Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,24 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import functools
|
2 |
import re
|
|
|
3 |
import PIL.Image
|
4 |
import gradio as gr
|
|
|
|
|
5 |
import jax
|
6 |
import jax.numpy as jnp
|
7 |
-
import numpy as np
|
8 |
import flax.linen as nn
|
9 |
-
from inference import PaliGemmaModel, VAEModel
|
10 |
|
11 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
-
# Instantiate the models
|
14 |
-
pali_gemma_model = PaliGemmaModel()
|
15 |
-
vae_model = VAEModel('vae-oid.npz')
|
16 |
|
17 |
-
|
18 |
-
##### Also returns bounding boxes with their labels
|
19 |
|
20 |
-
|
21 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
objs = extract_objs(out.lstrip("\n"), input_image.size[0], input_image.size[1], unique_labels=True)
|
23 |
labels = set(obj.get('name') for obj in objs if obj.get('name'))
|
24 |
color_map = {l: COLORS[i % len(COLORS)] for i, l in enumerate(labels)}
|
@@ -37,14 +102,166 @@ def parse_segmentation(input_image, input_text, max_new_tokens=100):
|
|
37 |
has_annotations = bool(annotated_img[1])
|
38 |
return annotated_img
|
39 |
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
|
46 |
def extract_objs(text, width, height, unique_labels=False):
|
47 |
-
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
objs = []
|
49 |
seen = set()
|
50 |
while text:
|
@@ -56,14 +273,14 @@ def extract_objs(text, width, height, unique_labels=False):
|
|
56 |
before = gs.pop(0)
|
57 |
name = gs.pop()
|
58 |
y1, x1, y2, x2 = [int(x) / 1024 for x in gs[:4]]
|
59 |
-
|
60 |
-
y1, x1, y2, x2 = map(round, (y1
|
61 |
seg_indices = gs[4:20]
|
62 |
if seg_indices[0] is None:
|
63 |
mask = None
|
64 |
else:
|
65 |
seg_indices = np.array([int(x) for x in seg_indices], dtype=np.int32)
|
66 |
-
m64, =
|
67 |
m64 = np.clip(np.array(m64) * 0.5 + 0.5, 0, 1)
|
68 |
m64 = PIL.Image.fromarray((m64 * 255).astype('uint8'))
|
69 |
mask = np.zeros([height, width])
|
@@ -86,12 +303,13 @@ def extract_objs(text, width, height, unique_labels=False):
|
|
86 |
|
87 |
return objs
|
88 |
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
|
|
95 |
|
96 |
with gr.Blocks(css="style.css") as demo:
|
97 |
gr.Markdown(INTRO_TEXT)
|
@@ -100,27 +318,25 @@ with gr.Blocks(css="style.css") as demo:
|
|
100 |
with gr.Column():
|
101 |
image = gr.Image(type="pil")
|
102 |
seg_input = gr.Text(label="Entities to Segment/Detect")
|
103 |
-
|
104 |
with gr.Column():
|
105 |
annotated_image = gr.AnnotatedImage(label="Output")
|
106 |
|
107 |
-
seg_btn = gr.Button("Submit")
|
108 |
-
examples = [
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
["./examples/cart3.jpg", "detect cells"]
|
115 |
-
]
|
116 |
gr.Examples(
|
117 |
examples=examples,
|
118 |
inputs=[image, seg_input],
|
119 |
)
|
120 |
seg_inputs = [
|
121 |
image,
|
122 |
-
seg_input
|
123 |
-
|
124 |
seg_outputs = [
|
125 |
annotated_image
|
126 |
]
|
@@ -133,6 +349,7 @@ with gr.Blocks(css="style.css") as demo:
|
|
133 |
with gr.Column():
|
134 |
image = gr.Image(type="pil")
|
135 |
text_input = gr.Text(label="Input Text")
|
|
|
136 |
text_output = gr.Text(label="Text Output")
|
137 |
chat_btn = gr.Button()
|
138 |
tokens = gr.Slider(
|
@@ -148,25 +365,25 @@ with gr.Blocks(css="style.css") as demo:
|
|
148 |
image,
|
149 |
text_input,
|
150 |
tokens
|
151 |
-
|
152 |
chat_outputs = [
|
153 |
text_output
|
154 |
]
|
155 |
chat_btn.click(
|
156 |
-
fn=
|
157 |
inputs=chat_inputs,
|
158 |
outputs=chat_outputs,
|
159 |
)
|
160 |
-
|
161 |
-
examples = [
|
162 |
-
|
163 |
-
|
164 |
-
["./examples/cart3.jpg", IMAGE_PROMPT]
|
165 |
-
]
|
166 |
gr.Examples(
|
167 |
examples=examples,
|
168 |
inputs=chat_inputs,
|
169 |
)
|
170 |
|
|
|
|
|
171 |
if __name__ == "__main__":
|
172 |
-
demo.queue(max_size=10).launch(debug=True)
|
|
|
1 |
+
"""
|
2 |
+
CellVision AI - Intelligent Cell Imaging Analysis
|
3 |
+
|
4 |
+
This module provides a Gradio web application for performing intelligent cell imaging analysis
|
5 |
+
using the PaliGemma model from Google. The app allows users to segment or detect cells in images
|
6 |
+
and generate descriptive text based on the input image and prompt.
|
7 |
+
|
8 |
+
Dependencies:
|
9 |
+
- gradio
|
10 |
+
- transformers
|
11 |
+
- torch
|
12 |
+
- jax
|
13 |
+
- flax
|
14 |
+
- spaces
|
15 |
+
- PIL
|
16 |
+
- numpy
|
17 |
+
- huggingface_hub
|
18 |
+
|
19 |
+
"""
|
20 |
+
|
21 |
+
import os
|
22 |
import functools
|
23 |
import re
|
24 |
+
|
25 |
import PIL.Image
|
26 |
import gradio as gr
|
27 |
+
import numpy as np
|
28 |
+
import torch
|
29 |
import jax
|
30 |
import jax.numpy as jnp
|
|
|
31 |
import flax.linen as nn
|
|
|
32 |
|
33 |
+
from transformers import PaliGemmaForConditionalGeneration, PaliGemmaProcessor
|
34 |
+
from huggingface_hub import login
|
35 |
+
import spaces
|
36 |
+
|
37 |
+
# Perform login using the token
|
38 |
+
hf_token = os.getenv("HF_TOKEN")
|
39 |
+
login(token=hf_token, add_to_git_credential=True)
|
40 |
|
|
|
|
|
|
|
41 |
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
|
43 |
|
44 |
+
model_id = "google/paligemma-3b-mix-448"
|
45 |
+
model = PaliGemmaForConditionalGeneration.from_pretrained(model_id).eval().to(device)
|
46 |
+
processor = PaliGemmaProcessor.from_pretrained(model_id)
|
47 |
+
|
48 |
+
@spaces.GPU
|
49 |
+
def infer(
|
50 |
+
image: PIL.Image.Image,
|
51 |
+
text: str,
|
52 |
+
max_new_tokens: int
|
53 |
+
) -> str:
|
54 |
+
"""
|
55 |
+
Perform inference using the PaliGemma model.
|
56 |
+
|
57 |
+
Args:
|
58 |
+
image (PIL.Image.Image): Input image.
|
59 |
+
text (str): Input text prompt.
|
60 |
+
max_new_tokens (int): Maximum number of new tokens to generate.
|
61 |
+
|
62 |
+
Returns:
|
63 |
+
str: Generated text based on the input image and prompt.
|
64 |
+
"""
|
65 |
+
inputs = processor(text=text, images=image, return_tensors="pt").to(device)
|
66 |
+
with torch.inference_mode():
|
67 |
+
generated_ids = model.generate(
|
68 |
+
**inputs,
|
69 |
+
max_new_tokens=max_new_tokens,
|
70 |
+
do_sample=False
|
71 |
+
)
|
72 |
+
result = processor.batch_decode(generated_ids, skip_special_tokens=True)
|
73 |
+
return result[0][len(text):].lstrip("\n")
|
74 |
+
|
75 |
+
def parse_segmentation(input_image, input_text):
|
76 |
+
"""
|
77 |
+
Parse segmentation output tokens into masks and bounding boxes.
|
78 |
+
|
79 |
+
Args:
|
80 |
+
input_image (PIL.Image.Image): Input image.
|
81 |
+
input_text (str): Input text specifying entities to segment or detect.
|
82 |
+
|
83 |
+
Returns:
|
84 |
+
tuple: A tuple containing the annotated image and a boolean indicating if annotations are present.
|
85 |
+
"""
|
86 |
+
out = infer(input_image, input_text, max_new_tokens=100)
|
87 |
objs = extract_objs(out.lstrip("\n"), input_image.size[0], input_image.size[1], unique_labels=True)
|
88 |
labels = set(obj.get('name') for obj in objs if obj.get('name'))
|
89 |
color_map = {l: COLORS[i % len(COLORS)] for i, l in enumerate(labels)}
|
|
|
102 |
has_annotations = bool(annotated_img[1])
|
103 |
return annotated_img
|
104 |
|
105 |
+
|
106 |
+
### Postprocessing Utils for Segmentation Tokens
|
107 |
+
|
108 |
+
_MODEL_PATH = 'vae-oid.npz'
|
109 |
+
|
110 |
+
_SEGMENT_DETECT_RE = re.compile(
|
111 |
+
r'(.*?)' +
|
112 |
+
r'<loc(\d{4})>' * 4 + r'\s*' +
|
113 |
+
'(?:%s)?' % (r'<seg(\d{3})>' * 16) +
|
114 |
+
r'\s*([^;<>]+)? ?(?:; )?',
|
115 |
+
)
|
116 |
+
|
117 |
+
COLORS = ['#4285f4', '#db4437', '#f4b400', '#0f9d58', '#e48ef1']
|
118 |
+
|
119 |
+
def _get_params(checkpoint):
|
120 |
+
"""
|
121 |
+
Convert PyTorch checkpoint to Flax params.
|
122 |
+
|
123 |
+
Args:
|
124 |
+
checkpoint (dict): PyTorch checkpoint dictionary.
|
125 |
+
|
126 |
+
Returns:
|
127 |
+
dict: Flax parameters.
|
128 |
+
"""
|
129 |
+
def transp(kernel):
|
130 |
+
return np.transpose(kernel, (2, 3, 1, 0))
|
131 |
+
|
132 |
+
def conv(name):
|
133 |
+
return {
|
134 |
+
'bias': checkpoint[name + '.bias'],
|
135 |
+
'kernel': transp(checkpoint[name + '.weight']),
|
136 |
+
}
|
137 |
+
|
138 |
+
def resblock(name):
|
139 |
+
return {
|
140 |
+
'Conv_0': conv(name + '.0'),
|
141 |
+
'Conv_1': conv(name + '.2'),
|
142 |
+
'Conv_2': conv(name + '.4'),
|
143 |
+
}
|
144 |
+
|
145 |
+
return {
|
146 |
+
'_embeddings': checkpoint['_vq_vae._embedding'],
|
147 |
+
'Conv_0': conv('decoder.0'),
|
148 |
+
'ResBlock_0': resblock('decoder.2.net'),
|
149 |
+
'ResBlock_1': resblock('decoder.3.net'),
|
150 |
+
'ConvTranspose_0': conv('decoder.4'),
|
151 |
+
'ConvTranspose_1': conv('decoder.6'),
|
152 |
+
'ConvTranspose_2': conv('decoder.8'),
|
153 |
+
'ConvTranspose_3': conv('decoder.10'),
|
154 |
+
'Conv_1': conv('decoder.12'),
|
155 |
+
}
|
156 |
+
|
157 |
+
|
158 |
+
def _quantized_values_from_codebook_indices(codebook_indices, embeddings):
|
159 |
+
"""
|
160 |
+
Get quantized values from codebook indices.
|
161 |
+
|
162 |
+
Args:
|
163 |
+
codebook_indices (jax.numpy.ndarray): Codebook indices.
|
164 |
+
embeddings (jax.numpy.ndarray): Embeddings.
|
165 |
+
|
166 |
+
Returns:
|
167 |
+
jax.numpy.ndarray: Quantized values.
|
168 |
+
"""
|
169 |
+
batch_size, num_tokens = codebook_indices.shape
|
170 |
+
assert num_tokens == 16, codebook_indices.shape
|
171 |
+
unused_num_embeddings, embedding_dim = embeddings.shape
|
172 |
+
|
173 |
+
encodings = jnp.take(embeddings, codebook_indices.reshape((-1)), axis=0)
|
174 |
+
encodings = encodings.reshape((batch_size, 4, 4, embedding_dim))
|
175 |
+
return encodings
|
176 |
+
|
177 |
+
|
178 |
+
@functools.cache
|
179 |
+
def _get_reconstruct_masks():
|
180 |
+
"""
|
181 |
+
Reconstruct masks from codebook indices.
|
182 |
+
|
183 |
+
Returns:
|
184 |
+
function: A function that expects indices shaped `[B, 16]` of dtype int32, each
|
185 |
+
ranging from 0 to 127 (inclusive), and returns decoded masks sized
|
186 |
+
`[B, 64, 64, 1]`, of dtype float32, in range [-1, 1].
|
187 |
+
"""
|
188 |
+
|
189 |
+
class ResBlock(nn.Module):
|
190 |
+
features: int
|
191 |
+
|
192 |
+
@nn.compact
|
193 |
+
def __call__(self, x):
|
194 |
+
original_x = x
|
195 |
+
x = nn.Conv(features=self.features, kernel_size=(3, 3), padding=1)(x)
|
196 |
+
x = nn.relu(x)
|
197 |
+
x = nn.Conv(features=self.features, kernel_size=(3, 3), padding=1)(x)
|
198 |
+
x = nn.relu(x)
|
199 |
+
x = nn.Conv(features=self.features, kernel_size=(1, 1), padding=0)(x)
|
200 |
+
return x + original_x
|
201 |
+
|
202 |
+
class Decoder(nn.Module):
|
203 |
+
"""Upscales quantized vectors to mask."""
|
204 |
+
|
205 |
+
@nn.compact
|
206 |
+
def __call__(self, x):
|
207 |
+
num_res_blocks = 2
|
208 |
+
dim = 128
|
209 |
+
num_upsample_layers = 4
|
210 |
+
|
211 |
+
x = nn.Conv(features=dim, kernel_size=(1, 1), padding=0)(x)
|
212 |
+
x = nn.relu(x)
|
213 |
+
|
214 |
+
for _ in range(num_res_blocks):
|
215 |
+
x = ResBlock(features=dim)(x)
|
216 |
+
|
217 |
+
for _ in range(num_upsample_layers):
|
218 |
+
x = nn.ConvTranspose(
|
219 |
+
features=dim,
|
220 |
+
kernel_size=(4, 4),
|
221 |
+
strides=(2, 2),
|
222 |
+
padding=2,
|
223 |
+
transpose_kernel=True,
|
224 |
+
)(x)
|
225 |
+
x = nn.relu(x)
|
226 |
+
dim //= 2
|
227 |
+
|
228 |
+
x = nn.Conv(features=1, kernel_size=(1, 1), padding=0)(x)
|
229 |
+
|
230 |
+
return x
|
231 |
+
|
232 |
+
def reconstruct_masks(codebook_indices):
|
233 |
+
"""
|
234 |
+
Reconstruct masks from codebook indices.
|
235 |
+
|
236 |
+
Args:
|
237 |
+
codebook_indices (jax.numpy.ndarray): Codebook indices.
|
238 |
+
|
239 |
+
Returns:
|
240 |
+
jax.numpy.ndarray: Reconstructed masks.
|
241 |
+
"""
|
242 |
+
quantized = _quantized_values_from_codebook_indices(
|
243 |
+
codebook_indices, params['_embeddings']
|
244 |
+
)
|
245 |
+
return Decoder().apply({'params': params}, quantized)
|
246 |
+
|
247 |
+
with open(_MODEL_PATH, 'rb') as f:
|
248 |
+
params = _get_params(dict(np.load(f)))
|
249 |
+
|
250 |
+
return jax.jit(reconstruct_masks, backend='cpu')
|
251 |
|
252 |
def extract_objs(text, width, height, unique_labels=False):
|
253 |
+
"""
|
254 |
+
Extract objects from text containing "<loc>" and "<seg>" tokens.
|
255 |
+
|
256 |
+
Args:
|
257 |
+
text (str): Input text containing "<loc>" and "<seg>" tokens.
|
258 |
+
width (int): Width of the image.
|
259 |
+
height (int): Height of the image.
|
260 |
+
unique_labels (bool, optional): Whether to enforce unique labels. Defaults to False.
|
261 |
+
|
262 |
+
Returns:
|
263 |
+
list: List of extracted objects.
|
264 |
+
"""
|
265 |
objs = []
|
266 |
seen = set()
|
267 |
while text:
|
|
|
273 |
before = gs.pop(0)
|
274 |
name = gs.pop()
|
275 |
y1, x1, y2, x2 = [int(x) / 1024 for x in gs[:4]]
|
276 |
+
|
277 |
+
y1, x1, y2, x2 = map(round, (y1*height, x1*width, y2*height, x2*width))
|
278 |
seg_indices = gs[4:20]
|
279 |
if seg_indices[0] is None:
|
280 |
mask = None
|
281 |
else:
|
282 |
seg_indices = np.array([int(x) for x in seg_indices], dtype=np.int32)
|
283 |
+
m64, = _get_reconstruct_masks()(seg_indices[None])[..., 0]
|
284 |
m64 = np.clip(np.array(m64) * 0.5 + 0.5, 0, 1)
|
285 |
m64 = PIL.Image.fromarray((m64 * 255).astype('uint8'))
|
286 |
mask = np.zeros([height, width])
|
|
|
303 |
|
304 |
return objs
|
305 |
|
306 |
+
#########
|
307 |
+
|
308 |
+
INTRO_TEXT="🔬🧠 CellVision AI -- Intelligent Cell Imaging Analysis 🤖🧫"
|
309 |
+
IMAGE_PROMPT="""
|
310 |
+
Describe the morphological characteristics and visible interactions between different cell types.
|
311 |
+
Assess the biological context to identify signs of cancer and the presence of antigens.
|
312 |
+
"""
|
313 |
|
314 |
with gr.Blocks(css="style.css") as demo:
|
315 |
gr.Markdown(INTRO_TEXT)
|
|
|
318 |
with gr.Column():
|
319 |
image = gr.Image(type="pil")
|
320 |
seg_input = gr.Text(label="Entities to Segment/Detect")
|
321 |
+
|
322 |
with gr.Column():
|
323 |
annotated_image = gr.AnnotatedImage(label="Output")
|
324 |
|
325 |
+
seg_btn = gr.Button("Submit")
|
326 |
+
examples = [["./examples/cart1.jpg", "segment cells"],
|
327 |
+
["./examples/cart1.jpg", "detect cells"],
|
328 |
+
["./examples/cart2.jpg", "segment cells"],
|
329 |
+
["./examples/cart2.jpg", "detect cells"],
|
330 |
+
["./examples/cart3.jpg", "segment cells"],
|
331 |
+
["./examples/cart3.jpg", "detect cells"]]
|
|
|
|
|
332 |
gr.Examples(
|
333 |
examples=examples,
|
334 |
inputs=[image, seg_input],
|
335 |
)
|
336 |
seg_inputs = [
|
337 |
image,
|
338 |
+
seg_input
|
339 |
+
]
|
340 |
seg_outputs = [
|
341 |
annotated_image
|
342 |
]
|
|
|
349 |
with gr.Column():
|
350 |
image = gr.Image(type="pil")
|
351 |
text_input = gr.Text(label="Input Text")
|
352 |
+
|
353 |
text_output = gr.Text(label="Text Output")
|
354 |
chat_btn = gr.Button()
|
355 |
tokens = gr.Slider(
|
|
|
365 |
image,
|
366 |
text_input,
|
367 |
tokens
|
368 |
+
]
|
369 |
chat_outputs = [
|
370 |
text_output
|
371 |
]
|
372 |
chat_btn.click(
|
373 |
+
fn=infer,
|
374 |
inputs=chat_inputs,
|
375 |
outputs=chat_outputs,
|
376 |
)
|
377 |
+
|
378 |
+
examples = [["./examples/cart1.jpg", IMAGE_PROMPT],
|
379 |
+
["./examples/cart2.jpg", IMAGE_PROMPT],
|
380 |
+
["./examples/cart3.jpg", IMAGE_PROMPT]]
|
|
|
|
|
381 |
gr.Examples(
|
382 |
examples=examples,
|
383 |
inputs=chat_inputs,
|
384 |
)
|
385 |
|
386 |
+
#########
|
387 |
+
|
388 |
if __name__ == "__main__":
|
389 |
+
demo.queue(max_size=10).launch(debug=True)
|