ckpt file for demo.py after training about tffrcnn

[ChrisDal]

In fact, the files correspond to :

• VGGnet_fast_rcnn_iter_70000.ckpt.data-00000-of-00001 : data itself (weigths and all)
• VGGnet_fast_rcnn_iter_70000.ckpt.index : (I actually have no idea haha)
• VGGnet_fast_rcnn_iter_70000.ckpt.meta : the graph definition and all metadata associated

Tensorflow can restore the model from a trio of model.ckpt.index , model.ckpt.meta, model.ckpt.data with for example here args.model = model.ckpt :

saver = tf.train.Saver()
saver.restore(sess, args.model)

But to run the demo from this training, you have to change the condition on the existence of the model which is either on the model.ckpt file (that may not exist), either on the model.ckpt.xxxx file(s) (where you can't restore the session). Theses lines test if the user gave a model.ckpt.xxxx file and then give to the saver the right model.ckpt.

if not os.path.splitext(args.model)[1] == 'ckpt':
    model_path = args.model
    args.model = os.path.splitext(args.model)[0] 

   
if args.model == ' ' or not os.path.exists(model_path):
    print ('current path is ' + os.path.abspath(__file__))
    raise IOError(('Error: Model not found.\n'))

You can add it/replace in the demo.py , just after the parsing of the args. Then give the path to an existing model file e.g :

python ./faster_rcnn/demo.py --model dir_model_path/model.ckpt.meta

A link to an article which explain how the data that is saved [1] . That's just a little solution, I am not a python expert neither a tensorflow expert.

[rgsachin]

Thanks @ChrisDal , I believe the trio files format introduced as part of the new saver of Tensorflow - tf.train.SaverDef.V2 which is defaulted now.
Ideal below is recommended way to read the new saver files, below code in demo.py should fix, but
saver = tf.train.import_meta_graph(args.model+'.meta') saver.restore(sess, args.model)
But I ran into some other issues. what seem to solve it forcing the saver to V1 format in train.py
self.saver = tf.train.Saver(max_to_keep=100,write_version=tf.train.SaverDef.V1)

[zyclarkcheng]

thanks @rgsachin.
hello, did you mean modify self.saver = tf.train.Saver(max_to_keep=100) in train.py to self.saver = tf.train.Saver(max_to_keep=100,write_version=tf.train.SaverDef.V1) ?
In this way how should I restore the result file?(I am now training the kitti and waiting for the result file )

[zyclarkcheng]

Thanks for @CharlesShang your work.
I have trained KITTI dataset and get ckpt.meta, ckpt.index files, and I switch to CLASSES = ('background','person','bike','motorbike','car','bus') in demo.py, and import the meta graph
saver = tf.train.import_meta_graph(model) .
BUT I still can not test demo images, so I change self.saver = tf.train.Saver(max_to_keep=100) in train.py to self.saver = tf.train.Saver(max_to_keep=100,write_version=tf.train.SaverDef.V1) based on @rgsachin , after training again yielding real ckpt file, I fail to run demo again...
Could you give me some suggestions ? thanks

[chrenlily]

@zyclarkcheng Did you solve the problem？could you help me ?I got the same problem