Commit
·
61d1cb6
1
Parent(s):
53eb83f
Simplify the names.
Browse files
vit_gpt2_image_caption.py
ADDED
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# https://huggingface.co/nlpconnect/vit-gpt2-image-captioning
|
2 |
+
|
3 |
+
import urllib.request
|
4 |
+
import modal
|
5 |
+
|
6 |
+
stub = modal.Stub("vit-gpt2-image-captioning")
|
7 |
+
volume = modal.SharedVolume().persist("shared_vol")
|
8 |
+
|
9 |
+
@stub.function(
|
10 |
+
gpu="any",
|
11 |
+
image=modal.Image.debian_slim().pip_install("Pillow", "transformers", "torch"),
|
12 |
+
shared_volumes={"/root/model_cache": volume},
|
13 |
+
retries=3,
|
14 |
+
)
|
15 |
+
def predict(image):
|
16 |
+
import io
|
17 |
+
from transformers import VisionEncoderDecoderModel, ViTImageProcessor, AutoTokenizer
|
18 |
+
import torch
|
19 |
+
from PIL import Image
|
20 |
+
|
21 |
+
model = VisionEncoderDecoderModel.from_pretrained(
|
22 |
+
"nlpconnect/vit-gpt2-image-captioning"
|
23 |
+
)
|
24 |
+
feature_extractor = ViTImageProcessor.from_pretrained(
|
25 |
+
"nlpconnect/vit-gpt2-image-captioning"
|
26 |
+
)
|
27 |
+
tokenizer = AutoTokenizer.from_pretrained("nlpconnect/vit-gpt2-image-captioning")
|
28 |
+
|
29 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
30 |
+
model.to(device)
|
31 |
+
|
32 |
+
max_length = 16
|
33 |
+
num_beams = 4
|
34 |
+
gen_kwargs = {"max_length": max_length, "num_beams": num_beams}
|
35 |
+
input_img = Image.open(io.BytesIO(image))
|
36 |
+
pixel_values = feature_extractor(
|
37 |
+
images=[input_img], return_tensors="pt"
|
38 |
+
).pixel_values
|
39 |
+
pixel_values = pixel_values.to(device)
|
40 |
+
|
41 |
+
output_ids = model.generate(pixel_values, **gen_kwargs)
|
42 |
+
|
43 |
+
preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
|
44 |
+
preds = [pred.strip() for pred in preds]
|
45 |
+
return preds
|
46 |
+
|
47 |
+
|
48 |
+
@stub.local_entrypoint()
|
49 |
+
def main():
|
50 |
+
from pathlib import Path
|
51 |
+
|
52 |
+
image_filepath = Path(__file__).parent / "sample.png"
|
53 |
+
if image_filepath.exists():
|
54 |
+
with open(image_filepath, "rb") as f:
|
55 |
+
image = f.read()
|
56 |
+
else:
|
57 |
+
try:
|
58 |
+
image = urllib.request.urlopen(
|
59 |
+
"https://drive.google.com/uc?id=0B0TjveMhQDhgLTlpOENiOTZ6Y00&export=download"
|
60 |
+
).read()
|
61 |
+
except urllib.error.URLError as e:
|
62 |
+
print(e.reason)
|
63 |
+
print(predict.call(image)[0])
|
vit_gpt2_image_caption_webapp.py
CHANGED
@@ -11,7 +11,7 @@ web_app = fastapi.FastAPI()
|
|
11 |
|
12 |
@web_app.post("/parse")
|
13 |
async def parse(request: fastapi.Request):
|
14 |
-
predict_step = Function.lookup("vit-gpt2-image-
|
15 |
|
16 |
form = await request.form()
|
17 |
image = await form["image"].read() # type: ignore
|
|
|
11 |
|
12 |
@web_app.post("/parse")
|
13 |
async def parse(request: fastapi.Request):
|
14 |
+
predict_step = Function.lookup("vit-gpt2-image-caption", "predict")
|
15 |
|
16 |
form = await request.form()
|
17 |
image = await form["image"].read() # type: ignore
|