Commit
·
5053a56
1
Parent(s):
3bc78d3
this is a total fucking mess.
Browse files
app.py
CHANGED
@@ -1,57 +1,69 @@
|
|
1 |
import gradio as gr
|
2 |
-
from transformers import AutoProcessor, Blip2ForConditionalGeneration
|
3 |
import torch
|
4 |
-
from
|
5 |
-
|
6 |
-
# Load the BLIP-2 model and processor
|
7 |
-
processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b")
|
8 |
-
model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b")
|
9 |
|
10 |
-
#
|
11 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
12 |
-
model.to(device)
|
13 |
|
14 |
-
|
15 |
-
|
16 |
-
image_input = Image.fromarray(image).convert('RGB')
|
17 |
-
inputs = processor(image_input, return_tensors="pt").to(device, torch.float16)
|
18 |
-
|
19 |
-
# Image Captioning
|
20 |
-
generated_ids = model.generate(**inputs, max_new_tokens=20)
|
21 |
-
image_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
22 |
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
prompted_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
27 |
-
|
28 |
-
# Visual Question Answering (VQA)
|
29 |
-
prompt = f"Question: {vqa_question} Answer:"
|
30 |
-
inputs = processor(image_input, text=prompt, return_tensors="pt").to(device, torch.float16)
|
31 |
-
generated_ids = model.generate(**inputs, max_new_tokens=10)
|
32 |
-
vqa_answer = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
33 |
-
|
34 |
-
# Chat-based Prompting
|
35 |
-
prompt = chat_context + " Answer:"
|
36 |
-
inputs = processor(image_input, text=prompt, return_tensors="pt").to(device, torch.float16)
|
37 |
-
generated_ids = model.generate(**inputs, max_new_tokens=10)
|
38 |
-
chat_response = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
39 |
|
40 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
41 |
|
42 |
# Define Gradio input and output components
|
43 |
-
image_input = gr.
|
44 |
-
text_input = gr.
|
45 |
output_text = gr.outputs.Textbox()
|
46 |
|
47 |
-
#
|
48 |
-
|
49 |
-
fn=
|
50 |
-
inputs=[image_input, text_input,
|
51 |
-
outputs=
|
52 |
-
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
|
|
2 |
import torch
|
3 |
+
from transformers import AutoProcessor, Blip2ForConditionalGeneration
|
|
|
|
|
|
|
|
|
4 |
|
5 |
+
# Check if CUDA is available
|
6 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
7 |
|
8 |
+
# Model ID
|
9 |
+
MODEL_ID_FLAN_T5_XXL = "Salesforce/blip2-flan-t5-xxl"
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
|
11 |
+
# Load the model and processor
|
12 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID_FLAN_T5_XXL)
|
13 |
+
model = Blip2ForConditionalGeneration.from_pretrained(MODEL_ID_FLAN_T5_XXL, load_in_8bit=True).to(device)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
+
# Define a function for generating captions and answering questions
|
16 |
+
def generate_text(image, text, decoding_method, temperature, length_penalty, repetition_penalty):
|
17 |
+
if text.startswith("Caption:"):
|
18 |
+
# Generate caption
|
19 |
+
inputs = processor(images=image, return_tensors="pt").to(device, torch.float16)
|
20 |
+
generated_ids = model.generate(
|
21 |
+
pixel_values=inputs.pixel_values,
|
22 |
+
do_sample=decoding_method == "Nucleus sampling",
|
23 |
+
temperature=temperature,
|
24 |
+
length_penalty=length_penalty,
|
25 |
+
repetition_penalty=repetition_penalty,
|
26 |
+
max_length=50,
|
27 |
+
min_length=1,
|
28 |
+
num_beams=5,
|
29 |
+
top_p=0.9,
|
30 |
+
)
|
31 |
+
result = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
32 |
+
return result
|
33 |
+
else:
|
34 |
+
# Answer question
|
35 |
+
inputs = processor(images=image, text=text, return_tensors="pt").to(device, torch.float16)
|
36 |
+
generated_ids = model.generate(
|
37 |
+
**inputs,
|
38 |
+
do_sample=decoding_method == "Nucleus sampling",
|
39 |
+
temperature=temperature,
|
40 |
+
length_penalty=length_penalty,
|
41 |
+
repetition_penalty=repetition_penalty,
|
42 |
+
max_length=30,
|
43 |
+
min_length=1,
|
44 |
+
num_beams=5,
|
45 |
+
top_p=0.9,
|
46 |
+
)
|
47 |
+
result = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
|
48 |
+
return result
|
49 |
|
50 |
# Define Gradio input and output components
|
51 |
+
image_input = gr.Image(type="numpy")
|
52 |
+
text_input = gr.Text()
|
53 |
output_text = gr.outputs.Textbox()
|
54 |
|
55 |
+
# Define Gradio interface
|
56 |
+
gr.Interface(
|
57 |
+
fn=generate_text,
|
58 |
+
inputs=[image_input, text_input, gr.inputs.Radio(["Beam search", "Nucleus sampling"]), gr.inputs.Slider(0.5, 1.0, 0.1), gr.inputs.Slider(-1.0, 2.0, 0.2), gr.inputs.Slider(1.0, 5.0, 0.5)],
|
59 |
+
outputs=output_text,
|
60 |
+
examples=[
|
61 |
+
["house.png", "Caption:"],
|
62 |
+
["flower.jpg", "What is this flower and where is its origin?"],
|
63 |
+
["pizza.jpg", "Caption:"],
|
64 |
+
["sunset.jpg", "Caption:"],
|
65 |
+
["forbidden_city.webp", "In what dynasties was this place built?"],
|
66 |
+
],
|
67 |
+
title="BLIP-2",
|
68 |
+
description="Gradio demo for BLIP-2, image-to-text generation from Salesforce Research.",
|
69 |
+
).launch()
|