Can't restore the given checkpoint for inception v3 in research/slim

[jlamypoirier]

I am trying to use the given pretrained inception v3 model in research/slim, but the weights don't match

import tensorflow as tf
from nets.inception_v3 import inception_v3_base
tf_images = tf.placeholder(tf.float32, [None, 299, 299, 3])
inception,inception_layers=inception_v3_base(tf_images)
saver = tf.train.Saver()
with tf.Session() as sess:
    saver.restore(sess,'models/weights/inception_v3.ckpt')

NotFoundError: Tensor name "InceptionV3/Mixed_5c/Branch_2/Conv2d_0a_1x1/biases" not found in checkpoint files models/weights/inception_v3.ckpt
	 [[Node: save/RestoreV2_30 = RestoreV2[dtypes=[DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"](_arg_save/Const_0_0, save/RestoreV2_30/tensor_names, save/RestoreV2_30/shape_and_slices)]]
	 [[Node: save/RestoreV2_184/_5 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_384_save/RestoreV2_184", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]

Looking at the variables they indeed don't match

from tensorflow.python import pywrap_tensorflow
reader = pywrap_tensorflow.NewCheckpointReader('models/weights/inception_v3.ckpt')
var_to_shape_map = reader.get_variable_to_shape_map()
print(sorted(var_to_shape_map))
print([var.name for var in saver._var_list])

For the checkpoint a subsample is:

'InceptionV3/Conv2d_1a_3x3/BatchNorm/beta',
 'InceptionV3/Conv2d_1a_3x3/BatchNorm/moving_mean',
 'InceptionV3/Conv2d_1a_3x3/BatchNorm/moving_variance',
 'InceptionV3/Conv2d_1a_3x3/weights',
 'InceptionV3/Conv2d_2a_3x3/BatchNorm/beta',
 'InceptionV3/Conv2d_2a_3x3/BatchNorm/moving_mean',
 'InceptionV3/Conv2d_2a_3x3/BatchNorm/moving_variance',
 'InceptionV3/Conv2d_2a_3x3/weights',
 'InceptionV3/Conv2d_2b_3x3/BatchNorm/beta',
 'InceptionV3/Conv2d_2b_3x3/BatchNorm/moving_mean',
 'InceptionV3/Conv2d_2b_3x3/BatchNorm/moving_variance',
 'InceptionV3/Conv2d_2b_3x3/weights',
 'InceptionV3/Conv2d_3b_1x1/BatchNorm/beta',
 'InceptionV3/Conv2d_3b_1x1/BatchNorm/moving_mean',
 'InceptionV3/Conv2d_3b_1x1/BatchNorm/moving_variance',
 'InceptionV3/Conv2d_3b_1x1/weights',
 'InceptionV3/Conv2d_4a_3x3/BatchNorm/beta',
 'InceptionV3/Conv2d_4a_3x3/BatchNorm/moving_mean',
 'InceptionV3/Conv2d_4a_3x3/BatchNorm/moving_variance',
 'InceptionV3/Conv2d_4a_3x3/weights',

And for the saver:

'InceptionV3/Conv2d_1a_3x3/weights:0',
 'InceptionV3/Conv2d_1a_3x3/biases:0',
 'InceptionV3/Conv2d_2a_3x3/weights:0',
 'InceptionV3/Conv2d_2a_3x3/biases:0',
 'InceptionV3/Conv2d_2b_3x3/weights:0',
 'InceptionV3/Conv2d_2b_3x3/biases:0',
 'InceptionV3/Conv2d_3b_1x1/weights:0',
 'InceptionV3/Conv2d_3b_1x1/biases:0',
 'InceptionV3/Conv2d_4a_3x3/weights:0',
 'InceptionV3/Conv2d_4a_3x3/biases:0',

Looks like a mismatch with the biases and the batchnorms. I am using tensorflow version 1.4.0 and models commit 69cf6fc if that is relevant. The checkpoint file is taken from inception_v3_2016_08_28.tar.gz

[reedwm]

@sguada can you resolve this?

[Steven-N-Hart]

+1

[jlamypoirier]

Can fix it using slim arg scope:

with slim.arg_scope([slim.conv2d, slim.fully_connected], normalizer_fn=slim.batch_norm,
                            normalizer_params={'is_training': is_training, 'updates_collections': None}):
    inception,inception_layers=inception_v3_base(tf_images)

Still, this should deserve a fix or at least some documentation.

[liuruijin17]

thank you @jlamypoirier
indeed, this should deserve some documentation

[vshmyhlo]

+1 have the same problem on InceptionV1

[Syzygy2048]

A better solution than the one proposed by @jlamypoirier is probably to use inception_v3_arg_scope() which can be found in slim.nets.inception:

       with slim.arg_scope(slim.nets.inception.inception_v3_arg_scope()):
            logits, endpoints = slim.nets.inception.inception_v3(input,
                                                         num_classes=1001,
                                                         is_training=False)

@v-shmyhlo inception_v1_arg_scope() should also work.

[ndrplz]

+1 same problem in resnet_v2

fixed it with with slim.arg_scope(resnet_v2.resnet_arg_scope()) but I agree it should be worth mentioning somewhere in the docs

[Elites2017]

@jlamypoirier @ndrplz @Syzygy2048 @v-shmyhlo @reedwm @Steven-N-Hart I'm having the same problems since 6 six days, but I don't know which file to edit. Could you help me guy?

I want to train " mobilenet_v1_0.5_128 " or " mobilenet_v1_1.0_224 " one of these pretained models, but the config files from the tensorflow official github repository seem not to be matched to the mentioned models. Could you give me a link to the right config file Guys? Or any help is welcomed.

Thanks in advance,

[vshmyhlo]

Hey @Elites2017 , have you tried using args_scope with your model's specific argument as mentioned in #2977 (comment)?

[Elites2017]

Hey @Elites2017 , have you tried using args_scope with your model's specific argument as mentioned in #2977 (comment)?

I read your comment above, but I don't know which file / field to edit. I'm using mobilenet_v1_1.0_224 with the config file: models/research/object_detection/samples/configs/ssd_mobilenet_v1_pets.config.

Thanks in advance for your help

[Elites2017]

@jlamypoirier @ndrplz @Syzygy2048 @v-shmyhlo @reedwm @Steven-N-Hart

What's the code to import model (.pb) files to TensorBoard so that I can see the right parameters to config my .pb model to .tflite?

Thanks,

[zhangbohnu]

I think I solved it, the code(slim) will record the num of last ckpt , so when you train other modles will lost the num. if you want train original model again, you should point to the specific file of ckpt rather than dir

[wt-huang]

Closing as the issue is resolved