Spaces:
Running
Running
run on cpu
Browse files
app.py
CHANGED
@@ -1,63 +1,33 @@
|
|
1 |
"""This space is taken and modified from https://huggingface.co/spaces/merve/compare_clip_siglip"""
|
2 |
-
import
|
3 |
-
from transformers import (
|
4 |
-
AutoModel,
|
5 |
-
AutoProcessor
|
6 |
-
)
|
7 |
import gradio as gr
|
8 |
|
9 |
################################################################################
|
10 |
# Load the models
|
11 |
################################################################################
|
12 |
sg1_ckpt = "google/siglip-so400m-patch14-384"
|
13 |
-
|
14 |
-
siglip1_processor = AutoProcessor.from_pretrained(sg1_ckpt)
|
15 |
|
16 |
sg2_ckpt = "google/siglip2-so400m-patch14-384"
|
17 |
-
|
18 |
-
siglip2_processor = AutoProcessor.from_pretrained(sg2_ckpt)
|
19 |
|
20 |
################################################################################
|
21 |
-
#
|
22 |
################################################################################
|
23 |
-
def postprocess(output):
|
24 |
-
return {out["label"]: float(out["score"]) for out in output}
|
25 |
-
|
26 |
-
|
27 |
-
def postprocess_siglip(sg1_probs, sg2_probs, labels):
|
28 |
-
sg1_output = {labels[i]: float(sg1_probs[0].cpu().numpy()[i]) for i in range(len(labels))}
|
29 |
-
sg2_output = {labels[i]: float(sg2_probs[0].cpu().numpy()[i]) for i in range(len(labels))}
|
30 |
-
return sg1_output, sg2_output
|
31 |
-
|
32 |
-
def siglip_detector(image, texts):
|
33 |
-
sg1_inputs = siglip1_processor(
|
34 |
-
text=texts, images=image, return_tensors="pt", padding="max_length", max_length=64
|
35 |
-
).to(siglip1_model.device)
|
36 |
-
|
37 |
-
sg2_inputs = siglip2_processor(
|
38 |
-
text=texts, images=image, return_tensors="pt", padding="max_length", max_length=64
|
39 |
-
).to(siglip2_model.device)
|
40 |
-
|
41 |
-
with torch.no_grad():
|
42 |
-
sg1_outputs = siglip1_model(**sg1_inputs)
|
43 |
-
sg2_outputs = siglip2_model(**sg2_inputs)
|
44 |
-
|
45 |
-
sg1_logits_per_image = sg1_outputs.logits_per_image
|
46 |
-
sg2_logits_per_image = sg2_outputs.logits_per_image
|
47 |
-
|
48 |
-
sg1_probs = torch.sigmoid(sg1_logits_per_image)
|
49 |
-
sg2_probs = torch.sigmoid(sg2_logits_per_image)
|
50 |
-
return sg1_probs, sg2_probs
|
51 |
-
|
52 |
-
|
53 |
def infer(image, candidate_labels):
|
54 |
candidate_labels = [label.lstrip(" ") for label in candidate_labels.split(",")]
|
55 |
-
sg1_probs, sg2_probs = siglip_detector(image, candidate_labels)
|
56 |
-
return postprocess_siglip(
|
57 |
-
sg1_probs, sg2_probs, labels=candidate_labels
|
58 |
-
)
|
59 |
|
|
|
|
|
|
|
|
|
|
|
60 |
|
|
|
|
|
|
|
|
|
|
|
61 |
with gr.Blocks() as demo:
|
62 |
gr.Markdown("# Compare SigLIP 1 and SigLIP 2")
|
63 |
gr.Markdown(
|
|
|
1 |
"""This space is taken and modified from https://huggingface.co/spaces/merve/compare_clip_siglip"""
|
2 |
+
from transformers import pipeline
|
|
|
|
|
|
|
|
|
3 |
import gradio as gr
|
4 |
|
5 |
################################################################################
|
6 |
# Load the models
|
7 |
################################################################################
|
8 |
sg1_ckpt = "google/siglip-so400m-patch14-384"
|
9 |
+
sg1_pipe = pipeline(task="zero-shot-image-classification", model=sg1_ckpt, device="cpu")
|
|
|
10 |
|
11 |
sg2_ckpt = "google/siglip2-so400m-patch14-384"
|
12 |
+
sg2_pipe = pipeline(task="zero-shot-image-classification", model=sg2_ckpt, device="cpu")
|
|
|
13 |
|
14 |
################################################################################
|
15 |
+
# Run inference
|
16 |
################################################################################
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
def infer(image, candidate_labels):
|
18 |
candidate_labels = [label.lstrip(" ") for label in candidate_labels.split(",")]
|
|
|
|
|
|
|
|
|
19 |
|
20 |
+
sg1_socres = sg1_pipe(image, candidate_labels=candidate_labels)
|
21 |
+
sg2_socres = sg2_pipe(image, candidate_labels=candidate_labels)
|
22 |
+
|
23 |
+
sg1_outputs = {element["label"]:element["score"] for element in sg1_socres}
|
24 |
+
sg2_outputs = {element["label"]:element["score"] for element in sg2_socres}
|
25 |
|
26 |
+
return sg1_outputs, sg2_outputs
|
27 |
+
|
28 |
+
################################################################################
|
29 |
+
# Gradio App
|
30 |
+
################################################################################
|
31 |
with gr.Blocks() as demo:
|
32 |
gr.Markdown("# Compare SigLIP 1 and SigLIP 2")
|
33 |
gr.Markdown(
|