Some personal tools for running ML experiments:
- Run variations of anything and everything (weight initializations, architecture, hyperpararmeters, optimizer choices, etc.).
- Fine-grained interventions (perturb the weights, gradients, activations, hyperparameters, etc.).
- Take checkpoints any time.
- Custom metrics and plotters (record any metric you can imagine — on the models themselves, between models, the test sets, training sets, etc.).
- Train in parallel on a cluster (ok, not yet, but, you know, eventually) or in serial on your local machine.
- Reuse past results when running a new intervention if you've already tested a particular condition/control.
- Consistent seed management.
The library is organized as follows:
Experimentis a collection ofLearnersdifferentiated byIntervention.Trialis a (Model,Optimizer,DataLoader,Metrics,Intervention) tuple. It is the basic unit of training.Interventionis a class that perturbs the model, optimizer, or data loader.Metricsis a class that records metrics and logs them.Plotteris a class that plots metrics.
Note: This library is still in development. The API is subject to change. I definitely don't recommend using it yet, as I haven't ironed out all the kinks.
Suppose I want to test the effect of a small weight perturbation at initialization. It's as simple as the following:
from torchvision import datasets, transforms
from perturb.experiments import Experiment
from perturb.interventions import PerturbWeights
from perturb.metrics import Metrics
from perturb.plotter import Plotter
from perturb.models import Lenet5
train_set = datasets.MNIST('data', train=True, download=True,transform=transforms.ToTensor())
test_set = datasets.MNIST('data', train=False, download=True,transform=transforms.ToTensor())
exp = Experiment(
model=Lenet5(),
datasets=(train_set, test_set),
interventions=[
PerturbWeights.make_variations(
epsilon=(0.001, 0.01, 0.1),
seed_weights=range(10)
)
]
)
# 1 control 3 x 10 interventions = 31 trials
exp.run(n_epochs=10)Or maybe I want to compare the behavior of different optimizers.
from perturb.interventions import ReinitWeights, ChangeOptimizer
exp = Experiment(
model=Lenet5(),
dataset=(train_set, test_set),
variations=[
ReinitWeights.make_variations(
seed_weights=range(5)
),
ChangeOptimizer.make_variations(
optimizer=torch.optim.SGD,
lr=(0.001, 0.01, 0.1),
)
],
interventions=[
ChangeOptimizer.make_variations(
optimizer=torch.optim.Adam,
)
],
)
# (5 x 3 variations) x (1 control 1 intervention) = 30 trialsThe distinction between variations and interventions is for measuring purposes: it allows us to write metrics that are relative to a control condition. This way we don't have to keep multiple large models in memory at once.
Maybe instead I'd like to vary the batch size.
exp = Experiment(
model=Lenet5(),
dataset=(train_set, test_set),
variations=[
ReinitWeights.make_variations(
seed_weights=range(5)
),
ChangeTrainLoader.make_variations(
batch_size=32,
seed_shuffle=(.1, .2, .3)
)
],
interventions=[
ChangeTrainLoader.make_variations(
batch_size=(64, 128, 256, 512, 1024),
)
]
)
# (5 x 3 variations) x (1 control 5 interventions) = 90 trialsOr maybe I want to test the effect of a temporary perturbation to momentum, depending on when it is applied during training.
exp = Experiment(
model=Lenet5(),
dataset=(train_set, test_set),
variations=[
ReinitWeights.make_variations(
seed_weights=range(10)
),
ChangeOptimizer.make_variations(
momentum=(0.9, 0.99, 0.999),
),
],
interventions=[
ChangeOptimizer.make_variations(
momentum=lambda m: (m * (1 epsilon) for epsilon in (0.001, -0.001, 0.01, -0.01, 0.1, -0.1)),
when=((0, 100), (1000, 1100), (5000, 5100)), # Step ranges to maintain perturbation
)
]
)
# (10 x 3 variations) x (1 control 6 interventions) = 210 trialsThat's a lot of trials. My computer will take several days to run all of them.
So I can get rid of the ReinitWeights, which leaves me with a more reasonable 21 trials.
After I've validated for a fixed choice of weight initialization, I can add it back in, and run the experiment again. Best of all, it'll automatically skip the trials that have already been run.
Alternatively, I can train for only a few epochs, validate the results, and then train for more epochs, picking up where I left off from the checkpoints.
This allows for a more iterative experimentation loop so you can explore more ground faster.
By default, the Metrics will record the loss and accuracy on the training and test sets at the end of each epoch. Plotting is disabled by default, but you can enable it by passing plot=True to Experiment.run.
Often, you'll want to compute performance relative to some control condition (e.g., cross entropy relative to an unperturbed model). The Trial class has a control attribute that points to the control condition. You can use this to compute any metric you want by subclassing Metrics.
class CustomMetrics(Metrics):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.register_metric('w', self.weight_norm)
self.register_metric('dw_control', self.weight_distance_from_control)
def weight_norm(self, trial):
norm = t.zeros(1)
for p in trial.model.parameters():
norm = torch.norm(p)
return norm
def weight_distance_from_control(self, trial):
distance = t.zeros(1)
for p1, p2 in zip(trial.model.parameters(), trial.control.model.parameters()):
distance = torch.norm(p1-p2) ** 2
return distance ** 0.5Let's take the example of the batch size experiment.
exp = Experiment(
model=Lenet5(),
dataset=(train_set, test_set),
variations=[
ReinitWeights.make_variations(
seed_weights=range(5)
),
ChangeTrainLoader.make_variations(
batch_size=32,
batch_order_seed_weights=(.1, .2, .3)
)
],
interventions=[
ChangeTrainLoader.make_variations(
batch_size=(64, 128, 256, 512, 1024),
)
],
metrics=CustomMetrics(),
plotter=Plotter(
average_over=('seed_shuffle', 'seed'),
)
)