add code

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

dockformerpp/__init__.py

@@ -0,0 +1,6 @@
+from . import model
+from . import utils
+from . import data
+from . import resources
+
+__all__ = ["model", "utils", "data", "resources"]

dockformerpp/config.py

@@ -0,0 +1,344 @@
+import copy
+import ml_collections as mlc
+
+from dockformerpp.utils.config_tools import set_inf, enforce_config_constraints
+
+
+def model_config(
+    name, 
+    train=False, 
+    low_prec=False, 
+    long_sequence_inference=False
+):
+    c = copy.deepcopy(config)
+    # TRAINING PRESETS
+    if name == "initial_training":
+        # AF2 Suppl. Table 4, "initial training" setting
+
+        pass
+    elif name == "finetune_affinity":
+        c.loss.affinity2d.weight = 0.5
+        c.loss.binding_site.weight = 0.5
+        c.loss.positions_inter_distogram.weight = 0.5  # this is not essential given fape?
+    else:
+        raise ValueError("Invalid model name")
+
+    c.globals.use_lma = False
+
+    if long_sequence_inference:
+        assert(not train)
+        c.globals.use_lma = True
+
+    if train:
+        c.globals.blocks_per_ckpt = 1
+        c.globals.use_lma = False
+
+    if low_prec:
+        c.globals.eps = 1e-4
+        # If we want exact numerical parity with the original, inf can't be
+        # a global constant
+        set_inf(c, 1e4)
+
+    enforce_config_constraints(c)
+
+    return c
+
+
+c_z = mlc.FieldReference(128, field_type=int)
+c_m = mlc.FieldReference(256, field_type=int)
+c_t = mlc.FieldReference(64, field_type=int)
+c_e = mlc.FieldReference(64, field_type=int)
+c_s = mlc.FieldReference(384, field_type=int)
+
+
+blocks_per_ckpt = mlc.FieldReference(None, field_type=int)
+aux_distogram_bins = mlc.FieldReference(64, field_type=int)
+aux_affinity_bins = mlc.FieldReference(32, field_type=int)
+eps = mlc.FieldReference(1e-8, field_type=float)
+
+NUM_RES = "num residues placeholder"
+NUM_TOKEN = "num tokens placeholder"
+
+
+config = mlc.ConfigDict(
+    {
+        "data": {
+            "common": {
+                "feat": {
+                    "aatype": [NUM_TOKEN],
+                    "all_atom_mask": [NUM_TOKEN, None],
+                    "all_atom_positions": [NUM_TOKEN, None, None],
+                    "atom14_alt_gt_exists": [NUM_TOKEN, None],
+                    "atom14_alt_gt_positions": [NUM_TOKEN, None, None],
+                    "atom14_atom_exists": [NUM_TOKEN, None],
+                    "atom14_atom_is_ambiguous": [NUM_TOKEN, None],
+                    "atom14_gt_exists": [NUM_TOKEN, None],
+                    "atom14_gt_positions": [NUM_TOKEN, None, None],
+                    "atom37_atom_exists": [NUM_TOKEN, None],
+                    "backbone_rigid_mask": [NUM_TOKEN],
+                    "backbone_rigid_tensor": [NUM_TOKEN, None, None],
+                    "chi_angles_sin_cos": [NUM_TOKEN, None, None],
+                    "chi_mask": [NUM_TOKEN, None],
+                    "no_recycling_iters": [],
+                    "pseudo_beta": [NUM_TOKEN, None],
+                    "pseudo_beta_mask": [NUM_TOKEN],
+                    "residue_index": [NUM_TOKEN],
+                    "in_chain_residue_index": [NUM_TOKEN],
+                    "chain_index": [NUM_TOKEN],
+                    "residx_atom14_to_atom37": [NUM_TOKEN, None],
+                    "residx_atom37_to_atom14": [NUM_TOKEN, None],
+                    "resolution": [],
+                    "rigidgroups_alt_gt_frames": [NUM_TOKEN, None, None, None],
+                    "rigidgroups_group_exists": [NUM_TOKEN, None],
+                    "rigidgroups_group_is_ambiguous": [NUM_TOKEN, None],
+                    "rigidgroups_gt_exists": [NUM_TOKEN, None],
+                    "rigidgroups_gt_frames": [NUM_TOKEN, None, None, None],
+                    "seq_length": [],
+                    "token_mask": [NUM_TOKEN],
+                    "target_feat": [NUM_TOKEN, None],
+                    "use_clamped_fape": [],
+                },
+                "max_recycling_iters": 1,
+                "unsupervised_features": [
+                    "aatype",
+                    "residue_index",
+                    "in_chain_residue_index",
+                    "chain_index",
+                    "seq_length",
+                    "no_recycling_iters",
+                    "all_atom_mask",
+                    "all_atom_positions",
+                ],
+            },
+            "supervised": {
+                "clamp_prob": 0.9,
+                "supervised_features": [
+                    "resolution",
+                    "use_clamped_fape",
+                ],
+            },
+            "predict": {
+                "fixed_size": True,
+                "crop": False,
+                "crop_size": None,
+                "supervised": False,
+                "uniform_recycling": False,
+            },
+            "eval": {
+                "fixed_size": True,
+                "crop": False,
+                "crop_size": None,
+                "supervised": True,
+                "uniform_recycling": False,
+            },
+            "train": {
+                "fixed_size": True,
+                "crop": True,
+                "crop_size": 355,
+                "supervised": True,
+                "clamp_prob": 0.9,
+                "uniform_recycling": True,
+                "distogram_mask_prob": 0.1,
+            },
+            "data_module": {
+                "data_loaders": {
+                    "batch_size": 1,
+                    # "batch_size": 2,
+                    "num_workers": 16,
+                    "pin_memory": True,
+                    "should_verify": False,
+                },
+            },
+        },
+        # Recurring FieldReferences that can be changed globally here
+        "globals": {
+            "blocks_per_ckpt": blocks_per_ckpt,
+            # Use Staats & Rabe's low-memory attention algorithm.
+            "use_lma": False,
+            "max_lr": 1e-3,
+            "c_z": c_z,
+            "c_m": c_m,
+            "c_t": c_t,
+            "c_e": c_e,
+            "c_s": c_s,
+            "eps": eps,
+        },
+        "model": {
+            "_mask_trans": False,
+            "structure_input_embedder": {
+                "protein_tf_dim": 20,
+                "additional_tf_dim": 3, # number of classes (prot_r, prot_l, aff)
+                "c_z": c_z,
+                "c_m": c_m,
+                "relpos_k": 32,
+                "prot_min_bin": 3.25,
+                "prot_max_bin": 20.75,
+                "prot_no_bins": 15,
+                "inf": 1e8,
+            },
+            "recycling_embedder": {
+                "c_z": c_z,
+                "c_m": c_m,
+                "min_bin": 3.25,
+                "max_bin": 20.75,
+                "no_bins": 15,
+                "inf": 1e8,
+            },
+            "evoformer_stack": {
+                "c_m": c_m,
+                "c_z": c_z,
+                "c_hidden_single_att": 32,
+                "c_hidden_mul": 128,
+                "c_hidden_pair_att": 32,
+                "c_s": c_s,
+                "no_heads_single": 8,
+                "no_heads_pair": 4,
+                # "no_blocks": 48,
+                "no_blocks": 2,
+                "transition_n": 4,
+                "single_dropout": 0.15,
+                "pair_dropout": 0.25,
+                "blocks_per_ckpt": blocks_per_ckpt,
+                "clear_cache_between_blocks": False,
+                "inf": 1e9,
+                "eps": eps,  # 1e-10,
+            },
+            "structure_module": {
+                "c_s": c_s,
+                "c_z": c_z,
+                "c_ipa": 16,
+                "c_resnet": 128,
+                "no_heads_ipa": 12,
+                "no_qk_points": 4,
+                "no_v_points": 8,
+                "dropout_rate": 0.1,
+                "no_blocks": 8,
+                "no_transition_layers": 1,
+                "no_resnet_blocks": 2,
+                "no_angles": 7,
+                "trans_scale_factor": 10,
+                "epsilon": eps,  # 1e-12,
+                "inf": 1e5,
+            },
+            "heads": {
+                "lddt": {
+                    "no_bins": 50,
+                    "c_in": c_s,
+                    "c_hidden": 128,
+                },
+                "distogram": {
+                    "c_z": c_z,
+                    "no_bins": aux_distogram_bins,
+                },
+                "affinity_2d": {
+                    "c_z": c_z,
+                    "num_bins": aux_affinity_bins,
+                },
+                "affinity_1d": {
+                    "c_s": c_s,
+                    "num_bins": aux_affinity_bins,
+                },
+                "affinity_cls": {
+                    "c_s": c_s,
+                    "num_bins": aux_affinity_bins,
+                },
+                "binding_site": {
+                    "c_s": c_s,
+                    "c_out": 1,
+                },
+                "inter_contact": {
+                    "c_s": c_s,
+                    "c_z": c_z,
+                    "c_out": 1,
+                },
+            },
+            # A negative value indicates that no early stopping will occur, i.e.
+            # the model will always run `max_recycling_iters` number of recycling
+            # iterations. A positive value will enable early stopping if the
+            # difference in pairwise distances is less than the tolerance between
+            # recycling steps.
+            "recycle_early_stop_tolerance": -1.
+        },
+        "relax": {
+            "max_iterations": 0,  # no max
+            "tolerance": 2.39,
+            "stiffness": 10.0,
+            "max_outer_iterations": 20,
+            "exclude_residues": [],
+        },
+        "loss": {
+            "distogram": {
+                "min_bin": 2.3125,
+                "max_bin": 21.6875,
+                "no_bins": 64,
+                "eps": eps,  # 1e-6,
+                "weight": 0.3,
+            },
+            "positions_inter_distogram": {
+                "max_dist": 20.0,
+                "weight": 0.0,
+            },
+            "positions_intra_distogram": {
+                "max_dist": 10.0,
+                "weight": 0.0,
+            },
+            "binding_site": {
+                "weight": 0.0,
+                "pos_class_weight": 20.0,
+            },
+            "inter_contact": {
+                "weight": 0.0,
+                "pos_class_weight": 200.0,
+            },
+            "affinity2d": {
+                "min_bin": 0,
+                "max_bin": 15,
+                "no_bins": aux_affinity_bins,
+                "weight": 0.0,
+            },
+            "affinity_cls": {
+                "min_bin": 0,
+                "max_bin": 15,
+                "no_bins": aux_affinity_bins,
+                "weight": 0.0,
+            },
+            "fape_backbone": {
+                "clamp_distance": 10.0,
+                "loss_unit_distance": 10.0,
+                "weight": 0.5,
+            },
+            "fape_sidechain": {
+                "clamp_distance": 10.0,
+                "length_scale": 10.0,
+                "weight": 0.5,
+            },
+            "fape_interface": {
+                "clamp_distance": 10.0,
+                "length_scale": 10.0,
+                "weight": 0.0,
+            },
+            "plddt_loss": {
+                "min_resolution": 0.1,
+                "max_resolution": 3.0,
+                "cutoff": 15.0,
+                "no_bins": 50,
+                "eps": eps,  # 1e-10,
+                "weight": 0.01,
+            },
+            "supervised_chi": {
+                "chi_weight": 0.5,
+                "angle_norm_weight": 0.01,
+                "eps": eps,  # 1e-6,
+                "weight": 1.0,
+            },
+            "chain_center_of_mass": {
+                "clamp_distance": -4.0,
+                "weight": 0.,
+                "eps": eps,
+                "enabled": False,
+            },
+            "eps": eps,
+        },
+        "ema": {"decay": 0.999},
+    }
+)

dockformerpp/data/data_modules.py

@@ -0,0 +1,643 @@
+import copy
+import itertools
+import time
+import traceback
+from collections import Counter
+from functools import partial
+import json
+import os
+import pickle
+from typing import Optional, Sequence, Any
+
+import ml_collections as mlc
+import lightning as L
+import torch
+from torch.utils.data import RandomSampler
+
+from dockformerpp.data.data_pipeline import parse_input_json
+from dockformerpp.data import data_pipeline
+from dockformerpp.utils.tensor_utils import dict_multimap
+from dockformerpp.utils.tensor_utils import (
+    tensor_tree_map,
+)
+
+
+class OpenFoldSingleDataset(torch.utils.data.Dataset):
+    def __init__(self,
+                 data_dir: str,
+                 config: mlc.ConfigDict,
+                 mode: str = "train",
+                 ):
+        """
+            Args:
+                data_dir:
+                    A path to a directory containing mmCIF files (in train
+                    mode) or FASTA files (in inference mode).
+                config:
+                    A dataset config object. See openfold.config
+                mode:
+                    "train", "val", or "predict"
+        """
+        super(OpenFoldSingleDataset, self).__init__()
+        self.data_dir = data_dir
+
+        self.config = config
+        self.mode = mode
+
+        valid_modes = ["train", "eval", "predict"]
+        if mode not in valid_modes:
+            raise ValueError(f'mode must be one of {valid_modes}')
+
+        self._all_input_files = [i for i in os.listdir(data_dir) if i.endswith(".json")]
+        if self.config.data_module.data_loaders.should_verify:
+            self._all_input_files = [i for i in self._all_input_files if self._verify_json_input_file(i)]
+
+        self.data_pipeline = data_pipeline.DataPipeline(config, mode)
+
+    def _verify_json_input_file(self, file_name: str) -> bool:
+        with open(os.path.join(self.data_dir, file_name), "r") as f:
+            try:
+                loaded = json.load(f)
+                for i in ["input_structure"]:
+                    if i not in loaded:
+                        return False
+                if self.mode != "predict":
+                    for i in ["gt_structure", "resolution"]:
+                        if i not in loaded:
+                            return False
+            except json.JSONDecodeError:
+                return False
+        return True
+
+    def get_metadata_for_idx(self, idx: int) -> dict:
+        input_path = os.path.join(self.data_dir, self._all_input_files[idx])
+        input_data = json.load(open(input_path, "r"))
+        metadata = {
+            "resolution": input_data.get("resolution", 99.0),
+            "input_path": input_path,
+            "input_name": os.path.basename(input_path).split(".json")[0],
+        }
+        return metadata
+
+    def __getitem__(self, idx):
+        return parse_input_json(
+            input_path=os.path.join(self.data_dir, self._all_input_files[idx]),
+            mode=self.mode,
+            config=self.config,
+            data_pipeline=self.data_pipeline,
+            data_dir=os.path.dirname(self.data_dir),
+            idx=idx,
+        )
+
+    def __len__(self):
+        return len(self._all_input_files)
+
+
+def resolution_filter(resolution: int, max_resolution: float) -> bool:
+    """Check that the resolution is <= max_resolution permitted"""
+    return resolution is not None and resolution <= max_resolution
+
+
+def all_seq_len_filter(seqs: list, minimum_number_of_residues: int) -> bool:
+    """Check if the total combined sequence lengths are >= minimum_numer_of_residues"""
+    total_len = sum([len(i) for i in seqs])
+    return total_len >= minimum_number_of_residues
+
+
+class OpenFoldDataset(torch.utils.data.Dataset):
+    """
+        Implements the stochastic filters applied during AlphaFold's training.
+        Because samples are selected from constituent datasets randomly, the
+        length of an OpenFoldFilteredDataset is arbitrary. Samples are selected
+        and filtered once at initialization.
+    """
+
+    def __init__(self,
+                 datasets: Sequence[OpenFoldSingleDataset],
+                 probabilities: Sequence[float],
+                 epoch_len: int,
+                 generator: torch.Generator = None,
+                 _roll_at_init: bool = True,
+                 ):
+        self.datasets = datasets
+        self.probabilities = probabilities
+        self.epoch_len = epoch_len
+        self.generator = generator
+
+        self._samples = [self.looped_samples(i) for i in range(len(self.datasets))]
+        if _roll_at_init:
+            self.reroll()
+
+    @staticmethod
+    def deterministic_train_filter(
+        cache_entry: Any,
+        max_resolution: float = 9.,
+        max_single_aa_prop: float = 0.8,
+        *args, **kwargs
+    ) -> bool:
+        # Hard filters
+        resolution = cache_entry["resolution"]
+
+        return all([
+            resolution_filter(resolution=resolution,
+                              max_resolution=max_resolution)
+        ])
+
+    @staticmethod
+    def get_stochastic_train_filter_prob(
+        cache_entry: Any,
+        *args, **kwargs
+    ) -> float:
+        # Stochastic filters
+        probabilities = []
+
+        cluster_size = cache_entry.get("cluster_size", None)
+        if cluster_size is not None and cluster_size > 0:
+            probabilities.append(1 / cluster_size)
+
+        # Risk of underflow here?
+        out = 1
+        for p in probabilities:
+            out *= p
+
+        return out
+
+    def looped_shuffled_dataset_idx(self, dataset_len):
+        while True:
+            # Uniformly shuffle each dataset's indices
+            weights = [1. for _ in range(dataset_len)]
+            shuf = torch.multinomial(
+                torch.tensor(weights),
+                num_samples=dataset_len,
+                replacement=False,
+                generator=self.generator,
+            )
+            for idx in shuf:
+                yield idx
+
+    def looped_samples(self, dataset_idx):
+        max_cache_len = int(self.epoch_len * self.probabilities[dataset_idx])
+        dataset = self.datasets[dataset_idx]
+        idx_iter = self.looped_shuffled_dataset_idx(len(dataset))
+        while True:
+            weights = []
+            idx = []
+            for _ in range(max_cache_len):
+                candidate_idx = next(idx_iter)
+                # chain_id = dataset.idx_to_chain_id(candidate_idx)
+                # chain_data_cache_entry = chain_data_cache[chain_id]
+                # data_entry = dataset[candidate_idx.item()]
+                entry_metadata_for_filter = dataset.get_metadata_for_idx(candidate_idx.item())
+                if not self.deterministic_train_filter(entry_metadata_for_filter):
+                    continue
+
+                p = self.get_stochastic_train_filter_prob(
+                    entry_metadata_for_filter,
+                )
+                weights.append([1. - p, p])
+                idx.append(candidate_idx)
+
+            samples = torch.multinomial(
+                torch.tensor(weights),
+                num_samples=1,
+                generator=self.generator,
+            )
+            samples = samples.squeeze()
+
+            cache = [i for i, s in zip(idx, samples) if s]
+
+            for datapoint_idx in cache:
+                yield datapoint_idx
+
+    def __getitem__(self, idx):
+        dataset_idx, datapoint_idx = self.datapoints[idx]
+        return self.datasets[dataset_idx][datapoint_idx]
+
+    def __len__(self):
+        return self.epoch_len
+
+    def reroll(self):
+        # TODO bshor: I have removed support for filters (currently done in preprocess) and to weighting clusters
+        # now it is much faster, because it doesn't call looped_samples
+        dataset_choices = torch.multinomial(
+            torch.tensor(self.probabilities),
+            num_samples=self.epoch_len,
+            replacement=True,
+            generator=self.generator,
+        )
+        self.datapoints = []
+        counter_datasets = Counter(dataset_choices.tolist())
+        for dataset_idx, num_samples in counter_datasets.items():
+            dataset = self.datasets[dataset_idx]
+            sample_choices = torch.randint(0, len(dataset), (num_samples,), generator=self.generator)
+            for datapoint_idx in sample_choices:
+                self.datapoints.append((dataset_idx, datapoint_idx))
+
+
+class OpenFoldBatchCollator:
+    def __call__(self, prots):
+        stack_fn = partial(torch.stack, dim=0)
+        return dict_multimap(stack_fn, prots)
+
+
+class OpenFoldDataLoader(torch.utils.data.DataLoader):
+    def __init__(self, *args, config, stage="train", generator=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.config = config
+        self.stage = stage
+        self.generator = generator
+        self._prep_batch_properties_probs()
+
+    def _prep_batch_properties_probs(self):
+        keyed_probs = []
+        stage_cfg = self.config[self.stage]
+
+        max_iters = self.config.common.max_recycling_iters
+
+        if stage_cfg.uniform_recycling:
+            recycling_probs = [
+                1. / (max_iters + 1) for _ in range(max_iters + 1)
+            ]
+        else:
+            recycling_probs = [
+                0. for _ in range(max_iters + 1)
+            ]
+            recycling_probs[-1] = 1.
+
+        keyed_probs.append(
+            ("no_recycling_iters", recycling_probs)
+        )
+
+        keys, probs = zip(*keyed_probs)
+        max_len = max([len(p) for p in probs])
+        padding = [[0.] * (max_len - len(p)) for p in probs]
+
+        self.prop_keys = keys
+        self.prop_probs_tensor = torch.tensor(
+            [p + pad for p, pad in zip(probs, padding)],
+            dtype=torch.float32,
+        )
+
+    def _add_batch_properties(self, batch):
+        # gt_features = batch.pop('gt_features', None)
+        samples = torch.multinomial(
+            self.prop_probs_tensor,
+            num_samples=1,  # 1 per row
+            replacement=True,
+            generator=self.generator
+        )
+
+        aatype = batch["aatype"]
+        batch_dims = aatype.shape[:-2]
+        recycling_dim = aatype.shape[-1]
+        no_recycling = recycling_dim
+        for i, key in enumerate(self.prop_keys):
+            sample = int(samples[i][0])
+            sample_tensor = torch.tensor(
+                sample,
+                device=aatype.device,
+                requires_grad=False
+            )
+            orig_shape = sample_tensor.shape
+            sample_tensor = sample_tensor.view(
+                (1,) * len(batch_dims) + sample_tensor.shape + (1,)
+            )
+            sample_tensor = sample_tensor.expand(
+                batch_dims + orig_shape + (recycling_dim,)
+            )
+            batch[key] = sample_tensor
+
+            if key == "no_recycling_iters":
+                no_recycling = sample
+
+        resample_recycling = lambda t: t[..., :no_recycling + 1]
+        batch = tensor_tree_map(resample_recycling, batch)
+        # batch['gt_features'] = gt_features
+
+        return batch
+
+    def __iter__(self):
+        it = super().__iter__()
+
+        def _batch_prop_gen(iterator):
+            for batch in iterator:
+                yield self._add_batch_properties(batch)
+
+        return _batch_prop_gen(it)
+
+
+class OpenFoldDataModule(L.LightningDataModule):
+    def __init__(self,
+                 config: mlc.ConfigDict,
+                 train_data_dir: Optional[str] = None,
+                 val_data_dir: Optional[str] = None,
+                 predict_data_dir: Optional[str] = None,
+                 batch_seed: Optional[int] = None,
+                 train_epoch_len: int = 50000,
+                 **kwargs
+                 ):
+        super(OpenFoldDataModule, self).__init__()
+
+        self.config = config
+        self.train_data_dir = train_data_dir
+        self.val_data_dir = val_data_dir
+        self.predict_data_dir = predict_data_dir
+        self.batch_seed = batch_seed
+        self.train_epoch_len = train_epoch_len
+
+        if self.train_data_dir is None and self.predict_data_dir is None:
+            raise ValueError(
+                'At least one of train_data_dir or predict_data_dir must be '
+                'specified'
+            )
+
+        self.training_mode = self.train_data_dir is not None
+
+        # if not self.training_mode and predict_alignment_dir is None:
+        #     raise ValueError(
+        #         'In inference mode, predict_alignment_dir must be specified'
+        #     )
+        # elif val_data_dir is not None and val_alignment_dir is None:
+        #     raise ValueError(
+        #         'If val_data_dir is specified, val_alignment_dir must '
+        #         'be specified as well'
+        #     )
+
+    def setup(self, stage):
+        # Most of the arguments are the same for the three datasets 
+        dataset_gen = partial(OpenFoldSingleDataset,
+                              config=self.config)
+
+        if self.training_mode:
+            train_dataset = dataset_gen(
+                data_dir=self.train_data_dir,
+                mode="train",
+            )
+
+            datasets = [train_dataset]
+            probabilities = [1.]
+
+            generator = None
+            if self.batch_seed is not None:
+                generator = torch.Generator()
+                generator = generator.manual_seed(self.batch_seed + 1)
+
+            self.train_dataset = OpenFoldDataset(
+                datasets=datasets,
+                probabilities=probabilities,
+                epoch_len=self.train_epoch_len,
+                generator=generator,
+                _roll_at_init=False,
+            )
+
+            if self.val_data_dir is not None:
+                self.eval_dataset = dataset_gen(
+                    data_dir=self.val_data_dir,
+                    mode="eval",
+                )
+            else:
+                self.eval_dataset = None
+        else:
+            self.predict_dataset = dataset_gen(
+                data_dir=self.predict_data_dir,
+                mode="predict",
+            )
+
+    def _gen_dataloader(self, stage):
+        generator = None
+        if self.batch_seed is not None:
+            generator = torch.Generator()
+            generator = generator.manual_seed(self.batch_seed)
+
+        if stage == "train":
+            dataset = self.train_dataset
+            # Filter the dataset, if necessary
+            dataset.reroll()
+        elif stage == "eval":
+            dataset = self.eval_dataset
+        elif stage == "predict":
+            dataset = self.predict_dataset
+        else:
+            raise ValueError("Invalid stage")
+
+        batch_collator = OpenFoldBatchCollator()
+
+        dl = OpenFoldDataLoader(
+            dataset,
+            config=self.config,
+            stage=stage,
+            generator=generator,
+            batch_size=self.config.data_module.data_loaders.batch_size,
+            # num_workers=self.config.data_module.data_loaders.num_workers,
+            num_workers=0, # TODO bshor: solve generator pickling issue and then bring back num_workers, or just remove generator
+            collate_fn=batch_collator,
+        )
+
+        return dl
+
+    def train_dataloader(self):
+        return self._gen_dataloader("train")
+
+    def val_dataloader(self):
+        if self.eval_dataset is not None:
+            return self._gen_dataloader("eval")
+        return None
+
+    def predict_dataloader(self):
+        return self._gen_dataloader("predict")
+
+
+class DummyDataset(torch.utils.data.Dataset):
+    def __init__(self, batch_path):
+        with open(batch_path, "rb") as f:
+            self.batch = pickle.load(f)
+
+    def __getitem__(self, idx):
+        return copy.deepcopy(self.batch)
+
+    def __len__(self):
+        return 1000
+
+
+class DummyDataLoader(L.LightningDataModule):
+    def __init__(self, batch_path):
+        super().__init__()
+        self.dataset = DummyDataset(batch_path)
+
+    def train_dataloader(self):
+        return torch.utils.data.DataLoader(self.dataset)
+
+
+class DockFormerSimpleDataset(torch.utils.data.Dataset):
+    def __init__(self, clusters_json: str, config: mlc.ConfigDict, mode: str = "train"):
+        clusters = json.load(open(clusters_json, "r"))
+        self.config = config
+        self.mode = mode
+        self._data_dir = os.path.dirname(clusters_json)
+        print("Data dir", self._data_dir)
+        self._clusters = clusters
+        self._all_input_files = sum(clusters.values(), [])
+        self.data_pipeline = data_pipeline.DataPipeline(config, mode)
+
+    def __getitem__(self, idx):
+        return parse_input_json(
+            input_path=os.path.join(self._data_dir, self._all_input_files[idx]),
+            mode=self.mode,
+            config=self.config,
+            data_pipeline=self.data_pipeline,
+            data_dir=self._data_dir,
+            idx=idx,
+        )
+
+    def __len__(self):
+        return len(self._all_input_files)
+
+
+class DockFormerClusteredDataset(torch.utils.data.Dataset):
+    def __init__(self, clusters_json: str, config: mlc.ConfigDict, mode: str = "train", generator=None):
+        clusters = json.load(open(clusters_json, "r"))
+        self.config = config
+        self.mode = mode
+        self._data_dir = os.path.dirname(clusters_json)
+        self._clusters = list(clusters.values())
+        self.data_pipeline = data_pipeline.DataPipeline(config, mode)
+        self._generator = generator
+
+    def __getitem__(self, idx):
+        try:
+            cluster = self._clusters[idx]
+            # choose random from cluster
+            input_file = cluster[torch.randint(0, len(cluster), (1,), generator=self._generator).item()]
+
+            return parse_input_json(
+                input_path=os.path.join(self._data_dir, input_file),
+                mode=self.mode,
+                config=self.config,
+                data_pipeline=self.data_pipeline,
+                data_dir=self._data_dir,
+                idx=idx,
+            )
+        except Exception as e:
+            print("ERROR in loading", e)
+            traceback.print_exc()
+            return parse_input_json(
+                input_path=os.path.join(self._data_dir, self._clusters[0][0]),
+                mode=self.mode,
+                config=self.config,
+                data_pipeline=self.data_pipeline,
+                data_dir=self._data_dir,
+                idx=idx,
+            )
+
+
+    def __len__(self):
+        return len(self._clusters)
+
+
+class DockFormerDataLoader(torch.utils.data.DataLoader):
+    def __init__(self, *args, config, stage="train", generator=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.config = config
+        self.stage = stage
+        # self.generator = generator
+
+    def _add_batch_properties(self, batch):
+        if self.config[self.stage].uniform_recycling:
+            aatype = batch["aatype"]
+            max_recycling_dim = aatype.shape[-1]
+
+            # num_recycles = torch.randint(0, max_recycling_dim, (1,), generator=self.generator)
+            num_recycles = torch.randint(0, max_recycling_dim, (1,)).item()
+
+            resample_recycling = lambda t: t[..., :num_recycles + 1]
+            batch = tensor_tree_map(resample_recycling, batch)
+
+        return batch
+
+    def __iter__(self):
+        it = super().__iter__()
+
+        def _batch_prop_gen(iterator):
+            for batch in iterator:
+                yield self._add_batch_properties(batch)
+
+        return _batch_prop_gen(it)
+
+
+class DockFormerDataModule(L.LightningDataModule):
+    def __init__(self,
+                 config: mlc.ConfigDict,
+                 train_data_file: Optional[str] = None,
+                 val_data_file: Optional[str] = None,
+                 batch_seed: Optional[int] = None,
+                 **kwargs
+                 ):
+        super(DockFormerDataModule, self).__init__()
+
+        self.config = config
+        self.train_data_file = train_data_file
+        self.val_data_file = val_data_file
+        self.batch_seed = batch_seed
+
+        assert self.train_data_file is not None, "train_data_file must be specified"
+        assert self.val_data_file is not None, "val_data_file must be specified"
+
+        self.train_dataset = None
+        self.val_dataset = None
+
+    def setup(self, stage):
+        generator = None
+        if self.batch_seed is not None:
+            generator = torch.Generator()
+            generator = generator.manual_seed(self.batch_seed + 1)
+
+        self.train_dataset = DockFormerClusteredDataset(
+            clusters_json=self.train_data_file,
+            config=self.config,
+            mode="train",
+            generator=generator,
+        )
+
+        self.val_dataset = DockFormerSimpleDataset(
+            clusters_json=self.val_data_file,
+            config=self.config,
+            mode="eval",
+        )
+
+    def _gen_dataloader(self, stage):
+        generator = None
+        if self.batch_seed is not None:
+            generator = torch.Generator()
+            generator = generator.manual_seed(self.batch_seed)
+
+        should_shuffle = stage == "train"
+        if stage == "train":
+            dataset = self.train_dataset
+        elif stage == "eval":
+            dataset = self.val_dataset
+        else:
+            raise ValueError("Invalid stage")
+
+        batch_collator = OpenFoldBatchCollator()
+
+        dl = DockFormerDataLoader(
+            dataset,
+            config=self.config,
+            stage=stage,
+            # generator=generator,
+            batch_size=self.config.data_module.data_loaders.batch_size,
+            # num_workers=self.config.data_module.data_loaders.num_workers,
+            num_workers=0, # TODO bshor: solve generator pickling issue and then bring back num_workers, or just remove generator
+            collate_fn=batch_collator,
+            shuffle=should_shuffle,
+        )
+
+        return dl
+
+    def train_dataloader(self):
+        return self._gen_dataloader("train")
+
+    def val_dataloader(self):
+        if self.val_dataset is not None:
+            return self._gen_dataloader("eval")
+        return None

dockformerpp/data/data_pipeline.py

@@ -0,0 +1,360 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import json
+import os
+import time
+from typing import List
+
+import numpy as np
+import torch
+import ml_collections as mlc
+from rdkit import Chem
+
+from dockformerpp.data import data_transforms
+from dockformerpp.data.data_transforms import get_restype_atom37_mask, get_restypes
+from dockformerpp.data.protein_features import make_protein_features
+from dockformerpp.data.utils import FeatureTensorDict, FeatureDict
+from dockformerpp.utils import protein
+
+
+def _np_filter_and_to_tensor_dict(np_example: FeatureDict, features_to_keep: List[str]) -> FeatureTensorDict:
+    """Creates dict of tensors from a dict of NumPy arrays.
+
+    Args:
+        np_example: A dict of NumPy feature arrays.
+        features: A list of strings of feature names to be returned in the dataset.
+
+    Returns:
+        A dictionary of features mapping feature names to features. Only the given
+        features are returned, all other ones are filtered out.
+    """
+    # torch generates warnings if feature is already a torch Tensor
+    to_tensor = lambda t: torch.tensor(t) if type(t) != torch.Tensor else t.clone().detach()
+    tensor_dict = {
+        k: to_tensor(v) for k, v in np_example.items() if k in features_to_keep
+    }
+
+    return tensor_dict
+
+
+def _add_protein_probablistic_features(features: FeatureDict, cfg: mlc.ConfigDict, mode: str) -> FeatureDict:
+    if mode == "train":
+        p = torch.rand(1).item()
+        use_clamped_fape_value = float(p < cfg.supervised.clamp_prob)
+        features["use_clamped_fape"] = np.float32(use_clamped_fape_value)
+    else:
+        features["use_clamped_fape"] = np.float32(0.0)
+    return features
+
+
+@data_transforms.curry1
+def compose(x, fs):
+    for f in fs:
+        x = f(x)
+    return x
+
+
+def _apply_protein_transforms(tensors: FeatureTensorDict) -> FeatureTensorDict:
+    transforms = [
+        data_transforms.cast_to_64bit_ints,
+        data_transforms.squeeze_features,
+        data_transforms.make_atom14_masks,
+        data_transforms.make_atom14_positions,
+        data_transforms.atom37_to_frames,
+        data_transforms.atom37_to_torsion_angles(""),
+        data_transforms.make_pseudo_beta(),
+        data_transforms.get_backbone_frames,
+        data_transforms.get_chi_angles,
+    ]
+
+    tensors = compose(transforms)(tensors)
+
+    return tensors
+
+
+def _apply_protein_probablistic_transforms(tensors: FeatureTensorDict, cfg: mlc.ConfigDict, mode: str) \
+        -> FeatureTensorDict:
+    transforms = [data_transforms.make_target_feat()]
+
+    crop_feats = dict(cfg.common.feat)
+
+    if cfg[mode].fixed_size:
+        transforms.append(data_transforms.select_feat(list(crop_feats)))
+        # TODO bshor: restore transforms for training on cropped proteins, need to handle pocket somehow
+        # if so, look for random_crop_to_size and make_fixed_size in data_transforms.py
+
+    compose(transforms)(tensors)
+
+    return tensors
+
+
+class DataPipeline:
+    """Assembles input features."""
+    def __init__(self, config: mlc.ConfigDict, mode: str):
+        self.config = config
+        self.mode = mode
+
+        self.feature_names = config.common.unsupervised_features
+        if config[mode].supervised:
+            self.feature_names += config.supervised.supervised_features
+
+    def process_pdb(self, pdb_path: str) -> FeatureTensorDict:
+        """
+            Assembles features for a protein in a PDB file.
+        """
+        with open(pdb_path, 'r') as f:
+            pdb_str = f.read()
+
+        protein_object = protein.from_pdb_string(pdb_str)
+        description = os.path.splitext(os.path.basename(pdb_path))[0].upper()
+        pdb_feats = make_protein_features(protein_object, description)
+        pdb_feats = _add_protein_probablistic_features(pdb_feats, self.config, self.mode)
+
+        tensor_feats = _np_filter_and_to_tensor_dict(pdb_feats, self.feature_names)
+
+        tensor_feats = _apply_protein_transforms(tensor_feats)
+        tensor_feats = _apply_protein_probablistic_transforms(tensor_feats, self.config, self.mode)
+
+        return tensor_feats
+
+def _prepare_recycles(feat: torch.Tensor, num_recycles: int) -> torch.Tensor:
+    return feat.unsqueeze(-1).repeat(*([1] * len(feat.shape)), num_recycles)
+
+
+def _fit_to_crop(target_tensor: torch.Tensor, crop_size: int, start_ind: int) -> torch.Tensor:
+    if len(target_tensor.shape) == 1:
+        ret = torch.zeros((crop_size, ), dtype=target_tensor.dtype)
+        ret[start_ind:start_ind + target_tensor.shape[0]] = target_tensor
+        return ret
+    elif len(target_tensor.shape) == 2:
+        ret = torch.zeros((crop_size, target_tensor.shape[-1]), dtype=target_tensor.dtype)
+        ret[start_ind:start_ind + target_tensor.shape[0], :] = target_tensor
+        return ret
+    else:
+        ret = torch.zeros((crop_size, *target_tensor.shape[1:]), dtype=target_tensor.dtype)
+        ret[start_ind:start_ind + target_tensor.shape[0], ...] = target_tensor
+        return ret
+
+
+def parse_input_json(input_path: str, mode: str, config: mlc.ConfigDict, data_pipeline: DataPipeline,
+                     data_dir: str, idx: int) -> FeatureTensorDict:
+    start_load_time = time.time()
+    input_data = json.load(open(input_path, "r"))
+    if mode == "train" or mode == "eval":
+        print("loading", input_data["pdb_id"], end=" ")
+
+    num_recycles = config.common.max_recycling_iters + 1
+
+    input_protein_r_feats = data_pipeline.process_pdb(pdb_path=os.path.join(data_dir, input_data["input_r_structure"]))
+    input_protein_l_feats = data_pipeline.process_pdb(pdb_path=os.path.join(data_dir, input_data["input_l_structure"]))
+
+    n_res_r = input_protein_r_feats["protein_target_feat"].shape[0]
+    n_res_l = input_protein_l_feats["protein_target_feat"].shape[0]
+    n_res_total = n_res_r + n_res_l
+    n_affinity = 1
+
+    # add 1 for affinity token
+    crop_size = n_res_total + n_affinity
+    if (mode == "train" or mode == "eval") and config.train.fixed_size:
+        crop_size = config.train.crop_size
+
+    assert crop_size >= n_res_total + n_affinity, f"crop_size: {crop_size}, n_res_r: {n_res_r}, n_res_l: {n_res_l}"
+
+    token_mask = torch.zeros((crop_size,), dtype=torch.float32)
+    token_mask[:n_res_total + n_affinity] = 1
+
+    protein_r_mask = torch.zeros((crop_size,), dtype=torch.float32)
+    protein_r_mask[:n_res_r] = 1
+
+    protein_l_mask = torch.zeros((crop_size,), dtype=torch.float32)
+    protein_l_mask[n_res_r:n_res_total] = 1
+
+    affinity_mask = torch.zeros((crop_size,), dtype=torch.float32)
+    affinity_mask[n_res_total] = 1
+
+    structural_mask = torch.zeros((crop_size,), dtype=torch.float32)
+    structural_mask[:n_res_total] = 1
+
+    inter_pair_mask = torch.zeros((crop_size, crop_size), dtype=torch.float32)
+    inter_pair_mask[:n_res_r, n_res_r:n_res_total] = 1
+    inter_pair_mask[n_res_r:n_res_total, :n_res_r] = 1
+
+    tf_dim = input_protein_r_feats["protein_target_feat"].shape[-1]
+
+    target_feat = torch.zeros((crop_size, tf_dim + 3), dtype=torch.float32)
+    target_feat[:n_res_r, :tf_dim] = input_protein_r_feats["protein_target_feat"]
+    target_feat[n_res_r:n_res_total, :tf_dim] = input_protein_l_feats["protein_target_feat"]
+
+    target_feat[:n_res_r, tf_dim] = 1  # Set "is_protein_r" flag for protein rows
+    target_feat[n_res_r:n_res_total, tf_dim + 1] = 1  # Set "is_protein_l" flag for ligand rows
+    target_feat[n_res_total, tf_dim + 2] = 1  # Set "is_affinity" flag for affinity row
+
+    input_positions = torch.zeros((crop_size, 3), dtype=torch.float32)
+    input_positions[:n_res_r] = input_protein_r_feats["pseudo_beta"]
+    input_positions[n_res_r:n_res_total] = input_protein_l_feats["pseudo_beta"]
+
+    distogram_mask = torch.zeros(crop_size)
+    if mode == "train":
+        ones_indices = torch.randperm(n_res_total)[:int(n_res_total * config.train.distogram_mask_prob)]
+        # print(ones_indices)
+        distogram_mask[ones_indices] = 1
+        input_positions = input_positions * (1 - distogram_mask).unsqueeze(-1)
+    elif mode == "predict":
+        # ignore all positions where pseudo_beta is 0, 0, 0
+        distogram_mask = (input_positions == 0).all(dim=-1).float()
+        # print("Ignoring residues", torch.nonzero(distogram_mask).flatten())
+
+    # Implement ligand as amino acid type 20
+    aatype = torch.cat([input_protein_r_feats["aatype"], input_protein_l_feats["aatype"]], dim=0)
+    residue_index = torch.cat([input_protein_r_feats["residue_index"], input_protein_l_feats["residue_index"]], dim=0)
+    residx_atom37_to_atom14 = torch.cat([input_protein_r_feats["residx_atom37_to_atom14"],
+                                         input_protein_l_feats["residx_atom37_to_atom14"]],
+                                        dim=0)
+    atom37_atom_exists = torch.cat([input_protein_r_feats["atom37_atom_exists"],
+                                    input_protein_l_feats["atom37_atom_exists"]], dim=0)
+
+    feats = {
+        "token_mask": token_mask,
+        "protein_r_mask": protein_r_mask,
+        "protein_l_mask": protein_l_mask,
+        "affinity_mask": affinity_mask,
+        "structural_mask": structural_mask,
+        "inter_pair_mask": inter_pair_mask,
+
+        "target_feat": target_feat,
+        "input_positions": input_positions,
+        "distogram_mask": distogram_mask,
+        "residue_index": _fit_to_crop(residue_index, crop_size, 0),
+        "aatype": _fit_to_crop(aatype, crop_size, 0),
+        "residx_atom37_to_atom14": _fit_to_crop(residx_atom37_to_atom14, crop_size, 0),
+        "atom37_atom_exists": _fit_to_crop(atom37_atom_exists, crop_size, 0),
+    }
+
+    if mode == "predict":
+        feats.update({
+            "in_chain_residue_index_r": input_protein_r_feats["in_chain_residue_index"],
+            "chain_index_r": input_protein_r_feats["chain_index"],
+            "in_chain_residue_index_l": input_protein_l_feats["in_chain_residue_index"],
+            "chain_index_l": input_protein_l_feats["chain_index"],
+        })
+
+    if mode == 'train' or mode == 'eval':
+        gt_protein_r_feats = data_pipeline.process_pdb(pdb_path=os.path.join(data_dir, input_data["gt_r_structure"]))
+        gt_protein_l_feats = data_pipeline.process_pdb(pdb_path=os.path.join(data_dir, input_data["gt_l_structure"]))
+
+        affinity_loss_factor = torch.tensor([1.0], dtype=torch.float32)
+        if input_data.get("affinity") is None:
+            eps = 1e-6
+            affinity_loss_factor = torch.tensor([eps], dtype=torch.float32)
+            affinity = torch.tensor([0.0], dtype=torch.float32)
+        else:
+            affinity = torch.tensor([input_data["affinity"]], dtype=torch.float32)
+
+        resolution = torch.tensor(input_data["resolution"], dtype=torch.float32)
+
+        # prepare inter_contacts
+        expanded_prot_r_pos = gt_protein_r_feats["pseudo_beta"].unsqueeze(1)  # Shape: (n_res_r, 1, 3)
+        expanded_prot_l_pos = gt_protein_l_feats["pseudo_beta"].unsqueeze(0)  # Shape: (1, n_res_l, 3)
+        distances = torch.sqrt(torch.sum((expanded_prot_r_pos - expanded_prot_l_pos) ** 2, dim=-1))
+        inter_contact = (distances < 8.0).float()
+        binding_site_mask_r = inter_contact.any(dim=1).float()
+        binding_site_mask_l = inter_contact.any(dim=0).float()
+        print("attaching binding masks", binding_site_mask_r.shape, binding_site_mask_l.shape)
+        binding_site_mask = torch.cat([binding_site_mask_r, binding_site_mask_l], dim=0)
+
+        inter_contact_reshaped_to_crop = torch.zeros((crop_size, crop_size), dtype=torch.float32)
+        inter_contact_reshaped_to_crop[:n_res_r, n_res_r:n_res_total] = inter_contact
+        inter_contact_reshaped_to_crop[n_res_r:n_res_total, :n_res_r] = inter_contact.T
+
+        # Use CA positions only
+        atom37_gt_positions = torch.cat([gt_protein_r_feats["all_atom_positions"],
+                                         gt_protein_l_feats["all_atom_positions"]], dim=0)
+        atom37_atom_exists_in_res = torch.cat([gt_protein_r_feats["atom37_atom_exists"],
+                                               gt_protein_l_feats["atom37_atom_exists"]], dim=0)
+        atom37_atom_exists_in_gt = torch.cat([gt_protein_r_feats["all_atom_mask"],
+                                              gt_protein_l_feats["all_atom_mask"]], dim=0)
+
+        atom14_gt_positions = torch.cat([gt_protein_r_feats["atom14_gt_positions"],
+                                         gt_protein_l_feats["atom14_gt_positions"]], dim=0)
+        atom14_atom_exists_in_res = torch.cat([gt_protein_r_feats["atom14_atom_exists"],
+                                               gt_protein_l_feats["atom14_atom_exists"]], dim=0)
+        atom14_atom_exists_in_gt = torch.cat([gt_protein_r_feats["atom14_gt_exists"],
+                                              gt_protein_l_feats["atom14_gt_exists"]], dim=0)
+
+        gt_pseudo_beta_joined = torch.cat([gt_protein_r_feats["pseudo_beta"], gt_protein_l_feats["pseudo_beta"]], dim=0)
+        gt_pseudo_beta_joined_mask = torch.cat([gt_protein_r_feats["pseudo_beta_mask"],
+                                         gt_protein_l_feats["pseudo_beta_mask"]], dim=0)
+
+        # IGNORES: residx_atom14_to_atom37, rigidgroups_group_exists,
+        # rigidgroups_group_is_ambiguous, pseudo_beta_mask, backbone_rigid_mask, protein_target_feat
+        gt_protein_feats = {
+            "atom37_gt_positions": atom37_gt_positions,  # torch.Size([n_struct, 37, 3])
+            "atom37_atom_exists_in_res": atom37_atom_exists_in_res,  # torch.Size([n_struct, 37])
+            "atom37_atom_exists_in_gt": atom37_atom_exists_in_gt,  # torch.Size([n_struct, 37])
+
+            "atom14_gt_positions": atom14_gt_positions,  # torch.Size([n_struct, 14, 3])
+            "atom14_atom_exists_in_res": atom14_atom_exists_in_res,  # torch.Size([n_struct, 14])
+            "atom14_atom_exists_in_gt": atom14_atom_exists_in_gt,  # torch.Size([n_struct, 14])
+
+            "gt_pseudo_beta_joined": gt_pseudo_beta_joined,  # torch.Size([n_struct, 3])
+            "gt_pseudo_beta_joined_mask": gt_pseudo_beta_joined_mask,  # torch.Size([n_struct])
+
+            # These we don't need to add the ligand to, because padding is sufficient (everything should be 0)
+            "atom14_alt_gt_positions": torch.cat([gt_protein_r_feats["atom14_alt_gt_positions"],
+                                                  gt_protein_l_feats["atom14_alt_gt_positions"]], dim=0),  # torch.Size([n_res, 14, 3])
+            "atom14_alt_gt_exists": torch.cat([gt_protein_r_feats["atom14_alt_gt_exists"],
+                                               gt_protein_l_feats["atom14_alt_gt_exists"]], dim=0),  # torch.Size([n_res, 14])
+            "atom14_atom_is_ambiguous": torch.cat([gt_protein_r_feats["atom14_atom_is_ambiguous"],
+                                                   gt_protein_l_feats["atom14_atom_is_ambiguous"]], dim=0),  # torch.Size([n_res, 14])
+            "rigidgroups_gt_frames": torch.cat([gt_protein_r_feats["rigidgroups_gt_frames"],
+                                                gt_protein_l_feats["rigidgroups_gt_frames"]], dim=0),  # torch.Size([n_res, 8, 4, 4])
+            "rigidgroups_gt_exists": torch.cat([gt_protein_r_feats["rigidgroups_gt_exists"],
+                                                gt_protein_l_feats["rigidgroups_gt_exists"]], dim=0),  # torch.Size([n_res, 8])
+            "rigidgroups_alt_gt_frames": torch.cat([gt_protein_r_feats["rigidgroups_alt_gt_frames"],
+                                                    gt_protein_l_feats["rigidgroups_alt_gt_frames"]], dim=0),  # torch.Size([n_res, 8, 4, 4])
+            "backbone_rigid_tensor": torch.cat([gt_protein_r_feats["backbone_rigid_tensor"],
+                                                gt_protein_l_feats["backbone_rigid_tensor"]], dim=0),  # torch.Size([n_res, 4, 4])
+            "backbone_rigid_mask": torch.cat([gt_protein_r_feats["backbone_rigid_mask"],
+                                              gt_protein_l_feats["backbone_rigid_mask"]], dim=0),  # torch.Size([n_res])
+            "chi_angles_sin_cos": torch.cat([gt_protein_r_feats["chi_angles_sin_cos"],
+                                             gt_protein_l_feats["chi_angles_sin_cos"]], dim=0),
+            "chi_mask": torch.cat([gt_protein_r_feats["chi_mask"], gt_protein_l_feats["chi_mask"]], dim=0),
+        }
+
+        for k, v in gt_protein_feats.items():
+            gt_protein_feats[k] = _fit_to_crop(v, crop_size, 0)
+
+        feats = {
+            **feats,
+            **gt_protein_feats,
+            "resolution": resolution,
+            "affinity": affinity,
+            "affinity_loss_factor": affinity_loss_factor,
+            "seq_length": torch.tensor(n_res_total),
+            "binding_site_mask": _fit_to_crop(binding_site_mask, crop_size, 0),
+            "gt_inter_contacts": inter_contact_reshaped_to_crop,
+        }
+
+    for k, v in feats.items():
+        # print(k, v.shape)
+        feats[k] = _prepare_recycles(v, num_recycles)
+
+    feats["batch_idx"] = torch.tensor(
+        [idx for _ in range(crop_size)], dtype=torch.int64, device=feats["aatype"].device
+    )
+
+    print("load time", round(time.time() - start_load_time, 4))
+
+    return feats

dockformerpp/data/data_transforms.py

@@ -0,0 +1,731 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import itertools
+from functools import reduce, wraps
+from operator import add
+
+import numpy as np
+import torch
+
+from dockformerpp.config import NUM_RES
+from dockformerpp.utils import residue_constants as rc
+from dockformerpp.utils.residue_constants import restypes
+from dockformerpp.utils.rigid_utils import Rotation, Rigid
+from dockformerpp.utils.geometry.rigid_matrix_vector import Rigid3Array
+from dockformerpp.utils.geometry.rotation_matrix import Rot3Array
+from dockformerpp.utils.geometry.vector import Vec3Array
+from dockformerpp.utils.tensor_utils import (
+    tree_map,
+    tensor_tree_map,
+    batched_gather,
+)
+
+
+def cast_to_64bit_ints(protein):
+    # We keep all ints as int64
+    for k, v in protein.items():
+        if v.dtype == torch.int32:
+            protein[k] = v.type(torch.int64)
+
+    return protein
+
+
+def make_one_hot(x, num_classes):
+    x_one_hot = torch.zeros(*x.shape, num_classes, device=x.device)
+    x_one_hot.scatter_(-1, x.unsqueeze(-1), 1)
+    return x_one_hot
+
+
+def curry1(f):
+    """Supply all arguments but the first."""
+    @wraps(f)
+    def fc(*args, **kwargs):
+        return lambda x: f(x, *args, **kwargs)
+
+    return fc
+
+
+def squeeze_features(protein):
+    """Remove singleton and repeated dimensions in protein features."""
+    protein["aatype"] = torch.argmax(protein["aatype"], dim=-1)
+    for k in [
+        "domain_name",
+        "seq_length",
+        "sequence",
+        "resolution",
+        "residue_index",
+    ]:
+        if k in protein:
+            final_dim = protein[k].shape[-1]
+            if isinstance(final_dim, int) and final_dim == 1:
+                if torch.is_tensor(protein[k]):
+                    protein[k] = torch.squeeze(protein[k], dim=-1)
+                else:
+                    protein[k] = np.squeeze(protein[k], axis=-1)
+
+    for k in ["seq_length"]:
+        if k in protein:
+            protein[k] = protein[k][0]
+
+    return protein
+
+
+def pseudo_beta_fn(aatype, all_atom_positions, all_atom_mask):
+    """Create pseudo beta features."""
+    is_gly = torch.eq(aatype, rc.restype_order["G"])
+    ca_idx = rc.atom_order["CA"]
+    cb_idx = rc.atom_order["CB"]
+    pseudo_beta = torch.where(
+        torch.tile(is_gly[..., None], [1] * len(is_gly.shape) + [3]),
+        all_atom_positions[..., ca_idx, :],
+        all_atom_positions[..., cb_idx, :],
+    )
+
+    if all_atom_mask is not None:
+        pseudo_beta_mask = torch.where(
+            is_gly, all_atom_mask[..., ca_idx], all_atom_mask[..., cb_idx]
+        )
+        return pseudo_beta, pseudo_beta_mask
+    else:
+        return pseudo_beta
+
+
+@curry1
+def make_pseudo_beta(protein):
+    """Create pseudo-beta (alpha for glycine) position and mask."""
+    (protein["pseudo_beta"], protein["pseudo_beta_mask"]) = pseudo_beta_fn(
+        protein["aatype"],
+        protein["all_atom_positions"],
+        protein["all_atom_mask"],
+    )
+    return protein
+
+
+@curry1
+def make_target_feat(protein):
+    """Create and concatenate protein features."""
+    # Whether there is a domain break. Always zero for chains, but keeping for
+    # compatibility with domain datasets.
+    aatype_1hot = make_one_hot(protein["aatype"], 20)
+
+    protein["protein_target_feat"] = aatype_1hot
+
+    return protein
+
+
+
+@curry1
+def select_feat(protein, feature_list):
+    return {k: v for k, v in protein.items() if k in feature_list}
+
+
+def get_restypes(device):
+    restype_atom14_to_atom37 = []
+    restype_atom37_to_atom14 = []
+    restype_atom14_mask = []
+
+    for rt in rc.restypes:
+        atom_names = rc.restype_name_to_atom14_names[rc.restype_1to3[rt]]
+        restype_atom14_to_atom37.append(
+            [(rc.atom_order[name] if name else 0) for name in atom_names]
+        )
+        atom_name_to_idx14 = {name: i for i, name in enumerate(atom_names)}
+        restype_atom37_to_atom14.append(
+            [
+                (atom_name_to_idx14[name] if name in atom_name_to_idx14 else 0)
+                for name in rc.atom_types
+            ]
+        )
+
+        restype_atom14_mask.append(
+            [(1.0 if name else 0.0) for name in atom_names]
+        )
+
+    # Add dummy mapping for restype 'UNK'
+    restype_atom14_to_atom37.append([0] * 14)
+    restype_atom37_to_atom14.append([0] * 37)
+    restype_atom14_mask.append([0.0] * 14)
+
+    restype_atom14_to_atom37 = torch.tensor(
+        restype_atom14_to_atom37,
+        dtype=torch.int32,
+        device=device,
+    )
+    restype_atom37_to_atom14 = torch.tensor(
+        restype_atom37_to_atom14,
+        dtype=torch.int32,
+        device=device,
+    )
+    restype_atom14_mask = torch.tensor(
+        restype_atom14_mask,
+        dtype=torch.float32,
+        device=device,
+    )
+
+    return restype_atom14_to_atom37, restype_atom37_to_atom14, restype_atom14_mask
+
+
+def get_restype_atom37_mask(device):
+    # create the corresponding mask
+    restype_atom37_mask = torch.zeros(
+        [len(restypes) + 1, 37], dtype=torch.float32, device=device
+    )
+    for restype, restype_letter in enumerate(rc.restypes):
+        restype_name = rc.restype_1to3[restype_letter]
+        atom_names = rc.residue_atoms[restype_name]
+        for atom_name in atom_names:
+            atom_type = rc.atom_order[atom_name]
+            restype_atom37_mask[restype, atom_type] = 1
+    return restype_atom37_mask
+
+
+def make_atom14_masks(protein):
+    """Construct denser atom positions (14 dimensions instead of 37)."""
+    restype_atom14_to_atom37, restype_atom37_to_atom14, restype_atom14_mask = get_restypes(protein["aatype"].device)
+
+    protein_aatype = protein['aatype'].to(torch.long)
+
+    # create the mapping for (residx, atom14) --> atom37, i.e. an array
+    # with shape (num_res, 14) containing the atom37 indices for this protein
+    residx_atom14_to_atom37 = restype_atom14_to_atom37[protein_aatype]
+    residx_atom14_mask = restype_atom14_mask[protein_aatype]
+
+    protein["atom14_atom_exists"] = residx_atom14_mask
+    protein["residx_atom14_to_atom37"] = residx_atom14_to_atom37.long()
+
+    # create the gather indices for mapping back
+    residx_atom37_to_atom14 = restype_atom37_to_atom14[protein_aatype]
+    protein["residx_atom37_to_atom14"] = residx_atom37_to_atom14.long()
+
+    restype_atom37_mask = get_restype_atom37_mask(protein["aatype"].device)
+
+    residx_atom37_mask = restype_atom37_mask[protein_aatype]
+    protein["atom37_atom_exists"] = residx_atom37_mask
+
+    return protein
+
+
+def make_atom14_positions(protein):
+    """Constructs denser atom positions (14 dimensions instead of 37)."""
+    residx_atom14_mask = protein["atom14_atom_exists"]
+    residx_atom14_to_atom37 = protein["residx_atom14_to_atom37"]
+
+    # Create a mask for known ground truth positions.
+    residx_atom14_gt_mask = residx_atom14_mask * batched_gather(
+        protein["all_atom_mask"],
+        residx_atom14_to_atom37,
+        dim=-1,
+        no_batch_dims=len(protein["all_atom_mask"].shape[:-1]),
+    )
+
+    # Gather the ground truth positions.
+    residx_atom14_gt_positions = residx_atom14_gt_mask[..., None] * (
+        batched_gather(
+            protein["all_atom_positions"],
+            residx_atom14_to_atom37,
+            dim=-2,
+            no_batch_dims=len(protein["all_atom_positions"].shape[:-2]),
+        )
+    )
+
+    protein["atom14_atom_exists"] = residx_atom14_mask
+    protein["atom14_gt_exists"] = residx_atom14_gt_mask
+    protein["atom14_gt_positions"] = residx_atom14_gt_positions
+
+    # As the atom naming is ambiguous for 7 of the 20 amino acids, provide
+    # alternative ground truth coordinates where the naming is swapped
+    restype_3 = [rc.restype_1to3[res] for res in rc.restypes]
+    restype_3 += ["UNK"]
+
+    # Matrices for renaming ambiguous atoms.
+    all_matrices = {
+        res: torch.eye(
+            14,
+            dtype=protein["all_atom_mask"].dtype,
+            device=protein["all_atom_mask"].device,
+        )
+        for res in restype_3
+    }
+    for resname, swap in rc.residue_atom_renaming_swaps.items():
+        correspondences = torch.arange(
+            14, device=protein["all_atom_mask"].device
+        )
+        for source_atom_swap, target_atom_swap in swap.items():
+            source_index = rc.restype_name_to_atom14_names[resname].index(
+                source_atom_swap
+            )
+            target_index = rc.restype_name_to_atom14_names[resname].index(
+                target_atom_swap
+            )
+            correspondences[source_index] = target_index
+            correspondences[target_index] = source_index
+            renaming_matrix = protein["all_atom_mask"].new_zeros((14, 14))
+            for index, correspondence in enumerate(correspondences):
+                renaming_matrix[index, correspondence] = 1.0
+        all_matrices[resname] = renaming_matrix
+
+    renaming_matrices = torch.stack(
+        [all_matrices[restype] for restype in restype_3]
+    )
+
+    # Pick the transformation matrices for the given residue sequence
+    # shape (num_res, 14, 14).
+    renaming_transform = renaming_matrices[protein["aatype"]]
+
+    # Apply it to the ground truth positions. shape (num_res, 14, 3).
+    alternative_gt_positions = torch.einsum(
+        "...rac,...rab->...rbc", residx_atom14_gt_positions, renaming_transform
+    )
+    protein["atom14_alt_gt_positions"] = alternative_gt_positions
+
+    # Create the mask for the alternative ground truth (differs from the
+    # ground truth mask, if only one of the atoms in an ambiguous pair has a
+    # ground truth position).
+    alternative_gt_mask = torch.einsum(
+        "...ra,...rab->...rb", residx_atom14_gt_mask, renaming_transform
+    )
+    protein["atom14_alt_gt_exists"] = alternative_gt_mask
+
+    # Create an ambiguous atoms mask.  shape: (21, 14).
+    restype_atom14_is_ambiguous = protein["all_atom_mask"].new_zeros((21, 14))
+    for resname, swap in rc.residue_atom_renaming_swaps.items():
+        for atom_name1, atom_name2 in swap.items():
+            restype = rc.restype_order[rc.restype_3to1[resname]]
+            atom_idx1 = rc.restype_name_to_atom14_names[resname].index(
+                atom_name1
+            )
+            atom_idx2 = rc.restype_name_to_atom14_names[resname].index(
+                atom_name2
+            )
+            restype_atom14_is_ambiguous[restype, atom_idx1] = 1
+            restype_atom14_is_ambiguous[restype, atom_idx2] = 1
+
+    # From this create an ambiguous_mask for the given sequence.
+    protein["atom14_atom_is_ambiguous"] = restype_atom14_is_ambiguous[
+        protein["aatype"]
+    ]
+
+    return protein
+
+
+def atom37_to_frames(protein, eps=1e-8):
+    aatype = protein["aatype"]
+    all_atom_positions = protein["all_atom_positions"]
+    all_atom_mask = protein["all_atom_mask"]
+
+    batch_dims = len(aatype.shape[:-1])
+
+    restype_rigidgroup_base_atom_names = np.full([21, 8, 3], "", dtype=object)
+    restype_rigidgroup_base_atom_names[:, 0, :] = ["C", "CA", "N"]
+    restype_rigidgroup_base_atom_names[:, 3, :] = ["CA", "C", "O"]
+
+    for restype, restype_letter in enumerate(rc.restypes):
+        resname = rc.restype_1to3[restype_letter]
+        for chi_idx in range(4):
+            if rc.chi_angles_mask[restype][chi_idx]:
+                names = rc.chi_angles_atoms[resname][chi_idx]
+                restype_rigidgroup_base_atom_names[
+                    restype, chi_idx + 4, :
+                ] = names[1:]
+
+    restype_rigidgroup_mask = all_atom_mask.new_zeros(
+        (*aatype.shape[:-1], 21, 8),
+    )
+    restype_rigidgroup_mask[..., 0] = 1
+    restype_rigidgroup_mask[..., 3] = 1
+    restype_rigidgroup_mask[..., :len(restypes), 4:] = all_atom_mask.new_tensor(
+        rc.chi_angles_mask
+    )
+
+    lookuptable = rc.atom_order.copy()
+    lookuptable[""] = 0
+    lookup = np.vectorize(lambda x: lookuptable[x])
+    restype_rigidgroup_base_atom37_idx = lookup(
+        restype_rigidgroup_base_atom_names,
+    )
+    restype_rigidgroup_base_atom37_idx = aatype.new_tensor(
+        restype_rigidgroup_base_atom37_idx,
+    )
+    restype_rigidgroup_base_atom37_idx = (
+        restype_rigidgroup_base_atom37_idx.view(
+            *((1,) * batch_dims), *restype_rigidgroup_base_atom37_idx.shape
+        )
+    )
+
+    residx_rigidgroup_base_atom37_idx = batched_gather(
+        restype_rigidgroup_base_atom37_idx,
+        aatype,
+        dim=-3,
+        no_batch_dims=batch_dims,
+    )
+
+    base_atom_pos = batched_gather(
+        all_atom_positions,
+        residx_rigidgroup_base_atom37_idx,
+        dim=-2,
+        no_batch_dims=len(all_atom_positions.shape[:-2]),
+    )
+
+    gt_frames = Rigid.from_3_points(
+        p_neg_x_axis=base_atom_pos[..., 0, :],
+        origin=base_atom_pos[..., 1, :],
+        p_xy_plane=base_atom_pos[..., 2, :],
+        eps=eps,
+    )
+
+    group_exists = batched_gather(
+        restype_rigidgroup_mask,
+        aatype,
+        dim=-2,
+        no_batch_dims=batch_dims,
+    )
+
+    gt_atoms_exist = batched_gather(
+        all_atom_mask,
+        residx_rigidgroup_base_atom37_idx,
+        dim=-1,
+        no_batch_dims=len(all_atom_mask.shape[:-1]),
+    )
+    gt_exists = torch.min(gt_atoms_exist, dim=-1)[0] * group_exists
+
+    rots = torch.eye(3, dtype=all_atom_mask.dtype, device=aatype.device)
+    rots = torch.tile(rots, (*((1,) * batch_dims), 8, 1, 1))
+    rots[..., 0, 0, 0] = -1
+    rots[..., 0, 2, 2] = -1
+
+    rots = Rotation(rot_mats=rots)
+    gt_frames = gt_frames.compose(Rigid(rots, None))
+
+    restype_rigidgroup_is_ambiguous = all_atom_mask.new_zeros(
+        *((1,) * batch_dims), 21, 8
+    )
+    restype_rigidgroup_rots = torch.eye(
+        3, dtype=all_atom_mask.dtype, device=aatype.device
+    )
+    restype_rigidgroup_rots = torch.tile(
+        restype_rigidgroup_rots,
+        (*((1,) * batch_dims), 21, 8, 1, 1),
+    )
+
+    for resname, _ in rc.residue_atom_renaming_swaps.items():
+        restype = rc.restype_order[rc.restype_3to1[resname]]
+        chi_idx = int(sum(rc.chi_angles_mask[restype]) - 1)
+        restype_rigidgroup_is_ambiguous[..., restype, chi_idx + 4] = 1
+        restype_rigidgroup_rots[..., restype, chi_idx + 4, 1, 1] = -1
+        restype_rigidgroup_rots[..., restype, chi_idx + 4, 2, 2] = -1
+
+    residx_rigidgroup_is_ambiguous = batched_gather(
+        restype_rigidgroup_is_ambiguous,
+        aatype,
+        dim=-2,
+        no_batch_dims=batch_dims,
+    )
+
+    residx_rigidgroup_ambiguity_rot = batched_gather(
+        restype_rigidgroup_rots,
+        aatype,
+        dim=-4,
+        no_batch_dims=batch_dims,
+    )
+
+    residx_rigidgroup_ambiguity_rot = Rotation(
+        rot_mats=residx_rigidgroup_ambiguity_rot
+    )
+    alt_gt_frames = gt_frames.compose(
+        Rigid(residx_rigidgroup_ambiguity_rot, None)
+    )
+
+    gt_frames_tensor = gt_frames.to_tensor_4x4()
+    alt_gt_frames_tensor = alt_gt_frames.to_tensor_4x4()
+
+    protein["rigidgroups_gt_frames"] = gt_frames_tensor
+    protein["rigidgroups_gt_exists"] = gt_exists
+    protein["rigidgroups_group_exists"] = group_exists
+    protein["rigidgroups_group_is_ambiguous"] = residx_rigidgroup_is_ambiguous
+    protein["rigidgroups_alt_gt_frames"] = alt_gt_frames_tensor
+
+    return protein
+
+
+def get_chi_atom_indices():
+    """Returns atom indices needed to compute chi angles for all residue types.
+
+    Returns:
+      A tensor of shape [residue_types=21, chis=4, atoms=4]. The residue types are
+      in the order specified in rc.restypes + unknown residue type
+      at the end. For chi angles which are not defined on the residue, the
+      positions indices are by default set to 0.
+    """
+    chi_atom_indices = []
+    for residue_name in rc.restypes:
+        residue_name = rc.restype_1to3[residue_name]
+        residue_chi_angles = rc.chi_angles_atoms[residue_name]
+        atom_indices = []
+        for chi_angle in residue_chi_angles:
+            atom_indices.append([rc.atom_order[atom] for atom in chi_angle])
+        for _ in range(4 - len(atom_indices)):
+            atom_indices.append(
+                [0, 0, 0, 0]
+            )  # For chi angles not defined on the AA.
+        chi_atom_indices.append(atom_indices)
+
+    chi_atom_indices.append([[0, 0, 0, 0]] * 4)  # For UNKNOWN residue.
+
+    return chi_atom_indices
+
+
+@curry1
+def atom37_to_torsion_angles(
+    protein,
+    prefix="",
+):
+    """
+    Convert coordinates to torsion angles.
+
+    This function is extremely sensitive to floating point imprecisions
+    and should be run with double precision whenever possible.
+
+    Args:
+        Dict containing:
+            * (prefix)aatype:
+                [*, N_res] residue indices
+            * (prefix)all_atom_positions:
+                [*, N_res, 37, 3] atom positions (in atom37
+                format)
+            * (prefix)all_atom_mask:
+                [*, N_res, 37] atom position mask
+    Returns:
+        The same dictionary updated with the following features:
+
+        "(prefix)torsion_angles_sin_cos" ([*, N_res, 7, 2])
+            Torsion angles
+        "(prefix)alt_torsion_angles_sin_cos" ([*, N_res, 7, 2])
+            Alternate torsion angles (accounting for 180-degree symmetry)
+        "(prefix)torsion_angles_mask" ([*, N_res, 7])
+            Torsion angles mask
+    """
+    aatype = protein[prefix + "aatype"]
+    all_atom_positions = protein[prefix + "all_atom_positions"]
+    all_atom_mask = protein[prefix + "all_atom_mask"]
+
+    aatype = torch.clamp(aatype, max=20)
+
+    pad = all_atom_positions.new_zeros(
+        [*all_atom_positions.shape[:-3], 1, 37, 3]
+    )
+    prev_all_atom_positions = torch.cat(
+        [pad, all_atom_positions[..., :-1, :, :]], dim=-3
+    )
+
+    pad = all_atom_mask.new_zeros([*all_atom_mask.shape[:-2], 1, 37])
+    prev_all_atom_mask = torch.cat([pad, all_atom_mask[..., :-1, :]], dim=-2)
+
+    pre_omega_atom_pos = torch.cat(
+        [prev_all_atom_positions[..., 1:3, :], all_atom_positions[..., :2, :]],
+        dim=-2,
+    )
+    phi_atom_pos = torch.cat(
+        [prev_all_atom_positions[..., 2:3, :], all_atom_positions[..., :3, :]],
+        dim=-2,
+    )
+    psi_atom_pos = torch.cat(
+        [all_atom_positions[..., :3, :], all_atom_positions[..., 4:5, :]],
+        dim=-2,
+    )
+
+    pre_omega_mask = torch.prod(
+        prev_all_atom_mask[..., 1:3], dim=-1
+    ) * torch.prod(all_atom_mask[..., :2], dim=-1)
+    phi_mask = prev_all_atom_mask[..., 2] * torch.prod(
+        all_atom_mask[..., :3], dim=-1, dtype=all_atom_mask.dtype
+    )
+    psi_mask = (
+        torch.prod(all_atom_mask[..., :3], dim=-1, dtype=all_atom_mask.dtype)
+        * all_atom_mask[..., 4]
+    )
+
+    chi_atom_indices = torch.as_tensor(
+        get_chi_atom_indices(), device=aatype.device
+    )
+
+    atom_indices = chi_atom_indices[..., aatype, :, :]
+    chis_atom_pos = batched_gather(
+        all_atom_positions, atom_indices, -2, len(atom_indices.shape[:-2])
+    )
+
+    chi_angles_mask = list(rc.chi_angles_mask)
+    chi_angles_mask.append([0.0, 0.0, 0.0, 0.0])
+    chi_angles_mask = all_atom_mask.new_tensor(chi_angles_mask)
+
+    chis_mask = chi_angles_mask[aatype, :]
+
+    chi_angle_atoms_mask = batched_gather(
+        all_atom_mask,
+        atom_indices,
+        dim=-1,
+        no_batch_dims=len(atom_indices.shape[:-2]),
+    )
+    chi_angle_atoms_mask = torch.prod(
+        chi_angle_atoms_mask, dim=-1, dtype=chi_angle_atoms_mask.dtype
+    )
+    chis_mask = chis_mask * chi_angle_atoms_mask
+
+    torsions_atom_pos = torch.cat(
+        [
+            pre_omega_atom_pos[..., None, :, :],
+            phi_atom_pos[..., None, :, :],
+            psi_atom_pos[..., None, :, :],
+            chis_atom_pos,
+        ],
+        dim=-3,
+    )
+
+    torsion_angles_mask = torch.cat(
+        [
+            pre_omega_mask[..., None],
+            phi_mask[..., None],
+            psi_mask[..., None],
+            chis_mask,
+        ],
+        dim=-1,
+    )
+
+    torsion_frames = Rigid.from_3_points(
+        torsions_atom_pos[..., 1, :],
+        torsions_atom_pos[..., 2, :],
+        torsions_atom_pos[..., 0, :],
+        eps=1e-8,
+    )
+
+    fourth_atom_rel_pos = torsion_frames.invert().apply(
+        torsions_atom_pos[..., 3, :]
+    )
+
+    torsion_angles_sin_cos = torch.stack(
+        [fourth_atom_rel_pos[..., 2], fourth_atom_rel_pos[..., 1]], dim=-1
+    )
+
+    denom = torch.sqrt(
+        torch.sum(
+            torch.square(torsion_angles_sin_cos),
+            dim=-1,
+            dtype=torsion_angles_sin_cos.dtype,
+            keepdims=True,
+        )
+        + 1e-8
+    )
+    torsion_angles_sin_cos = torsion_angles_sin_cos / denom
+
+    torsion_angles_sin_cos = torsion_angles_sin_cos * all_atom_mask.new_tensor(
+        [1.0, 1.0, -1.0, 1.0, 1.0, 1.0, 1.0],
+    )[((None,) * len(torsion_angles_sin_cos.shape[:-2])) + (slice(None), None)]
+
+    chi_is_ambiguous = torsion_angles_sin_cos.new_tensor(
+        rc.chi_pi_periodic,
+    )[aatype, ...]
+
+    mirror_torsion_angles = torch.cat(
+        [
+            all_atom_mask.new_ones(*aatype.shape, 3),
+            1.0 - 2.0 * chi_is_ambiguous,
+        ],
+        dim=-1,
+    )
+
+    alt_torsion_angles_sin_cos = (
+        torsion_angles_sin_cos * mirror_torsion_angles[..., None]
+    )
+
+    protein[prefix + "torsion_angles_sin_cos"] = torsion_angles_sin_cos
+    protein[prefix + "alt_torsion_angles_sin_cos"] = alt_torsion_angles_sin_cos
+    protein[prefix + "torsion_angles_mask"] = torsion_angles_mask
+
+    return protein
+
+
+def get_backbone_frames(protein):
+    # DISCREPANCY: AlphaFold uses tensor_7s here. I don't know why.
+    protein["backbone_rigid_tensor"] = protein["rigidgroups_gt_frames"][
+        ..., 0, :, :
+    ]
+    protein["backbone_rigid_mask"] = protein["rigidgroups_gt_exists"][..., 0]
+
+    return protein
+
+
+def get_chi_angles(protein):
+    dtype = protein["all_atom_mask"].dtype
+    protein["chi_angles_sin_cos"] = (
+        protein["torsion_angles_sin_cos"][..., 3:, :]
+    ).to(dtype)
+    protein["chi_mask"] = protein["torsion_angles_mask"][..., 3:].to(dtype)
+
+    return protein
+
+
+@curry1
+def random_crop_to_size(
+    protein,
+    crop_size,
+    shape_schema,
+    seed=None,
+):
+    """Crop randomly to `crop_size`, or keep as is if shorter than that."""
+    # We want each ensemble to be cropped the same way
+
+    g = None
+    if seed is not None:
+        g = torch.Generator(device=protein["seq_length"].device)
+        g.manual_seed(seed)
+
+    seq_length = protein["seq_length"]
+
+    num_res_crop_size = min(int(seq_length), crop_size)
+
+    def _randint(lower, upper):
+        return int(torch.randint(
+                lower,
+                upper + 1,
+                (1,),
+                device=protein["seq_length"].device,
+                generator=g,
+        )[0])
+
+    n = seq_length - num_res_crop_size
+    if "use_clamped_fape" in protein and protein["use_clamped_fape"] == 1.:
+        right_anchor = n
+    else:
+        x = _randint(0, n)
+        right_anchor = n - x
+
+    num_res_crop_start = _randint(0, right_anchor)
+
+    for k, v in protein.items():
+        if k not in shape_schema or (NUM_RES not in shape_schema[k]):
+            continue
+
+        slices = []
+        for i, (dim_size, dim) in enumerate(zip(shape_schema[k], v.shape)):
+            is_num_res = dim_size == NUM_RES
+            crop_start = num_res_crop_start if is_num_res else 0
+            crop_size = num_res_crop_size if is_num_res else dim
+            slices.append(slice(crop_start, crop_start + crop_size))
+        protein[k] = v[slices]
+
+    protein["seq_length"] = protein["seq_length"].new_tensor(num_res_crop_size)
+    
+    return protein
+

dockformerpp/data/errors.py

@@ -0,0 +1,22 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""General-purpose errors used throughout the data pipeline"""
+class Error(Exception):
+    """Base class for exceptions."""
+
+
+class MultipleChainsError(Error):
+    """An error indicating that multiple chains were found for a given ID."""

dockformerpp/data/parsers.py

@@ -0,0 +1,53 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Functions for parsing various file formats."""
+import collections
+import dataclasses
+import itertools
+import re
+import string
+from typing import Dict, Iterable, List, Optional, Sequence, Tuple, Set
+
+
+def parse_fasta(fasta_string: str) -> Tuple[Sequence[str], Sequence[str]]:
+    """Parses FASTA string and returns list of strings with amino-acid sequences.
+
+    Arguments:
+        fasta_string: The string contents of a FASTA file.
+
+    Returns:
+        A tuple of two lists:
+        * A list of sequences.
+        * A list of sequence descriptions taken from the comment lines. In the
+            same order as the sequences.
+    """
+    sequences = []
+    descriptions = []
+    index = -1
+    for line in fasta_string.splitlines():
+        line = line.strip()
+        if line.startswith(">"):
+            index += 1
+            descriptions.append(line[1:])  # Remove the '>' at the beginning.
+            sequences.append("")
+            continue
+        elif line.startswith("#"):
+            continue
+        elif not line:
+            continue  # Skip blank lines.
+        sequences[index] += line
+
+    return sequences, descriptions

dockformerpp/data/protein_features.py

@@ -0,0 +1,71 @@
+import numpy as np
+
+from dockformerpp.data.utils import FeatureDict
+from dockformerpp.utils import residue_constants, protein
+
+
+def _make_sequence_features(sequence: str, description: str, num_res: int) -> FeatureDict:
+    """Construct a feature dict of sequence features."""
+    features = {}
+    features["aatype"] = residue_constants.sequence_to_onehot(
+        sequence=sequence,
+        mapping=residue_constants.restype_order_with_x,
+        map_unknown_to_x=True,
+    )
+    features["domain_name"] = np.array(
+        [description.encode("utf-8")], dtype=object
+    )
+    # features["residue_index"] = np.array(range(num_res), dtype=np.int32)
+    features["seq_length"] = np.array([num_res] * num_res, dtype=np.int32)
+    features["sequence"] = np.array(
+        [sequence.encode("utf-8")], dtype=object
+    )
+    return features
+
+
+def _aatype_to_str_sequence(aatype):
+    return ''.join([
+        residue_constants.restypes_with_x[aatype[i]]
+        for i in range(len(aatype))
+    ])
+
+
+def _make_protein_structure_features(protein_object: protein.Protein) -> FeatureDict:
+    pdb_feats = {}
+
+    all_atom_positions = protein_object.atom_positions
+    all_atom_mask = protein_object.atom_mask
+
+    pdb_feats["all_atom_positions"] = all_atom_positions.astype(np.float32)
+    pdb_feats["all_atom_mask"] = all_atom_mask.astype(np.float32)
+    pdb_feats["in_chain_residue_index"] = protein_object.residue_index.astype(np.int32)
+
+    gapped_res_indexes = []
+    prev_chain_index = protein_object.chain_index[0]
+    chain_start_res_ind = 0
+    for relative_res_ind, chain_index in zip(protein_object.residue_index, protein_object.chain_index):
+        if chain_index != prev_chain_index:
+            chain_start_res_ind = gapped_res_indexes[-1] + 50
+            prev_chain_index = chain_index
+        gapped_res_indexes.append(relative_res_ind + chain_start_res_ind)
+
+    pdb_feats["residue_index"] = np.array(gapped_res_indexes).astype(np.int32)
+    pdb_feats["chain_index"] = np.array(protein_object.chain_index).astype(np.int32)
+    pdb_feats["resolution"] = np.array([0.]).astype(np.float32)
+
+    return pdb_feats
+
+
+def make_protein_features(protein_object: protein.Protein, description: str) -> FeatureDict:
+    feats = {}
+    aatype = protein_object.aatype
+    sequence = _aatype_to_str_sequence(aatype)
+    feats.update(
+        _make_sequence_features(sequence=sequence, description=description, num_res=len(protein_object.aatype))
+    )
+
+    feats.update(
+        _make_protein_structure_features(protein_object=protein_object)
+    )
+
+    return feats

dockformerpp/data/utils.py

@@ -0,0 +1,54 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Common utilities for data pipeline tools."""
+import contextlib
+import datetime
+import logging
+import shutil
+import tempfile
+import time
+from typing import Optional, Mapping, Dict
+
+import numpy as np
+import torch
+
+FeatureDict = Dict[str, np.ndarray]
+FeatureTensorDict = Dict[str, torch.Tensor]
+
+
+@contextlib.contextmanager
+def tmpdir_manager(base_dir: Optional[str] = None):
+    """Context manager that deletes a temporary directory on exit."""
+    tmpdir = tempfile.mkdtemp(dir=base_dir)
+    try:
+        yield tmpdir
+    finally:
+        shutil.rmtree(tmpdir, ignore_errors=True)
+
+
+@contextlib.contextmanager
+def timing(msg: str):
+    logging.info("Started %s", msg)
+    tic = time.perf_counter()
+    yield
+    toc = time.perf_counter()
+    logging.info("Finished %s in %.3f seconds", msg, toc - tic)
+
+
+def to_date(s: str):
+    return datetime.datetime(
+        year=int(s[:4]), month=int(s[5:7]), day=int(s[8:10])
+    )

dockformerpp/model/dropout.py

@@ -0,0 +1,69 @@
+# Copyright 2021 AlQuraishi Laboratory
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+from functools import partialmethod
+from typing import Union, List
+
+
+class Dropout(nn.Module):
+    """
+    Implementation of dropout with the ability to share the dropout mask
+    along a particular dimension.
+
+    If not in training mode, this module computes the identity function.
+    """
+
+    def __init__(self, r: float, batch_dim: Union[int, List[int]]):
+        """
+        Args:
+            r:
+                Dropout rate
+            batch_dim:
+                Dimension(s) along which the dropout mask is shared
+        """
+        super(Dropout, self).__init__()
+
+        self.r = r
+        if type(batch_dim) == int:
+            batch_dim = [batch_dim]
+        self.batch_dim = batch_dim
+        self.dropout = nn.Dropout(self.r)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x:
+                Tensor to which dropout is applied. Can have any shape
+                compatible with self.batch_dim
+        """
+        shape = list(x.shape)
+        if self.batch_dim is not None:
+            for bd in self.batch_dim:
+                shape[bd] = 1
+        mask = x.new_ones(shape)
+        mask = self.dropout(mask)
+        x *= mask
+        return x
+
+
+class DropoutRowwise(Dropout):
+    """
+    Convenience class for rowwise dropout as described in subsection
+    1.11.6.
+    """
+
+    __init__ = partialmethod(Dropout.__init__, batch_dim=-3)

dockformerpp/model/embedders.py

@@ -0,0 +1,320 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partial
+
+import torch
+import torch.nn as nn
+from typing import Tuple, Optional
+
+from dockformerpp.model.primitives import Linear, LayerNorm
+from dockformerpp.utils.tensor_utils import add
+
+
+class StructureInputEmbedder(nn.Module):
+    """
+    Embeds a subset of the input features.
+
+    Implements a merge of Algorithms 3 and Algorithm 32.
+    """
+
+    def __init__(
+        self,
+        protein_tf_dim: int,
+        additional_tf_dim: int,
+        c_z: int,
+        c_m: int,
+        relpos_k: int,
+        prot_min_bin: float,
+        prot_max_bin: float,
+        prot_no_bins: int,
+        inf: float = 1e8,
+        **kwargs,
+    ):
+        """
+        Args:
+            tf_dim:
+                Final dimension of the target features
+            c_z:
+                Pair embedding dimension
+            c_m:
+                Single embedding dimension
+            relpos_k:
+                Window size used in relative positional encoding
+        """
+        super(StructureInputEmbedder, self).__init__()
+
+        self.tf_dim = protein_tf_dim + additional_tf_dim
+
+        self.c_z = c_z
+        self.c_m = c_m
+
+        self.linear_tf_z_i = Linear(self.tf_dim, c_z)
+        self.linear_tf_z_j = Linear(self.tf_dim, c_z)
+        self.linear_tf_m = Linear(self.tf_dim, c_m)
+
+        # RPE stuff
+        self.relpos_k = relpos_k
+        self.no_bins = 2 * relpos_k + 1
+        self.linear_relpos = Linear(self.no_bins, c_z)
+
+        # Recycling stuff
+        self.prot_min_bin = prot_min_bin
+        self.prot_max_bin = prot_max_bin
+        self.prot_no_bins = prot_no_bins
+        self.inf = inf
+
+        self.prot_recycling_linear = Linear(self.prot_no_bins + 1, self.c_z)
+        self.layer_norm_m = LayerNorm(self.c_m)
+        self.layer_norm_z = LayerNorm(self.c_z)
+
+    def relpos(self, ri: torch.Tensor):
+        """
+        Computes relative positional encodings
+
+        Implements Algorithm 4.
+
+        Args:
+            ri:
+                "residue_index" features of shape [*, N]
+        """
+        d = ri[..., None] - ri[..., None, :]
+        boundaries = torch.arange(
+            start=-self.relpos_k, end=self.relpos_k + 1, device=d.device
+        )
+        reshaped_bins = boundaries.view(((1,) * len(d.shape)) + (len(boundaries),))
+        d = d[..., None] - reshaped_bins
+        d = torch.abs(d)
+        d = torch.argmin(d, dim=-1)
+        d = nn.functional.one_hot(d, num_classes=len(boundaries)).float()
+        d = d.to(ri.dtype)
+        return self.linear_relpos(d)
+
+    def _get_binned_distogram(self, x, min_bin, max_bin, no_bins, recycling_linear, prot_distogram_mask=None):
+        # This squared method might become problematic in FP16 mode.
+        bins = torch.linspace(
+            min_bin,
+            max_bin,
+            no_bins,
+            dtype=x.dtype,
+            device=x.device,
+            requires_grad=False,
+        )
+        squared_bins = bins ** 2
+        upper = torch.cat(
+            [squared_bins[1:], squared_bins.new_tensor([self.inf])], dim=-1
+        )
+        d = torch.sum((x[..., None, :] - x[..., None, :, :]) ** 2, dim=-1, keepdims=True)
+
+        # [*, N, N, no_bins]
+        d = ((d > squared_bins) * (d < upper)).type(x.dtype)
+        # print("d shape", d.shape, d[0][0][:10])
+
+        if prot_distogram_mask is not None:
+            expanded_d = torch.cat([d, torch.zeros(*d.shape[:-1], 1, device=d.device)], dim=-1)
+
+            # Step 2: Create a mask where `input_positions_masked` is 0
+            # Use broadcasting and tensor operations directly without additional variables
+            input_positions_mask = (prot_distogram_mask == 1).float()  # Shape [N, crop_size]
+            mask_i = input_positions_mask.unsqueeze(2)  # Shape [N, crop_size, 1]
+            mask_j = input_positions_mask.unsqueeze(1)  # Shape [N, 1, crop_size]
+
+            # Step 3: Combine masks for both [N, :, i, :] and [N, i, :, :]
+            combined_mask = mask_i + mask_j  # Shape [N, crop_size, crop_size]
+            combined_mask = combined_mask.clamp(max=1)  # Ensure binary mask
+
+            # Step 4: Apply the mask
+            # a. Set all but the last position in the `no_bins + 1` dimension to 0 where the mask is 1
+            expanded_d[..., :-1] *= (1 - combined_mask).unsqueeze(-1)  # Shape [N, crop_size, crop_size, no_bins]
+
+            # print("expanded_d shape1", expanded_d.shape, expanded_d[0][0][:10])
+
+            # b. Set the last position in the `no_bins + 1` dimension to 1 where the mask is 1
+            expanded_d[..., -1] += combined_mask  # Shape [N, crop_size, crop_size, 1]
+            d = expanded_d
+            # print("expanded_d shape2", d.shape, d[0][0][:10])
+
+        return recycling_linear(d)
+
+    def forward(
+        self,
+        token_mask: torch.Tensor,
+        protein_r_mask: torch.Tensor,
+        protein_l_mask: torch.Tensor,
+        target_feat: torch.Tensor,
+        input_positions: torch.Tensor,
+        residue_index: torch.Tensor,
+        distogram_mask: torch.Tensor,
+        inplace_safe: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            batch: Dict containing
+                "protein_target_feat":
+                    Features of shape [*, N_res + N_lig_atoms, tf_dim]
+                "residue_index":
+                    Features of shape [*, N_res]
+                input_protein_coords:
+                    [*, N_res, 3] AF predicted C_beta coordinates supplied as input
+                ligand_bonds_feat:
+                    [*, N_lig_atoms, N_lig_atoms, tf_dim] ligand bonds features
+        Returns:
+            single_emb:
+                [*, N_res + N_lig_atoms, C_m] single embedding
+            pair_emb:
+                [*, N_res + N_lig_atoms, N_res + N_lig_atoms, C_z] pair embedding
+
+        """
+        device = token_mask.device
+        pair_protein_r_mask = protein_r_mask[..., None] * protein_r_mask[..., None, :]
+        pair_protein_l_mask = protein_l_mask[..., None] * protein_l_mask[..., None, :]
+        intra_pair_protein_mask = pair_protein_r_mask + pair_protein_l_mask
+
+        # Single representation embedding - Algorithm 3
+        tf_m = self.linear_tf_m(target_feat)
+        tf_m = self.layer_norm_m(tf_m)  # previously this happened in the do_recycle function
+
+        # Pair representation
+        # protein pair embedding - Algorithm 3
+        # [*, N_res, c_z]
+        tf_emb_i = self.linear_tf_z_i(target_feat)
+        tf_emb_j = self.linear_tf_z_j(target_feat)
+
+        pair_emb = torch.zeros(*pair_protein_r_mask.shape, self.c_z, device=device)
+        pair_emb = add(pair_emb, tf_emb_i[..., None, :], inplace=inplace_safe)
+        pair_emb = add(pair_emb, tf_emb_j[..., None, :, :], inplace=inplace_safe)
+
+        # Apply relpos
+        relpos = self.relpos(residue_index.type(tf_emb_i.dtype))
+        pair_emb += relpos * intra_pair_protein_mask[..., None]
+
+        del relpos
+
+        # before recycles, do z_norm, this previously was a part of the recycles
+        pair_emb = self.layer_norm_z(pair_emb)
+
+        # apply protein recycle
+        prot_distogram_embed = self._get_binned_distogram(input_positions, self.prot_min_bin, self.prot_max_bin,
+                                                          self.prot_no_bins, self.prot_recycling_linear,
+                                                          distogram_mask)
+
+        pair_emb = add(pair_emb, prot_distogram_embed * intra_pair_protein_mask.unsqueeze(-1), inplace_safe)
+
+        del prot_distogram_embed
+
+        return tf_m, pair_emb
+
+
+class RecyclingEmbedder(nn.Module):
+    """
+    Embeds the output of an iteration of the model for recycling.
+
+    Implements Algorithm 32.
+    """
+    def __init__(
+        self,
+        c_m: int,
+        c_z: int,
+        min_bin: float,
+        max_bin: float,
+        no_bins: int,
+        inf: float = 1e8,
+        **kwargs,
+    ):
+        """
+        Args:
+            c_m:
+                Single channel dimension
+            c_z:
+                Pair embedding channel dimension
+            min_bin:
+                Smallest distogram bin (Angstroms)
+            max_bin:
+                Largest distogram bin (Angstroms)
+            no_bins:
+                Number of distogram bins
+        """
+        super(RecyclingEmbedder, self).__init__()
+
+        self.c_m = c_m
+        self.c_z = c_z
+        self.min_bin = min_bin
+        self.max_bin = max_bin
+        self.no_bins = no_bins
+        self.inf = inf
+
+        self.linear = Linear(self.no_bins, self.c_z)
+        self.layer_norm_m = LayerNorm(self.c_m)
+        self.layer_norm_z = LayerNorm(self.c_z)
+
+    def forward(
+        self,
+        m: torch.Tensor,
+        z: torch.Tensor,
+        x: torch.Tensor,
+        inplace_safe: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            m:
+                First row of the single embedding. [*, N_res, C_m]
+            z:
+                [*, N_res, N_res, C_z] pair embedding
+            x:
+                [*, N_res, 3] predicted C_beta coordinates
+        Returns:
+            m:
+                [*, N_res, C_m] single embedding update
+            z:
+                [*, N_res, N_res, C_z] pair embedding update
+        """
+        # [*, N, C_m]
+        m_update = self.layer_norm_m(m)
+        if(inplace_safe):
+            m.copy_(m_update)
+            m_update = m
+
+        # [*, N, N, C_z]
+        z_update = self.layer_norm_z(z)
+        if(inplace_safe):
+            z.copy_(z_update)
+            z_update = z
+
+        # This squared method might become problematic in FP16 mode.
+        bins = torch.linspace(
+            self.min_bin,
+            self.max_bin,
+            self.no_bins,
+            dtype=x.dtype,
+            device=x.device,
+            requires_grad=False,
+        )
+        squared_bins = bins ** 2
+        upper = torch.cat(
+            [squared_bins[1:], squared_bins.new_tensor([self.inf])], dim=-1
+        )
+        d = torch.sum(
+            (x[..., None, :] - x[..., None, :, :]) ** 2, dim=-1, keepdims=True
+        )
+
+        # [*, N, N, no_bins]
+        d = ((d > squared_bins) * (d < upper)).type(x.dtype)
+
+        # [*, N, N, C_z]
+        d = self.linear(d)
+        z_update = add(z_update, d, inplace_safe)
+
+        return m_update, z_update
+

dockformerpp/model/evoformer.py

@@ -0,0 +1,468 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+import sys
+import torch
+import torch.nn as nn
+from typing import Tuple, Sequence, Optional
+from functools import partial
+from abc import ABC, abstractmethod
+
+from dockformerpp.model.primitives import Linear, LayerNorm
+from dockformerpp.model.dropout import DropoutRowwise
+from dockformerpp.model.single_attention import SingleRowAttentionWithPairBias
+
+from dockformerpp.model.pair_transition import PairTransition
+from dockformerpp.model.triangular_attention import (
+    TriangleAttention,
+)
+from dockformerpp.model.triangular_multiplicative_update import (
+    TriangleMultiplicationOutgoing,
+    TriangleMultiplicationIncoming,
+)
+from dockformerpp.utils.checkpointing import checkpoint_blocks
+from dockformerpp.utils.tensor_utils import add
+
+
+class SingleRepTransition(nn.Module):
+    """
+    Feed-forward network applied to single representation activations after attention.
+
+    Implements Algorithm 9
+    """
+    def __init__(self, c_m, n):
+        """
+        Args:
+            c_m:
+                channel dimension
+            n:
+                Factor multiplied to c_m to obtain the hidden channel dimension
+        """
+        super(SingleRepTransition, self).__init__()
+
+        self.c_m = c_m
+        self.n = n
+
+        self.layer_norm = LayerNorm(self.c_m)
+        self.linear_1 = Linear(self.c_m, self.n * self.c_m, init="relu")
+        self.relu = nn.ReLU()
+        self.linear_2 = Linear(self.n * self.c_m, self.c_m, init="final")
+
+    def _transition(self, m, mask):
+        m = self.layer_norm(m)
+        m = self.linear_1(m)
+        m = self.relu(m)
+        m = self.linear_2(m) * mask
+        return m
+
+    def forward(
+        self,
+        m: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            m:
+                [*, N_res, C_m] activation after attention
+            mask:
+                [*, N_res, C_m] mask
+        Returns:
+            m:
+                [*, N_res, C_m] activation update
+        """
+        # DISCREPANCY: DeepMind forgets to apply the mask here.
+        if mask is None:
+            mask = m.new_ones(m.shape[:-1])
+
+        mask = mask.unsqueeze(-1)
+
+        m = self._transition(m, mask)
+
+        return m
+
+
+class PairStack(nn.Module):
+    def __init__(
+        self,
+        c_z: int,
+        c_hidden_mul: int,
+        c_hidden_pair_att: int,
+        no_heads_pair: int,
+        transition_n: int,
+        pair_dropout: float,
+        inf: float,
+        eps: float
+    ):
+        super(PairStack, self).__init__()
+
+        self.tri_mul_out = TriangleMultiplicationOutgoing(
+            c_z,
+            c_hidden_mul,
+        )
+        self.tri_mul_in = TriangleMultiplicationIncoming(
+            c_z,
+            c_hidden_mul,
+        )
+
+        self.tri_att_start = TriangleAttention(
+            c_z,
+            c_hidden_pair_att,
+            no_heads_pair,
+            inf=inf,
+        )
+        self.tri_att_end = TriangleAttention(
+            c_z,
+            c_hidden_pair_att,
+            no_heads_pair,
+            inf=inf,
+        )
+
+        self.pair_transition = PairTransition(
+            c_z,
+            transition_n,
+        )
+
+        self.ps_dropout_row_layer = DropoutRowwise(pair_dropout)
+
+    def forward(self,
+        z: torch.Tensor,
+        pair_mask: torch.Tensor,
+        use_lma: bool = False,
+        inplace_safe: bool = False,
+        _mask_trans: bool = True,
+    ) -> torch.Tensor:
+        # DeepMind doesn't mask these transitions in the source, so _mask_trans
+        # should be disabled to better approximate the exact activations of
+        # the original.
+        pair_trans_mask = pair_mask if _mask_trans else None
+
+        tmu_update = self.tri_mul_out(
+            z,
+            mask=pair_mask,
+            inplace_safe=inplace_safe,
+            _add_with_inplace=True,
+        )
+        if (not inplace_safe):
+            z = z + self.ps_dropout_row_layer(tmu_update)
+        else:
+            z = tmu_update
+
+        del tmu_update
+
+        tmu_update = self.tri_mul_in(
+            z,
+            mask=pair_mask,
+            inplace_safe=inplace_safe,
+            _add_with_inplace=True,
+        )
+        if (not inplace_safe):
+            z = z + self.ps_dropout_row_layer(tmu_update)
+        else:
+            z = tmu_update
+
+        del tmu_update
+
+        z = add(z,
+                self.ps_dropout_row_layer(
+                    self.tri_att_start(
+                        z,
+                        mask=pair_mask,
+                        use_memory_efficient_kernel=False,
+                        use_lma=use_lma,
+                    )
+                ),
+                inplace=inplace_safe,
+                )
+
+        z = z.transpose(-2, -3)
+        if (inplace_safe):
+            z = z.contiguous()
+
+        z = add(z,
+                self.ps_dropout_row_layer(
+                    self.tri_att_end(
+                        z,
+                        mask=pair_mask.transpose(-1, -2),
+                        use_memory_efficient_kernel=False,
+                        use_lma=use_lma,
+                    )
+                ),
+                inplace=inplace_safe,
+                )
+
+        z = z.transpose(-2, -3)
+        if (inplace_safe):
+            z = z.contiguous()
+
+        z = add(z,
+                self.pair_transition(
+                    z, mask=pair_trans_mask,
+                ),
+                inplace=inplace_safe,
+        )
+
+        return z
+
+
+class EvoformerBlock(nn.Module, ABC):
+    def __init__(self,
+        c_m: int,
+        c_z: int,
+        c_hidden_single_att: int,
+        c_hidden_mul: int,
+        c_hidden_pair_att: int,
+        no_heads_single: int,
+        no_heads_pair: int,
+        transition_n: int,
+        single_dropout: float,
+        pair_dropout: float,
+        inf: float,
+        eps: float,
+    ):
+        super(EvoformerBlock, self).__init__()
+
+        self.single_att_row = SingleRowAttentionWithPairBias(
+            c_m=c_m,
+            c_z=c_z,
+            c_hidden=c_hidden_single_att,
+            no_heads=no_heads_single,
+            inf=inf,
+        )
+
+        self.single_dropout_layer = DropoutRowwise(single_dropout)
+
+        self.single_transition = SingleRepTransition(
+            c_m=c_m,
+            n=transition_n,
+        )
+
+        self.pair_stack = PairStack(
+            c_z=c_z,
+            c_hidden_mul=c_hidden_mul,
+            c_hidden_pair_att=c_hidden_pair_att,
+            no_heads_pair=no_heads_pair,
+            transition_n=transition_n,
+            pair_dropout=pair_dropout,
+            inf=inf,
+            eps=eps
+        )
+
+    def forward(self,
+        m: Optional[torch.Tensor],
+        z: Optional[torch.Tensor],
+        single_mask: torch.Tensor,
+        pair_mask: torch.Tensor,
+        use_lma: bool = False,
+        inplace_safe: bool = False,
+        _mask_trans: bool = True,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+
+        single_trans_mask = single_mask if _mask_trans else None
+
+        input_tensors = [m, z]
+
+        m, z = input_tensors
+
+        z = self.pair_stack(
+            z=z,
+            pair_mask=pair_mask,
+            use_lma=use_lma,
+            inplace_safe=inplace_safe,
+            _mask_trans=_mask_trans,
+        )
+
+        m = add(m,
+                self.single_dropout_layer(
+                    self.single_att_row(
+                        m,
+                        z=z,
+                        mask=single_mask,
+                        use_memory_efficient_kernel=False,
+                        use_lma=use_lma,
+                    )
+                ),
+                inplace=inplace_safe,
+                )
+
+        m = add(m, self.single_transition(m, mask=single_mask), inplace=inplace_safe)
+
+        return m, z
+
+
+class EvoformerStack(nn.Module):
+    """
+    Main Evoformer trunk.
+
+    Implements Algorithm 6.
+    """
+
+    def __init__(
+        self,
+        c_m: int,
+        c_z: int,
+        c_hidden_single_att: int,
+        c_hidden_mul: int,
+        c_hidden_pair_att: int,
+        c_s: int,
+        no_heads_single: int,
+        no_heads_pair: int,
+        no_blocks: int,
+        transition_n: int,
+        single_dropout: float,
+        pair_dropout: float,
+        blocks_per_ckpt: int,
+        inf: float,
+        eps: float,
+        clear_cache_between_blocks: bool = False, 
+        **kwargs,
+    ):
+        """
+        Args:
+            c_m:
+                single channel dimension
+            c_z:
+                Pair channel dimension
+            c_hidden_single_att:
+                Hidden dimension in single representation attention
+            c_hidden_mul:
+                Hidden dimension in multiplicative updates
+            c_hidden_pair_att:
+                Hidden dimension in triangular attention
+            c_s:
+                Channel dimension of the output "single" embedding
+            no_heads_single:
+                Number of heads used for single attention
+            no_heads_pair:
+                Number of heads used for pair attention
+            no_blocks:
+                Number of Evoformer blocks in the stack
+            transition_n:
+                Factor by which to multiply c_m to obtain the SingleTransition
+                hidden dimension
+            single_dropout:
+                Dropout rate for single activations
+            pair_dropout:
+                Dropout used for pair activations
+            blocks_per_ckpt:
+                Number of Evoformer blocks in each activation checkpoint
+            clear_cache_between_blocks:
+                Whether to clear CUDA's GPU memory cache between blocks of the
+                stack. Slows down each block but can reduce fragmentation
+        """
+        super(EvoformerStack, self).__init__()
+
+        self.blocks_per_ckpt = blocks_per_ckpt
+        self.clear_cache_between_blocks = clear_cache_between_blocks
+
+        self.blocks = nn.ModuleList()
+
+        for _ in range(no_blocks):
+            block = EvoformerBlock(
+                c_m=c_m,
+                c_z=c_z,
+                c_hidden_single_att=c_hidden_single_att,
+                c_hidden_mul=c_hidden_mul,
+                c_hidden_pair_att=c_hidden_pair_att,
+                no_heads_single=no_heads_single,
+                no_heads_pair=no_heads_pair,
+                transition_n=transition_n,
+                single_dropout=single_dropout,
+                pair_dropout=pair_dropout,
+                inf=inf,
+                eps=eps,
+            )
+            self.blocks.append(block)
+
+        self.linear = Linear(c_m, c_s)
+
+    def _prep_blocks(self,
+        use_lma: bool,
+        single_mask: Optional[torch.Tensor],
+        pair_mask: Optional[torch.Tensor],
+        inplace_safe: bool,
+        _mask_trans: bool,
+    ):
+        blocks = [
+            partial(
+                b,
+                single_mask=single_mask,
+                pair_mask=pair_mask,
+                use_lma=use_lma,
+                inplace_safe=inplace_safe,
+                _mask_trans=_mask_trans,
+            )
+            for b in self.blocks
+        ]
+
+        if self.clear_cache_between_blocks:
+            def block_with_cache_clear(block, *args, **kwargs):
+                torch.cuda.empty_cache()
+                return block(*args, **kwargs)
+
+            blocks = [partial(block_with_cache_clear, b) for b in blocks]
+
+        return blocks
+
+    def forward(self,
+        m: torch.Tensor,
+        z: torch.Tensor,
+        single_mask: torch.Tensor,
+        pair_mask: torch.Tensor,
+        use_lma: bool = False,
+        inplace_safe: bool = False,
+        _mask_trans: bool = True,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            m:
+                [*, N_res, C_m] single embedding
+            z:
+                [*, N_res, N_res, C_z] pair embedding
+            single_mask:
+                [*, N_res] single mask
+            pair_mask:
+                [*, N_res, N_res] pair mask
+            use_lma:
+                Whether to use low-memory attention during inference.
+
+        Returns:
+            m:
+                [*, N_res, C_m] single embedding
+            z:
+                [*, N_res, N_res, C_z] pair embedding
+            s:
+                [*, N_res, C_s] single embedding after linear layer
+        """ 
+        blocks = self._prep_blocks(
+            use_lma=use_lma,
+            single_mask=single_mask,
+            pair_mask=pair_mask,
+            inplace_safe=inplace_safe,
+            _mask_trans=_mask_trans,
+        )
+
+        blocks_per_ckpt = self.blocks_per_ckpt
+        if(not torch.is_grad_enabled()):
+            blocks_per_ckpt = None
+        
+        m, z = checkpoint_blocks(
+            blocks,
+            args=(m, z),
+            blocks_per_ckpt=blocks_per_ckpt,
+        )
+
+        s = self.linear(m)
+
+        return m, z, s

dockformerpp/model/heads.py

@@ -0,0 +1,233 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import torch
+import torch.nn as nn
+from torch.nn import Parameter
+
+from dockformerpp.model.primitives import Linear, LayerNorm
+from dockformerpp.utils.loss import (
+    compute_plddt,
+    compute_tm,
+    compute_predicted_aligned_error,
+)
+from dockformerpp.utils.precision_utils import is_fp16_enabled
+
+
+class AuxiliaryHeads(nn.Module):
+    def __init__(self, config):
+        super(AuxiliaryHeads, self).__init__()
+
+        self.plddt = PerResidueLDDTCaPredictor(
+            **config["lddt"],
+        )
+
+        self.distogram = DistogramHead(
+            **config["distogram"],
+        )
+
+        self.affinity_2d = Affinity2DPredictor(
+            **config["affinity_2d"],
+        )
+
+        self.affinity_cls = AffinityClsTokenPredictor(
+            **config["affinity_cls"],
+        )
+
+        self.binding_site = BindingSitePredictor(
+            **config["binding_site"],
+        )
+
+        self.inter_contact = InterContactHead(
+            **config["inter_contact"],
+        )
+
+        self.config = config
+
+    def forward(self, outputs, inter_mask, affinity_mask):
+        aux_out = {}
+        lddt_logits = self.plddt(outputs["sm"]["single"])
+        aux_out["lddt_logits"] = lddt_logits
+
+        # Required for relaxation later on
+        aux_out["plddt"] = compute_plddt(lddt_logits)
+
+        distogram_logits = self.distogram(outputs["pair"])
+        aux_out["distogram_logits"] = distogram_logits
+
+        aux_out["inter_contact_logits"] = self.inter_contact(outputs["single"], outputs["pair"])
+
+        aux_out["affinity_2d_logits"] = self.affinity_2d(outputs["pair"], aux_out["inter_contact_logits"], inter_mask)
+
+        aux_out["affinity_cls_logits"] = self.affinity_cls(outputs["single"], affinity_mask)
+
+        aux_out["binding_site_logits"] = self.binding_site(outputs["single"])
+
+        return aux_out
+
+
+class Affinity2DPredictor(nn.Module):
+    def __init__(self, c_z, num_bins):
+        super(Affinity2DPredictor, self).__init__()
+
+        self.c_z = c_z
+
+        self.weight_linear = Linear(self.c_z + 1, 1)
+        self.embed_linear = Linear(self.c_z, self.c_z)
+        self.bins_linear = Linear(self.c_z, num_bins)
+
+    def forward(self, z, inter_contacts_logits, inter_pair_mask):
+        z_with_inter_contacts = torch.cat((z, inter_contacts_logits), dim=-1)  # [*, N, N, c_z + 1]
+        weights = self.weight_linear(z_with_inter_contacts)  # [*, N, N, 1]
+
+        x = self.embed_linear(z)  # [*, N, N, c_z]
+        batch_size, N, M, _ = x.shape
+
+        flat_weights = weights.reshape(batch_size, N*M, -1)  # [*, N*M, 1]
+        flat_x = x.reshape(batch_size, N*M, -1)  # [*, N*M, c_z]
+        flat_inter_pair_mask = inter_pair_mask.reshape(batch_size, N*M, 1)
+
+        flat_weights = flat_weights.masked_fill(~(flat_inter_pair_mask.bool()), float('-inf'))  # [*, N*N, 1]
+        flat_weights = torch.nn.functional.softmax(flat_weights, dim=1)  # [*, N*N, 1]
+        flat_weights = torch.nan_to_num(flat_weights, nan=0.0)  # [*, N*N, 1]
+        weighted_sum = torch.sum((flat_weights * flat_x).reshape(batch_size, N*M, -1), dim=1)  # [*, c_z]
+
+        return self.bins_linear(weighted_sum)
+
+
+class AffinityClsTokenPredictor(nn.Module):
+    def __init__(self, c_s, num_bins, **kwargs):
+        super(AffinityClsTokenPredictor, self).__init__()
+
+        self.c_s = c_s
+        self.linear = Linear(self.c_s, num_bins, init="final")
+
+    def forward(self, s, affinity_mask):
+        affinity_tokens = (s * affinity_mask.unsqueeze(-1)).sum(dim=1)
+        return self.linear(affinity_tokens)
+
+
+class BindingSitePredictor(nn.Module):
+    def __init__(self, c_s, c_out, **kwargs):
+        super(BindingSitePredictor, self).__init__()
+
+        self.c_s = c_s
+        self.c_out = c_out
+
+        self.linear = Linear(self.c_s, self.c_out, init="final")
+
+    def forward(self, s):
+        # [*, N, C_out]
+        return self.linear(s)
+
+
+class InterContactHead(nn.Module):
+    def __init__(self, c_s, c_z, c_out, **kwargs):
+        """
+        Args:
+            c_z:
+                Input channel dimension
+            c_out:
+                Number of bins, but since boolean should be 1
+        """
+        super(InterContactHead, self).__init__()
+
+        self.c_s = c_s
+        self.c_z = c_z
+        self.c_out = c_out
+
+        self.linear = Linear(2 * self.c_s + self.c_z, self.c_out, init="final")
+
+    def forward(self, s, z):  # [*, N, N, C_z]
+        # [*, N, N, no_bins]
+        batch_size, n, s_dim = s.shape
+
+        s_i = s.unsqueeze(2).expand(batch_size, n, n, s_dim)
+        s_j = s.unsqueeze(1).expand(batch_size, n, n, s_dim)
+        joined = torch.cat((s_i, s_j, z), dim=-1)
+
+        logits = self.linear(joined)
+
+        return logits
+
+
+class PerResidueLDDTCaPredictor(nn.Module):
+    def __init__(self, no_bins, c_in, c_hidden):
+        super(PerResidueLDDTCaPredictor, self).__init__()
+
+        self.no_bins = no_bins
+        self.c_in = c_in
+        self.c_hidden = c_hidden
+
+        self.layer_norm = LayerNorm(self.c_in)
+
+        self.linear_1 = Linear(self.c_in, self.c_hidden, init="relu")
+        self.linear_2 = Linear(self.c_hidden, self.c_hidden, init="relu")
+        self.linear_3 = Linear(self.c_hidden, self.no_bins, init="final")
+
+        self.relu = nn.ReLU()
+
+    def forward(self, s):
+        s = self.layer_norm(s)
+        s = self.linear_1(s)
+        s = self.relu(s)
+        s = self.linear_2(s)
+        s = self.relu(s)
+        s = self.linear_3(s)
+
+        return s
+
+
+class DistogramHead(nn.Module):
+    """
+    Computes a distogram probability distribution.
+
+    For use in computation of distogram loss, subsection 1.9.8
+    """
+
+    def __init__(self, c_z, no_bins, **kwargs):
+        """
+        Args:
+            c_z:
+                Input channel dimension
+            no_bins:
+                Number of distogram bins
+        """
+        super(DistogramHead, self).__init__()
+
+        self.c_z = c_z
+        self.no_bins = no_bins
+
+        self.linear = Linear(self.c_z, self.no_bins, init="final")
+
+    def _forward(self, z):  # [*, N, N, C_z]
+        """
+        Args:
+            z:
+                [*, N_res, N_res, C_z] pair embedding
+        Returns:
+            [*, N, N, no_bins] distogram probability distribution
+        """
+        # [*, N, N, no_bins]
+        logits = self.linear(z)
+        logits = logits + logits.transpose(-2, -3)
+        return logits
+    
+    def forward(self, z): 
+        if(is_fp16_enabled()):
+            with torch.cuda.amp.autocast(enabled=False):
+                return self._forward(z.float())
+        else:
+            return self._forward(z)

dockformerpp/model/model.py

@@ -0,0 +1,317 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from functools import partial
+import weakref
+
+import torch
+import torch.nn as nn
+
+from dockformerpp.utils.tensor_utils import masked_mean
+from dockformerpp.model.embedders import (
+    StructureInputEmbedder,
+    RecyclingEmbedder,
+)
+from dockformerpp.model.evoformer import EvoformerStack
+from dockformerpp.model.heads import AuxiliaryHeads
+from dockformerpp.model.structure_module import StructureModule
+import dockformerpp.utils.residue_constants as residue_constants
+from dockformerpp.utils.feats import (
+    pseudo_beta_fn,
+    atom14_to_atom37,
+)
+from dockformerpp.utils.tensor_utils import (
+    add,
+    tensor_tree_map,
+)
+
+
+class AlphaFold(nn.Module):
+    """
+    Alphafold 2.
+
+    Implements Algorithm 2 (but with training).
+    """
+
+    def __init__(self, config):
+        """
+        Args:
+            config:
+                A dict-like config object (like the one in config.py)
+        """
+        super(AlphaFold, self).__init__()
+
+        self.globals = config.globals
+        self.config = config.model
+
+        # Main trunk + structure module
+        self.input_embedder = StructureInputEmbedder(
+            **self.config["structure_input_embedder"],
+        )
+
+        self.recycling_embedder = RecyclingEmbedder(
+            **self.config["recycling_embedder"],
+        )
+
+        self.evoformer = EvoformerStack(
+            **self.config["evoformer_stack"],
+        )
+
+        self.structure_module = StructureModule(
+            **self.config["structure_module"],
+        )
+        self.aux_heads = AuxiliaryHeads(
+            self.config["heads"],
+        )
+
+    def tolerance_reached(self, prev_pos, next_pos, mask, eps=1e-8) -> bool:
+        """
+        Early stopping criteria based on criteria used in
+        AF2Complex: https://www.nature.com/articles/s41467-022-29394-2
+        Args:
+          prev_pos: Previous atom positions in atom37/14 representation
+          next_pos: Current atom positions in atom37/14 representation
+          mask: 1-D sequence mask
+          eps: Epsilon used in square root calculation
+        Returns:
+          Whether to stop recycling early based on the desired tolerance.
+        """
+
+        def distances(points):
+            """Compute all pairwise distances for a set of points."""
+            d = points[..., None, :] - points[..., None, :, :]
+            return torch.sqrt(torch.sum(d ** 2, dim=-1))
+
+        if self.config.recycle_early_stop_tolerance < 0:
+            return False
+
+        ca_idx = residue_constants.atom_order['CA']
+        sq_diff = (distances(prev_pos[..., ca_idx, :]) - distances(next_pos[..., ca_idx, :])) ** 2
+        mask = mask[..., None] * mask[..., None, :]
+        sq_diff = masked_mean(mask=mask, value=sq_diff, dim=list(range(len(mask.shape))))
+        diff = torch.sqrt(sq_diff + eps).item()
+        return diff <= self.config.recycle_early_stop_tolerance
+
+    def iteration(self, feats, prevs, _recycle=True):
+        # Primary output dictionary
+        outputs = {}
+
+        # This needs to be done manually for DeepSpeed's sake
+        dtype = next(self.parameters()).dtype
+        for k in feats:
+            if feats[k].dtype == torch.float32:
+                feats[k] = feats[k].to(dtype=dtype)
+
+        # Grab some data about the input
+        batch_dims, n_total = feats["token_mask"].shape
+        device = feats["token_mask"].device
+
+        print("doing sample of size", feats["token_mask"].shape,
+              feats["protein_r_mask"].sum(dim=1), feats["protein_l_mask"].sum(dim=1))
+
+        # Controls whether the model uses in-place operations throughout
+        # The dual condition accounts for activation checkpoints
+        # inplace_safe = not (self.training or torch.is_grad_enabled())
+        inplace_safe = False  # so we don't need attn_core_inplace_cuda
+
+        # Prep some features
+        token_mask = feats["token_mask"]
+        pair_mask = token_mask[..., None] * token_mask[..., None, :]
+
+        # Initialize the single and pair representations
+        # m: [*, 1, n_total, C_m]
+        # z: [*, n_total, n_total, C_z]
+        m, z = self.input_embedder(
+            feats["token_mask"],
+            feats["protein_r_mask"],
+            feats["protein_l_mask"],
+            feats["target_feat"],
+            feats["input_positions"],
+            feats["residue_index"],
+            feats["distogram_mask"],
+            inplace_safe=inplace_safe,
+        )
+
+        # Unpack the recycling embeddings. Removing them from the list allows 
+        # them to be freed further down in this function, saving memory
+        m_1_prev, z_prev, x_prev = reversed([prevs.pop() for _ in range(3)])
+
+        # Initialize the recycling embeddings, if needs be 
+        if None in [m_1_prev, z_prev, x_prev]:
+            # [*, N, C_m]
+            m_1_prev = m.new_zeros(
+                (batch_dims, n_total, self.config.structure_input_embedder.c_m),
+                requires_grad=False,
+            )
+
+            # [*, N, N, C_z]
+            z_prev = z.new_zeros(
+                (batch_dims, n_total, n_total, self.config.structure_input_embedder.c_z),
+                requires_grad=False,
+            )
+
+            # [*, N, 3]
+            x_prev = z.new_zeros(
+                (batch_dims, n_total, residue_constants.atom_type_num, 3),
+                requires_grad=False,
+            )
+
+        # shape == [1, n_total, 37, 3]
+        pseudo_beta_or_lig_x_prev = pseudo_beta_fn(feats["aatype"], x_prev, None).to(dtype=z.dtype)
+
+        # m_1_prev_emb: [*, N, C_m]
+        # z_prev_emb: [*, N, N, C_z]
+        m_1_prev_emb, z_prev_emb = self.recycling_embedder(
+            m_1_prev,
+            z_prev,
+            pseudo_beta_or_lig_x_prev,
+            inplace_safe=inplace_safe,
+        )
+
+        del pseudo_beta_or_lig_x_prev
+
+        # [*, S_c, N, C_m]
+        m += m_1_prev_emb
+
+        # [*, N, N, C_z]
+        z = add(z, z_prev_emb, inplace=inplace_safe)
+
+        # Deletions like these become significant for inference with large N,
+        # where they free unused tensors and remove references to others such
+        # that they can be offloaded later
+        del m_1_prev, z_prev, m_1_prev_emb, z_prev_emb
+
+        # Run single + pair embeddings through the trunk of the network
+        # m: [*, N, C_m]
+        # z: [*, N, N, C_z]
+        # s: [*, N, C_s]
+        m, z, s = self.evoformer(
+            m,
+            z,
+            single_mask=token_mask.to(dtype=m.dtype),
+            pair_mask=pair_mask.to(dtype=z.dtype),
+            use_lma=self.globals.use_lma,
+            inplace_safe=inplace_safe,
+            _mask_trans=self.config._mask_trans,
+        )
+
+        outputs["pair"] = z
+        outputs["single"] = s
+
+        del z
+
+        # Predict 3D structure
+        outputs["sm"] = self.structure_module(
+            outputs,
+            feats["aatype"],
+            mask=token_mask.to(dtype=s.dtype),
+            inplace_safe=inplace_safe,
+        )
+        outputs["final_atom_positions"] = atom14_to_atom37(
+            outputs["sm"]["positions"][-1], feats
+        )
+        outputs["final_atom_mask"] = feats["atom37_atom_exists"]
+
+        # Save embeddings for use during the next recycling iteration
+
+        # [*, N, C_m]
+        m_1_prev = m[..., 0, :, :]
+
+        # [*, N, N, C_z]
+        z_prev = outputs["pair"]
+
+        # TODO bshor: early stop depends on is_multimer, but I don't think it must
+        early_stop = False
+        # if self.globals.is_multimer:
+        #     early_stop = self.tolerance_reached(x_prev, outputs["final_atom_positions"], seq_mask)
+
+        del x_prev
+
+        # [*, N, 3]
+        x_prev = outputs["final_atom_positions"]
+
+        return outputs, m_1_prev, z_prev, x_prev, early_stop
+
+    def forward(self, batch):
+        """
+        Args:
+            batch:
+                Dictionary of arguments outlined in Algorithm 2. Keys must
+                include the official names of the features in the
+                supplement subsection 1.2.9.
+
+                The final dimension of each input must have length equal to
+                the number of recycling iterations.
+
+                Features (without the recycling dimension):
+
+                    "aatype" ([*, N_res]):
+                        Contrary to the supplement, this tensor of residue
+                        indices is not one-hot.
+                    "protein_target_feat" ([*, N_res, C_tf])
+                        One-hot encoding of the target sequence. C_tf is
+                        config.model.input_embedder.tf_dim.
+                    "residue_index" ([*, N_res])
+                        Tensor whose final dimension consists of
+                        consecutive indices from 0 to N_res.
+                    "token_mask" ([*, N_token])
+                        1-D token mask
+                    "pair_mask" ([*, N_token, N_token])
+                        2-D pair mask
+        """
+        # Initialize recycling embeddings
+        m_1_prev, z_prev, x_prev = None, None, None
+        prevs = [m_1_prev, z_prev, x_prev]
+
+        is_grad_enabled = torch.is_grad_enabled()
+
+        # Main recycling loop
+        num_iters = batch["aatype"].shape[-1]
+        early_stop = False
+        num_recycles = 0
+        for cycle_no in range(num_iters):
+            # Select the features for the current recycling cycle
+            fetch_cur_batch = lambda t: t[..., cycle_no]
+            feats = tensor_tree_map(fetch_cur_batch, batch)
+
+            # Enable grad iff we're training and it's the final recycling layer
+            is_final_iter = cycle_no == (num_iters - 1) or early_stop
+            with torch.set_grad_enabled(is_grad_enabled and is_final_iter):
+                if is_final_iter:
+                    # Sidestep AMP bug (PyTorch issue #65766)
+                    if torch.is_autocast_enabled():
+                        torch.clear_autocast_cache()
+
+                # Run the next iteration of the model
+                outputs, m_1_prev, z_prev, x_prev, early_stop = self.iteration(
+                    feats,
+                    prevs,
+                    _recycle=(num_iters > 1)
+                )
+
+                num_recycles += 1
+
+                if not is_final_iter:
+                    del outputs
+                    prevs = [m_1_prev, z_prev, x_prev]
+                    del m_1_prev, z_prev, x_prev
+                else:
+                    break
+
+        outputs["num_recycles"] = torch.tensor(num_recycles, device=feats["aatype"].device)
+
+        # Run auxiliary heads, remove the recycling dimension batch properties
+        outputs.update(self.aux_heads(outputs, batch["inter_pair_mask"][..., 0], batch["affinity_mask"][..., 0]))
+
+        return outputs

dockformerpp/model/pair_transition.py

@@ -0,0 +1,81 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from dockformerpp.model.primitives import Linear, LayerNorm
+
+
+class PairTransition(nn.Module):
+    """
+    Implements Algorithm 15.
+    """
+
+    def __init__(self, c_z, n):
+        """
+        Args:
+            c_z:
+                Pair transition channel dimension
+            n:
+                Factor by which c_z is multiplied to obtain hidden channel
+                dimension
+        """
+        super(PairTransition, self).__init__()
+
+        self.c_z = c_z
+        self.n = n
+
+        self.layer_norm = LayerNorm(self.c_z)
+        self.linear_1 = Linear(self.c_z, self.n * self.c_z, init="relu")
+        self.relu = nn.ReLU()
+        self.linear_2 = Linear(self.n * self.c_z, c_z, init="final")
+
+    def _transition(self, z, mask):
+        # [*, N_res, N_res, C_z]
+        z = self.layer_norm(z)
+        
+        # [*, N_res, N_res, C_hidden]
+        z = self.linear_1(z)
+        z = self.relu(z)
+
+        # [*, N_res, N_res, C_z]
+        z = self.linear_2(z) 
+        z = z * mask
+
+        return z
+
+    def forward(self, 
+        z: torch.Tensor, 
+        mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        """
+        Args:
+            z:
+                [*, N_res, N_res, C_z] pair embedding
+        Returns:
+            [*, N_res, N_res, C_z] pair embedding update
+        """
+        # DISCREPANCY: DeepMind forgets to apply the mask in this module.
+        if mask is None:
+            mask = z.new_ones(z.shape[:-1])
+
+        # [*, N_res, N_res, 1]
+        mask = mask.unsqueeze(-1)
+
+        z = self._transition(z=z, mask=mask)
+
+        return z

dockformerpp/model/primitives.py

@@ -0,0 +1,598 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+import math
+from typing import Optional, Callable, List, Tuple
+import numpy as np
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from scipy.stats import truncnorm
+
+from dockformerpp.utils.kernel.attention_core import attention_core
+from dockformerpp.utils.precision_utils import is_fp16_enabled
+from dockformerpp.utils.tensor_utils import (
+    permute_final_dims,
+    flatten_final_dims,
+)
+
+
+# Suited for 40gb GPU
+# DEFAULT_LMA_Q_CHUNK_SIZE = 1024
+# DEFAULT_LMA_KV_CHUNK_SIZE = 4096
+# Suited for 10gb GPU
+DEFAULT_LMA_Q_CHUNK_SIZE = 64
+DEFAULT_LMA_KV_CHUNK_SIZE = 256
+
+
+def _prod(nums):
+    out = 1
+    for n in nums:
+        out = out * n
+    return out
+
+
+def _calculate_fan(linear_weight_shape, fan="fan_in"):
+    fan_out, fan_in = linear_weight_shape
+
+    if fan == "fan_in":
+        f = fan_in
+    elif fan == "fan_out":
+        f = fan_out
+    elif fan == "fan_avg":
+        f = (fan_in + fan_out) / 2
+    else:
+        raise ValueError("Invalid fan option")
+
+    return f
+
+
+def trunc_normal_init_(weights, scale=1.0, fan="fan_in"):
+    shape = weights.shape
+    f = _calculate_fan(shape, fan)
+    scale = scale / max(1, f)
+    a = -2
+    b = 2
+    std = math.sqrt(scale) / truncnorm.std(a=a, b=b, loc=0, scale=1)
+    size = _prod(shape)
+    samples = truncnorm.rvs(a=a, b=b, loc=0, scale=std, size=size)
+    samples = np.reshape(samples, shape)
+    with torch.no_grad():
+        weights.copy_(torch.tensor(samples, device=weights.device))
+
+
+def lecun_normal_init_(weights):
+    trunc_normal_init_(weights, scale=1.0)
+
+
+def he_normal_init_(weights):
+    trunc_normal_init_(weights, scale=2.0)
+
+
+def glorot_uniform_init_(weights):
+    nn.init.xavier_uniform_(weights, gain=1)
+
+
+def final_init_(weights):
+    with torch.no_grad():
+        weights.fill_(0.0)
+
+
+def gating_init_(weights):
+    with torch.no_grad():
+        weights.fill_(0.0)
+
+
+def normal_init_(weights):
+    torch.nn.init.kaiming_normal_(weights, nonlinearity="linear")
+
+
+def ipa_point_weights_init_(weights):
+    with torch.no_grad():
+        softplus_inverse_1 = 0.541324854612918
+        weights.fill_(softplus_inverse_1)
+
+
+class Linear(nn.Linear):
+    """
+    A Linear layer with built-in nonstandard initializations. Called just
+    like torch.nn.Linear.
+
+    Implements the initializers in 1.11.4, plus some additional ones found
+    in the code.
+    """
+
+    def __init__(
+        self,
+        in_dim: int,
+        out_dim: int,
+        bias: bool = True,
+        init: str = "default",
+        init_fn: Optional[Callable[[torch.Tensor, torch.Tensor], None]] = None,
+        precision=None
+    ):
+        """
+        Args:
+            in_dim:
+                The final dimension of inputs to the layer
+            out_dim:
+                The final dimension of layer outputs
+            bias:
+                Whether to learn an additive bias. True by default
+            init:
+                The initializer to use. Choose from:
+
+                "default": LeCun fan-in truncated normal initialization
+                "relu": He initialization w/ truncated normal distribution
+                "glorot": Fan-average Glorot uniform initialization
+                "gating": Weights=0, Bias=1
+                "normal": Normal initialization with std=1/sqrt(fan_in)
+                "final": Weights=0, Bias=0
+
+                Overridden by init_fn if the latter is not None.
+            init_fn:
+                A custom initializer taking weight and bias as inputs.
+                Overrides init if not None.
+        """
+        super(Linear, self).__init__(in_dim, out_dim, bias=bias)
+
+        if bias:
+            with torch.no_grad():
+                self.bias.fill_(0)
+
+        with torch.no_grad():
+            if init_fn is not None:
+                init_fn(self.weight, self.bias)
+            else:
+                if init == "default":
+                    lecun_normal_init_(self.weight)
+                elif init == "relu":
+                    he_normal_init_(self.weight)
+                elif init == "glorot":
+                    glorot_uniform_init_(self.weight)
+                elif init == "gating":
+                    gating_init_(self.weight)
+                    if bias:
+                        self.bias.fill_(1.0)
+                elif init == "normal":
+                    normal_init_(self.weight)
+                elif init == "final":
+                    final_init_(self.weight)
+                else:
+                    raise ValueError("Invalid init string.")
+
+        self.precision = precision
+
+    def forward(self, input: torch.Tensor) -> torch.Tensor:
+        d = input.dtype
+        if self.precision is not None:
+            with torch.cuda.amp.autocast(enabled=False):
+                bias = self.bias.to(dtype=self.precision) if self.bias is not None else None
+                return nn.functional.linear(input.to(dtype=self.precision),
+                                            self.weight.to(dtype=self.precision),
+                                            bias).to(dtype=d)
+
+        if d is torch.bfloat16:
+            with torch.cuda.amp.autocast(enabled=False):
+                bias = self.bias.to(dtype=d) if self.bias is not None else None
+                return nn.functional.linear(input, self.weight.to(dtype=d), bias)
+
+        return nn.functional.linear(input, self.weight, self.bias)
+
+
+class LayerNorm(nn.Module):
+    def __init__(self, c_in, eps=1e-5):
+        super(LayerNorm, self).__init__()
+        
+        self.c_in = (c_in,)
+        self.eps = eps
+
+        self.weight = nn.Parameter(torch.ones(c_in))
+        self.bias = nn.Parameter(torch.zeros(c_in))
+
+    def forward(self, x): 
+        d = x.dtype
+        if d is torch.bfloat16:
+            with torch.cuda.amp.autocast(enabled=False):
+                out = nn.functional.layer_norm(
+                    x, 
+                    self.c_in, 
+                    self.weight.to(dtype=d), 
+                    self.bias.to(dtype=d), 
+                    self.eps
+                )
+        else:
+            out = nn.functional.layer_norm(
+                x,
+                self.c_in,
+                self.weight,
+                self.bias,
+                self.eps,
+            )
+
+        return out
+
+
+@torch.jit.ignore
+def softmax_no_cast(t: torch.Tensor, dim: int = -1) -> torch.Tensor:
+    """
+        Softmax, but without automatic casting to fp32 when the input is of
+        type bfloat16
+    """
+    d = t.dtype
+    if d is torch.bfloat16:
+        with torch.cuda.amp.autocast(enabled=False):
+            s = torch.nn.functional.softmax(t, dim=dim)
+    else:
+        s = torch.nn.functional.softmax(t, dim=dim)
+
+    return s
+
+
+#@torch.jit.script
+def _attention(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, biases: List[torch.Tensor]) -> torch.Tensor:
+    # [*, H, C_hidden, K]
+    key = permute_final_dims(key, (1, 0))
+
+    # [*, H, Q, K]
+    a = torch.matmul(query, key)
+
+    for b in biases:
+        a += b
+
+    a = softmax_no_cast(a, -1)
+
+    # [*, H, Q, C_hidden]
+    a = torch.matmul(a, value)
+
+    return a
+
+
+class Attention(nn.Module):
+    """
+    Standard multi-head attention using AlphaFold's default layer
+    initialization. Allows multiple bias vectors.
+    """
+    def __init__(
+        self,
+        c_q: int,
+        c_k: int,
+        c_v: int,
+        c_hidden: int,
+        no_heads: int,
+        gating: bool = True,
+    ):
+        """
+        Args:
+            c_q:
+                Input dimension of query data
+            c_k:
+                Input dimension of key data
+            c_v:
+                Input dimension of value data
+            c_hidden:
+                Per-head hidden dimension
+            no_heads:
+                Number of attention heads
+            gating:
+                Whether the output should be gated using query data
+        """
+        super(Attention, self).__init__()
+
+        self.c_q = c_q
+        self.c_k = c_k
+        self.c_v = c_v
+        self.c_hidden = c_hidden
+        self.no_heads = no_heads
+        self.gating = gating
+
+        # DISCREPANCY: c_hidden is not the per-head channel dimension, as
+        # stated in the supplement, but the overall channel dimension.
+
+        self.linear_q = Linear(
+            self.c_q, self.c_hidden * self.no_heads, bias=False, init="glorot"
+        )
+        self.linear_k = Linear(
+            self.c_k, self.c_hidden * self.no_heads, bias=False, init="glorot"
+        )
+        self.linear_v = Linear(
+            self.c_v, self.c_hidden * self.no_heads, bias=False, init="glorot"
+        )
+        self.linear_o = Linear(
+            self.c_hidden * self.no_heads, self.c_q, init="final"
+        )
+
+        self.linear_g = None
+        if self.gating:
+            self.linear_g = Linear(
+                self.c_q, self.c_hidden * self.no_heads, init="gating"
+            )
+
+        self.sigmoid = nn.Sigmoid()
+
+    def _prep_qkv(self,
+        q_x: torch.Tensor, 
+        kv_x: torch.Tensor,
+        apply_scale: bool = True
+    ) -> Tuple[
+        torch.Tensor, torch.Tensor, torch.Tensor
+    ]:
+        # [*, Q/K/V, H * C_hidden]
+        q = self.linear_q(q_x)
+        k = self.linear_k(kv_x)
+        v = self.linear_v(kv_x)
+
+        # [*, Q/K, H, C_hidden]
+        q = q.view(q.shape[:-1] + (self.no_heads, -1))
+        k = k.view(k.shape[:-1] + (self.no_heads, -1))
+        v = v.view(v.shape[:-1] + (self.no_heads, -1))
+
+        # [*, H, Q/K, C_hidden]
+        q = q.transpose(-2, -3)
+        k = k.transpose(-2, -3)
+        v = v.transpose(-2, -3)
+
+        if apply_scale:
+            q /= math.sqrt(self.c_hidden)
+
+        return q, k, v
+
+    def _wrap_up(self,
+        o: torch.Tensor, 
+        q_x: torch.Tensor
+    ) -> torch.Tensor:
+        if self.linear_g is not None:
+            g = self.sigmoid(self.linear_g(q_x))
+        
+            # [*, Q, H, C_hidden]
+            g = g.view(g.shape[:-1] + (self.no_heads, -1))
+            o = o * g
+
+        # [*, Q, H * C_hidden]
+        o = flatten_final_dims(o, 2)
+
+        # [*, Q, C_q]
+        o = self.linear_o(o)
+
+        return o
+
+    def forward(
+        self,
+        q_x: torch.Tensor,
+        kv_x: torch.Tensor,
+        biases: Optional[List[torch.Tensor]] = None,
+        use_memory_efficient_kernel: bool = False,
+        use_lma: bool = False,
+        lma_q_chunk_size: int = DEFAULT_LMA_Q_CHUNK_SIZE,
+        lma_kv_chunk_size: int = DEFAULT_LMA_KV_CHUNK_SIZE,
+    ) -> torch.Tensor:
+        """
+        Args:
+            q_x:
+                [*, Q, C_q] query data
+            kv_x:
+                [*, K, C_k] key data
+            biases:
+                List of biases that broadcast to [*, H, Q, K]
+            use_memory_efficient_kernel:
+                Whether to use a custom memory-efficient attention kernel.
+                This should be the default choice for most. If none of the
+                "use_<...>" flags are True, a stock PyTorch implementation
+                is used instead
+            use_lma:
+                Whether to use low-memory attention (Staats & Rabe 2021). If
+                none of the "use_<...>" flags are True, a stock PyTorch 
+                implementation is used instead
+            lma_q_chunk_size:
+                Query chunk size (for LMA)
+            lma_kv_chunk_size:
+                Key/Value chunk size (for LMA)
+        Returns
+            [*, Q, C_q] attention update
+        """
+        if use_lma and (lma_q_chunk_size is None or lma_kv_chunk_size is None):
+            raise ValueError(
+                "If use_lma is specified, lma_q_chunk_size and "
+                "lma_kv_chunk_size must be provided"
+            )
+
+        attn_options = [use_memory_efficient_kernel, use_lma]
+        if sum(attn_options) > 1:
+            raise ValueError(
+                "Choose at most one alternative attention algorithm"
+            )
+
+        if biases is None:
+            biases = []
+        
+        q, k, v = self._prep_qkv(q_x, kv_x, apply_scale=True)
+
+        if is_fp16_enabled():
+            use_memory_efficient_kernel = False
+        
+        if use_memory_efficient_kernel:
+            if len(biases) > 2:
+                raise ValueError(
+                    "If use_memory_efficient_kernel is True, you may only "
+                    "provide up to two bias terms"
+                )
+            o = attention_core(q, k, v, *((biases + [None] * 2)[:2]))
+            o = o.transpose(-2, -3)
+        elif use_lma:
+            biases = [
+                b.expand(b.shape[:-2] + (q_x.shape[-2],) + (kv_x.shape[-2],)) 
+                for b in biases
+            ]
+            o = _lma(q, k, v, biases, lma_q_chunk_size, lma_kv_chunk_size)
+            o = o.transpose(-2, -3)
+        else:
+            o = _attention(q, k, v, biases)
+            o = o.transpose(-2, -3)
+
+        o = self._wrap_up(o, q_x)
+
+        return o
+
+
+class GlobalAttention(nn.Module):
+    def __init__(self, c_in, c_hidden, no_heads, inf, eps):
+        super(GlobalAttention, self).__init__()
+
+        self.c_in = c_in
+        self.c_hidden = c_hidden
+        self.no_heads = no_heads
+        self.inf = inf
+        self.eps = eps
+
+        self.linear_q = Linear(
+            c_in, c_hidden * no_heads, bias=False, init="glorot"
+        )
+
+        self.linear_k = Linear(
+            c_in, c_hidden, bias=False, init="glorot",
+        )
+        self.linear_v = Linear(
+            c_in, c_hidden, bias=False, init="glorot",
+        )
+        self.linear_g = Linear(c_in, c_hidden * no_heads, init="gating")
+        self.linear_o = Linear(c_hidden * no_heads, c_in, init="final")
+
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, 
+        m: torch.Tensor, 
+        mask: torch.Tensor,
+        use_lma: bool = False,
+    ) -> torch.Tensor:
+        # [*, N_res, C_in]
+        q = torch.sum(m * mask.unsqueeze(-1), dim=-2) / (
+            torch.sum(mask, dim=-1)[..., None] + self.eps
+        )
+
+        # [*, N_res, H * C_hidden]
+        q = self.linear_q(q)
+        q *= (self.c_hidden ** (-0.5))
+
+        # [*, N_res, H, C_hidden]
+        q = q.view(q.shape[:-1] + (self.no_heads, -1))
+
+        # [*, N_res, C_hidden]
+        k = self.linear_k(m)
+        v = self.linear_v(m)
+
+        bias = (self.inf * (mask - 1))[..., :, None, :]
+        if not use_lma:
+            # [*, N_res, H, N_seq]
+            a = torch.matmul(
+                q,
+                k.transpose(-1, -2),  # [*, N_res, C_hidden, N_seq]
+            )
+            a += bias
+            a = softmax_no_cast(a)
+
+            # [*, N_res, H, C_hidden]
+            o = torch.matmul(
+                a,
+                v,
+            )
+        else:
+            o = _lma(
+                q, 
+                k, 
+                v, 
+                [bias], 
+                DEFAULT_LMA_Q_CHUNK_SIZE, 
+                DEFAULT_LMA_KV_CHUNK_SIZE
+            )
+
+        # [*, N_res, C_hidden]
+        g = self.sigmoid(self.linear_g(m))
+
+        # [*, N_res, H, C_hidden]
+        g = g.view(g.shape[:-1] + (self.no_heads, -1))
+
+        # [*, N_res, H, C_hidden]
+        o = o.unsqueeze(-3) * g
+
+        # [*, N_res, H * C_hidden]
+        o = o.reshape(o.shape[:-2] + (-1,))
+
+        # [*, N_res, C_in]
+        m = self.linear_o(o)
+
+        return m
+
+
+def _lma(
+    q: torch.Tensor, 
+    k: torch.Tensor, 
+    v: torch.Tensor, 
+    biases: List[torch.Tensor], 
+    q_chunk_size: int, 
+    kv_chunk_size: int,
+):
+    no_q, no_kv = q.shape[-2], k.shape[-2]
+
+    # [*, H, Q, C_hidden]
+    o = q.new_zeros(q.shape)
+    for q_s in range(0, no_q, q_chunk_size):
+        q_chunk = q[..., q_s: q_s + q_chunk_size, :]
+        large_bias_chunks = [
+            b[..., q_s: q_s + q_chunk_size, :] for b in biases
+        ]
+
+        maxes = []
+        weights = []
+        values = []
+        for kv_s in range(0, no_kv, kv_chunk_size):
+            k_chunk = k[..., kv_s: kv_s + kv_chunk_size, :]
+            v_chunk = v[..., kv_s: kv_s + kv_chunk_size, :]
+            small_bias_chunks = [
+                b[..., kv_s: kv_s + kv_chunk_size] for b in large_bias_chunks
+            ]
+
+            a = torch.einsum(
+                "...hqd,...hkd->...hqk", q_chunk, k_chunk,
+            )
+       
+            for b in small_bias_chunks:
+                a += b
+        
+            max_a = torch.max(a, dim=-1, keepdim=True)[0]
+            exp_a = torch.exp(a - max_a)
+            exp_v = torch.einsum("...hvf,...hqv->...hqf", v_chunk, exp_a)
+ 
+            maxes.append(max_a.detach().squeeze(-1))
+            weights.append(torch.sum(exp_a, dim=-1))
+            values.append(exp_v)
+
+        chunk_max = torch.stack(maxes, dim=-3)
+        chunk_weights = torch.stack(weights, dim=-3)
+        chunk_values = torch.stack(values, dim=-4)
+
+        global_max = torch.max(chunk_max, dim=-3, keepdim=True)[0]
+        max_diffs = torch.exp(chunk_max - global_max)
+        chunk_values = chunk_values * max_diffs.unsqueeze(-1)
+        chunk_weights = chunk_weights * max_diffs
+
+        all_values = torch.sum(chunk_values, dim=-4)
+        all_weights = torch.sum(chunk_weights.unsqueeze(-1), dim=-4)
+
+        q_chunk_out = all_values / all_weights
+
+        o[..., q_s: q_s + q_chunk_size, :] = q_chunk_out
+
+    return o

dockformerpp/model/single_attention.py

@@ -0,0 +1,184 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from functools import partial
+import math
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+from typing import Optional, List, Tuple
+
+from dockformerpp.model.primitives import (
+    Linear, 
+    LayerNorm,
+    Attention,
+)
+from dockformerpp.utils.tensor_utils import permute_final_dims
+
+
+class SingleAttention(nn.Module):
+    def __init__(
+        self,
+        c_in,
+        c_hidden,
+        no_heads,
+        pair_bias=False,
+        c_z=None,
+        inf=1e9,
+    ):
+        """
+        Args:
+            c_in:
+                Input channel dimension
+            c_hidden:
+                Per-head hidden channel dimension
+            no_heads:
+                Number of attention heads
+            pair_bias:
+                Whether to use pair embedding bias
+            c_z:
+                Pair embedding channel dimension. Ignored unless pair_bias
+                is true
+            inf:
+                A large number to be used in computing the attention mask
+        """
+        super(SingleAttention, self).__init__()
+
+        self.c_in = c_in
+        self.c_hidden = c_hidden
+        self.no_heads = no_heads
+        self.pair_bias = pair_bias
+        self.c_z = c_z
+        self.inf = inf
+
+        self.layer_norm_m = LayerNorm(self.c_in)
+
+        self.layer_norm_z = None
+        self.linear_z = None
+        if self.pair_bias:
+            self.layer_norm_z = LayerNorm(self.c_z)
+            self.linear_z = Linear(
+                self.c_z, self.no_heads, bias=False, init="normal"
+            )
+        
+        self.mha = Attention(
+            self.c_in, 
+            self.c_in, 
+            self.c_in, 
+            self.c_hidden, 
+            self.no_heads,
+        )
+
+    def _prep_inputs(self,
+        m: torch.Tensor,
+        z: Optional[torch.Tensor],
+        mask: Optional[torch.Tensor],
+        inplace_safe: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: 
+        if mask is None:
+            # [*, N_res]
+            mask = m.new_ones(m.shape[:-1])
+
+        # [*, 1, 1, N_res]
+        mask_bias = (self.inf * (mask - 1))[..., :, None, None, :]
+
+        if (self.pair_bias and 
+            z is not None and                       # For the 
+            self.layer_norm_z is not None and       # benefit of
+            self.linear_z is not None               # TorchScript
+        ):
+            chunks = []
+
+            for i in range(0, z.shape[-3], 256):
+                z_chunk = z[..., i: i + 256, :, :]
+
+                # [*, N_res, N_res, C_z]
+                z_chunk = self.layer_norm_z(z_chunk)
+            
+                # [*, N_res, N_res, no_heads]
+                z_chunk = self.linear_z(z_chunk)
+
+                chunks.append(z_chunk)
+            
+            z = torch.cat(chunks, dim=-3)
+            
+            # [*, no_heads, N_res, N_res]
+            z = permute_final_dims(z, (2, 0, 1))
+
+        return m, mask_bias, z
+
+    def forward(self, 
+        m: torch.Tensor, 
+        z: Optional[torch.Tensor] = None, 
+        mask: Optional[torch.Tensor] = None, 
+        use_memory_efficient_kernel: bool = False,
+        use_lma: bool = False,
+        inplace_safe: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+            m:
+                [*, N_res, C_m] single embedding
+            z:
+                [*, N_res, N_res, C_z] pair embedding. Required only if pair_bias is True
+            mask:
+                [*, N_res] single mask
+        """
+        m, mask_bias, z = self._prep_inputs(
+            m, z, mask, inplace_safe=inplace_safe
+        )
+
+        biases = [mask_bias]
+        if(z is not None):
+            biases.append(z)
+
+        m = self.layer_norm_m(m)
+        m = self.mha(
+            q_x=m,
+            kv_x=m,
+            biases=biases,
+            use_memory_efficient_kernel=use_memory_efficient_kernel,
+            use_lma=use_lma,
+        )
+
+        return m
+
+
+class SingleRowAttentionWithPairBias(SingleAttention):
+    """
+    Implements Algorithm 7.
+    """
+
+    def __init__(self, c_m, c_z, c_hidden, no_heads, inf=1e9):
+        """
+        Args:
+            c_m:
+                Input channel dimension
+            c_z:
+                Pair embedding channel dimension
+            c_hidden:
+                Per-head hidden channel dimension
+            no_heads:
+                Number of attention heads
+            inf:
+                Large number used to construct attention masks
+        """
+        super(SingleRowAttentionWithPairBias, self).__init__(
+            c_m,
+            c_hidden,
+            no_heads,
+            pair_bias=True,
+            c_z=c_z,
+            inf=inf,
+        )

dockformerpp/model/structure_module.py

@@ -0,0 +1,837 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from functools import reduce
+import importlib
+import math
+import sys
+from operator import mul
+
+import torch
+import torch.nn as nn
+from typing import Optional, Tuple, Sequence, Union
+
+from dockformerpp.model.primitives import Linear, LayerNorm, ipa_point_weights_init_
+from dockformerpp.utils.residue_constants import (
+    restype_rigid_group_default_frame,
+    restype_atom14_to_rigid_group,
+    restype_atom14_mask,
+    restype_atom14_rigid_group_positions,
+)
+from dockformerpp.utils.geometry.quat_rigid import QuatRigid
+from dockformerpp.utils.geometry.rigid_matrix_vector import Rigid3Array
+from dockformerpp.utils.geometry.vector import Vec3Array, square_euclidean_distance
+from dockformerpp.utils.feats import (
+    frames_and_literature_positions_to_atom14_pos,
+    torsion_angles_to_frames,
+)
+from dockformerpp.utils.precision_utils import is_fp16_enabled
+from dockformerpp.utils.rigid_utils import Rotation, Rigid
+from dockformerpp.utils.tensor_utils import (
+    dict_multimap,
+    permute_final_dims,
+    flatten_final_dims,
+)
+
+import importlib.util
+attn_core_is_installed = importlib.util.find_spec("attn_core_inplace_cuda") is not None
+attn_core_inplace_cuda = None
+if attn_core_is_installed:
+    attn_core_inplace_cuda = importlib.import_module("attn_core_inplace_cuda")
+
+
+class AngleResnetBlock(nn.Module):
+    def __init__(self, c_hidden):
+        """
+        Args:
+            c_hidden:
+                Hidden channel dimension
+        """
+        super(AngleResnetBlock, self).__init__()
+
+        self.c_hidden = c_hidden
+
+        self.linear_1 = Linear(self.c_hidden, self.c_hidden, init="relu")
+        self.linear_2 = Linear(self.c_hidden, self.c_hidden, init="final")
+
+        self.relu = nn.ReLU()
+
+    def forward(self, a: torch.Tensor) -> torch.Tensor:
+
+        s_initial = a
+
+        a = self.relu(a)
+        a = self.linear_1(a)
+        a = self.relu(a)
+        a = self.linear_2(a)
+
+        return a + s_initial
+
+
+class AngleResnet(nn.Module):
+    """
+    Implements Algorithm 20, lines 11-14
+    """
+
+    def __init__(self, c_in, c_hidden, no_blocks, no_angles, epsilon):
+        """
+        Args:
+            c_in:
+                Input channel dimension
+            c_hidden:
+                Hidden channel dimension
+            no_blocks:
+                Number of resnet blocks
+            no_angles:
+                Number of torsion angles to generate
+            epsilon:
+                Small constant for normalization
+        """
+        super(AngleResnet, self).__init__()
+
+        self.c_in = c_in
+        self.c_hidden = c_hidden
+        self.no_blocks = no_blocks
+        self.no_angles = no_angles
+        self.eps = epsilon
+
+        self.linear_in = Linear(self.c_in, self.c_hidden)
+        self.linear_initial = Linear(self.c_in, self.c_hidden)
+
+        self.layers = nn.ModuleList()
+        for _ in range(self.no_blocks):
+            layer = AngleResnetBlock(c_hidden=self.c_hidden)
+            self.layers.append(layer)
+
+        self.linear_out = Linear(self.c_hidden, self.no_angles * 2)
+
+        self.relu = nn.ReLU()
+
+    def forward(
+        self, s: torch.Tensor, s_initial: torch.Tensor
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            s:
+                [*, C_hidden] single embedding
+            s_initial:
+                [*, C_hidden] single embedding as of the start of the
+                StructureModule
+        Returns:
+            [*, no_angles, 2] predicted angles
+        """
+        # NOTE: The ReLU's applied to the inputs are absent from the supplement
+        # pseudocode but present in the source. For maximal compatibility with
+        # the pretrained weights, I'm going with the source.
+
+        # [*, C_hidden]
+        s_initial = self.relu(s_initial)
+        s_initial = self.linear_initial(s_initial)
+        s = self.relu(s)
+        s = self.linear_in(s)
+        s = s + s_initial
+
+        for l in self.layers:
+            s = l(s)
+
+        s = self.relu(s)
+
+        # [*, no_angles * 2]
+        s = self.linear_out(s)
+
+        # [*, no_angles, 2]
+        s = s.view(s.shape[:-1] + (-1, 2))
+
+        unnormalized_s = s
+        norm_denom = torch.sqrt(
+            torch.clamp(
+                torch.sum(s ** 2, dim=-1, keepdim=True),
+                min=self.eps,
+            )
+        )
+        s = s / norm_denom
+
+        return unnormalized_s, s
+
+
+class PointProjection(nn.Module):
+    def __init__(self,
+        c_hidden: int,
+        num_points: int,
+        no_heads: int,
+        return_local_points: bool = False,
+    ):
+        super().__init__()
+        self.return_local_points = return_local_points
+        self.no_heads = no_heads
+        self.num_points = num_points
+
+        # Multimer requires this to be run with fp32 precision during training
+        precision = None
+        self.linear = Linear(c_hidden, no_heads * 3 * num_points, precision=precision)
+
+    def forward(self, 
+        activations: torch.Tensor, 
+        rigids: Union[Rigid, Rigid3Array],
+    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+        # TODO: Needs to run in high precision during training
+        points_local = self.linear(activations)
+        out_shape = points_local.shape[:-1] + (self.no_heads, self.num_points, 3)
+
+        points_local = torch.split(
+            points_local, points_local.shape[-1] // 3, dim=-1
+        )
+
+        points_local = torch.stack(points_local, dim=-1).view(out_shape)
+
+        points_global = rigids[..., None, None].apply(points_local)
+
+        if(self.return_local_points):
+            return points_global, points_local
+
+        return points_global
+
+
+class InvariantPointAttention(nn.Module):
+    """
+    Implements Algorithm 22.
+    """
+    def __init__(
+        self,
+        c_s: int,
+        c_z: int,
+        c_hidden: int,
+        no_heads: int,
+        no_qk_points: int,
+        no_v_points: int,
+        inf: float = 1e5,
+        eps: float = 1e-8,
+    ):
+        """
+        Args:
+            c_s:
+                Single representation channel dimension
+            c_z:
+                Pair representation channel dimension
+            c_hidden:
+                Hidden channel dimension
+            no_heads:
+                Number of attention heads
+            no_qk_points:
+                Number of query/key points to generate
+            no_v_points:
+                Number of value points to generate
+        """
+        super(InvariantPointAttention, self).__init__()
+
+        self.c_s = c_s
+        self.c_z = c_z
+        self.c_hidden = c_hidden
+        self.no_heads = no_heads
+        self.no_qk_points = no_qk_points
+        self.no_v_points = no_v_points
+        self.inf = inf
+        self.eps = eps
+
+        # These linear layers differ from their specifications in the
+        # supplement. There, they lack bias and use Glorot initialization.
+        # Here as in the official source, they have bias and use the default
+        # Lecun initialization.
+        hc = self.c_hidden * self.no_heads
+        self.linear_q = Linear(self.c_s, hc, bias=True)
+
+        self.linear_q_points = PointProjection(
+            self.c_s,
+            self.no_qk_points,
+            self.no_heads,
+        )
+
+
+        self.linear_kv = Linear(self.c_s, 2 * hc)
+        self.linear_kv_points = PointProjection(
+            self.c_s,
+            self.no_qk_points + self.no_v_points,
+            self.no_heads,
+        )
+
+        self.linear_b = Linear(self.c_z, self.no_heads)
+
+        self.head_weights = nn.Parameter(torch.zeros((no_heads)))
+        ipa_point_weights_init_(self.head_weights)
+
+        concat_out_dim = self.no_heads * (
+            self.c_z + self.c_hidden + self.no_v_points * 4
+        )
+        self.linear_out = Linear(concat_out_dim, self.c_s, init="final")
+
+        self.softmax = nn.Softmax(dim=-1)
+        self.softplus = nn.Softplus()
+
+    def forward(
+        self,
+        s: torch.Tensor,
+        z: torch.Tensor,
+        r: Union[Rigid, Rigid3Array],
+        mask: torch.Tensor,
+        inplace_safe: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+            s:
+                [*, N_res, C_s] single representation
+            z:
+                [*, N_res, N_res, C_z] pair representation
+            r:
+                [*, N_res] transformation object
+            mask:
+                [*, N_res] mask
+        Returns:
+            [*, N_res, C_s] single representation update
+        """
+        z = [z]
+
+        #######################################
+        # Generate scalar and point activations
+        #######################################
+        # [*, N_res, H * C_hidden]
+        q = self.linear_q(s)
+
+        # [*, N_res, H, C_hidden]
+        q = q.view(q.shape[:-1] + (self.no_heads, -1))
+
+        # [*, N_res, H, P_qk]
+        q_pts = self.linear_q_points(s, r)
+
+        # The following two blocks are equivalent
+        # They're separated only to preserve compatibility with old AF weights
+
+        # [*, N_res, H * 2 * C_hidden]
+        kv = self.linear_kv(s)
+
+        # [*, N_res, H, 2 * C_hidden]
+        kv = kv.view(kv.shape[:-1] + (self.no_heads, -1))
+
+        # [*, N_res, H, C_hidden]
+        k, v = torch.split(kv, self.c_hidden, dim=-1)
+
+        kv_pts = self.linear_kv_points(s, r)
+
+        # [*, N_res, H, P_q/P_v, 3]
+        k_pts, v_pts = torch.split(
+            kv_pts, [self.no_qk_points, self.no_v_points], dim=-2
+        )
+
+        ##########################
+        # Compute attention scores
+        ##########################
+        # [*, N_res, N_res, H]
+        b = self.linear_b(z[0])
+
+        # [*, H, N_res, N_res]
+        if (is_fp16_enabled()):
+            with torch.cuda.amp.autocast(enabled=False):
+                a = torch.matmul(
+                    permute_final_dims(q.float(), (1, 0, 2)),  # [*, H, N_res, C_hidden]
+                    permute_final_dims(k.float(), (1, 2, 0)),  # [*, H, C_hidden, N_res]
+                )
+        else:
+            a = torch.matmul(
+                permute_final_dims(q, (1, 0, 2)),  # [*, H, N_res, C_hidden]
+                permute_final_dims(k, (1, 2, 0)),  # [*, H, C_hidden, N_res]
+            )
+
+        a *= math.sqrt(1.0 / (3 * self.c_hidden))
+        a += (math.sqrt(1.0 / 3) * permute_final_dims(b, (2, 0, 1)))
+
+        # [*, N_res, N_res, H, P_q, 3]
+        pt_att = q_pts.unsqueeze(-4) - k_pts.unsqueeze(-5)
+
+        if (inplace_safe):
+            pt_att *= pt_att
+        else:
+            pt_att = pt_att ** 2
+
+        pt_att = sum(torch.unbind(pt_att, dim=-1))
+
+        head_weights = self.softplus(self.head_weights).view(
+            *((1,) * len(pt_att.shape[:-2]) + (-1, 1))
+        )
+        head_weights = head_weights * math.sqrt(
+            1.0 / (3 * (self.no_qk_points * 9.0 / 2))
+        )
+
+        if (inplace_safe):
+            pt_att *= head_weights
+        else:
+            pt_att = pt_att * head_weights
+
+        # [*, N_res, N_res, H]
+        pt_att = torch.sum(pt_att, dim=-1) * (-0.5)
+
+        # [*, N_res, N_res]
+        square_mask = mask.unsqueeze(-1) * mask.unsqueeze(-2)
+        square_mask = self.inf * (square_mask - 1)
+
+        # [*, H, N_res, N_res]
+        pt_att = permute_final_dims(pt_att, (2, 0, 1))
+
+        if (inplace_safe):
+            a += pt_att
+            del pt_att
+            a += square_mask.unsqueeze(-3)
+            # in-place softmax
+            attn_core_inplace_cuda.forward_(
+                a,
+                reduce(mul, a.shape[:-1]),
+                a.shape[-1],
+            )
+        else:
+            a = a + pt_att
+            a = a + square_mask.unsqueeze(-3)
+            a = self.softmax(a)
+
+        ################
+        # Compute output
+        ################
+        # [*, N_res, H, C_hidden]
+        o = torch.matmul(
+            a, v.transpose(-2, -3).to(dtype=a.dtype)
+        ).transpose(-2, -3)
+
+        # [*, N_res, H * C_hidden]
+        o = flatten_final_dims(o, 2)
+
+        # [*, H, 3, N_res, P_v]
+        if (inplace_safe):
+            v_pts = permute_final_dims(v_pts, (1, 3, 0, 2))
+            o_pt = [
+                torch.matmul(a, v.to(a.dtype))
+                for v in torch.unbind(v_pts, dim=-3)
+            ]
+            o_pt = torch.stack(o_pt, dim=-3)
+        else:
+            o_pt = torch.sum(
+                (
+                        a[..., None, :, :, None]
+                        * permute_final_dims(v_pts, (1, 3, 0, 2))[..., None, :, :]
+                ),
+                dim=-2,
+            )
+
+        # [*, N_res, H, P_v, 3]
+        o_pt = permute_final_dims(o_pt, (2, 0, 3, 1))
+        o_pt = r[..., None, None].invert_apply(o_pt)
+
+        # [*, N_res, H * P_v]
+        o_pt_norm = flatten_final_dims(
+            torch.sqrt(torch.sum(o_pt ** 2, dim=-1) + self.eps), 2
+        )
+
+        # [*, N_res, H * P_v, 3]
+        o_pt = o_pt.reshape(*o_pt.shape[:-3], -1, 3)
+        o_pt = torch.unbind(o_pt, dim=-1)
+
+        # [*, N_res, H, C_z]
+        o_pair = torch.matmul(a.transpose(-2, -3), z[0].to(dtype=a.dtype))
+
+        # [*, N_res, H * C_z]
+        o_pair = flatten_final_dims(o_pair, 2)
+
+        # [*, N_res, C_s]
+        s = self.linear_out(
+            torch.cat(
+                (o, *o_pt, o_pt_norm, o_pair), dim=-1
+            ).to(dtype=z[0].dtype)
+        )
+
+        return s
+
+
+class BackboneUpdate(nn.Module):
+    """
+    Implements part of Algorithm 23.
+    """
+
+    def __init__(self, c_s):
+        """
+        Args:
+            c_s:
+                Single representation channel dimension
+        """
+        super(BackboneUpdate, self).__init__()
+
+        self.c_s = c_s
+
+        self.linear = Linear(self.c_s, 6, init="final")
+
+    def forward(self, s: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            [*, N_res, C_s] single representation
+        Returns:
+            [*, N_res, 6] update vector 
+        """
+        # [*, 6]
+        update = self.linear(s)
+
+        return update 
+
+
+class StructureModuleTransitionLayer(nn.Module):
+    def __init__(self, c):
+        super(StructureModuleTransitionLayer, self).__init__()
+
+        self.c = c
+
+        self.linear_1 = Linear(self.c, self.c, init="relu")
+        self.linear_2 = Linear(self.c, self.c, init="relu")
+        self.linear_3 = Linear(self.c, self.c, init="final")
+
+        self.relu = nn.ReLU()
+
+    def forward(self, s):
+        s_initial = s
+        s = self.linear_1(s)
+        s = self.relu(s)
+        s = self.linear_2(s)
+        s = self.relu(s)
+        s = self.linear_3(s)
+
+        s = s + s_initial
+
+        return s
+
+
+class StructureModuleTransition(nn.Module):
+    def __init__(self, c, num_layers, dropout_rate):
+        super(StructureModuleTransition, self).__init__()
+
+        self.c = c
+        self.num_layers = num_layers
+        self.dropout_rate = dropout_rate
+
+        self.layers = nn.ModuleList()
+        for _ in range(self.num_layers):
+            l = StructureModuleTransitionLayer(self.c)
+            self.layers.append(l)
+
+        self.dropout = nn.Dropout(self.dropout_rate)
+        self.layer_norm = LayerNorm(self.c)
+
+    def forward(self, s):
+        for l in self.layers:
+            s = l(s)
+
+        s = self.dropout(s)
+        s = self.layer_norm(s)
+
+        return s
+
+
+class StructureModule(nn.Module):
+    def __init__(
+        self,
+        c_s,
+        c_z,
+        c_ipa,
+        c_resnet,
+        no_heads_ipa,
+        no_qk_points,
+        no_v_points,
+        dropout_rate,
+        no_blocks,
+        no_transition_layers,
+        no_resnet_blocks,
+        no_angles,
+        trans_scale_factor,
+        epsilon,
+        inf,
+        **kwargs,
+    ):
+        """
+        Args:
+            c_s:
+                Single representation channel dimension
+            c_z:
+                Pair representation channel dimension
+            c_ipa:
+                IPA hidden channel dimension
+            c_resnet:
+                Angle resnet (Alg. 23 lines 11-14) hidden channel dimension
+            no_heads_ipa:
+                Number of IPA heads
+            no_qk_points:
+                Number of query/key points to generate during IPA
+            no_v_points:
+                Number of value points to generate during IPA
+            dropout_rate:
+                Dropout rate used throughout the layer
+            no_blocks:
+                Number of structure module blocks
+            no_transition_layers:
+                Number of layers in the single representation transition
+                (Alg. 23 lines 8-9)
+            no_resnet_blocks:
+                Number of blocks in the angle resnet
+            no_angles:
+                Number of angles to generate in the angle resnet
+            trans_scale_factor:
+                Scale of single representation transition hidden dimension
+            epsilon:
+                Small number used in angle resnet normalization
+            inf:
+                Large number used for attention masking
+        """
+        super(StructureModule, self).__init__()
+
+        self.c_s = c_s
+        self.c_z = c_z
+        self.c_ipa = c_ipa
+        self.c_resnet = c_resnet
+        self.no_heads_ipa = no_heads_ipa
+        self.no_qk_points = no_qk_points
+        self.no_v_points = no_v_points
+        self.dropout_rate = dropout_rate
+        self.no_blocks = no_blocks
+        self.no_transition_layers = no_transition_layers
+        self.no_resnet_blocks = no_resnet_blocks
+        self.no_angles = no_angles
+        self.trans_scale_factor = trans_scale_factor
+        self.epsilon = epsilon
+        self.inf = inf
+
+        # Buffers to be lazily initialized later
+        # self.default_frames
+        # self.group_idx
+        # self.atom_mask
+        # self.lit_positions
+
+        self.layer_norm_s = LayerNorm(self.c_s)
+        self.layer_norm_z = LayerNorm(self.c_z)
+
+        self.linear_in = Linear(self.c_s, self.c_s)
+
+        self.ipa = InvariantPointAttention(
+            self.c_s,
+            self.c_z,
+            self.c_ipa,
+            self.no_heads_ipa,
+            self.no_qk_points,
+            self.no_v_points,
+            inf=self.inf,
+            eps=self.epsilon,
+        )
+
+        self.ipa_dropout = nn.Dropout(self.dropout_rate)
+        self.layer_norm_ipa = LayerNorm(self.c_s)
+
+        self.transition = StructureModuleTransition(
+            self.c_s,
+            self.no_transition_layers,
+            self.dropout_rate,
+        )
+
+        self.bb_update = BackboneUpdate(self.c_s)
+
+        self.angle_resnet = AngleResnet(
+            self.c_s,
+            self.c_resnet,
+            self.no_resnet_blocks,
+            self.no_angles,
+            self.epsilon,
+        )
+
+    def forward(
+        self,
+        evoformer_output_dict,
+        aatype,
+        mask=None,
+        inplace_safe=False,
+    ):
+        """
+        Args:
+            evoformer_output_dict:
+                Dictionary containing:
+                    "single":
+                        [*, N_res, C_s] single representation
+                    "pair":
+                        [*, N_res, N_res, C_z] pair representation
+            aatype:
+                [*, N_res] amino acid indices
+            mask:
+                Optional [*, N_res] sequence mask
+        Returns:
+            A dictionary of outputs
+        """
+        s = evoformer_output_dict["single"]
+
+        if mask is None:
+            # [*, N]
+            mask = s.new_ones(s.shape[:-1])
+
+        # [*, N, C_s]
+        s = self.layer_norm_s(s)
+
+        # [*, N, N, C_z]
+        z = self.layer_norm_z(evoformer_output_dict["pair"])
+
+        # [*, N, C_s]
+        s_initial = s
+        s = self.linear_in(s)
+
+        # [*, N]
+        rigids = Rigid.identity(
+            s.shape[:-1],
+            s.dtype, 
+            s.device, 
+            self.training,
+            fmt="quat",
+        )
+        outputs = []
+        for i in range(self.no_blocks):
+            # [*, N, C_s]
+            s = s + self.ipa(
+                s, 
+                z, 
+                rigids,
+                mask, 
+                inplace_safe=inplace_safe,
+            )
+            s = self.ipa_dropout(s)
+            s = self.layer_norm_ipa(s)
+            s = self.transition(s)
+           
+            # [*, N]
+
+            # [*, N_res, 6] vector of translations and rotations
+            bb_update_output = self.bb_update(s)
+
+            rigids = rigids.compose_q_update_vec(bb_update_output)
+
+
+            # To hew as closely as possible to AlphaFold, we convert our
+            # quaternion-based transformations to rotation-matrix ones
+            # here
+            backb_to_global = Rigid(
+                Rotation(
+                    rot_mats=rigids.get_rots().get_rot_mats(),
+                    quats=None
+                ),
+                rigids.get_trans(),
+            )
+
+            backb_to_global = backb_to_global.scale_translation(
+                self.trans_scale_factor
+            )
+
+            # [*, N, 7, 2]
+            unnormalized_angles, angles = self.angle_resnet(s, s_initial)
+
+            all_frames_to_global = self.torsion_angles_to_frames(
+                backb_to_global,
+                angles,
+                aatype,
+            )
+
+            pred_xyz = self.frames_and_literature_positions_to_atom14_pos(
+                all_frames_to_global,
+                aatype,
+            )
+
+            scaled_rigids = rigids.scale_translation(self.trans_scale_factor)
+            
+            preds = {
+                "frames": scaled_rigids.to_tensor_7(),
+                "sidechain_frames": all_frames_to_global.to_tensor_4x4(),
+                "unnormalized_angles": unnormalized_angles,
+                "angles": angles,
+                "positions": pred_xyz,
+                "states": s,
+            }
+
+            outputs.append(preds)
+
+            rigids = rigids.stop_rot_gradient()
+
+        del z
+
+        outputs = dict_multimap(torch.stack, outputs)
+        outputs["single"] = s
+
+        return outputs
+
+    def _init_residue_constants(self, float_dtype, device):
+        if not hasattr(self, "default_frames"):
+            self.register_buffer(
+                "default_frames",
+                torch.tensor(
+                    restype_rigid_group_default_frame,
+                    dtype=float_dtype,
+                    device=device,
+                    requires_grad=False,
+                ),
+                persistent=False,
+            )
+        if not hasattr(self, "group_idx"):
+            self.register_buffer(
+                "group_idx",
+                torch.tensor(
+                    restype_atom14_to_rigid_group,
+                    device=device,
+                    requires_grad=False,
+                ),
+                persistent=False,
+            )
+        if not hasattr(self, "atom_mask"):
+            self.register_buffer(
+                "atom_mask",
+                torch.tensor(
+                    restype_atom14_mask,
+                    dtype=float_dtype,
+                    device=device,
+                    requires_grad=False,
+                ),
+                persistent=False,
+            )
+        if not hasattr(self, "lit_positions"):
+            self.register_buffer(
+                "lit_positions",
+                torch.tensor(
+                    restype_atom14_rigid_group_positions,
+                    dtype=float_dtype,
+                    device=device,
+                    requires_grad=False,
+                ),
+                persistent=False,
+            )
+
+    def torsion_angles_to_frames(self, r, alpha, f):
+        # Lazily initialize the residue constants on the correct device
+        self._init_residue_constants(alpha.dtype, alpha.device)
+        # Separated purely to make testing less annoying
+        return torsion_angles_to_frames(r, alpha, f, self.default_frames)
+
+    def frames_and_literature_positions_to_atom14_pos(
+            self, r, f  # [*, N, 8]  # [*, N]
+    ):
+        # Lazily initialize the residue constants on the correct device
+        self._init_residue_constants(r.dtype, r.device)
+        return frames_and_literature_positions_to_atom14_pos(
+            r,
+            f,
+            self.default_frames,
+            self.group_idx,
+            self.atom_mask,
+            self.lit_positions,
+        )

dockformerpp/model/torchscript.py

@@ -0,0 +1,171 @@
+# Copyright 2021 AlQuraishi Laboratory
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional, Sequence, Tuple
+
+import torch
+import torch.nn as nn
+
+from dockformerpp.model.evoformer import (
+    EvoformerBlock,
+    EvoformerStack,
+)
+from dockformerpp.model.single_attention import SingleRowAttentionWithPairBias
+from dockformerpp.model.primitives import Attention, GlobalAttention
+
+
+def script_preset_(model: torch.nn.Module):
+    """
+    TorchScript a handful of low-level but frequently used submodule types 
+    that are known to be scriptable.
+
+    Args:
+        model: 
+            A torch.nn.Module. It should contain at least some modules from 
+            this repository, or this function won't do anything.
+    """
+    script_submodules_(
+        model, 
+        [
+            nn.Dropout,
+            Attention,
+            GlobalAttention,
+            EvoformerBlock,
+        ],
+        attempt_trace=False,
+        batch_dims=None,
+    ) 
+
+    
+def _get_module_device(module: torch.nn.Module) -> torch.device:
+    """
+    Fetches the device of a module, assuming that all of the module's
+    parameters reside on a single device
+
+    Args:
+        module: A torch.nn.Module
+    Returns:
+        The module's device
+    """
+    return next(module.parameters()).device
+
+
+def _trace_module(module, batch_dims=None):
+    if(batch_dims is None):
+        batch_dims = ()
+
+    # Stand-in values
+    n_seq = 10
+    n_res = 10
+
+    device = _get_module_device(module)
+
+    def msa(channel_dim):
+        return torch.rand(
+            (*batch_dims, n_seq, n_res, channel_dim),
+            device=device,
+        )
+
+    def pair(channel_dim):
+        return torch.rand(
+            (*batch_dims, n_res, n_res, channel_dim),
+            device=device,
+        )
+
+    if(isinstance(module, SingleRowAttentionWithPairBias)):
+        inputs = {
+            "forward": (
+                msa(module.c_in), # m
+                pair(module.c_z), # z
+                torch.randint(
+                    0, 2, 
+                    (*batch_dims, n_seq, n_res)
+                ), # mask
+            ),
+        }
+    else:
+        raise TypeError(
+            f"tracing is not supported for modules of type {type(module)}"
+        )
+
+    return torch.jit.trace_module(module, inputs)
+
+
+def _script_submodules_helper_(
+    model,
+    types,
+    attempt_trace,
+    to_trace,
+):
+    for name, child in model.named_children():
+        if(types is None or any(isinstance(child, t) for t in types)):
+            try:
+                scripted = torch.jit.script(child)
+                setattr(model, name, scripted)
+                continue
+            except (RuntimeError, torch.jit.frontend.NotSupportedError) as e:
+                if(attempt_trace):
+                    to_trace.add(type(child))
+                else:
+                    raise e
+        
+        _script_submodules_helper_(child, types, attempt_trace, to_trace)
+
+
+def _trace_submodules_(
+    model,
+    types,
+    batch_dims=None,
+):
+    for name, child in model.named_children():
+        if(any(isinstance(child, t) for t in types)):
+            traced = _trace_module(child, batch_dims=batch_dims)
+            setattr(model, name, traced)
+        else:
+            _trace_submodules_(child, types, batch_dims=batch_dims)
+
+
+def script_submodules_(
+    model: nn.Module,
+    types: Optional[Sequence[type]] = None,
+    attempt_trace: Optional[bool] = True,
+    batch_dims: Optional[Tuple[int]] = None,
+):
+    """
+    Convert all submodules whose types match one of those in the input 
+    list to recursively scripted equivalents in place. To script the entire
+    model, just call torch.jit.script on it directly.
+
+    When types is None, all submodules are scripted.
+
+    Args:
+        model: 
+            A torch.nn.Module
+        types: 
+            A list of types of submodules to script
+        attempt_trace: 
+            Whether to attempt to trace specified modules if scripting 
+            fails. Recall that tracing eliminates all conditional 
+            logic---with great tracing comes the mild responsibility of 
+            having to remember to ensure that the modules in question 
+            perform the same computations no matter what.
+    """
+    to_trace = set()
+
+    # Aggressively script as much as possible first...
+    _script_submodules_helper_(model, types, attempt_trace, to_trace)
+  
+    # ... and then trace stragglers.
+    if(attempt_trace and len(to_trace) > 0):
+        _trace_submodules_(model, to_trace, batch_dims=batch_dims)

dockformerpp/model/triangular_attention.py

@@ -0,0 +1,104 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partialmethod, partial
+import math
+from typing import Optional, List
+
+import torch
+import torch.nn as nn
+
+from dockformerpp.model.primitives import Linear, LayerNorm, Attention
+from dockformerpp.utils.tensor_utils import permute_final_dims
+
+
+class TriangleAttention(nn.Module):
+    def __init__(
+        self, c_in, c_hidden, no_heads, starting=True, inf=1e9
+    ):
+        """
+        Args:
+            c_in:
+                Input channel dimension
+            c_hidden:
+                Overall hidden channel dimension (not per-head)
+            no_heads:
+                Number of attention heads
+        """
+        super(TriangleAttention, self).__init__()
+
+        self.c_in = c_in
+        self.c_hidden = c_hidden
+        self.no_heads = no_heads
+        self.starting = starting
+        self.inf = inf
+
+        self.layer_norm = LayerNorm(self.c_in)
+
+        self.linear = Linear(c_in, self.no_heads, bias=False, init="normal")
+
+        self.mha = Attention(
+            self.c_in, self.c_in, self.c_in, self.c_hidden, self.no_heads
+        )
+
+    def forward(self, 
+        x: torch.Tensor, 
+        mask: Optional[torch.Tensor] = None,
+        use_memory_efficient_kernel: bool = False,
+        use_lma: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+            x:
+                [*, I, J, C_in] input tensor (e.g. the pair representation)
+        Returns:
+            [*, I, J, C_in] output tensor
+        """ 
+        if mask is None:
+            # [*, I, J]
+            mask = x.new_ones(
+                x.shape[:-1],
+            )
+
+        if(not self.starting):
+            x = x.transpose(-2, -3)
+            mask = mask.transpose(-1, -2)
+
+        # [*, I, J, C_in]
+        x = self.layer_norm(x)
+
+        # [*, I, 1, 1, J]
+        mask_bias = (self.inf * (mask - 1))[..., :, None, None, :]
+
+        # [*, H, I, J]
+        triangle_bias = permute_final_dims(self.linear(x), (2, 0, 1))
+
+        # [*, 1, H, I, J]
+        triangle_bias = triangle_bias.unsqueeze(-4)
+
+        biases = [mask_bias, triangle_bias]
+
+        x = self.mha(
+            q_x=x,
+            kv_x=x,
+            biases=biases,
+            use_memory_efficient_kernel=use_memory_efficient_kernel,
+            use_lma=use_lma
+        )
+
+        if(not self.starting):
+            x = x.transpose(-2, -3)
+
+        return x

dockformerpp/model/triangular_multiplicative_update.py

@@ -0,0 +1,173 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from functools import partialmethod
+from typing import Optional
+from abc import ABC, abstractmethod
+
+import torch
+import torch.nn as nn
+
+from dockformerpp.model.primitives import Linear, LayerNorm
+from dockformerpp.utils.precision_utils import is_fp16_enabled
+from dockformerpp.utils.tensor_utils import permute_final_dims
+
+
+class BaseTriangleMultiplicativeUpdate(nn.Module, ABC):
+    """
+    Implements Algorithms 11 and 12.
+    """
+    @abstractmethod
+    def __init__(self, c_z, c_hidden, _outgoing):
+        """
+        Args:
+            c_z:
+                Input channel dimension
+            c:
+                Hidden channel dimension
+        """
+        super(BaseTriangleMultiplicativeUpdate, self).__init__()
+        self.c_z = c_z
+        self.c_hidden = c_hidden
+        self._outgoing = _outgoing
+
+        self.linear_g = Linear(self.c_z, self.c_z, init="gating")
+        self.linear_z = Linear(self.c_hidden, self.c_z, init="final")
+
+        self.layer_norm_in = LayerNorm(self.c_z)
+        self.layer_norm_out = LayerNorm(self.c_hidden)
+
+        self.sigmoid = nn.Sigmoid()
+
+    def _combine_projections(self,
+        a: torch.Tensor,
+        b: torch.Tensor,
+    ) -> torch.Tensor:
+        if(self._outgoing):
+            a = permute_final_dims(a, (2, 0, 1))
+            b = permute_final_dims(b, (2, 1, 0))
+        else:
+            a = permute_final_dims(a, (2, 1, 0))
+            b = permute_final_dims(b,  (2, 0, 1))
+
+        p = torch.matmul(a, b)
+
+        return permute_final_dims(p, (1, 2, 0))
+
+    @abstractmethod
+    def forward(self,
+        z: torch.Tensor,
+        mask: Optional[torch.Tensor] = None,
+        inplace_safe: bool = False,
+        _add_with_inplace: bool = False
+    ) -> torch.Tensor:
+        """
+        Args:
+            x:
+                [*, N_res, N_res, C_z] input tensor
+            mask:
+                [*, N_res, N_res] input mask
+        Returns:
+            [*, N_res, N_res, C_z] output tensor
+        """
+        pass
+
+
+class TriangleMultiplicativeUpdate(BaseTriangleMultiplicativeUpdate):
+    """
+    Implements Algorithms 11 and 12.
+    """
+    def __init__(self, c_z, c_hidden, _outgoing=True):
+        """
+        Args:
+            c_z:
+                Input channel dimension
+            c:
+                Hidden channel dimension
+        """
+        super(TriangleMultiplicativeUpdate, self).__init__(c_z=c_z,
+                                                           c_hidden=c_hidden,
+                                                           _outgoing=_outgoing)
+
+        self.linear_a_p = Linear(self.c_z, self.c_hidden)
+        self.linear_a_g = Linear(self.c_z, self.c_hidden, init="gating")
+        self.linear_b_p = Linear(self.c_z, self.c_hidden)
+        self.linear_b_g = Linear(self.c_z, self.c_hidden, init="gating")
+
+    def forward(self, 
+        z: torch.Tensor, 
+        mask: Optional[torch.Tensor] = None,
+        inplace_safe: bool = False,
+        _add_with_inplace: bool = False,
+    ) -> torch.Tensor:
+        """
+        Args:
+            x:
+                [*, N_res, N_res, C_z] input tensor
+            mask:
+                [*, N_res, N_res] input mask
+        Returns:
+            [*, N_res, N_res, C_z] output tensor
+        """
+
+        if mask is None:
+            mask = z.new_ones(z.shape[:-1])
+
+        mask = mask.unsqueeze(-1)
+        
+        z = self.layer_norm_in(z)
+        a = mask
+        a = a * self.sigmoid(self.linear_a_g(z)) 
+        a = a * self.linear_a_p(z)
+        b = mask
+        b = b * self.sigmoid(self.linear_b_g(z))
+        b = b * self.linear_b_p(z)
+
+        # Prevents overflow of torch.matmul in combine projections in
+        # reduced-precision modes
+        a_std = a.std()
+        b_std = b.std()
+        if(is_fp16_enabled() and a_std != 0. and b_std != 0.):
+            a = a / a.std()
+            b = b / b.std()
+
+        if(is_fp16_enabled()):
+            with torch.cuda.amp.autocast(enabled=False):
+                x = self._combine_projections(a.float(), b.float())
+        else:
+            x = self._combine_projections(a, b)
+        
+        del a, b
+        x = self.layer_norm_out(x)
+        x = self.linear_z(x)
+        g = self.sigmoid(self.linear_g(z))
+        x = x * g
+
+        return x
+
+
+class TriangleMultiplicationOutgoing(TriangleMultiplicativeUpdate):
+    """
+    Implements Algorithm 11.
+    """
+    __init__ = partialmethod(TriangleMultiplicativeUpdate.__init__, _outgoing=True)
+
+
+class TriangleMultiplicationIncoming(TriangleMultiplicativeUpdate):
+    """
+    Implements Algorithm 12.
+    """
+    __init__ = partialmethod(TriangleMultiplicativeUpdate.__init__, _outgoing=False)
+

dockformerpp/resources/stereo_chemical_props.txt

@@ -0,0 +1,345 @@
+Bond			Residue		Mean		StdDev
+CA-CB			ALA		1.520		0.021
+N-CA			ALA		1.459		0.020
+CA-C			ALA		1.525		0.026
+C-O			ALA		1.229		0.019
+CA-CB			ARG		1.535		0.022
+CB-CG			ARG		1.521		0.027
+CG-CD			ARG		1.515		0.025
+CD-NE			ARG		1.460		0.017
+NE-CZ			ARG		1.326		0.013
+CZ-NH1			ARG		1.326		0.013
+CZ-NH2			ARG		1.326		0.013
+N-CA			ARG		1.459		0.020
+CA-C			ARG		1.525		0.026
+C-O			ARG		1.229		0.019
+CA-CB			ASN		1.527		0.026
+CB-CG			ASN		1.506		0.023
+CG-OD1			ASN		1.235		0.022
+CG-ND2			ASN		1.324		0.025
+N-CA			ASN		1.459		0.020
+CA-C			ASN		1.525		0.026
+C-O			ASN		1.229		0.019
+CA-CB			ASP		1.535		0.022
+CB-CG			ASP		1.513		0.021
+CG-OD1			ASP		1.249		0.023
+CG-OD2			ASP		1.249		0.023
+N-CA			ASP		1.459		0.020
+CA-C			ASP		1.525		0.026
+C-O			ASP		1.229		0.019
+CA-CB			CYS		1.526		0.013
+CB-SG			CYS		1.812		0.016
+N-CA			CYS		1.459		0.020
+CA-C			CYS		1.525		0.026
+C-O			CYS		1.229		0.019
+CA-CB			GLU		1.535		0.022
+CB-CG			GLU		1.517		0.019
+CG-CD			GLU		1.515		0.015
+CD-OE1			GLU		1.252		0.011
+CD-OE2			GLU		1.252		0.011
+N-CA			GLU		1.459		0.020
+CA-C			GLU		1.525		0.026
+C-O			GLU		1.229		0.019
+CA-CB			GLN		1.535		0.022
+CB-CG			GLN		1.521		0.027
+CG-CD			GLN		1.506		0.023
+CD-OE1			GLN		1.235		0.022
+CD-NE2			GLN		1.324		0.025
+N-CA			GLN		1.459		0.020
+CA-C			GLN		1.525		0.026
+C-O			GLN		1.229		0.019
+N-CA			GLY		1.456		0.015
+CA-C			GLY		1.514		0.016
+C-O			GLY		1.232		0.016
+CA-CB			HIS		1.535		0.022
+CB-CG			HIS		1.492		0.016
+CG-ND1			HIS		1.369		0.015
+CG-CD2			HIS		1.353		0.017
+ND1-CE1			HIS		1.343		0.025
+CD2-NE2			HIS		1.415		0.021
+CE1-NE2			HIS		1.322		0.023
+N-CA			HIS		1.459		0.020
+CA-C			HIS		1.525		0.026
+C-O			HIS		1.229		0.019
+CA-CB			ILE		1.544		0.023
+CB-CG1			ILE		1.536		0.028
+CB-CG2			ILE		1.524		0.031
+CG1-CD1			ILE		1.500		0.069
+N-CA			ILE		1.459		0.020
+CA-C			ILE		1.525		0.026
+C-O			ILE		1.229		0.019
+CA-CB			LEU		1.533		0.023
+CB-CG			LEU		1.521		0.029
+CG-CD1			LEU		1.514		0.037
+CG-CD2			LEU		1.514		0.037
+N-CA			LEU		1.459		0.020
+CA-C			LEU		1.525		0.026
+C-O			LEU		1.229		0.019
+CA-CB			LYS		1.535		0.022
+CB-CG			LYS		1.521		0.027
+CG-CD			LYS		1.520		0.034
+CD-CE			LYS		1.508		0.025
+CE-NZ			LYS		1.486		0.025
+N-CA			LYS		1.459		0.020
+CA-C			LYS		1.525		0.026
+C-O			LYS		1.229		0.019
+CA-CB			MET		1.535		0.022
+CB-CG			MET		1.509		0.032
+CG-SD			MET		1.807		0.026
+SD-CE			MET		1.774		0.056
+N-CA			MET		1.459		0.020
+CA-C			MET		1.525		0.026
+C-O			MET		1.229		0.019
+CA-CB			PHE		1.535		0.022
+CB-CG			PHE		1.509		0.017
+CG-CD1			PHE		1.383		0.015
+CG-CD2			PHE		1.383		0.015
+CD1-CE1			PHE		1.388		0.020
+CD2-CE2			PHE		1.388		0.020
+CE1-CZ			PHE		1.369		0.019
+CE2-CZ			PHE		1.369		0.019
+N-CA			PHE		1.459		0.020
+CA-C			PHE		1.525		0.026
+C-O			PHE		1.229		0.019
+CA-CB			PRO		1.531		0.020
+CB-CG			PRO		1.495		0.050
+CG-CD			PRO		1.502		0.033
+CD-N			PRO		1.474		0.014
+N-CA			PRO		1.468		0.017
+CA-C			PRO		1.524		0.020
+C-O			PRO		1.228		0.020
+CA-CB			SER		1.525		0.015
+CB-OG			SER		1.418		0.013
+N-CA			SER		1.459		0.020
+CA-C			SER		1.525		0.026
+C-O			SER		1.229		0.019
+CA-CB			THR		1.529		0.026
+CB-OG1			THR		1.428		0.020
+CB-CG2			THR		1.519		0.033
+N-CA			THR		1.459		0.020
+CA-C			THR		1.525		0.026
+C-O			THR		1.229		0.019
+CA-CB			TRP		1.535		0.022
+CB-CG			TRP		1.498		0.018
+CG-CD1			TRP		1.363		0.014
+CG-CD2			TRP		1.432		0.017
+CD1-NE1			TRP		1.375		0.017
+NE1-CE2			TRP		1.371		0.013
+CD2-CE2			TRP		1.409		0.012
+CD2-CE3			TRP		1.399		0.015
+CE2-CZ2			TRP		1.393		0.017
+CE3-CZ3			TRP		1.380		0.017
+CZ2-CH2			TRP		1.369		0.019
+CZ3-CH2			TRP		1.396		0.016
+N-CA			TRP		1.459		0.020
+CA-C			TRP		1.525		0.026
+C-O			TRP		1.229		0.019
+CA-CB			TYR		1.535		0.022
+CB-CG			TYR		1.512		0.015
+CG-CD1			TYR		1.387		0.013
+CG-CD2			TYR		1.387		0.013
+CD1-CE1			TYR		1.389		0.015
+CD2-CE2			TYR		1.389		0.015
+CE1-CZ			TYR		1.381		0.013
+CE2-CZ			TYR		1.381		0.013
+CZ-OH			TYR		1.374		0.017
+N-CA			TYR		1.459		0.020
+CA-C			TYR		1.525		0.026
+C-O			TYR		1.229		0.019
+CA-CB			VAL		1.543		0.021
+CB-CG1			VAL		1.524		0.021
+CB-CG2			VAL		1.524		0.021
+N-CA			VAL		1.459		0.020
+CA-C			VAL		1.525		0.026
+C-O			VAL		1.229		0.019
+-
+
+Angle			Residue		Mean		StdDev
+N-CA-CB			ALA		110.1		1.4
+CB-CA-C			ALA		110.1		1.5
+N-CA-C			ALA		111.0		2.7
+CA-C-O			ALA		120.1		2.1
+N-CA-CB			ARG		110.6		1.8
+CB-CA-C			ARG		110.4		2.0
+CA-CB-CG		ARG		113.4		2.2
+CB-CG-CD		ARG		111.6		2.6
+CG-CD-NE		ARG		111.8		2.1
+CD-NE-CZ		ARG		123.6		1.4
+NE-CZ-NH1		ARG		120.3		0.5
+NE-CZ-NH2		ARG		120.3		0.5
+NH1-CZ-NH2		ARG		119.4		1.1
+N-CA-C			ARG		111.0		2.7
+CA-C-O			ARG		120.1		2.1
+N-CA-CB			ASN		110.6		1.8
+CB-CA-C			ASN		110.4		2.0
+CA-CB-CG		ASN		113.4		2.2
+CB-CG-ND2		ASN		116.7		2.4
+CB-CG-OD1		ASN		121.6		2.0
+ND2-CG-OD1		ASN		121.9		2.3
+N-CA-C			ASN		111.0		2.7
+CA-C-O			ASN		120.1		2.1
+N-CA-CB			ASP		110.6		1.8
+CB-CA-C			ASP		110.4		2.0
+CA-CB-CG		ASP		113.4		2.2
+CB-CG-OD1		ASP		118.3		0.9
+CB-CG-OD2		ASP		118.3		0.9
+OD1-CG-OD2		ASP		123.3		1.9
+N-CA-C			ASP		111.0		2.7
+CA-C-O			ASP		120.1		2.1
+N-CA-CB			CYS		110.8		1.5
+CB-CA-C			CYS		111.5		1.2
+CA-CB-SG		CYS		114.2		1.1
+N-CA-C			CYS		111.0		2.7
+CA-C-O			CYS		120.1		2.1
+N-CA-CB			GLU		110.6		1.8
+CB-CA-C			GLU		110.4		2.0
+CA-CB-CG		GLU		113.4		2.2
+CB-CG-CD		GLU		114.2		2.7
+CG-CD-OE1		GLU		118.3		2.0
+CG-CD-OE2		GLU		118.3		2.0
+OE1-CD-OE2		GLU		123.3		1.2
+N-CA-C			GLU		111.0		2.7
+CA-C-O			GLU		120.1		2.1
+N-CA-CB			GLN		110.6		1.8
+CB-CA-C			GLN		110.4		2.0
+CA-CB-CG		GLN		113.4		2.2
+CB-CG-CD		GLN		111.6		2.6
+CG-CD-OE1		GLN		121.6		2.0
+CG-CD-NE2		GLN		116.7		2.4
+OE1-CD-NE2		GLN		121.9		2.3
+N-CA-C			GLN		111.0		2.7
+CA-C-O			GLN		120.1		2.1
+N-CA-C			GLY		113.1		2.5
+CA-C-O			GLY		120.6		1.8
+N-CA-CB			HIS		110.6		1.8
+CB-CA-C			HIS		110.4		2.0
+CA-CB-CG		HIS		113.6		1.7
+CB-CG-ND1		HIS		123.2		2.5
+CB-CG-CD2		HIS		130.8		3.1
+CG-ND1-CE1		HIS		108.2		1.4
+ND1-CE1-NE2		HIS		109.9		2.2
+CE1-NE2-CD2		HIS		106.6		2.5
+NE2-CD2-CG		HIS		109.2		1.9
+CD2-CG-ND1		HIS		106.0		1.4
+N-CA-C			HIS		111.0		2.7
+CA-C-O			HIS		120.1		2.1
+N-CA-CB			ILE		110.8		2.3
+CB-CA-C			ILE		111.6		2.0
+CA-CB-CG1		ILE		111.0		1.9
+CB-CG1-CD1		ILE		113.9		2.8
+CA-CB-CG2		ILE		110.9		2.0
+CG1-CB-CG2		ILE		111.4		2.2
+N-CA-C			ILE		111.0		2.7
+CA-C-O			ILE		120.1		2.1
+N-CA-CB			LEU		110.4		2.0
+CB-CA-C			LEU		110.2		1.9
+CA-CB-CG		LEU		115.3		2.3
+CB-CG-CD1		LEU		111.0		1.7
+CB-CG-CD2		LEU		111.0		1.7
+CD1-CG-CD2		LEU		110.5		3.0
+N-CA-C			LEU		111.0		2.7
+CA-C-O			LEU		120.1		2.1
+N-CA-CB			LYS		110.6		1.8
+CB-CA-C			LYS		110.4		2.0
+CA-CB-CG		LYS		113.4		2.2
+CB-CG-CD		LYS		111.6		2.6
+CG-CD-CE		LYS		111.9		3.0
+CD-CE-NZ		LYS		111.7		2.3
+N-CA-C			LYS		111.0		2.7
+CA-C-O			LYS		120.1		2.1
+N-CA-CB			MET		110.6		1.8
+CB-CA-C			MET		110.4		2.0
+CA-CB-CG		MET		113.3		1.7
+CB-CG-SD		MET		112.4		3.0
+CG-SD-CE		MET		100.2		1.6
+N-CA-C			MET		111.0		2.7
+CA-C-O			MET		120.1		2.1
+N-CA-CB			PHE		110.6		1.8
+CB-CA-C			PHE		110.4		2.0
+CA-CB-CG		PHE		113.9		2.4
+CB-CG-CD1		PHE		120.8		0.7
+CB-CG-CD2		PHE		120.8		0.7
+CD1-CG-CD2		PHE		118.3		1.3
+CG-CD1-CE1		PHE		120.8		1.1
+CG-CD2-CE2		PHE		120.8		1.1
+CD1-CE1-CZ		PHE		120.1		1.2
+CD2-CE2-CZ		PHE		120.1		1.2
+CE1-CZ-CE2		PHE		120.0		1.8
+N-CA-C			PHE		111.0		2.7
+CA-C-O			PHE		120.1		2.1
+N-CA-CB			PRO		103.3		1.2
+CB-CA-C			PRO		111.7		2.1
+CA-CB-CG		PRO		104.8		1.9
+CB-CG-CD		PRO		106.5		3.9
+CG-CD-N			PRO		103.2		1.5
+CA-N-CD			PRO		111.7		1.4
+N-CA-C			PRO		112.1		2.6
+CA-C-O			PRO		120.2		2.4
+N-CA-CB			SER		110.5		1.5
+CB-CA-C			SER		110.1		1.9
+CA-CB-OG		SER		111.2		2.7
+N-CA-C			SER		111.0		2.7
+CA-C-O			SER		120.1		2.1
+N-CA-CB			THR		110.3		1.9
+CB-CA-C			THR		111.6		2.7
+CA-CB-OG1		THR		109.0		2.1
+CA-CB-CG2		THR		112.4		1.4
+OG1-CB-CG2		THR		110.0		2.3
+N-CA-C			THR		111.0		2.7
+CA-C-O			THR		120.1		2.1
+N-CA-CB			TRP		110.6		1.8
+CB-CA-C			TRP		110.4		2.0
+CA-CB-CG		TRP		113.7		1.9
+CB-CG-CD1		TRP		127.0		1.3
+CB-CG-CD2		TRP		126.6		1.3
+CD1-CG-CD2		TRP		106.3		0.8
+CG-CD1-NE1		TRP		110.1		1.0
+CD1-NE1-CE2		TRP		109.0		0.9
+NE1-CE2-CD2		TRP		107.3		1.0
+CE2-CD2-CG		TRP		107.3		0.8
+CG-CD2-CE3		TRP		133.9		0.9
+NE1-CE2-CZ2		TRP		130.4		1.1
+CE3-CD2-CE2		TRP		118.7		1.2
+CD2-CE2-CZ2		TRP		122.3		1.2
+CE2-CZ2-CH2		TRP		117.4		1.0
+CZ2-CH2-CZ3		TRP		121.6		1.2
+CH2-CZ3-CE3		TRP		121.2		1.1
+CZ3-CE3-CD2		TRP		118.8		1.3
+N-CA-C			TRP		111.0		2.7
+CA-C-O			TRP		120.1		2.1
+N-CA-CB			TYR		110.6		1.8
+CB-CA-C			TYR		110.4		2.0
+CA-CB-CG		TYR		113.4		1.9
+CB-CG-CD1		TYR		121.0		0.6
+CB-CG-CD2		TYR		121.0		0.6
+CD1-CG-CD2		TYR		117.9		1.1
+CG-CD1-CE1		TYR		121.3		0.8
+CG-CD2-CE2		TYR		121.3		0.8
+CD1-CE1-CZ		TYR		119.8		0.9
+CD2-CE2-CZ		TYR		119.8		0.9
+CE1-CZ-CE2		TYR		119.8		1.6
+CE1-CZ-OH		TYR		120.1		2.7
+CE2-CZ-OH		TYR		120.1		2.7
+N-CA-C			TYR		111.0		2.7
+CA-C-O			TYR		120.1		2.1
+N-CA-CB			VAL		111.5		2.2
+CB-CA-C			VAL		111.4		1.9
+CA-CB-CG1		VAL		110.9		1.5
+CA-CB-CG2		VAL		110.9		1.5
+CG1-CB-CG2		VAL		110.9		1.6
+N-CA-C			VAL		111.0		2.7
+CA-C-O			VAL		120.1		2.1
+-
+
+Non-bonded distance     Minimum Dist    Tolerance
+C-C                     3.4             1.5
+C-N                     3.25            1.5
+C-S                     3.5             1.5
+C-O                     3.22            1.5
+N-N                     3.1             1.5
+N-S                     3.35            1.5
+N-O                     3.07            1.5
+O-S                     3.32            1.5
+O-O                     3.04            1.5
+S-S                     2.03            1.0
+-

dockformerpp/utils/callbacks.py

@@ -0,0 +1,15 @@
+from lightning.pytorch.callbacks import EarlyStopping
+from lightning_utilities.core.rank_zero import rank_zero_info
+
+
+class EarlyStoppingVerbose(EarlyStopping):
+    """
+        The default EarlyStopping callback's verbose mode is too verbose.
+        This class outputs a message only when it's getting ready to stop. 
+    """
+    def _evalute_stopping_criteria(self, *args, **kwargs):
+        should_stop, reason = super()._evalute_stopping_criteria(*args, **kwargs)
+        if(should_stop):
+            rank_zero_info(f"{reason}\n")
+
+        return should_stop, reason

dockformerpp/utils/checkpointing.py

@@ -0,0 +1,78 @@
+# Copyright 2021 AlQuraishi Laboratory
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import importlib
+from typing import Any, Tuple, List, Callable, Optional
+
+
+import torch
+import torch.utils.checkpoint
+
+
+BLOCK_ARG = Any
+BLOCK_ARGS = List[BLOCK_ARG]
+
+
+@torch.jit.ignore
+def checkpoint_blocks(
+    blocks: List[Callable],
+    args: BLOCK_ARGS,
+    blocks_per_ckpt: Optional[int],
+) -> BLOCK_ARGS:
+    """
+    Chunk a list of blocks and run each chunk with activation
+    checkpointing. We define a "block" as a callable whose only inputs are
+    the outputs of the previous block.
+
+    Implements Subsection 1.11.8
+
+    Args:
+        blocks:
+            List of blocks
+        args:
+            Tuple of arguments for the first block.
+        blocks_per_ckpt:
+            Size of each chunk. A higher value corresponds to fewer 
+            checkpoints, and trades memory for speed. If None, no checkpointing 
+            is performed.
+    Returns:
+        The output of the final block
+    """
+    def wrap(a):
+        return (a,) if type(a) is not tuple else a
+
+    def exec(b, a):
+        for block in b:
+            a = wrap(block(*a))
+        return a
+
+    def chunker(s, e):
+        def exec_sliced(*a):
+            return exec(blocks[s:e], a)
+
+        return exec_sliced
+
+    # Avoids mishaps when the blocks take just one argument
+    args = wrap(args)
+
+    if blocks_per_ckpt is None or not torch.is_grad_enabled():
+        return exec(blocks, args)
+    elif blocks_per_ckpt < 1 or blocks_per_ckpt > len(blocks):
+        raise ValueError("blocks_per_ckpt must be between 1 and len(blocks)")
+
+    for s in range(0, len(blocks), blocks_per_ckpt):
+        e = s + blocks_per_ckpt
+        args = torch.utils.checkpoint.checkpoint(chunker(s, e), *args, use_reentrant=True)
+        args = wrap(args)
+
+    return args

dockformerpp/utils/config_tools.py

@@ -0,0 +1,32 @@
+import importlib.util
+
+import ml_collections as mlc
+
+
+def set_inf(c, inf):
+    for k, v in c.items():
+        if isinstance(v, mlc.ConfigDict):
+            set_inf(v, inf)
+        elif k == "inf":
+            c[k] = inf
+
+
+def enforce_config_constraints(config):
+    def string_to_setting(s):
+        path = s.split('.')
+        setting = config
+        for p in path:
+            setting = setting.get(p)
+
+        return setting
+
+    mutually_exclusive_bools = [
+        (
+            "globals.use_lma",
+        ),
+    ]
+
+    for options in mutually_exclusive_bools:
+        option_settings = [string_to_setting(o) for o in options]
+        if sum(option_settings) > 1:
+            raise ValueError(f"Only one of {', '.join(options)} may be set at a time")

dockformerpp/utils/consts.py

@@ -0,0 +1,25 @@
+from rdkit.Chem.rdchem import ChiralType, BondType
+
+# Survey of atom types in the PDBBind
+# {'C': 403253, 'O': 101283, 'N': 81325, 'S': 6262, 'F': 5256, 'P': 3378, 'Cl': 2920, 'Br': 552, 'B': 237, 'I': 185,
+#  'H': 181, 'Fe': 19, 'Se': 15, 'Ru': 10, 'Si': 5, 'Co': 4, 'Ir': 4, 'As': 2, 'Pt': 2, 'V': 1, 'Mg': 1, 'Be': 1,
+#  'Rh': 1, 'Cu': 1, 'Re': 1}
+# I have changed the uncommon types to common ions for the plinder dataset
+# {'As': "Zn", 'Pt': "Mn", 'V': "Ca", 'Mg': "Mg", 'Be': "Na", 'Rh': "Al", 'Cu': "K", 'Re': "Ni"}
+
+POSSIBLE_ATOM_TYPES = ['C', 'O', 'N', 'S', 'F', 'P', 'Cl', 'Br', 'B', 'I', 'H', 'Fe', 'Se', 'Ru', 'Si', 'Co', 'Ir',
+                       'Zn', 'Mn', 'Ca', 'Mg', 'Na', 'Al', 'K', 'Ni']
+
+# bonds Counter({BondType.SINGLE: 366857, BondType.AROMATIC: 214238, BondType.DOUBLE: 59725, BondType.TRIPLE: 866,
+# BondType.UNSPECIFIED: 18, BondType.DATIVE: 8})
+POSSIBLE_BOND_TYPES = [BondType.SINGLE, BondType.DOUBLE, BondType.TRIPLE, BondType.AROMATIC, BondType.UNSPECIFIED,
+                       BondType.DATIVE]
+
+# {0: 580061, 1: 13273, -1: 11473, 2: 44, 7: 17, -2: 8, 9: 7, 10: 7, 5: 3, 3: 3, 4: 1, 6: 1, 8: 1}
+POSSIBLE_CHARGES = [-1, 0, 1]
+
+# {ChiralType.CHI_UNSPECIFIED: 551374, ChiralType.CHI_TETRAHEDRAL_CCW: 27328, ChiralType.CHI_TETRAHEDRAL_CW: 26178,
+# ChiralType.CHI_OCTAHEDRAL: 13, ChiralType.CHI_SQUAREPLANAR: 3, ChiralType.CHI_TRIGONALBIPYRAMIDAL: 3}
+POSSIBLE_CHIRALITIES = [ChiralType.CHI_UNSPECIFIED, ChiralType.CHI_TETRAHEDRAL_CCW, ChiralType.CHI_TETRAHEDRAL_CW,
+                        ChiralType.CHI_OCTAHEDRAL, ChiralType.CHI_SQUAREPLANAR, ChiralType.CHI_TRIGONALBIPYRAMIDAL]
+

dockformerpp/utils/exponential_moving_average.py

@@ -0,0 +1,71 @@
+from collections import OrderedDict
+import copy
+import torch
+import torch.nn as nn
+
+from dockformerpp.utils.tensor_utils import tensor_tree_map
+
+
+class ExponentialMovingAverage:
+    """
+    Maintains moving averages of parameters with exponential decay
+
+    At each step, the stored copy `copy` of each parameter `param` is
+    updated as follows:
+
+        `copy = decay * copy + (1 - decay) * param`
+
+    where `decay` is an attribute of the ExponentialMovingAverage object.
+    """
+
+    def __init__(self, model: nn.Module, decay: float):
+        """
+        Args:
+            model:
+                A torch.nn.Module whose parameters are to be tracked
+            decay:
+                A value (usually close to 1.) by which updates are
+                weighted as part of the above formula
+        """
+        super(ExponentialMovingAverage, self).__init__()
+
+        clone_param = lambda t: t.clone().detach()
+        self.params = tensor_tree_map(clone_param, model.state_dict())
+        self.decay = decay
+        self.device = next(model.parameters()).device
+
+    def to(self, device):
+        self.params = tensor_tree_map(lambda t: t.to(device), self.params)
+        self.device = device
+
+    def _update_state_dict_(self, update, state_dict):
+        with torch.no_grad():
+            for k, v in update.items():
+                stored = state_dict[k]
+                if not isinstance(v, torch.Tensor):
+                    self._update_state_dict_(v, stored)
+                else:
+                    diff = stored - v
+                    diff *= 1 - self.decay
+                    stored -= diff
+
+    def update(self, model: torch.nn.Module) -> None:
+        """
+        Updates the stored parameters using the state dict of the provided
+        module. The module should have the same structure as that used to
+        initialize the ExponentialMovingAverage object.
+        """
+        self._update_state_dict_(model.state_dict(), self.params)
+
+    def load_state_dict(self, state_dict: OrderedDict) -> None:
+        for k in state_dict["params"].keys():
+            self.params[k] = state_dict["params"][k].clone()
+        self.decay = state_dict["decay"]
+
+    def state_dict(self) -> OrderedDict:
+        return OrderedDict(
+            {
+                "params": self.params,
+                "decay": self.decay,
+            }
+        )

dockformerpp/utils/feats.py

@@ -0,0 +1,174 @@
+# Copyright 2021 AlQuraishi Laboratory
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import math
+
+import numpy as np
+import torch
+import torch.nn as nn
+from typing import Dict, Union
+
+from dockformerpp.utils import protein
+import dockformerpp.utils.residue_constants as rc
+from dockformerpp.utils.geometry import rigid_matrix_vector, rotation_matrix, vector
+from dockformerpp.utils.rigid_utils import Rotation, Rigid
+from dockformerpp.utils.tensor_utils import (
+    batched_gather,
+    one_hot,
+    tree_map,
+    tensor_tree_map,
+)
+
+
+def pseudo_beta_fn(aatype, all_atom_positions, all_atom_masks):
+    # rc.restype_order["X"] defines a ligand, and the atom position used is the CA
+    is_gly_or_lig = (aatype == rc.restype_order["G"]) | (aatype == rc.restype_order["Z"])
+    ca_idx = rc.atom_order["CA"]
+    cb_idx = rc.atom_order["CB"]
+    pseudo_beta = torch.where(
+        is_gly_or_lig[..., None].expand(*((-1,) * len(is_gly_or_lig.shape)), 3),
+        all_atom_positions[..., ca_idx, :],
+        all_atom_positions[..., cb_idx, :],
+    )
+
+    if all_atom_masks is not None:
+        pseudo_beta_mask = torch.where(
+            is_gly_or_lig,
+            all_atom_masks[..., ca_idx],
+            all_atom_masks[..., cb_idx],
+        )
+        return pseudo_beta, pseudo_beta_mask
+    else:
+        return pseudo_beta
+
+
+def atom14_to_atom37(atom14, batch):
+    atom37_data = batched_gather(
+        atom14,
+        batch["residx_atom37_to_atom14"],
+        dim=-2,
+        no_batch_dims=len(atom14.shape[:-2]),
+    )
+
+    atom37_data = atom37_data * batch["atom37_atom_exists"][..., None]
+
+    return atom37_data
+
+
+def torsion_angles_to_frames(
+    r: Union[Rigid, rigid_matrix_vector.Rigid3Array],
+    alpha: torch.Tensor,
+    aatype: torch.Tensor,
+    rrgdf: torch.Tensor,
+):
+
+    rigid_type = type(r)
+
+    # [*, N, 8, 4, 4]
+    default_4x4 = rrgdf[aatype, ...]
+
+    # [*, N, 8] transformations, i.e.
+    #   One [*, N, 8, 3, 3] rotation matrix and
+    #   One [*, N, 8, 3]    translation matrix
+    default_r = rigid_type.from_tensor_4x4(default_4x4)
+
+    bb_rot = alpha.new_zeros((*((1,) * len(alpha.shape[:-1])), 2))
+    bb_rot[..., 1] = 1
+
+    # [*, N, 8, 2]
+    alpha = torch.cat(
+        [bb_rot.expand(*alpha.shape[:-2], -1, -1), alpha], dim=-2
+    )
+
+    # [*, N, 8, 3, 3]
+    # Produces rotation matrices of the form:
+    # [
+    #   [1, 0  , 0  ],
+    #   [0, a_2,-a_1],
+    #   [0, a_1, a_2]
+    # ]
+    # This follows the original code rather than the supplement, which uses
+    # different indices.
+
+    all_rots = alpha.new_zeros(default_r.shape + (4, 4))
+    all_rots[..., 0, 0] = 1
+    all_rots[..., 1, 1] = alpha[..., 1]
+    all_rots[..., 1, 2] = -alpha[..., 0]
+    all_rots[..., 2, 1:3] = alpha
+
+    all_rots = rigid_type.from_tensor_4x4(all_rots)
+    all_frames = default_r.compose(all_rots)
+
+    chi2_frame_to_frame = all_frames[..., 5]
+    chi3_frame_to_frame = all_frames[..., 6]
+    chi4_frame_to_frame = all_frames[..., 7]
+
+    chi1_frame_to_bb = all_frames[..., 4]
+    chi2_frame_to_bb = chi1_frame_to_bb.compose(chi2_frame_to_frame)
+    chi3_frame_to_bb = chi2_frame_to_bb.compose(chi3_frame_to_frame)
+    chi4_frame_to_bb = chi3_frame_to_bb.compose(chi4_frame_to_frame)
+
+    all_frames_to_bb = rigid_type.cat(
+        [
+            all_frames[..., :5],
+            chi2_frame_to_bb.unsqueeze(-1),
+            chi3_frame_to_bb.unsqueeze(-1),
+            chi4_frame_to_bb.unsqueeze(-1),
+        ],
+        dim=-1,
+    )
+    
+    all_frames_to_global = r[..., None].compose(all_frames_to_bb)
+
+    return all_frames_to_global
+
+
+def frames_and_literature_positions_to_atom14_pos(
+    r: Union[Rigid, rigid_matrix_vector.Rigid3Array],
+    aatype: torch.Tensor,
+    default_frames,
+    group_idx,
+    atom_mask,
+    lit_positions,
+):
+    # [*, N, 14, 4, 4]
+    default_4x4 = default_frames[aatype, ...]
+
+    # [*, N, 14]
+    group_mask = group_idx[aatype, ...]
+
+    # [*, N, 14, 8]
+    group_mask = nn.functional.one_hot(
+        group_mask,
+        num_classes=default_frames.shape[-3],
+    )
+
+    # [*, N, 14, 8]
+    t_atoms_to_global = r[..., None, :] * group_mask
+
+    # [*, N, 14]
+    t_atoms_to_global = t_atoms_to_global.map_tensor_fn(
+        lambda x: torch.sum(x, dim=-1)
+    )
+
+    # [*, N, 14]
+    atom_mask = atom_mask[aatype, ...].unsqueeze(-1)
+
+    # [*, N, 14, 3]
+    lit_positions = lit_positions[aatype, ...]
+    pred_positions = t_atoms_to_global.apply(lit_positions)
+    pred_positions = pred_positions * atom_mask
+
+    return pred_positions

dockformerpp/utils/geometry/__init__.py

@@ -0,0 +1,28 @@
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Geometry Module."""
+
+from dockformerpp.utils.geometry import rigid_matrix_vector
+from dockformerpp.utils.geometry import rotation_matrix
+from dockformerpp.utils.geometry import vector
+
+Rot3Array = rotation_matrix.Rot3Array
+Rigid3Array = rigid_matrix_vector.Rigid3Array
+
+Vec3Array = vector.Vec3Array
+square_euclidean_distance = vector.square_euclidean_distance
+euclidean_distance = vector.euclidean_distance
+dihedral_angle = vector.dihedral_angle
+dot = vector.dot
+cross = vector.cross

dockformerpp/utils/geometry/quat_rigid.py

@@ -0,0 +1,38 @@
+import torch
+import torch.nn as nn
+
+from dockformerpp.model.primitives import Linear
+from dockformerpp.utils.geometry.rigid_matrix_vector import Rigid3Array
+from dockformerpp.utils.geometry.rotation_matrix import Rot3Array
+from dockformerpp.utils.geometry.vector import Vec3Array
+
+
+class QuatRigid(nn.Module):
+    def __init__(self, c_hidden, full_quat):
+        super().__init__()
+        self.full_quat = full_quat
+        if self.full_quat:
+            rigid_dim = 7
+        else:
+            rigid_dim = 6
+
+        self.linear = Linear(c_hidden, rigid_dim, init="final", precision=torch.float32)
+
+    def forward(self, activations: torch.Tensor) -> Rigid3Array:
+        # NOTE: During training, this needs to be run in higher precision
+        rigid_flat = self.linear(activations)
+        
+        rigid_flat = torch.unbind(rigid_flat, dim=-1)
+        if(self.full_quat):
+            qw, qx, qy, qz = rigid_flat[:4]
+            translation = rigid_flat[4:]
+        else:
+            qx, qy, qz = rigid_flat[:3]
+            qw = torch.ones_like(qx)
+            translation = rigid_flat[3:]
+
+        rotation = Rot3Array.from_quaternion(
+            qw, qx, qy, qz, normalize=True,
+        )
+        translation = Vec3Array(*translation)
+        return Rigid3Array(rotation, translation)

dockformerpp/utils/geometry/rigid_matrix_vector.py

@@ -0,0 +1,181 @@
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#   http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Rigid3Array Transformations represented by a Matrix and a Vector."""
+
+from __future__ import annotations
+import dataclasses
+from typing import Union, List
+
+import torch
+
+from dockformerpp.utils.geometry import rotation_matrix
+from dockformerpp.utils.geometry import vector
+
+
+Float = Union[float, torch.Tensor]
+
+
+@dataclasses.dataclass(frozen=True)
+class Rigid3Array:
+    """Rigid Transformation, i.e. element of special euclidean group."""
+
+    rotation: rotation_matrix.Rot3Array
+    translation: vector.Vec3Array
+
+    def __matmul__(self, other: Rigid3Array) -> Rigid3Array:
+        new_rotation = self.rotation @ other.rotation # __matmul__
+        new_translation = self.apply_to_point(other.translation)
+        return Rigid3Array(new_rotation, new_translation)
+
+    def __getitem__(self, index) -> Rigid3Array:
+        return Rigid3Array(
+            self.rotation[index],
+            self.translation[index],
+        )
+
+    def __mul__(self, other: torch.Tensor) -> Rigid3Array:
+        return Rigid3Array(
+            self.rotation * other,
+            self.translation * other,
+        )
+
+    def map_tensor_fn(self, fn) -> Rigid3Array:
+        return Rigid3Array(
+            self.rotation.map_tensor_fn(fn),
+            self.translation.map_tensor_fn(fn),
+        )
+
+    def inverse(self) -> Rigid3Array:
+        """Return Rigid3Array corresponding to inverse transform."""
+        inv_rotation = self.rotation.inverse()
+        inv_translation = inv_rotation.apply_to_point(-self.translation)
+        return Rigid3Array(inv_rotation, inv_translation)
+
+    def apply_to_point(self, point: vector.Vec3Array) -> vector.Vec3Array:
+        """Apply Rigid3Array transform to point."""
+        return self.rotation.apply_to_point(point) + self.translation
+
+    def apply(self, point: torch.Tensor) -> torch.Tensor:
+        return self.apply_to_point(vector.Vec3Array.from_array(point)).to_tensor()
+
+    def apply_inverse_to_point(self, point: vector.Vec3Array) -> vector.Vec3Array:
+        """Apply inverse Rigid3Array transform to point."""
+        new_point = point - self.translation
+        return self.rotation.apply_inverse_to_point(new_point)
+
+    def invert_apply(self, point: torch.Tensor) -> torch.Tensor:
+        return self.apply_inverse_to_point(vector.Vec3Array.from_array(point)).to_tensor()
+
+    def compose_rotation(self, other_rotation):
+        rot = self.rotation @ other_rotation
+        return Rigid3Array(rot, self.translation.clone())
+
+    def compose(self, other_rigid):
+        return self @ other_rigid
+
+    def unsqueeze(self, dim: int):
+        return Rigid3Array(
+            self.rotation.unsqueeze(dim),
+            self.translation.unsqueeze(dim),
+        )
+
+    @property
+    def shape(self) -> torch.Size:
+        return self.rotation.xx.shape
+
+    @property
+    def dtype(self) -> torch.dtype:
+        return self.rotation.xx.dtype
+
+    @property
+    def device(self) -> torch.device:
+        return self.rotation.xx.device
+
+    @classmethod
+    def identity(cls, shape, device) -> Rigid3Array:
+        """Return identity Rigid3Array of given shape."""
+        return cls(
+            rotation_matrix.Rot3Array.identity(shape, device),
+            vector.Vec3Array.zeros(shape, device)
+        )
+
+    @classmethod
+    def cat(cls, rigids: List[Rigid3Array], dim: int) -> Rigid3Array:
+        return cls(
+            rotation_matrix.Rot3Array.cat(
+                [r.rotation for r in rigids], dim=dim
+            ),
+            vector.Vec3Array.cat(
+                [r.translation for r in rigids], dim=dim
+            ),
+        ) 
+
+    def scale_translation(self, factor: Float) -> Rigid3Array:
+        """Scale translation in Rigid3Array by 'factor'."""
+        return Rigid3Array(self.rotation, self.translation * factor)
+
+    def to_tensor(self) -> torch.Tensor:
+        rot_array = self.rotation.to_tensor()
+        vec_array = self.translation.to_tensor()
+        array = torch.zeros(
+            rot_array.shape[:-2] + (4, 4), 
+            device=rot_array.device, 
+            dtype=rot_array.dtype
+        )
+        array[..., :3, :3] = rot_array
+        array[..., :3, 3] = vec_array
+        array[..., 3, 3] = 1.
+        return array
+
+    def to_tensor_4x4(self) -> torch.Tensor:
+        return self.to_tensor()
+
+    def reshape(self, new_shape) -> Rigid3Array:
+        rots = self.rotation.reshape(new_shape)
+        trans = self.translation.reshape(new_shape)
+        return Rigid3Array(rots, trans)
+
+    def stop_rot_gradient(self) -> Rigid3Array:
+        return Rigid3Array(
+            self.rotation.stop_gradient(),
+            self.translation,
+        )
+
+    @classmethod
+    def from_array(cls, array):
+        rot = rotation_matrix.Rot3Array.from_array(
+            array[..., :3, :3],
+        )
+        vec = vector.Vec3Array.from_array(array[..., :3, 3])
+        return cls(rot, vec)
+
+    @classmethod
+    def from_tensor_4x4(cls, array):
+        return cls.from_array(array)
+
+    @classmethod
+    def from_array4x4(cls, array: torch.tensor) -> Rigid3Array:
+        """Construct Rigid3Array from homogeneous 4x4 array."""
+        rotation = rotation_matrix.Rot3Array(
+            array[..., 0, 0], array[..., 0, 1], array[..., 0, 2],
+            array[..., 1, 0], array[..., 1, 1], array[..., 1, 2],
+            array[..., 2, 0], array[..., 2, 1], array[..., 2, 2]
+        )
+        translation = vector.Vec3Array(
+            array[..., 0, 3], array[..., 1, 3], array[..., 2, 3]
+        )
+        return cls(rotation, translation)
+
+    def cuda(self) -> Rigid3Array:
+        return Rigid3Array.from_tensor_4x4(self.to_tensor_4x4().cuda())

dockformerpp/utils/geometry/rotation_matrix.py

@@ -0,0 +1,208 @@
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#            http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Rot3Array Matrix Class."""
+
+from __future__ import annotations
+import dataclasses
+from typing import List
+
+import torch
+
+from dockformerpp.utils.geometry import utils
+from dockformerpp.utils.geometry import vector
+from dockformerpp.utils.tensor_utils import tensor_tree_map
+
+
+COMPONENTS = ['xx', 'xy', 'xz', 'yx', 'yy', 'yz', 'zx', 'zy', 'zz']
+
+@dataclasses.dataclass(frozen=True)
+class Rot3Array:
+    """Rot3Array Matrix in 3 dimensional Space implemented as struct of arrays."""
+    xx: torch.Tensor = dataclasses.field(metadata={'dtype': torch.float32})
+    xy: torch.Tensor
+    xz: torch.Tensor
+    yx: torch.Tensor
+    yy: torch.Tensor
+    yz: torch.Tensor
+    zx: torch.Tensor
+    zy: torch.Tensor
+    zz: torch.Tensor
+
+    __array_ufunc__ = None
+
+    def __getitem__(self, index):
+        field_names = utils.get_field_names(Rot3Array) 
+        return Rot3Array(
+            **{
+                name: getattr(self, name)[index] 
+                for name in field_names
+            }
+        )
+    
+    def __mul__(self, other: torch.Tensor):
+        field_names = utils.get_field_names(Rot3Array)
+        return Rot3Array(
+            **{
+                name: getattr(self, name) * other
+                for name in field_names
+            }
+        )
+    
+    def __matmul__(self, other: Rot3Array) -> Rot3Array:
+        """Composes two Rot3Arrays."""
+        c0 = self.apply_to_point(vector.Vec3Array(other.xx, other.yx, other.zx))
+        c1 = self.apply_to_point(vector.Vec3Array(other.xy, other.yy, other.zy))
+        c2 = self.apply_to_point(vector.Vec3Array(other.xz, other.yz, other.zz))
+        return Rot3Array(c0.x, c1.x, c2.x, c0.y, c1.y, c2.y, c0.z, c1.z, c2.z)
+
+    def map_tensor_fn(self, fn) -> Rot3Array:
+        field_names = utils.get_field_names(Rot3Array) 
+        return Rot3Array(
+            **{
+                name: fn(getattr(self, name))
+                for name in field_names
+            }
+        )
+
+    def inverse(self) -> Rot3Array:
+        """Returns inverse of Rot3Array."""
+        return Rot3Array(
+            self.xx, self.yx, self.zx,
+            self.xy, self.yy, self.zy,
+            self.xz, self.yz, self.zz
+        )
+
+    def apply_to_point(self, point: vector.Vec3Array) -> vector.Vec3Array:
+        """Applies Rot3Array to point."""
+        return vector.Vec3Array(
+            self.xx * point.x + self.xy * point.y + self.xz * point.z,
+            self.yx * point.x + self.yy * point.y + self.yz * point.z,
+            self.zx * point.x + self.zy * point.y + self.zz * point.z
+        )
+
+    def apply_inverse_to_point(self, point: vector.Vec3Array) -> vector.Vec3Array:
+        """Applies inverse Rot3Array to point."""
+        return self.inverse().apply_to_point(point)
+
+
+    def unsqueeze(self, dim: int):
+        return Rot3Array(
+            *tensor_tree_map(
+                lambda t: t.unsqueeze(dim), 
+                [getattr(self, c) for c in COMPONENTS]
+            )
+        )
+
+    def stop_gradient(self) -> Rot3Array:
+        return Rot3Array(
+            *[getattr(self, c).detach() for c in COMPONENTS]
+        )
+
+    @classmethod
+    def identity(cls, shape, device) -> Rot3Array:
+        """Returns identity of given shape."""
+        ones = torch.ones(shape, dtype=torch.float32, device=device)
+        zeros = torch.zeros(shape, dtype=torch.float32, device=device)
+        return cls(ones, zeros, zeros, zeros, ones, zeros, zeros, zeros, ones)
+
+    @classmethod
+    def from_two_vectors(
+        cls, e0: vector.Vec3Array,
+        e1: vector.Vec3Array
+    ) -> Rot3Array:
+        """Construct Rot3Array from two Vectors.
+
+        Rot3Array is constructed such that in the corresponding frame 'e0' lies on
+        the positive x-Axis and 'e1' lies in the xy plane with positive sign of y.
+
+        Args:
+            e0: Vector
+            e1: Vector
+        Returns:
+            Rot3Array
+        """
+        # Normalize the unit vector for the x-axis, e0.
+        e0 = e0.normalized()
+        # make e1 perpendicular to e0.
+        c = e1.dot(e0)
+        e1 = (e1 - c * e0).normalized()
+        # Compute e2 as cross product of e0 and e1.
+        e2 = e0.cross(e1)
+        return cls(e0.x, e1.x, e2.x, e0.y, e1.y, e2.y, e0.z, e1.z, e2.z)
+
+    @classmethod
+    def from_array(cls, array: torch.Tensor) -> Rot3Array:
+        """Construct Rot3Array Matrix from array of shape. [..., 3, 3]."""
+        rows = torch.unbind(array, dim=-2)
+        rc = [torch.unbind(e, dim=-1) for e in rows]
+        return cls(*[e for row in rc for e in row])
+
+    def to_tensor(self) -> torch.Tensor:
+        """Convert Rot3Array to array of shape [..., 3, 3]."""
+        return torch.stack(
+            [
+                torch.stack([self.xx, self.xy, self.xz], dim=-1),
+                torch.stack([self.yx, self.yy, self.yz], dim=-1),
+                torch.stack([self.zx, self.zy, self.zz], dim=-1)
+            ],
+            dim=-2
+        )
+
+    @classmethod
+    def from_quaternion(cls,
+        w: torch.Tensor,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        z: torch.Tensor,
+        normalize: bool = True,
+        eps: float = 1e-6
+    ) -> Rot3Array:
+        """Construct Rot3Array from components of quaternion."""
+        if normalize:
+            inv_norm = torch.rsqrt(torch.clamp(w**2 + x**2 + y**2 + z**2, min=eps))
+            w = w * inv_norm
+            x = x * inv_norm
+            y = y * inv_norm
+            z = z * inv_norm
+        xx = 1.0 - 2.0 * (y ** 2 + z ** 2)
+        xy = 2.0 * (x * y - w * z)
+        xz = 2.0 * (x * z + w * y)
+        yx = 2.0 * (x * y + w * z)
+        yy = 1.0 - 2.0 * (x ** 2 + z ** 2)
+        yz = 2.0 * (y * z - w * x)
+        zx = 2.0 * (x * z - w * y)
+        zy = 2.0 * (y * z + w * x)
+        zz = 1.0 - 2.0 * (x ** 2 + y ** 2)
+        return cls(xx, xy, xz, yx, yy, yz, zx, zy, zz)
+
+    def reshape(self, new_shape):
+        field_names = utils.get_field_names(Rot3Array) 
+        reshape_fn = lambda t: t.reshape(new_shape)
+        return Rot3Array(
+            **{
+                name: reshape_fn(getattr(self, name))
+                for name in field_names
+            }
+        )
+
+    @classmethod
+    def cat(cls, rots: List[Rot3Array], dim: int) -> Rot3Array:
+        field_names = utils.get_field_names(Rot3Array) 
+        cat_fn = lambda l: torch.cat(l, dim=dim)
+        return cls(
+            **{
+                name: cat_fn([getattr(r, name) for r in rots])
+                for name in field_names
+            }
+        )

dockformerpp/utils/geometry/test_utils.py

@@ -0,0 +1,97 @@
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Shared utils for tests."""
+
+import dataclasses
+import torch
+
+from dockformerpp.utils.geometry import rigid_matrix_vector
+from dockformerpp.utils.geometry import rotation_matrix
+from dockformerpp.utils.geometry import vector
+
+
+def assert_rotation_matrix_equal(matrix1: rotation_matrix.Rot3Array,
+                                 matrix2: rotation_matrix.Rot3Array):
+    for field in dataclasses.fields(rotation_matrix.Rot3Array):
+        field = field.name
+        assert torch.equal(
+            getattr(matrix1, field), getattr(matrix2, field))
+
+
+def assert_rotation_matrix_close(mat1: rotation_matrix.Rot3Array,
+                                 mat2: rotation_matrix.Rot3Array):
+    assert torch.allclose(mat1.to_tensor(), mat2.to_tensor(), atol=1e-6)
+
+
+def assert_array_equal_to_rotation_matrix(array: torch.Tensor,
+                                          matrix: rotation_matrix.Rot3Array):
+    """Check that array and Matrix match."""
+    assert torch.equal(matrix.xx, array[..., 0, 0])
+    assert torch.equal(matrix.xy, array[..., 0, 1])
+    assert torch.equal(matrix.xz, array[..., 0, 2])
+    assert torch.equal(matrix.yx, array[..., 1, 0])
+    assert torch.equal(matrix.yy, array[..., 1, 1])
+    assert torch.equal(matrix.yz, array[..., 1, 2])
+    assert torch.equal(matrix.zx, array[..., 2, 0])
+    assert torch.equal(matrix.zy, array[..., 2, 1])
+    assert torch.equal(matrix.zz, array[..., 2, 2])
+
+
+def assert_array_close_to_rotation_matrix(array: torch.Tensor,
+                                          matrix: rotation_matrix.Rot3Array):
+    assert torch.allclose(matrix.to_tensor(), array, atol=1e-6)
+
+
+def assert_vectors_equal(vec1: vector.Vec3Array, vec2: vector.Vec3Array):
+    assert torch.equal(vec1.x, vec2.x)
+    assert torch.equal(vec1.y, vec2.y)
+    assert torch.equal(vec1.z, vec2.z)
+
+
+def assert_vectors_close(vec1: vector.Vec3Array, vec2: vector.Vec3Array):
+    assert torch.allclose(vec1.x, vec2.x, atol=1e-6, rtol=0.)
+    assert torch.allclose(vec1.y, vec2.y, atol=1e-6, rtol=0.)
+    assert torch.allclose(vec1.z, vec2.z, atol=1e-6, rtol=0.)
+
+
+def assert_array_close_to_vector(array: torch.Tensor, vec: vector.Vec3Array):
+    assert torch.allclose(vec.to_tensor(), array, atol=1e-6, rtol=0.)
+
+
+def assert_array_equal_to_vector(array: torch.Tensor, vec: vector.Vec3Array):
+    assert torch.equal(vec.to_tensor(), array)
+
+
+def assert_rigid_equal_to_rigid(rigid1: rigid_matrix_vector.Rigid3Array,
+                                rigid2: rigid_matrix_vector.Rigid3Array):
+    assert_rot_trans_equal_to_rigid(rigid1.rotation, rigid1.translation, rigid2)
+
+
+def assert_rigid_close_to_rigid(rigid1: rigid_matrix_vector.Rigid3Array,
+                                rigid2: rigid_matrix_vector.Rigid3Array):
+    assert_rot_trans_close_to_rigid(rigid1.rotation, rigid1.translation, rigid2)
+
+
+def assert_rot_trans_equal_to_rigid(rot: rotation_matrix.Rot3Array,
+                                    trans: vector.Vec3Array,
+                                    rigid: rigid_matrix_vector.Rigid3Array):
+    assert_rotation_matrix_equal(rot, rigid.rotation)
+    assert_vectors_equal(trans, rigid.translation)
+
+
+def assert_rot_trans_close_to_rigid(rot: rotation_matrix.Rot3Array,
+                                    trans: vector.Vec3Array,
+                                    rigid: rigid_matrix_vector.Rigid3Array):
+    assert_rotation_matrix_close(rot, rigid.rotation)
+    assert_vectors_close(trans, rigid.translation)

dockformerpp/utils/geometry/utils.py

@@ -0,0 +1,22 @@
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Utils for geometry library."""
+
+import dataclasses
+
+
+def get_field_names(cls):
+    fields = dataclasses.fields(cls)
+    field_names = [f.name for f in fields]
+    return field_names

dockformerpp/utils/geometry/vector.py

@@ -0,0 +1,261 @@
+# Copyright 2021 DeepMind Technologies Limited
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Vec3Array Class."""
+
+from __future__ import annotations
+import dataclasses
+from typing import Union, List
+
+import torch
+
+Float = Union[float, torch.Tensor]
+
+@dataclasses.dataclass(frozen=True)
+class Vec3Array:
+    x: torch.Tensor = dataclasses.field(metadata={'dtype': torch.float32})
+    y: torch.Tensor
+    z: torch.Tensor
+
+    def __post_init__(self):
+        if hasattr(self.x, 'dtype'):
+            assert self.x.dtype == self.y.dtype
+            assert self.x.dtype == self.z.dtype
+            assert all([x == y for x, y in zip(self.x.shape, self.y.shape)])
+            assert all([x == z for x, z in zip(self.x.shape, self.z.shape)])
+
+    def __add__(self, other: Vec3Array) -> Vec3Array:
+        return Vec3Array(
+            self.x + other.x,
+            self.y + other.y,
+            self.z + other.z,
+        )
+
+    def __sub__(self, other: Vec3Array) -> Vec3Array:
+        return Vec3Array(
+            self.x - other.x,
+            self.y - other.y,
+            self.z - other.z,
+        )
+
+    def __mul__(self, other: Float) -> Vec3Array:
+        return Vec3Array(
+            self.x * other,
+            self.y * other,
+            self.z * other,
+        )
+
+    def __rmul__(self, other: Float) -> Vec3Array:
+        return self * other
+
+    def __truediv__(self, other: Float) -> Vec3Array:
+        return Vec3Array(
+            self.x / other,
+            self.y / other,
+            self.z / other,
+        )
+
+    def __neg__(self) -> Vec3Array:
+        return self * -1 
+
+    def __pos__(self) -> Vec3Array:
+        return self * 1
+
+    def __getitem__(self, index) -> Vec3Array:
+        return Vec3Array(
+            self.x[index],
+            self.y[index],
+            self.z[index],
+        )
+
+    def __iter__(self):
+        return iter((self.x, self.y, self.z))
+
+    @property
+    def shape(self):
+        return self.x.shape
+
+    def map_tensor_fn(self, fn) -> Vec3Array:
+        return Vec3Array(
+            fn(self.x),
+            fn(self.y),
+            fn(self.z),
+        )
+        
+    def cross(self, other: Vec3Array) -> Vec3Array:
+        """Compute cross product between 'self' and 'other'."""
+        new_x = self.y * other.z - self.z * other.y
+        new_y = self.z * other.x - self.x * other.z
+        new_z = self.x * other.y - self.y * other.x
+        return Vec3Array(new_x, new_y, new_z)
+
+    def dot(self, other: Vec3Array) -> Float:
+        """Compute dot product between 'self' and 'other'."""
+        return self.x * other.x + self.y * other.y + self.z * other.z
+
+    def norm(self, epsilon: float = 1e-6) -> Float:
+        """Compute Norm of Vec3Array, clipped to epsilon."""
+        # To avoid NaN on the backward pass, we must use maximum before the sqrt
+        norm2 = self.dot(self)
+        if epsilon:
+            norm2 = torch.clamp(norm2, min=epsilon**2)
+        return torch.sqrt(norm2)
+
+    def norm2(self):
+        return self.dot(self)
+
+    def normalized(self, epsilon: float = 1e-6) -> Vec3Array:
+        """Return unit vector with optional clipping."""
+        return self / self.norm(epsilon)
+
+    def clone(self) -> Vec3Array:
+        return Vec3Array(
+            self.x.clone(),
+            self.y.clone(),
+            self.z.clone(),
+        )
+
+    def reshape(self, new_shape) -> Vec3Array:
+        x = self.x.reshape(new_shape)
+        y = self.y.reshape(new_shape)
+        z = self.z.reshape(new_shape)
+
+        return Vec3Array(x, y, z)
+
+    def sum(self, dim: int) -> Vec3Array:
+        return Vec3Array(
+            torch.sum(self.x, dim=dim),
+            torch.sum(self.y, dim=dim),
+            torch.sum(self.z, dim=dim),
+        )
+
+    def unsqueeze(self, dim: int):
+        return Vec3Array(
+            self.x.unsqueeze(dim),
+            self.y.unsqueeze(dim),
+            self.z.unsqueeze(dim),
+        )
+
+    @classmethod
+    def zeros(cls, shape, device="cpu"):
+        """Return Vec3Array corresponding to zeros of given shape."""
+        return cls(
+            torch.zeros(shape, dtype=torch.float32, device=device), 
+            torch.zeros(shape, dtype=torch.float32, device=device),
+            torch.zeros(shape, dtype=torch.float32, device=device)
+        )
+
+    def to_tensor(self) -> torch.Tensor:
+        return torch.stack([self.x, self.y, self.z], dim=-1)
+
+    @classmethod
+    def from_array(cls, tensor):
+        return cls(*torch.unbind(tensor, dim=-1))
+
+    @classmethod
+    def cat(cls, vecs: List[Vec3Array], dim: int) -> Vec3Array:
+        return cls(
+            torch.cat([v.x for v in vecs], dim=dim),
+            torch.cat([v.y for v in vecs], dim=dim),
+            torch.cat([v.z for v in vecs], dim=dim),
+        )
+
+
+def square_euclidean_distance(
+    vec1: Vec3Array,
+    vec2: Vec3Array,
+    epsilon: float = 1e-6
+) -> Float:
+    """Computes square of euclidean distance between 'vec1' and 'vec2'.
+
+    Args:
+        vec1: Vec3Array to compute    distance to
+        vec2: Vec3Array to compute    distance from, should be
+                    broadcast compatible with 'vec1'
+        epsilon: distance is clipped from below to be at least epsilon
+
+    Returns:
+        Array of square euclidean distances;
+        shape will be result of broadcasting 'vec1' and 'vec2'
+    """
+    difference = vec1 - vec2
+    distance = difference.dot(difference)
+    if epsilon:
+        distance = torch.clamp(distance, min=epsilon)
+    return distance
+
+
+def dot(vector1: Vec3Array, vector2: Vec3Array) -> Float:
+    return vector1.dot(vector2)
+
+
+def cross(vector1: Vec3Array, vector2: Vec3Array) -> Float:
+    return vector1.cross(vector2)
+
+
+def norm(vector: Vec3Array, epsilon: float = 1e-6) -> Float:
+    return vector.norm(epsilon)
+
+
+def normalized(vector: Vec3Array, epsilon: float = 1e-6) -> Vec3Array:
+    return vector.normalized(epsilon)
+
+
+def euclidean_distance(
+    vec1: Vec3Array,
+    vec2: Vec3Array,
+    epsilon: float = 1e-6
+) -> Float:
+    """Computes euclidean distance between 'vec1' and 'vec2'.
+
+    Args:
+        vec1: Vec3Array to compute euclidean distance to
+        vec2: Vec3Array to compute euclidean distance from, should be
+                    broadcast compatible with 'vec1'
+        epsilon: distance is clipped from below to be at least epsilon
+
+    Returns:
+        Array of euclidean distances;
+        shape will be result of broadcasting 'vec1' and 'vec2'
+    """
+    distance_sq = square_euclidean_distance(vec1, vec2, epsilon**2)
+    distance = torch.sqrt(distance_sq)
+    return distance
+
+
+def dihedral_angle(a: Vec3Array, b: Vec3Array, c: Vec3Array,
+                                     d: Vec3Array) -> Float:
+    """Computes torsion angle for a quadruple of points.
+
+    For points (a, b, c, d), this is the angle between the planes defined by
+    points (a, b, c) and (b, c, d). It is also known as the dihedral angle.
+
+    Arguments:
+        a: A Vec3Array of coordinates.
+        b: A Vec3Array of coordinates.
+        c: A Vec3Array of coordinates.
+        d: A Vec3Array of coordinates.
+
+    Returns:
+        A tensor of angles in radians: [-pi, pi].
+    """
+    v1 = a - b
+    v2 = b - c
+    v3 = d - c
+
+    c1 = v1.cross(v2)
+    c2 = v3.cross(v2)
+    c3 = c2.cross(c1)
+
+    v2_mag = v2.norm()
+    return torch.atan2(c3.dot(v2), v2_mag * c1.dot(c2))