Convert torchtyping to jaxtyping for PyTorch>=2.4

Author: iwyoo

chytorch/nn/losses.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Bool, Float
 #
 # Copyright 2023, 2024 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -23,7 +25,7 @@
 from torch import float32, zeros_like, exp, Tensor
 from torch.nn import Parameter, MSELoss
 from torch.nn.modules.loss import _Loss
-from torchtyping import TensorType
+
 
 
 class MultiTaskLoss(_Loss):
@@ -32,15 +34,15 @@ class MultiTaskLoss(_Loss):
 
     https://arxiv.org/abs/1705.07115
     """
-    def __init__(self, loss_type: TensorType['loss_type', bool], *, reduction='mean'):
+    def __init__(self, loss_type: Bool[torch.Tensor, "loss_type"], *, reduction='mean'):
         """
         :param loss_type: vector equal to the number of tasks losses. True for regression and False for classification.
         """
         super().__init__(reduction=reduction)
         self.log = Parameter(zeros_like(loss_type, dtype=float32))
         self.register_buffer('coefficient', (loss_type + 1.).to(float32))
 
-    def forward(self, x: TensorType['loss', float]):
+    def forward(self, x: Float[torch.Tensor, "loss"]):
         """
         :param x: 1d vector of losses or 2d matrix of batch X losses.
         """

chytorch/nn/molecule/encoder.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Float
 #
 # Copyright 2021-2024 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -22,7 +24,7 @@
 #
 from itertools import repeat
 from torch.nn import GELU, Module, ModuleList, LayerNorm
-from torchtyping import TensorType
+
 from typing import Tuple, Optional, List
 from warnings import warn
 from ._embedding import EmbeddingBag
@@ -118,9 +120,9 @@ def __init__(self, max_neighbors: int = 14, max_distance: int = 10, d_model: int
         self._register_load_state_dict_pre_hook(_update)
 
     def forward(self, batch: MoleculeDataBatch, /, *,
-                cache: Optional[List[Tuple[TensorType['batch', 'atoms+conditions', 'embedding'],
-                                           TensorType['batch', 'atoms+conditions', 'embedding']]]] = None) -> \
-            TensorType['batch', 'atoms', 'embedding']:
+                cache: Optional[List[Tuple[Float[torch.Tensor, "batch atoms+conditions embedding"],
+                                           Float[torch.Tensor, "batch atoms+conditions embedding"]]]] = None) -> \
+            Float[torch.Tensor, "batch atoms embedding"]:
         """
         Use 0 for padding.
         Atoms should be coded by atomic numbers + 2.

chytorch/nn/reaction/encoder.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Float
 #
 # Copyright 2021-2023 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -23,7 +25,7 @@
 from math import inf
 from torch import zeros_like, float as t_float
 from torch.nn import Embedding, GELU, Module
-from torchtyping import TensorType
+
 from ..molecule import MoleculeEncoder
 from ..transformer import EncoderLayer
 from ...utils.data import ReactionEncoderDataBatch
@@ -58,7 +60,7 @@ def max_distance(self):
         """
         return self.molecule_encoder.max_distance
 
-    def forward(self, batch: ReactionEncoderDataBatch) -> TensorType['batch', 'atoms', 'embedding']:
+    def forward(self, batch: ReactionEncoderDataBatch) -> Float[torch.Tensor, "batch atoms embedding"]:
         """
         Use 0 for padding. Roles should be coded by 2 for reactants, 3 for products and 1 for special cls token.
         Distances - same as molecular encoder distances but batched diagonally.

chytorch/nn/voting/binary.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Float, Int
 #
 # Copyright 2022, 2023 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -24,7 +26,7 @@
 from torch import sigmoid, no_grad
 from torch.nn import GELU
 from torch.nn.functional import binary_cross_entropy_with_logits
-from torchtyping import TensorType
+
 from typing import Union, Optional
 from ._kfold import k_fold_mask
 from .regressor import VotingRegressor
@@ -42,8 +44,8 @@ def __init__(self, ensemble: int = 10, output: int = 1, hidden: int = 256, input
                          layer_norm_eps, loss_function, norm_first)
 
     @no_grad()
-    def predict(self, x: TensorType['batch', 'embedding'], *,
-                k_fold: Optional[int] = None) -> Union[TensorType['batch', int], TensorType['batch', 'output', int]]:
+    def predict(self, x: Float[torch.Tensor, "batch embedding"], *,
+                k_fold: Optional[int] = None) -> Union[Int[torch.Tensor, "batch"], Int[torch.Tensor, "batch output"]]:
         """
         Average class prediction
 
@@ -53,9 +55,9 @@ def predict(self, x: TensorType['batch', 'embedding'], *,
         return (self.predict_proba(x, k_fold=k_fold) > .5).long()
 
     @no_grad()
-    def predict_proba(self, x: TensorType['batch', 'embedding'], *,
-                      k_fold: Optional[int] = None) -> Union[TensorType['batch', float],
-                                                             TensorType['batch', 'output', float]]:
+    def predict_proba(self, x: Float[torch.Tensor, "batch embedding"], *,
+                      k_fold: Optional[int] = None) -> Union[Float[torch.Tensor, "batch"],
+                                                             Float[torch.Tensor, "batch output"]]:
         """
         Average probability
 

chytorch/nn/voting/classifier.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Float, Int
 #
 # Copyright 2022, 2023 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -24,7 +26,7 @@
 from torch import bmm, no_grad, Tensor
 from torch.nn import Dropout, GELU, LayerNorm, LazyLinear, Linear, Module
 from torch.nn.functional import cross_entropy, softmax
-from torchtyping import TensorType
+
 from typing import Optional, Union
 from ._kfold import k_fold_mask
 
@@ -83,8 +85,8 @@ def forward(self, x):
             return x.view(-1, self._output, self._ensemble, self._n_classes)  # B x O x E x C
         return x  # B x E x C
 
-    def loss(self, x: TensorType['batch', 'embedding'],
-             y: Union[TensorType['batch', 1, int], TensorType['batch', 'output', int]],
+    def loss(self, x: Float[torch.Tensor, "batch embedding"],
+             y: Union[Int[torch.Tensor, "batch 1 int] TensorType[batch output"]],
              k_fold: Optional[int] = None, ignore_index: int = -100) -> Tensor:
         """
         Apply loss function to ensemble of predictions.
@@ -120,8 +122,8 @@ def loss(self, x: TensorType['batch', 'embedding'],
         return self.loss_function(p, y)
 
     @no_grad()
-    def predict(self, x: TensorType['batch', 'embedding'], *,
-                k_fold: Optional[int] = None) -> Union[TensorType['batch', int], TensorType['batch', 'output', int]]:
+    def predict(self, x: Float[torch.Tensor, "batch embedding"], *,
+                k_fold: Optional[int] = None) -> Union[Int[torch.Tensor, "batch"], Int[torch.Tensor, "batch output"]]:
         """
         Average class prediction
 
@@ -130,9 +132,9 @@ def predict(self, x: TensorType['batch', 'embedding'], *,
         return self.predict_proba(x, k_fold=k_fold).argmax(-1)  # B or B x O
 
     @no_grad()
-    def predict_proba(self, x: TensorType['batch', 'embedding'], *,
-                      k_fold: Optional[int] = None) -> Union[TensorType['batch', 'classes', float],
-                                                             TensorType['batch', 'output', 'classes', float]]:
+    def predict_proba(self, x: Float[torch.Tensor, "batch embedding"], *,
+                      k_fold: Optional[int] = None) -> Union[Float[torch.Tensor, "batch classes"],
+                                                             Float[torch.Tensor, "batch output classes"]]:
         """
         Average probability
 

chytorch/nn/voting/regressor.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Float
 #
 # Copyright 2022, 2023 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -24,7 +26,7 @@
 from torch import bmm, no_grad, Tensor
 from torch.nn import Dropout, GELU, LayerNorm, LazyLinear, Linear, Module
 from torch.nn.functional import smooth_l1_loss
-from torchtyping import TensorType
+
 from typing import Optional, Union
 from ._kfold import k_fold_mask
 
@@ -81,8 +83,8 @@ def forward(self, x):
             return x.view(-1, self._output, self._ensemble)  # B x O x E
         return x  # B x E
 
-    def loss(self, x: TensorType['batch', 'embedding'],
-             y: Union[TensorType['batch', 1, float], TensorType['batch', 'output', float]],
+    def loss(self, x: Float[torch.Tensor, "batch embedding"],
+             y: Union[Float[torch.Tensor, "batch 1 float] TensorType[batch output"]],
              k_fold: Optional[int] = None) -> Tensor:
         """
         Apply loss function to ensemble of predictions.
@@ -110,9 +112,9 @@ def loss(self, x: TensorType['batch', 'embedding'],
         return self.loss_function(p, y)
 
     @no_grad()
-    def predict(self, x: TensorType['batch', 'embedding'], *,
-                k_fold: Optional[int] = None) -> Union[TensorType['batch', float],
-                                                       TensorType['batch', 'output', float]]:
+    def predict(self, x: Float[torch.Tensor, "batch embedding"], *,
+                k_fold: Optional[int] = None) -> Union[Float[torch.Tensor, "batch"],
+                                                       Float[torch.Tensor, "batch output"]]:
         """
         Average prediction
 

chytorch/utils/data/molecule/conformer.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Int
 #
 # Copyright 2022-2024 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -27,21 +29,21 @@
 from torch import IntTensor, Size, zeros, ones as t_ones, int32 as t_int32, eye
 from torch.nn.utils.rnn import pad_sequence
 from torch.utils.data import Dataset
-from torchtyping import TensorType
+
 from typing import Sequence, Tuple, Union, NamedTuple
 from .._abc import default_collate_fn_map
 
 
 class ConformerDataPoint(NamedTuple):
-    atoms: TensorType['atoms', int]
-    hydrogens: TensorType['atoms', int]
-    distances: TensorType['atoms', 'atoms', int]
+    atoms: Int[torch.Tensor, "atoms"]
+    hydrogens: Int[torch.Tensor, "atoms"]
+    distances: Int[torch.Tensor, "atoms atoms"]
 
 
 class ConformerDataBatch(NamedTuple):
-    atoms: TensorType['batch', 'atoms', int]
-    hydrogens: TensorType['batch', 'atoms', int]
-    distances: TensorType['batch', 'atoms', 'atoms', int]
+    atoms: Int[torch.Tensor, "batch atoms"]
+    hydrogens: Int[torch.Tensor, "batch atoms"]
+    distances: Int[torch.Tensor, "batch atoms atoms"]
 
     def to(self, *args, **kwargs):
         return ConformerDataBatch(*(x.to(*args, **kwargs) for x in self))

chytorch/utils/data/molecule/encoder.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Int
 #
 # Copyright 2021-2024 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -26,22 +28,22 @@
 from torch import IntTensor, Size, int32, ones, zeros, eye, empty, full
 from torch.nn.utils.rnn import pad_sequence
 from torch.utils.data import Dataset
-from torchtyping import TensorType
+
 from typing import Sequence, Union, NamedTuple, Optional, Tuple
 from zlib import decompress
 from .._abc import default_collate_fn_map
 
 
 class MoleculeDataPoint(NamedTuple):
-    atoms: TensorType['atoms', int]
-    neighbors: TensorType['atoms', int]
-    distances: TensorType['atoms', 'atoms', int]
+    atoms: Int[torch.Tensor, "atoms"]
+    neighbors: Int[torch.Tensor, "atoms"]
+    distances: Int[torch.Tensor, "atoms atoms"]
 
 
 class MoleculeDataBatch(NamedTuple):
-    atoms: TensorType['batch', 'atoms', int]
-    neighbors: TensorType['batch', 'atoms', int]
-    distances: TensorType['batch', 'atoms', 'atoms', int]
+    atoms: Int[torch.Tensor, "batch atoms"]
+    neighbors: Int[torch.Tensor, "batch atoms"]
+    distances: Int[torch.Tensor, "batch atoms atoms"]
 
     def to(self, *args, **kwargs):
         return MoleculeDataBatch(*(x.to(*args, **kwargs) for x in self))

chytorch/utils/data/reaction/encoder.py

@@ -1,4 +1,6 @@
 # -*- coding: utf-8 -*-
+import torch
+from jaxtyping import Int
 #
 # Copyright 2021-2023 Ramil Nugmanov <nougmanoff@protonmail.com>
 #
@@ -25,24 +27,24 @@
 from torch import IntTensor, cat, zeros, int32, Size, eye
 from torch.nn.utils.rnn import pad_sequence
 from torch.utils.data import Dataset
-from torchtyping import TensorType
+
 from typing import Sequence, Union, NamedTuple
 from ..molecule import MoleculeDataset
 from .._abc import default_collate_fn_map
 
 
 class ReactionEncoderDataPoint(NamedTuple):
-    atoms: TensorType['atoms', int]
-    neighbors: TensorType['atoms', int]
-    distances: TensorType['atoms', 'atoms', int]
-    roles: TensorType['atoms', int]
+    atoms: Int[torch.Tensor, "atoms"]
+    neighbors: Int[torch.Tensor, "atoms"]
+    distances: Int[torch.Tensor, "atoms atoms"]
+    roles: Int[torch.Tensor, "atoms"]
 
 
 class ReactionEncoderDataBatch(NamedTuple):
-    atoms: TensorType['batch', 'atoms', int]
-    neighbors: TensorType['batch', 'atoms', int]
-    distances: TensorType['batch', 'atoms', 'atoms', int]
-    roles: TensorType['batch', 'atoms', int]
+    atoms: Int[torch.Tensor, "batch atoms"]
+    neighbors: Int[torch.Tensor, "batch atoms"]
+    distances: Int[torch.Tensor, "batch atoms atoms"]
+    roles: Int[torch.Tensor, "batch atoms"]
 
     def to(self, *args, **kwargs):
         return ReactionEncoderDataBatch(*(x.to(*args, **kwargs) for x in self))

pyproject.toml

@@ -29,7 +29,7 @@ include = [
 
 [tool.poetry.dependencies]
 python = '>=3.8,<3.12'
-torchtyping = '^0.1.4'
+jaxtyping = '^0.3.2'
 chython = '^1.70'
 scipy = '^1.10'
 torch = '>=1.8'