-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
312 lines (251 loc) · 13.2 KB
/
utils.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
312
import cv2
import matplotlib.pyplot as plt
import numpy as np
import torch
import torch.nn.functional as F
from PIL import Image
from losses import get_accum_card_error
font = cv2.FONT_HERSHEY_SIMPLEX
org = (12, 25)
fontScale = 0.5
color = (0, 255, 0)
thickness = 2
points_colours = {'upper_left_face_visible': (0, 0, 255),
'upper_right_face_visible': (0, 255, 0),
'upper_left_back_visible': (255, 0, 0),
'upper_right_back_visible': (0, 255, 255),
'lower_left_face_visible': (255, 0, 255),
'lower_right_face_visible': (255, 255, 0),
'lower_left_back_visible': (50, 128, 255),
'lower_right_back_visible': (255, 128, 0)}
points_colours_abbreviations = {'upper_left_face_visible': 'ulf',
'upper_right_face_visible': 'urf',
'upper_left_back_visible': 'ulb',
'upper_right_back_visible': 'urb',
'lower_left_face_visible': 'llf',
'lower_right_face_visible': 'lrf',
'lower_left_back_visible': 'llb',
'lower_right_back_visible': 'lrb'}
card_edges = {'upper_back_edge': ['upper_left_back_x',
'upper_left_back_y',
'upper_right_back_x',
'upper_right_back_y'],
'upper_face_edge': ['upper_left_face_x',
'upper_left_face_y',
'upper_right_face_x',
'upper_right_face_y'],
'upper_left_edge': ['upper_left_face_x',
'upper_left_face_y',
'upper_left_back_x',
'upper_left_back_y'],
'upper_right_edge': ['upper_right_face_x',
'upper_right_face_y',
'upper_right_back_x',
'upper_right_back_y'],
'lower_back_edge': ['lower_left_back_x',
'lower_left_back_y',
'lower_right_back_x',
'lower_right_back_y'],
'lower_face_edge': ['lower_left_face_x',
'lower_left_face_y',
'lower_right_face_x',
'lower_right_face_y'],
'lower_left_edge': ['lower_left_face_x',
'lower_left_face_y',
'lower_left_back_x',
'lower_left_back_y'],
'lower_right_edge': ['lower_right_face_x',
'lower_right_face_y',
'lower_right_back_x',
'lower_right_back_y'],
}
# TODO: Fix the issue with denormalisation?
def show_learning_sample_random(inputs, values, output_value, output_detections, output_points, mean, std, waitTime=0):
global points_colours, font, org, fontScale, color, thickness
idx = np.random.randint(0, inputs.shape[0])
image = np.transpose(inputs[idx].numpy(), [1, 2, 0]) * std + mean
image = image[:, :, [2, 1, 0]]
image = Image.fromarray((image * 255).astype(np.uint8))
image = np.array(image)
target_value = values[idx].item()
value = round(output_value[idx].item(), 4)
detections = output_detections[idx].numpy()
points = output_points[idx].numpy().reshape(-1, 2)
points[:, 0] = points[:, 0] * image.shape[0]
points[:, 1] = points[:, 1] * image.shape[1]
for i in range(len(points)):
if detections[i] > 0.5:
x = int(points[i, 0])
y = int(points[i, 1])
image = cv2.circle(image, (y, x), 3, points_colours[list(points_colours.keys())[i]], thickness=-1)
image = cv2.putText(image, str(int(value * 52)) + "-" + str(int(target_value * 52)), org, font, fontScale, color,
thickness, cv2.LINE_AA)
cv2.imshow('Sample image', image)
cv2.waitKey(waitTime)
# cv2.destroyAllWindows()
# TODO: Fix the issue with denormalisation?
def show_learning_sample_batch(inputs, values, output_value, output_detections, output_points, mean, std, waitTime=0):
global points_colours, font, org, fontScale, color, thickness
for idx in range(inputs.shape[0]):
image = np.transpose(inputs[idx].numpy(), [1, 2, 0]) # RGB?
image = image[:, :, [2, 1, 0]] # BGR?
image = Image.fromarray((image * 255).astype(np.uint8))
image = np.array(image)
target_value = values[idx].item()
value = round(output_value[idx].item(), 4)
detections = output_detections[idx].numpy()
points = output_points[idx].numpy().reshape(-1, 2)
points[:, 0] = points[:, 0] * image.shape[0]
points[:, 1] = points[:, 1] * image.shape[1]
for i in range(len(points)):
if detections[i] > 0.5:
x = int(points[i, 0])
y = int(points[i, 1])
image = cv2.circle(image, (y, x), 3, points_colours[list(points_colours.keys())[i]], thickness=-1)
image = cv2.putText(image, str(int(value * 52)) + "-" + str(int(target_value * 52)), org, font, fontScale,
color, thickness, cv2.LINE_AA)
cv2.imshow('Validation image', image)
cv2.waitKey(waitTime)
# cv2.destroyAllWindows()
# TODO: Fix the issue with denormalisation?
def show_mask_points_sample_batch(inputs, values, output_value, output_detections, output_points, output_masks, mean,
std, waitTime=0):
global points_colours, font, org, fontScale, color, thickness
for idx in range(inputs.shape[0]):
image = np.transpose(inputs[idx].numpy(), [1, 2, 0]) # RGB?
image = image[:, :, [2, 1, 0]] # BGR?
image = Image.fromarray((image * 255).astype(np.uint8))
image = np.array(image)
masks = np.transpose(output_masks[idx].numpy(), [1, 2, 0])
# masks = masks/masks.max(axis=(0,1))
grid = create_mask_grid(masks)
cv2.imshow('Validation image - Masks', grid)
target_value = values[idx].item()
value = round(output_value[idx].item(), 4)
detections = output_detections[idx].numpy()
points = output_points[idx].numpy().reshape(-1, 2)
points[:, 0] = points[:, 0] * image.shape[0]
points[:, 1] = points[:, 1] * image.shape[1]
for i in range(len(points)):
if detections[i] > 0.5:
x = int(points[i, 0])
y = int(points[i, 1])
image = cv2.circle(image, (y, x), 3, points_colours[list(points_colours.keys())[i]], thickness=-1)
image = cv2.putText(image, str(int(value * 52)) + "-" + str(int(target_value * 52)), org, font, fontScale,
color, thickness, cv2.LINE_AA)
cv2.imshow('Validation image', image)
cv2.waitKey(waitTime)
# TODO: Fix the issue with denormalisation?
def show_points_from_mask_sample_batch(inputs, values, output_value, output_detections, output_masks, mean, std,
threshold=0.1, waitTime=0):
global points_colours, font, org, fontScale, color, thickness
for idx in range(inputs.shape[0]):
image = np.transpose(inputs[idx].numpy(), [1, 2, 0]) # RGB?
image = image[:, :, [2, 1, 0]] # BGR?
image = Image.fromarray((image * 255).astype(np.uint8))
image = np.array(image)
masks = np.transpose(output_masks[idx].numpy(), [1, 2, 0])
# masks = masks/masks.max(axis=(0,1))
grid = create_mask_grid(masks)
cv2.imshow('Validation image - Masks', grid)
points = get_points_from_mask(masks)
points[:, 0] = points[:, 0] * image.shape[0]
points[:, 1] = points[:, 1] * image.shape[1]
target_value = values[idx].item()
out_value = torch.argmax(F.softmax(output_value, dim=1), axis=1)
value = round(out_value[idx].item(), 4)
# detections = output_detections[idx].numpy()
for i in range(len(points)):
if output_detections[idx, i] >= threshold:
x = int(points[i, 0])
y = int(points[i, 1])
image = cv2.circle(image, (y, x), 3, points_colours[list(points_colours.keys())[i]], thickness=-1)
# image = cv2.putText(image,
# str(int(value * 52)) + "-" + str(int(target_value * 52)),
# org, font, fontScale,
# color, thickness, cv2.LINE_AA)
image = cv2.putText(image,
str(int(value)) + "-" + str(int(target_value * 52)),
org, font, fontScale,
color, thickness, cv2.LINE_AA)
cv2.imshow('Validation image', image)
cv2.waitKey(waitTime)
def show_mask_only_sample(output_masks, idx, save=True, waitTime=33):
masks = np.transpose(output_masks[idx].cpu().numpy(), [1, 2, 0])
grid = create_mask_grid(masks)
cv2.imshow('Validation image - Masks', grid)
cv2.waitKey(waitTime)
if save:
cv2.imwrite('masks_output.jpg', (np.clip(grid, 0, 1) * 255).astype(np.uint8))
def create_mask_grid(masks, padding=((1, 1), (1, 1)), padding_values=((0.5, 0.5), (0.5, 0.5))):
global points_colours, font, points_colours_abbreviations
grid1_stack = []
for i in range(4):
mask_i = np.pad(masks[:, :, i], padding, 'constant', constant_values=padding_values)
mask_i = cv2.cvtColor(mask_i, cv2.COLOR_GRAY2BGR)
colour = points_colours[list(points_colours.keys())[i]]
cv2.putText(mask_i, points_colours_abbreviations[list(points_colours.keys())[i]], (35, 8), font, 0.25, colour,
thickness=1)
grid1_stack.append(mask_i)
grid1 = np.hstack(grid1_stack)
grid2_stack = []
for i in range(4, 8):
mask_i = np.pad(masks[:, :, i], ((1, 1), (1, 1)), 'constant', constant_values=((0.5, 0.5), (0.5, 0.5)))
mask_i = cv2.cvtColor(mask_i, cv2.COLOR_GRAY2BGR)
colour = points_colours[list(points_colours.keys())[i]]
cv2.putText(mask_i, points_colours_abbreviations[list(points_colours.keys())[i]], (35, 8), font, 0.25, colour,
thickness=1)
grid2_stack.append(mask_i)
grid2 = np.hstack(grid2_stack)
grid = np.vstack([grid1, grid2])
return grid
def get_points_from_mask(masks):
mask_points = []
mask_shape = masks[:, :, 0].shape
for i in range(8):
mask_i = masks[:, :, i]
coordy, coordx = np.where(mask_i == mask_i.max())
mask_points.append(np.array([[coordy[0] / mask_shape[0], coordx[0] / mask_shape[1]]]))
mask_points = np.stack(mask_points, axis=1)
points = mask_points.reshape(-1, 2)
return points # normalised! they still need to be multiplied by the real shape of the image
def show_model_validation(model, validation_loader, mean, std, waitTime=0):
running_card_error = 0.0
batch_size = 0
for batch_index, batch in enumerate(validation_loader):
if batch_size == 0:
batch_size = batch['image'].shape[0]
inputs = batch['image']
value = batch['value']
model.eval()
with torch.no_grad():
out_mask, out_detections, out_value = model(inputs)
show_points_from_mask_sample_batch(inputs, value,
out_value, out_detections, out_mask,
mean, std, waitTime=waitTime)
running_card_error += get_accum_card_error(value, out_value).item()
cv2.destroyAllWindows()
card_error_epoch = int(round(running_card_error / (len(validation_loader) * batch_size), 1))
print('Test - Mean Card error: {} cards.'.format(card_error_epoch))
def get_train_val_indices(dataset, random_seed, validation_split, shuffle_dataset=True):
# Splitting the dataset into train and validation
dataset_size = len(dataset)
indices = list(range(dataset_size))
split = int(np.floor(validation_split * dataset_size))
if shuffle_dataset:
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]
return train_indices, val_indices
def get_point_edges(data):
global card_edges
edge_cols = []
for edge, pts in card_edges.items():
ptAx, ptAy = data[pts[0]], data[pts[1]]
ptBx, ptBy = data[pts[2]], data[pts[3]]
vector = np.array([ptAx - ptBx, ptAy - ptBy]).T
norm = np.linalg.norm(vector, axis=1)
data[edge] = norm
data.loc[(ptAx == 0) | (ptAy == 0) | (ptBx == 0) | (ptBy == 0), edge] = 0 # Maybe -1 was not working?
edge_cols.append(edge)
return data, edge_cols