-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmetrics.py
311 lines (239 loc) · 9.33 KB
/
metrics.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
import cv2
import sys
import lpips
import numpy as np
from matplotlib import pyplot as plt
import torch
import os
from skimage.metrics import structural_similarity as ssim
class LMDMeter:
def __init__(self, backend='dlib', region='mouth'):
self.backend = backend
self.region = region # mouth or face
if self.backend == 'dlib':
import dlib
# load checkpoint manually
self.predictor_path = './shape_predictor_68_face_landmarks.dat'
if not os.path.exists(self.predictor_path):
raise FileNotFoundError('Please download dlib checkpoint from http://dlib.net/files/shape_predictor_68_face_landmarks.dat.bz2')
self.detector = dlib.get_frontal_face_detector()
self.predictor = dlib.shape_predictor(self.predictor_path)
else:
import face_alignment
try:
self.predictor = face_alignment.FaceAlignment(face_alignment.LandmarksType._2D, flip_input=False)
except:
self.predictor = face_alignment.FaceAlignment(face_alignment.LandmarksType.TWO_D, flip_input=False)
self.V = 0
self.N = 0
def get_landmarks(self, img):
if self.backend == 'dlib':
dets = self.detector(img, 1)
for det in dets:
shape = self.predictor(img, det)
# ref: https://github.com/PyImageSearch/imutils/blob/c12f15391fcc945d0d644b85194b8c044a392e0a/imutils/face_utils/helpers.py
lms = np.zeros((68, 2), dtype=np.int32)
for i in range(0, 68):
lms[i, 0] = shape.part(i).x
lms[i, 1] = shape.part(i).y
break
else:
lms = self.predictor.get_landmarks(img)[-1]
# self.vis_landmarks(img, lms)
lms = lms.astype(np.float32)
return lms
def vis_landmarks(self, img, lms):
plt.imshow(img)
plt.plot(lms[48:68, 0], lms[48:68, 1], marker='o', markersize=1, linestyle='-', lw=2)
plt.show()
def clear(self):
self.V = 0
self.N = 0
def prepare_inputs(self, *inputs):
outputs = []
for i, inp in enumerate(inputs):
inp = inp.detach().cpu().numpy()
inp = (inp * 255).astype(np.uint8)
outputs.append(inp)
return outputs
def update(self, preds, truths):
# assert B == 1
preds, truths = self.prepare_inputs(preds[0], truths[0]) # [H, W, 3] numpy array
try:
# get lms
lms_pred = self.get_landmarks(preds)
lms_truth = self.get_landmarks(truths)
if self.region == 'mouth':
lms_pred = lms_pred[48:68]
lms_truth = lms_truth[48:68]
# avarage
lms_pred = lms_pred - lms_pred.mean(0)
lms_truth = lms_truth - lms_truth.mean(0)
# distance
dist = np.sqrt(((lms_pred - lms_truth) ** 2).sum(1)).mean(0)
self.V += dist
self.N += 1
except Exception as e:
print(e)
def measure(self):
return self.V / self.N
def write(self, writer, global_step, prefix=""):
writer.add_scalar(os.path.join(prefix, f"LMD ({self.backend})"), self.measure(), global_step)
def report(self):
return f'LMD ({self.backend}) = {self.measure():.6f}'
class PSNRMeter:
def __init__(self):
self.V = 0
self.N = 0
def clear(self):
self.V = 0
self.N = 0
def prepare_inputs(self, *inputs):
outputs = []
for i, inp in enumerate(inputs):
if torch.is_tensor(inp):
inp = inp.detach().cpu().numpy()
outputs.append(inp)
return outputs
def update(self, preds, truths):
preds, truths = self.prepare_inputs(preds, truths) # [B, N, 3] or [B, H, W, 3], range in [0, 1]
# simplified since max_pixel_value is 1 here.
psnr = -10 * np.log10(np.mean((preds - truths) ** 2))
self.V += psnr
self.N += 1
def measure(self):
return self.V / self.N
def write(self, writer, global_step, prefix=""):
writer.add_scalar(os.path.join(prefix, "PSNR"), self.measure(), global_step)
def report(self):
return f'PSNR = {self.measure():.6f}'
class LPIPSMeter:
def __init__(self, net='alex', device=None):
self.V = 0
self.N = 0
self.net = net
self.device = device if device is not None else torch.device('cuda' if torch.cuda.is_available() else 'cpu')
self.fn = lpips.LPIPS(net=net).eval().to(self.device)
def clear(self):
self.V = 0
self.N = 0
def prepare_inputs(self, *inputs):
outputs = []
for i, inp in enumerate(inputs):
inp = inp.permute(0, 3, 1, 2).contiguous() # [B, 3, H, W]
inp = inp.to(self.device)
outputs.append(inp)
return outputs
def update(self, preds, truths):
preds, truths = self.prepare_inputs(preds, truths) # [B, H, W, 3] --> [B, 3, H, W], range in [0, 1]
v = self.fn(truths, preds, normalize=True).item() # normalize=True: [0, 1] to [-1, 1]
self.V += v
self.N += 1
def measure(self):
return self.V / self.N
def write(self, writer, global_step, prefix=""):
writer.add_scalar(os.path.join(prefix, f"LPIPS ({self.net})"), self.measure(), global_step)
def report(self):
return f'LPIPS ({self.net}) = {self.measure():.6f}'
class SSIMMeter:
def __init__(self):
self.V = 0
self.N = 0
def clear(self):
self.V = 0
self.N = 0
def prepare_inputs(self, *inputs):
outputs = []
for i, inp in enumerate(inputs):
if torch.is_tensor(inp):
inp = inp.detach().cpu().numpy()
inp = (inp * 255).astype(np.uint8)
outputs.append(inp)
return outputs
def update(self, preds, truths):
preds, truths = self.prepare_inputs(preds[0], truths[0]) # [H, W, 3] numpy array
ssim_value = ssim(preds, truths, channel_axis = 2)
self.V += ssim_value
self.N += 1
def measure(self):
return self.V / self.N
def write(self, writer, global_step, prefix=""):
writer.add_scalar(os.path.join(prefix, "SSIM"), self.measure(), global_step)
def report(self):
return f'SSIM = {self.measure():.6f}'
class FIDMeter:
def __init__(self):
self.preds = []
self.truths = []
def clear(self):
self.preds = []
self.truths = []
def prepare_inputs(self, *inputs):
outputs = []
for inp in inputs:
inp = inp.permute(0, 3, 1, 2).contiguous() # [B, H, W, 3] --> [B, 3, H, W]
inp = inp.detach().cpu().numpy()
outputs.append(inp)
return outputs
def update(self, preds, truths):
preds, truths = self.prepare_inputs(preds, truths)
self.preds.append(preds[0])
self.truths.append(truths[0])
def measure(self):
import tempfile
from pytorch_fid import fid_score
with tempfile.TemporaryDirectory() as tmpdirname:
pred_path = os.path.join(tmpdirname, 'preds')
truth_path = os.path.join(tmpdirname, 'truths')
os.makedirs(pred_path)
os.makedirs(truth_path)
for i, (pred, truth) in enumerate(zip(self.preds, self.truths)):
pred_image_path = os.path.join(pred_path, f'{i}.png')
truth_image_path = os.path.join(truth_path, f'{i}.png')
cv2.imwrite(pred_image_path, pred.transpose(1, 2, 0)[:, :, ::-1])
cv2.imwrite(truth_image_path, truth.transpose(1, 2, 0)[:, :, ::-1])
fid_value = fid_score.calculate_fid_given_paths([pred_path, truth_path], batch_size=1, device='cuda', dims=2048)
return fid_value
def write(self, writer, global_step, prefix=""):
writer.add_scalar(os.path.join(prefix, "FID"), self.measure(), global_step)
def report(self):
return f'FID = {self.measure():.6f}'
lmd_meter = LMDMeter(backend='fan')
psnr_meter = PSNRMeter()
lpips_meter = LPIPSMeter()
ssim_meter = SSIMMeter()
lmd_meter.clear()
psnr_meter.clear()
lpips_meter.clear()
ssim_meter.clear()
vid_path_1 = sys.argv[1]
vid_path_2 = sys.argv[2]
capture_1 = cv2.VideoCapture(vid_path_1)
capture_2 = cv2.VideoCapture(vid_path_2)
# Get total frame count
total_frames_1 = int(capture_1.get(cv2.CAP_PROP_FRAME_COUNT))
total_frames_2 = int(capture_2.get(cv2.CAP_PROP_FRAME_COUNT))
print(f"Total Frames in Video 1: {total_frames_1}")
print(f"Total Frames in Video 2: {total_frames_2}")
counter = 0
while True:
ret_1, frame_1 = capture_1.read()
ret_2, frame_2 = capture_2.read()
if not ret_1 * ret_2:
break
# plt.imshow(frame_1[:, :, ::-1])
# plt.show()
inp_1 = torch.FloatTensor(frame_1[..., ::-1] / 255.0)[None, ...].cuda()
inp_2 = torch.FloatTensor(frame_2[..., ::-1] / 255.0)[None, ...].cuda()
lmd_meter.update(inp_1, inp_2)
psnr_meter.update(inp_1, inp_2)
lpips_meter.update(inp_1, inp_2)
ssim_meter.update(inp_1, inp_2)
counter += 1
if counter % 100 == 0:
print(counter)
print("Frame Count", counter)
print(lmd_meter.report())
print(psnr_meter.report())
print(lpips_meter.report())
print(ssim_meter.report())