TF Serving loading error of self trained model (TFv2.x)

[lodemo]

Hi,

i trained my own model like described in the wiki "Training using the VGGFace2 dataset". I used a recent TFv2.x version and Python 3.8 and replaced all v1 calls with tensorflow.compat.v1 imports. The training was successful and everything works using the resulting model checkpoints.

My main goal however is to use the model in TF Serving, thus i first froze the model using "freeze_graph.py" and then transformed the model into the SavedModel format using the snippet provided in #1055. This all worked without errors, but when i try to load the SavedModel, TF Serving throws this error:

Reading SavedModel debug info (if present) from: /models/facenet/2
tfserving_1  | 2022-01-19 19:56:58.221357: I external/org_tensorflow/tensorflow/cc/saved_model/loader.cc:277] SavedModel load for tags { serve }; Status: fail: Invalid argument: input resource[0] expected type resource != float, the type of assignmovingavg_readvariableop_inceptionresnetv1_conv2d_1a_3x3_batchnorm_moving_mean[0]
tfserving_1  |  In {{node AssignMovingAvg/AssignSubVariableOp}}. Took 17064957 microseconds.
tfserving_1  | 2022-01-19 19:56:58.222254: E tensorflow_serving/util/retrier.cc:37] Loading servable: {name: facenet version: 2} failed: Invalid argument: input resource[0] expected type resource != float, the type of assignmovingavg_readvariableop_inceptionresnetv1_conv2d_1a_3x3_batchnorm_moving_mean[0]
tfserving_1  |  In {{node AssignMovingAvg/AssignSubVariableOp}}

It seems some tensor type is not what is expected but im not sure whats the reason for this.

The same process as above using the 20180402-114759 checkpoints works and can be loaded by TF Serving.

Ive used the latest TF Serving version using the docker-image tensorflow/serving:latest-gpu

Any ideas? Could this come from the fact that i trained the model using a recent TFv2.x version instead of v1.7?

[AlexeyAL1e]

hello, what commands did you use to learn?

[ygpusr]

I also met the same problem. Have you solved it?

[ygpusr]

this problom has been solved, The method is as follows,:you need to add tf.compat.v1.disable_v2_behavior () to train_tripletloss.py
in order to Shielding tf2.x Version