Spaces:
Running
on
Zero
Running
on
Zero
update dataset
Browse files- app.py +131 -58
- requirements.txt +1 -1
app.py
CHANGED
@@ -1,4 +1,10 @@
|
|
1 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
import gradio as gr
|
3 |
|
4 |
import torch
|
@@ -8,8 +14,11 @@ import time
|
|
8 |
|
9 |
import gradio as gr
|
10 |
|
11 |
-
|
12 |
-
from
|
|
|
|
|
|
|
13 |
|
14 |
|
15 |
def compute_ncut(
|
@@ -24,8 +33,17 @@ def compute_ncut(
|
|
24 |
perplexity=150,
|
25 |
n_neighbors=150,
|
26 |
min_dist=0.1,
|
|
|
27 |
):
|
28 |
logging_str = ""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
start = time.time()
|
30 |
eigvecs, eigvals = NCUT(
|
31 |
num_eig=num_eig,
|
@@ -33,33 +51,24 @@ def compute_ncut(
|
|
33 |
device="cuda" if torch.cuda.is_available() else "cpu",
|
34 |
affinity_focal_gamma=affinity_focal_gamma,
|
35 |
knn=knn_ncut,
|
|
|
36 |
).fit_transform(features.reshape(-1, features.shape[-1]))
|
37 |
# print(f"NCUT time: {time.time() - start:.2f}s")
|
38 |
logging_str += f"NCUT time: {time.time() - start:.2f}s\n"
|
39 |
|
40 |
start = time.time()
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
-
|
53 |
-
num_sample=num_sample_tsne,
|
54 |
-
perplexity=perplexity,
|
55 |
-
knn=knn_tsne,
|
56 |
-
device="cuda" if torch.cuda.is_available() else "cpu",
|
57 |
-
)
|
58 |
-
# print(f"t-SNE time: {time.time() - start:.2f}s")
|
59 |
-
logging_str += f"t-SNE time: {time.time() - start:.2f}s\n"
|
60 |
-
else:
|
61 |
-
raise ValueError(f"Embedding method {embedding_method} not supported.")
|
62 |
-
|
63 |
rgb = rgb.reshape(features.shape[:3] + (3,))
|
64 |
return rgb, logging_str
|
65 |
|
@@ -76,7 +85,7 @@ def dont_use_too_much_green(image_rgb):
|
|
76 |
|
77 |
def to_pil_images(images):
|
78 |
return [
|
79 |
-
Image.fromarray((image * 255).cpu().numpy().astype(np.uint8)).resize((256, 256), Image.NEAREST)
|
80 |
for image in images
|
81 |
]
|
82 |
|
@@ -103,11 +112,13 @@ def ncut_run(
|
|
103 |
perplexity=500,
|
104 |
n_neighbors=500,
|
105 |
min_dist=0.1,
|
|
|
106 |
):
|
107 |
logging_str = ""
|
108 |
if perplexity >= num_sample_tsne or n_neighbors >= num_sample_tsne:
|
109 |
# raise gr.Error("Perplexity must be less than the number of samples for t-SNE.")
|
110 |
-
gr.Warning("Perplexity/n_neighbors must be less than the number of samples.\n" f"Setting to {num_sample_tsne-1}.")
|
|
|
111 |
perplexity = num_sample_tsne - 1
|
112 |
n_neighbors = num_sample_tsne - 1
|
113 |
|
@@ -135,26 +146,48 @@ def ncut_run(
|
|
135 |
perplexity=perplexity,
|
136 |
n_neighbors=n_neighbors,
|
137 |
min_dist=min_dist,
|
|
|
138 |
)
|
139 |
logging_str += _logging_str
|
140 |
rgb = dont_use_too_much_green(rgb)
|
141 |
return to_pil_images(rgb), logging_str
|
142 |
|
143 |
-
|
144 |
-
|
145 |
-
|
|
|
|
|
|
|
146 |
|
147 |
-
|
148 |
-
|
149 |
-
|
|
|
150 |
|
151 |
-
@spaces.GPU(duration=
|
152 |
-
def
|
153 |
-
|
154 |
|
155 |
-
@spaces.GPU(duration=
|
156 |
-
def
|
157 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
158 |
|
159 |
def run_fn(
|
160 |
images,
|
@@ -171,11 +204,15 @@ def run_fn(
|
|
171 |
perplexity=500,
|
172 |
n_neighbors=500,
|
173 |
min_dist=0.1,
|
|
|
174 |
):
|
175 |
if images is None:
|
176 |
gr.Warning("No images selected.")
|
177 |
return [], "No images selected."
|
178 |
|
|
|
|
|
|
|
179 |
kwargs = {
|
180 |
"model_name": model_name,
|
181 |
"layer": layer,
|
@@ -190,6 +227,7 @@ def run_fn(
|
|
190 |
"perplexity": perplexity,
|
191 |
"n_neighbors": n_neighbors,
|
192 |
"min_dist": min_dist,
|
|
|
193 |
}
|
194 |
num_images = len(images)
|
195 |
if num_images > 100:
|
@@ -216,34 +254,48 @@ with gr.Blocks() as demo:
|
|
216 |
submit_button = gr.Button("🔴RUN", elem_id="submit_button")
|
217 |
clear_images_button = gr.Button("🗑️Clear", elem_id='clear_button')
|
218 |
|
219 |
-
gr.Markdown('### Load
|
220 |
-
load_images_button = gr.Button("Load", elem_id="load-images-button")
|
221 |
-
hide_button = gr.Button("Hide", elem_id="hide-button")
|
222 |
example_gallery = gr.Gallery(value=example_items, label="Example Set A", show_label=False, columns=[3], rows=[2], object_fit="scale-down", height="200px", show_share_button=False, elem_id="example-gallery")
|
|
|
223 |
|
224 |
hide_button.click(
|
225 |
fn=lambda: gr.update(visible=False),
|
226 |
outputs=example_gallery
|
227 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
228 |
|
229 |
with gr.Column(scale=5, min_width=200):
|
230 |
gr.Markdown('### Output Images')
|
231 |
output_gallery = gr.Gallery(value=[], label="NCUT Embedding", show_label=False, elem_id="ncut", columns=[3], rows=[1], object_fit="contain", height="auto")
|
232 |
-
model_dropdown = gr.Dropdown(["SAM(sam_vit_b)", "MobileSAM", "DiNO(dinov2_vitb14_reg)", "CLIP(openai/clip-vit-base-patch16)"], label="
|
233 |
-
layer_slider = gr.Slider(0, 11, step=1, label="Layer", value=11, elem_id="layer")
|
234 |
-
|
235 |
-
|
|
|
236 |
|
237 |
-
with gr.Accordion("
|
238 |
-
|
239 |
-
|
240 |
-
knn_ncut_slider = gr.Slider(1, 100, step=1, label="KNN
|
241 |
-
embedding_method_dropdown = gr.Dropdown(["
|
242 |
-
num_sample_tsne_slider = gr.Slider(100, 1000, step=100, label="
|
243 |
-
knn_tsne_slider = gr.Slider(1, 100, step=1, label="
|
244 |
-
perplexity_slider = gr.Slider(10, 500, step=10, label="
|
245 |
-
n_neighbors_slider = gr.Slider(10, 500, step=10, label="n_neighbors
|
246 |
-
min_dist_slider = gr.Slider(0.1, 1, step=0.1, label="min_dist
|
247 |
|
248 |
# logging text box
|
249 |
logging_text = gr.Textbox("Logging information", label="Logging", elem_id="logging", type="text", placeholder="Logging information")
|
@@ -253,19 +305,40 @@ with gr.Blocks() as demo:
|
|
253 |
|
254 |
def empty_input_and_output():
|
255 |
return [], []
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
256 |
|
257 |
load_images_button.click(load_default_images, outputs=[input_gallery, output_gallery])
|
258 |
clear_images_button.click(empty_input_and_output, outputs=[input_gallery, output_gallery])
|
|
|
259 |
submit_button.click(
|
260 |
run_fn,
|
261 |
inputs=[
|
262 |
input_gallery, model_dropdown, layer_slider, num_eig_slider, node_type_dropdown,
|
263 |
affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider,
|
264 |
embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider,
|
265 |
-
perplexity_slider, n_neighbors_slider, min_dist_slider
|
266 |
],
|
267 |
outputs=[output_gallery, logging_text]
|
268 |
)
|
269 |
|
270 |
|
271 |
-
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Author: Huzheng Yang
|
2 |
+
# %%
|
3 |
+
USE_SPACES = False
|
4 |
+
|
5 |
+
if USE_SPACES:
|
6 |
+
import spaces
|
7 |
+
|
8 |
import gradio as gr
|
9 |
|
10 |
import torch
|
|
|
14 |
|
15 |
import gradio as gr
|
16 |
|
17 |
+
if USE_SPACES:
|
18 |
+
from backbone import extract_features
|
19 |
+
else:
|
20 |
+
from draft_gradio_backbone import extract_features
|
21 |
+
from ncut_pytorch import NCUT, eigenvector_to_rgb
|
22 |
|
23 |
|
24 |
def compute_ncut(
|
|
|
33 |
perplexity=150,
|
34 |
n_neighbors=150,
|
35 |
min_dist=0.1,
|
36 |
+
sampling_method="fps",
|
37 |
):
|
38 |
logging_str = ""
|
39 |
+
|
40 |
+
num_nodes = np.prod(features.shape[:3])
|
41 |
+
if num_nodes / 2 < num_eig:
|
42 |
+
# raise gr.Error("Number of eigenvectors should be less than half the number of nodes.")
|
43 |
+
gr.Warning("Number of eigenvectors should be less than half the number of nodes.\n" f"Setting num_eig to {num_nodes // 2 - 1}.")
|
44 |
+
num_eig = num_nodes // 2 - 1
|
45 |
+
logging_str += f"Number of eigenvectors should be less than half the number of nodes.\n" f"Setting num_eig to {num_nodes // 2 - 1}.\n"
|
46 |
+
|
47 |
start = time.time()
|
48 |
eigvecs, eigvals = NCUT(
|
49 |
num_eig=num_eig,
|
|
|
51 |
device="cuda" if torch.cuda.is_available() else "cpu",
|
52 |
affinity_focal_gamma=affinity_focal_gamma,
|
53 |
knn=knn_ncut,
|
54 |
+
sample_method=sampling_method,
|
55 |
).fit_transform(features.reshape(-1, features.shape[-1]))
|
56 |
# print(f"NCUT time: {time.time() - start:.2f}s")
|
57 |
logging_str += f"NCUT time: {time.time() - start:.2f}s\n"
|
58 |
|
59 |
start = time.time()
|
60 |
+
_, rgb = eigenvector_to_rgb(
|
61 |
+
eigvecs,
|
62 |
+
method=embedding_method,
|
63 |
+
num_sample=num_sample_tsne,
|
64 |
+
perplexity=perplexity,
|
65 |
+
n_neighbors=n_neighbors,
|
66 |
+
min_distance=min_dist,
|
67 |
+
knn=knn_tsne,
|
68 |
+
device="cuda" if torch.cuda.is_available() else "cpu",
|
69 |
+
)
|
70 |
+
logging_str += f"{embedding_method} time: {time.time() - start:.2f}s\n"
|
71 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
rgb = rgb.reshape(features.shape[:3] + (3,))
|
73 |
return rgb, logging_str
|
74 |
|
|
|
85 |
|
86 |
def to_pil_images(images):
|
87 |
return [
|
88 |
+
Image.fromarray((image * 255).cpu().numpy().astype(np.uint8)).resize((256, 256), Image.Resampling.NEAREST)
|
89 |
for image in images
|
90 |
]
|
91 |
|
|
|
112 |
perplexity=500,
|
113 |
n_neighbors=500,
|
114 |
min_dist=0.1,
|
115 |
+
sampling_method="fps",
|
116 |
):
|
117 |
logging_str = ""
|
118 |
if perplexity >= num_sample_tsne or n_neighbors >= num_sample_tsne:
|
119 |
# raise gr.Error("Perplexity must be less than the number of samples for t-SNE.")
|
120 |
+
gr.Warning("Perplexity/n_neighbors must be less than the number of samples.\n" f"Setting Perplexity to {num_sample_tsne-1}.")
|
121 |
+
logging_str += f"Perplexity/n_neighbors must be less than the number of samples.\n" f"Setting Perplexity to {num_sample_tsne-1}.\n"
|
122 |
perplexity = num_sample_tsne - 1
|
123 |
n_neighbors = num_sample_tsne - 1
|
124 |
|
|
|
146 |
perplexity=perplexity,
|
147 |
n_neighbors=n_neighbors,
|
148 |
min_dist=min_dist,
|
149 |
+
sampling_method=sampling_method,
|
150 |
)
|
151 |
logging_str += _logging_str
|
152 |
rgb = dont_use_too_much_green(rgb)
|
153 |
return to_pil_images(rgb), logging_str
|
154 |
|
155 |
+
def _ncut_run(*args, **kwargs):
|
156 |
+
try:
|
157 |
+
return ncut_run(*args, **kwargs)
|
158 |
+
except Exception as e:
|
159 |
+
gr.Error(str(e))
|
160 |
+
return [], "Error: " + str(e)
|
161 |
|
162 |
+
if USE_SPACES:
|
163 |
+
@spaces.GPU(duration=13)
|
164 |
+
def quick_run(*args, **kwargs):
|
165 |
+
return _ncut_run(*args, **kwargs)
|
166 |
|
167 |
+
@spaces.GPU(duration=30)
|
168 |
+
def long_run(*args, **kwargs):
|
169 |
+
return _ncut_run(*args, **kwargs)
|
170 |
|
171 |
+
@spaces.GPU(duration=60)
|
172 |
+
def longer_run(*args, **kwargs):
|
173 |
+
return _ncut_run(*args, **kwargs)
|
174 |
+
|
175 |
+
@spaces.GPU(duration=120)
|
176 |
+
def super_duper_long_run(*args, **kwargs):
|
177 |
+
return _ncut_run(*args, **kwargs)
|
178 |
+
|
179 |
+
if not USE_SPACES:
|
180 |
+
def quick_run(*args, **kwargs):
|
181 |
+
return _ncut_run(*args, **kwargs)
|
182 |
+
|
183 |
+
def long_run(*args, **kwargs):
|
184 |
+
return _ncut_run(*args, **kwargs)
|
185 |
+
|
186 |
+
def longer_run(*args, **kwargs):
|
187 |
+
return _ncut_run(*args, **kwargs)
|
188 |
+
|
189 |
+
def super_duper_long_run(*args, **kwargs):
|
190 |
+
return _ncut_run(*args, **kwargs)
|
191 |
|
192 |
def run_fn(
|
193 |
images,
|
|
|
204 |
perplexity=500,
|
205 |
n_neighbors=500,
|
206 |
min_dist=0.1,
|
207 |
+
sampling_method="fps",
|
208 |
):
|
209 |
if images is None:
|
210 |
gr.Warning("No images selected.")
|
211 |
return [], "No images selected."
|
212 |
|
213 |
+
if sampling_method == "fps":
|
214 |
+
sampling_method = "farthest"
|
215 |
+
|
216 |
kwargs = {
|
217 |
"model_name": model_name,
|
218 |
"layer": layer,
|
|
|
227 |
"perplexity": perplexity,
|
228 |
"n_neighbors": n_neighbors,
|
229 |
"min_dist": min_dist,
|
230 |
+
"sampling_method": sampling_method,
|
231 |
}
|
232 |
num_images = len(images)
|
233 |
if num_images > 100:
|
|
|
254 |
submit_button = gr.Button("🔴RUN", elem_id="submit_button")
|
255 |
clear_images_button = gr.Button("🗑️Clear", elem_id='clear_button')
|
256 |
|
257 |
+
gr.Markdown('### Load from Cloud Dataset 👇')
|
258 |
+
load_images_button = gr.Button("Load Example", elem_id="load-images-button")
|
|
|
259 |
example_gallery = gr.Gallery(value=example_items, label="Example Set A", show_label=False, columns=[3], rows=[2], object_fit="scale-down", height="200px", show_share_button=False, elem_id="example-gallery")
|
260 |
+
hide_button = gr.Button("Hide Example", elem_id="hide-button")
|
261 |
|
262 |
hide_button.click(
|
263 |
fn=lambda: gr.update(visible=False),
|
264 |
outputs=example_gallery
|
265 |
)
|
266 |
+
|
267 |
+
with gr.Accordion("➜ Load from dataset", open=True):
|
268 |
+
dataset_names = [
|
269 |
+
'UCSC-VLAA/Recap-COCO-30K',
|
270 |
+
'nateraw/pascal-voc-2012',
|
271 |
+
'johnowhitaker/imagenette2-320',
|
272 |
+
'JapanDegitalMaterial/Places_in_Japan',
|
273 |
+
'Borismile/Anime-dataset',
|
274 |
+
]
|
275 |
+
dataset_dropdown = gr.Dropdown(dataset_names, label="Dataset name", value="UCSC-VLAA/Recap-COCO-30K", elem_id="dataset")
|
276 |
+
num_images_slider = gr.Slider(1, 200, step=1, label="Number of images", value=9, elem_id="num_images")
|
277 |
+
random_seed_slider = gr.Number(0, label="Random seed", value=42, elem_id="random_seed")
|
278 |
+
load_dataset_button = gr.Button("Load Dataset", elem_id="load-dataset-button")
|
279 |
|
280 |
with gr.Column(scale=5, min_width=200):
|
281 |
gr.Markdown('### Output Images')
|
282 |
output_gallery = gr.Gallery(value=[], label="NCUT Embedding", show_label=False, elem_id="ncut", columns=[3], rows=[1], object_fit="contain", height="auto")
|
283 |
+
model_dropdown = gr.Dropdown(["SAM(sam_vit_b)", "MobileSAM", "DiNO(dinov2_vitb14_reg)", "CLIP(openai/clip-vit-base-patch16)", "MAE(vit_base)"], label="Backbone", value="SAM(sam_vit_b)", elem_id="model_name")
|
284 |
+
layer_slider = gr.Slider(0, 11, step=1, label="Backbone: Layer index", value=11, elem_id="layer")
|
285 |
+
node_type_dropdown = gr.Dropdown(["attn: attention output", "mlp: mlp output", "block: sum of residual"], label="Backbone: Layer type", value="block: sum of residual", elem_id="node_type", info="which feature to take from each layer?")
|
286 |
+
num_eig_slider = gr.Slider(1, 1000, step=1, label="NCUT: Number of eigenvectors", value=100, elem_id="num_eig", info='increase for more clusters')
|
287 |
+
affinity_focal_gamma_slider = gr.Slider(0.01, 1, step=0.01, label="NCUT: Affinity focal gamma", value=0.5, elem_id="affinity_focal_gamma", info="decrease for shaper segmentation")
|
288 |
|
289 |
+
with gr.Accordion("➜ Click to expand: more parameters", open=False):
|
290 |
+
num_sample_ncut_slider = gr.Slider(100, 50000, step=100, label="NCUT: num_sample", value=10000, elem_id="num_sample_ncut", info="Nyström approximation")
|
291 |
+
sampling_method_dropdown = gr.Dropdown(["fps", "random"], label="NCUT: Sampling method", value="fps", elem_id="sampling_method")
|
292 |
+
knn_ncut_slider = gr.Slider(1, 100, step=1, label="NCUT: KNN", value=10, elem_id="knn_ncut", info="Nyström approximation")
|
293 |
+
embedding_method_dropdown = gr.Dropdown(["tsne_3d", "umap_3d", "umap_shpere", "tsne_2d", "umap_2d"], label="Coloring method", value="tsne_3d", elem_id="embedding_method")
|
294 |
+
num_sample_tsne_slider = gr.Slider(100, 1000, step=100, label="t-SNE/UMAP: num_sample", value=300, elem_id="num_sample_tsne", info="Nyström approximation")
|
295 |
+
knn_tsne_slider = gr.Slider(1, 100, step=1, label="t-SNE/UMAP: KNN", value=10, elem_id="knn_tsne", info="Nyström approximation")
|
296 |
+
perplexity_slider = gr.Slider(10, 500, step=10, label="t-SNE: Perplexity", value=150, elem_id="perplexity")
|
297 |
+
n_neighbors_slider = gr.Slider(10, 500, step=10, label="UMAP: n_neighbors", value=150, elem_id="n_neighbors")
|
298 |
+
min_dist_slider = gr.Slider(0.1, 1, step=0.1, label="UMAP: min_dist", value=0.1, elem_id="min_dist")
|
299 |
|
300 |
# logging text box
|
301 |
logging_text = gr.Textbox("Logging information", label="Logging", elem_id="logging", type="text", placeholder="Logging information")
|
|
|
305 |
|
306 |
def empty_input_and_output():
|
307 |
return [], []
|
308 |
+
|
309 |
+
def load_dataset_images(dataset_name, num_images=10, random_seed=42):
|
310 |
+
from datasets import load_dataset
|
311 |
+
try:
|
312 |
+
dataset = load_dataset(dataset_name)['train']
|
313 |
+
except Exception as e:
|
314 |
+
gr.Error(f"Error loading dataset {dataset_name}: {e}")
|
315 |
+
return None
|
316 |
+
if num_images > len(dataset):
|
317 |
+
num_images = len(dataset)
|
318 |
+
image_idx = np.random.RandomState(random_seed).choice(len(dataset), num_images, replace=False)
|
319 |
+
image_idx = image_idx.tolist()
|
320 |
+
images = [dataset[i]['image'] for i in image_idx]
|
321 |
+
return images
|
322 |
+
|
323 |
|
324 |
load_images_button.click(load_default_images, outputs=[input_gallery, output_gallery])
|
325 |
clear_images_button.click(empty_input_and_output, outputs=[input_gallery, output_gallery])
|
326 |
+
load_dataset_button.click(load_dataset_images, inputs=[dataset_dropdown, num_images_slider, random_seed_slider], outputs=[input_gallery])
|
327 |
submit_button.click(
|
328 |
run_fn,
|
329 |
inputs=[
|
330 |
input_gallery, model_dropdown, layer_slider, num_eig_slider, node_type_dropdown,
|
331 |
affinity_focal_gamma_slider, num_sample_ncut_slider, knn_ncut_slider,
|
332 |
embedding_method_dropdown, num_sample_tsne_slider, knn_tsne_slider,
|
333 |
+
perplexity_slider, n_neighbors_slider, min_dist_slider, sampling_method_dropdown
|
334 |
],
|
335 |
outputs=[output_gallery, logging_text]
|
336 |
)
|
337 |
|
338 |
|
339 |
+
if USE_SPACES:
|
340 |
+
demo.launch()
|
341 |
+
else:
|
342 |
+
demo.launch(share=True)
|
343 |
+
|
344 |
+
# %%
|
requirements.txt
CHANGED
@@ -2,7 +2,7 @@ torch
|
|
2 |
torchvision
|
3 |
ncut-pytorch
|
4 |
transformers
|
|
|
5 |
segment-anything @ git+https://github.com/facebookresearch/segment-anything.git
|
6 |
mobile-sam @ git+https://github.com/ChaoningZhang/MobileSAM.git
|
7 |
-
|
8 |
timm
|
|
|
2 |
torchvision
|
3 |
ncut-pytorch
|
4 |
transformers
|
5 |
+
datasets
|
6 |
segment-anything @ git+https://github.com/facebookresearch/segment-anything.git
|
7 |
mobile-sam @ git+https://github.com/ChaoningZhang/MobileSAM.git
|
|
|
8 |
timm
|