Error while executing relay.frontend.from_tflite

rankyung · June 26, 2019, 11:23pm

Hello,

I got the following error while executing “relay.frontend.from_tflite(tflite_model, shape_dict, dtype_dict)”. The tflite model does not use “depthwise_conv2d” operator.
Can anybody advise how to resolve it?
Thanks,

Traceback (most recent call last):
File “/Desktop/quantization/lenet_mnist/compile_tflite_int8.py”, line 39, in
dtype_dict=dtype)
File “/usr/local/lib/python3.7/site-packages/tvm-0.6.dev0-py3.7-macosx-10.13-x86_64.egg/tvm/relay/frontend/tflite.py”, line 709, in from_tflite
op_converter.convert_op_to_relay()
File “/usr/local/lib/python3.7/site-packages/tvm-0.6.dev0-py3.7-macosx-10.13-x86_64.egg/tvm/relay/frontend/tflite.py”, line 93, in convert_op_to_relay
ret = self.convert_mapop_code_str
File “/usr/local/lib/python3.7/site-packages/tvm-0.6.dev0-py3.7-macosx-10.13-x86_64.egg/tvm/relay/frontend/tflite.py”, line 182, in convert_depthwise_conv2d
return self.convert_conv(op, “depthwise”)
File “/usr/local/lib/python3.7/site-packages/tvm-0.6.dev0-py3.7-macosx-10.13-x86_64.egg/tvm/relay/frontend/tflite.py”, line 445, in convert_conv
assert depth_multiplier == 1, "TF frontend have transformed it be 1 "
AssertionError: TF frontend have transformed it be 1 no matter original value be set by 0.25, 0.5 or any else

apivovarov · June 26, 2019, 11:38pm

AFAIK quantized int8 models are not supported yet in TVM in general and in relay from_tflite in particular.

Have you tried to compile non-quantized tflite mobilenet v1 models with your script first?
tflite files can be downloaded from

github.com

tensorflow/models/blob/master/research/slim/nets/mobilenet_v1.md

# Mobilenet_v2
For Mobilenet V2 see this file [mobilenet/README.md]

# MobileNet_v1

[MobileNets](https://arxiv.org/abs/1704.04861) are small, low-latency, low-power models parameterized to meet the resource constraints of a variety of use cases. They can be built upon for classification, detection, embeddings and segmentation similar to how other popular large scale models, such as Inception, are used. MobileNets can be run efficiently on mobile devices with [TensorFlow Mobile](https://www.tensorflow.org/mobile/).

MobileNets trade off between latency, size and accuracy while comparing favorably with popular models from the literature.

![alt text](mobilenet_v1.png "MobileNet Graph")

# Pre-trained Models

Choose the right MobileNet model to fit your latency and size budget. The size of the network in memory and on disk is proportional to the number of parameters. The latency and power usage of the network scales with the number of Multiply-Accumulates (MACs) which measures the number of fused Multiplication and Addition operations. These MobileNet models have been trained on the
[ILSVRC-2012-CLS](http://www.image-net.org/challenges/LSVRC/2012/)
image classification dataset. Accuracies were computed by evaluating using a single image crop.

Model  | Million MACs | Million Parameters | Top-1 Accuracy| Top-5 Accuracy |
:----:|:------------:|:----------:|:-------:|:-------:|
[MobileNet_v1_1.0_224](http://download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_1.0_224.tgz)|569|4.24|70.9|89.9|

This file has been truncated. show original

shape={"input": (1, 224, 224, 3)}

rankyung · June 27, 2019, 2:57pm

@apivovarov Thanks for the comment. The model is not int8 quantized model. The file name is *int8 though. I will try with mobilenet_v1 model to check if my model caused the issue. Thanks!

rankyung · June 27, 2019, 10:43pm

As of now, tf.lite.TFLiteConverter converts “Con2D” operator of the original model to “DepthwiseConv2D” in the TFlite model. Such conversion is not an issue when you infer with the converted TFLite model. However, when I parse and convert the TFlite model into a relay computation graph(calls relay.frontend.from_tflite), it returns an error because “filters” of Conv2D op of the origial model was converted to “depth_multiplier” of DepthwiseConv2D op of the TFlite model, which is violating this (assert depth_multiplier == 1).

I tried to convert .h5 --> .tflite and .pb --> .tflite. Both converted “Conv2D (filters=6)” to “DepthwiseConv2D (depth_multiplier=6)”, and it causes an error when executing relay.frontend.from_tflite.

rankyung · June 27, 2019, 10:58pm

BTW, Mobilenet works.
My model not using Conv2D works as well.