ml_glaucoma

Originally this repo started off as a custom CNN for glaucoma diagnoses, but has since expanded into something more.

This repo is no longer maintained, but has been split into (at least):

ml-prepare
ml-params
ml-params-keras
ml-params-tensorflow

Install dependencies

pip install -r requirements.txt

Install package

pip install .

CLI usage

$ python -m ml_glaucoma --help

usage: python -m ml_glaucoma [-h] [--version]
                         {download,vis,train,evaluate,parser,info} ...

CLI for a Glaucoma diagnosing CNN

positional arguments:
  {download,vis,train,evaluate,parser,info}
    download            Download and prepare required data
    vis                 Visualise data
    train               Train model
    evaluate            Evaluate model
    parser              Parse out metrics from log output. Default: per epoch
                        sensitivity & specificity.
    info                Info subcommand

optional arguments:
  -h, --help            show this help message and exit
  --version             show program's version number and exit

`download`

$ python -m ml_glaucoma download --help

usage: python -m ml_glaucoma download [-h]
                                  [-ds {bmes,refuge} [{bmes,refuge} ...]]
                                  [--data_dir DATA_DIR]
                                  [--download_dir DOWNLOAD_DIR]
                                  [--extract_dir EXTRACT_DIR]
                                  [--manual_dir MANUAL_DIR]
                                  [--download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}]
                                  [-r RESOLUTION RESOLUTION]
                                  [--gray_on_disk] [--bmes_init]
                                  [--bmes_parent_dir BMES_PARENT_DIR]

optional arguments:
  -h, --help            show this help message and exit
  -ds {bmes,refuge} [{bmes,refuge} ...], --dataset {bmes,refuge} [{bmes,refuge} ...]
                        dataset key
  --data_dir DATA_DIR   root directory to store processed tfds records
  --download_dir DOWNLOAD_DIR
                        directory to store downloaded files
  --extract_dir EXTRACT_DIR
                        directory where extracted files are stored
  --manual_dir MANUAL_DIR
                        directory where manually downloaded files are saved
  --download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}
                        tfds.GenerateMode
  -r RESOLUTION RESOLUTION, --resolution RESOLUTION RESOLUTION
                        image resolution
  --gray_on_disk        whether or not to save data as grayscale on disk
  --bmes_init           initial bmes get_data
  --bmes_parent_dir BMES_PARENT_DIR
                        parent directory of bmes data

`vis`

$ python -m ml_glaucoma vis --help

usage: python -m ml_glaucoma vis [-h] [-ds {bmes,refuge} [{bmes,refuge} ...]]
                                 [--data_dir DATA_DIR]
                                 [--download_dir DOWNLOAD_DIR]
                                 [--extract_dir EXTRACT_DIR]
                                 [--manual_dir MANUAL_DIR]
                                 [--download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}]
                                 [-r RESOLUTION RESOLUTION] [--gray_on_disk]
                                 [--bmes_init]
                                 [--bmes_parent_dir BMES_PARENT_DIR] [-fv]
                                 [-fh] [--gray]
                                 [-l {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}]
                                 [-m [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]]]
                                 [-pt [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]]]
                                 [-rt [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]]]
                                 [--shuffle_buffer SHUFFLE_BUFFER]
                                 [--use_inverse_freq_weights]

optional arguments:
  -h, --help            show this help message and exit
  -ds {bmes,refuge} [{bmes,refuge} ...], --dataset {bmes,refuge} [{bmes,refuge} ...]
                        dataset key
  --data_dir DATA_DIR   root directory to store processed tfds records
  --download_dir DOWNLOAD_DIR
                        directory to store downloaded files
  --extract_dir EXTRACT_DIR
                        directory where extracted files are stored
  --manual_dir MANUAL_DIR
                        directory where manually downloaded files are saved
  --download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}
                        tfds.GenerateMode
  -r RESOLUTION RESOLUTION, --resolution RESOLUTION RESOLUTION
                        image resolution
  --gray_on_disk        whether or not to save data as grayscale on disk
  --bmes_init           initial bmes get_data
  --bmes_parent_dir BMES_PARENT_DIR
                        parent directory of bmes data
  -fv, --maybe_vertical_flip
                        randomly flip training input vertically
  -fh, --maybe_horizontal_flip
                        randomly flip training input horizontally
  --gray                use grayscale
  -l {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}, --loss {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}
                        loss function to use
  -m [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]], --metrics [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]]
                        metric functions to use
  -pt [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]], --precision_thresholds [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]]
                        precision thresholds
  -rt [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]], --recall_thresholds [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]]
                        recall thresholds
  --shuffle_buffer SHUFFLE_BUFFER
                        buffer used in tf.data.Dataset.shuffle
  --use_inverse_freq_weights
                        weight loss according to inverse class frequency

`train`

$ python -m ml_glaucoma train --help

usage: python -m ml_glaucoma train [-h]
                                   [-ds {bmes,refuge} [{bmes,refuge} ...]]
                                   [--data_dir DATA_DIR]
                                   [--download_dir DOWNLOAD_DIR]
                                   [--extract_dir EXTRACT_DIR]
                                   [--manual_dir MANUAL_DIR]
                                   [--download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}]
                                   [-r RESOLUTION RESOLUTION] [--gray_on_disk]
                                   [--bmes_init]
                                   [--bmes_parent_dir BMES_PARENT_DIR] [-fv]
                                   [-fh] [--gray]
                                   [-l {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}]
                                   [-m [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]]]
                                   [-pt [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]]]
                                   [-rt [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]]]
                                   [--shuffle_buffer SHUFFLE_BUFFER]
                                   [--use_inverse_freq_weights]
                                   [--model_file [MODEL_FILE [MODEL_FILE ...]]]
                                   [--model_param [MODEL_PARAM [MODEL_PARAM ...]]]
                                   [-o {Adadelta,Adagrad,Adam,Adamax,Ftrl,Nadam,Optimizer,RMSprop,SGD}]
                                   [-lr LEARNING_RATE]
                                   [--optimizer_params OPTIMIZER_PARAMS]
                                   [--exp_lr_decay EXP_LR_DECAY]
                                   [-b BATCH_SIZE] [-e EPOCHS]
                                   [--class-weight CLASS_WEIGHT]
                                   [--callback [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} ...]]]
                                   [--model_dir MODEL_DIR]
                                   [-c CHECKPOINT_FREQ]
                                   [--summary_freq SUMMARY_FREQ]
                                   [-tb TB_LOG_DIR] [--write_images]
                                   [--seed SEED] [--disable-gpu]
                                   [--continuous] [--delete-lt DELETE_LT]
                                   [--model-dir-autoincrement MODEL_DIR_AUTOINCREMENT]

optional arguments:
  -h, --help            show this help message and exit
  -ds {bmes,refuge} [{bmes,refuge} ...], --dataset {bmes,refuge} [{bmes,refuge} ...]
                        dataset key
  --data_dir DATA_DIR   root directory to store processed tfds records
  --download_dir DOWNLOAD_DIR
                        directory to store downloaded files
  --extract_dir EXTRACT_DIR
                        directory where extracted files are stored
  --manual_dir MANUAL_DIR
                        directory where manually downloaded files are saved
  --download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}
                        tfds.GenerateMode
  -r RESOLUTION RESOLUTION, --resolution RESOLUTION RESOLUTION
                        image resolution
  --gray_on_disk        whether or not to save data as grayscale on disk
  --bmes_init           initial bmes get_data
  --bmes_parent_dir BMES_PARENT_DIR
                        parent directory of bmes data
  -fv, --maybe_vertical_flip
                        randomly flip training input vertically
  -fh, --maybe_horizontal_flip
                        randomly flip training input horizontally
  --gray                use grayscale
  -l {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}, --loss {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}
                        loss function to use
  -m [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]], --metrics [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]]
                        metric functions to use
  -pt [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]], --precision_thresholds [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]]
                        precision thresholds
  -rt [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]], --recall_thresholds [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]]
                        recall thresholds
  --shuffle_buffer SHUFFLE_BUFFER
                        buffer used in tf.data.Dataset.shuffle
  --use_inverse_freq_weights
                        weight loss according to inverse class frequency
  --model_file [MODEL_FILE [MODEL_FILE ...]]
                        gin files for model definition. Should define
                        `model_fn` macro either here or in --gin_param
  --model_param [MODEL_PARAM [MODEL_PARAM ...]]
                        gin_params for model definition. Should define
                        `model_fn` macro either here or in --gin_file
  -o {Adadelta,Adagrad,Adam,Adamax,Ftrl,Nadam,Optimizer,RMSprop,SGD}, --optimizer {Adadelta,Adagrad,Adam,Adamax,Ftrl,Nadam,Optimizer,RMSprop,SGD}
                        class name of optimizer to use
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        base optimizer learning rate
  --optimizer_params OPTIMIZER_PARAMS
                        Extra optimiser args, e.g.: '{epsilon: 1e-7, amsgrad:
                        true}'
  --exp_lr_decay EXP_LR_DECAY
                        exponential learning rate decay factor applied per
                        epoch, e.g. 0.98. None is interpreted as no decay
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        size of each batch
  -e EPOCHS, --epochs EPOCHS
                        number of epochs to run training from
  --class-weight CLASS_WEIGHT
                        Optional dictionary mapping class indices (integers)to
                        a weight (float) value, used for weighting the loss
                        function(during training only).This can be useful to
                        tell the model to"pay more attention" to samples
                        froman under-represented class.
  --callback [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} ...]]
                        Keras callback function(s) to use. Extends default
                        callback list.
  --model_dir MODEL_DIR
                        model directory in which to save weights and
                        tensorboard summaries
  -c CHECKPOINT_FREQ, --checkpoint_freq CHECKPOINT_FREQ
                        epoch frequency at which to save model weights
  --summary_freq SUMMARY_FREQ
                        batch frequency at which to save tensorboard summaries
  -tb TB_LOG_DIR, --tb_log_dir TB_LOG_DIR
                        tensorboard_log_dir (defaults to model_dir)
  --write_images        whether or not to write images to tensorboard
  --seed SEED           Set the seed, combine with `--disable-gpu` to disable
                        GPU for added determinism
  --disable-gpu         Set the seed, combine with `--disable-gpu` to disable
                        GPU for added determinism
  --continuous          after each successful train, run again
  --delete-lt DELETE_LT
                        delete *.h5 files that are less than this threshold
  --model-dir-autoincrement MODEL_DIR_AUTOINCREMENT
                        autoincrement rather than overwrite the model dir
                        (when --continuous is set)

`evaluate`

$ python -m ml_glaucoma evaluate --help

usage: python -m ml_glaucoma evaluate [-h]
                                      [-ds {bmes,refuge} [{bmes,refuge} ...]]
                                      [--data_dir DATA_DIR]
                                      [--download_dir DOWNLOAD_DIR]
                                      [--extract_dir EXTRACT_DIR]
                                      [--manual_dir MANUAL_DIR]
                                      [--download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}]
                                      [-r RESOLUTION RESOLUTION]
                                      [--gray_on_disk] [--bmes_init]
                                      [--bmes_parent_dir BMES_PARENT_DIR]
                                      [-fv] [-fh] [--gray]
                                      [-l {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}]
                                      [-m [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]]]
                                      [-pt [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]]]
                                      [-rt [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]]]
                                      [--shuffle_buffer SHUFFLE_BUFFER]
                                      [--use_inverse_freq_weights]
                                      [--model_file [MODEL_FILE [MODEL_FILE ...]]]
                                      [--model_param [MODEL_PARAM [MODEL_PARAM ...]]]
                                      [-o {Adadelta,Adagrad,Adam,Adamax,Ftrl,Nadam,Optimizer,RMSprop,SGD}]
                                      [-lr LEARNING_RATE]
                                      [--optimizer_params OPTIMIZER_PARAMS]
                                      [--exp_lr_decay EXP_LR_DECAY]
                                      [-b BATCH_SIZE] [-e EPOCHS]
                                      [--class-weight CLASS_WEIGHT]
                                      [--callback [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} ...]]]
                                      [--model_dir MODEL_DIR]
                                      [-c CHECKPOINT_FREQ]
                                      [--summary_freq SUMMARY_FREQ]
                                      [-tb TB_LOG_DIR] [--write_images]
                                      [--seed SEED] [--disable-gpu]
                                      [--continuous] [--delete-lt DELETE_LT]
                                      [--model-dir-autoincrement MODEL_DIR_AUTOINCREMENT]

optional arguments:
  -h, --help            show this help message and exit
  -ds {bmes,refuge} [{bmes,refuge} ...], --dataset {bmes,refuge} [{bmes,refuge} ...]
                        dataset key
  --data_dir DATA_DIR   root directory to store processed tfds records
  --download_dir DOWNLOAD_DIR
                        directory to store downloaded files
  --extract_dir EXTRACT_DIR
                        directory where extracted files are stored
  --manual_dir MANUAL_DIR
                        directory where manually downloaded files are saved
  --download_mode {reuse_dataset_if_exists,reuse_cache_if_exists,force_redownload}
                        tfds.GenerateMode
  -r RESOLUTION RESOLUTION, --resolution RESOLUTION RESOLUTION
                        image resolution
  --gray_on_disk        whether or not to save data as grayscale on disk
  --bmes_init           initial bmes get_data
  --bmes_parent_dir BMES_PARENT_DIR
                        parent directory of bmes data
  -fv, --maybe_vertical_flip
                        randomly flip training input vertically
  -fh, --maybe_horizontal_flip
                        randomly flip training input horizontally
  --gray                use grayscale
  -l {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}, --loss {AdaBound,BinaryCrossentropy,BinaryCrossentropyWithRanking,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,DiceLoss,Hinge,Huber,JaccardDistance,KLD,KLDivergence,Kappa,LogCosh,MAE,MAPE,MSE,MSLE,MeanAbsoluteError,MeanAbsolutePercentageError,MeanSquaredError,MeanSquaredLogarithmicError,PairLoss,Poisson,Reduction,SmoothL1,SoftAUC,SparseCategoricalCrossentropy,SquaredHinge,Yogi,binary_crossentropy,categorical_crossentropy,categorical_hinge,cosine_similarity,hinge,kld,kullback_leibler_divergence,logcosh,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,serialize,sparse_categorical_crossentropy,squared_hinge}
                        loss function to use
  -m [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]], --metrics [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} [{AUC,Accuracy,BinaryAccuracy,BinaryCrossentropy,CategoricalAccuracy,CategoricalCrossentropy,CategoricalHinge,CosineSimilarity,FalseNegatives,FalsePositives,Hinge,KLD,KLDivergence,LogCoshError,MAE,MAPE,MSE,MSLE,Mean,MeanAbsoluteError,MeanAbsolutePercentageError,MeanIoU,MeanRelativeError,MeanSquaredError,MeanSquaredLogarithmicError,MeanTensor,Metric,Poisson,Precision,Recall,RootMeanSquaredError,SensitivityAtSpecificity,SparseCategoricalAccuracy,SparseCategoricalCrossentropy,SparseTopKCategoricalAccuracy,SpecificityAtSensitivity,SquaredHinge,Sum,TopKCategoricalAccuracy,TrueNegatives,TruePositives,binary_accuracy,binary_crossentropy,categorical_accuracy,categorical_crossentropy,hinge,kld,kullback_leibler_divergence,mae,mape,mean_absolute_error,mean_absolute_percentage_error,mean_squared_error,mean_squared_logarithmic_error,mse,msle,poisson,sparse_categorical_accuracy,sparse_categorical_crossentropy,sparse_top_k_categorical_accuracy,squared_hinge,top_k_categorical_accuracy} ...]]
                        metric functions to use
  -pt [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]], --precision_thresholds [PRECISION_THRESHOLDS [PRECISION_THRESHOLDS ...]]
                        precision thresholds
  -rt [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]], --recall_thresholds [RECALL_THRESHOLDS [RECALL_THRESHOLDS ...]]
                        recall thresholds
  --shuffle_buffer SHUFFLE_BUFFER
                        buffer used in tf.data.Dataset.shuffle
  --use_inverse_freq_weights
                        weight loss according to inverse class frequency
  --model_file [MODEL_FILE [MODEL_FILE ...]]
                        gin files for model definition. Should define
                        `model_fn` macro either here or in --gin_param
  --model_param [MODEL_PARAM [MODEL_PARAM ...]]
                        gin_params for model definition. Should define
                        `model_fn` macro either here or in --gin_file
  -o {Adadelta,Adagrad,Adam,Adamax,Ftrl,Nadam,Optimizer,RMSprop,SGD}, --optimizer {Adadelta,Adagrad,Adam,Adamax,Ftrl,Nadam,Optimizer,RMSprop,SGD}
                        class name of optimizer to use
  -lr LEARNING_RATE, --learning_rate LEARNING_RATE
                        base optimizer learning rate
  --optimizer_params OPTIMIZER_PARAMS
                        Extra optimiser args, e.g.: '{epsilon: 1e-7, amsgrad:
                        true}'
  --exp_lr_decay EXP_LR_DECAY
                        exponential learning rate decay factor applied per
                        epoch, e.g. 0.98. None is interpreted as no decay
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        size of each batch
  -e EPOCHS, --epochs EPOCHS
                        number of epochs to run training from
  --class-weight CLASS_WEIGHT
                        Optional dictionary mapping class indices (integers)to
                        a weight (float) value, used for weighting the loss
                        function(during training only).This can be useful to
                        tell the model to"pay more attention" to samples
                        froman under-represented class.
  --callback [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} [{AucRocCallback,BaseLogger,CSVLogger,Callback,EarlyStopping,ExponentialDecayLrSchedule,History,LambdaCallback,LearningRateScheduler,LoadingModelCheckpoint,ModelCheckpoint,ProgbarLogger,ReduceLROnPlateau,RemoteMonitor,SGDRScheduler,TensorBoard,TerminateOnNaN} ...]]
                        Keras callback function(s) to use. Extends default
                        callback list.
  --model_dir MODEL_DIR
                        model directory in which to save weights and
                        tensorboard summaries
  -c CHECKPOINT_FREQ, --checkpoint_freq CHECKPOINT_FREQ
                        epoch frequency at which to save model weights
  --summary_freq SUMMARY_FREQ
                        batch frequency at which to save tensorboard summaries
  -tb TB_LOG_DIR, --tb_log_dir TB_LOG_DIR
                        tensorboard_log_dir (defaults to model_dir)
  --write_images        whether or not to write images to tensorboard
  --seed SEED           Set the seed, combine with `--disable-gpu` to disable
                        GPU for added determinism
  --disable-gpu         Set the seed, combine with `--disable-gpu` to disable
                        GPU for added determinism
  --continuous          after each successful train, run again
  --delete-lt DELETE_LT
                        delete *.h5 files that are less than this threshold
  --model-dir-autoincrement MODEL_DIR_AUTOINCREMENT
                        autoincrement rather than overwrite the model dir
                        (when --continuous is set)

`parser`

You can pipe or include a filename.

$ python -m ml_glaucoma parser --help

usage: python -m ml_glaucoma parser [-h] [-d DIRECTORY]
                                    [--threshold THRESHOLD] [--top TOP]
                                    [--by-diff] [--tag TAG]
                                    [infile]

Show metrics from output. Default: per epoch sensitivity & specificity.

positional arguments:
  infile                File to work from. Defaults to stdin. So can pipe.

optional arguments:
  -h, --help            show this help message and exit
  -d DIRECTORY, --directory DIRECTORY
                        Directory. Searches here rather than infile.
  --threshold THRESHOLD
                        E.g.: 0.7 for sensitivity & specificity >= 70%
  --top TOP             Show top k results
  --by-diff             Sort by lowest difference between sensitivity &
                        specificity
  --tag TAG             Tag to filter by

Project Structure

Training/validation scripts are provided in data_preparation_scripts and each call a function defined in ml_glaucoma.runners. We aim to provide highly-configurable runs, but the main parts to consider are:

problem: the dataset, loss and metrics used during training
model_fn: the function that takes one or more tf.keras.layers.Inputs and returns a learnable keras model.

model_fns are configured using using a forked TF2.0 compatible gin-config (awaiting on this PR before reverting to the google version. See example configs in model_configs and the gin user guide.

Example usage:

python -m ml_glaucoma vis --dataset=refuge
python -m ml_glaucoma train \
  --model_file 'model_configs/dc.gin'  \
  --model_param 'import ml_glaucoma.gin_keras' 'dc0.kernel_regularizer=@tf.keras.regularizers.l2()' 'tf.keras.regularizers.l2.l = 1e-2' \
  --model_dir /tmp/ml_glaucoma/dc0-reg \
  -m BinaryAccuracy AUC \
  -pt 0.1 0.2 0.5 -rt 0.1 0.2 0.5 \
  --use_inverse_freq_weights
# ...
tensorboard --logdir=/tmp/ml_glaucoma

Tensorflow Datasets

The main Problem implementation is backed by tensorflow_datasets. This should manage dataset downloads, extraction, sha256 checks, on-disk shuffling/sharding and other best practices. Consequently it takes slightly longer to process initially, but the benefits in the long run are worth it.

BMES Initialization

The current implementation leverages the existing ml_glaucoma.utils.bmes_data_prep.get_data method to separate files. This uses tf.contrib so requires tf < 2.0. It can be run using the --bmes_init flag within python -m ml_glaucoma download. This must be run prior to the standard tfds.DatasetBuilder.download_and_prepare which is run automatically if necessary. Once the tfds files have been generated, the original get_data directories are no longer required.

If the test/train/validation split here is just a random split, this could be done more easily by creating a single tfds split and using tfds.Split.subsplit - see this post.

Status

Automatic model saving/loading via modified ModelCheckpoint.
Automatic tensorboard updates (fairly hacky interoperability with ModelCheckpoint to ensure restarted training runs have the appropriate step count).
Loss re-weighting according to inverse class frequency (TfdsProblem.use_inverse_freq_weights).
Only dc0, applications (Keras applications), efficientnet and squeeze_excite_resnet model verified to work. dr0, dc1, dc2, dc3 and other squeeze excite networks implemented but untested.
Only refuge and bmes dataset implemented, and only tested the classification task.
BMES dataset: currently requires 2-stage preparation: bmes_init which is based on ml_glaucoma.utils.bmes_data_prep.get_data and the standard tfds.DatasetBuilder.download_and_prepare. The first stage will only be run if --bmes_init is used in python -m ml_glaucoma download arguments.

License

Licensed under either of

Apache License, Version 2.0 (LICENSE-APACHE or https://www.apache.org/licenses/LICENSE-2.0)
MIT license (LICENSE-MIT or https://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

README.md

README.md

ml_glaucoma

Install dependencies

Install package

CLI usage

`download`

`vis`

`train`

`evaluate`

`parser`

Project Structure

Example usage:

Tensorflow Datasets

BMES Initialization

Status

License

Contribution

Files

README.md

Latest commit

History

README.md

File metadata and controls

ml_glaucoma

Install dependencies

Install package

CLI usage

download

vis

train

evaluate

parser

Project Structure

Example usage:

Tensorflow Datasets

BMES Initialization

Status

License

Contribution

`download`

`vis`

`train`

`evaluate`

`parser`