95 lines
2.8 KiB
Python
95 lines
2.8 KiB
Python
|
"""Dynamically prune prototypes in GLVQ-type models."""
|
||
|
|
|
||
|
|
import argparse
|
||
|
|
|
||
|
|
import prototorch as pt
|
||
|
|
import pytorch_lightning as pl
|
||
|
|
import torch
|
||
|
|
from pytorch_lightning.callbacks import Callback
|
||
|
|
|
||
|
|
|
||
|
|
class PrototypePruning(Callback):
|
||
|
|
def __init__(self, threshold=0.01, prune_after=10, verbose=False):
|
||
|
|
self.threshold = threshold
|
||
|
|
self.prune_after = prune_after
|
||
|
|
self.verbose = verbose
|
||
|
|
|
||
|
|
def on_epoch_start(self, trainer, pl_module):
|
||
|
|
pl_module.initialize_prototype_win_ratios()
|
||
|
|
|
||
|
|
def on_epoch_end(self, trainer, pl_module):
|
||
|
|
if (trainer.current_epoch + 1) > self.prune_after:
|
||
|
|
ratios = pl_module.prototype_win_ratios.mean(dim=0)
|
||
|
|
to_prune = torch.arange(len(ratios))[ratios < self.threshold]
|
||
|
|
if len(to_prune) > 0:
|
||
|
|
if self.verbose:
|
||
|
|
print(f"\nPrototype win ratios: {ratios}")
|
||
|
|
print(f"Pruning prototypes at indices: {to_prune}")
|
||
|
|
cur_num_protos = pl_module.num_prototypes
|
||
|
|
pl_module.remove_prototypes(indices=to_prune)
|
||
|
|
new_num_protos = pl_module.num_prototypes
|
||
|
|
if self.verbose:
|
||
|
|
print(f"`num_prototypes` reduced from {cur_num_protos} "
|
||
|
|
f"to {new_num_protos}.")
|
||
|
|
|
||
|
|
|
||
|
|
if __name__ == "__main__":
|
||
|
|
# Command-line arguments
|
||
|
|
parser = argparse.ArgumentParser()
|
||
|
|
parser = pl.Trainer.add_argparse_args(parser)
|
||
|
|
args = parser.parse_args()
|
||
|
|
|
||
|
|
# Dataset
|
||
|
|
num_classes = 4
|
||
|
|
num_features = 2
|
||
|
|
num_clusters = 1
|
||
|
|
train_ds = pt.datasets.Random(num_samples=500,
|
||
|
|
num_classes=num_classes,
|
||
|
|
num_features=num_features,
|
||
|
|
num_clusters=num_clusters,
|
||
|
|
separation=3.0,
|
||
|
|
seed=42)
|
||
|
|
|
||
|
|
# Dataloaders
|
||
|
|
train_loader = torch.utils.data.DataLoader(train_ds, batch_size=256)
|
||
|
|
|
||
|
|
# Hyperparameters
|
||
|
|
prototypes_per_class = num_clusters * 5
|
||
|
|
hparams = dict(
|
||
|
|
distribution=(num_classes, prototypes_per_class),
|
||
|
|
lr=0.3,
|
||
|
|
)
|
||
|
|
|
||
|
|
# Initialize the model
|
||
|
|
model = pt.models.CELVQ(
|
||
|
|
hparams,
|
||
|
|
prototype_initializer=pt.components.Ones(2, scale=3),
|
||
|
|
)
|
||
|
|
|
||
|
|
# Summary
|
||
|
|
print(model)
|
||
|
|
|
||
|
|
# Callbacks
|
||
|
|
vis = pt.models.VisGLVQ2D(train_ds)
|
||
|
|
pruning = PrototypePruning(
|
||
|
|
threshold=0.01, # prune prototype if it wins less than 1%
|
||
|
|
prune_after=50,
|
||
|
|
verbose=True,
|
||
|
|
)
|
||
|
|
|
||
|
|
# Setup trainer
|
||
|
|
trainer = pl.Trainer.from_argparse_args(
|
||
|
|
args,
|
||
|
|
max_epochs=100,
|
||
|
|
callbacks=[
|
||
|
|
vis,
|
||
|
|
pruning,
|
||
|
|
],
|
||
|
|
terminate_on_nan=True,
|
||
|
|
weights_summary=None,
|
||
|
|
accelerator="ddp",
|
||
|
|
)
|
||
|
|
|
||
|
|
# Training loop
|
||
|
|
trainer.fit(model, train_loader)
|