ssd+squeezenet

[kaishijeng]

Jay

Do you have a plan to generate a builder for ssd+squeezenet?
I am looking for a low computational complexity of SSD detector and am thinking ssd+squeezenet may be a good compromise between accuracy and speed.

Thanks,

[jay-mahadeokar]

@kaishijeng I believe the complexity of squeezenet in terms of flops is ~800 Million (though not sure, need to run it through complexity module) and the corresponding top1 accuracy on imagenet is ~58%, its advantage is lesser no of params (which affects memory and not speed). In comparison, thin resnet 50 (or resnet_50_1by2) which I trained has ~10k M flops with top1 accuracy of 66.79 on imagenet. See this comparison table. I had run experiment to train resnet_50_1by2 with SSD and got around 64-65% mAP on voc dataset, as compared to 70.4 using full resnet 50 described here. If you want even faster network (and not smaller in size), I suppose using tweaked resnet variants could be useful.
That said, it would be interesting to see how squeezenet can be used as base network for SSD (which layers /feature maps to use etc). There is a quick guide on how it can be done.

[kaishijeng]

Jay,

Thanks for the info about squeezenet vs resnet50. My understanding is
squeezenet is faster than alexnet and also has smaller size of parameters.

Do you have speed comparison between ssd+vgg16 vs ssd+resnet50?
Can you share pretrained models of ssd+resnet_50 or ssd+resnet_50_1by2?
I will try to train ssd+resnet_50 this weekend.

Thanks,

On Mon, Aug 8, 2016 at 11:12 PM, Jay Mahadeokar notifications@github.com
wrote:

@kaishijeng https://github.com/kaishijeng I believe the complexity of
squeezenet in terms of flops is ~800 Million (though not sure, need to run
it through complexity module) and the corresponding top1 accuracy on
imagenet is ~58%, its advantage is lesser no of params (which affects
memory and not speed). In comparison, thin resnet 50 (or resnet_50_1by2)
which I trained has ~10k M flops with top1 accuracy of 66.79 on imagenet.
See this (comparison table)[https://github.com/jay-
mahadeokar/pynetbuilder/tree/master/models/imagenet#basic-
residual-network-results]. I had run experiment to train resnet_50_1by2
with SSD and got around 64% mAP on voc dataset, as compared to 70.4 using
full resnet 50 described here
https://github.com/jay-mahadeokar/pynetbuilder/tree/master/models/voc2007_ssd.
If you want even faster network (and not smaller in size), I suppose using
tweaked resnet variants could be useful.

That said, it would be interesting to see how squeezenet can be used as
base network for SSD (which layers /feature maps to use etc). This is a
quick guide on how it can be done: https://github.com/jay-
mahadeokar/pynetbuilder/tree/master/models/voc2007_ssd#
building-other-detection-networks

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AMGg3lF5HJ3q6Ct0lSrDRbAHPHmXpP7Zks5qeBo2gaJpZM4Jfvjr
.

[jay-mahadeokar]

Please refer this table for ssd+vgg16 and ssd+resnet50. I have also shared the caffemodel.. This table also compares resnet 50 and resnet_50_1by2. Though I havent yet added model files object detection using for resnet_50_1by2 + ssd, it should be easy to train it (since I have added the model pre-trained on imagenet). Let me know if the training ssd+resnet doc is sufficient, or you run into any bugs.

[kaishijeng]

According to your table, ssd+resnet50 shpould be 2 or 3 times faster than
ssd+vgg16.
Is this what you have observed?

Thanks,

On Tue, Aug 9, 2016 at 12:05 AM, Jay Mahadeokar notifications@github.com
wrote:

Please refer this table
https://github.com/jay-mahadeokar/pynetbuilder/tree/master/models/voc2007_ssd#comparing-vgg-and-resnet-50-ssd-based-detection-networks
for ssd+vgg16 and ssd+resnet50. I have also shared the caffemodel.. This
table
https://github.com/jay-mahadeokar/pynetbuilder/tree/master/models/imagenet#basic-residual-network-results
also compares resnet 50 and resnet_50_1by2. Though I havent yet added model
files object detection using for resnet_50_1by2 + ssd, it should be easy to
train it (since I have added the model pre-trained on imagenet). Let me
know if the training ssd+resnet
https://github.com/jay-mahadeokar/pynetbuilder/tree/master/models/voc2007_ssd
doc is sufficient, or you run into any bugs.

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AMGg3qzyLqGRvaVWlLOx1JXRJ2ENOe76ks5qeCaigaJpZM4Jfvjr
.

[jay-mahadeokar]

I haven't done thorough benchmarking on cpu, since I only tested validation set on gpu machines. but I guess that should be true! I will run it on CPU and will update here.

[kaishijeng]

Jay

No need to benchmark on CPU because I have a GPU, TitanX.
What parameters do I need to use with create_ssdnet.py to create ssd+resnet50_1by2 instead of ssd+resnet50?

python app/ssd/create_ssdnet.py --type Resnet -n 256 -b 3 4 6 3 --no-fc_layers -m bottleneck --extra_blocks 3 3 --extra_num_outputs 2048 2048 --mbox_source_layers relu_stage1_block3 relu_stage2_block5 relu_stage3_block2 relu_stage4_block2 relu_stage5_block2 pool_last --extra_layer_attach pool -c 21 -o ./

Thanks,

[jay-mahadeokar]

--extra_num_outputs could be reduced to 1024 1024, and -n to 128. Rest of the params should remain same I think. Use -h for more help on params.

[jay-mahadeokar]

@kaishijeng did the above params work for you? I am closing this for now, feel free to re-open it if you have additional questions.

[kaishijeng]

Jay,

Yes, it works

Thanks

[kaishijeng]

Jay,

I am able to train ssd_resent50 and ssd_resnet50_1by2 and try out inference on TitanX and Jetson TX1.
For TitanX, I can see the speed improvement, but not much difference on Jetson TX1. I think that it is due to memory bandwidth because of parameter size.
If it is not much effort for you to create ssd_squeezenet, I can do the training and measure inference time on TitanX and Jetson TX1.

Thanks,

[jay-mahadeokar]

@kaishijeng
Squeezenet architecture is a lot different than resnet / vgg in terms of feature map sizes. I am not sure which layers would we attach the detection heads.

If you want to try some experiments, id suggest:

• Look at code to generate base squeezenet which is available in this app
• Follow these steps for adding detection heads to a base network.
• Look at this code on how AssembleLego can be used to attach detection heads to base network.
• Main part is figuring out which layers we should attach the SSD detection heads. (example see this table for how I attached it to resnet_50), it will need some experimentation.

Please give it a try and I can help out if you have any further questions.

[kaishijeng]

Jay,

 It looks like not s simple exercise to create a ssd+squeezenet network. So  I like to try ssd+resnet18 first. I need to train resnet18 imagenet first and use it a pretained model for ssd+resnet18 training.. 

I plan to use the following command to create resnet18 for imagenet , but not sure the parameters are correct or not. Can you help me to check it:

python app/imagenet/build_resnet.py -m bottleneck -b 2 2 2 2 -n 256 --no-fc_layers -o ./

Also I got an error to use the following command to generate ssd+resnet18. Do you know which parameters are incorrect?
python app/ssd/create_ssdnet.py --type Resnet -n 256 -b 2 2 2 2 --no-fc_layers -m bottleneck --extra_blocks 3 3 --extra_num_outputs 2048 2048 --mbox_source_layers relu_stage1_block3 relu_stage2_block5 relu_stage3_block2 relu_stage4_block2 relu_stage5_block2 pool_last --extra_layer_attach pool -c 21 -o ./

Thanks,

[jay-mahadeokar]

Sounds good!

You need to modify relu_stage1_block3 relu_stage2_block5 relu_stage3_block2 relu_stage4_block2 relu_stage5_block2 params to relu_stage1_block1 relu_stage2_block1 relu_stage3_block1 relu_stage4_block1 relu_stage5_block1

Also, extra_blocks could be 2 2 (or your choice, more blocks will increase runtime). Notice that resnet 18 has only 2 blocks in each stage (index starts with 0). Read more here

[kaishijeng]

Jay

Shouldn't main_branch of resnet18 of imagenet be normal instead of
bottleneck? If yes, I use the following command to generate resnet18, there
is an error.
python app/imagenet/build_resnet.py -m normal -b 2 2 2 2 -n 256
--no-fc_layers -o ./

The error is:

F0814 01:05:54.747488 14253 eltwise_layer.cpp:34] Check failed:
bottom[i]->shape() == bottom[0]->shape()
*** Check failure stack trace: ***
Aborted (core dumped)

Thanks,

On Sun, Aug 14, 2016 at 12:41 AM, Jay Mahadeokar notifications@github.com
wrote:

Sounds good!

You need to modify relu_stage1_block3 relu_stage2_block5
relu_stage3_block2 relu_stage4_block2 relu_stage5_block2 params to
relu_stage1_block1 relu_stage2_block1 relu_stage3_block1 relu_stage4_block1
relu_stage5_block1

Also, extra_blocks could be 2 2 (or your choice, more blocks will increase
runtime). Notice that resnet 18 has only 2 blocks in each stage (index
starts with 0). Read more here
https://github.com/jay-mahadeokar/pynetbuilder/tree/master/models/imagenet#creating-residual-networks

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AMGg3iqkohWZrbrfXIMAxCpVa8Sh-t8Aks5qfsbBgaJpZM4Jfvjr
.

[jay-mahadeokar]

Please specify -n as 64. Note that the bottleneck block has 3 layers 64,64,256 filters, whereas normal block has 2 layers with 64,64 filters. Since 1st conv layer has 64 filters, it gives error. I should do this check somewhere!

FYI, resnet_18 has:

python app/imagenet/build_resnet.py -m normal -b 2 2 2 2 -n 64 --no-fc_layers -o ./
Number of params:  11.688512  Million
Number of flops:  1814.082944  Million

The flops is larger than resnet_50_1by2. Not sure if it will be faster, but I havent benchmarked.

[poorfriend]

@kaishijeng, can you tell me how many times the speed of ssd+resnet50 is on the ssd+vgg16 using a GPU, TitanX. Thank you

[MisayaZ]

@kaishijeng , hi, I have test the Benchmarking by command line caffe time and found that the forward time of ssd+resnet50 is more than the forward time of ssd+vgg16. I do not how you see the speed improvement?

[kaishijeng]

MisayaZ,

  Your data is correct. This has been a while since I did the test last

time.
My impression is ssd+resnet50 is slower to ssd+vgg16. By ssd+resnet50-1by2
is slightly faster than ssd+vgg16, but lower memory footprint

On Wed, Nov 2, 2016 at 12:13 AM, MisayaZ notifications@github.com wrote:

kaishijeng , hi, I have test the of by the Benchmarking by command line
caffe time and found that the forward time of ssd+resnet50 is more than the
forward time of ssd+vgg16. I do not how you see the speed improvement?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#1 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AMGg3tZ55OuXK42iRlVzDJ2veWiSFCHAks5q6DgIgaJpZM4Jfvjr
.

[mrgloom]

SqeezeNet is not fast (compare to AlexNet), it just have small on disk size.
See table https://github.com/mrgloom/kaggle-dogs-vs-cats-solution

[KevinYuk]

@kaishijeng Hi kaishijeng,

Have you successfully build the resnet18+SSD and get a good mAP?
If so, could you please share your related resnet18+SSD prototxt file and resnet18 pre-train weights?
Thanks a lot.