Add test_forward_ssd_mobilenet_v1 to tflite/test_forward

[apivovarov]

This tests uses official non-quantized model ssd_mobilenet_v1_coco downloaded from https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md

ssd_mobilenet_v1_coco_2018_01_28.tar.gz contains checkpoint.
To generate tflite file I exported the checkpoint to tflite_graph.pb file using object_detection/export_tflite_ssd_graph.py form `tensorflow/models' repo.

Then I converted tflite_graph.pb file to TFLITE format using the command below.
Because TFLite frontend does not support custom operators I did not include custom operator TFLite_Detection_PostProcess which are the last operator in the graph.
I did that by specifying --output_arrays as raw_outputs/class_predictions,raw_outputs/box_encodings which are inputs for TFLite_Detection_PostProcess op.
Users can do PostProcessing outside of TVM.
post-processing code is here

# FLOAT no Post Processing
export TFM_PATH=<path_to_dir_containing_tflite_graph.pb>
export FMT=TFLITE
tflite_convert \
--graph_def_file=$TFM_PATH/tflite_graph.pb \
--output_file=$TFM_PATH/model_nopp.tflite \
--output_format=$FMT \
--input_arrays=normalized_input_image_tensor \
--input_shapes=1,300,300,3 \
--inference_type=FLOAT \
--output_arrays="raw_outputs/class_predictions,raw_outputs/box_encodings"

Graph visualization

I uploaded resulting model_nopp.tflite file to s3 bucket belonging to neo-ai organization

[apivovarov]

ci_2: All checks have passed

ci_1: test_vm.test_sum_loop ... ./tests/scripts/task_python_integration.sh: line 41: 31003 Segmentation fault (core dumped) TVM_FFI=ctypes python3 -m nose -v tests/python/relay
http://ci.tvm.ai:8080/blue/organizations/jenkins/tvm/detail/PR-3350/1/pipeline

[apivovarov]

@FrozenGene @kevinthesun @yongwww Could you review?

[FrozenGene]

better to be

# SSD Mobilenet
---------------

Keep the same length

[apivovarov]

fixed

[FrozenGene]

Move this to our dmlc / web-data : https://github.com/dmlc/web-data

[yongwww]

Or download the model directly from tf lite community if possible

[apivovarov]

TF provides tflite files only for quantized models, but TVM does not support quantized models yet.
For non-quantized models they provide tflite_graph.pb file which is compatible with tflite_convert util and can be converted to tflite file.

For now I can put my tflite converted file to dmlc/web-data.

I can also investigate if Jenkins worker has tflite_convert util or python module tensorflow.lite.TFLiteConverter available. In that case I can download tgz model file from Tensorflow zoo, unpack it and convert tflite_graph.pb to tflite file.

[apivovarov]

Actually, official non-quantized tar.gz ssd model files do not have tflite_graph.pb. They only have inference pb file and a checkpoint.
tflite_graph.pb can be created from checkpoint using object_detection/export_tflite_ssd_graph.py from tensorflow/models repo.
We do not have this repo and util on Jenkins worker.

I opened PR to add tflite model packed to tar.gz ssd_mobilenet_v1_coco_2018_01_28_nopp.tgz to dmlc/web-data
dmlc/web-data#186

[apivovarov]

will change it to dmlc once dmlc/web-data#186 is merged

[apivovarov]

updated

[apivovarov]

@FrozenGene CI build/test passed. Can you merge it?

[FrozenGene]

@FrozenGene CI build/test passed. Can you merge it?

I don't have merge privilege. ping @tqchen @srkreddy1238 help to handle.

[apivovarov]

@FrozenGene @yongwww Can you approve the PR?

[yongwww]

LGTM

[kevinthesun]

Thanks!