msypetkowski/snr-proj

Bird recognition using deep neural networks

SNR project

Dependencies

Packages for Arch linux:

tensorflow-opt-cuda opencv python-scikit-learn

Dataset

Data should be like this (with default path):

user ~/snrdata $ ls
bounding_boxes.txt  classes.txt  image_class_labels.txt  SET_C
user ~/snrdata $ cat `find . -type f | sort` | md5sum
2d7256832dffd6d06043fc742661e860  -

Preparing for fine-tuning pretrained VGG16 model

Downloading checkpoint:

wget http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
tar -xvf vgg_16_2016_08_28.tar.gz

Examples

Train 8 layer convolutional network (model is defined in config.py - MyConvConfig)

./train.py -n Model8Conv -t MyConv

Preview results on testset for previously trained VGG16Pretrained model (latest checkpoint)

./data_preview.py --test --eval -n Model5 -t VGG16Pretrained

Train SVM after last hidden layer:

./trainSVM.py -n Model8Conv -l "Model/batch_normalization_8/batchnorm/add_1:0"

Train SVM after softmax:

./trainSVM.py -n Model8Conv -l "Model/Softmax:0"