thierryherrmann

public class TrainAndServeSavedModel {
    public static void main(String[] args) throws Exception {

        // args[0]: saved model directory
        SavedModelBundle savedModel = SavedModelBundle.load(args[0], "serve");
        Map<String, SignatureDef> signatureMap = savedModel.metaGraphDef().getSignatureDefMap();

        Tensor<TFloat32> inputTensor = TFloat32.tensorOf(StdArrays.ndCopyOf(new float[][] { { 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f, 1.0f } }));
        Tensor<TFloat32> labelTensor = TFloat32.tensorOf(StdArrays.ndCopyOf(new float[] { 1.0f }));
        

thierryherrmann

input_X = graph.get_tensor_by_name('my_train_X:0')
input_y = graph.get_tensor_by_name('my_train_y:0')
output_1 = graph.get_tensor_by_name('StatefulPartitionedCall_1:0')
output_2 = graph.get_tensor_by_name('StatefulPartitionedCall_1:1')

out_val_1, out_val_2 = session.run([output_1, output_2], 
                                   feed_dict={input_X: X_train[0:1], input_y: y_train[0:1]})

thierryherrmann

def train_predict_serve(model_dir):
    tf.compat.v1.reset_default_graph()
    session = tf.compat.v1.Session()
    tf.compat.v1.saved_model.loader.load(session, tags=[tf.saved_model.SERVING], export_dir=model_dir)
    graph = session.graph
    operations=graph.get_operations()
    
    input_X = graph.get_tensor_by_name('my_train_X:0')
    input_y = graph.get_tensor_by_name('my_train_y:0')
    output_loss = graph.get_tensor_by_name('StatefulPartitionedCall_1:0')

thierryherrmann

tensor    :  model/layer_with_weights-0/bias/.OPTIMIZER_SLOT/opt/m/.ATTRIBUTES/VARIABLE_VALUE (30,) 
[ 3.4107354e-05 -6.1855838e-04 -1.6536651e-06  1.6930330e-06
 -9.1597438e-05 -3.8669934e-04  5.6557164e-05  7.1755665e-08
 -1.2517045e-04 -1.0449246e-03  5.9954262e-05  7.3613039e-05
  6.6272205e-06 -5.7156640e-04  5.4908687e-06 -7.3699164e-05
 -8.7973615e-04 -3.6661630e-04  5.2946081e-05 -5.7122961e-04
 -8.7792240e-04 -4.1600107e-04 -1.2562575e-03 -2.4318745e-06
  7.0880642e-06  9.7999236e-06 -6.5629813e-04  1.1121790e-05
 -1.3819840e-03  6.7142719e-06]

thierryherrmann

del new_module
new_module_2 = tf.keras.models.load_model(model_dir)

loss_hist = train_module(new_module_2, train_dataset, valid_dataset)
plot_loss(loss_hist)

save_module(new_module_2, model_dir)

thierryherrmann

tensor    :  model/layer_with_weights-0/bias/.OPTIMIZER_SLOT/opt/m/.ATTRIBUTES/VARIABLE_VALUE (30,) 
[ 1.3162548e-04 -1.0862495e-03  8.3323405e-04  8.4080239e-06
 -1.6426330e-04 -9.0881845e-04  4.7971989e-04 -6.0352772e-06
 -9.3550794e-04 -3.1544755e-03  5.4244534e-04  1.0909925e-03
  1.3340317e-03 -1.0700974e-03  3.7469756e-04 -1.5879219e-03
 -2.1641832e-03 -1.7716389e-03  2.8458738e-04 -6.3899945e-04
 -2.9655998e-03 -1.7114554e-03 -3.9885961e-03  2.6567639e-05
 -3.6036890e-05  6.1224034e-04 -1.0181948e-03  1.6523007e-04
 -4.8340447e-03  1.5539475e-03]

thierryherrmann

inspect_checkpoint(model_dir + '/variables/variables', print_values=True, 
    variables=['model/layer_with_weights-0/bias/.OPTIMIZER_SLOT/opt/m/.ATTRIBUTES/VARIABLE_VALUE'])

thierryherrmann

loss_hist = train_module(new_module, train_dataset, valid_dataset)
plot_loss(loss_hist)

save_module(new_module, model_dir)

thierryherrmann

type of reloaded module: <class 'tensorflow.python.saved_model.load.Loader._recreate_base_user_object.<locals>._UserObject'>
type of instantiated module: <class '__main__.CustomModule'>
my_train function: <tensorflow.python.saved_model.function_deserialization.RestoredFunction object at 0x7fc1c5872390>
__call__ function: <tensorflow.python.saved_model.function_deserialization.RestoredFunction object at 0x7fc1c58cbfd0>
sample prediction:  [[0.54084957]]

thierryherrmann

print('type of reloaded module:', type(new_module))
print('type of instantiated module:', type(CustomModule()))
print('my_train function:', new_module.my_train)
print('__call__ function:', new_module.__call__)

# demo a call to the module. (calls the __call__() method)
print('sample prediction: ', new_module(X_train[0:1]).numpy())