MAML++: Per-Layer & Per-Layer Per-Step learning rate transforms #328

DubiousCactus · 2022-03-25T16:53:02Z

Description

Learning Per-Layer Per-Step Learning Rates (LSLR, MAML++). These GBML transforms can be used to reproduce this MAML++ functionality for MAML, MetaSGD and other algorithms which can be reproduced with GBML.

Contribution Checklist

If your contribution modifies code in the core library (not docs, tests, or examples), please fill the following checklist.

My contribution is listed in CHANGELOG.md with attribution.
My contribution modifies code in the main library.
My modifications are tested.
My modifications are documented.

Optional

If you make major changes to the core library, please run make alltests and copy-paste the content of alltests.txt below.

make[1]: Entering directory '/home/cactus/Code/learn2learn'
OMP_NUM_THREADS=1 \
MKL_NUM_THREADS=1 \
python -W ignore -m unittest discover -s 'tests' -p '*_test.py' -v
test_final_accuracy (integration.maml_omniglot_test.MAMLOmniglotIntegrationTests) ... ok
test_adaptation (unit.algorithms.gbml_test.TestGBMLgorithm) ... ok
test_allow_nograd (unit.algorithms.gbml_test.TestGBMLgorithm) ... Traceback (most recent call last):
  File "/home/cactus/Code/learn2learn/learn2learn/optim/parameter_update.py", line 119, in forward
    gradients = torch.autograd.grad(
  File "/home/cactus/anaconda3/envs/torch/lib/python3.8/site-packages/torch/autograd/__init__.py", line 234, in grad
    return Variable._execution_engine.run_backward(
RuntimeError: One of the differentiated Tensors does not require grad
ok
test_allow_unused (unit.algorithms.gbml_test.TestGBMLgorithm) ... ok
test_clone_module (unit.algorithms.gbml_test.TestGBMLgorithm) ... ok
test_graph_connection (unit.algorithms.gbml_test.TestGBMLgorithm) ... ok
test_adaptation (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_allow_nograd (unit.algorithms.maml_test.TestMAMLAlgorithm) ... Traceback (most recent call last):
  File "/home/cactus/Code/learn2learn/learn2learn/algorithms/maml.py", line 160, in adapt
    gradients = grad(loss,
  File "/home/cactus/anaconda3/envs/torch/lib/python3.8/site-packages/torch/autograd/__init__.py", line 234, in grad
    return Variable._execution_engine.run_backward(
RuntimeError: One of the differentiated Tensors does not require grad
ok
test_allow_unused (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_clone_module (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_first_order_adaptation (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_graph_connection (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_memory_consumption (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_module_shared_params (unit.algorithms.maml_test.TestMAMLAlgorithm) ... ok
test_adaptation (unit.algorithms.metasgd_test.TestMetaSGDAlgorithm) ... ok
test_clone_module (unit.algorithms.metasgd_test.TestMetaSGDAlgorithm) ... ok
test_graph_connection (unit.algorithms.metasgd_test.TestMetaSGDAlgorithm) ... ok
test_memory_consumption (unit.algorithms.metasgd_test.TestMetaSGDAlgorithm) ... ok
test_meta_lr (unit.algorithms.metasgd_test.TestMetaSGDAlgorithm) ... ok
Files already downloaded and verified
Files already downloaded and verified
0 Meta Train Accuracy 0.3937500095926225
1 Meta Train Accuracy 0.5000000102445483
2 Meta Train Accuracy 0.5125000113621354
3 Meta Train Accuracy 0.5312500107102096
4 Meta Train Accuracy 0.5812500161118805
learn2learn: Maybe try with allow_nograd=True and/orallow_unused=True ?
learn2learn: Maybe try with allow_nograd=True and/or allow_unused=True ?
Downloading https://raw.githubusercontent.com/brendenlake/omniglot/master/python/images_background.zip to /tmp/datasets/omniglot-py/images_background.zip

  0%|          | 0/9464212 [00:00<?, ?it/s]
  3%|▎         | 319488/9464212 [00:00<00:02, 3176299.16it/s]
 24%|██▍       | 2285568/9464212 [00:00<00:00, 12845223.70it/s]
 58%|█████▊    | 5480448/9464212 [00:00<00:00, 21536194.07it/s]
 91%|█████████ | 8569856/9464212 [00:00<00:00, 25217203.90it/s]
9464832it [00:00, 21873074.09it/s]                             
Extracting /tmp/datasets/omniglot-py/images_background.zip to /tmp/datasets/omniglot-py
Downloading https://raw.githubusercontent.com/brendenlake/omniglot/master/python/images_evaluation.zip to /tmp/datasets/omniglot-py/images_evaluation.zip

  0%|          | 0/6462886 [00:00<?, ?it/s]
  5%|▌         | 327680/6462886 [00:00<00:01, 3260524.99it/s]
 36%|███▌      | 2326528/6462886 [00:00<00:00, 13038351.78it/s]
 85%|████████▍ | 5472256/6462886 [00:00<00:00, 21400815.70it/s]
6463488it [00:00, 19254143.16it/s]                             
test_data_labels_length (unit.data.metadataset_test.TestMetaDataset) ... ok
test_data_labels_values (unit.data.metadataset_test.TestMetaDataset) ... ok
test_data_length (unit.data.metadataset_test.TestMetaDataset) ... ok
test_fails_with_non_torch_dataset (unit.data.metadataset_test.TestMetaDataset) ... ok
test_filtered_metadataset (unit.data.metadataset_test.TestMetaDataset) ... Extracting /tmp/datasets/omniglot-py/images_evaluation.zip to /tmp/datasets/omniglot-py
Files already downloaded and verified
Files already downloaded and verified
Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/train-images-idx3-ubyte.gz to /tmp/datasets/MNIST/raw/train-images-idx3-ubyte.gz

  0%|          | 0/9912422 [00:00<?, ?it/s]
  4%|▍         | 374784/9912422 [00:00<00:02, 3743015.66it/s]
 18%|█▊        | 1787904/9912422 [00:00<00:00, 9836908.54it/s]
 50%|████▉     | 4931584/9912422 [00:00<00:00, 19689376.20it/s]
 80%|████████  | 7957504/9912422 [00:00<00:00, 23661710.36it/s]
9913344it [00:00, 21682749.88it/s]                             
Extracting /tmp/datasets/MNIST/raw/train-images-idx3-ubyte.gz to /tmp/datasets/MNIST/raw

Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/train-labels-idx1-ubyte.gz to /tmp/datasets/MNIST/raw/train-labels-idx1-ubyte.gz

  0%|          | 0/28881 [00:00<?, ?it/s]
29696it [00:00, 49445832.31it/s]         
Extracting /tmp/datasets/MNIST/raw/train-labels-idx1-ubyte.gz to /tmp/datasets/MNIST/raw

Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/t10k-images-idx3-ubyte.gz to /tmp/datasets/MNIST/raw/t10k-images-idx3-ubyte.gz

  0%|          | 0/1648877 [00:00<?, ?it/s]
 24%|██▍       | 394240/1648877 [00:00<00:00, 3922912.19it/s]
1649664it [00:00, 9436849.35it/s]                            
Extracting /tmp/datasets/MNIST/raw/t10k-images-idx3-ubyte.gz to /tmp/datasets/MNIST/raw

Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz
Downloading http://yann.lecun.com/exdb/mnist/t10k-labels-idx1-ubyte.gz to /tmp/datasets/MNIST/raw/t10k-labels-idx1-ubyte.gz

  0%|          | 0/4542 [00:00<?, ?it/s]
5120it [00:00, 39116277.74it/s]         
ok
test_get_item (unit.data.metadataset_test.TestMetaDataset) ... ok
test_labels_to_indices (unit.data.metadataset_test.TestMetaDataset) ... ok
test_union_metadataset (unit.data.metadataset_test.TestMetaDataset) ... ok
test_dataloader (unit.data.task_dataset_test.TestTaskDataset) ... Extracting /tmp/datasets/MNIST/raw/t10k-labels-idx1-ubyte.gz to /tmp/datasets/MNIST/raw

Files already downloaded and verified
Files already downloaded and verified
ok
test_infinite_tasks (unit.data.task_dataset_test.TestTaskDataset) ... ok
test_instanciation (unit.data.task_dataset_test.TestTaskDataset) ... ok
test_task_caching (unit.data.task_dataset_test.TestTaskDataset) ... ok
test_task_transforms (unit.data.task_dataset_test.TestTaskDataset) ... ok
test_filter_labels (unit.data.transforms_test.TestTransforms) ... ok
test_k_shots (unit.data.transforms_test.TestTransforms) ... ok
test_load_data (unit.data.transforms_test.TestTransforms) ... ok
test_n_ways (unit.data.transforms_test.TestTransforms) ... ok
test_remap_labels (unit.data.transforms_test.TestTransforms) ... ok
test_infinite_iterator (unit.data.utils_test.DataUtilsTests) ... ok
test_partition_task (unit.data.utils_test.DataUtilsTests) ... ok
test_illegal_dimensions (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_illegal_dimensions_1d (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_m_edge (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_m_edge_1d (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_m_n_edge (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_m_n_edge_1d (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_n_edge (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_n_edge_1d (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_simple (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_simple_1d (unit.nn.kroneckers_test.KroneckerLinearTests) ... ok
test_cosine_distance (unit.nn.protonet_test.PrototypicalClassifierTests) ... ok
test_euclidean_distance (unit.nn.protonet_test.PrototypicalClassifierTests) ... ok
test_simple (unit.nn.protonet_test.PrototypicalClassifierTests) ... ok
test_clone_module_basics (unit.utils_test.UtilTests) ... ok
test_clone_module_models (unit.utils_test.UtilTests) ... ok
test_clone_module_nomodule (unit.utils_test.UtilTests) ... ok
test_distribution_clone (unit.utils_test.UtilTests) ... ok
test_distribution_detach (unit.utils_test.UtilTests) ... ok
test_module_clone_shared_params (unit.utils_test.UtilTests) ... ok
test_module_detach (unit.utils_test.UtilTests) ... ok
test_module_detach_keep_requires_grad (unit.utils_test.UtilTests) ... ok
test_module_update_shared_params (unit.utils_test.UtilTests) ... ok
test_rnn_clone (unit.utils_test.UtilTests) ... ok

----------------------------------------------------------------------
Ran 62 tests in 40.781s

OK
make lint
make[2]: Entering directory '/home/cactus/Code/learn2learn'
pycodestyle learn2learn/ --max-line-length=160
make[2]: Leaving directory '/home/cactus/Code/learn2learn'
make[1]: Leaving directory '/home/cactus/Code/learn2learn'
make[1]: Entering directory '/home/cactus/Code/learn2learn'
OMP_NUM_THREADS=1 \
MKL_NUM_THREADS=1 \
python -W ignore -m unittest discover -s 'tests' -p '*_test_notravis.py' -v
test_final_accuracy (integration.maml_miniimagenet_test_notravis.MAMLMiniImagenetIntegrationTests) ... ok
test_final_accuracy (integration.protonets_miniimagenet_test_notravis.ProtoNetMiniImageNetIntegrationTests) ... make[1]: *** [Makefile:37: notravis-tests] Killed
make[1]: Leaving directory '/home/cactus/Code/learn2learn'

lgtm-com · 2022-03-25T17:13:22Z

This pull request introduces 1 alert when merging baf38e0 into df3c329 - view on LGTM.com

new alerts:

1 for Unused import

…to mamlpp_LSLR

* Add load_state_dict(), state_dict() for saving/restoring of LRs * Add optional argument to PerLayerPerStepLRTransform to only attach per-step LRs to given layer names

lgtm-com · 2022-07-28T17:43:24Z

This pull request fixes 2 alerts when merging 98a118c into f099ddc - view on LGTM.com

fixed alerts:

1 for Unused local variable
1 for Unused import

lgtm-com · 2022-07-29T16:19:36Z

This pull request fixes 2 alerts when merging c4ea713 into f099ddc - view on LGTM.com

fixed alerts:

1 for Unused local variable
1 for Unused import

lgtm-com · 2022-07-30T16:29:51Z

This pull request fixes 2 alerts when merging 811beb0 into f099ddc - view on LGTM.com

fixed alerts:

1 for Unused local variable
1 for Unused import

DubiousCactus added 3 commits March 24, 2022 18:27

Implement LSLR as an algorithm wrapper

2122d30

[WIP] LSLR wrapper

bda9cff

Remove debug stuff

baf38e0

DubiousCactus added 5 commits April 22, 2022 20:13

Implement two LSLR flavours as GBML transforms

bb86034

Merge branch 'mamlpp_LSLR' of github.com:DubiousCactus/learn2learn in…

7ec2011

…to mamlpp_LSLR

Undo debug changes to examples

89e78e9

Add my changes to the CHANGELOG

d60d4fa

Rename transform file and add documentation

0120f79

DubiousCactus marked this pull request as ready for review April 22, 2022 19:42

DubiousCactus changed the title ~~[WIP] MAML++: LSLR~~ MAML++: Per-Layer & Per-Layer Per-Step learning rate transforms Apr 22, 2022

DubiousCactus added 5 commits April 22, 2022 20:52

Undo example changes

5e1b506

If the steps counter and learning rates sign

4849f2c

Use the new LSLR transform for the maml++ example

88681d0

Improve transforms

ce0da39

* Add load_state_dict(), state_dict() for saving/restoring of LRs * Add optional argument to PerLayerPerStepLRTransform to only attach per-step LRs to given layer names

Fix layer_names argument and use it for the example

98a118c

Fix state_dict() in LSLR

c4ea713

Fix load_state_dict() to handle None layer LRs

811beb0

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

MAML++: Per-Layer & Per-Layer Per-Step learning rate transforms #328

MAML++: Per-Layer & Per-Layer Per-Step learning rate transforms #328

DubiousCactus commented Mar 25, 2022 •

edited

Loading

lgtm-com bot commented Mar 25, 2022

lgtm-com bot commented Jul 28, 2022

lgtm-com bot commented Jul 29, 2022

lgtm-com bot commented Jul 30, 2022

MAML++: Per-Layer & Per-Layer Per-Step learning rate transforms #328

Are you sure you want to change the base?

MAML++: Per-Layer & Per-Layer Per-Step learning rate transforms #328

Conversation

DubiousCactus commented Mar 25, 2022 • edited Loading

Description

Contribution Checklist

Optional

lgtm-com bot commented Mar 25, 2022

lgtm-com bot commented Jul 28, 2022

lgtm-com bot commented Jul 29, 2022

lgtm-com bot commented Jul 30, 2022

DubiousCactus commented Mar 25, 2022 •

edited

Loading