Overview
I know this subject has been discussed many times, but I am having a hard time understanding the workflow, or rather, the variations of the workflow.
For example, imagine you are installing TensorFlow on Windows 10. The main goal being to train a custom model, convert to TensorFlow Lite, and copy the converted .tflite file to a Raspberry Pi running TensorFlow Lite.
The confusion for me starts with the conversion process. After following along with multiple guides, it seems TensorFlow is often install with pip, or Anaconda. But then I see detailed tutorials which indicate it needs to be built from source in order to convert from TensorFlow models to TFLite models.
To make things more interesting, I've also seen models which are converted via Python scripts as seen here.
Question
So far I have seen 3 ways to do this conversion, and it could just be that I don't have a grasp on the full picture. Below are the abbreviated methods I have seen:
Build from source, and use the TensorFlow Lite Optimizing Converter (TOCO):
bazel run --config=opt tensorflow/lite/toco:toco -- --input_file=$OUTPUT_DIR/tflite_graph.pb --output_file=$OUTPUT_DIR/detect.tflite ...
Use the TensorFlow Lite Converter Python API:
converter = tf.lite.TFLiteConverter.from_saved_model(export_dir)
tflite_model = converter.convert()
with tf.io.gfile.GFile('model.tflite', 'wb') as f:
f.write(tflite_model)
Use the tflite_convert CLI utilities:
tflite_convert --saved_model_dir=/tmp/mobilenet_saved_model --output_file=/tmp/mobilenet.tflite
I *think I understand that options 2/3 are the same, in the sense that the tflite_convert utility is installed, and can be invoked either from the command line, or through a Python script. But is there a specific reason you should choose one over the other?
And lastly, what really gets me confused is option 1. And maybe it's a version thing (1.x vs 2.x)? But what's the difference between the TensorFlow Lite Optimizing Converter (TOCO) and the TensorFlow Lite Converter. It appears that in order to use TOCO you would have to build TensorFlow from source, so is there is a reason you would use one over the other?
There is no difference in the output from different conversion methods, as long as the parameters remain the same. The Python API is better if you want to generate TFLite models in an automated way (for eg a Python script that's run periodically).
The TensorFlow Lite Optimizing Converter (TOCO) was the first version of the TF->TFLite converter. It was recently deprecated and replaced with a new converter that can handle more ops/models. So I wouldn't recommend using toco:toco via bazel, but rather use tflite_convert as mentioned here.
You should never have to build the converter from source, unless you are making some changes to it and want to test them out.
Related
I am currently trying to train a custom model for use in Unity (Barracuda) for object detection and I am struggling near what I believe to be the last part of the pipeline. Following various tutorials and git-repos I have done the following...
Using Darknet, I have trained a custom-model using the Tiny-Yolov2 model. (model tested successfully on a webcam python script)
I have taken the final weights from that training and converted them
to a Keras (h5) file. (model tested successfully on a webcam python
script)
From Keras, I then use tf.save_model to turn it into a
save_model.pd.
From save_model.pd I then convert it using tf2onnx.convert to change
it to an onnx file.
Supposedly from there it can then work in one of a few Unity sample
projects...
...however, this project fails to read in the Unity Sample projects I've tried to use. From various posts it seems that I may need to use a 'frozen' save_model.pd before converting it to ONNX. However all the guides and python functions that seem to be used for freezing save_models require a lot more arguments than I have awareness of or data for after going through so many systems. https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py - for example, after converting into Keras, I only get left with a h5 file, with no knowledge of what an input_graph_def, or output_node_names might refer to.
Additionally, for whatever reason, I cannot find any TF version (1 or 2) that can successfully run this python script using 'from tensorflow.python.checkpoint import checkpoint_management' it genuinely seems like it not longer exists.
I am not sure why I am going through all of these conversions and steps but every attempt to find a cleaner process between training and unity seemed to lead only to dead ends.
Any help or guidance on this topic would be sincerely appreciated, thank you.
What I'm trying to do
I'm trying to learn TensorFlow object recognition and as usual with new things, I scoured the web for tutorials. I don't want to involve any third party cloud service or web development framework, I want to learn to do it with just native JavaScript, Python, and the TensorFlow library.
What I have so far
So far, I've followed a TensorFlow object detection tutorial (accompanied by a 5+ hour video) to the point where I've trained a model in Tensorflow (python) and want to convert it to run in a browser via TensorflowJS. I've also tried other tutorials and haven't seemed to find one that explains how to do this without a third party cloud / tool and React.
I know in order to use this model with tensorflow.js my goal is to get files like:
group1-shard1of2.bin
group1-shard2of2.bin
labels.json
model.json
I've gotten to the point where I created my TFRecord files and started training:
py Tensorflow\models\research\object_detection\model_main_tf2.py --model_dir=Tensorflow\workspace\models\my_ssd_mobnet --pipeline_config_path=Tensorflow\workspace\models\my_ssd_mobnet\pipeline.config --num_train_steps=100
It seems after training the model, I'm left with:
files named checkpoint, ckpt-1.data-00000-of-00001, ckpt-1.index, pipeline.config
the pre-trained model (which I believe isn't the file that changes during training, right?) ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8
I'm sure it's not hard to get from this step to the files I need, but I honestly browsed a lot of documentation and tutorials and google and didn't see an example of doing it without some third party cloud service. Maybe it's in the documentation, I'm missing something obvious.
The project directory structure looks like this:
Where I've looked for an answer
For some reason, frustratingly, every single tutorial I've found (including the one linked above) for using a pre-trained Tensorflow model for object detection via TensorFlowJS has required the use of IBM Cloud and ReactJS. Maybe they're all copying from some tutorial they found and now all the tutorials include this, I don't know. What I do know is I'm building an Electron.js desktop app and object detection shouldn't require network connectivity assuming the compute is happening on the user's device. To clarify: I'm creating an app where the user trains the model, so it's not just a matter of one time conversion. I want to be able to train with Python Tensorflow and convert the model to run on JavaScript Tensorflow without any cloud API.
So I stopped looking for tutorials and tried looking directly at the documentation at https://github.com/tensorflow/tfjs.
When you get to the section about importing pre-trained models, it says:
Importing pre-trained models
We support porting pre-trained models from:
TensorFlow SavedModel
Keras
So I followed that link to Tensorflow SavedModel, which brings us to a project called tfjs-converter. That repo says:
This repository has been archived in favor of tensorflow/tfjs.
This repo will remain around for some time to keep history but all
future PRs should be sent to tensorflow/tfjs inside the tfjs-core
folder.
All history and contributions have been preserved in the monorepo.
Which sounds a bit like a circular reference to me, considering it's directing me to the page that just told me to go here. So at this point you're wondering well is this whole library deprecated, will it work or what? I look around in this repo anyway, into: https://github.com/tensorflow/tfjs-converter/tree/master/tfjs-converter
It says:
A 2-step process to import your model:
A python pip package to convert a TensorFlow SavedModel or TensorFlow Hub module to a web friendly format. If you already have a converted model, or are using an already hosted model (e.g. MobileNet), skip this step.
JavaScript API, for loading and running inference.
And basically says to create a venv and do:
pip install tensorflowjs
tensorflowjs_converter \
--input_format=tf_saved_model \
--output_format=tfjs_graph_model \
--signature_name=serving_default \
--saved_model_tags=serve \
/mobilenet/saved_model \
/mobilenet/web_model
But wait, are the checkpoint files I have a "TensorFlow SavedModel"? This doesn't seem clear, the documentation doesn't explain. So I google it, find the documentation, and it says:
You can save and load a model in the SavedModel format using the
following APIs:
Low-level tf.saved_model API. This document describes how to use this
API in detail. Save: tf.saved_model.save(model, path_to_dir)
The linked syntax extrapolates somewhat:
tf.saved_model.save(
obj, export_dir, signatures=None, options=None
)
with an example:
class Adder(tf.Module):
#tf.function(input_signature=[tf.TensorSpec(shape=[], dtype=tf.float32)])
def add(self, x):
return x + x
model = Adder()
tf.saved_model.save(model, '/tmp/adder')
But so far, this isn't familiar at all. I don't understand how to take the results of my training process so far (the checkpoints) to load it into a variable model so I can pass it to this function.
This passage seems important:
Variables must be tracked by assigning them to an attribute of a
tracked object or to an attribute of obj directly. TensorFlow objects
(e.g. layers from tf.keras.layers, optimizers from tf.train) track
their variables automatically. This is the same tracking scheme that
tf.train.Checkpoint uses, and an exported Checkpoint object may be
restored as a training checkpoint by pointing
tf.train.Checkpoint.restore to the SavedModel's "variables/"
subdirectory.
And it might be the answer, but I'm not really clear on what it means as far as being "restored", or where I go from there, if that's even the right step to take. All of this is very confusing to someone learning TF which is why I looked for a tutorial that does it, but again, I can't seem to find one without third party cloud services / React.
Please help me connect the dots.
You can convert your model to TensorFlowJS format without any cloud services. I have laid out the steps below.
I'm sure it's not hard to get from this step to the files I need.
The checkpoints you see are in tf.train.Checkpoint format (relevant source code that creates these checkpoints in the object detection model code). This is different from the SavedModel and Keras formats.
We will go through these steps:
Checkpoint (current) --> SavedModel --> TensorFlowJS
Converting from tf.train.Checkpoint to SavedModel
Please see the script models/research/object_detection/export_inference_graph.py to convert the Checkpoint files to SavedModel.
The code below is taken from the docs of that script. Please adjust the paths to your specific project. --input_type should remain as image_tensor.
python export_inference_graph.py \
--input_type image_tensor \
--pipeline_config_path path/to/ssd_inception_v2.config \
--trained_checkpoint_prefix path/to/model.ckpt \
--output_directory path/to/exported_model_directory
In the output directory, you should see a savedmodel directory. We will use this in the next step.
Converting SavedModel to TensorFlowJS
Follow the instructions at https://github.com/tensorflow/tfjs/tree/master/tfjs-converter, specifically paying attention to the "TensorFlow SavedModel example". The example conversion code is copied below. Please modify the input and output paths for your project. The --signature_name and --saved_model_tags might have to be changed, but hopefully not.
tensorflowjs_converter \
--input_format=tf_saved_model \
--output_format=tfjs_graph_model \
--signature_name=serving_default \
--saved_model_tags=serve \
/mobilenet/saved_model \
/mobilenet/web_model
Using the TensorFlowJS model
I know in order to use this model with tensorflow.js my goal is to get files like:
group1-shard1of2.bin
group1-shard2of2.bin
labels.json
model.json
The steps above should create these files for you, though I don't think labels.json will be created. I am not sure what that file should contain. TensorFlowJS will use model.json to construct the inference graph, and it will load the weights from the .bin files.
Because we converted a TensorFlow SavedModel to a TensorFlowJS model, we will need to load the JS model with tf.loadGraphModel(). See the tfjs converter page for more information.
Note that for TensorFlowJS, there is a difference between a TensorFlow SavedModel and a Keras SavedModel. Here, we are dealing with a TensorFlow SavedModel.
The Javascript code to run the model is probably out of scope for this answer, but I would recommend reading this TensorFlowJS tutorial. I have included a representative javascript portion below.
import * as tf from '#tensorflow/tfjs';
import {loadGraphModel} from '#tensorflow/tfjs-converter';
const MODEL_URL = 'model_directory/model.json';
const model = await loadGraphModel(MODEL_URL);
const cat = document.getElementById('cat');
model.execute(tf.browser.fromPixels(cat));
Extra notes
... Which sounds a bit like a circular reference to me,
The TensorFlowJS ecosystem has been consolidated in the tensorflow/tfjs GitHub repository. The tfjs-converter documentation lives there now. You can create a pull request to https://github.com/tensorflow/tfjs to fix the SavedModel link to point to the tensorflow/tfjs repository.
I am currently working with the YoloV3-tiny.
Repository: https://github.com/AlexeyAB/darknet
To import the network into C++ project I use OpenVINO-Toolkit. In more detail I use the following procedure to convert the network:
Converting YOLO* Models to the Intermediate Representation (IR)
This procedure carries out a conversion and an optimization to proceed with the inference.
Now, I would like to try the YoloV4 because it seems to be more effective for the purpose of the project. The problem is that OpenVINO Toolkit does not yet support this version and does not report the .json (file needed for optimization) file relative to version 4 but only up to version 3.
What has changed in terms of structure between version 3 and version 4 of the Yolo?
Can I hopefully hope that the conversion of the YoloV3-tiny (or YoloV3) is the same as the YoloV4?
Is the YoloV4 much slower than the YoloV3-tiny using only the CPU for inference?
When will the YoloV4-tiny be available?
Does anyone have information about it?
The difference between YoloV4 and YoloV3 is the backbone. YoloV4 has CSPDarknet53, whilst YoloV3 has Darknet53 backbone.
See https://arxiv.org/pdf/2004.10934.pdf.
Also, YoloV4 is not supported officially by OpenVINO. However, you can still test and validate YoloV4 on your end with some workaround. There is one way for now to run YoloV4 through OpenCV which will build network using nGraph API and then pass to Inference Engine. See https://github.com/opencv/opencv/pull/17185.
The key problem is the Mish activation function - there is no optimized implementation yet, which is why we have to implement it by definition with tanh and exponential functions. Unfortunately, one-to-one topology comparison shows significant performance degradation. The performance results are also available in the github link above.
https://github.com/TNTWEN/OpenVINO-YOLOV4
This is my project based on v3's converter (darknet -> tensorflow ->IR)and i have finished the adaptation of OpenVINO Yolov4,v4-relu,v4-tiny.
You could have a try. And you can use V4's IRmodel and run on v3's c++ demo directly
I would like to convert a TRT optimized frozen model to saved model for tensorflow serving. Are there any suggestions or sources to share?
Or are there any other ways to deploy a TRT optimized model in tensorflow serving?
Thanks.
Assuming you have a TRT optimized model (i.e., the model is represented already in UFF) you can simply follow the steps outlined here: https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#python_topics. Pay special attention to section 3.3 and 3.4, since in these sections you actually build the TRT engine and then save it to a file for later use. From that point forward, you can just re-use the serialized engine (aka. a PLAN file) to do inference.
Basically, the workflow looks something like this:
Build/train model in TensorFlow.
Freeze model (you get a protobuf representation).
Convert model to UFF so TensorRT can understand it.
Use the UFF representation to build a TensorRT engine.
Serialize the engine and save it to a PLAN file.
Once those steps are done (and you should have sufficient example code in the link I provided) you can just load the PLAN file and re-use it over and over again for inference operations.
If you are still stuck, there is an excellent example that is installed by default here: /usr/src/tensorrt/samples/python/end_to_end_tensorflow_mnist. You should be able to use that example to see how to get to the UFF format. Then you can just combine that with the example code found in the link I provided.
I have trained a model for detection, which is doing great when embedded in tensorflow sample app.
After freezing with export_tflite_ssd_graph and conversion to tflite using toco the results do perform rather bad and have a huge "variety".
Reading this answer on a similar problem with loss of accuracy I wanted to try tflite_diff_example_test on a tensorflow docker machine.
As the documentation is not that evolved right now, I build the tool referencing this SO Post
using:
bazel build tensorflow/contrib/lite/testing/tflite_diff_example_test.cc which ran smooth.
After figuring out all my needed input parameters I tried the testscript with following commands:
~/.cache/bazel/_bazel_root/68a62076e91007a7908bc42a32e4cff9/external/bazel_tools/tools/test/test-setup.sh tensorflow/contrib/lite/testing/tflite_diff_example_test '--tensorflow_model=/tensorflow/shared/exported/tflite_graph.pb' '--tflite_model=/tensorflow/shared/exported/detect.tflite' '--input_layer=a,b,c,d' '--input_layer_type=float,float,float,float' '--input_layer_shape=1,3,4,3:1,3,4,3:1,3,4,3:1,3,4,3' '--output_layer=x,y'
and
bazel-bin/tensorflow/contrib/lite/testing/tflite_diff_example_test --tensorflow_model="/tensorflow/shared/exported/tflite_graph.pb" --tflite_model="/tensorflow/shared/exported/detect.tflite" --input_layer=a,b,c,d --input_layer_type=float,float,float,float --input_layer_shape=1,3,4,3:1,3,4,3:1,3,4,3:1,3,4,3 --output_layer=x,y
Both ways are failing. Errors:
way:
tflite_diff_example_test.cc:line 1: /bazel: Is a directory
tflite_diff_example_test.cc: line 3: syntax error near unexpected token '('
tflite_diff_example_test.cc: line 3: 'Licensed under the Apache License, Version 2.0 (the "License");'
/root/.cache/bazel/_bazel_root/68a62076e91007a7908bc42a32e4cff9/external/bazel_tools/tools/test/test-setup.sh: line 184: /tensorflow/: Is a directory
/root/.cache/bazel/_bazel_root/68a62076e91007a7908bc42a32e4cff9/external/bazel_tools/tools/test/test-setup.sh: line 276: /tensorflow/: Is a directory
way:
2018-09-10 09:34:27.650473: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
Failed to create session. Op type not registered 'TFLite_Detection_PostProcess' in binary running on d36de5b65187. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib, accessing (e.g.)tf.contrib.resamplershould be done before importing the graph, as contrib ops are lazily registered when the module is first accessed.
I would really appreciate any help, that enables me to compare the output of two graphs using tensorflows given tests.
The second way you mentioned is the correct way to use tflite_diff. However, the object detection model containing the TFLite_Detection_PostProcess op cannot be run via tflite_diff.
tflite_diff runs the provided TensorFlow (.pb) model in the TensorFlow runtime and runs the provided TensorFlow Lite (.tflite) model in the TensorFlow Lite runtime. In order to run the .pb model in the TensorFlow runtime, all of the operations must be implemented in TensorFlow.
However, in the model you provided, the TFLite_Detection_PostProcess op is not implemented in TensorFlow runtime - it is only available in the TensorFlow Lite runtime. Therefore, TensorFlow cannot resolve the op. Therefore, you unfortunately cannot use the tflite_diff tool with this model.