aiforsea-cv-cars

This project is to do car recognition by transfer learning, fine-tuning ResNet models with Cars Dataset from Stanford. It was done in pursuance of finding a solution to the Grab AI for SEA challenge.

Table of contents

• Summary
• Recent state-of-the-art results
• Dependencies
• Development platform
• Usage
  • Prerequisites
  • Steps
  • Some examples
  • Setup docker
• Citations and Acknowledgements

Summary

• Trained and tested with ResNet50, Resnet101, ResNet152. Different notebooks can be found in experiments folder. However, the solution that I intend to submit is in cv-stanford-cars.ipynb notebook and the model file is export-rn101_train_stage2-50e.pkl. It achieves 91.4% accuracy on test set provided by Stanford. For inference, please refer to the usages of inference.py in Usage section.
• Ceteris paribus, a model with denser networks performed better. But, there is a trade-off between resources and accuracy. So, the goal of this project was to arrive at a good model (say > 90% accuracy, without overfitting) within a faster training time. Given more time and resources, we could definitely experiment and fine-tune more to improve accuracy.
• With this goal in mind, I could come up with 90.7% on validation set and 91.4% on test set by fine-tuning ResNet101. Specifically, I used a differential learning rate between slice(1e-4,1e-3) over 40 epochs, with a momentum of Point(0.95, 0.85). The process of training and intuition behind tuning could be found in cv-stanford-cars.ipynb notebook. With ResNet152, I was able to get to 91.8% accuracy with mixup but the model was overfitting.

To learn about how to run and test, jump to Usage section below. Here's a screenshot of my submission of result.

Recent state-of-the-art results

• Deep CNNs With Spatially Weighted Pooling for Fine-Grained Car Recognition: 93.6% (April, 2018)
• Weakly Supervised Data Augmentation Network for Fine-Grained Visual Classification: 94.5% (March, 2019)

Dependencies

• NumPy
• PyTorch
• Fastai
• OpenCV

Development platform

• Google Deep Learning VM based on: Debian GNU/Linux 9.9 (stretch) (GNU/Linux 4.9.0-9-amd64 x86_64\n)
• Machine Type: n1-highmem-8 (8 vCPUs, 52 GB memory)
• Python 3.7, PyTorch 1.1, Fastai 1.0.52

$ python
Python 3.7.0 (default, Oct  9 2018, 10:31:47)
[GCC 7.3.0] :: Anaconda, Inc. on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'1.1.0'
>>> import fastai
>>> fastai.__version__
'1.0.52'

• GPU: NVIDIA Tesla P4, CUDA version 10.0

$ nvidia-smi
Sat Jun 15 06:42:17 2019
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 410.104      Driver Version: 410.104      CUDA Version: 10.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  Tesla P4            Off  | 00000000:00:04.0 Off |                    0 |
| N/A   53C    P0    24W /  75W |   5953MiB /  7611MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0      8503      C   /opt/anaconda3/bin/python                   5943MiB |
+-----------------------------------------------------------------------------+

Usage

As mentioned by the AI challenge, the project uses the Cars Dataset, which contains 16,185 images of 196 classes of cars. The data is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in a 50-50 split.

In addition, I have also setup a docker, which can be built to start a server locally where you can upload an image or give the image url to get the class prediction.

To setup and run the project follow the steps below.

Prerequisites

• Any last stable standard Linux distribution. Below steps have been tested on Debian GNU/Linux 9.9 (stretch). But, Ubuntu 16.04 or 18.04 should be fine too. Note that, Fastai, that is being used in this project, currently supports Linux only, and requires PyTorch v1 and Python 3.6 or later. Moreover, CUDA support on macOS is only available by building PyTorch from source and I have not tested that. So, it is recommended that Linux is used.
• Install PyTorch and Fastai latest stable version as mentioned in Development platform section above.

Steps

• Clone the repo and move to the root directory of the repo. Alternatively, you can download if you don't want to clone.

$ git clone https://github.com/codope/aiforsea-cv-cars.git 
$ cd aiforsea-cv-cars/
$ git status
On branch master
Your branch is up-to-date with 'origin/master'.
nothing to commit, working tree clean

• Do a pip install of the dependencies in requirements.txt. If you already have all the dependencies then you should see an output like below, otherwise sit back and relax until all the dependencies are installed.

$ pip3 install -r requirements.txt
Requirement already satisfied: fbpca==1.0 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 1)) (1.0)
Requirement already satisfied: imageio==2.4.1 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 2)) (2.4.1)
Requirement already satisfied: numpy==1.16.4 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 3)) (1.16.4)
Requirement already satisfied: opencv-python==4.1.0.25 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 4)) (4.1.0.25)
Requirement already satisfied: pandas==0.24.2 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 5)) (0.24.2)
Requirement already satisfied: Pillow==6.0.0 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 6)) (6.0.0)
Requirement already satisfied: python-dateutil==2.8.0 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 7)) (2.8.0)
Requirement already satisfied: pytz==2019.1 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 8)) (2019.1)
Requirement already satisfied: scipy==1.3.0 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 9)) (1.3.0)
Requirement already satisfied: six==1.12.0 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 10)) (1.12.0)
Requirement already satisfied: starlette==0.12.0 in /usr/local/lib/python3.7/site-packages (from -r requirements.txt (line 11)) (0.12.0)

• Run the below script to get the data and organise it into ImageNet folder architecture style, the train/valid split is specified in 'train_split = 0.8' in preprocess.py. Thanks to foamliu for providing the preprocessing script on this github repo

$ sh prepare_data.sh

Sit back and relax until you see the following message on the terminal:

The number of different cars is 196
Training data saved.
Processing test data...
Testing data saved.

NOTE: Please check the aiforsea-cv-cars/model directory. It must have export-rn101_train_stage2-50e.pkl file and it should be 171.15MB in size. If you notice that the file was not downloaded completely then you can download the model from here.

• Execute the inference.py script to do inference from the pre-trained model. It accepts some arguments and syntax is as follows:

$ python inference.py [OPTION] <file_name>

[OPTION] is either one of :

1. predict-one
2. with-accuracy
3. with-confidence

file_name is the absolute path to either one of (depending on OPTION):

1. A .jpg image, if you want to predict class for one image using predict-one option.
2. cars_test_annos_withlabels.mat file containing the test data annotations, if you want to just know the accuracy of the model using with-accuracy option. Note that, the cars_test_annos_withlabels.mat file would have already been saved under the root directory of this project if prepare_data.sh script ran successfully. In case, it isn't there then you can download this file here.
3. A .txt file, which will be the output file, if you want to get predictions for all test data in test set, including the confidence score using with-confidence option.

NOTE: OPTION is optional but file_name is mandatory argument. If you do not provide any option, the code will write the output to the file_name with only the class_id of the predictions for all the test data in the format that is acceptable at Cars 196 Submission Site. Check the last example in the examples section below.

Some examples

• This one below shows the output of prediction of one image. Output is ine the format <class_id, class_name, confidence>.

$ python inference.py predict-one /home/jupyter/ai-for-sea/data/2012-toyota-camry-hybrid-test-review-car-and-driver-photo-434698-s-original.jpg
187, Toyota Camry Sedan 2012, 0.9165221452713013

• This one shows the output of using with-accuracy option:

$ python inference.py with-accuracy /home/jupyter/aiforsea-cv-cars/cars_test_annos_withlabels.mat
torch.Size([8041, 196])
The accuracy is 91.39410257339478%

• And this one shows the usage of with-confidence option. Note the output file result_score.txt in the ls -l output.

$ python inference.py with-confidence result_score.txt
avg fps: 55.15086955505329

$ ls -l /home/jupyter/aiforsea-cv-cars/*.txt
-rw-r--r-- 1 jupyter jupyter       172 Jun 16 10:47 /home/jupyter/aiforsea-cv-cars/requirements.txt
-rw-r--r-- 1 jupyter jupyter    417172 Jun 16 11:05 /home/jupyter/aiforsea-cv-cars/result_score.txt

For your reference, you can find result_score.txt in the repo.

• This one shows the usage without any OPTION.

$ python inference.py result.txt
avg fps: 55.26469256130214

$ ls -l /home/jupyter/aiforsea-cv-cars/*.txt
-rw-r--r-- 1 jupyter jupyter       172 Jun 16 10:47 /home/jupyter/aiforsea-cv-cars/requirements.txt
-rw-r--r-- 1 jupyter jupyter     27774 Jun 16 11:08 /home/jupyter/aiforsea-cv-cars/result.txt
-rw-r--r-- 1 jupyter jupyter    417172 Jun 16 11:05 /home/jupyter/aiforsea-cv-cars/result_score.txt

For your reference, you can find result.txt in the repo.

Setup docker

• First install docker if you don't have it already.
• Next go to the root directory of the project and execute the following commands. Make sure you have the devkit directory after the execution of prepare_data.sh script. In case, the execution of above script failed, you can download devkit here.

$ cd aiforsea-cv-cars/
$ docker build -t testcars .

This will take some time and in the end, you should see a result as follows:

Successfully built 48bc4daadfd7
Successfully tagged testcars:latest

• Next, run the docker with the following command and you should see the result as below:

$ docker run -p 8008:8008 testcars
INFO: Started server process [1]
INFO: Waiting for application startup.
INFO: Uvicorn running on http://0.0.0.0:8008 (Press CTRL+C to quit)

• Now, if you have a car image, then you can upload at localhost:8008. Else, you can also give the url of an image as follows: http://localhost:8008/classify-url?url=https%3A%2F%2Fcars.usnews.com%2Fstatic%2Fimages%2FAuto%2Fizmo%2F338244%2F2012_toyota_corolla_angularfront.jpg The image looks like this:

And your output will look like this:

Try with the Tesla image in images folder for fun!

Citations and Acknowledgements

• 3D Object Representations for Fine-Grained Categorization by Jonathan Krause et. al for the annotated data.
• One cycle learning paper by Leslie Smith.
• mixup: Beyond Empirical Risk Minimization by Hongyi Zhang et. al, I used this for experiments with ResNets to reduce sensitivity to adverserial examples. You can find the relevant notebook in experiments folder.
• Current state-of-the-art results using ResNet101 on the same dataset: Deep CNNs With Spatially Weighted Pooling for Fine-Grained Car Recognition by Qichang Hu et. al and Weakly Supervised Data Augmentation Network for Fine-Grained Visual Classification by Tao Hu et. al
• foamliu for the preprocessing script.
• Fastai for democratising deep learning and making neural nets uncool again!