Spaces:
Running
on
Zero
Running
on
Zero
update app.py
Browse files
app.py
CHANGED
@@ -10,6 +10,10 @@ import time
|
|
10 |
|
11 |
import gradio as gr
|
12 |
|
|
|
|
|
|
|
|
|
13 |
|
14 |
class SAM(torch.nn.Module):
|
15 |
def __init__(self, checkpoint="/data/sam_model/sam_vit_b_01ec64.pth", **kwargs):
|
@@ -50,7 +54,8 @@ class SAM(torch.nn.Module):
|
|
50 |
|
51 |
self.image_encoder = sam.image_encoder
|
52 |
self.image_encoder.eval()
|
53 |
-
|
|
|
54 |
|
55 |
@torch.no_grad()
|
56 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
@@ -98,8 +103,9 @@ def image_sam_feature(
|
|
98 |
outputs = []
|
99 |
for i, image in enumerate(images):
|
100 |
torch_image = transform(image)
|
|
|
|
|
101 |
attn_output, mlp_output, block_output = feat_extractor(
|
102 |
-
# torch_image.unsqueeze(0).cuda()
|
103 |
torch_image.unsqueeze(0)
|
104 |
)
|
105 |
out_dict = {
|
@@ -120,7 +126,8 @@ class DiNOv2(torch.nn.Module):
|
|
120 |
self.dinov2 = torch.hub.load("facebookresearch/dinov2", ver)
|
121 |
self.dinov2.requires_grad_(False)
|
122 |
self.dinov2.eval()
|
123 |
-
|
|
|
124 |
|
125 |
def new_block_forward(self, x: torch.Tensor) -> torch.Tensor:
|
126 |
def attn_residual_func(x):
|
@@ -173,8 +180,9 @@ def image_dino_feature(images, resolution=(448, 448), node_type="block", layer=-
|
|
173 |
outputs = []
|
174 |
for i, image in enumerate(images):
|
175 |
torch_image = transform(image)
|
|
|
|
|
176 |
attn_output, mlp_output, block_output = feat_extractor(
|
177 |
-
# torch_image.unsqueeze(0).cuda()
|
178 |
torch_image.unsqueeze(0)
|
179 |
)
|
180 |
out_dict = {
|
@@ -199,8 +207,9 @@ class CLIP(torch.nn.Module):
|
|
199 |
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch16")
|
200 |
# processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch16")
|
201 |
self.model = model.eval()
|
202 |
-
|
203 |
-
|
|
|
204 |
def new_forward(
|
205 |
self,
|
206 |
hidden_states: torch.Tensor,
|
@@ -277,8 +286,9 @@ def image_clip_feature(
|
|
277 |
outputs = []
|
278 |
for i, image in enumerate(images):
|
279 |
torch_image = transform(image)
|
|
|
|
|
280 |
attn_output, mlp_output, block_output = feat_extractor(
|
281 |
-
# torch_image.unsqueeze(0).cuda()
|
282 |
torch_image.unsqueeze(0)
|
283 |
)
|
284 |
out_dict = {
|
@@ -321,7 +331,7 @@ def compute_ncut(
|
|
321 |
eigvecs, eigvals = NCUT(
|
322 |
num_eig=num_eig,
|
323 |
num_sample=num_sample_ncut,
|
324 |
-
|
325 |
affinity_focal_gamma=affinity_focal_gamma,
|
326 |
knn=knn_ncut,
|
327 |
).fit_transform(features.reshape(-1, features.shape[-1]))
|
|
|
10 |
|
11 |
import gradio as gr
|
12 |
|
13 |
+
use_cuda = torch.cuda.is_available()
|
14 |
+
|
15 |
+
print("CUDA is available:", use_cuda)
|
16 |
+
|
17 |
|
18 |
class SAM(torch.nn.Module):
|
19 |
def __init__(self, checkpoint="/data/sam_model/sam_vit_b_01ec64.pth", **kwargs):
|
|
|
54 |
|
55 |
self.image_encoder = sam.image_encoder
|
56 |
self.image_encoder.eval()
|
57 |
+
if use_cuda:
|
58 |
+
self.image_encoder = self.image_encoder.cuda()
|
59 |
|
60 |
@torch.no_grad()
|
61 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
103 |
outputs = []
|
104 |
for i, image in enumerate(images):
|
105 |
torch_image = transform(image)
|
106 |
+
if use_cuda:
|
107 |
+
torch_image = torch_image.cuda()
|
108 |
attn_output, mlp_output, block_output = feat_extractor(
|
|
|
109 |
torch_image.unsqueeze(0)
|
110 |
)
|
111 |
out_dict = {
|
|
|
126 |
self.dinov2 = torch.hub.load("facebookresearch/dinov2", ver)
|
127 |
self.dinov2.requires_grad_(False)
|
128 |
self.dinov2.eval()
|
129 |
+
if use_cuda:
|
130 |
+
self.dinov2 = self.dinov2.cuda()
|
131 |
|
132 |
def new_block_forward(self, x: torch.Tensor) -> torch.Tensor:
|
133 |
def attn_residual_func(x):
|
|
|
180 |
outputs = []
|
181 |
for i, image in enumerate(images):
|
182 |
torch_image = transform(image)
|
183 |
+
if use_cuda:
|
184 |
+
torch_image = torch_image.cuda()
|
185 |
attn_output, mlp_output, block_output = feat_extractor(
|
|
|
186 |
torch_image.unsqueeze(0)
|
187 |
)
|
188 |
out_dict = {
|
|
|
207 |
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch16")
|
208 |
# processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch16")
|
209 |
self.model = model.eval()
|
210 |
+
if use_cuda:
|
211 |
+
self.model = self.model.cuda()
|
212 |
+
|
213 |
def new_forward(
|
214 |
self,
|
215 |
hidden_states: torch.Tensor,
|
|
|
286 |
outputs = []
|
287 |
for i, image in enumerate(images):
|
288 |
torch_image = transform(image)
|
289 |
+
if use_cuda:
|
290 |
+
torch_image = torch_image.cuda()
|
291 |
attn_output, mlp_output, block_output = feat_extractor(
|
|
|
292 |
torch_image.unsqueeze(0)
|
293 |
)
|
294 |
out_dict = {
|
|
|
331 |
eigvecs, eigvals = NCUT(
|
332 |
num_eig=num_eig,
|
333 |
num_sample=num_sample_ncut,
|
334 |
+
device="cuda" if use_cuda else "cpu",
|
335 |
affinity_focal_gamma=affinity_focal_gamma,
|
336 |
knn=knn_ncut,
|
337 |
).fit_transform(features.reshape(-1, features.shape[-1]))
|