3D-DenseNet with TensorFlow (Train in AWS sagemaker)

Get started !!

Dependencies

• Tensorflow 1.11
• python 3.6.5
• opencv-python 3.4.3.18
• Pillow 5.3.0
• sagemaker 1.15.2

Data preparation

1. Download the video dataset and make sure it has the following folder structure (../video/<action_name>/<video1.avi> KTH ex: ../kth_video/boxing/person01_boxing_d1_uncomp.avi)
2. Run the prepare_data_main.py. You need to specify the data_dir, train_output_path, and eval_output_path.

• data_dir: ../kth_video

3. When the script finished. It will print out the following informations

• Total clips in train dataset: AAAA (Take a record of this number)
• Total clips in eval dataset: BBBB (Take a record of this number)

Train (Local)

1. Paste the number AAAA from previous step to train_total_video_clip in the debug_train.py file.
2. Paste the number BBBB from previous step to eval_total_video_clip in the debug_train.py file.
3. Copy and paste the eval.tfrecord and train.tfrecord file generated from the previous step to a folder named ../tfrecord.
4. Set the DATA_DIR in the debug_train.py to the proper folder name in the previous step.
5. Run python debug_train.py (Make sure you have all the dependencies).

Train (AWS sagemaker)

Note: It turns out Sagemaker doesn't support python3 for Tensorflow script at this moment (2018.Nov.1)!!! So I will stop working on this part and left the sagemaker_main.template file as it is for now.

1. Register AWS account. AWS Console
2. Create an IAM user with only Programmatic access and attached AmazonS3FullAccess and AmazonSageMakerFullAccess to this IAM user. Keep a record of your Access Key ID and Secret Access Key (Don't tell anyone this information!!! Even your husband/wife).
3. Install boto3 on your local desktop. Run aws configure in your console and paste the Access Key ID and Secret Access Key from previous step. Keep in mind the region (ex: us-west-2) that you used.
4. Create a new Role with name sagemaker-full-access-role and attach an inline policy with the following JSON
5. Create a new S3 bucekt with whatever name you want in the same region in Step3. Let said the S3 bucket name is machine_leaning_data_bucket.
6. Rename the sagemaker_main.template to sagemaker_main.py
7. Copy the new Role ARN (ex: arn:aws:iam::<aws_account_id>:role/sagemaker-full-access-role) and paste it to the role value in the sagemaker_main.py
8. Replace the <s3_bucket_name> in sagemaker_main.py with S3 bucket name machine_leaning_data_bucket (Whatever S3 bucket name you have).
9. Chooes one option in the sagemaker_main.py and run python sagemaker_main.py. Notice that if you choose

As I said at the beginning, sagemaker doesn't support tensorflow docker image with python version 3, so you will get error Attempted relative import in non-package at this moment. I will try to rework this file once sagemaker support it.

Background

Expand the `Densely Connected Convolutional Networks DenseNets to 3D-DenseNet for action recognition (video classification):

• 3D-DenseNet - without bottleneck layers
• 3D-DenseNet-BC - with bottleneck layers

Each model can be tested on such datasets:

• KTH
• MERL

A number of layers, blocks, growth rate, video normalization and other training params may be changed trough shell or inside the source code.

There are also many other implementations, they may be useful also.

Reference

Thesis