tensorflow lite: error when convert frozen model to lite format

[mrbrantofgithub]

System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
  No
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
  Ubuntu 14.04
• TensorFlow installed from (source or binary):
  source
• TensorFlow version (use command below):
  1.3.0
• Python version:
  2.7
• Bazel version (if compiling from source):
  0.7.0
• GCC/Compiler version (if compiling from source):
  gcc (Ubuntu 4.8.4-2ubuntu1~14.04.3) 4.8.4
• CUDA/cuDNN version:
  cuda8.0/cudnn6.0

I tried to convert squeezenet frozen model to lite format with the following command:
"bazel run --config=opt tensorflow/contrib/lite/toco:toco -- --input_file=/home/xxx/caffe-tensorflow/npy2ckpt/squeezenet/frozen_model.pb --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --output_file=/home/xxx/caffe-tensorflow/npy2ckpt/squeezenet/squeezenet.lite --inference_type=FLOAT --input_type=FLOAT --input_arrays=input --output_arrays=prob --input_shapes=1,227,227,3"

the output is shown below:
2017-11-21 18:35:29.977505: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before general graph transformations: 170 operators, 231 arrays (0 quantized)
2017-11-21 18:35:29.981856: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 1: 40 operators, 93 arrays (0 quantized)
2017-11-21 18:35:29.982061: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before dequantization graph transformations: 40 operators, 93 arrays (0 quantized)
2017-11-21 18:35:29.982201: I tensorflow/contrib/lite/toco/allocate_transient_arrays.cc:312] Total transient array allocated size: 4071680 bytes, theoretical optimal value: 4071680 bytes.
2017-11-21 18:35:29.982317: I tensorflow/contrib/lite/toco/toco_tooling.cc:255] Estimated count of arithmetic ops: 0.781679 billion (note that a multiply-add is counted as 2 ops).
2017-11-21 18:35:29.982482: F tensorflow/contrib/lite/toco/tflite/export.cc:192] Unsupported operator: Squeeze

Then I tried to convert mobilenet_v1_1.0_224.pb to lite format, the same error as above.
"bazel run --config=opt tensorflow/contrib/lite/toco:toco -- --input_file=/home/xxx/Downloads/freeze_mobilenet/MobileNet/img224/mobilenet_v1_1.0_224.pb --input_format=TENSORFLOW_GRAPHDEF --output_format=TFLITE --output_file=/home/xxx/Downloads/freeze_mobilenet/MobileNet/img224/mobilenet.lite --inference_type=FLOAT --input_type=FLOAT --input_arrays=input --output_arrays=output --input_shapes=1,224,224,3"

output:
2017-11-21 22:07:39.747095: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before general graph transformations: 418 operators, 584 arrays (0 quantized)
2017-11-21 22:07:39.766175: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 1: 31 operators, 88 arrays (0 quantized)
2017-11-21 22:07:39.766390: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before dequantization graph transformations: 31 operators, 88 arrays (0 quantized)
2017-11-21 22:07:39.766592: I tensorflow/contrib/lite/toco/allocate_transient_arrays.cc:312] Total transient array allocated size: 6422528 bytes, theoretical optimal value: 4816896 bytes.
2017-11-21 22:07:39.766751: I tensorflow/contrib/lite/toco/toco_tooling.cc:255] Estimated count of arithmetic ops: 1.14264 billion (note that a multiply-add is counted as 2 ops).
2017-11-21 22:07:39.766952: F tensorflow/contrib/lite/toco/tflite/export.cc:192] Unsupported operator: Squeeze

Although I installed tensorflow with "pip install tensorflow-gpu", in order to convert model to lite format, I git clone the tensorflow files and configure, bazel to compile the files. I don't know whether this affect the converting of models, but the error is really strange!

[miaout17]

For now, Squeeze is supported only when it can be fused into other ops.

Please refer to this doc to see the ops compatibility. The team is working on supporting more ops in Toco and hopefully this will be supported in the near future.

[mpeniak]

Thanks for the info @miaout17, is ssd-mobilenet supported please? I cannot tell based on that document.

[gargn]

The model ssd-mobilenet is currently not supported. The team is working on supporting more ops and models.

Regarding mobilenet_v1_1.0_224, feel free to reference this doc for an example on converting an available mobilenet_v1_1.0_224 model using toco.

A pretrained frozen mobilenet_v1_1.0_224 can be found here: https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_0.25_128_frozen.tgz. Additionally, the list of available tflite models is found in this doc.

[mrbrantofgithub]

Now that I convert the mobilenet model to lite format failed, how tensorflow offered us the converted mobilenet lite model on the github? @miaout17

[aselle]

As @gargn mentioned, we haven't yet provided an example of Mobilenet SSD, only Mobilenet classification. We will likely provide support for that in the future. If you are adventurous, we are happy to help you along.

[tensorflowbutler]

It has been 14 days with no activity and the awaiting response label was assigned. Is this still an issue? Please update the label and/or status accordingly.

[offbye]

when the ssd-mobilenet can be supported by tflite ?

[smitshilu]

@aselle any update on mobilenet SSD support?

[mpeniak]

or any anitcipated release date for this feature? I see you added "contribution welcome" tag, does it mean you are not planning developing any object detection support in-house?

[dsfour]

how to fuse squeeze to other ops? @miaout17