Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -36,19 +36,113 @@ def infer(image: PIL.Image.Image, text: str, max_new_tokens: int) -> str:
|
|
36 |
result = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
37 |
return result[0][len(text):].lstrip("\n")
|
38 |
|
39 |
-
# Image Captioning
|
40 |
-
def generate_caption(image: PIL.Image.Image) -> str:
|
41 |
-
return infer(image, "caption", max_new_tokens=50)
|
42 |
|
43 |
-
# Object Detection
|
44 |
-
def
|
45 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
-
#
|
48 |
-
def
|
49 |
-
|
|
|
50 |
|
51 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
with gr.Blocks() as demo:
|
53 |
gr.Markdown("# PaliGemma Multi-Modal App")
|
54 |
gr.Markdown("Upload an image and explore its features using the PaliGemma model!")
|
@@ -59,43 +153,23 @@ with gr.Blocks() as demo:
|
|
59 |
with gr.Row():
|
60 |
with gr.Column():
|
61 |
caption_image = gr.Image(type="pil", label="Upload Image", width=512, height=512)
|
|
|
62 |
caption_btn = gr.Button("Generate Caption")
|
63 |
with gr.Column():
|
64 |
caption_output = gr.Text(label="Generated Caption")
|
65 |
-
caption_btn.click(fn=generate_caption, inputs=[caption_image], outputs=[caption_output])
|
66 |
|
67 |
-
# Tab 2:
|
68 |
-
with gr.Tab("
|
69 |
with gr.Row():
|
70 |
with gr.Column():
|
71 |
detect_image = gr.Image(type="pil", label="Upload Image", width=512, height=512)
|
72 |
-
|
73 |
-
|
74 |
-
detect_output = gr.Text(label="Detected Objects")
|
75 |
-
detect_btn.click(fn=detect_objects, inputs=[detect_image], outputs=[detect_output])
|
76 |
-
|
77 |
-
# Tab 3: Visual Question Answering (VQA)
|
78 |
-
with gr.Tab("Visual Question Answering"):
|
79 |
-
with gr.Row():
|
80 |
-
with gr.Column():
|
81 |
-
vqa_image = gr.Image(type="pil", label="Upload Image", width=512, height=512)
|
82 |
-
vqa_question = gr.Text(label="Ask a Question", placeholder="What is in the image?")
|
83 |
-
vqa_btn = gr.Button("Ask")
|
84 |
-
with gr.Column():
|
85 |
-
vqa_output = gr.Text(label="Answer")
|
86 |
-
vqa_btn.click(fn=vqa, inputs=[vqa_image, vqa_question], outputs=[vqa_output])
|
87 |
-
|
88 |
-
# Tab 4: Text Generation (Original Feature)
|
89 |
-
with gr.Tab("Text Generation"):
|
90 |
-
with gr.Row():
|
91 |
-
with gr.Column():
|
92 |
-
text_image = gr.Image(type="pil", label="Upload Image", width=512, height=512)
|
93 |
-
text_input = gr.Text(label="Input Text", placeholder="Describe the image...")
|
94 |
-
text_btn = gr.Button("Generate Text")
|
95 |
with gr.Column():
|
96 |
-
|
97 |
-
|
98 |
|
99 |
# Launch the App
|
100 |
if __name__ == "__main__":
|
101 |
-
demo.queue(max_size=10).launch(debug=True)
|
|
|
36 |
result = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
37 |
return result[0][len(text):].lstrip("\n")
|
38 |
|
39 |
+
# Image Captioning (with user input for improvement)
|
40 |
+
def generate_caption(image: PIL.Image.Image, caption_improvement: str) -> str:
|
41 |
+
return infer(image, f"caption: {caption_improvement}", max_new_tokens=50)
|
42 |
|
43 |
+
# Object Detection/Segmentation
|
44 |
+
def parse_segmentation(input_image, input_text):
|
45 |
+
out = infer(input_image, input_text, max_new_tokens=200)
|
46 |
+
objs = extract_objs(out.lstrip("\n"), input_image.size[0], input_image.size[1], unique_labels=True)
|
47 |
+
labels = set(obj.get('name') for obj in objs if obj.get('name'))
|
48 |
+
color_map = {l: COLORS[i % len(COLORS)] for i, l in enumerate(labels)}
|
49 |
+
highlighted_text = [(obj['content'], obj.get('name')) for obj in objs]
|
50 |
+
annotated_img = (
|
51 |
+
input_image,
|
52 |
+
[
|
53 |
+
(
|
54 |
+
obj['mask'] if obj.get('mask') is not None else obj['xyxy'],
|
55 |
+
obj['name'] or '',
|
56 |
+
)
|
57 |
+
for obj in objs
|
58 |
+
if 'mask' in obj or 'xyxy' in obj
|
59 |
+
],
|
60 |
+
)
|
61 |
+
has_annotations = bool(annotated_img[1])
|
62 |
+
return annotated_img
|
63 |
|
64 |
+
# Helper functions for object detection/segmentation
|
65 |
+
def _get_params(checkpoint):
|
66 |
+
def transp(kernel):
|
67 |
+
return np.transpose(kernel, (2, 3, 1, 0))
|
68 |
|
69 |
+
def conv(name):
|
70 |
+
return {
|
71 |
+
'bias': checkpoint[name + '.bias'],
|
72 |
+
'kernel': transp(checkpoint[name + '.weight']),
|
73 |
+
}
|
74 |
+
|
75 |
+
def resblock(name):
|
76 |
+
return {
|
77 |
+
'Conv_0': conv(name + '.0'),
|
78 |
+
'Conv_1': conv(name + '.2'),
|
79 |
+
'Conv_2': conv(name + '.4'),
|
80 |
+
}
|
81 |
+
|
82 |
+
return {
|
83 |
+
'_embeddings': checkpoint['_vq_vae._embedding'],
|
84 |
+
'Conv_0': conv('decoder.0'),
|
85 |
+
'ResBlock_0': resblock('decoder.2.net'),
|
86 |
+
'ResBlock_1': resblock('decoder.3.net'),
|
87 |
+
'ConvTranspose_0': conv('decoder.4'),
|
88 |
+
'ConvTranspose_1': conv('decoder.6'),
|
89 |
+
'ConvTranspose_2': conv('decoder.8'),
|
90 |
+
'ConvTranspose_3': conv('decoder.10'),
|
91 |
+
'Conv_1': conv('decoder.12'),
|
92 |
+
}
|
93 |
+
|
94 |
+
def _quantized_values_from_codebook_indices(codebook_indices, embeddings):
|
95 |
+
batch_size, num_tokens = codebook_indices.shape
|
96 |
+
assert num_tokens == 16, codebook_indices.shape
|
97 |
+
unused_num_embeddings, embedding_dim = embeddings.shape
|
98 |
+
|
99 |
+
encodings = jnp.take(embeddings, codebook_indices.reshape((-1)), axis=0)
|
100 |
+
encodings = encodings.reshape((batch_size, 4, 4, embedding_dim))
|
101 |
+
return encodings
|
102 |
+
|
103 |
+
def extract_objs(text, width, height, unique_labels=False):
|
104 |
+
objs = []
|
105 |
+
seen = set()
|
106 |
+
while text:
|
107 |
+
m = _SEGMENT_DETECT_RE.match(text)
|
108 |
+
if not m:
|
109 |
+
break
|
110 |
+
|
111 |
+
gs = list(m.groups())
|
112 |
+
before = gs.pop(0)
|
113 |
+
name = gs.pop()
|
114 |
+
y1, x1, y2, x2 = [int(x) / 1024 for x in gs[:4]]
|
115 |
+
|
116 |
+
y1, x1, y2, x2 = map(round, (y1*height, x1*width, y2*height, x2*width))
|
117 |
+
seg_indices = gs[4:20]
|
118 |
+
if seg_indices[0] is None:
|
119 |
+
mask = None
|
120 |
+
else:
|
121 |
+
seg_indices = np.array([int(x) for x in seg_indices], dtype=np.int32)
|
122 |
+
m64, = _get_reconstruct_masks()(seg_indices[None])[..., 0]
|
123 |
+
m64 = np.clip(np.array(m64) * 0.5 + 0.5, 0, 1)
|
124 |
+
m64 = PIL.Image.fromarray((m64 * 255).astype('uint8'))
|
125 |
+
mask = np.zeros([height, width])
|
126 |
+
if y2 > y1 and x2 > x1:
|
127 |
+
mask[y1:y2, x1:x2] = np.array(m64.resize([x2 - x1, y2 - y1])) / 255.0
|
128 |
+
|
129 |
+
content = m.group()
|
130 |
+
if before:
|
131 |
+
objs.append(dict(content=before))
|
132 |
+
content = content[len(before):]
|
133 |
+
while unique_labels and name in seen:
|
134 |
+
name = (name or '') + "'"
|
135 |
+
seen.add(name)
|
136 |
+
objs.append(dict(
|
137 |
+
content=content, xyxy=(x1, y1, x2, y2), mask=mask, name=name))
|
138 |
+
text = text[len(before) + len(content):]
|
139 |
+
|
140 |
+
if text:
|
141 |
+
objs.append(dict(content=text))
|
142 |
+
|
143 |
+
return objs
|
144 |
+
|
145 |
+
# Gradio Interface
|
146 |
with gr.Blocks() as demo:
|
147 |
gr.Markdown("# PaliGemma Multi-Modal App")
|
148 |
gr.Markdown("Upload an image and explore its features using the PaliGemma model!")
|
|
|
153 |
with gr.Row():
|
154 |
with gr.Column():
|
155 |
caption_image = gr.Image(type="pil", label="Upload Image", width=512, height=512)
|
156 |
+
caption_improvement_input = gr.Textbox(label="Improvement Input", placeholder="Enter description to improve caption")
|
157 |
caption_btn = gr.Button("Generate Caption")
|
158 |
with gr.Column():
|
159 |
caption_output = gr.Text(label="Generated Caption")
|
160 |
+
caption_btn.click(fn=generate_caption, inputs=[caption_image, caption_improvement_input], outputs=[caption_output])
|
161 |
|
162 |
+
# Tab 2: Segment/Detect
|
163 |
+
with gr.Tab("Segment/Detect"):
|
164 |
with gr.Row():
|
165 |
with gr.Column():
|
166 |
detect_image = gr.Image(type="pil", label="Upload Image", width=512, height=512)
|
167 |
+
detect_text = gr.Textbox(label="Entities to Detect", placeholder="List entities to segment/detect")
|
168 |
+
detect_btn = gr.Button("Detect/Segment")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
with gr.Column():
|
170 |
+
detect_output = gr.AnnotatedImage(label="Annotated Image")
|
171 |
+
detect_btn.click(fn=parse_segmentation, inputs=[detect_image, detect_text], outputs=[detect_output])
|
172 |
|
173 |
# Launch the App
|
174 |
if __name__ == "__main__":
|
175 |
+
demo.queue(max_size=10).launch(debug=True)
|