Spaces:
Running
Running
File size: 3,513 Bytes
1e868bb 0553ee9 1e868bb 0553ee9 1e868bb 0553ee9 1e868bb 0553ee9 1e868bb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 |
"""This space is taken and modified from https://huggingface.co/spaces/merve/compare_clip_siglip"""
import torch
from transformers import (
AutoModel,
AutoProcessor
)
import gradio as gr
################################################################################
# Load the models
################################################################################
sg1_ckpt = "google/siglip-so400m-patch14-384"
siglip1_model = AutoModel.from_pretrained(sg1_ckpt, device_map="auto").eval()
siglip1_processor = AutoProcessor.from_pretrained(sg1_ckpt)
sg2_ckpt = "google/siglip2-so400m-patch14-384"
siglip2_model = AutoModel.from_pretrained(sg2_ckpt, device_map="auto").eval()
siglip2_processor = AutoProcessor.from_pretrained(sg2_ckpt)
################################################################################
# Utilities
################################################################################
def postprocess(output):
return {out["label"]: float(out["score"]) for out in output}
def postprocess_siglip(sg1_probs, sg2_probs, labels):
sg1_output = {labels[i]: float(sg1_probs[0].cpu().numpy()[i]) for i in range(len(labels))}
sg2_output = {labels[i]: float(sg2_probs[0].cpu().numpy()[i]) for i in range(len(labels))}
return sg1_output, sg2_output
def siglip_detector(image, texts):
sg1_inputs = siglip1_processor(
text=texts, images=image, return_tensors="pt", padding="max_length", max_length=64
).to(siglip1_model.device)
sg2_inputs = siglip2_processor(
text=texts, images=image, return_tensors="pt", padding="max_length", max_length=64
).to(siglip2_model.device)
with torch.no_grad():
sg1_outputs = siglip1_model(**sg1_inputs)
sg2_outputs = siglip2_model(**sg2_inputs)
sg1_logits_per_image = sg1_outputs.logits_per_image
sg2_logits_per_image = sg2_outputs.logits_per_image
sg1_probs = torch.sigmoid(sg1_logits_per_image)
sg2_probs = torch.sigmoid(sg2_logits_per_image)
return sg1_probs, sg2_probs
def infer(image, candidate_labels):
candidate_labels = [label.lstrip(" ") for label in candidate_labels.split(",")]
sg1_probs, sg2_probs = siglip_detector(image, candidate_labels)
return postprocess_siglip(
sg1_probs, sg2_probs, labels=candidate_labels
)
with gr.Blocks() as demo:
gr.Markdown("# Compare SigLIP 1 and SigLIP 2")
gr.Markdown(
"Compare the performance of SigLIP 1 and SigLIP 2 on zero-shot classification in this Space 👇"
)
with gr.Row():
with gr.Column():
image_input = gr.Image(type="pil")
text_input = gr.Textbox(label="Input a list of labels (comma seperated)")
run_button = gr.Button("Run", visible=True)
with gr.Column():
siglip1_output = gr.Label(label="SigLIP 1 Output", num_top_classes=3)
siglip2_output = gr.Label(label="SigLIP 2 Output", num_top_classes=3)
examples = [
["./baklava.jpg", "desser on a plate, a serving of baklava, a plate and spoon"],
["./baklava.jpg", "a cat, two cats, three cats"],
["./baklava.jpg", "two sleeping cats, two cats playing, three cats laying down"],
]
gr.Examples(
examples=examples,
inputs=[image_input, text_input],
outputs=[siglip1_output, siglip2_output],
fn=infer,
)
run_button.click(
fn=infer, inputs=[image_input, text_input], outputs=[siglip1_output, siglip2_output]
)
demo.launch()
|