Visual-search-nets

Changed
- change `searchnets.test` so it saves a test results .csv for all datasets
[84](https://github.com/NickleDave/visual-search-nets/pull/84)

Fixed
- fix `searchnets.tensorboard.logdir2csv` so it doesn't duplicate part of save path
[82](https://github.com/NickleDave/visual-search-nets/pull/82)

Removed
- remove unnecessary dependencies from setup.py, fix 79
[2248c34](https://github.com/NickleDave/visual-search-nets/commit/2248c341635908994a11491df34f534501766eb8)

Added
- `transforms` sub-package [58](https://github.com/NickleDave/visual-search-nets/pull/58)
+ decouples transforms from datasets
+ documents transforms used with each combination of datasets and loss function, in `util.get_transforms` function
- transforms for VOC target that pick either largest class (based on bounding box)
or a random bounding box [59](https://github.com/NickleDave/visual-search-nets/pull/59)
- other `CORnet` models [66](https://github.com/NickleDave/visual-search-nets/pull/66)
- `tensorboard` module with functions for converting tensorboard events files to
Pandas `DataFrame`s and `.csv` files [69](https://github.com/NickleDave/visual-search-nets/pull/69)
- `analysis` sub-package with functions used to analyze results from
`searchstims` and `VSD` experiments
[71](https://github.com/NickleDave/visual-search-nets/pull/71)

Changed
- change `dataset`s so they return dictionaries, enabling more
flexibility in what items are present / used for training and testing
[67](https://github.com/NickleDave/visual-search-nets/pull/67)
- change how `AbstractTrainer` and `Tester` compute metrics
for the `VOCDetection` dataset;
now compute *all* metrics during validation and testing
[66](https://github.com/NickleDave/visual-search-nets/pull/66)
- remove `utils.munge` and `utils.metrics` modules, refactor
functions from them into `analysis` sub-package
[71](https://github.com/NickleDave/visual-search-nets/pull/71)
- make validation and checkpointing happen on steps, not epochs, so
they can happen more frequently and earlier in training
[73](https://github.com/NickleDave/visual-search-nets/pull/73)

Fixed
- now actually using the correct target corresponding to different loss
functions when training on `VOCDetection` dataset, e.g. `CE-largest`
uses `largest` from the batch dictionary
[66](https://github.com/NickleDave/visual-search-nets/pull/66)

Removed
- automatic setting of defaults for plotting, that were
in `plot.__init__` which made it hard to override them
[72](https://github.com/NickleDave/visual-search-nets/pull/72)
- factor out everything related to papers, making `searchnets` a separate library
[78](https://github.com/NickleDave/visual-search-nets/pull/78)

This is the version used for SfN 2019 poster, and for the paper
Added
- logging to a `tf.events` file [49](https://github.com/NickleDave/visual-search-nets/pull/49)

Fixed
- fix checkpoint saving [46](https://github.com/NickleDave/visual-search-nets/pull/46)
+ save "best" checkpoint as well as intermittent / last epoch checkpoint
+ don't save a separate model file

Changed
- switched to `torch` [33](https://github.com/NickleDave/visual-search-nets/pull/33)
- use `pyprojroot` [45](https://github.com/NickleDave/visual-search-nets/pull/33)
- clean up codebase [44](https://github.com/NickleDave/visual-search-nets/pull/33)
- rename `classes` subpackage to `engine` [48](https://github.com/NickleDave/visual-search-nets/pull/48)

This is the version used for presentation at SciPy 2019
Added
- functionality in `utils.results_csv` that computes d prime and adds it
to the .csv, as well as accuracy across both target present and target
absent conditions (i.e. what most people would call just "accuracy")
- single-source version
- summary results files and files with paths to training/validation/test
data are part of repository

Changed
- `figures.acc_v_set_size` re-written as more general `metric_v_set_size`,
works with d-prime metric and can plot accuracy, means, etc., for both
conditions (instead of always separating out target present and target
absent conditions into two lines)

This is the version used for paper submitted to ccneuro 2019
Added
- DOI badge to README.md
- `tf.data.dataset` pipeline to be able to use larger training data
- ability to 'shard' datasets to use even larger training data
- paper in ``./docs/ccneuro2019`
- scripts for making visual search stimuli in ``./src/bin`
- configs for new training (not fine-tuning) and training on multiple
stimuli
- ability to specify types of visual search stimuli to use when a single
run of `searchstims` places paths to all types in a single `json` file
+ using `stim_types` option in `config.ini` file
- ability to specify number of samples per (visual search stimulus)
"set size" in training, validation, and test sets
+ enables "balancing" data set
- sub-module for running learning curves (needs to be updated to use
additions to `searchstims.train`)

Fixed
- `searchnets.train` uses `MomentumOptimizer` like original AlexNet and
VGG16 papers

Changed
- how `searchnets.test` saves results file; includes name of 'config.ini`
in saved filename