Awesome

mxnet-finetuner

An all-in-one Deep Learning toolkit for image classification to fine-tuning pretrained models using MXNet.

Prerequisites

• docker
• docker-compose
• jq
• wget or curl

When using NVIDIA GPUs

• nvidia-docker (Both version 1.0 and 2.0 are acceptable)

If you are using nvidia-docker version 1.0 and have never been running the nvidia-docker command after installing it, run the following command at least once to create the volume for GPU container.

$ nvidia-docker run --rm nvidia/cuda nvidia-smi

Setup

$ git clone https://github.com/knjcode/mxnet-finetuner
$ cd mxnet-finetuner
$ bash setup.sh

setup.sh will automatically generate docker-compose.yml and config.yml which are necessary for executing this tool according to your environment such as existence of the GPU. Please see common settings on how to run with GPU image on NVIDIA GPUs.

When updating the GPU driver of the host machine, re-running setup.sh. Please see After updating the GPU driver of the host machine for details.

Example usage

1. Arrange images into their respective directories

A training data directory (images/train), validation data directory (images/valid), and test data directory (images/test) should containing one subdirectory per image class.

For example, arrange training, validation, and test data as follows.

images/
    train/
        airplanes/
            airplane001.jpg
            airplane002.jpg
            ...
        watch/
            watch001.jpg
            watch002.jpg
            ...
    valid/
        airplanes/
            airplane101.jpg
            airplane102.jpg
            ...
        watch/
            watch101.jpg
            watch102.jpg
            ...
    test/
        airplanes/
            airplane201.jpg
            airplane202.jpg
            ...
        watch/
            watch201.jpg
            watch202.jpg
            ...

2. Edit config.yml

Edit config.yml as you like.

For example

common:
  num_threads: 4
  gpus: 0

data:
  quality: 100
  shuffle: 1
  center_crop: 0

finetune:
  models:
    - imagenet1k-resnet-50
  optimizers:
    - sgd
  num_epochs: 30
  lr: 0.0001
  lr_factor: 0.1
  lr_step_epochs: 10,20
  mom: 0.9
  wd: 0.00001
  batch_size: 10

Please see common settings on how to run with GPU image on NVIDIA GPUs.

3. Do Fine-tuning

$ docker-compose run finetuner

mxnet-finetuner will automatically execute the followings according to config.yml.

• Create RecordIO data from images
• Download pretrained models
• Replace the last fully-connected layer with a new one that outputs the desired number of classes
• Data augumentaion
• Do Fine-tuning
• Make training accuracy/loss graph
• Make confusion matrix
• Upload training accuracy/loss graph and confusion matrix to Slack

Training accuracy/loss graph and/or confusion matrix are save at logs/ directory.
Trained models are save at model/ directory.

Trained models are saved with the following file name for each epoch.

model/201705292200-imagenet1k-nin-sgd-0000.params

If you want to upload results to Slack, set SLACK_API_TOKEN environment variable and edit config.yml as below.

finetune:
  train_accuracy_graph_slack_upload: 1
  train_loss_graph_slack_upload: 1
test:
  confusion_matrix_slack_upload: 1

4. Predict with trained models

Select the trained model and epoch you want to use for testing and edit config.yml

If you want to use model/201705292200-imagenet1k-nin-sgd-0001.params, edit config.yml as blow.

test:
  model: 201705292200-imagenet1k-nin-sgd-0001

When you want to use the latest highest validation accuracy trained model, edit config.yml as below.

test:
  use_latest: 1

If set this option, model is ignored.

When you are done, you can predict with the following command

$ docker-compose run finetuner test

Predict result and classification report and/or confusion matrix are save at logs/ directory.

Available pretrained models

To use these pretrained models, specify the following pretrained model name in config.yml.

For details, please check Available pretrained models

Available optimizers

• SGD
• NAG
• RMSProp
• Adam
• AdaGrad
• AdaDelta
• Adamax
• Nadam
• DCASGD
• SGLD
• Signum
• FTML
• Ftrl

To use these optimizers, specify the optimizer name in lowercase in config.yml.

Benchmark (Speed and Memory Footprint)

Single TITAN X (Maxwell) with batch size 40

For details, please check Benchmark

Utilities

util/counter.sh

Count the number of files in each subdirectory.

$ util/counter.sh testdir
testdir contains 4 directories
Leopards    197
Motorbikes  198
airplanes   199
watch       200

util/move_images.sh

Move the specified number of jpeg images from the target directory to the output directory while maintaining the directory structure.

$ util/move_images.sh 20 testdir newdir
processing Leopards
processing Motorbikes
processing airplanes
processing watch
$ util/counter.sh newdir
newdir contains 4 directories
Leopards    20
Motorbikes  20
airplanes   20
watch       20

Prepare sample images for fine-tuning

Download Caltech 101 dataset, and split part of it into the example_images directory.

$ util/caltech101_prepare.sh

• example_images/train is train set of 60 images for each classes
• example_images/valid is validation set of 20 images for each classes
• example_imags/test is test set of 20 images for each classes

$ util/counter.sh example_images/train
example_images/train contains 10 directories
Faces       60
Leopards    60
Motorbikes  60
airplanes   60
bonsai      60
car_side    60
chandelier  60
hawksbill   60
ketch       60
watch       60

With this data you can immediately try fine-tuning.

$ util/caltech101_prepare.sh
$ rm -rf images
$ mv exmaple_images images
$ docker-compose run finetuner

Misc

How to freeze layers during fine-tuning

If you set the number of target layer to finetune.num_active_layers in config.yml as below, only layers whose number is not greater than the number of the specified layer will be train.

finetune:
  models:
    - imagenet1k-nin
  optimizers:
    - sgd
  num_active_layers: 6

The default for finetune.num_active_layers is 0, in which case all layers are trained.

If you set 1 to finetune.num_active_layers, only the last fully-connected layers are trained.

You can check the layer numbers of various pretrained models with num_layers command.

$ docker-compose run finetuner num_layers <pretrained model name>

For details, please check How to freeze layers during fine-tuning

Training from scratch

Edit config.yml as below.

finetune:
  models:
    - scratch-alexnet

You can also run fine-tuning and training from scratch together.

finetune:
  models:
    - imagenet1k-inception-v3
    - scratch-inception-v3

For details, please check Available models training from scratch

Averaging ensemble test with trained models

You can do averaging ensemble test using multiple trained models.

If you want to use the following the three trained models,

• model/20180130074818-imagenet1k-nin-nadam-0003.params
• model/20180130075252-imagenet1k-squeezenet-nadam-0003.params
• model/20180131105109-imagenet1k-caffenet-nadam-0003.params

edit config.yml as blow.

ensemble:
  models:
    - 20180130074818-imagenet1k-nin-nadam-0003
    - 20180130075252-imagenet1k-squeezenet-nadam-0003
    - 20180131105109-imagenet1k-caffenet-nadam-0003

When you are done, you can do averaging ensemble test with the following command.

$ docker-compose run finetuner ensemble test

If you want to use validation dataset, do as follows.

$ docker-compose run finetuner ensemble valid

Averaging ensemble test result and classification report and/or confusion matrix are save at logs/ directory.

Export your trained model

You can export your trained model in a format that can be used with Model Server for Apache MXNet as follows.

$ docker-compose run finetuner export

The exported file (extension is .model) is saved at model/ directory.

Please check export settings for export settings.

Serve your exported model

You can serve your exported model as API server.

With the following command, launch the API server with the last exported model using pre-configured Docker image of Model Server for Apache MXNet.

$ docker-compose up -d mms

The API server is started at port 8080 of your local host.

Then you will curl a POST to the MMS predict endpoint with the test image. (For exmple, use airlane.jpg).

$ curl -X POST http://127.0.0.1:8080/model/predict -F "data=@airplane.jpg"

The predict endpoint will return a prediction response in JSON. It will look something like the following result:

{
  "prediction": [
    [
      {
        "class": "airplane",
        "probability": 0.9950716495513916
      },
      {
        "class": "watch",
        "probability": 0.004928381647914648
      }
    ]
  ]
}

Try image classification with jupyter notebook

$ docker-compose run --service-ports finetuner jupyter

Please note that the --service-port option is required

Replace the IP address of the displayed URL with the IP address of the host machine and access it from the browser.

Open the classify_example/classify_example.ipynb and try out the image classification sample using the VGG-16 pretrained model pretrained with ImageNet.

config.yml options

common settings

common:
  num_threads: 4
  gpus: 0  # list of gpus to run, e.g. 0 or 0,2,5.

If a machine has one or more GPU cards installed, then each card is labeled by a number starting from 0. To use GPU for training or inference, specify GPU number in common.gpus.

If you do not use the GPU or you can not use it, please comment out common.gpus.

In the environment where GPU can not be used, common.gpus in config.yml generated by setup.sh is automatically commented out.

data settings

train, validation and test RecordIO data generation settings.

mxnet-finetuner resize and pack the image files into a recordIO file for increased performance.

By setting the resize_short, you can resize shorter edge of images to that size.

If resize_short is not specified, it is automatically determined according to the model you are using.

data:
  quality: 100
  shuffle: 1
  center_crop: 0
  # test_center_crop: 1
  # resize_short: 256

finetune settings

finetune:
  models:  # specify models to use
    - imagenet1k-nin
    # - imagenet1k-inception-v3
    # - imagenet1k-vgg16
    # - imagenet1k-resnet-50
    # - imagenet11k-resnet-152
    # - imagenet1k-resnext-101
    # - imagenet1k-se-resnext-50
    # etc
  optimizers:  # specify optimizers to use
    - sgd
    # optimizers: sgd, nag, rmsprop, adam, adagrad, adadelta, adamax, nadam, dcasgd, signum, etc.
  # num_active_layers: 1  # train last n-layers without last fully-connected layer
  num_epochs: 10  # max num of epochs
  # load_epoch: 0  # specify when using user fine-tuned model
  lr: 0.0001  # initial learning rate
  lr_factor: 0.1  # the ratio to reduce lr on each step
  lr_step_epochs: 10  # the epochs to reduce the lr, e.g. 30,60
  mom: 0.9  # momentum for sgd
  wd: 0.00001  # weight decay for sgd
  batch_size: 10  # the batch size
  disp_batches: 10  # show progress for every n batches
  # top_k: 0  # report the top-k accuracy. 0 means no report.
  # data_aug_level: 3  # preset data augumentation level
  # random_crop: 0  # if or not randomly crop the image
  # random_mirror: 0  # if or not randomly flip horizontally
  # max_random_h: 0  # max change of hue, whose range is [0, 180]
  # max_random_s: 0  # max change of saturation, whose range is [0, 255]
  # max_random_l: 0  # max change of intensity, whose range is [0, 255]
  # max_random_aspect_ratio: 0  # max change of aspect ratio, whose range is [0, 1]
  # max_random_rotate_angle: 0  # max angle to rotate, whose range is [0, 360]
  # max_random_shear_ratio: 0  # max ratio to shear, whose range is [0, 1]
  # max_random_scale: 1  # max ratio to scale
  # min_random_scale: 1  # min ratio to scale, should >= img_size/input_shape. otherwise use --pad-size
  # rgb_mean: '123.68,116.779,103.939'  # a tuple of size 3 for the mean rgb
  # monitor: 0  # log network parameters every N iters if larger than 0
  # pad_size: 0  # padding the input image
  auto_test: 1  # if or not test with validation data after fine-tuneing is completed
  train_accuracy_graph_output: 1
  # train_accuracy_graph_fontsize: 12
  # train_accuracy_graph_figsize: 8,6
  # train_accuracy_graph_slack_upload: 1
  # train_accuracy_graph_slack_channels:
  #   - general
  train_loss_graph_output: 1
  # train_loss_graph_fontsize: 12
  # train_loss_graph_figsize: 8,6
  # train_loss_graph_slack_upload: 1
  # train_loss_graph_slack_channels:
  #   - general

data_aug_level

By setting the data_aug_level parameter, you can set the data augumentation settings collectively.

If data_aug_level is set, parameters related to data augumentation will be overwritten.

test settings

test:
  use_latest: 1  # Use last trained model. If set this option, model is ignored
  model: 201705292200-imagenet1k-nin-sgd-0001
  # model_epoch_up_to: 10  # test from epoch of model to model_epoch_up_to respectively
  test_batch_size: 10
  # top_k: 10
  # rgb_mean: '123.68,116.779,103.939'  # a tuple of size 3 for the mean rgb
  classification_report_output: 1
  # classification_report_digits: 3
  confusion_matrix_output: 1
  # confusion_matrix_fontsize: 12
  # confusion_matrix_figsize: 16,12
  # confusion_matrix_slack_upload: 1
  # confusion_matrix_slack_channels:
  #   - general

ensemble settings

# ensemble settings
ensemble:
  models:
    - 20180130074818-imagenet1k-nin-nadam-0003
    - 20180130075252-imagenet1k-squeezenet-nadam-0003
    - 20180131105109-imagenet1k-caffenet-nadam-0003
  # weights: 1,1,1
  ensemble_batch_size: 10
  # top_k: 10
  # rgb_mean: '123.68,116.779,103.939'  # a tuple of size 3 for the mean rgb
  classification_report_output: 1
  # classification_report_digits: 3
  confusion_matrix_output: 1
  # confusion_matrix_fontsize: 12
  # confusion_matrix_figsize: 16,12
  # confusion_matrix_slack_upload: 1
  # confusion_matrix_slack_channels:
  #   - general

export settings

# export settings
export:
  use_latest: 1 # Use last trained model. If set this option, model is ignored
  model: 201705292200-imagenet1k-nin-sgd-0001
  # top_k: 10  # report the top-k accuracy
  # rgb_mean: '123.68,116.779,103.939'  # a tuple of size 3 for the mean rgb
  # center_crop: 1  # if or not center crop at image preprocessing
  # model_name: model

Acknowledgement

• MXNet
• Mo - Mustache Templates in Bash
• A MXNet implementation of DenseNet with BC structure
• SENet.mxnet
• Model Server for Apache MXNet

Licnese

Apache-2.0 license.