Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -23,18 +23,18 @@ def load_model():
|
|
23 |
)
|
24 |
|
25 |
# Load the processor and model using the correct identifier
|
26 |
-
model_id = "google/paligemma2-
|
27 |
-
processor = PaliGemmaProcessor.from_pretrained(model_id,
|
28 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
29 |
model = PaliGemmaForConditionalGeneration.from_pretrained(
|
30 |
-
model_id, torch_dtype=torch.bfloat16,
|
31 |
).to(device).eval()
|
32 |
|
33 |
return processor, model
|
34 |
|
35 |
|
36 |
@spaces.GPU(duration=120) # Increased timeout to 120 seconds
|
37 |
-
def process_image_and_text(image_pil, text_input):
|
38 |
"""Extract text from image using PaliGemma2."""
|
39 |
try:
|
40 |
processor, model = load_model()
|
@@ -43,6 +43,9 @@ def process_image_and_text(image_pil, text_input):
|
|
43 |
# Load the image using load_image
|
44 |
image = load_image(image_pil)
|
45 |
|
|
|
|
|
|
|
46 |
# Use the provided text input
|
47 |
model_inputs = processor(text=text_input, images=image, return_tensors="pt").to(
|
48 |
device, dtype=torch.bfloat16
|
@@ -50,7 +53,7 @@ def process_image_and_text(image_pil, text_input):
|
|
50 |
input_len = model_inputs["input_ids"].shape[-1]
|
51 |
|
52 |
with torch.inference_mode():
|
53 |
-
generation = model.generate(**model_inputs, max_new_tokens=
|
54 |
generation = generation[0][input_len:]
|
55 |
decoded = processor.decode(generation, skip_special_tokens=True)
|
56 |
|
@@ -66,6 +69,7 @@ if __name__ == "__main__":
|
|
66 |
inputs=[
|
67 |
gr.Image(type="pil", label="Upload an image"),
|
68 |
gr.Textbox(label="Enter Text Prompt"),
|
|
|
69 |
],
|
70 |
outputs=gr.Textbox(label="Generated Text"),
|
71 |
title="PaliGemma2 Image and Text to Text",
|
|
|
23 |
)
|
24 |
|
25 |
# Load the processor and model using the correct identifier
|
26 |
+
model_id = "google/paligemma2-28b-pt-896"
|
27 |
+
processor = PaliGemmaProcessor.from_pretrained(model_id, use_auth_token=token)
|
28 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
29 |
model = PaliGemmaForConditionalGeneration.from_pretrained(
|
30 |
+
model_id, torch_dtype=torch.bfloat16, use_auth_token=token
|
31 |
).to(device).eval()
|
32 |
|
33 |
return processor, model
|
34 |
|
35 |
|
36 |
@spaces.GPU(duration=120) # Increased timeout to 120 seconds
|
37 |
+
def process_image_and_text(image_pil, text_input, num_beams):
|
38 |
"""Extract text from image using PaliGemma2."""
|
39 |
try:
|
40 |
processor, model = load_model()
|
|
|
43 |
# Load the image using load_image
|
44 |
image = load_image(image_pil)
|
45 |
|
46 |
+
# Add <image> token to the beginning of the text prompt
|
47 |
+
text_input = "<image> " + text_input
|
48 |
+
|
49 |
# Use the provided text input
|
50 |
model_inputs = processor(text=text_input, images=image, return_tensors="pt").to(
|
51 |
device, dtype=torch.bfloat16
|
|
|
53 |
input_len = model_inputs["input_ids"].shape[-1]
|
54 |
|
55 |
with torch.inference_mode():
|
56 |
+
generation = model.generate(**model_inputs, max_new_tokens=200, do_sample=False, num_beams=num_beams)
|
57 |
generation = generation[0][input_len:]
|
58 |
decoded = processor.decode(generation, skip_special_tokens=True)
|
59 |
|
|
|
69 |
inputs=[
|
70 |
gr.Image(type="pil", label="Upload an image"),
|
71 |
gr.Textbox(label="Enter Text Prompt"),
|
72 |
+
gr.Slider(minimum=1, maximum=10, step=1, value=1, label="Number of Beams"),
|
73 |
],
|
74 |
outputs=gr.Textbox(label="Generated Text"),
|
75 |
title="PaliGemma2 Image and Text to Text",
|