This repository contains code for the paper Scalable One-Pass Optimisation of High-Dimensional Weight-Update Hyperparameters by Implicit Differentiation.
Installation
Our dependencies are fully specified in Pipfile
, which can be supplied to pipenv
to install the environment. One failsafe approach is to install pipenv
in a fresh virtual environment, then run pipenv install
in this directory. Note the Pipfile specifies our Python 3.9 development environment; most experiments were run in an identical environment under Python 3.7 instead.
Difficulties with CUDA versions meant we had to manually install PyTorch and Torchvision rather than use pipenv
--- the corresponding lines in Pipfile
may need adjustment for your use case. Alternatively, use the list of dependencies as a guide to what to install yourself with pip
, or use the full dump of our development environment in final_requirements.txt
.
Datasets may not be bundled with the repository, but are expected to be found at locations specified in datasets.py
, preprocessed into single PyTorch tensors of all the input and output data (generally data/<dataset>/data.pt
and data/<dataset>/targets.pt
).
Configuration
Training code is controlled with YAML configuration files, as per the examples in configs/
. Generally one file is required to specify the dataset, and a second to specify the algorithm, using the obvious naming convention. Brief help text is available on the command line, but the meanings of each option should be reasonably self-explanatory.
For Ours (WD+LR), use the file Ours_LR.yaml
; for Ours (WD+LR+M), use the file Ours_LR_Momentum.yaml
; for Ours (WD+HDLR+M), use the file Ours_HDLR_Momentum.yaml
. For Long/Medium/Full Diff-through-Opt, we provide separate configuration files for the UCI cases and the Fashion-MNIST cases.
We provide two additional helper configurations. Random_Validation.yaml
copies Random.yaml
, but uses the entire validation set to compute the validation loss at each logging step. This allows for stricter analysis of the best-performing run at particular time steps, for instance while constructing Random (3-batched). Random_Validation_BayesOpt.yaml
only forces the use of the entire dataset for the very last validation loss computation, so that Bayesian Optimisation runs can access reliable performance metrics without adversely affecting runtime.
The configurations provided match those necessary to replicate the main experiments in our paper (in Section 4: Experiments). Other trials, such as those in the Appendix, will require these configurations to be modified as we describe in the paper. Note especially that our three short-horizon bias studies all require different modifications to the LongDiffThroughOpt_*.yaml
configurations.
Running
Individual runs are commenced by executing train.py
and passing the desired configuration files with the -c
flag. For example, to run the default Fashion-MNIST experiments using Diff-through-Opt, use:
$ python train.py -c ./configs/fashion_mnist.yaml ./configs/DiffThroughOpt.yaml
Bayesian Optimisation runs are started in a similar way, but with a call to bayesopt.py
rather than train.py
.
For executing multiple runs in parallel, parallel_exec.py
may be useful: modify the main function call at the bottom of the file as required, then call this file instead of train.py
at the command line. The number of parallel workers may be specified by num_workers
. Any configurations passed at the command line are used as a base, to which modifications may be added by override_generator
. The latter should either be a function which generates one override dictionary per call (in which case num_repetitions
sets the number of overrides to generate), or a function which returns a generator over configurations (in which case set num_repetitions = None
). Each configuration override is run once for each of algorithms
, whose configurations are read automatically from the corresponding files and should not be explicitly passed at the command line. Finally, main_function
may be used to switch between parallel calls to train.py
and bayesopt.py
as required.
For blank-slate replications, the most useful override generators will be natural_sgd_generator
, which generates a full SGD initialisation in the ranges we use, and iteration_id
, which should be used with Bayesian Optimisation runs to name each parallel run using a counter. Other generators may be useful if you wish to supplement existing results with additional algorithms etc.
PennTreebank and CIFAR-10 were executed on clusters running SLURM; the corresponding subfolders contain configuration scripts for these experiments, and submit.sh
handles the actual job submission.
Analysis
By default, runs are logged in Tensorboard format to the ./runs
directory, where Tensorboard may be used to inspect the results. If desired, a descriptive name can be appended to a particular execution using the -n
switch on the command line. Runs can optionally be written to a dedicated subfolder specified with the -g
switch, and the base folder for logging can be changed with the -l
switch.
If more precise analysis is desired, pass the directory containing the desired results to util.get_tags()
, which will return a dictionary of the evolution of each logged scalar in the results. Note that this function uses Tensorboard calls which predate its --load_fast
option, so may take tens of minutes to return.
This data dictionary can be passed to one of the more involved plotting routines in figures.py
to produce specific plots. The script paper_plots.py
generates all the plots we use in our paper, and may be inspected for details of any particular plot.