To lessen the burden of re-running the benchmark, would it be possible to publish the optimal hyper-parameters somewhere?

By-reusing those hyper-parameters, one would avoid the most computationally-demanding part of reproducing the results (by 1-2 orders of magnitude).

• Adding parameter skip_if_exists to runner.run
  • Default value is set such that the current behavior is maintained
  • By setting to True, runs that already have a .json output file will not be executed again
• Possible extensions
  • Make skip_if_exists arg-parsable

(Apologies for creating multiple issues in a row -- it seemed more clean to keep them separate.)

I downloaded the data from DeepOBS_Baselines, and attempted to run example_analyze_pytorch.py. Unfortunately DeepOBS seems to look for keys in the JSON files that don't exist:

$ python example_analyze_pytorch.py
/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py:144: RuntimeWarning: Metric valid_accu
racies does not exist for testproblem quadratic_deep. We now use fallback metric valid_losses
  default_metric), RuntimeWarning)
/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py:229: RuntimeWarning: All settings for
/scratch/local/ssd/user/data/deepobs/quadratic_deep/SGD on test problem quadratic_deep have the same
 number of seeds runs. Mode 'most' does not make sense and we use the fallback mode 'final'
  .format(optimizer_path, testproblem_name), RuntimeWarning)
{'Performance': 127.96759578159877, 'Speed': 'N.A.', 'Hyperparameters': {'lr': 0.01, 'momentum': 0.9
9, 'nesterov': False}, 'Training Parameters': {}}
/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py:144: RuntimeWarning: Metric valid_accu
racies does not exist for testproblem quadratic_deep. We now use fallback metric valid_losses
  default_metric), RuntimeWarning)
/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py:229: RuntimeWarning: All settings for
/scratch/local/ssd/user/data/deepobs/quadratic_deep/SGD on test problem quadratic_deep have the same
 number of seeds runs. Mode 'most' does not make sense and we use the fallback mode 'final'
  .format(optimizer_path, testproblem_name), RuntimeWarning)
/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py:150: RuntimeWarning: Cannot fallback t
o metric valid_losses for optimizer MomentumOptimizer on testproblem quadratic_deep. Will now fallba
ck to metric test_losses
  testproblem_name), RuntimeWarning)
/users/user/miniconda3/lib/python3.7/site-packages/numpy/core/_methods.py:193: RuntimeWarning: inva$
id value encountered in subtract
  x = asanyarray(arr - arrmean)
/users/user/miniconda3/lib/python3.7/site-packages/numpy/lib/function_base.py:3949: RuntimeWarning:
invalid value encountered in multiply
  x2 = take(ap, indices_above, axis=axis) * weights_above
Traceback (most recent call last):
  File "example_analyze_pytorch.py", line 17, in <module>
    analyzer.plot_optimizer_performance(result_path, reference_path=base + '/deepobs/baselines/quad$
atic_deep/MomentumOptimizer')
  File "/users/user/Research/deepobs/deepobs/analyzer/analyze.py", line 514, in plot_optimizer_perfo
rmance
    which=which)
  File "/users/user/Research/deepobs/deepobs/analyzer/analyze.py", line 462, in _plot_optimizer_perf
ormance
    optimizer_path, mode, metric)
  File "/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py", line 206, in create_setting_
analyzer_ranking
    setting_analyzers = _get_all_setting_analyzer(optimizer_path)
  File "/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py", line 184, in _get_all_settin
g_analyzer
    setting_analyzers.append(SettingAnalyzer(sett_path))
  File "/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py", line 260, in __init__
    self.aggregate = aggregate_runs(path)
  File "/users/user/Research/deepobs/deepobs/analyzer/shared_utils.py", line 101, in aggregate_runs
    aggregate['optimizer_hyperparams'] = json_data['optimizer_hyperparams']
KeyError: 'optimizer_hyperparams'

One of the JSON files in question looks like this (data points snipped for brevity):

{
"train_losses": [353.9337594168527, 347.5994306291853, 331.35902622767856, 307.2468915666853, ... 97.28871154785156, 91.45470428466797, 96.45774841308594, 86.27237701416016],
"optimizer": "MomentumOptimizer",
"testproblem": "quadratic_deep",
"weight_decay": null,
"batch_size": 128,
"num_epochs": 100,
"learning_rate": 1e-05,
"lr_sched_epochs": null,
"lr_sched_factors": null,
"random_seed": 42,
"train_log_interval": 1,
"hyperparams": {"momentum": 0.99, "use_nesterov": false}
}

The obvious key seems to be hyperparams as opposed to optimizer_hyperparams; this occurs only for some JSON files.

Edit: Having fixed this, there is a further key error on training_params. Perhaps these were generated with different versions of the package.

Attempting to install by following the documentation's instructions, after installing all the mentioned dependencies with conda, results in the following error:

(base) [email protected]:~$ pip install -e git+https://github.com/abahde/[email protected]#egg=DeepOBS
Obtaining DeepOBS from git+https://github.com/abahde/[email protected]#egg=DeepOBS
  Cloning https://github.com/abahde/DeepOBS.git (to revision master) to ./src/deepobs
  Running command git clone -q https://github.com/abahde/DeepOBS.git /users/user/src/deepobs
    ERROR: Complete output from command python setup.py egg_info:
    ERROR: Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/users/user/src/deepobs/setup.py", line 5, in <module>
        from deepobs import __version__
      File "/users/user/src/deepobs/deepobs/__init__.py", line 5, in <module>
        from . import analyzer
      File "/users/user/src/deepobs/deepobs/analyzer/__init__.py", line 2, in <module>
        from . import analyze
      File "/users/user/src/deepobs/deepobs/analyzer/analyze.py", line 12, in <module>
        from ..tuner.tuner_utils import generate_tuning_summary
      File "/users/user/src/deepobs/deepobs/tuner/__init__.py", line 4, in <module>
        from .bayesian import GP
      File "/users/user/src/deepobs/deepobs/tuner/bayesian.py", line 3, in <module>
        from bayes_opt import UtilityFunction
    ModuleNotFoundError: No module named 'bayes_opt'
    ----------------------------------------
ERROR: Command "python setup.py egg_info" failed with error code 1 in /users/user/src/deepobs/

Is this bayes_opt package really necessary? It seems a bit tangential to the package's purpose (or at most optional).

Edit: It turns out that bayesian-optimization has relatively few requirements so this is not a big issue; perhaps just the docs need updating.

As an aside, it might be possible to suggest a single conda command that installs everything: conda install -c conda-forge seaborn matplotlib2tikz bayesian-optimization.

Optimizers can have very different runtimes per iteration, especially 2nd-order ones.

This means that sometimes, despite promises of "faster" convergence, the wall-clock time taken to converge is disappointingly larger.

Is there any chance DeepOBS could implement wall-clock time plots, in addition to per-epoch ones? (E.g. X axis in minutes or hours.)

There are a couple of improvements that I suggest:

• [ ] Return the results not as a string, but as a dict or an object.
• [ ] (Maybe, think about that) Include the ability to test multiple optimizers simultaneously.
• [ ] Report standard deviation and individual runtimes for SGD.
• [ ] Add a function that generates a figure, similar to https://github.com/ludwigbald/probprec/blob/master/code/exp_perf_prec/analyze.py

In PyTorch we split the validation set from the training set randomly. It has the size of the test set. The validation performance is used by the tuner and analyzer to obtain the best instance. This split should be implemented in the TensorFlow data sets as well. We have already prepared the test problem and the runner implementations for this change. The only change that needs to be done to the runner is marked in the code with a ToDo flag.