How to prepare training data for retraining?

[xxlxx1]

I use the icdar data , i have changed it into : 0 0.5203125 0.21458333333333332 0.5325 0.135 0 PROPER 0 0.5234375 0.5104166666666666 0.535 0.205 0 FOOD 0 0.521875 0.7947916666666667 0.53 0.13125 0 PRONTO But how should i do next? I am new to caffe, i saw mnist demo should change data to imdb format. I don't know how to prepare data for this project.

And i have another question: in tiny.proto , data layer named "OnDiskData", but i can't find this layer in project file. I think it is strange.

Thanks a lot.

[Gitchenguang]

You can see this issue https://github.com/MichalBusta/DeepTextSpotter/issues/10

[mattroos]

ondisk_data_layer.cpp is in the caffe code. In caffe/src/caffe/layers/.

Make a text file with the list of image filenames (png or jpg), with either absolute pathnames or names relative to the directory that holds that text file. In the tiny.prototext, set the source as your text file of image filenames. data_param { source: "/path/to/your/list/list_icdar2015.txt" }

[ghost]

@xxlxx1 How did you change your icdar data into that???

[xxlxx1]

@a41888936 https://github.com/MichalBusta/dataset_conversions in dup_boxes_icdar17.py

[xxlxx1]

@linchenguang @mattroos Thanks a lot. Now i can train, but there is still some question.

1. What is the role of cmp_trie in validation? I am not vary clear to icdar competition
2. There is too many "Train net output #484763: trans" when trainning, why? I just train for 20 itert but there is hundreds of thousands of "Train net output #484763: trans".

[MichalBusta]

1. What is the role of cmp_trie in validation? I am not vary clear to icdar competition
   • none, it is just for dictionary decoding
2. There is too many "Train net output #484763: trans" when trainning, why? I just train for 20 itert but there is hundreds of thousands of "Train net output #484763: trans". default caffe stuff: it is expected that leaf blobs have just one output (loss value)
   • to suppress it, you can change log level

[mattroos]

In python:

os.environ['GLOG_minloglevel'] = '2'
import caffe

You have to set it before importing caffe.

The levels are: 0 - debug 1 - info (still a LOT of outputs) 2 - warnings 3 - errors https://stackoverflow.com/questions/29788075/setting-glog-minloglevel-1-to-prevent-output-in-shell-from-caffe