Facenet: Transfer learning with the pretrained model.

[hector246288]

Hi there, I just tried to load the pretrained model (20170512-110547) as the initial state to perform the transfer learning with my own database, which has 367 identities. What I did is following the instructions you've shown on the page Classifier training of Inception-ResNet-v1. Here was my command:

$ python src/train_softmax.py --logs_base_dir logs/ --models_base_dir models/ --data_dir ~/Database/my_own_database/ --image_size 160 --model_def models.inception_resnet_v1 --pretrained_model models/20170512-110547/model-20170512-110547.ckpt-250000 --optimizer RMSPROP --learning_rate -1 --max_nrof_epochs 80 --keep_probability 0.8 --random_crop --random_flip --learning_rate_schedule_file data/learning_rate_schedule_classifier_casia.txt --weight_decay 5e-5 --center_loss_factor 1e-2 --center_loss_alfa 0.9

However, my process met a size dismatch problem. The error message was:

`Traceback (most recent call last):
File "src/train_softmax.py", line 447, in
main(parse_arguments(sys.argv[1:]))
File "src/train_softmax.py", line 206, in main
saver.restore(sess, pretrained_model)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 1457, in restore
{self.saver_def.filename_tensor_name: save_path})
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 778, in run
run_metadata_ptr)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 982, in _run
feed_dict_string, options, run_metadata)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 1032, in _do_run
target_list, options, run_metadata)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/client/session.py", line 1052, in _do_call
raise type(e)(node_def, op, message)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Assign requires shapes of both tensors to match. lhs shape= [128,367] rhs shape= [128,44052]
[[Node: save/Assign_491 = Assign[T=DT_FLOAT, _class=["loc:@Logits/weights"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](Logits/weights, save/RestoreV2_491/_77)]]

Caused by op u'save/Assign_491', defined at:
File "src/train_softmax.py", line 447, in
main(parse_arguments(sys.argv[1:]))
File "src/train_softmax.py", line 188, in main
saver = tf.train.Saver(tf.trainable_variables(), max_to_keep=3)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 1056, in __init__
self.build()
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 1086, in build
restore_sequentially=self._restore_sequentially)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 691, in build
restore_sequentially, reshape)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 419, in _AddRestoreOps
assign_ops.append(saveable.restore(tensors, shapes))
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/training/saver.py", line 155, in restore
self.op.get_shape().is_fully_defined())
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/state_ops.py", line 270, in assign
validate_shape=validate_shape)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_state_ops.py", line 47, in assign
use_locking=use_locking, name=name)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/op_def_library.py", line 768, in apply_op
op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 2336, in create_op
original_op=self._default_original_op, op_def=op_def)
File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.py", line 1228, in __init__
self._traceback = _extract_stack()

InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [128,367] rhs shape= [128,44052]
[[Node: save/Assign_491 = Assign[T=DT_FLOAT, _class=["loc:@Logits/weights"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](Logits/weights, save/RestoreV2_491/_77)]]
`

I have check the input image type(.png) and size(182x182), which I believe were not the major reason causing the error. Is it possible for you to offer the training command you used to train the model (20170512-110547). Or should I change any parameters of fully-connected layers in the model?

[AKSHAYUBHAT]

Typically with 367 identities you don't need transfer learning, you can simply extract embedding and train a simple Multi-class linear SVM on top.

[knvpk]

Hi @AKSHAYUBHAT, you said for 367 identities don't need a transfer learning and use SVM classifier for it. can you please tell me up to which number we can use SVM I'm developing an application where sometimes only 500 faces are present and sometimes there is more than 5000 faces.

Did you mean for 500 identities SVM is better than the inception training as classifier.

[knvpk]

@hector246288, have you solved the problem.

[fanyuzeng]

I have the same problem. who solve this?

[hector246288]

Actually....I have found an easy way to reach the goal.
Assume that you want the layers in set A to inherit the variables from the pre-trained model.
The remaining set B contains the layers which are newly defined or with different input/output size.
Here I just use FaceNet (Incep-Res-v1) as an example:

set_A_vars = [v for v in tf.trainable_variables() if v.name.startswith('InceptionResnetV1')]
saver_set_A = tf.train.Saver(set_A_vars, max_to_keep=3)
saver_set_A_and_B = tf.train.Saver(tf.trainable_variables(), max_to_keep=3)

In my approach, I firstly do the global initialization as below:
sess.run(tf.global_variables_initializer())
sess.run(tf.local_variables_initializer())

Then I do the transfer learning by restoring the variables in set A:
saver_set_A.restore(sess, pretrained_model)

And please don't forget to save your trained model completely:
saver_set_A_and_B.save(sess, checkpoint_path[, ])

If you have any more efficient way for transfer learning, please let me know. =)

[davidsandberg]

I think @hector246288's way of doing it is the correct one. Thanks!

[lvmeng8]

thinks @hector246288