Hello,
Thank you for building this tool, we are thinking of using it at our company. I am having trouble converting a very simple model which only includes two Conv2D and a Dense Layer.
The code for tensorflow is:
with tf.variable_scope('layer1') as layer1:
conv1 = tf.layers.conv2d(
inputs=images,
filters=M,
kernel_size=[W, W],
padding='same',
name='conv1',
use_bias=False
)
activation1 = tf.nn.relu(conv1, name='relu1')
with tf.variable_scope('layer2') as layer2:
conv2 = tf.layers.conv2d(
inputs=activation1,
filters=N,
kernel_size=[1,1],
activation=tf.nn.relu,
padding='same',
name='conv2',
use_bias=False
)
activation2 = tf.nn.relu(conv2, name='relu2')
with tf.variable_scope('layer3') as layer3:
out = tf.layers.dense(activation2, units=1, name='FullyConnected')
The output I am getting:
2018-06-28 14:11:13.477127: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/Tensordot/transpose_1) not consumed by converter: Transpose.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/Tensordot/MatMul) not consumed by converter: MatMul.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer2/conv2/Conv2D) not consumed by converter: Conv2D.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer1/conv1/Conv2D) not consumed by converter: Conv2D.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer2/conv2/Relu) not consumed by converter: Relu.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/Tensordot) not consumed by converter: Reshape.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/BiasAdd) not consumed by converter: BiasAdd.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/Tensordot/transpose) not consumed by converter: Transpose.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer2/relu2) not consumed by converter: Relu.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer1/relu1) not consumed by converter: Relu.
2018-06-28 14:11:13,830 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/Tensordot/Reshape) not consumed by converter: Reshape.
2018-06-28 14:11:13,831 - 366 - WARNING - WARNING_TF_SCOPE_OP_NOT_CONSUMED: Operation (layer3/FullyConnected/Tensordot/Reshape_1) not consumed by converter: Reshape.
2018-06-28 14:11:13,831 - 122 - ERROR - Conversion failed: ERROR_TF_OPERATION_NOT_MAPPED_TO_LAYER: Some operations in the Tensorflow graph were not resolved to a layer. You can use --allow_unconsumed_nodes for partial graph resolution!
<strong> However, when I remove the last dense layer, the model compiles fine . I find this interesting because when I add the last dense layer, it says a Conv2D operation is not consumed < /strong> . It seems like a few people have experienced this "NOT CONSUMED" issue. Do you think you can help me clarify the rules regarding <strong> What defines resolving an operation to an layer ? < / strong>.
Thanks!