equiformer pytorchダウンロード - equiformer pytorchソースコードのダウンロード

Equiformer - パイトーチ (wip)

新しい SOTA に到達する Equiformer、SE3/E3 equivariant tention network の実装、およびタンパク質フォールディングのための EquiFold (Prescient Design) による使用に採用

この設計は SE3 Transformers から構築されているようで、ドット積アテンションが MLP アテンションと GATv2 からの非線形メッセージパッシングに置き換えられています。また、もう少し効率を上げるために深さ方向のテンソル積も実行します。私が間違っていると思われる場合は、お気軽にメールでご連絡ください。

更新: SE3 等変ネットワークの次数のスケーリングを劇的に改善する新しい開発が行われました。この論文では、最初に、z 軸 (または他の慣例により y 軸) に沿って表現を整列させることにより、球面調和関数が疎になることに注目しました。これにより、方程式から m _f次元が削除されます。 Passaroらのフォローアップ論文。クレブシュゴルダン行列も疎になり、 _miと l _fが削除されることに注意してください。彼らはまた、担当者を 1 つの軸に揃えた後、問題が SO(3) から SO(2) に軽減されたと関連付けました。 Equiformer v2 (公式リポジトリ) は、トランスフォーマーのようなフレームワークでこれを利用して、新しい SOTA に到達します。

間違いなく、これにさらに多くの作業/探索を費やすことになるでしょう。今のところ、SO(2) への完全な変換を除いて、Equiformer v1 の最初の 2 つの論文のトリックを組み込みました。

アップデート 2: より高次のレップ間の相互作用のない新しい SOTA があるようです (言い換えれば、すべてのテンソル積/クレブシュゴルダン数学はなくなります)。 GotenNet、HEGNN のトランスフォーマー版と思われます

インストール

$ pip install equiformer-pytorch

使用法

 import torch
from equiformer_pytorch import Equiformer

model = Equiformer (
    num_tokens = 24 ,
    dim = ( 4 , 4 , 2 ),               # dimensions per type, ascending, length must match number of degrees (num_degrees)
    dim_head = ( 4 , 4 , 4 ),          # dimension per attention head
    heads = ( 2 , 2 , 2 ),             # number of attention heads
    num_linear_attn_heads = 0 ,     # number of global linear attention heads, can see all the neighbors
    num_degrees = 3 ,               # number of degrees
    depth = 4 ,                     # depth of equivariant transformer
    attend_self = True ,            # attending to self or not
    reduce_dim_out = True ,         # whether to reduce out to dimension of 1, say for predicting new coordinates for type 1 features
    l2_dist_attention = False      # set to False to try out MLP attention
). cuda ()

feats = torch . randint ( 0 , 24 , ( 1 , 128 )). cuda ()
coors = torch . randn ( 1 , 128 , 3 ). cuda ()
mask  = torch . ones ( 1 , 128 ). bool (). cuda ()

out = model ( feats , coors , mask ) # (1, 128)

out . type0 # invariant type 0    - (1, 128)
out . type1 # equivariant type 1  - (1, 128, 3)

このリポジトリには、可逆ネットワークを使用してメモリ使用量を深さから切り離す方法も含まれています。言い換えれば、深さを増やした場合、メモリコストは 1 つの等価変換器ブロック (注意とフィードフォワード) の使用量で一定に保たれます。

 import torch
from equiformer_pytorch import Equiformer

model = Equiformer (
    num_tokens = 24 ,
    dim = ( 4 , 4 , 2 ),
    dim_head = ( 4 , 4 , 4 ),
    heads = ( 2 , 2 , 2 ),
    num_degrees = 3 ,
    depth = 48 ,          # depth of 48 - just to show that it runs - in reality, seems to be quite unstable at higher depths, so architecture stil needs more work
    reversible = True ,   # just set this to True to use https://arxiv.org/abs/1707.04585
). cuda ()

feats = torch . randint ( 0 , 24 , ( 1 , 128 )). cuda ()
coors = torch . randn ( 1 , 128 , 3 ). cuda ()
mask  = torch . ones ( 1 , 128 ). bool (). cuda ()

out = model ( feats , coors , mask )

out . type0 . sum (). backward ()

エッジ

エッジ付き、例:原子結合

 import torch
from equiformer_pytorch import Equiformer

model = Equiformer (
    num_tokens = 28 ,
    dim = 64 ,
    num_edge_tokens = 4 ,       # number of edge type, say 4 bond types
    edge_dim = 16 ,             # dimension of edge embedding
    depth = 2 ,
    input_degrees = 1 ,
    num_degrees = 3 ,
    reduce_dim_out = True
)

atoms = torch . randint ( 0 , 28 , ( 2 , 32 ))
bonds = torch . randint ( 0 , 4 , ( 2 , 32 , 32 ))
coors = torch . randn ( 2 , 32 , 3 )
mask  = torch . ones ( 2 , 32 ). bool ()

out = model ( atoms , coors , mask , edges = bonds )

out . type0 # (2, 32)
out . type1 # (2, 32, 3)

隣接行列あり

 import torch
from equiformer_pytorch import Equiformer

model = Equiformer (
    dim = 32 ,
    heads = 8 ,
    depth = 1 ,
    dim_head = 64 ,
    num_degrees = 2 ,
    valid_radius = 10 ,
    reduce_dim_out = True ,
    attend_sparse_neighbors = True ,  # this must be set to true, in which case it will assert that you pass in the adjacency matrix
    num_neighbors = 0 ,               # if you set this to 0, it will only consider the connected neighbors as defined by the adjacency matrix. but if you set a value greater than 0, it will continue to fetch the closest points up to this many, excluding the ones already specified by the adjacency matrix
    num_adj_degrees_embed = 2 ,       # this will derive the second degree connections and embed it correctly
    max_sparse_neighbors = 8         # you can cap the number of neighbors, sampled from within your sparse set of neighbors as defined by the adjacency matrix, if specified
)

feats = torch . randn ( 1 , 128 , 32 )
coors = torch . randn ( 1 , 128 , 3 )
mask  = torch . ones ( 1 , 128 ). bool ()

# placeholder adjacency matrix
# naively assuming the sequence is one long chain (128, 128)

i = torch . arange ( 128 )
adj_mat = ( i [:, None ] <= ( i [ None , :] + 1 )) & ( i [:, None ] >= ( i [ None , :] - 1 ))

out = model ( feats , coors , mask , adj_mat = adj_mat )

out . type0 # (1, 128)
out . type1 # (1, 128, 3)

感謝

StabilityAI の寛大なスポンサーシップと、他のスポンサーの皆様

テスト

等分散性のテストなど

$ python setup.py test

例

まずsidechainnetをインストールします

$ pip install sidechainnet

次に、タンパク質バックボーンのノイズ除去タスクを実行します。

$ python denoise.py

藤堂

引用

 @article { Liao2022EquiformerEG ,
    title   = { Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs } ,
    author  = { Yi Liao and Tess E. Smidt } ,
    journal = { ArXiv } ,
    year    = { 2022 } ,
    volume  = { abs/2206.11990 }
}

 @article { Lee2022.10.07.511322 ,
    author  = { Lee, Jae Hyeon and Yadollahpour, Payman and Watkins, Andrew and Frey, Nathan C. and Leaver-Fay, Andrew and Ra, Stephen and Cho, Kyunghyun and Gligorijevic, Vladimir and Regev, Aviv and Bonneau, Richard } ,
    title   = { EquiFold: Protein Structure Prediction with a Novel Coarse-Grained Structure Representation } ,
    elocation-id = { 2022.10.07.511322 } ,
    year    = { 2022 } ,
    doi     = { 10.1101/2022.10.07.511322 } ,
    publisher = { Cold Spring Harbor Laboratory } ,
    URL     = { https://www.biorxiv.org/content/early/2022/10/08/2022.10.07.511322 } ,
    eprint  = { https://www.biorxiv.org/content/early/2022/10/08/2022.10.07.511322.full.pdf } ,
    journal = { bioRxiv }
}

 @article { Shazeer2019FastTD ,
    title   = { Fast Transformer Decoding: One Write-Head is All You Need } ,
    author  = { Noam M. Shazeer } ,
    journal = { ArXiv } ,
    year    = { 2019 } ,
    volume  = { abs/1911.02150 }
}

 @misc { ding2021cogview ,
    title   = { CogView: Mastering Text-to-Image Generation via Transformers } ,
    author  = { Ming Ding and Zhuoyi Yang and Wenyi Hong and Wendi Zheng and Chang Zhou and Da Yin and Junyang Lin and Xu Zou and Zhou Shao and Hongxia Yang and Jie Tang } ,
    year    = { 2021 } ,
    eprint  = { 2105.13290 } ,
    archivePrefix = { arXiv } ,
    primaryClass = { cs.CV }
}

 @inproceedings { Kim2020TheLC ,
    title   = { The Lipschitz Constant of Self-Attention } ,
    author  = { Hyunjik Kim and George Papamakarios and Andriy Mnih } ,
    booktitle = { International Conference on Machine Learning } ,
    year    = { 2020 }
}

 @article { Zitnick2022SphericalCF ,
    title   = { Spherical Channels for Modeling Atomic Interactions } ,
    author  = { C. Lawrence Zitnick and Abhishek Das and Adeesh Kolluru and Janice Lan and Muhammed Shuaibi and Anuroop Sriram and Zachary W. Ulissi and Brandon C. Wood } ,
    journal = { ArXiv } ,
    year    = { 2022 } ,
    volume  = { abs/2206.14331 }
}

 @article { Passaro2023ReducingSC ,
  title     = { Reducing SO(3) Convolutions to SO(2) for Efficient Equivariant GNNs } ,
  author    = { Saro Passaro and C. Lawrence Zitnick } ,
  journal   = { ArXiv } ,
  year      = { 2023 } ,
  volume    = { abs/2302.03655 }
}

 @inproceedings { Gomez2017TheRR ,
    title   = { The Reversible Residual Network: Backpropagation Without Storing Activations } ,
    author  = { Aidan N. Gomez and Mengye Ren and Raquel Urtasun and Roger Baker Grosse } ,
    booktitle = { NIPS } ,
    year    = { 2017 }
}

 @article { Bondarenko2023QuantizableTR ,
    title   = { Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothing } ,
    author  = { Yelysei Bondarenko and Markus Nagel and Tijmen Blankevoort } ,
    journal = { ArXiv } ,
    year    = { 2023 } ,
    volume  = { abs/2306.12929 } ,
    url     = { https://api.semanticscholar.org/CorpusID:259224568 }
}

 @inproceedings { Arora2023ZoologyMA ,
  title   = { Zoology: Measuring and Improving Recall in Efficient Language Models } ,
  author  = { Simran Arora and Sabri Eyuboglu and Aman Timalsina and Isys Johnson and Michael Poli and James Zou and Atri Rudra and Christopher R'e } ,
  year    = { 2023 } ,
  url     = { https://api.semanticscholar.org/CorpusID:266149332 }
}