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)
my simple v4 with addition deform cnn at the end	5392aaf6c14fdd910f52096dbb921bed7470c4f7	Thai Thien	2020-03-07 18:15:22
fix the scheduler	77e6737a040f5aa5745b8a8830f5bec12322b10f	Thai Thien	2020-03-07 17:46:02
t4 lr 1e-5	acd41ed30c95f63e01a05a6d9929410637852d9e	Thai Thien	2020-03-06 19:41:49
no more lr scheduler	7289adb41de7807258eb8c29e6108fa65f59525a	Thai Thien	2020-03-06 19:35:49
reduce learning rate	bc8241e5b88b91c18bb7999a8d5d12fc79a5e3f7	Thai Thien	2020-03-06 19:28:27
dilated ccnn	5c5d92bdc0a288dd5d4ec5f1367d8cb928175bbe	Thai Thien	2020-03-06 19:04:01
done	9f05e093ec7c10284a4aedf0738f9e61d5ac6fb6	Thai Thien	2020-03-06 18:02:34
with lr scheduler	466c364b60ed22c77319b14ccc9a201614b908bf	Thai Thien	2020-03-04 17:57:49
train with learning rate scheduler	fcd5a3c8da2dd6763e0d40742edf47b49c95fcfb	Thai Thien	2020-03-04 17:55:11
ccnn no padding at output layer	57563fc07f656c63f807de4d80712ff11345109d	Thai Thien	2020-03-04 17:36:43
fix dimension of ccnn	f4439d9a78273ab3ba450f31a528509816b4352f	Thai Thien	2020-03-04 17:32:48
ready to train	dbe0d6c3271dbb22490f0877fa31ba9cd7852b99	Thai Thien	2020-03-04 15:55:05
done implement c-cnn	2deecef953baf1e07ce5cf5477d208bc7ffa34cf	Thai Thien	2020-03-03 17:25:03
fix script	0e9d372b9ad60b32939f1e558b2a59fc7d518fa2	Thai Thien	2020-03-02 16:23:55
simple v3 to 91 epoch	539fdd03c3e3497fd22b7db2aaa14f067cbf6f8d	Thai Thien	2020-03-02 16:09:43
we train on all training data and validate on test data	9407ef8d5b7c47c53d6f98dcb3c20208aad1d7a9	Thai Thien	2020-03-01 15:36:46
load and continue train v3	12421fb7330e5c9d2eed4f6e574dfe69bdfddefc	Thai Thien	2020-03-01 14:50:01
add env file	b1ed02088b01af42efc8d6963b3699e0a5c31c01	Thai Thien	2020-03-01 11:10:56
sanity check dataloader	034daa8bef69daff92891cc42b988a6c77b010f9	Thai Thien	2020-03-01 10:24:38
print train loader len	eef3995f63a631e0ec5d92e31f5d7db27fd04401	Thai Thien	2020-03-01 05:17:55

Commit 5392aaf6c14fdd910f52096dbb921bed7470c4f7 - my simple v4 with addition deform cnn at the end
Author: Thai Thien
Author date (UTC): 2020-03-07 18:15
Committer name: Thai Thien
Committer date (UTC): 2020-03-07 18:15
Parent(s): 77e6737a040f5aa5745b8a8830f5bec12322b10f
Signing key:
Tree: ad8e31ebda24e28d02998e77d2a127a55768d127

File	Lines added	Lines deleted
models/__init__.py	1	1
models/attn_can_adcrowdnet_simple.py	132	0
train_attn_can_adcrowdnet_simple.py	2	2
train_script/my_simple1/my_simple_v4_t1.sh	2	2

File models/__init__.py changed (mode: 100644) (index f1cf370..e02e0ec)
...	...	from .context_aware_network import CANNet
4	4	from .deform_conv_v2 import DeformConv2d	from .deform_conv_v2 import DeformConv2d
5	5	from .attn_can_adcrowdnet import AttnCanAdcrowdNet	from .attn_can_adcrowdnet import AttnCanAdcrowdNet
6	6	from .attn_can_adcrowdnet_freeze_vgg import AttnCanAdcrowdNetFreezeVgg	from .attn_can_adcrowdnet_freeze_vgg import AttnCanAdcrowdNetFreezeVgg
7		from .attn_can_adcrowdnet_simple import AttnCanAdcrowdNetSimpleV1, AttnCanAdcrowdNetSimpleV2, AttnCanAdcrowdNetSimpleV3
	7		from .attn_can_adcrowdnet_simple import AttnCanAdcrowdNetSimpleV1, AttnCanAdcrowdNetSimpleV2, AttnCanAdcrowdNetSimpleV3, AttnCanAdcrowdNetSimpleV4
8	8	from .compact_cnn import CompactCNN, CompactDilatedCNN	from .compact_cnn import CompactCNN, CompactDilatedCNN

File models/attn_can_adcrowdnet_simple.py changed (mode: 100644) (index 135c6ca..85b6f95)
...	...	class AttnCanAdcrowdNetSimpleV3(nn.Module):
392	392	nn.init.constant_(m.bias, 0)	nn.init.constant_(m.bias, 0)
393	393
394	394
	395		class AttnCanAdcrowdNetSimpleV4(nn.Module):
	396		"""
	397		compare with v3: add 1 layer (1 branch) of deformable cnn before output layer
	398		"""
	399		def __init__(self, load_weights=False):
	400		super(AttnCanAdcrowdNetSimpleV4, self).__init__()
	401		self.frontend_feat = [64, 64, 'M', 128, 128, 'M', 256, 256, 256, 'M', 512, 512, 512]
	402		self.frontend = make_layers(self.frontend_feat)
	403
	404		# freeze vgg layer
	405		for param in self.frontend.parameters():
	406		param.requires_grad = False
	407
	408		self.sSE = SpatialSELayer(num_channels=512)
	409
	410		self.concat_filter_layer = nn.Conv2d(1024, 512, kernel_size=3, padding=2, dilation=2)
	411
	412		# we skip one formation of deformconv
	413		# self.deform_conv_1_3 = DeformConv2d(512, 256, kernel_size=3, stride=1, padding=1)
	414		# self.deform_conv_1_5 = DeformConv2d(512, 256, kernel_size=5, stride=1, padding=2)
	415		# self.deform_conv_1_7 = DeformConv2d(512, 256, kernel_size=7, stride=1, padding=3)
	416		self.concat_filter_layer_1 = nn.Conv2d(512, 256, kernel_size=3, padding=2, dilation=2)
	417
	418		self.dilated_conv_2_3 = nn.Conv2d(256, 128, kernel_size=3, stride=1, dilation=2, padding=2)
	419		self.dilated_conv_2_5 = nn.Conv2d(256, 128, kernel_size=3, stride=1, dilation=4, padding=4)
	420		# self.deform_conv_2_7 = DeformConv2d(256, 128, kernel_size=7, stride=1, padding=3)
	421		self.concat_filter_layer_2 = nn.Conv2d(128 * 2, 128, kernel_size=3, padding=2, dilation=2)
	422
	423		self.deform_conv_3_3 = DeformConv2d(128, 64, kernel_size=3, stride=1, padding=1)
	424		self.deform_conv_3_5 = DeformConv2d(128, 64, kernel_size=5, stride=1, padding=2)
	425		# self.deform_conv_3_7 = DeformConv2d(128, 64, kernel_size=7, stride=1, padding=3)
	426		self.concat_filter_layer_3 = nn.Conv2d(64 * 2, 64, kernel_size=3, padding=2, dilation=2)
	427
	428		self.deform_conv_4_3 = DeformConv2d(64, 32, kernel_size=3, stride=1, padding=1)
	429
	430		self.output_layer = nn.Conv2d(32, 1, kernel_size=1)
	431		self.conv1_1 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	432		self.conv1_2 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	433		self.conv2_1 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	434		self.conv2_2 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	435		self.conv3_1 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	436		self.conv3_2 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	437		self.conv6_1 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	438		self.conv6_2 = nn.Conv2d(512, 512, kernel_size=1, bias=False)
	439		if not load_weights:
	440		mod = models.vgg16(pretrained=True)
	441		self._initialize_weights()
	442		fsd = collections.OrderedDict()
	443		for i in range(len(self.frontend.state_dict().items())):
	444		temp_key = list(self.frontend.state_dict().items())[i][0]
	445		fsd[temp_key] = list(mod.state_dict().items())[i][1]
	446		self.frontend.load_state_dict(fsd)
	447
	448		def forward(self, x):
	449		fv = self.frontend(x)
	450
	451		# spatial squeeze & excitation
	452		fv = self.sSE(fv)
	453
	454		# S=1
	455		ave1 = nn.functional.adaptive_avg_pool2d(fv, (1, 1))
	456		ave1 = self.conv1_1(ave1)
	457		s1 = nn.functional.upsample(ave1, size=(fv.shape[2], fv.shape[3]), mode='bilinear')
	458		c1 = s1 - fv
	459		w1 = self.conv1_2(c1)
	460		w1 = nn.functional.sigmoid(w1)
	461		# S=2
	462		ave2 = nn.functional.adaptive_avg_pool2d(fv, (2, 2))
	463		ave2 = self.conv2_1(ave2)
	464		s2 = nn.functional.upsample(ave2, size=(fv.shape[2], fv.shape[3]), mode='bilinear')
	465		c2 = s2 - fv
	466		w2 = self.conv2_2(c2)
	467		w2 = nn.functional.sigmoid(w2)
	468		# S=3
	469		ave3 = nn.functional.adaptive_avg_pool2d(fv, (3, 3))
	470		ave3 = self.conv3_1(ave3)
	471		s3 = nn.functional.upsample(ave3, size=(fv.shape[2], fv.shape[3]), mode='bilinear')
	472		c3 = s3 - fv
	473		w3 = self.conv3_2(c3)
	474		w3 = nn.functional.sigmoid(w3)
	475		# S=6
	476		ave6 = nn.functional.adaptive_avg_pool2d(fv, (6, 6))
	477		ave6 = self.conv6_1(ave6)
	478		s6 = nn.functional.upsample(ave6, size=(fv.shape[2], fv.shape[3]), mode='bilinear')
	479		c6 = s6 - fv
	480		w6 = self.conv6_2(c6)
	481		w6 = nn.functional.sigmoid(w6)
	482
	483		fi = (w1 * s1 + w2 * s2 + w3 * s3 + w6 * s6) / (w1 + w2 + w3 + w6 + 0.000000000001)
	484		x = torch.cat((fv, fi), 1)
	485		x = F.relu(self.concat_filter_layer(x), inplace=True)
	486
	487		# x3 = self.deform_conv_1_3(x)
	488		# x5 = self.deform_conv_1_5(x)
	489		# x7 = self.deform_conv_1_7(x)
	490		# x = torch.cat((x3, x5, x7), 1)
	491		# x = torch.cat((x3, x5), 1)
	492		x = F.relu(self.concat_filter_layer_1(x), inplace=True)
	493
	494		x3 = self.dilated_conv_2_3(x)
	495		x5 = self.dilated_conv_2_5(x)
	496		# x7 = self.deform_conv_2_7(x)
	497		# x = torch.cat((x3, x5, x7), 1)
	498		x = F.relu(torch.cat((x3, x5), 1), inplace=True)
	499		x = F.relu(self.concat_filter_layer_2(x), inplace=True)
	500
	501		x3 = self.deform_conv_3_3(x)
	502		x5 = self.deform_conv_3_5(x)
	503		# x7 = self.deform_conv_3_7(x)
	504		# x = torch.cat((x3, x5, x7), 1)
	505		x = F.relu(torch.cat((x3, x5), 1), inplace=True)
	506		x = F.relu(self.concat_filter_layer_3(x), inplace=True)
	507
	508		x = F.relu(self.deform_conv_4_3(x))
	509		x = self.output_layer(x)
	510
	511		# this cause too much dimension mismatch problem
	512		# so we desampling label instead
	513		# x = nn.functional.upsample(x, scale_factor=8, mode='bilinear') / 64.0
	514		return x
	515
	516		def _initialize_weights(self):
	517		for m in self.modules():
	518		if isinstance(m, nn.Conv2d):
	519		nn.init.normal_(m.weight, std=0.01)
	520		if m.bias is not None:
	521		nn.init.constant_(m.bias, 0)
	522		elif isinstance(m, nn.BatchNorm2d):
	523		nn.init.constant_(m.weight, 1)
	524		nn.init.constant_(m.bias, 0)
	525
	526
395	527	def make_layers(cfg, in_channels=3, batch_norm=False, dilation=False):	def make_layers(cfg, in_channels=3, batch_norm=False, dilation=False):
396	528	if dilation:	if dilation:
397	529	d_rate = 2	d_rate = 2

File train_attn_can_adcrowdnet_simple.py changed (mode: 100644) (index 4096800..31334b0)
...	...	from visualize_util import get_readable_time
9	9
10	10	import torch	import torch
11	11	from torch import nn	from torch import nn
12		from models import AttnCanAdcrowdNetSimpleV3
	12		from models import AttnCanAdcrowdNetSimpleV4
13	13	import os	import os
14	14
15	15

...	...	if __name__ == "__main__":
40	40	print("len train_loader ", len(train_loader))	print("len train_loader ", len(train_loader))
41	41
42	42	# model	# model
43		model = AttnCanAdcrowdNetSimpleV3()
	43		model = AttnCanAdcrowdNetSimpleV4()
44	44	model = model.to(device)	model = model.to(device)
45	45
46	46	# loss function	# loss function

File train_script/my_simple1/my_simple_v4_t1.sh copied from file train_script/my_simple1/my_simplev3_noload_t1.sh (similarity 75%) (mode: 100644) (index fe6e65a..423cead)
1	1	CUDA_VISIBLE_DEVICES=4 nohup python train_attn_can_adcrowdnet_simple.py \	CUDA_VISIBLE_DEVICES=4 nohup python train_attn_can_adcrowdnet_simple.py \
2		--task_id simple_v3_t1 \
	2		--task_id simple_v4_t1 \
3	3	--input /data/rnd/thient/thient_data/ShanghaiTech/part_A \	--input /data/rnd/thient/thient_data/ShanghaiTech/part_A \
4	4	--lr 1e-5 \	--lr 1e-5 \
5	5	--decay 5e-4 \	--decay 5e-4 \
6	6	--datasetname shanghaitech_keepfull \	--datasetname shanghaitech_keepfull \
7		--epochs 22 > logs/simple_v3_t1.log &
	7		--epochs 32 > logs/simple_v4_t1.log &

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