xiankgx
commented
4 years ago Hi,
I am having exactly this problem as well.
I was able to avoid it by replacing my UpSampling2D layer to something else (tf.image.resize) but what this means is the model loses the ability to take arbitrary sized inputs because in tf.image.resize I have to specify the target size (height and width) instead of specifying a width and height scale as in UpSampling2D. As a result from this, the converted model is basically useless.
These are the versions of libraries that I used: tensorflow : 2.0.0 coremltools: 3.3 (<- the latest version is 3.4 but with that, conversion doesn't even begin with another error) tfcoreml: 1.1
🐞Describe the bug
Trace
[SSAConverter] [197/239] Converting op type: 'LeakyRelu', name: 'conv_dw_15_3_leaky_relu/LeakyRelu', output_shape: (-1, 256, -1, -1). [SSAConverter] [198/239] Converting op type: 'Shape', name: 'up_sampling2d_1/Shape', output_shape: (4,). [SSAConverter] [199/239] Converting op type: 'Conv2D', name: 'conv_pw_15_3/Conv2D', output_shape: (-1, 512, -1, -1). [SSAConverter] [200/239] Converting op type: 'StridedSlice', name: 'up_sampling2d_1/strided_slice', output_shape: (2,). [SSAConverter] [201/239] Converting op type: 'BatchNorm', name: 'conv_pw_15_3_bn/FusedBatchNormV3/Add_batch_norm', output_shape: (-1, 512, -1, -1). [SSAConverter] [202/239] Converting op type: 'Mul', name: 'up_sampling2d_1/mul', output_shape: (2,). [SSAConverter] [203/239] Converting op type: 'LeakyRelu', name: 'conv_pw_15_3_leaky_relu/LeakyRelu', output_shape: (-1, 512, -1, -1). [SSAConverter] [204/239] Converting op type: 'ResizeBilinear', name: 'up_sampling2d_1/resize/ResizeBilinear', output_shape: (-1, 128, -1, -1). Traceback (most recent call last): File "/Users/opt/anaconda3/envs/py_374/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3331, in run_code exec(code_obj, self.user_global_ns, self.user_ns) File "", line 58, in
class_labels='/Users/Desktop/saved_model/loto_classes.txt',
File "/Users/opt/anaconda3/envs/py_374/lib/python3.7/site-packages/tfcoreml/_tf_coreml_converter.py", line 691, in convert
custom_shape_functions=custom_shape_functions)
File "/Users/opt/anaconda3/envs/py_374/lib/python3.7/site-packages/coremltools/converters/tensorflow/_tf_converter.py", line 193, in convert
optional_inputs=optional_inputs)
File "/Users/opt/anaconda3/envs/py_374/lib/python3.7/site-packages/coremltools/converters/nnssa/coreml/ssa_converter.py", line 154, in ssa_convert
converter.convert()
File "/Users/opt/anaconda3/envs/py_374/lib/python3.7/site-packages/coremltools/converters/nnssa/coreml/ssa_converter.py", line 578, in convert
convert_func(node)
File "/Users/opt/anaconda3/envs/py_374/lib/python3.7/site-packages/coremltools/converters/nnssa/coreml/ssa_converter.py", line 2193, in _convert_resize_bilinear
raise ValueError('[SSAConverter] Unable to determine target size'
ValueError: [SSAConverter] Unable to determine target size for ResizeBilinear## To Reproduce
import tfcoreml tfcoreml.convert( tf_model_path='/Users/Desktop/saved_model/saved_model1.pb', mlmodel_path='/Users/Desktop/saved_model/test.mlmodel', input_name_shape_dict={"image_input": [1, 416, 416, 3]}, output_feature_names= ["conv2d_3/BiasAdd","conv2d_8/BiasAdd","conv2d_13/BiasAdd"], image_input_names="image_input", red_bias=-1, green_bias=-1, blue_bias=-1, image_scale=2.0 / 255.0, minimum_ios_deployment_target='13', class_labels='/Users/Desktop/saved_model/classes.txt', )
System environment (please complete the following information):
Additional context
We tried using coremltools 3.1 and 3.3 and also with tensorflow 1.15 and 2.2 and have the same issue .
we also tried TF OP "ResizeNearestNeighbor" which is involved by keras UpSampling2D layer with default "interpolation='nearest'" option then tried to change the interpolation method to 'bilinear' to see if it's ok for the CoreML convert tool but it didn't work as well.