+
pixel_values = pixel_values.to(self.device)
|
55 |
+
|
56 |
+
output_ids = self.model.generate(pixel_values, **self.gen_kwargs)
|
57 |
+
|
58 |
+
preds = self.tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
59 |
+
preds = [pred.strip() for pred in preds]
|
60 |
+
else:
|
61 |
+
|
62 |
+
inputs = self.processor(images, return_tensors="pt")
|
63 |
+
output_ids = self.model.generate(**inputs)
|
64 |
+
preds = self.processor.batch_decode(output_ids, skip_special_tokens=True)
|
65 |
+
preds = [pred.strip() for pred in preds]
|
66 |
+
|
67 |
+
# captions: List[Document] = []
|
68 |
+
# for caption, image_file_path in zip(preds, image_file_paths):
|
69 |
+
# document = Document(content=caption, meta={"image_path": image_file_path})
|
70 |
+
# captions.append(document)
|
71 |
+
return {"captions": preds}
|
72 |
+
|
73 |
+
# captioner = ImageCaptioner(model_name="Salesforce/blip-image-captioning-base")
|
74 |
+
# result = captioner.run(image_file_paths=["selfie.png"])
|
75 |
+
# print(result)
|