sanket kheni
commited on
Commit
·
7576d48
1
Parent(s):
6b2b650
- app.py +7 -8
- retina_model/anchor.py +296 -0
- retina_model/models.py +301 -0
- retina_model/ops.py +27 -0
app.py
CHANGED
@@ -1,17 +1,16 @@
|
|
|
|
|
|
1 |
import os
|
2 |
|
3 |
-
import
|
4 |
-
import
|
|
|
5 |
import numpy as np
|
6 |
-
|
7 |
from scipy.ndimage import gaussian_filter
|
8 |
-
from tensorflow.keras.models import load_model
|
9 |
-
from tensorflow_addons.layers import InstanceNormalization
|
10 |
|
11 |
-
from
|
12 |
from options.swap_options import SwapOptions
|
13 |
-
from utils.utils import (estimate_norm, get_lm, inverse_estimate_norm,
|
14 |
-
norm_crop, transform_landmark_points)
|
15 |
|
16 |
# .
|
17 |
# token = os.environ['model_fetch']
|
|
|
1 |
+
import gradio
|
2 |
+
from huggingface_hub import Repository
|
3 |
import os
|
4 |
|
5 |
+
from utils.utils import norm_crop, estimate_norm, inverse_estimate_norm, transform_landmark_points, get_lm
|
6 |
+
from networks.layers import AdaIN, AdaptiveAttention
|
7 |
+
from tensorflow_addons.layers import InstanceNormalization
|
8 |
import numpy as np
|
9 |
+
import cv2
|
10 |
from scipy.ndimage import gaussian_filter
|
|
|
|
|
11 |
|
12 |
+
from tensorflow.keras.models import load_model
|
13 |
from options.swap_options import SwapOptions
|
|
|
|
|
14 |
|
15 |
# .
|
16 |
# token = os.environ['model_fetch']
|
retina_model/anchor.py
ADDED
@@ -0,0 +1,296 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Anchor utils modified from https://github.com/biubug6/Pytorch_Retinaface"""
|
2 |
+
import math
|
3 |
+
import tensorflow as tf
|
4 |
+
import numpy as np
|
5 |
+
from itertools import product as product
|
6 |
+
|
7 |
+
|
8 |
+
###############################################################################
|
9 |
+
# Tensorflow / Numpy Priors #
|
10 |
+
###############################################################################
|
11 |
+
def prior_box(image_sizes, min_sizes, steps, clip=False):
|
12 |
+
"""prior box"""
|
13 |
+
feature_maps = [
|
14 |
+
[math.ceil(image_sizes[0] / step), math.ceil(image_sizes[1] / step)]
|
15 |
+
for step in steps]
|
16 |
+
|
17 |
+
anchors = []
|
18 |
+
for k, f in enumerate(feature_maps):
|
19 |
+
for i, j in product(range(f[0]), range(f[1])):
|
20 |
+
for min_size in min_sizes[k]:
|
21 |
+
s_kx = min_size / image_sizes[1]
|
22 |
+
s_ky = min_size / image_sizes[0]
|
23 |
+
cx = (j + 0.5) * steps[k] / image_sizes[1]
|
24 |
+
cy = (i + 0.5) * steps[k] / image_sizes[0]
|
25 |
+
anchors += [cx, cy, s_kx, s_ky]
|
26 |
+
|
27 |
+
output = np.asarray(anchors).reshape([-1, 4])
|
28 |
+
|
29 |
+
if clip:
|
30 |
+
output = np.clip(output, 0, 1)
|
31 |
+
|
32 |
+
return output
|
33 |
+
|
34 |
+
|
35 |
+
def prior_box_tf(image_sizes, min_sizes, steps, clip=False):
|
36 |
+
"""prior box"""
|
37 |
+
image_sizes = tf.cast(tf.convert_to_tensor(image_sizes), tf.float32)
|
38 |
+
feature_maps = tf.math.ceil(
|
39 |
+
tf.reshape(image_sizes, [1, 2]) /
|
40 |
+
tf.reshape(tf.cast(steps, tf.float32), [-1, 1]))
|
41 |
+
|
42 |
+
anchors = []
|
43 |
+
for k in range(len(min_sizes)):
|
44 |
+
grid_x, grid_y = _meshgrid_tf(tf.range(feature_maps[k][1]),
|
45 |
+
tf.range(feature_maps[k][0]))
|
46 |
+
cx = (grid_x + 0.5) * steps[k] / image_sizes[1]
|
47 |
+
cy = (grid_y + 0.5) * steps[k] / image_sizes[0]
|
48 |
+
cxcy = tf.stack([cx, cy], axis=-1)
|
49 |
+
cxcy = tf.reshape(cxcy, [-1, 2])
|
50 |
+
cxcy = tf.repeat(cxcy, repeats=tf.shape(min_sizes[k])[0], axis=0)
|
51 |
+
|
52 |
+
sx = min_sizes[k] / image_sizes[1]
|
53 |
+
sy = min_sizes[k] / image_sizes[0]
|
54 |
+
sxsy = tf.stack([sx, sy], 1)
|
55 |
+
sxsy = tf.repeat(sxsy[tf.newaxis],
|
56 |
+
repeats=tf.shape(grid_x)[0] * tf.shape(grid_x)[1],
|
57 |
+
axis=0)
|
58 |
+
sxsy = tf.reshape(sxsy, [-1, 2])
|
59 |
+
|
60 |
+
anchors.append(tf.concat([cxcy, sxsy], 1))
|
61 |
+
|
62 |
+
output = tf.concat(anchors, axis=0)
|
63 |
+
|
64 |
+
if clip:
|
65 |
+
output = tf.clip_by_value(output, 0, 1)
|
66 |
+
|
67 |
+
return output
|
68 |
+
|
69 |
+
|
70 |
+
def _meshgrid_tf(x, y):
|
71 |
+
""" workaround solution of the tf.meshgrid() issue:
|
72 |
+
https://github.com/tensorflow/tensorflow/issues/34470"""
|
73 |
+
grid_shape = [tf.shape(y)[0], tf.shape(x)[0]]
|
74 |
+
grid_x = tf.broadcast_to(tf.reshape(x, [1, -1]), grid_shape)
|
75 |
+
grid_y = tf.broadcast_to(tf.reshape(y, [-1, 1]), grid_shape)
|
76 |
+
return grid_x, grid_y
|
77 |
+
|
78 |
+
|
79 |
+
###############################################################################
|
80 |
+
# Tensorflow Encoding #
|
81 |
+
###############################################################################
|
82 |
+
def encode_tf(labels, priors, match_thresh, ignore_thresh,
|
83 |
+
variances=[0.1, 0.2]):
|
84 |
+
"""tensorflow encoding"""
|
85 |
+
assert ignore_thresh <= match_thresh
|
86 |
+
priors = tf.cast(priors, tf.float32)
|
87 |
+
bbox = labels[:, :4]
|
88 |
+
landm = labels[:, 4:-1]
|
89 |
+
landm_valid = labels[:, -1] # 1: with landm, 0: w/o landm.
|
90 |
+
|
91 |
+
# jaccard index
|
92 |
+
overlaps = _jaccard(bbox, _point_form(priors))
|
93 |
+
|
94 |
+
# (Bipartite Matching)
|
95 |
+
# [num_objects] best prior for each ground truth
|
96 |
+
best_prior_overlap, best_prior_idx = tf.math.top_k(overlaps, k=1)
|
97 |
+
best_prior_overlap = best_prior_overlap[:, 0]
|
98 |
+
best_prior_idx = best_prior_idx[:, 0]
|
99 |
+
|
100 |
+
# [num_priors] best ground truth for each prior
|
101 |
+
overlaps_t = tf.transpose(overlaps)
|
102 |
+
best_truth_overlap, best_truth_idx = tf.math.top_k(overlaps_t, k=1)
|
103 |
+
best_truth_overlap = best_truth_overlap[:, 0]
|
104 |
+
best_truth_idx = best_truth_idx[:, 0]
|
105 |
+
|
106 |
+
# ensure best prior
|
107 |
+
def _loop_body(i, bt_idx, bt_overlap):
|
108 |
+
bp_mask = tf.one_hot(best_prior_idx[i], tf.shape(bt_idx)[0])
|
109 |
+
bp_mask_int = tf.cast(bp_mask, tf.int32)
|
110 |
+
new_bt_idx = bt_idx * (1 - bp_mask_int) + bp_mask_int * i
|
111 |
+
bp_mask_float = tf.cast(bp_mask, tf.float32)
|
112 |
+
new_bt_overlap = bt_overlap * (1 - bp_mask_float) + bp_mask_float * 2
|
113 |
+
return tf.cond(best_prior_overlap[i] > match_thresh,
|
114 |
+
lambda: (i + 1, new_bt_idx, new_bt_overlap),
|
115 |
+
lambda: (i + 1, bt_idx, bt_overlap))
|
116 |
+
_, best_truth_idx, best_truth_overlap = tf.while_loop(
|
117 |
+
lambda i, bt_idx, bt_overlap: tf.less(i, tf.shape(best_prior_idx)[0]),
|
118 |
+
_loop_body, [tf.constant(0), best_truth_idx, best_truth_overlap])
|
119 |
+
|
120 |
+
matches_bbox = tf.gather(bbox, best_truth_idx) # [num_priors, 4]
|
121 |
+
matches_landm = tf.gather(landm, best_truth_idx) # [num_priors, 10]
|
122 |
+
matches_landm_v = tf.gather(landm_valid, best_truth_idx) # [num_priors]
|
123 |
+
|
124 |
+
loc_t = _encode_bbox(matches_bbox, priors, variances)
|
125 |
+
landm_t = _encode_landm(matches_landm, priors, variances)
|
126 |
+
landm_valid_t = tf.cast(matches_landm_v > 0, tf.float32)
|
127 |
+
conf_t = tf.cast(best_truth_overlap > match_thresh, tf.float32)
|
128 |
+
conf_t = tf.where(
|
129 |
+
tf.logical_and(best_truth_overlap < match_thresh,
|
130 |
+
best_truth_overlap > ignore_thresh),
|
131 |
+
tf.ones_like(conf_t) * -1, conf_t) # 1: pos, 0: neg, -1: ignore
|
132 |
+
|
133 |
+
return tf.concat([loc_t, landm_t, landm_valid_t[..., tf.newaxis],
|
134 |
+
conf_t[..., tf.newaxis]], axis=1)
|
135 |
+
|
136 |
+
|
137 |
+
def _encode_bbox(matched, priors, variances):
|
138 |
+
"""Encode the variances from the priorbox layers into the ground truth
|
139 |
+
boxes we have matched (based on jaccard overlap) with the prior boxes.
|
140 |
+
Args:
|
141 |
+
matched: (tensor) Coords of ground truth for each prior in point-form
|
142 |
+
Shape: [num_priors, 4].
|
143 |
+
priors: (tensor) Prior boxes in center-offset form
|
144 |
+
Shape: [num_priors,4].
|
145 |
+
variances: (list[float]) Variances of priorboxes
|
146 |
+
Return:
|
147 |
+
encoded boxes (tensor), Shape: [num_priors, 4]
|
148 |
+
"""
|
149 |
+
|
150 |
+
# dist b/t match center and prior's center
|
151 |
+
g_cxcy = (matched[:, :2] + matched[:, 2:]) / 2 - priors[:, :2]
|
152 |
+
# encode variance
|
153 |
+
g_cxcy /= (variances[0] * priors[:, 2:])
|
154 |
+
# match wh / prior wh
|
155 |
+
g_wh = (matched[:, 2:] - matched[:, :2]) / priors[:, 2:]
|
156 |
+
g_wh = tf.math.log(g_wh) / variances[1]
|
157 |
+
# return target for smooth_l1_loss
|
158 |
+
return tf.concat([g_cxcy, g_wh], 1) # [num_priors,4]
|
159 |
+
|
160 |
+
|
161 |
+
def _encode_landm(matched, priors, variances):
|
162 |
+
"""Encode the variances from the priorbox layers into the ground truth
|
163 |
+
boxes we have matched (based on jaccard overlap) with the prior boxes.
|
164 |
+
Args:
|
165 |
+
matched: (tensor) Coords of ground truth for each prior in point-form
|
166 |
+
Shape: [num_priors, 10].
|
167 |
+
priors: (tensor) Prior boxes in center-offset form
|
168 |
+
Shape: [num_priors,4].
|
169 |
+
variances: (list[float]) Variances of priorboxes
|
170 |
+
Return:
|
171 |
+
encoded landm (tensor), Shape: [num_priors, 10]
|
172 |
+
"""
|
173 |
+
|
174 |
+
# dist b/t match center and prior's center
|
175 |
+
matched = tf.reshape(matched, [tf.shape(matched)[0], 5, 2])
|
176 |
+
priors = tf.broadcast_to(
|
177 |
+
tf.expand_dims(priors, 1), [tf.shape(matched)[0], 5, 4])
|
178 |
+
g_cxcy = matched[:, :, :2] - priors[:, :, :2]
|
179 |
+
# encode variance
|
180 |
+
g_cxcy /= (variances[0] * priors[:, :, 2:])
|
181 |
+
# g_cxcy /= priors[:, :, 2:]
|
182 |
+
g_cxcy = tf.reshape(g_cxcy, [tf.shape(g_cxcy)[0], -1])
|
183 |
+
# return target for smooth_l1_loss
|
184 |
+
return g_cxcy
|
185 |
+
|
186 |
+
|
187 |
+
def _point_form(boxes):
|
188 |
+
""" Convert prior_boxes to (xmin, ymin, xmax, ymax)
|
189 |
+
representation for comparison to point form ground truth data.
|
190 |
+
Args:
|
191 |
+
boxes: (tensor) center-size default boxes from priorbox layers.
|
192 |
+
Return:
|
193 |
+
boxes: (tensor) Converted xmin, ymin, xmax, ymax form of boxes.
|
194 |
+
"""
|
195 |
+
return tf.concat((boxes[:, :2] - boxes[:, 2:] / 2,
|
196 |
+
boxes[:, :2] + boxes[:, 2:] / 2), axis=1)
|
197 |
+
|
198 |
+
|
199 |
+
def _intersect(box_a, box_b):
|
200 |
+
""" We resize both tensors to [A,B,2]:
|
201 |
+
[A,2] -> [A,1,2] -> [A,B,2]
|
202 |
+
[B,2] -> [1,B,2] -> [A,B,2]
|
203 |
+
Then we compute the area of intersect between box_a and box_b.
|
204 |
+
Args:
|
205 |
+
box_a: (tensor) bounding boxes, Shape: [A,4].
|
206 |
+
box_b: (tensor) bounding boxes, Shape: [B,4].
|
207 |
+
Return:
|
208 |
+
(tensor) intersection area, Shape: [A,B].
|
209 |
+
"""
|
210 |
+
A = tf.shape(box_a)[0]
|
211 |
+
B = tf.shape(box_b)[0]
|
212 |
+
max_xy = tf.minimum(
|
213 |
+
tf.broadcast_to(tf.expand_dims(box_a[:, 2:], 1), [A, B, 2]),
|
214 |
+
tf.broadcast_to(tf.expand_dims(box_b[:, 2:], 0), [A, B, 2]))
|
215 |
+
min_xy = tf.maximum(
|
216 |
+
tf.broadcast_to(tf.expand_dims(box_a[:, :2], 1), [A, B, 2]),
|
217 |
+
tf.broadcast_to(tf.expand_dims(box_b[:, :2], 0), [A, B, 2]))
|
218 |
+
inter = tf.maximum((max_xy - min_xy), tf.zeros_like(max_xy - min_xy))
|
219 |
+
return inter[:, :, 0] * inter[:, :, 1]
|
220 |
+
|
221 |
+
|
222 |
+
def _jaccard(box_a, box_b):
|
223 |
+
"""Compute the jaccard overlap of two sets of boxes. The jaccard overlap
|
224 |
+
is simply the intersection over union of two boxes. Here we operate on
|
225 |
+
ground truth boxes and default boxes.
|
226 |
+
E.g.:
|
227 |
+
A ∩ B / A ∪ B = A ∩ B / (area(A) + area(B) - A ∩ B)
|
228 |
+
Args:
|
229 |
+
box_a: (tensor) Ground truth bounding boxes, Shape: [num_objects,4]
|
230 |
+
box_b: (tensor) Prior boxes from priorbox layers, Shape: [num_priors,4]
|
231 |
+
Return:
|
232 |
+
jaccard overlap: (tensor) Shape: [box_a.size(0), box_b.size(0)]
|
233 |
+
"""
|
234 |
+
inter = _intersect(box_a, box_b)
|
235 |
+
area_a = tf.broadcast_to(
|
236 |
+
tf.expand_dims(
|
237 |
+
(box_a[:, 2] - box_a[:, 0]) * (box_a[:, 3] - box_a[:, 1]), 1),
|
238 |
+
tf.shape(inter)) # [A,B]
|
239 |
+
area_b = tf.broadcast_to(
|
240 |
+
tf.expand_dims(
|
241 |
+
(box_b[:, 2] - box_b[:, 0]) * (box_b[:, 3] - box_b[:, 1]), 0),
|
242 |
+
tf.shape(inter)) # [A,B]
|
243 |
+
union = area_a + area_b - inter
|
244 |
+
return inter / union # [A,B]
|
245 |
+
|
246 |
+
|
247 |
+
###############################################################################
|
248 |
+
# Tensorflow Decoding #
|
249 |
+
###############################################################################
|
250 |
+
def decode_tf(labels, priors, variances=[0.1, 0.2]):
|
251 |
+
"""tensorflow decoding"""
|
252 |
+
bbox = _decode_bbox(labels[:, :4], priors, variances)
|
253 |
+
landm = _decode_landm(labels[:, 4:14], priors, variances)
|
254 |
+
landm_valid = labels[:, 14][:, tf.newaxis]
|
255 |
+
conf = labels[:, 15][:, tf.newaxis]
|
256 |
+
|
257 |
+
return tf.concat([bbox, landm, landm_valid, conf], axis=1)
|
258 |
+
|
259 |
+
|
260 |
+
def _decode_bbox(pre, priors, variances=[0.1, 0.2]):
|
261 |
+
"""Decode locations from predictions using priors to undo
|
262 |
+
the encoding we did for offset regression at train time.
|
263 |
+
Args:
|
264 |
+
pre (tensor): location predictions for loc layers,
|
265 |
+
Shape: [num_priors,4]
|
266 |
+
priors (tensor): Prior boxes in center-offset form.
|
267 |
+
Shape: [num_priors,4].
|
268 |
+
variances: (list[float]) Variances of priorboxes
|
269 |
+
Return:
|
270 |
+
decoded bounding box predictions
|
271 |
+
"""
|
272 |
+
centers = priors[:, :2] + pre[:, :2] * variances[0] * priors[:, 2:]
|
273 |
+
sides = priors[:, 2:] * tf.math.exp(pre[:, 2:] * variances[1])
|
274 |
+
|
275 |
+
return tf.concat([centers - sides / 2, centers + sides / 2], axis=1)
|
276 |
+
|
277 |
+
|
278 |
+
def _decode_landm(pre, priors, variances=[0.1, 0.2]):
|
279 |
+
"""Decode landm from predictions using priors to undo
|
280 |
+
the encoding we did for offset regression at train time.
|
281 |
+
Args:
|
282 |
+
pre (tensor): landm predictions for loc layers,
|
283 |
+
Shape: [num_priors,10]
|
284 |
+
priors (tensor): Prior boxes in center-offset form.
|
285 |
+
Shape: [num_priors,4].
|
286 |
+
variances: (list[float]) Variances of priorboxes
|
287 |
+
Return:
|
288 |
+
decoded landm predictions
|
289 |
+
"""
|
290 |
+
landms = tf.concat(
|
291 |
+
[priors[:, :2] + pre[:, :2] * variances[0] * priors[:, 2:],
|
292 |
+
priors[:, :2] + pre[:, 2:4] * variances[0] * priors[:, 2:],
|
293 |
+
priors[:, :2] + pre[:, 4:6] * variances[0] * priors[:, 2:],
|
294 |
+
priors[:, :2] + pre[:, 6:8] * variances[0] * priors[:, 2:],
|
295 |
+
priors[:, :2] + pre[:, 8:10] * variances[0] * priors[:, 2:]], axis=1)
|
296 |
+
return landms
|
retina_model/models.py
ADDED
@@ -0,0 +1,301 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import tensorflow as tf
|
2 |
+
from tensorflow.keras import Model
|
3 |
+
from tensorflow.keras.applications import MobileNetV2, ResNet50
|
4 |
+
from tensorflow.keras.layers import Input, Conv2D, ReLU, LeakyReLU
|
5 |
+
from retinaface.anchor import decode_tf, prior_box_tf
|
6 |
+
|
7 |
+
|
8 |
+
def _regularizer(weights_decay):
|
9 |
+
"""l2 regularizer"""
|
10 |
+
return tf.keras.regularizers.l2(weights_decay)
|
11 |
+
|
12 |
+
|
13 |
+
def _kernel_init(scale=1.0, seed=None):
|
14 |
+
"""He normal initializer"""
|
15 |
+
return tf.keras.initializers.he_normal()
|
16 |
+
|
17 |
+
|
18 |
+
class BatchNormalization(tf.keras.layers.BatchNormalization):
|
19 |
+
"""Make trainable=False freeze BN for real (the og version is sad).
|
20 |
+
ref: https://github.com/zzh8829/yolov3-tf2
|
21 |
+
"""
|
22 |
+
def __init__(self, axis=-1, momentum=0.9, epsilon=1e-5, center=True,
|
23 |
+
scale=True, name=None, **kwargs):
|
24 |
+
super(BatchNormalization, self).__init__(
|
25 |
+
axis=axis, momentum=momentum, epsilon=epsilon, center=center,
|
26 |
+
scale=scale, name=name, **kwargs)
|
27 |
+
|
28 |
+
def call(self, x, training=False):
|
29 |
+
if training is None:
|
30 |
+
training = tf.constant(False)
|
31 |
+
training = tf.logical_and(training, self.trainable)
|
32 |
+
|
33 |
+
return super().call(x, training)
|
34 |
+
|
35 |
+
|
36 |
+
def Backbone(backbone_type='ResNet50', use_pretrain=True):
|
37 |
+
"""Backbone Model"""
|
38 |
+
weights = None
|
39 |
+
if use_pretrain:
|
40 |
+
weights = 'imagenet'
|
41 |
+
|
42 |
+
def backbone(x):
|
43 |
+
if backbone_type == 'ResNet50':
|
44 |
+
extractor = ResNet50(
|
45 |
+
input_shape=x.shape[1:], include_top=False, weights=weights)
|
46 |
+
pick_layer1 = 80 # [80, 80, 512]
|
47 |
+
pick_layer2 = 142 # [40, 40, 1024]
|
48 |
+
pick_layer3 = 174 # [20, 20, 2048]
|
49 |
+
preprocess = tf.keras.applications.resnet.preprocess_input
|
50 |
+
elif backbone_type == 'MobileNetV2':
|
51 |
+
extractor = MobileNetV2(
|
52 |
+
input_shape=x.shape[1:], include_top=False, weights=weights)
|
53 |
+
pick_layer1 = 54 # [80, 80, 32]
|
54 |
+
pick_layer2 = 116 # [40, 40, 96]
|
55 |
+
pick_layer3 = 143 # [20, 20, 160]
|
56 |
+
preprocess = tf.keras.applications.mobilenet_v2.preprocess_input
|
57 |
+
else:
|
58 |
+
raise NotImplementedError(
|
59 |
+
'Backbone type {} is not recognized.'.format(backbone_type))
|
60 |
+
|
61 |
+
return Model(extractor.input,
|
62 |
+
(extractor.layers[pick_layer1].output,
|
63 |
+
extractor.layers[pick_layer2].output,
|
64 |
+
extractor.layers[pick_layer3].output),
|
65 |
+
name=backbone_type + '_extrator')(preprocess(x))
|
66 |
+
|
67 |
+
return backbone
|
68 |
+
|
69 |
+
|
70 |
+
class ConvUnit(tf.keras.layers.Layer):
|
71 |
+
"""Conv + BN + Act"""
|
72 |
+
def __init__(self, f, k, s, wd, act=None, **kwargs):
|
73 |
+
super(ConvUnit, self).__init__(**kwargs)
|
74 |
+
self.conv = Conv2D(filters=f, kernel_size=k, strides=s, padding='same',
|
75 |
+
kernel_initializer=_kernel_init(),
|
76 |
+
kernel_regularizer=_regularizer(wd),
|
77 |
+
use_bias=False)
|
78 |
+
self.bn = BatchNormalization()
|
79 |
+
|
80 |
+
if act is None:
|
81 |
+
self.act_fn = tf.identity
|
82 |
+
elif act == 'relu':
|
83 |
+
self.act_fn = ReLU()
|
84 |
+
elif act == 'lrelu':
|
85 |
+
self.act_fn = LeakyReLU(0.1)
|
86 |
+
else:
|
87 |
+
raise NotImplementedError(
|
88 |
+
'Activation function type {} is not recognized.'.format(act))
|
89 |
+
|
90 |
+
def call(self, x):
|
91 |
+
return self.act_fn(self.bn(self.conv(x)))
|
92 |
+
|
93 |
+
|
94 |
+
class FPN(tf.keras.layers.Layer):
|
95 |
+
"""Feature Pyramid Network"""
|
96 |
+
def __init__(self, out_ch, wd, **kwargs):
|
97 |
+
super(FPN, self).__init__(**kwargs)
|
98 |
+
act = 'relu'
|
99 |
+
self.out_ch = out_ch
|
100 |
+
self.wd = wd
|
101 |
+
if (out_ch <= 64):
|
102 |
+
act = 'lrelu'
|
103 |
+
|
104 |
+
self.output1 = ConvUnit(f=out_ch, k=1, s=1, wd=wd, act=act)
|
105 |
+
self.output2 = ConvUnit(f=out_ch, k=1, s=1, wd=wd, act=act)
|
106 |
+
self.output3 = ConvUnit(f=out_ch, k=1, s=1, wd=wd, act=act)
|
107 |
+
self.merge1 = ConvUnit(f=out_ch, k=3, s=1, wd=wd, act=act)
|
108 |
+
self.merge2 = ConvUnit(f=out_ch, k=3, s=1, wd=wd, act=act)
|
109 |
+
|
110 |
+
def call(self, x):
|
111 |
+
output1 = self.output1(x[0]) # [80, 80, out_ch]
|
112 |
+
output2 = self.output2(x[1]) # [40, 40, out_ch]
|
113 |
+
output3 = self.output3(x[2]) # [20, 20, out_ch]
|
114 |
+
|
115 |
+
up_h, up_w = tf.shape(output2)[1], tf.shape(output2)[2]
|
116 |
+
up3 = tf.image.resize(output3, [up_h, up_w], method='nearest')
|
117 |
+
output2 = output2 + up3
|
118 |
+
output2 = self.merge2(output2)
|
119 |
+
|
120 |
+
up_h, up_w = tf.shape(output1)[1], tf.shape(output1)[2]
|
121 |
+
up2 = tf.image.resize(output2, [up_h, up_w], method='nearest')
|
122 |
+
output1 = output1 + up2
|
123 |
+
output1 = self.merge1(output1)
|
124 |
+
|
125 |
+
return output1, output2, output3
|
126 |
+
|
127 |
+
def get_config(self):
|
128 |
+
config = {
|
129 |
+
'out_ch': self.out_ch,
|
130 |
+
'wd': self.wd,
|
131 |
+
}
|
132 |
+
base_config = super(FPN, self).get_config()
|
133 |
+
return dict(list(base_config.items()) + list(config.items()))
|
134 |
+
|
135 |
+
|
136 |
+
class SSH(tf.keras.layers.Layer):
|
137 |
+
"""Single Stage Headless Layer"""
|
138 |
+
def __init__(self, out_ch, wd, **kwargs):
|
139 |
+
super(SSH, self).__init__(**kwargs)
|
140 |
+
assert out_ch % 4 == 0
|
141 |
+
self.out_ch = out_ch
|
142 |
+
self.wd = wd
|
143 |
+
act = 'relu'
|
144 |
+
if (out_ch <= 64):
|
145 |
+
act = 'lrelu'
|
146 |
+
|
147 |
+
self.conv_3x3 = ConvUnit(f=out_ch // 2, k=3, s=1, wd=wd, act=None)
|
148 |
+
|
149 |
+
self.conv_5x5_1 = ConvUnit(f=out_ch // 4, k=3, s=1, wd=wd, act=act)
|
150 |
+
self.conv_5x5_2 = ConvUnit(f=out_ch // 4, k=3, s=1, wd=wd, act=None)
|
151 |
+
|
152 |
+
self.conv_7x7_2 = ConvUnit(f=out_ch // 4, k=3, s=1, wd=wd, act=act)
|
153 |
+
self.conv_7x7_3 = ConvUnit(f=out_ch // 4, k=3, s=1, wd=wd, act=None)
|
154 |
+
|
155 |
+
self.relu = ReLU()
|
156 |
+
|
157 |
+
def call(self, x):
|
158 |
+
conv_3x3 = self.conv_3x3(x)
|
159 |
+
|
160 |
+
conv_5x5_1 = self.conv_5x5_1(x)
|
161 |
+
conv_5x5 = self.conv_5x5_2(conv_5x5_1)
|
162 |
+
|
163 |
+
conv_7x7_2 = self.conv_7x7_2(conv_5x5_1)
|
164 |
+
conv_7x7 = self.conv_7x7_3(conv_7x7_2)
|
165 |
+
|
166 |
+
output = tf.concat([conv_3x3, conv_5x5, conv_7x7], axis=3)
|
167 |
+
output = self.relu(output)
|
168 |
+
|
169 |
+
return output
|
170 |
+
|
171 |
+
def get_config(self):
|
172 |
+
config = {
|
173 |
+
'out_ch': self.out_ch,
|
174 |
+
'wd': self.wd,
|
175 |
+
}
|
176 |
+
base_config = super(SSH, self).get_config()
|
177 |
+
return dict(list(base_config.items()) + list(config.items()))
|
178 |
+
|
179 |
+
|
180 |
+
class BboxHead(tf.keras.layers.Layer):
|
181 |
+
"""Bbox Head Layer"""
|
182 |
+
def __init__(self, num_anchor, wd, **kwargs):
|
183 |
+
super(BboxHead, self).__init__(**kwargs)
|
184 |
+
self.num_anchor = num_anchor
|
185 |
+
self.wd = wd
|
186 |
+
self.conv = Conv2D(filters=num_anchor * 4, kernel_size=1, strides=1)
|
187 |
+
|
188 |
+
def call(self, x):
|
189 |
+
h, w = tf.shape(x)[1], tf.shape(x)[2]
|
190 |
+
x = self.conv(x)
|
191 |
+
|
192 |
+
return tf.reshape(x, [-1, h * w * self.num_anchor, 4])
|
193 |
+
|
194 |
+
def get_config(self):
|
195 |
+
config = {
|
196 |
+
'num_anchor': self.num_anchor,
|
197 |
+
'wd': self.wd,
|
198 |
+
}
|
199 |
+
base_config = super(BboxHead, self).get_config()
|
200 |
+
return dict(list(base_config.items()) + list(config.items()))
|
201 |
+
|
202 |
+
|
203 |
+
class LandmarkHead(tf.keras.layers.Layer):
|
204 |
+
"""Landmark Head Layer"""
|
205 |
+
def __init__(self, num_anchor, wd, name='LandmarkHead', **kwargs):
|
206 |
+
super(LandmarkHead, self).__init__(name=name, **kwargs)
|
207 |
+
self.num_anchor = num_anchor
|
208 |
+
self.wd = wd
|
209 |
+
self.conv = Conv2D(filters=num_anchor * 10, kernel_size=1, strides=1)
|
210 |
+
|
211 |
+
def call(self, x):
|
212 |
+
h, w = tf.shape(x)[1], tf.shape(x)[2]
|
213 |
+
x = self.conv(x)
|
214 |
+
|
215 |
+
return tf.reshape(x, [-1, h * w * self.num_anchor, 10])
|
216 |
+
|
217 |
+
def get_config(self):
|
218 |
+
config = {
|
219 |
+
'num_anchor': self.num_anchor,
|
220 |
+
'wd': self.wd,
|
221 |
+
}
|
222 |
+
base_config = super(LandmarkHead, self).get_config()
|
223 |
+
return dict(list(base_config.items()) + list(config.items()))
|
224 |
+
|
225 |
+
|
226 |
+
class ClassHead(tf.keras.layers.Layer):
|
227 |
+
"""Class Head Layer"""
|
228 |
+
def __init__(self, num_anchor, wd, name='ClassHead', **kwargs):
|
229 |
+
super(ClassHead, self).__init__(name=name, **kwargs)
|
230 |
+
self.num_anchor = num_anchor
|
231 |
+
self.wd = wd
|
232 |
+
self.conv = Conv2D(filters=num_anchor * 2, kernel_size=1, strides=1)
|
233 |
+
|
234 |
+
def call(self, x):
|
235 |
+
h, w = tf.shape(x)[1], tf.shape(x)[2]
|
236 |
+
x = self.conv(x)
|
237 |
+
|
238 |
+
return tf.reshape(x, [-1, h * w * self.num_anchor, 2])
|
239 |
+
|
240 |
+
def get_config(self):
|
241 |
+
config = {
|
242 |
+
'num_anchor': self.num_anchor,
|
243 |
+
'wd': self.wd,
|
244 |
+
}
|
245 |
+
base_config = super(ClassHead, self).get_config()
|
246 |
+
return dict(list(base_config.items()) + list(config.items()))
|
247 |
+
|
248 |
+
|
249 |
+
def RetinaFaceModel(cfg, training=False, iou_th=0.4, score_th=0.02,
|
250 |
+
name='RetinaFaceModel'):
|
251 |
+
"""Retina Face Model"""
|
252 |
+
input_size = cfg['input_size'] if training else None
|
253 |
+
wd = cfg['weights_decay']
|
254 |
+
out_ch = cfg['out_channel']
|
255 |
+
num_anchor = len(cfg['min_sizes'][0])
|
256 |
+
backbone_type = cfg['backbone_type']
|
257 |
+
|
258 |
+
# define model
|
259 |
+
x = inputs = Input([input_size, input_size, 3], name='input_image')
|
260 |
+
|
261 |
+
x = Backbone(backbone_type=backbone_type)(x)
|
262 |
+
|
263 |
+
fpn = FPN(out_ch=out_ch, wd=wd)(x)
|
264 |
+
|
265 |
+
features = [SSH(out_ch=out_ch, wd=wd)(f)
|
266 |
+
for i, f in enumerate(fpn)]
|
267 |
+
|
268 |
+
bbox_regressions = tf.concat(
|
269 |
+
[BboxHead(num_anchor, wd=wd)(f)
|
270 |
+
for i, f in enumerate(features)], axis=1)
|
271 |
+
landm_regressions = tf.concat(
|
272 |
+
[LandmarkHead(num_anchor, wd=wd, name=f'LandmarkHead_{i}')(f)
|
273 |
+
for i, f in enumerate(features)], axis=1)
|
274 |
+
classifications = tf.concat(
|
275 |
+
[ClassHead(num_anchor, wd=wd, name=f'ClassHead_{i}')(f)
|
276 |
+
for i, f in enumerate(features)], axis=1)
|
277 |
+
|
278 |
+
classifications = tf.keras.layers.Softmax(axis=-1)(classifications)
|
279 |
+
|
280 |
+
if training:
|
281 |
+
out = (bbox_regressions, landm_regressions, classifications)
|
282 |
+
else:
|
283 |
+
# only for batch size 1
|
284 |
+
preds = tf.concat( # [bboxes, landms, landms_valid, conf]
|
285 |
+
[bbox_regressions[0],
|
286 |
+
landm_regressions[0],
|
287 |
+
tf.ones_like(classifications[0, :, 0][..., tf.newaxis]),
|
288 |
+
classifications[0, :, 1][..., tf.newaxis]], 1)
|
289 |
+
priors = prior_box_tf((tf.shape(inputs)[1], tf.shape(inputs)[2]), cfg['min_sizes'], cfg['steps'], cfg['clip'])
|
290 |
+
decode_preds = decode_tf(preds, priors, cfg['variances'])
|
291 |
+
|
292 |
+
selected_indices = tf.image.non_max_suppression(
|
293 |
+
boxes=decode_preds[:, :4],
|
294 |
+
scores=decode_preds[:, -1],
|
295 |
+
max_output_size=tf.shape(decode_preds)[0],
|
296 |
+
iou_threshold=iou_th,
|
297 |
+
score_threshold=score_th)
|
298 |
+
|
299 |
+
out = tf.gather(decode_preds, selected_indices)
|
300 |
+
|
301 |
+
return Model(inputs, out, name=name), Model(inputs, [bbox_regressions, landm_regressions, classifications], name=name + '_bb_only')
|
retina_model/ops.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from retinaface.anchor import decode_tf, prior_box_tf
|
2 |
+
import tensorflow as tf
|
3 |
+
|
4 |
+
|
5 |
+
def extract_detections(bbox_regressions, landm_regressions, classifications, image_sizes, iou_th=0.4, score_th=0.02):
|
6 |
+
min_sizes = [[16, 32], [64, 128], [256, 512]]
|
7 |
+
steps = [8, 16, 32]
|
8 |
+
variances = [0.1, 0.2]
|
9 |
+
preds = tf.concat( # [bboxes, landms, landms_valid, conf]
|
10 |
+
[bbox_regressions,
|
11 |
+
landm_regressions,
|
12 |
+
tf.ones_like(classifications[:, 0][..., tf.newaxis]),
|
13 |
+
classifications[:, 1][..., tf.newaxis]], 1)
|
14 |
+
priors = prior_box_tf(image_sizes, min_sizes, steps, False)
|
15 |
+
decode_preds = decode_tf(preds, priors, variances)
|
16 |
+
|
17 |
+
selected_indices = tf.image.non_max_suppression(
|
18 |
+
boxes=decode_preds[:, :4],
|
19 |
+
scores=decode_preds[:, -1],
|
20 |
+
max_output_size=tf.shape(decode_preds)[0],
|
21 |
+
iou_threshold=iou_th,
|
22 |
+
score_threshold=score_th)
|
23 |
+
|
24 |
+
out = tf.gather(decode_preds, selected_indices)
|
25 |
+
|
26 |
+
return out
|
27 |
+
|