-
Notifications
You must be signed in to change notification settings - Fork 0
/
inference_core_MiVOS_HoliMem.py
336 lines (254 loc) · 12.8 KB
/
inference_core_MiVOS_HoliMem.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
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
"""
This file can handle DAVIS 2016/2017 evaluation.
"""
import torch
import numpy as np
import cv2
import torch.nn.functional as F
from MiVOS.model.propagation.prop_net import PropagationNetwork
from model.aggregate import aggregate_wbg
from util.tensor_util import pad_divide_by
# Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ
# modified by Stéphane Vujasinovic is enclaved in such manner
from icecream import ic
import yaml
import os
import sys
# ic(os.getcwd())
from HoliMem.HoliMem import HoliMem
from atomic_crop import super_crop, uncrop_mask, extract_size_of_target_for_window_filtering
import torchvision.transforms as T
from PIL import Image
transform = T.ToPILImage()
import lovely_tensors as lt
from Toolbox_Eteph.Debugging.ST_LT_memory_plot.Plot_ST_and_LT_Memory import *
# Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ Σ
class InferenceCore:
def __init__(self, prop_net: PropagationNetwork, num_objects, mem_config:str, debugging_flag=False, record_det_flag=False, use_super_crop=False):
self.prop_net = prop_net
self.device = 'cuda'
self.k = num_objects
# Initialize HoliMem attributes
self.HoliMem = HoliMem(mem_config, debugging_flag=True)
self.HoliMem.nbr_of_objects_working_with = num_objects
# Initialize flags for helping by the debugging process
self.debugging_flag = debugging_flag
self.record_det_flag = record_det_flag
# self.use_super_crop = True if use_super_crop !=1.0 else False ## TODO SET A VARIABLE THAT ALLOWS ME TO CHOSE THIS
# self.use_super_crop = False ## TODO SET A VARIABLE THAT ALLOWS ME TO CHOSE THIS
# if self.use_super_crop:
self.Affinity_mode = 0 # Affinity 0 is the one with query keys and memory keys, for STCN this does not matter
self.mode = 1 # mode 1 default , 2 padding, 3 crops
# self.mode = 3 # mode 1 default , 2 padding, 3 crop + padding based on THOR
if 3 == self.mode:
self.hypo_coeff = 1.44
# self.use_cosine_sim = True # True default, if False use dot product normalized on the annotated frame
# self.use_cosine_sim = False # True default, if False use dot product normalized on the annotated frame
# self.use_cosine_sim = True
def get_path_2_image_folder(self, img_folder: str):
self.path_2_image_folder = img_folder
def unpad(self, data, pad):
if len(data.shape) == 4:
if pad[2] + pad[3] > 0:
data = data[:, :, pad[2]:-pad[3], :]
if pad[0] + pad[1] > 0:
data = data[:, :, :, pad[0]:-pad[1]]
elif len(data.shape) == 3:
if pad[2] + pad[3] > 0:
data = data[:, pad[2]:-pad[3], :]
if pad[0] + pad[1] > 0:
data = data[:, :, pad[0]:-pad[1]]
else:
raise NotImplementedError
return data
def _get_query_kv_buffered(self, image):
# not actually buffered
return self.prop_net.get_query_values(image.cuda())
# f16, f8, f4, k16, v16 = self.prop_net.get_query_values(image.cuda())
# _, _, h, w = f16.size()
# pre_mask, _ = torch.max(self.prob[1:], dim=0)
# pre_mask = pre_mask.unsqueeze(0)
# pre_mask = F.interpolate(pre_mask, size=[h, w], mode='bilinear')
# concat_f16 = torch.cat([f16, pre_mask], dim=1)
# concat_f16 = self.prop_net.concat_conv(concat_f16)
# concat_f16 = torch.sigmoid(concat_f16)
# concat_f16 = f16 * concat_f16
# k16, v16 = self.prop_net.kv_q_f16(concat_f16)
# result = (concat_f16, f8, f4, k16, v16)
# return result
def _set_image(self, sequence_length, OG_image):
# True dimensions
OG_image = OG_image.unsqueeze(dim=0).cuda()
# self.t = sequence_length
self.image, self.pad = pad_divide_by(OG_image, 16)
def set_annotated_frame(self, idx, sequence_length, image, anno_mask):
self.HoliMem.reset_HoliMem() # Reset the ST and LT memories
# print(image.shape)
self._set_image(sequence_length, image)
self.annotated_image = self.image.clone()
anno_mask = anno_mask.unsqueeze(dim=1)
mask, _ = pad_divide_by(anno_mask.cuda(), 16)
self.prob = aggregate_wbg(mask, keep_bg=True)
# print(anno_mask.shape)
# print(mask.shape)
# print(self.prob.shape)
# # CROP the IMAGE region ?
# if not self.use_super_crop:
# self.prob = aggregate_wbg(mask, keep_bg=True)
# else:
# # CROP AND RESIZE
# self.IMG_RESOLUTION = self.image.shape[-2:]
# # print(self.IMG_RESOLUTION)
# self.image, crop_prob, _ , _= super_crop(self.image.clone(), self.prob.clone(), self.k, self.hypo_coeff, None)
# # self.prev_mask_to_use_for_crops = self.mask
#
# # img_PIL = transform(self.image.clone().squeeze())
# # img_PIL.show()
#
# self.prob = crop_prob
# KV pair for the interacting frame
print(self.image.shape)
print(self.prob.shape)
# anno_key_k, anno_key_v = self.prop_net.memorize(self.image.cuda(),self.prob[1:].cuda())
# Add the annotated frame to the ST and LT Memory
if 1 == self.mode:
anno_key_k, anno_key_v = self.prop_net.memorize(self.image.cuda(), self.prob[1:].cuda())
self.HoliMem.update_HoliMem(idx, anno_key_k, anno_key_v)
# elif 3 == self.mode:
# # self.idx_for_mode_3 = -2 # 0 for features from the bakcbone ResNet or -2 for keys
# self.idx_for_mode_3 = 0 # 0 for features from the bakcbone ResNet or -2 for keys
# self.IMG_RESOLUTION = self.image.shape
# crop_img, crop_prob, crop_vector, pad_vector = super_crop(self.image.clone(), self.prob, self.k, self.hypo_coeff, None)
#
# c_query = self._get_query_kv_buffered(crop_img)
# feature_f16 = c_query[self.idx_for_mode_3]
# anno_key_k, anno_key_v = self.prop_net.memorize(crop_img.cuda(), crop_prob[1:].cuda())
# # self.HoliMem.update_HoliMem_based_on_crop_mode_3(idx, anno_key_k, anno_key_v, feature_f16)
# self.HoliMem.update_HoliMem(idx, anno_key_k, anno_key_v)
#
# self.prob = uncrop_mask(crop_prob, crop_vector, pad_vector, self.IMG_RESOLUTION)
# if not self.use_cosine_sim:
# self.HoliMem.LT_HoliMem.reset_LT_gram_matrix(anno_key_k)
return self.unpad(self.prob,self.pad)
def _adapt_img(self,OG_image):
return pad_divide_by(OG_image.unsqueeze(dim=0).cuda(), 16)
def step(self, idx, image):
print('idx:',idx)
# print(image.shape)
# Extract the key and values of the current frame
img, pad = self._adapt_img(image)
ic('check1')
# if idx == 3237:
# transform(img[0]).show()
# print(img)
ori_img = img.clone()
# CROP the IMAGE region ?
# if 3 == self.mode:
# img, crop_prob, crop_vector, pad_vector = super_crop(ori_img.clone(), self.prob.clone(), self.k, self.hypo_coeff, None)
# if idx == 3237:
# print(img.shape)
# transform(img[0]).show()
# print('hi')
query = self._get_query_kv_buffered(img)
ic('check2')
# Extract the holistic representation
HoliMem_idx_list, HoliMem_keys, HoliMem_values = self.HoliMem.get_holistic_memory()
ic('check3')
# Infer the segmentation mask based on the holistic representation stored in the memory
Holi_out_mask = self.prop_net.segment_with_query(HoliMem_keys, HoliMem_values, *query)
Holi_out_mask = aggregate_wbg(Holi_out_mask, keep_bg=True)
self.prob = Holi_out_mask
ic('check4')
# self.display_an_embending(self.prob)# Display memory values or others
# if self.use_super_crop:
# self.prev_mask_to_use_for_crops = uncrop(Holi_out_mask.clone(), crop_vector, self.IMG_RESOLUTION)
# # self.prob = uncrop(self.prob.clone(), crop_vector, self.IMG_RESOLUTION)
# Extract the features of the current frame through the memory network
# prev_key, prev_value = self.prop_net.memorize(img, Holi_out_mask[1:])
# print(Holi_out_mask.shape)
# prev_key, prev_value = self.prop_net.memorize(ori_img, Holi_out_mask[1:])
# # Update the holistic memory
# if 0 == self.Affinity_mode:
# self.HoliMem.set_affinity_matrices(self.prop_net.get_affinity())
# elif 1 == self.Affinity_mode:
# Ddd = self.prop_net.compute_similarites_for_memory_key_types(HoliMem_keys,prev_key)
# self.HoliMem.set_affinity_matrices(Ddd)
# TESTING !!
if 1 == self.mode:
prev_key, prev_value = self.prop_net.memorize(ori_img, Holi_out_mask[1:])
# print(prev_key.shape)
ic('check5')
# Update the holistic memory
if 0 == self.Affinity_mode:
self.HoliMem.set_affinity_matrices(self.prop_net.get_affinity())
elif 1 == self.Affinity_mode:
Ddd = self.prop_net.compute_similarites_for_memory_key_types(HoliMem_keys, prev_key)
self.HoliMem.set_affinity_matrices(Ddd)
ic('check6')
self.HoliMem.update_HoliMem(idx, prev_key, prev_value)
# elif 2 == self.mode:
# ###########################################
# # crop_region = extract_size_of_target_for_window_filtering(self.prob,1.0)
# c_img, _, crop_vector = super_crop(img.clone(), self.prob, self.k,
# 1.1, None)
# c_prev_key, c_prev_value = self.prop_net.memorize(c_img, Holi_out_mask[1:])
# ###########################################
# self.HoliMem.find_best_LT_idx(idx, prev_key, prev_value, c_prev_key, c_prev_value)
# elif 3 == self.mode:
# print(img.shape)
# self.prob = uncrop_mask(self.prob, crop_vector, pad_vector, self.IMG_RESOLUTION)
#
# img, crop_prob, crop_vector, pad_vector = super_crop(ori_img.clone(), self.prob, self.k, self.hypo_coeff, None)
#
#
# prev_key, prev_value = self.prop_net.memorize(img.cuda(), crop_prob[1:].cuda())
# c_query = self._get_query_kv_buffered(img)
#
# # print(c_query[0].shape) # Directly the feature space
# # print(c_query[-2].shape) # Using the key vectors
#
# feature_f16 = c_query[self.idx_for_mode_3]
# # self.HoliMem.update_HoliMem_based_on_crop_mode_3(idx, prev_key, prev_value, feature_f16)
# self.HoliMem.update_HoliMem(idx, prev_key, prev_value)#, feature_f16)
ic('check7')
return self.unpad(self.prob, pad)
@property
def return_lt_det(self):
return self.HoliMem.LT_gram_det.copy()
@property
def ST_N_LT_Memories(self):
return self.HoliMem.ST_Memory_indexes.copy(), self.HoliMem.LT_Memory_indexes.copy()
@ST_N_LT_Memories.setter
def ST_N_LT_Memories(self, new_st_indexes, new_lt_indexes):
self.HoliMem.ST_Memory_indexes = new_st_indexes
self.HoliMem.LT_Memory_indexes = new_lt_indexes
@property
def get_size_of_ST_N_LT_memory(self):
return self.HoliMem.get_size_of_ST_N_LT_memory
def display_an_embending(self, input):
# Take a look at the memory value of the annotated frame
print(input)
print(input.shape)
# numpy_memory_value = input[0,:,0].clone().permute(1,2,0).detach().cpu().numpy()
numpy_memory_value = input[:,0].clone().permute(1,2,0).detach().cpu().numpy()[:,:,1]# take the obj, or 0 for the background
print(numpy_memory_value.shape)
# np_2_img = numpy_memory_value.std(axis=2)
np_2_img = numpy_memory_value
# np_2_img = np_2_img*255
# np_2_img = np_2_img - np_2_img.min()
# np_2_img = np_2_img/(np_2_img.max())*255
# np_2_img = numpy_memory_value.min(axis=2)
# np_2_img = numpy_memory_value.std(axis=2)
print(np_2_img.shape)
resized = cv2.resize(np_2_img, (912,480), interpolation=cv2.INTER_AREA)
#
while True:
cv2.imshow('Memory_value', resized)#.astype(np.int8))
# waits for user to press any key
# (this is necessary to avoid Python kernel form crashing)
cv2.waitKey(0)
if cv2.key == ord('q'):
# closing all open windows
cv2.destroyAllWindows()
break
print('hi')