TensorFlow website claims that Quantization provides up to 3x lower latency on mobile devices:
https://www.tensorflow.org/lite/performance/post_training_quantization
I tried to verify this claim, and found that Quantized models are 45%-75% SLOWER than Float models in spite of being almost 4 times smaller in size. Needless to say, this is very disappointing and conflicts with Google's claims.
My test uses Google's official MnasNet model: https://storage.googleapis.com/mnasnet/checkpoints/mnasnet-a1.tar.gz
Here is the average latency based on 100 inference operations on a freshly rebooted phone:
Pixel 2: float=81ms, quant=118ms
Moto E: float=337ms, quant=590ms
LG Treasure: float=547ms, quant=917ms
My test app measures the timing for only one method (tfLite.runForMultipleInputsOutputs). The results are very consistent (within 1% across multiple executions).
I am hoping to see some comments from the Tensorflow team or anybody who can share their metrics. The numbers above are based on image classifier model. I also tested an SSD MobileNetV2 object detector with similar results (quantized model being substantially slower).
Related
I faced problem regarding Yolo object detection deployment on TX2.
I use pre-trained Yolo3 (trained on Coco dataset) to detect some limited objects (I mostly concern on five classes, not all the classes), The speed is low for real-time detection, and the accuracy is not perfect (but acceptable) on my laptop. I’m thinking to make it faster by multithreading or multiprocessing on my laptop, is it possible for yolo?
But my main problem is that algorithm is not running on raspberry pi and nvidia TX2.
Here are my questions:
In general, is it possible to run yolov3 on TX2 without any modification like accelerators and model compression techniques?
I cannot run the model on TX2. Firstly I got error regarding camera, so I decided to run the model on a video, this time I got the 'cannot allocate memory in static TLS block' error, what is the reason of getting this error? the model is too big. It uses 16 GB GPU memory on my laptop.The GPU memory of raspberry and TX2 are less than 8GB. As far as I know there are two solutions, using a smaller model or using tensor RT or pruning. Do you have any idea if there is any other way?
if I use tiny-yolo I will get lower accuracy and this is not what I want. Is there any way to run any object detection model with high performance for real-time in terms of both accuracy and speed (FPS) on raspberry pi or NVIDIA TX2?
If I clean the coco data for just the objects I concern and then train the same model, I would get higher accuracy and speed but the size would not change, Am I correct?
In general, what is the best model in terms of accuracy for real-time detection and what is the best in terms of speed?
How is Mobilenet? Is it better than YOLOs in terms of both accuracy and speed?
1- Yes it is possible. I already run Yolov3 on Jetson Nano.
2- It depends on model and input resolution of data. You can decrease input resolution. Input images are transferred to GPU VRAM to use on model. Big input sizes can allocate much memory. As far as I remember I have run normal Yolov3 on Jetson Nano(which is worse than tx2) 2 years ago. Also, you can use Yolov3-tiny and Tensorrt as you mention. There are many sources on the web like this & this.
3- I suggest you to have a look at here. In this repo, you can make transfer learning with your dataset & optimize the model with TensorRT & run it on Jetson.
4- Size not dependent to dataset. It depend the model architecture(because it contains weights). Speed probably does not change. Accuracy depends on your dataset. It can be better or worser. If any class on COCO is similiar to your dataset's any class, I suggest you to transfer learning.
5- You have to find right model with small size, enough accuracy, gracefully speed. There is not best model. There is best model for your case which depend on also your dataset. You can compare some of the model's accuracy and fps here.
6- Most people uses mobilenet as feature extractor. Read this paper. You will see Yolov3 have better accuracy, SSD with MobileNet backbone have better FPS. I suggest you to use jetson-inference repo.
By using jetson-inference repo, I get enough accuracy on SSD model & get 30 FPS. Also, I suggest you to use MIPI-CSI camera on Jetson. It is faster than USB cameras.
I fixed the problem 1 and 2 only by replacing import order of the opencv and tensorflow inside the script.Now I can run Yolov3 without any modification on tx2. I got average FPS of 3.
I am using YOLOv4 to train my custom detector. Source: https://github.com/AlexeyAB/darknet
One of the issues while training is the computing power of GPU and available video RAM. What is the relationship between number of object classes and the time it takes to train the model? Also, is it possible to significantly reduce the inference time of images by reducing the number of object classes? The goal is to run inference on a Raspberry Pi or a Jetson Nano.
Any help is much appreciated. Thanks.
Change is number of classes doesn't have significant impact on
inference time.
For example in case of Yolov4, which has got 3 Yolo layers, change in classes leads to change in filter size for conv layers preceding Yolo layers and some computation reduction within Yolo layers that's all. This is very minute compared to overall inference time as conv layers preceding Yolo layers are bottom layers with very small width and hight and also time spent on logic that depends upon number of classes within Yolo layer is very less.
Here:
filters=(classes + 5)x3
Note that tinier version of yolov4 i.e tiny-yolov4 have got two Yolo layers only, instead of 3.
If your intent is to reduce inference time, especially on raspberry pi or a jetson nano, without losing on accuracy/mAP, do following things:
Quantisation: Run inference with INT8 instead of FP32. You can use this repo for this purpose. You can do this for both Jetson nano and raspberry pi.
Use inference library such as tkDNN, which is a Deep Neural Network library built with cuDNN and tensorRT primitives, specifically thought to work on NVIDIA Jetson Boards. You can use this for Jetson nano. Note that with TensorRT, you can use INT8 and FP16 instead of FP32 to reduce detection time.
Following techniques can be used to reduce inference time, but they come at the cost of significant drop in accuracy/mAP:
You can train the models with tinier versions rather than full Yolo versions.
Model Pruning - If you could rank the neurons in the network according to how much they contribute, you could then remove the low ranking neurons from the network, resulting in a smaller and faster network. Pruned yolov3 research paper and it's implementation. This is another pruned Yolov3 implementation.
I tried reducing the number of classes from 80 to 5 classes, I was aiming to detect vehicles only, on YOLOv3 and found a reduction in time. For example, using Intel Core i5-6300U CPU # 2.40 GHz, the time was reduced by 50%, and for Nvidia GeForce 930M, it was reduced by 20%. Generally, the stronger the processor, the less reduction in time you get.
I'm new in everithing about CNN and tensorflow. Im training a pretrained ssd-mobilenev1-pets.config to detect columns of buildings, about one day but the loss is between 2-1 and doesnt decrease since 10 hours ago.
I realized that my input images are 128x128 and SSD resize de image to 300*300.
Does the size of the input images affect the training?
If that is the case, should I retrain the network with larger input images? or what would be another option to decrease the loss? my train dataset has 660 images and test 166 I dont Know if there are enough images
I really aprecciate your help ....
Loss values of ssd_mobilenet can be different from faster_rcnn. From EdjeElectronics' TensorFlow Object Detection Tutorial:
For my training on the Faster-RCNN-Inception-V2 model, it started at
about 3.0 and quickly dropped below 0.8. I recommend allowing your
model to train until the loss consistently drops below 0.05, which
will take about 40,000 steps, or about 2 hours (depending on how
powerful your CPU and GPU are). Note: The loss numbers will be
different if a different model is used. MobileNet-SSD starts with a
loss of about 20, and should be trained until the loss is consistently
under 2.
For more information: https://github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10#6-run-the-training
The SSD Mobilnet architecture demands additional training to suffice
the loss accuracy values of the R-CNN model, however, offers
practicality, scalability, and easy accessibility on smaller devices
which reveals the SSD model as a promising candidate for further
assessment (Fleury and Fleury, 2018).
For more information: Fleury, D. & Fleury, A. (2018). Implementation of Regional-CNN and SSD machine learning object detection architectures for the real time analysis of blood borne pathogens in dark field microscopy. MDPI AG.
I would recommend you to take 15%-20% images for testing which cover all the variety present in training data. As you said you have 650+ images for training and 150+ for testing. That is roughly 25% of testing images. It looks like you have enough images to start with. I know the more, the merrier but make sure your model also has sufficient data to learn from!
Resizing the images does not contribute to the loss. It makes sure there is consistency across all images for the model to recognize them without bias. The loss has nothing to do with image resizing as long as every image is resized identically.
You have to make stops and recover checkpoints again and again if you want your model to be perfectly fit. Usually, you can get away with good accuracy by re-training the ssd mobilenet until the loss consistently becomes under 1.Ideally we want the loss to be as lower as possible but we want to make sure the model is not over-fitting. It is all about trial and error. (Loss between 0.5 and 1 seems to be doing the job well but again it all depends on you.)
The reason I think your model is underperforming is due to the fact that you have variety of testing data and not enough training data to suffice.
The model has not been given enough knowledge in training data to make the model learn for new variety of testing data. (For example : Your test data has some images of new angles of buildings which are not sufficiently present in training data). In that case, I recommend you to put variety of all images in training data and then picking images to test making sure you still have sufficient training data of new postures. That's why I recommend you to take 15%-20% test data.
I’m working on a few side projects that involve deploying ML models to the edge. One of them is a photo-editing app that includes CNN’s for facial recognition, object detection, classification, and style transfer. The other is a NLP app that assists in the writing process by suggesting words and sentence completions..
Once I have a trained model that’s accurate, it ends up being really slow on one or more mobile devices that I'm testing on (usually the lower end Android). I’ve read that there are optimizations one can do to speed models up, but I don’t know how. Is there a standard, go-to tool for optimizing models for mobile/edge?
I will be talking about TensorFlow Lite specifically it is a platform for running TensorFlow ops on Android and iOS. There are several optimisation techniques mentioned on their website but I will discuss the ones which feel important to me.
Constructing relevant models for platforms:
The first step in model optimization is its construction from scratch meaning TensorFlow. We need to create a model which can be used exported to a memory constrained device.
We definitely need to train different models for different machines. A model constructed to work on a high-end TPU will never run efficiently on a Mobile processor.
Create a model which has minimum layers and ops.
Do this without compromising the model's accuracy.
For this, you will need expertise in ML and also which ops are the best to preprocess data.
Also, extra preprocessing of input data brings down the model complexity to a great extent.
Model quantization:
We convert the high precision floats or decimals to lower precision floats. It affects the model's performance slightly but greatly reduces the model size and then holds less of the memory.
Post-training quantization is a general technique to reduce model size while also providing up to 3x lower latency with little degradation in model accuracy. Post-training quantization quantizes weights from floating point to 8-bits of precision - from TF docs.
You can see the TensorFlow Lite TFLiteConverter example:
import tensorflow as tf
converter = tf.lite.TFLiteConverter.from_saved_model(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]
tflite_quant_model = converter.convert()
Also you should try using the post_training_quantize= flag which reduces the model size considerably.
Hope it helps.
I am setting up an object detection pipeline based on recently released tensorflow object detection API. I am using the arXiv as guidance. I am looking to understand the below for training on my own dataset.
It is not clear how they selected the learning rate schedules and how that would change based on the number of GPUs available for training. How do the training rate schedule change based on number of GPU's available for training? The paper mentions 9 GPUs are used. How should I change the training rate if I only want to use 1 GPU?
The released sample training config file for Pascal VOC using Faster R-CNN has initial learning rate = 0.0001. This is 10x lower than what was published in the original Faster-RCNN paper. Is this due to an assumption on the number of GPU's available for training or due to a different reason?
When I start training from the COCO detection checkpoint, how should the training loss decrease? Looking at tensorboard, on my dataset training loss is low - between 0.8 to 1.2 per iteration (with batch size of 1). Below image shows the various losses from tensorboard. . Is this expected behavior?
For questions 1 and 2: our implementation differs in a few small details compared to the original paper and internally we train all of our detectors with asynchronous SGD with ~10 GPUs. Our learning rates are calibrated for this setting (which you will also have if you decide to train via Cloud ML Engine as in the Pets walkthrough). If you use another setting, you will have to do a bit of hyperparameter exploration. For a single GPU, leaving the learning rate alone probably won't hurt performance, but you may be able to get faster convergence by increasing it.
For question 3: Training losses decrease erratically and you can only see the decrease if you smooth the plots quite a bit over time. Moreover, it's hard to explicitly say how well you are doing with respect to eval metrics just by looking at the training losses. I recommend looking at the mAP plots over time as well as the image visualizations to really get an idea of whether your model has "lifted off".
Hope this helps.