Restoring model from checkpoint

[dianamurgulet]

Hi @HadoopIt ,

Thank you for publishing the code for the paper.
I am trying to use a stored pre-trained model to generate the intent and slots for a new sentence. However, based on the outputs it generates, it ends up using a new, untrained model.

saver = tf.train.import_meta_graph('/tmp/model.ckpt-1900.meta')
saver.restore(session, '/tmp/model.ckpt-1900')

model_train, model_test = create_model(session, 139, 36, 6)
step_outputs  = model_test.joint_step(session, encoder_inputs, tags, tag_weights, labels,sequence_length, bucket_id, True)

Any suggestions on how to use a trained model from a stored file?

[wqdong]

The train and test model share parameters, since you didn't load the train model, the test model would be randomly initialized, also you should create the models before restore.

[HadoopIt]

Hi @dianamurgulet, the "create_mode" function finds the model checkpoint from FLAGS.train_dir with ckpt = tf.train.get_checkpoint_state(FLAGS.train_dir). If you have a pre-trained model, just pass the model directory to this get_checkpoint_state function should make it work.

[programmer9208]

Hi @HadoopIt , I tried to pass the model directory to "create_model" func at your suggestion, but it still didn't work. Then I read your code for restoring pre-trained model from disk in run_multi-task_rnn.py, and made some change:

original code in run_multi-task_rnn.py, line 219:
if ckpt and tf.gfile.Exists(ckpt.model_checkpoint_path):
I simply delete the second condition:
if ckpt:

Then it works. I guess the reason may be ckpt.model_checkpoint_path = './model/model.ckpt-29900' and it doesn't exist. The files we need are with suffix '.data-00000-of-00001' , 'index', 'meta', which are exist in $train_dir.

[bringtree]

@pachirayz tf.gfile.Exists(ckpt.model_checkpoint_path) always false. maybe you should edit it

[bringtree]

ckpt.model_checkpoint_path
Out[2]: u'model_tmp/model.ckpt-15000'
sss = '/Users/huangpeisong/Desktop/py2 2/rnn-nlu/model_tmp/model.ckpt-15000'
tf.gfile.Exists(sss)
Out[4]: False
sss = '/Users/huangpeisong/Desktop/py2 2/rnn-nlu/model_tmp/model.ckpt-15000.index'
tf.gfile.Exists(sss)
Out[6]: True

[HadoopIt]

@bringtree Thanks for pointing this out. I have just pushed a fix.

[programmer9208]

@bringtree Yeah, that's what I mean. The second condition (tf.gfile.Exists(ckpt.model_checkpoint_path)) is always false because ckpt.model_checkpoint_path does not contains suffix , e.g 'index' as your example.
Thank you all ! @HadoopIt @bringtree