I want to know how to train the model

[LiuDong777]

hi I want to reproduce your experiment，in the code to generate .h5 file POOLFEAT_FILEPATTERN = 'data/coco2014/coco2014{0}_vgg_fc7.txt' SENTS_FILE_PATTERN = 'data/coco2014/sents/cocosentences{0}_tokens.txt' LABEL_FILE_PATTERN = 'data/coco2014/sents/labelsglove72k{0}.txt' #train2014

IMAGEID_FILE_PATTERN = 'data/coco2014/coco_rm8objs_imagelist{0}.txt'

IMAGEID_FILE_PATTERN = 'data/coco2014/cvpr17_rm8newobjs/coco_rm8newobjs_imagelist{0}.txt' i want to know how can i get these handled documents like coco2014_{0}_vgg_fc7.txt? I am a beginner，I am so sorry to take up your time. Can you tell me how to reproduce your experiment? thank you!

[vsubhashini]

Sorry about that, I should probably include more comprehensive instructions.

I have provided the links to download the raw training data: https://github.com/LisaAnne/DCC This imageid file pattern is just a list of image IDs for the train, test, and validation splits.

The file names in the scripts (within data_utils) should point to the train, validation, and test splits of the data. I have also included the script that extracts labels from the captions.

vgg_fc7: should contain the image features (from VGG net). Can be extracted using: https://github.com/vsubhashini/noc/blob/recurrent/examples/noc/extract_vgg_features.py -- Just create a file with a list of images and this script will help you extract VGG features.

Hope this helps.

On Jan 9, 2018 2:52 AM, "dong798" notifications@github.com wrote:

hi I want to reproduce your experiment，in the code to generate .h5 file POOLFEAT_FILEPATTERN = 'data/coco2014/coco2014{0} vgg_fc7.txt' SENTS_FILE_PATTERN = 'data/coco2014/sents/coco_sentences{0} tokens.txt' LABEL_FILE_PATTERN = 'data/coco2014/sents/labels_glove72k{0}.txt'

train2014

IMAGEID_FILE_PATTERN = 'data/coco2014/coco_rm8objs_imagelist{0}.txt'

IMAGEID_FILE_PATTERN = 'data/coco2014/cvpr17_rm8newobjs/cocorm8newobjs imagelist{0}.txt' i want to know how can i get these handled documents like coco2014_{0}_vgg_fc7.txt? I am a beginner，I am so sorry to take up your time. Can you tell me how to reproduce your experiment? thank you!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/vsubhashini/noc/issues/6, or mute the thread https://github.com/notifications/unsubscribe-auth/AAWOch4Nl7uwYt2ksnE23K6j-83hH_cmks5tIyiwgaJpZM4RXftp .

[LiuDong777]

What your means is that POOLFEAT_FILEPATTERN = 'data / coco2014 / coco2014 {0} _vgg_fc7.txt' is the feature of the training data(train2014), and IMAGEID_FILE_PATTERN is the list of data right? but how about the LABEL_FILE_PATTERN?

[vsubhashini]

Yes. The coco ID lists are actually all here: https://github.com/ vsubhashini/noc/blob/recurrent/examples/noc/data_utils/image_list/ https://github.com/vsubhashini/noc/blob/recurrent/examples/noc/data_utils/image_list/coco2014_cocoid.train.txt

The label list is the list of labels corresponding to the images. For Imagenet it would just be a single label for each image, for COCO it's extracted from the captions. But you make a good point looks like I have not included the code that extracts the labels or the labels list. Will add them. Thanks!

On Jan 12, 2018 2:45 AM, "dong798" notifications@github.com wrote:

What your means is that POOLFEAT_FILEPATTERN = 'data / coco2014 / coco2014 {0} _vgg_fc7.txt' is the feature of the training data(train2014), and IMAGEID_FILE_PATTERN is the list of data right? but how about the LABEL_FILE_PATTERN?

— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/vsubhashini/noc/issues/6#issuecomment-357177669, or mute the thread https://github.com/notifications/unsubscribe-auth/AAWOcknxbA1CYvICdc0zQHjNadEhz2Ogks5tJxvDgaJpZM4RXftp .

[Sally-fyk]

Where can I find the code that extracts the labels or the labels list ?