Last active
November 27, 2015 02:30
-
-
Save georgy7/82738ca1c98650fa9725 to your computer and use it in GitHub Desktop.
To train https://github.com/karpathy/neuraltalk2 on my own data. Just run from a folder with tagged JPEG images. I mean the keywords in the file properties.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
# pip install exifread | |
import exifread | |
import array | |
import string | |
import glob | |
import json | |
KEYWORDS = 'Image XPKeywords' | |
OUTPUT = 'neuraltalk.json' | |
def filter_str(str): | |
s = ''.join(filter(lambda x: (x in string.printable) and (x not in '\t\n\r'), str)) | |
return s.strip() | |
def tags(filename): | |
try: | |
f = open(filename, 'rb') | |
exif = exifread.process_file(f) | |
tags = array.array('B', exif[KEYWORDS].values).tostring().decode('utf-16').split(';') | |
tags = [filter_str(t) for t in tags] | |
return tags | |
except Exception as e: | |
print(filename, 'has no keywords.') | |
return [] | |
def save(arr): | |
f = open(OUTPUT, 'w') | |
json.dump(arr, f, sort_keys=True, indent=2, separators=(',', ': ')) | |
f.close() | |
print('Finished.') | |
def main(): | |
files = glob.glob('*.[jJ][pP][gG]') + glob.glob('*.[jJ][pP][eE][gG]') | |
result = [] | |
for filename in files: | |
image_tags = tags(filename) | |
if len(image_tags) > 0: | |
result.append({'file_path': filename, 'captions': image_tags}) | |
save(result) | |
if __name__ == "__main__": | |
main() |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
NEUROTALK=../neuraltalk2 | |
JSON_OUTPUT=neuraltalk_output.json | |
H5_OUTPUT=neuraltalk.h5 | |
CHECKPOINTS_OUTPUT=./checkpoints | |
# Making the paths absolute. | |
JSON_OUTPUT=`readlink -f $JSON_OUTPUT` | |
H5_OUTPUT=`readlink -f $H5_OUTPUT` | |
CHECKPOINTS_OUTPUT=`readlink -f $CHECKPOINTS_OUTPUT` | |
START_FOLDER=`readlink -f .` | |
CNN_MODEL=`readlink -f $NEUROTALK/model/VGG_ILSVRC_16_layers.caffemodel` | |
CNN_PROTO=`readlink -f $NEUROTALK/model/VGG_ILSVRC_16_layers_deploy.prototxt` | |
if [ ! -f $H5_OUTPUT ] || [ ! -f $JSON_OUTPUT ]; then | |
rm -f $H5_OUTPUT | |
rm -f $JSON_OUTPUT | |
python $NEUROTALK/prepro.py \ | |
--input_json neuraltalk.json \ | |
--num_val 5000 --num_test 5000 \ | |
--images_root . \ | |
--word_count_threshold 5 \ | |
--output_json $JSON_OUTPUT \ | |
--output_h5 $H5_OUTPUT | |
fi | |
mkdir -p $CHECKPOINTS_OUTPUT | |
cd $NEUROTALK | |
th train.lua -input_h5 $H5_OUTPUT -input_json $JSON_OUTPUT -checkpoint_path $CHECKPOINTS_OUTPUT \ | |
-gpuid -1 \ | |
-cnn_model $CNN_MODEL \ | |
-cnn_proto $CNN_PROTO | |
cd $START_FOLDER |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment