RocketGit

hahattpro / crowd_counting_framework (public) (License: Unspecified) (since 2020-12-23) (hash sha1)

backupfrom https://github.com/ttpro1995/crowd_counting_framework

Clone URLs: https://rocketgit.com/user/hahattpro/crowd_counting_framework ssh://rocketgit@ssh.rocketgit.com/user/hahattpro/crowd_counting_framework git://git.rocketgit.com/user/hahattpro/crowd_counting_framework

List of commits:

Subject	Hash	Author	Date (UTC)
add perspective	642d6fff8c9f31e510fda85a7fb631fb855d8a6d	Thai Thien	2019-10-06 16:54:44
fix padding with p	86c2fa07822d956a34b3b37e14da485a4249f01b	Thai Thien	2019-10-06 02:52:58
pacnn perspective loss	fb673e38a5f24ae9004fe2b7b93c88991e0c2304	Thai Thien	2019-10-06 01:38:28
data_flow shanghaitech_pacnn_with_perspective seem working	91d350a06f358e03223966297d124daee94123d0	Thai Thien	2019-10-06 01:31:11
multiscale loss and final loss only mode	c65dd0e74ad28503821e5c8651a3b47b4a0c7c64	Thai Thien	2019-10-05 15:58:19
wip : perspective map	eac63f2671dc5b064753acc4f40bf0f9f216ad2a	Thai Thien	2019-10-04 16:26:56
shell script	f2106e700b6f6174d4dd276f25ec6f3d9ff239bb	thient	2019-10-04 07:42:51
WIP	42c7c8e1d772fbbda61a4bdf9e329f74e1efb600	tthien	2019-10-03 17:52:47
add readme	580cf43d1edddd67b1f6a2c57fdd5cee3dba925c	Thai Thien	2019-10-02 17:44:49
update script, debug	ddb68b95389be1c1d398118677dd227a8bb2b70b	Thai Thien	2019-10-02 15:52:31
add d (output density map) to loss function)	a0c71bf4bf2ab7393d60b06a84db8dfbbfb1a6c2	tthien	2019-09-30 16:32:39
fix the args, add save interval for model, so we don't save them all	9fdf9daa2ac4bd12b7b62521d81e520db0debd01	tthien	2019-09-30 16:30:00
meow	1ad19a22a310992e27a26471feeb37375124d075	tthien	2019-09-29 18:25:43
fix pacnn perspective map	453ece3ccb818889ba895bfc4285f7905d33cba5	Thai Thien	2019-09-25 17:20:33
apex not work so well	da8c0dd57297f972201f31d57e66897177922f48	Thai Thien	2019-09-24 17:25:59
fix data loader pacnn so it will scale up with correct number of people	11d55b50d764511f2491291f0208fee0905dec49	Thai Thien	2019-09-24 15:40:56
add comet ml	a9d4b89ce594f5e241168ccafdcdf0f150ea0ebb	Thai Thien	2019-09-23 17:07:58
fix pacnn avg schema	c2140a96886195782e5689c24aeeb4fe7a2db7ad	Thai Thien	2019-09-22 17:35:01
debug number not divisible by 8	a568fd7f294a8bd31b3db78437b4b6b51b5b41b9	Thai Thien	2019-09-22 04:36:06
pacnn	967074890d14ab0eefc277801860270a468e8f9f	Thai Thien	2019-09-22 03:54:48

Commit 642d6fff8c9f31e510fda85a7fb631fb855d8a6d - add perspective
Author: Thai Thien
Author date (UTC): 2019-10-06 16:54
Committer name: Thai Thien
Committer date (UTC): 2019-10-06 16:54
Parent(s): 86c2fa07822d956a34b3b37e14da485a4249f01b
Signing key:
Tree: 3816f82b3c01435472a8758e7e4322db1b3a6a6f

File	Lines added	Lines deleted
args_util.py	1	0
data_flow.py	4	2
main_pacnn.py	72	7
train_script/test_pacnn_shanghaitechA.sh	16	0
train_script/train_pacnn_shanghaitechA.sh	51	9

File args_util.py changed (mode: 100644) (index 3fa2c42..2e74c07)
...	...	def real_args_parse():
72	72	parser.add_argument('--momentum', action="store", default=0.9, type=float)	parser.add_argument('--momentum', action="store", default=0.9, type=float)
73	73	parser.add_argument('--decay', action="store", default=5*1e-3, type=float)	parser.add_argument('--decay', action="store", default=5*1e-3, type=float)
74	74	parser.add_argument('--epochs', action="store", default=1, type=int)	parser.add_argument('--epochs', action="store", default=1, type=int)
	75		parser.add_argument('--test', action="store_true", default=False)
75	76
76	77	# pacnn setting only	# pacnn setting only
77	78	parser.add_argument('--PACNN_PERSPECTIVE_AWARE_MODEL', action="store", default=0, type=int)	parser.add_argument('--PACNN_PERSPECTIVE_AWARE_MODEL', action="store", default=0, type=int)

File main_pacnn.py changed (mode: 100644) (index 823774f..cb90fab)
...	...	if __name__ == "__main__":
24	24
25	25	# Add the following code anywhere in your machine learning file	# Add the following code anywhere in your machine learning file
26	26	experiment = Experiment(api_key="S3mM1eMq6NumMxk2QJAXASkUM",	experiment = Experiment(api_key="S3mM1eMq6NumMxk2QJAXASkUM",
27		project_name="pacnn-dev2", workspace="ttpro1995", disabled=True)
	27		project_name="pacnn-dev2", workspace="ttpro1995")
28	28
29	29	args = real_args_parse()	args = real_args_parse()
30		device = "cpu"
31	30	print(device)	print(device)
32	31	print(args)	print(args)
33	32
34	33
35	34
36	35	MODEL_SAVE_NAME = args.task_id	MODEL_SAVE_NAME = args.task_id
	36		TEST = args.test
37	37	MODEL_SAVE_INTERVAL = 5	MODEL_SAVE_INTERVAL = 5
38	38	DATA_PATH = args.input	DATA_PATH = args.input
39	39	TOTAL_EPOCH = args.epochs	TOTAL_EPOCH = args.epochs
40	40	PACNN_PERSPECTIVE_AWARE_MODEL = args.PACNN_PERSPECTIVE_AWARE_MODEL	PACNN_PERSPECTIVE_AWARE_MODEL = args.PACNN_PERSPECTIVE_AWARE_MODEL
41	41	PACNN_MUTILPLE_SCALE_LOSS = args.PACNN_MUTILPLE_SCALE_LOSS	PACNN_MUTILPLE_SCALE_LOSS = args.PACNN_MUTILPLE_SCALE_LOSS
42	42	DATASET_NAME = "shanghaitech_pacnn"	DATASET_NAME = "shanghaitech_pacnn"
	43
43	44	if PACNN_PERSPECTIVE_AWARE_MODEL:	if PACNN_PERSPECTIVE_AWARE_MODEL:
44	45	DATASET_NAME = "shanghaitech_pacnn_with_perspective"	DATASET_NAME = "shanghaitech_pacnn_with_perspective"
45	46

...	...	if __name__ == "__main__":
87	88	num_workers=4, dataset_name="shanghaitech_pacnn"),	num_workers=4, dataset_name="shanghaitech_pacnn"),
88	89	batch_size=1, num_workers=4)	batch_size=1, num_workers=4)
89	90
	91		test_loader_pacnn = torch.utils.data.DataLoader(
	92		ListDataset(test_list,
	93		shuffle=False,
	94		transform=transforms.Compose([
	95		transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406],
	96		std=[0.229, 0.224, 0.225]),
	97		]),
	98		train=False,
	99		batch_size=1,
	100		num_workers=4, dataset_name="shanghaitech_pacnn"),
	101		batch_size=1, num_workers=4)
	102
90	103	# create model	# create model
91	104	net = PACNNWithPerspectiveMap(perspective_aware_mode=PACNN_PERSPECTIVE_AWARE_MODEL).to(device)	net = PACNNWithPerspectiveMap(perspective_aware_mode=PACNN_PERSPECTIVE_AWARE_MODEL).to(device)
92	105	criterion_mse = nn.MSELoss(size_average=False).to(device)	criterion_mse = nn.MSELoss(size_average=False).to(device)

...	...	if __name__ == "__main__":
112	125	else:	else:
113	126	print("new model")	print("new model")
114	127
	128		if TEST:
	129		print("test model")
	130		mae_calculator_d1 = MAECalculator()
	131		mae_calculator_d2 = MAECalculator()
	132		mae_calculator_d3 = MAECalculator()
	133		mae_calculator_final = MAECalculator()
	134		with torch.no_grad():
	135		for val_img, label in test_loader_pacnn:
	136		net.eval()
	137		# load data
	138		d1_label, d2_label, d3_label = label
	139
	140		# forward pass
	141		d1, d2, d3, p_s, p, d = net(val_img.to(device))
	142
	143		d1_label = d1_label.to(device)
	144		d2_label = d2_label.to(device)
	145		d3_label = d3_label.to(device)
	146
	147		# score
	148		mae_calculator_d1.eval(d1.cpu().detach().numpy(), d1_label.cpu().detach().numpy())
	149		mae_calculator_d2.eval(d2.cpu().detach().numpy(), d2_label.cpu().detach().numpy())
	150		mae_calculator_d3.eval(d3.cpu().detach().numpy(), d3_label.cpu().detach().numpy())
	151		mae_calculator_final.eval(d.cpu().detach().numpy(), d1_label.cpu().detach().numpy())
	152		print("count ", mae_calculator_d1.count)
	153		print("d1_val ", mae_calculator_d1.get_mae())
	154		print("d2_val ", mae_calculator_d2.get_mae())
	155		print("d3_val ", mae_calculator_d3.get_mae())
	156		print("dfinal_val ", mae_calculator_final.get_mae())
	157		experiment.log_metric("d1_val", mae_calculator_d1.get_mae())
	158		experiment.log_metric("d2_val", mae_calculator_d2.get_mae())
	159		experiment.log_metric("d3_val", mae_calculator_d3.get_mae())
	160		experiment.log_metric("dfinal_val", mae_calculator_final.get_mae())
	161		exit()
	162
115	163	while current_epoch < TOTAL_EPOCH:	while current_epoch < TOTAL_EPOCH:
116	164	experiment.log_current_epoch(current_epoch)	experiment.log_current_epoch(current_epoch)
117	165	current_epoch += 1	current_epoch += 1

...	...	if __name__ == "__main__":
153	201	# TODO: loss for perspective map here	# TODO: loss for perspective map here
154	202	pad_p_0 = perspective_p.size()[2] - p.size()[2]	pad_p_0 = perspective_p.size()[2] - p.size()[2]
155	203	pad_p_1 = perspective_p.size()[3] - p.size()[3]	pad_p_1 = perspective_p.size()[3] - p.size()[3]
156		p_pad = F.pad(p, (0, pad_p_1, 0, pad_p_0), mode='replicate')
	204		if pad_p_0 == 0:
	205		pad_p_0 = -perspective_p.size()[2]
	206		if pad_p_1 == 0:
	207		pad_p_1 = -perspective_p.size()[3]
	208
	209		# p_pad = F.pad(p, (0, pad_p_1, 0, pad_p_0), mode='replicate')
	210		perspective_p_pad = perspective_p[:,:, 0:-pad_p_0, 0:-pad_p_1]
157	211
158		loss_p = criterion_mse(p_pad, perspective_p) + criterion_ssim(p_pad, perspective_p)
	212		# print(p.shape)
	213		# print(perspective_p.shape)
	214		# print(pad_p_0, pad_p_1)
	215		# print(perspective_p_pad.shape)
	216		loss_p = criterion_mse(p, perspective_p_pad) + criterion_ssim(p, perspective_p_pad)
159	217
160	218	loss += loss_p	loss += loss_p
161	219	if PACNN_MUTILPLE_SCALE_LOSS:	if PACNN_MUTILPLE_SCALE_LOSS:
162	220	pad_s_0 = perspective_s.size()[2] - p_s.size()[2]	pad_s_0 = perspective_s.size()[2] - p_s.size()[2]
163	221	pad_s_1 = perspective_s.size()[3] - p_s.size()[3]	pad_s_1 = perspective_s.size()[3] - p_s.size()[3]
164		p_s_pad = F.pad(perspective_s, (0, pad_s_1, 0, pad_s_0),
165		mode='replicate')
	222		# p_s_pad = F.pad(perspective_s, (0, pad_s_1, 0, pad_s_0),
	223		# mode='replicate')
	224
	225		if pad_s_0 == 0:
	226		pad_s_0 = -perspective_s.size()[2]
	227		if pad_s_1 == 0:
	228		pad_s_1 = -perspective_s.size()[3]
	229
	230		perspective_s_pad = perspective_s[:,:, 0:-pad_s_0, 0:-pad_s_1]
166	231
167		loss_p_s = criterion_mse(p_s_pad, perspective_s) + criterion_ssim(p_s_pad, perspective_s)
	232		loss_p_s = criterion_mse(p_s, perspective_s_pad) + criterion_ssim(p_s, perspective_s_pad)
168	233	loss += loss_p_s	loss += loss_p_s
169	234
170	235	# what is this, loss_d count 2 ?	# what is this, loss_d count 2 ?

File train_script/test_pacnn_shanghaitechA.sh added (mode: 100644) (index 0000000..b372569)
	1	#python main_pacnn.py \
	2	#--input data/ShanghaiTech/part_A \
	3	#--load_model saved_model/train_state2_attemp4_265_checkpoint.pth.tar \
	4	#--PACNN_PERSPECTIVE_AWARE_MODEL 0 \
	5	#--PACNN_MUTILPLE_SCALE_LOSS 0 \
	6	#--test \
	7	#--task_id test
	8
	9	python main_pacnn.py \
	10	--input data/ShanghaiTech/part_A \
	11	--load_model saved_model/train_state1_attemp7_180_checkpoint.pth.tar \
	12	--PACNN_PERSPECTIVE_AWARE_MODEL 0 \
	13	--PACNN_MUTILPLE_SCALE_LOSS 0 \
	14	--test \
	15	--task_id test
	16

File train_script/train_pacnn_shanghaitechA.sh changed (mode: 100644) (index a405a96..ae75a9f)
58	58
59	59	####################	####################
60	60
61		python main_pacnn.py \
62		--input data/ShanghaiTech/part_A \
63		--load_model saved_model/train_state1_attemp7_180_checkpoint.pth.tar \
64		--epochs 500 \
65		--lr 1e-9 \
66		--PACNN_PERSPECTIVE_AWARE_MODEL 1 \
67		--PACNN_MUTILPLE_SCALE_LOSS 1 \
68		--task_id train_state2_attemp1
	61		#python main_pacnn.py \
	62		#--input data/ShanghaiTech/part_A \
	63		#--load_model saved_model/train_state1_attemp7_180_checkpoint.pth.tar \
	64		#--epochs 500 \
	65		#--lr 1e-9 \
	66		#--PACNN_PERSPECTIVE_AWARE_MODEL 1 \
	67		#--PACNN_MUTILPLE_SCALE_LOSS 1 \
	68		#--task_id train_state2_attemp1
	69
	70
	71		#python main_pacnn.py \
	72		#--input data/ShanghaiTech/part_A \
	73		#--load_model saved_model/train_state2_attemp1_185_checkpoint.pth.tar \
	74		#--epochs 500 \
	75		#--lr 3e-10 \
	76		#--momentum 0.7 \
	77		#--PACNN_PERSPECTIVE_AWARE_MODEL 1 \
	78		#--PACNN_MUTILPLE_SCALE_LOSS 1 \
	79		#--task_id train_state2_attemp2
	80
69	81
	82		# momentum
70	83
71	84	#--input data/ShanghaiTech/part_A \	#--input data/ShanghaiTech/part_A \
72	85	#--load_model saved_model/train_state1_attemp7_180_checkpoint.pth.tar	#--load_model saved_model/train_state1_attemp7_180_checkpoint.pth.tar

...	...	python main_pacnn.py \
74	87	#--lr 1e-9	#--lr 1e-9
75	88	#--PACNN_PERSPECTIVE_AWARE_MODEL 1	#--PACNN_PERSPECTIVE_AWARE_MODEL 1
76	89	#--PACNN_MUTILPLE_SCALE_LOSS 1	#--PACNN_MUTILPLE_SCALE_LOSS 1
77		#--task_id dev
	90		#--task_id dev
	91
	92
	93		#python main_pacnn.py \
	94		#--input data/ShanghaiTech/part_A \
	95		#--load_model saved_model/train_state2_attemp2_220_checkpoint.pth.tar \
	96		#--epochs 500 \
	97		#--lr 3e-10 \
	98		#--momentum 0.7 \
	99		#--PACNN_PERSPECTIVE_AWARE_MODEL 1 \
	100		#--PACNN_MUTILPLE_SCALE_LOSS 0 \
	101		#--task_id train_state2_attemp3
	102
	103		#python main_pacnn.py \
	104		#--input data/ShanghaiTech/part_A \
	105		#--load_model saved_model/train_state2_attemp3_240_checkpoint.pth.tar \
	106		#--epochs 600 \
	107		#--lr 1e-9 \
	108		#--momentum 0.9 \
	109		#--PACNN_PERSPECTIVE_AWARE_MODEL 1 \
	110		#--PACNN_MUTILPLE_SCALE_LOSS 0 \
	111		#--task_id train_state2_attemp4
	112
	113		python main_pacnn.py \
	114		--input data/ShanghaiTech/part_A \
	115		--load_model saved_model/train_state2_attemp4_265_checkpoint.pth.tar \
	116		--PACNN_PERSPECTIVE_AWARE_MODEL 1 \
	117		--PACNN_MUTILPLE_SCALE_LOSS 0 \
	118		--test \
	119		--task_id test

Hints:
Before first commit, do not forget to setup your git environment:

git config --global user.name "your_name_here"
git config --global user.email "your@email_here"

Clone this repository using HTTP(S):

git clone https://rocketgit.com/user/hahattpro/crowd_counting_framework

Clone this repository using ssh (do not forget to upload a key first):

git clone ssh://rocketgit@ssh.rocketgit.com/user/hahattpro/crowd_counting_framework

Clone this repository using git:

git clone git://git.rocketgit.com/user/hahattpro/crowd_counting_framework

You are allowed to anonymously push to this repository.
This means that your pushed commits will automatically be transformed into a merge request:

... clone the repository ...
... make some changes and some commits ...
git push origin main

File data_flow.py changed (mode: 100644) (index 6faebf8..a3f3d3c)
...	...	def load_data_shanghaitech_pacnn_with_perspective(img_path, train=True):
138	138	img = img.transpose(Image.FLIP_LEFT_RIGHT)	img = img.transpose(Image.FLIP_LEFT_RIGHT)
139	139	perspective = np.fliplr(perspective)	perspective = np.fliplr(perspective)
140	140
	141		perspective /= np.max(perspective)
	142
141	143	target1 = cv2.resize(target, (int(target.shape[1] / 8), int(target.shape[0] / 8)),	target1 = cv2.resize(target, (int(target.shape[1] / 8), int(target.shape[0] / 8)),
142	144	interpolation=cv2.INTER_CUBIC) * 64	interpolation=cv2.INTER_CUBIC) * 64
143	145	target2 = cv2.resize(target, (int(target.shape[1] / 16), int(target.shape[0] / 16)),	target2 = cv2.resize(target, (int(target.shape[1] / 16), int(target.shape[0] / 16)),

...	...	def load_data_shanghaitech_pacnn_with_perspective(img_path, train=True):
146	148	interpolation=cv2.INTER_CUBIC) * 1024	interpolation=cv2.INTER_CUBIC) * 1024
147	149
148	150	perspective_s = cv2.resize(perspective, (int(perspective.shape[1] / 16), int(perspective.shape[0] / 16)),	perspective_s = cv2.resize(perspective, (int(perspective.shape[1] / 16), int(perspective.shape[0] / 16)),
149		interpolation=cv2.INTER_CUBIC) * 256
	151		interpolation=cv2.INTER_CUBIC)
150	152
151	153	perspective_p = cv2.resize(perspective, (int(perspective.shape[1] / 8), int(perspective.shape[0] / 8)),	perspective_p = cv2.resize(perspective, (int(perspective.shape[1] / 8), int(perspective.shape[0] / 8)),
152		interpolation=cv2.INTER_CUBIC) * 64
	154		interpolation=cv2.INTER_CUBIC)
153	155
154	156	return img, (target1, target2, target3, perspective_s, perspective_p)	return img, (target1, target2, target3, perspective_s, perspective_p)
155	157