Uncertainty estimation for deep supervised models

My master thesis project.

Project plan (in Croatian)

Related (uncertainty, bayesian neural networks, out-of-distribution and misclassified examples)

• MacKay
• neal95 Bayesian deep learning for neural networks
• gal15arxiv Dropout as a Bayesian Approximation
• gal16cam Uncertainty in Deep Learning
• kendall17arxiv What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision
• smith18arxiv What Understanding Measures of Uncertainty for Adversarial Example Detection
• hendryks2017arxiv A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks
• guo17arxiv On Calibration of Modern Neural Networks
• liang18iclr Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks
• devries18arxiv Learning Confidence for Out-of-Distribution Detection in Neural Networks
• lee18iclr Training Confidence-calibrated Classifiers for Detecting Out-of-Distribution Samples

Code structure

.
├── data  # non-code
│   ├── nets                # trained model parameters
│   └── plots
├── dl_uncertainty          # library
│   ├── data
│   │   ├── data.py         # `Dataset`, `DataLoader`
│   │   ├── datasets.py     # raw `Dataset` implementations
│   ├── ioutils
│   ├── models
│   │   ├── components.py   # abstractions/wrappers of `tf_utils/layers.py` with an experimantal programming design
│   │   ├── modeldef.py     # `InferenceComponent` (model graph), `TrainingComponent` (inductive bias), `ModelDef` ((almost) all information that needs to be given to the `Model` inintializer)
│   │   ├── model.py        # `Model` with methods such as `train`, `test`, `predict`, `save_state`, `load_state`, `load_parameters`
│   │   └── tf_utils
│   │       ├── evaluation.py
│   │       ├── layers.py   # functional core implementtions of layers and larger model components
│   │       ├── losses.py
│   │       ├── regularization.py
│   │       ├── update_ops.py
│   │       └── variables.py
│   ├── processing          # data augmentation, preprocessing, ...
│   ├── utils               # visualization, ...
│   ├── data_utils.py       # dataset factories `get_dataset` and `get_cached_dataset_with_normalized_inputs`
│   ├── dirs.py  
│   ├── evaluation.py
│   ├── model_utils.py      # factories `get_inference_component`, `get_training_component`, model factory `get_model`
│   ├── parameter_loading.py
│   └── training.py         # `train`, training loop function
└── scripts
    ├── _context.py
    ├── test.py             # e.g. python test.py cifar wrn 28 10 ../data/nets/cifar-trainval/wrn-28-10-e100/2018-05-28-0956/Model
    ├── train.py            # e.g. python train.py cifar rn 18 64 --epochs 100 --trainval
    └── view_dataset.py     # e.g. python view_dataset.py cifar trainval

Prerequisites

dl_uncertainty/dirs.py requires that directories matching the following regular expressions exist ('.' and '/' are not escaped for readability).

.(/..)+/data/cache       # some ancestor directory contains `data/cache`
.(/..)+/data/datasets    # some ancestor directory contains `data/datasets` ...
.(/..)+/data/pretrained_parameters
.(/..)+/data/nets

Dataset directory

The <ancestor>/data/datasets directory (stored in the dirs.DATASETS variable in dl_uncertainty/dirs.py) needs to contain the required datasets. For example, for CIFAR-10 it needs to contain this directory subtree:

cifar-10-batches-py/
├── batches.meta
├── data_batch_1
├── data_batch_2
├── data_batch_3
├── data_batch_4
├── data_batch_5
└── test_batch

Preprocessed data caching directory

The <ancestor>/data/cache directory is used for caching preprocessed data in the get_cached_dataset_with_normalized_inputs factory function in dl_uncertainty/data_utils.py (data_utils.LazyNormalizer stores average image channel values and data.Dataset.cache_hdd_only stores preprocessed data). python scripts/clear_cache.py <dataset_id> can be used to clear cache for a dataset.

Trained model parameters directory

The <ancestor>/data/nets directory is used storing trained model parameters and training log data.

Usage

Training

To train a model, the script scripts/train.py needs to be run. An example command for training a model is

python train.py <dataset_id> <model_id> <model_depth> <model_width_parameter> --epochs <epoch_count> [--trainval]

Some examples of <dataset_id> are cifar, inaturalist, tinyimagenet. Other dataset IDs can be seen in dl_uncertainty/data_utils.py.

<model_id> <model_depth> <model_width_parameter> describe the general model structure. Some examples are

• ResNets (V2), where <model_width_parameter> represents the number of feature maps of the first group of residual blocks rn 18 64 (ResNetV2 18), rn 50 64 (ResNetV2 50);
• wide ResNets, where <model_width_parameter> represents 1/16 of the number of feature maps of the first group of residual blocks: wrn 28 10 (WRN-28-10), wrn 16 4 (WRN-16-4);
• DenseNets (BC variant) where <model_width_parameter> represents the growth rate (k): dn 100 12, dn 121 32.

If the optional --trainval argument is passed, the model is trained on the trainval set and tested on the test set. Otherwise, it is trained on the train set and tested on the val set.

When the training is finished the path of the trained model (a path ending with "/Model") is saved is printed on the screen. It can be copied for use in other scripts, like test.py.

Testing

To test a trained model on some new dataset scripts/test.py needs to be run. An example command for testing is

python test.py <dataset_id> <model_id> <model_depth> <model_width_parameter> <path_to_model>

<dataset_id> is the ID of the dataset that the model was trained on. <path_to_model> is the path to the saved parameters of the model (the path ending with "/Model)).

TODO