I updated from 1.0.0 to 1.1.9 today and now I get the following error when running pruner.step():
pruner.step()
File "/home/mle/prune_data_collection/pruning_benchmark_tool/.venv/lib/python3.8/site-packages/torch_pruning/pruner/algorithms/metapruner.py", line 159, in step
for group in self.prune_global():
File "/home/mle/prune_data_collection/pruning_benchmark_tool/.venv/lib/python3.8/site-packages/torch_pruning/pruner/algorithms/metapruner.py", line 257, in prune_global
imp = torch.cat([local_imp[-1]
RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 512 but got size 256 for tensor number 1 in the list.
I downgraded the version to check back and my code still works on 1.0.0. This is the setup I use for pruning a ResNet trained on CIFAR10:
criterion = tp.importance.MagnitudeImportance(p=2, group_reduction=None)
ignored_layers = []
for m in model.modules():
if isinstance(m, torch.nn.Linear) and m.out_features == 10:
ignored_layers.append(m)
# Pruner initialization
iterative_steps = 20
pruner = tp.pruner.MagnitudePruner(
model,
example_inputs,
global_pruning=True,
importance=criterion,
iterative_steps=iterative_steps,
ch_sparsity=0.9,
ignored_layers=ignored_layers,
)
Currently, I'm fine with running 1.0.0 so I've not spent time on debugging it but I might do so later.
Maybe it's just a minor thing and easy to change.
Hi,
I updated from 1.0.0 to 1.1.9 today and now I get the following error when running pruner.step():
I downgraded the version to check back and my code still works on 1.0.0. This is the setup I use for pruning a ResNet trained on CIFAR10:
Currently, I'm fine with running 1.0.0 so I've not spent time on debugging it but I might do so later. Maybe it's just a minor thing and easy to change.
Thanks