search

alexgkendall / SegNet-Tutorial

Files for a tutorial to train SegNet for road scenes using the CamVid dataset
http://mi.eng.cam.ac.uk/projects/segnet/tutorial.html
847 stars 517 forks source link

Modifying compute_bn_statistics.py to work with LMDB data #77

Closed ilia-nikiforov closed 7 years ago

ilia-nikiforov commented 7 years ago

As several users have before me, I would like to use SegNet to classify data with an arbitrary number of channels, and using LMDB seems to be the way to go. The compute_bn_statistics.py script needs to be modified in order to use LMDB data. In issues https://github.com/alexgkendall/caffe-segnet/issues/8 https://github.com/alexgkendall/caffe-segnet/issues/32 @mtrth, @cslxiao, and @alexgkendall speak about having modified the script, but I have not been able to find any specifics regarding the modifications. Can anyone help?

ilia-nikiforov commented 7 years ago

The script doesn't actually do anything with the input data itself, except use it to calculate the number of iterations needed to cover the entire test dataset. So the easiest way to make it work is just to I passed the number of iterations required to cover the data sets to the script manually.

kislayabhi commented 7 years ago

@ilia-nikiforov Yeah, very true. If someone else gets here, find the line train_ims, train_labs = extract_dataset(testable_msg). Here the only place train_ims is used is to provide the train_size.

Better do it manually in case of dealing with k-channels. Also, change the minibatch_size initialization to : minibatch_size = testable_msg.layer[0].data_param.batch_size

hftsai commented 7 years ago

Hi guys, I'm trying to do the same. I avoided the assert in the compute_bn_statistics.py for denseimage data and modified according to kislayabhi.

But i got an error 13687 insert_splits.cpp:35] unknown blob input data to layer 0 .

Sorry i'm really new to caffe and segnet. Can you give me a hint other things i should modify to adapt for lmdb data input to segnet (i'm also thinking of changing to hdf5 for 16bit data)

Thank you.

manyaafonso commented 6 years ago

Hi, I am trying to run SegNet with RGBD data stored as an lmdb. I modified compute_bn_statistics.py to avoid the assertion for DenseImageData (in the function extract_dataset), but get the following error. Has anyone managed to get this to run? It will be highly appreciated, if anyone shares their modified compute_bn_statistics.py.

Building BN calc net... Calculate BN stats... Traceback (most recent call last): File "./Scripts/compute_bn_statistics_modified.py", line 181, in train_ims, train_labs = extract_dataset(testable_msg) File "./Scripts/compute_bn_statistics_modified.py", line 23, in extract_dataset with open(source) as f: IOError: [Errno 21] Is a directory: u'/home/mafonso/SegNet/example_lmdb'