https://www.kaggle.com/c/whale-categorization-playground

Categorizer for this Kaggle competition. To be used as part of my personal project website whale game 'computer' player.

Requires: Python 3+ as use type hints

Folder Structure:

• Data: where the Pascal XML files with annotations are located. conversion.py assumes that this is where the directory is
• faster_rcnn_resnet101_coco_2018_01_28: holds the initial model configurations and checkpoints. This is where the config file points to so it knows where to start off from

Major Resources Used (other than documentation from pandas, tensorflow and numpy):

• https://github.com/tensorflow/models/tree/master/research/object_detection
• https://towardsdatascience.com/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9
• https://github.com/priya-dwivedi/Deep-Learning/tree/master/tensorflow_toy_detector
• https://towardsdatascience.com/building-a-toy-detector-with-tensorflow-object-detection-api-63c0fdf2ac95
• https://github.com/datitran/raccoon_dataset

General Steps: 1 Download labelimg to manually annonate the images that had whale flukes in it 2 labelimg outputs a type of file (XML) that holds all the necessary data but needs to get converted to a TFrecord format (manual step using LabelImg) 3 We do this by first translating them to CSV files / dataframe objects(see xml_to_csv.py then taking these and converting them to TFrecords (conversion.py). Don't have to save to CSV file. Only do it if you want a nice readable CSV of all the images and their corresponding details. (This work is done in conversion.py and then user interacts with it through run_converter.py via command line to ensure that the relevant images are converted to TFRecord). NOTE: this method creates a subdir conversion.output_path within this home directory where it puts the TFRecord files and any CSV files 4 Then train the model using: https://github.com/tensorflow/models/blob/master/research/object_detection/train.py#L17 -Note: git cloned https://github.com/tensorflow/models -Note: I cloned the repo to some other directory not within this 'Whales' directory as I wanted it to be easily accessible to other projects. And thus in order to get this to work on your machine as described you'd have to clone the same repo onto your computer and follow the installation steps accordingly -Note: then followed installation instructions - https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md -Note: To run the protobuf compilation command: protoc object_detection/protos/*.proto --python_out=. found here: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md I had to make sure my conda environment that I installed all of the necessary dependencies was activated before being able to successfully compile protobuf -Note: Running this to add to path (at least on Ubuntu): https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md#add-libraries-to-pythonpath is required before the test installation step will work. Or else if you don't add it when you open a new terminal the test step will give you an error. Make sure you are in the research directory and if you have closed terminals then you have to rerun it again or else running train.py (in the object_detection dir) won't work and get an error like 'no module named object_detection found'. Also, key that you run that adding path code in the research directory. Seemingly if you do it anywhere else (ex. i did it in object_detection dir) it won't do the job Decided not to wrap it with my own command line function. user then runs train.py in the terminal with the appriorate flags see below: 5 Run the eval.py code

ex.) For the 'simple usage' do not have a good example of the 2nd usage yet as don't understand the last input of 'input_config_path' and where that comes from as see 'model_config' and 'train_config' sections within the pipeline config file (we use faster_rcnn_resnet101_coco.config) but don't see 'input_config_path'

python train.py --logtostderr --train_dir=path/to/train_dir --pipeline_config_path=pipeline_config.pbtxt

---this was my actual command within the object detection directory: python train.py --logtostderr --train_dir=/home/harrison/Whales/training/ --pipeline_config_path=/home/harrison/Whales/training/faster_rcnn_resnet101_coco.config

---below is my tensorboard command also while in the object_detection directory tensorboard --logdir=/home/harrison/Whales/training

Explaining the above.

• run this in the directory of objection_detection in the models repo
• for me I had to put the absolute path (not using tilda, but the actual path) because my training directory and pipeline config path were not within the object_detection directory
• note had to apply this fix: tensorflow/models#3705 (comment) the first part, not the fix to the config file --train_dir value is the directory where all the training output for the model will go, the actual model graph file and any tensorboard related files. So for us we have 'training/'. This is the directory that we point to when we run tensorboard and want to see the training progress - (tensorboard --logdir=training/) -- pipeline_config_path value is the path to the 'TrainEvalPipelineConfig configuration file' that we use to config the file will whatever preliminary settings we want. This file also tells the model how many labels it needs to classify, where to save model checkpoints, and where to find the data to be used for testing and validation ->i.e note that when calling train.py we do not tell it where the model can find the training and testing data that's because that info is in the config file itself (in this project I use the faster_rcnn_resnet101_coco.config under the faster_rcnn_resnet101_coco_2018_01_28 directory) -- Note: For me the train_dir path had to be directed to the directory within 'Whales' which was in another directoy as well as the pipeline path which had to point to the faster_rcnn_resnet101_coco_2018_01_28 directory

---running eval.py while in object_detection directory: see: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_locally.md#running-the-evaluation-job python eval.py --logtostderr --checkpoint_dir=/home/harrison/Whales/training/ --pipeline_config_path=/home/harrison/Whales/training/faster_rcnn_resnet101_coco.config --eval_dir=/home/harrison/Whales/training/

Explaining above:

• checkpoint_dir is where the train.py is outputting all of its model related checkpoints and output data while training
• pipeline_config_path is the same path to the config file used in the train.py file above (ie the config file that is customized for our purpose with paths to specific labels and train and test record files)
• eval_dir is the directory where the eval.py outputs its 'evaluation data' that will be used for tensorboard to display in the tensorboard UI. I.e whatever directory you put in this when you run tensorboard on the evaluation data you should have the tensorboard --logdir flag match (for ease you can just make eval_dir == the train_dir so that you can just run one tensorboard instance - I think that works that way?):

tensorboard --logdir=

Other notes:

• note the ft.summary.image code that allows tensorboard to visualize your model on an actual image is in the eval_util file. Doesn't exist in the trainer.py file. I guess only makes sense to visualize on an image when doing validation