Spaces:
Build error
Build error
englert
commited on
Commit
β’
ce0d4fb
1
Parent(s):
fc405ab
add all files
Browse files- README.md +1 -1
- app.py +61 -15
- fastdist2.py +332 -0
- requirements.txt +4 -0
- sampling_util.py +30 -0
- video_reader.py +35 -0
README.md
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
---
|
2 |
title: Visdif
|
3 |
-
emoji:
|
4 |
colorFrom: green
|
5 |
colorTo: yellow
|
6 |
sdk: gradio
|
|
|
1 |
---
|
2 |
title: Visdif
|
3 |
+
emoji: π
|
4 |
colorFrom: green
|
5 |
colorTo: yellow
|
6 |
sdk: gradio
|
app.py
CHANGED
@@ -3,23 +3,69 @@ import torch
|
|
3 |
import requests
|
4 |
from torchvision import transforms
|
5 |
|
6 |
-
|
7 |
-
|
8 |
-
labels = response.text.split("\n")
|
9 |
|
10 |
-
torch.
|
|
|
11 |
|
12 |
-
def predict(inp):
|
13 |
-
inp = transforms.ToTensor()(inp).unsqueeze(0)
|
14 |
-
with torch.no_grad():
|
15 |
-
prediction = torch.nn.functional.softmax(model(inp)[0], dim=0)
|
16 |
-
confidences = {labels[i]: float(prediction[i]) for i in range(1000)}
|
17 |
-
return confidences
|
18 |
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
demo.launch()
|
|
|
3 |
import requests
|
4 |
from torchvision import transforms
|
5 |
|
6 |
+
from sampling_util import furthest_neighbours
|
7 |
+
from video_reader import video_reader
|
|
|
8 |
|
9 |
+
model = torch.load("model").eval()
|
10 |
+
avg_pool = nn.AdaptiveAvgPool2d((1, 1))
|
11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
+
def predict(input_file):
|
14 |
+
base_directory = os.getcwd()
|
15 |
+
selected_directory = os.path.join(base_directory, "selected_images")
|
16 |
+
if os.path.isdir(selected_directory):
|
17 |
+
shutil.rmtree(selected_directory)
|
18 |
+
os.mkdir(selected_directory)
|
19 |
+
|
20 |
+
zip_path = os.path.join(input_file.split('/')[-1][:-4] + ".zip")
|
21 |
+
|
22 |
+
mean = [0.3156024, 0.33569682, 0.34337464]
|
23 |
+
std = [0.16568947, 0.17827448, 0.18925823]
|
24 |
+
|
25 |
+
img_vecs = []
|
26 |
+
with torch.no_grad():
|
27 |
+
for fp_i, file_path in enumerate([input_file]):
|
28 |
+
for i, in_img in enumerate(video_reader(file_path,
|
29 |
+
targetFPS=9,
|
30 |
+
targetWidth=100,
|
31 |
+
to_rgb=True)):
|
32 |
+
in_img = (in_img.astype(np.float32) / 255.)
|
33 |
+
in_img = (in_img - mean) / std
|
34 |
+
in_img = np.transpose(in_img, (0, 3, 1, 2))
|
35 |
+
in_img = torch.from_numpy(in_img)
|
36 |
+
encoded = avg_pool(model(in_img))[0, :, 0, 0].cpu().numpy()
|
37 |
+
img_vecs += [encoded]
|
38 |
+
|
39 |
+
img_vecs = np.asarray(img_vecs)
|
40 |
+
rv_indices, _ = furthest_neighbours(
|
41 |
+
img_vecs,
|
42 |
+
downsample_size,
|
43 |
+
seed=0)
|
44 |
+
indices = np.zeros((img_vecs.shape[0],))
|
45 |
+
indices[np.asarray(rv_indices)] = 1
|
46 |
+
|
47 |
+
global_ctr = 0
|
48 |
+
for fp_i, file_path in enumerate([input_file]):
|
49 |
+
for i, img in enumerate(video_reader(file_path,
|
50 |
+
targetFPS=9,
|
51 |
+
targetWidth=None,
|
52 |
+
to_rgb=False)):
|
53 |
+
if indices[global_ctr] == 1:
|
54 |
+
cv2.imwrite(join(selected_directory, str(global_ctr) + ".jpg"), img)
|
55 |
+
global_ctr += 1
|
56 |
+
|
57 |
+
all_selected_imgs_path = [join(selected_directory, f) for f in listdir(selected_directory) if isfile(join(selected_directory, f))]
|
58 |
+
if 0 < len(all_file_paths):
|
59 |
+
zipf = zipfile.ZipFile(zip_path, 'w', zipfile.ZIP_DEFLATED)
|
60 |
+
for i, f in enumerate(all_selected_imgs_path):
|
61 |
+
zipf.write(f, basename(f))
|
62 |
+
zipf.close()
|
63 |
+
|
64 |
+
return zip_path
|
65 |
+
|
66 |
+
demo = gr.Interface(
|
67 |
+
fn=predict,
|
68 |
+
inputs=gr.inputs.Video(label="Upload Video File"),
|
69 |
+
outputs=gr.outputs.File(label="Zip"))
|
70 |
|
71 |
demo.launch()
|
fastdist2.py
ADDED
@@ -0,0 +1,332 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
from numba import jit, prange, cuda, float32
|
5 |
+
|
6 |
+
|
7 |
+
# https://github.com/talboger/fastdist
|
8 |
+
|
9 |
+
@jit(nopython=True, fastmath=True)
|
10 |
+
def cosine(u, v, w=None):
|
11 |
+
"""
|
12 |
+
:purpose:
|
13 |
+
Computes the cosine similarity between two 1D arrays
|
14 |
+
Unlike scipy's cosine distance, this returns similarity, which is 1 - distance
|
15 |
+
|
16 |
+
:params:
|
17 |
+
u, v : input arrays, both of shape (n,)
|
18 |
+
w : weights at each index of u and v. array of shape (n,)
|
19 |
+
if no w is set, it is initialized as an array of ones
|
20 |
+
such that it will have no impact on the output
|
21 |
+
|
22 |
+
:returns:
|
23 |
+
cosine : float, the cosine similarity between u and v
|
24 |
+
|
25 |
+
:example:
|
26 |
+
>>> from fastdist import fastdist
|
27 |
+
>>> import numpy as np
|
28 |
+
>>> u, v, w = np.random.RandomState(seed=0).rand(10000, 3).T
|
29 |
+
>>> fastdist.cosine(u, v, w)
|
30 |
+
0.7495065944399267
|
31 |
+
"""
|
32 |
+
n = len(u)
|
33 |
+
num = 0
|
34 |
+
u_norm, v_norm = 0, 0
|
35 |
+
for i in range(n):
|
36 |
+
num += u[i] * v[i] * w[i]
|
37 |
+
u_norm += abs(u[i]) ** 2 * w[i]
|
38 |
+
v_norm += abs(v[i]) ** 2 * w[i]
|
39 |
+
|
40 |
+
denom = (u_norm * v_norm) ** (1 / 2)
|
41 |
+
return num / denom
|
42 |
+
|
43 |
+
|
44 |
+
@jit(nopython=True, fastmath=True)
|
45 |
+
def cosine_vector_to_matrix(u, m):
|
46 |
+
"""
|
47 |
+
:purpose:
|
48 |
+
Computes the cosine similarity between a 1D array and rows of a matrix
|
49 |
+
|
50 |
+
:params:
|
51 |
+
u : input vector of shape (n,)
|
52 |
+
m : input matrix of shape (m, n)
|
53 |
+
|
54 |
+
:returns:
|
55 |
+
cosine vector : np.array, of shape (m,) vector containing cosine similarity between u
|
56 |
+
and the rows of m
|
57 |
+
|
58 |
+
:example:
|
59 |
+
>>> from fastdist import fastdist
|
60 |
+
>>> import numpy as np
|
61 |
+
>>> u = np.random.RandomState(seed=0).rand(10)
|
62 |
+
>>> m = np.random.RandomState(seed=0).rand(100, 10)
|
63 |
+
>>> fastdist.cosine_vector_to_matrix(u, m)
|
64 |
+
(returns an array of shape (100,))
|
65 |
+
"""
|
66 |
+
norm = 0
|
67 |
+
for i in range(len(u)):
|
68 |
+
norm += abs(u[i]) ** 2
|
69 |
+
u = u / norm ** (1 / 2)
|
70 |
+
for i in range(m.shape[0]):
|
71 |
+
norm = 0
|
72 |
+
for j in range(len(m[i])):
|
73 |
+
norm += abs(m[i][j]) ** 2
|
74 |
+
m[i] = m[i] / norm ** (1 / 2)
|
75 |
+
return np.dot(u, m.T)
|
76 |
+
|
77 |
+
|
78 |
+
@jit(nopython=True, fastmath=True)
|
79 |
+
def cosine_matrix_to_matrix(a, b):
|
80 |
+
"""
|
81 |
+
:purpose:
|
82 |
+
Computes the cosine similarity between the rows of two matrices
|
83 |
+
|
84 |
+
:params:
|
85 |
+
a, b : input matrices of shape (m, n) and (k, n)
|
86 |
+
the matrices must share a common dimension at index 1
|
87 |
+
|
88 |
+
:returns:
|
89 |
+
cosine matrix : np.array, an (m, k) array of the cosine similarity
|
90 |
+
between the rows of a and b
|
91 |
+
|
92 |
+
:example:
|
93 |
+
>>> from fastdist import fastdist
|
94 |
+
>>> import numpy as np
|
95 |
+
>>> a = np.random.RandomState(seed=0).rand(10, 50)
|
96 |
+
>>> b = np.random.RandomState(seed=0).rand(100, 50)
|
97 |
+
>>> fastdist.cosine_matrix_to_matrix(a, b)
|
98 |
+
(returns an array of shape (10, 100))
|
99 |
+
"""
|
100 |
+
for i in range(a.shape[0]):
|
101 |
+
norm = 0
|
102 |
+
for j in range(len(a[i])):
|
103 |
+
norm += abs(a[i][j]) ** 2
|
104 |
+
a[i] = a[i] / norm ** (1 / 2)
|
105 |
+
for i in range(b.shape[0]):
|
106 |
+
norm = 0
|
107 |
+
for j in range(len(b[i])):
|
108 |
+
norm += abs(b[i][j]) ** 2
|
109 |
+
b[i] = b[i] / norm ** (1 / 2)
|
110 |
+
return np.dot(a, b.T)
|
111 |
+
|
112 |
+
|
113 |
+
@jit(nopython=True, fastmath=True)
|
114 |
+
def euclidean(u, v):
|
115 |
+
"""
|
116 |
+
:purpose:
|
117 |
+
Computes the Euclidean distance between two 1D arrays
|
118 |
+
|
119 |
+
:params:
|
120 |
+
u, v : input arrays, both of shape (n,)
|
121 |
+
w : weights at each index of u and v. array of shape (n,)
|
122 |
+
if no w is set, it is initialized as an array of ones
|
123 |
+
such that it will have no impact on the output
|
124 |
+
|
125 |
+
:returns:
|
126 |
+
euclidean : float, the Euclidean distance between u and v
|
127 |
+
|
128 |
+
:example:
|
129 |
+
>>> from fastdist import fastdist
|
130 |
+
>>> import numpy as np
|
131 |
+
>>> u, v, w = np.random.RandomState(seed=0).rand(10000, 3).T
|
132 |
+
>>> fastdist.euclidean(u, v, w)
|
133 |
+
28.822558591834163
|
134 |
+
"""
|
135 |
+
n = len(u)
|
136 |
+
dist = 0
|
137 |
+
for i in range(n):
|
138 |
+
dist += abs(u[i] - v[i]) ** 2
|
139 |
+
return dist ** (1 / 2)
|
140 |
+
|
141 |
+
|
142 |
+
@jit(nopython=True, fastmath=True)
|
143 |
+
def euclidean_vector_to_matrix_distance(u, m):
|
144 |
+
"""
|
145 |
+
:purpose:
|
146 |
+
Computes the distance between a vector and the rows of a matrix using any given metric
|
147 |
+
|
148 |
+
:params:
|
149 |
+
u : input vector of shape (n,)
|
150 |
+
m : input matrix of shape (m, n)
|
151 |
+
|
152 |
+
distance vector : np.array, of shape (m,) vector containing the distance between u
|
153 |
+
and the rows of m
|
154 |
+
|
155 |
+
:example:
|
156 |
+
>>> from fastdist import fastdist
|
157 |
+
>>> import numpy as np
|
158 |
+
>>> u = np.random.RandomState(seed=0).rand(10)
|
159 |
+
>>> m = np.random.RandomState(seed=0).rand(100, 10)
|
160 |
+
>>> fastdist.vector_to_matrix_distance(u, m)
|
161 |
+
(returns an array of shape (100,))
|
162 |
+
|
163 |
+
:note:
|
164 |
+
the cosine similarity uses its own function, cosine_vector_to_matrix.
|
165 |
+
this is because normalizing the rows and then taking the dot product
|
166 |
+
of the vector and matrix heavily optimizes the computation. the other similarity
|
167 |
+
metrics do not have such an optimization, so we loop through them
|
168 |
+
"""
|
169 |
+
|
170 |
+
n = m.shape[0]
|
171 |
+
out = np.zeros((n), dtype=np.float32)
|
172 |
+
for i in prange(n):
|
173 |
+
dist = 0
|
174 |
+
for l in range(len(u)):
|
175 |
+
dist += abs(u[l] - m[i][l]) ** 2
|
176 |
+
out[i] = dist ** (1 / 2)
|
177 |
+
|
178 |
+
return out
|
179 |
+
|
180 |
+
|
181 |
+
@cuda.jit
|
182 |
+
def gpu_kernel_euclidean_vector_to_matrix_distance(u, m, u_dim0, m_dim0, out):
|
183 |
+
# Thread id in a 1D block
|
184 |
+
tx = cuda.threadIdx.x
|
185 |
+
# Block id in a 1D grid
|
186 |
+
ty = cuda.blockIdx.x
|
187 |
+
# Block width, i.e. number of threads per block
|
188 |
+
bw = cuda.blockDim.x
|
189 |
+
# Compute flattened index inside the array
|
190 |
+
pos = tx + ty * bw
|
191 |
+
if pos < m_dim0: # Check array boundaries
|
192 |
+
dist = 0
|
193 |
+
for l in range(u_dim0):
|
194 |
+
d = abs(u[l] - m[pos][l])
|
195 |
+
dist += d * d
|
196 |
+
out[pos] = dist ** (1 / 2)
|
197 |
+
|
198 |
+
|
199 |
+
def euclidean_vector_to_matrix_distance_gpu(u, m):
|
200 |
+
m_dim0 = m.shape[0]
|
201 |
+
u_dim0 = u.shape[0]
|
202 |
+
out = np.zeros((m_dim0), dtype=np.float32)
|
203 |
+
|
204 |
+
threadsperblock = 16
|
205 |
+
blockspergrid = (m_dim0 + (threadsperblock - 1)) // threadsperblock
|
206 |
+
gpu_kernel_euclidean_vector_to_matrix_distance[blockspergrid, threadsperblock](u, m, u_dim0, m_dim0, out)
|
207 |
+
|
208 |
+
return out
|
209 |
+
|
210 |
+
|
211 |
+
# https://numba.readthedocs.io/en/stable/cuda/examples.html
|
212 |
+
@cuda.jit
|
213 |
+
def gpu_kernel_euclidean_matrix_to_matrix_distance_fast(A, B, C):
|
214 |
+
TPB = 16
|
215 |
+
|
216 |
+
# Define an array in the shared memory
|
217 |
+
# The size and type of the arrays must be known at compile time
|
218 |
+
sA = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
|
219 |
+
|
220 |
+
sB = cuda.shared.array(shape=(TPB, TPB), dtype=float32)
|
221 |
+
|
222 |
+
x, y = cuda.grid(2)
|
223 |
+
|
224 |
+
tx = cuda.threadIdx.x
|
225 |
+
|
226 |
+
ty = cuda.threadIdx.y
|
227 |
+
|
228 |
+
bpg = cuda.gridDim.x # blocks per grid
|
229 |
+
|
230 |
+
# Each thread computes one element in the result matrix.
|
231 |
+
|
232 |
+
# The dot product is chunked into dot products of TPB-long vectors.
|
233 |
+
|
234 |
+
tmp = float32(0.)
|
235 |
+
|
236 |
+
for i in range(bpg):
|
237 |
+
|
238 |
+
# Preload data into shared memory
|
239 |
+
|
240 |
+
sA[ty, tx] = 0
|
241 |
+
|
242 |
+
sB[ty, tx] = 0
|
243 |
+
|
244 |
+
if y < A.shape[0] and (tx + i * TPB) < A.shape[1]:
|
245 |
+
sA[ty, tx] = A[y, tx + i * TPB]
|
246 |
+
|
247 |
+
if x < B.shape[1] and (ty + i * TPB) < B.shape[0]:
|
248 |
+
sB[ty, tx] = B[ty + i * TPB, x]
|
249 |
+
|
250 |
+
# Wait until all threads finish preloading
|
251 |
+
|
252 |
+
cuda.syncthreads()
|
253 |
+
|
254 |
+
# Computes partial product on the shared memory
|
255 |
+
|
256 |
+
for j in range(TPB):
|
257 |
+
d = abs(sA[ty, j] - sB[j, tx])
|
258 |
+
tmp += d * d
|
259 |
+
# Wait until all threads finish computing
|
260 |
+
|
261 |
+
cuda.syncthreads()
|
262 |
+
|
263 |
+
if y < C.shape[0] and x < C.shape[1]:
|
264 |
+
C[y, x] = tmp ** (1 / 2)
|
265 |
+
|
266 |
+
|
267 |
+
def euclidean_matrix_to_matrix_distance_gpu_fast(u, m):
|
268 |
+
u_dim0 = u.shape[0]
|
269 |
+
m_dim1 = m.shape[1]
|
270 |
+
|
271 |
+
# vec_dim = u.shape[1]
|
272 |
+
# assert vec_dim == m.shape[1]
|
273 |
+
out = np.zeros((u_dim0, m_dim1), dtype=np.float32)
|
274 |
+
|
275 |
+
threadsperblock = (16, 16)
|
276 |
+
grid_y_max = max(u.shape[0], m.shape[0])
|
277 |
+
grid_x_max = max(u.shape[1], m.shape[1])
|
278 |
+
blockspergrid_x = math.ceil(grid_x_max / threadsperblock[0])
|
279 |
+
blockspergrid_y = math.ceil(grid_y_max / threadsperblock[1])
|
280 |
+
|
281 |
+
blockspergrid = (blockspergrid_x, blockspergrid_y)
|
282 |
+
|
283 |
+
u_d = cuda.to_device(u)
|
284 |
+
m_d = cuda.to_device(m)
|
285 |
+
out_d = cuda.to_device(out)
|
286 |
+
|
287 |
+
gpu_kernel_euclidean_matrix_to_matrix_distance_fast[blockspergrid, threadsperblock](u_d, m_d, out_d)
|
288 |
+
out = out_d.copy_to_host()
|
289 |
+
return out
|
290 |
+
|
291 |
+
|
292 |
+
@jit(cache=True, nopython=True, parallel=True, fastmath=True, boundscheck=False, nogil=True)
|
293 |
+
def euclidean_matrix_to_matrix_distance(a, b):
|
294 |
+
"""
|
295 |
+
:purpose:
|
296 |
+
Computes the distance between the rows of two matrices using any given metric
|
297 |
+
|
298 |
+
:params:
|
299 |
+
a, b : input matrices either of shape (m, n) and (k, n)
|
300 |
+
the matrices must share a common dimension at index 1
|
301 |
+
metric : the function used to calculate the distance
|
302 |
+
metric_name : str of the function name. this is only used for
|
303 |
+
the if statement because cosine similarity has its
|
304 |
+
own function
|
305 |
+
|
306 |
+
:returns:
|
307 |
+
distance matrix : np.array, an (m, k) array of the distance
|
308 |
+
between the rows of a and b
|
309 |
+
|
310 |
+
:example:
|
311 |
+
>>> from fastdist import fastdist
|
312 |
+
>>> import numpy as np
|
313 |
+
>>> a = np.random.RandomState(seed=0).rand(10, 50)
|
314 |
+
>>> b = np.random.RandomState(seed=0).rand(100, 50)
|
315 |
+
>>> fastdist.matrix_to_matrix_distance(a, b, fastdist.cosine, "cosine")
|
316 |
+
(returns an array of shape (10, 100))
|
317 |
+
|
318 |
+
:note:
|
319 |
+
the cosine similarity uses its own function, cosine_matrix_to_matrix.
|
320 |
+
this is because normalizing the rows and then taking the dot product
|
321 |
+
of the two matrices heavily optimizes the computation. the other similarity
|
322 |
+
metrics do not have such an optimization, so we loop through them
|
323 |
+
"""
|
324 |
+
n, m = a.shape[0], b.shape[0]
|
325 |
+
out = np.zeros((n, m), dtype=np.float32)
|
326 |
+
for i in prange(n):
|
327 |
+
for j in range(m):
|
328 |
+
dist = 0
|
329 |
+
for l in range(len(a[i])):
|
330 |
+
dist += abs(a[i][l] - b[j][l]) ** 2
|
331 |
+
out[i][j] = dist ** (1 / 2)
|
332 |
+
return out
|
requirements.txt
CHANGED
@@ -1,3 +1,7 @@
|
|
1 |
torch
|
2 |
Pillow
|
3 |
torchvision
|
|
|
|
|
|
|
|
|
|
1 |
torch
|
2 |
Pillow
|
3 |
torchvision
|
4 |
+
numpy
|
5 |
+
opencv-python
|
6 |
+
umap-learn
|
7 |
+
numba
|
sampling_util.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
|
3 |
+
from fastdist2 import euclidean_vector_to_matrix_distance
|
4 |
+
|
5 |
+
|
6 |
+
def furthest_neighbours(x, downsampled_size, seed):
|
7 |
+
x = x.astype(np.float32)
|
8 |
+
np.random.seed(seed)
|
9 |
+
length = x.shape[0]
|
10 |
+
img_vecs_dims = x.shape[-1]
|
11 |
+
|
12 |
+
rv_indices = [np.random.randint(low=0, high=downsampled_size - 1, size=1)[0]]
|
13 |
+
selected_points = np.zeros((downsampled_size, img_vecs_dims), np.float32)
|
14 |
+
selected_points[0, :] = x[rv_indices[0], :]
|
15 |
+
|
16 |
+
distance_for_selected_min = np.ones((length,)) * 1e15
|
17 |
+
|
18 |
+
inactive_points = np.zeros(length, dtype=bool)
|
19 |
+
inactive_points[rv_indices[0]] = True
|
20 |
+
|
21 |
+
for i in (range(downsampled_size - 1)):
|
22 |
+
distance_for_selected = euclidean_vector_to_matrix_distance(selected_points[i, :], x)
|
23 |
+
distance_for_selected_min = np.minimum(distance_for_selected_min, distance_for_selected)
|
24 |
+
furthest_point_idx = np.argmax(np.ma.array(distance_for_selected_min, mask=inactive_points))
|
25 |
+
|
26 |
+
rv_indices.append(furthest_point_idx)
|
27 |
+
selected_points[i + 1, :] = x[furthest_point_idx, :]
|
28 |
+
inactive_points[furthest_point_idx] = True
|
29 |
+
|
30 |
+
return rv_indices, selected_points
|
video_reader.py
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import cv2
|
2 |
+
|
3 |
+
|
4 |
+
def video_reader(file_path, targetFPS=9, targetWidth=None, to_rgb=True, prompt=None):
|
5 |
+
cap = cv2.VideoCapture(file_path)
|
6 |
+
sourceFPS = int(cap.get(cv2.CAP_PROP_FPS))
|
7 |
+
|
8 |
+
if sourceFPS < targetFPS:
|
9 |
+
raise ValueError("sourceFPS < targetFPS: {} < {}".format(sourceFPS, targetFPS))
|
10 |
+
|
11 |
+
fpsDiv = 3 # sourceFPS // targetFPS
|
12 |
+
print("sourceFPS: {}, targetFPS: {}, fpsDiv: {}".format(sourceFPS, targetFPS, fpsDiv))
|
13 |
+
|
14 |
+
frameCount = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
|
15 |
+
frameWidth = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
|
16 |
+
frameHeight = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
17 |
+
print("frameCount: {}, frameWidth: {}, fpsDiv: {}".format(frameCount, frameWidth, frameHeight))
|
18 |
+
|
19 |
+
if targetWidth:
|
20 |
+
targetHeight = int(targetWidth * frameHeight / frameWidth)
|
21 |
+
|
22 |
+
fc = 0
|
23 |
+
ret = True
|
24 |
+
|
25 |
+
while fc < frameCount and ret:
|
26 |
+
ret, img = cap.read()
|
27 |
+
if fc % fpsDiv == 0:
|
28 |
+
if targetWidth:
|
29 |
+
img = cv2.resize(img, (targetWidth, targetHeight), interpolation=cv2.INTER_AREA)
|
30 |
+
if to_rgb:
|
31 |
+
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
|
32 |
+
yield img
|
33 |
+
fc += 1
|
34 |
+
|
35 |
+
cap.release()
|