没有注册 Opkernel 来支持 Op 'SparseSoftmaxCrossEntropyWithLogits'

Question

I want use Tensorflow on Android by using pb file trained on python with GPU.

Official example of Tensorflow-demo on Android use pre-trained 'tensorflow_inception_graph.pb'. Like above, I want use mine '*.pb' on Android.

First, I have trained simple Tensorflow code which similar to mnist-example on python like below.

python code

def deepnn(x):
    with tf.name_scope('reshape'):
    x_image = tf.reshape(x, [-1, 30, 6, 1])

    with tf.name_scope('conv1'):
        W_conv1 = weight_variable([5, 5, 1, 32])
        b_conv1 = bias_variable([32])
        h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)

    with tf.name_scope('conv2'):
        W_conv2 = weight_variable([5, 5, 32, 64])
        b_conv2 = bias_variable([64])
        h_conv2 = tf.nn.relu(conv2d(h_conv1, W_conv2) + b_conv2)

    with tf.name_scope('pool2'):
        h_pool2 = max_pool_2x2(h_conv2)

    with tf.name_scope('fc1'):
        W_fc1 = weight_variable([15 * 3 * 64, 1024])
        b_fc1 = bias_variable([1024])

        h_pool2_flat = tf.reshape(h_pool2, [-1, 15 * 3 * 64])
        h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

    with tf.name_scope('dropout'):
        keep_prob = tf.placeholder(tf.float32)
        h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

    with tf.name_scope('fc2'):
        W_fc2 = weight_variable([1024, 28])
        b_fc2 = bias_variable([28])

        y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2

    return y_conv, keep_prob


def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')

def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)

def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)



x = tf.placeholder(tf.float32, [None, 180], name='input')
y_ = tf.placeholder(tf.int64, [None])

y_conv, keep_prob = deepnn(x)
y_conv = tf.identity(y_conv, 'output')

with tf.name_scope('loss'):
    cross_entropy = tf.losses.sparse_softmax_cross_entropy( labels=y_, logits=y_conv )
cross_entropy = tf.reduce_mean(cross_entropy, name='loss')

y_argout = tf.argmax(input=y_conv, axis=1, name="y_argout")

with tf.name_scope('adam_optimizer'):
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

with tf.name_scope('accuracy'):
    correct_prediction = tf.equal(tf.argmax(y_conv, 1), y_)
    correct_prediction = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(correct_prediction)

graph_location = tempfile.mkdtemp()
print('Saving graph to: %s' % graph_location)
train_writer = tf.summary.FileWriter(graph_location)
train_writer.add_graph(tf.get_default_graph())

init_op = tf.global_variables_initializer()
saver = tf.train.Saver()

with tf.Session() as sess:
    sess.run(init_op)

    ...
    ...

    tf.train.write_graph(sess.graph_def, './', 'mlp.pb', as_text=False)

x = tf.placeholder(tf.float32, [None, 180], name='input') and y_conv = tf.identity(y_conv, 'output') naming is to use on Android Studio.

Then, I used outputted 'mlp.pb' on Android Studio like below.

java code on android studio

inferenceInterface = new TensorFlowInferenceInterface(getAssets(), 'mlp.pb');

...

inferenceInterface.feed('input', floatValues, 180);
inferenceInterface.run(new String[] {'output'});　// an error will occur here.
inferenceInterface.fetch('output', outputs);

Occurred Errors

06-08 09:33:59.500 22971-22971/com.example.tkt.tf_test E/TensorFlowInferenceInterface: Failed to run TensorFlow inference with inputs:[input], outputs:[output]
06-08 09:33:59.501 22971-22971/com.example.tkt.tf_test D/AndroidRuntime: Shutting down VM
06-08 09:33:59.584 22971-22971/com.example.tkt.tf_test E/AndroidRuntime: FATAL EXCEPTION: main
Process: com.example.tkt.tf_test, PID: 22971
java.lang.IllegalArgumentException: No OpKernel was registered to support Op 'SparseSoftmaxCrossEntropyWithLogits' with these attrs.  Registered devices: [CPU], Registered kernels:
  <no registered kernels>

     [[Node: loss/sparse_softmax_cross_entropy_loss/xentropy/xentropy = SparseSoftmaxCrossEntropyWithLogits[T=DT_FLOAT, Tlabels=DT_INT64](output, Placeholder)]]

Questions

1. I just want to get variable named 'output' but why does 'SparseSoftmaxCrossEntropyWithLogits' involved?
2. What does Registered kernels means.
3. What should I do?

I would really appreciate for your reply.

Answer 1

Question 1:
tf.losses.sparse_softmax_cross_entropy

Question 2:
Registered kernel means what op kernel could be supported in device.

Question3:
In android, there is only a subset of kernels to keep base library small. If you want to support more custom TensorFlow operators, you should build the libraries yourself following this link
To support SparseSoftmaxCrossEntropyWithLogits in android, you could add sparse_xent_op.cc and sparse_xent_op.h in BUILD file like xent_op.cc and xent_op.h in source file.