In the retrosynthesis tutorial

this code of synthon completion

synthon_optimizer = torch.optim.Adam(synthon_task.parameters(), lr=1e-3)
synthon_solver = core.Engine(synthon_task, synthon_train, synthon_valid,
                             synthon_test, synthon_optimizer,
                             gpus=[0], batch_size=128)
synthon_solver.train(num_epoch=1)
synthon_solver.evaluate("valid")
synthon_solver.save("g2gs_synthon_model.pth")

gives below error

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
Input In [13], in <cell line: 11>()
      7 synthon_optimizer = torch.optim.Adam(synthon_task.parameters(), lr=1e-3)
      8 synthon_solver = core.Engine(synthon_task, synthon_train, synthon_valid,
      9                              synthon_test, synthon_optimizer,
     10                              gpus=[0], batch_size=128)
---> 11 synthon_solver.train(num_epoch=1)
     12 synthon_solver.evaluate("valid")
     13 synthon_solver.save("g2gs_synthon_model.pth")

File /usr/local/lib/python3.9/dist-packages/torchdrug/core/engine.py:155, in Engine.train(self, num_epoch, batch_per_epoch)
    152 if self.device.type == "cuda":
    153     batch = utils.cuda(batch, device=self.device)
--> 155 loss, metric = model(batch)
    156 if not loss.requires_grad:
    157     raise RuntimeError("Loss doesn't require grad. Did you define any loss in the task?")

File /usr/local/lib/python3.9/dist-packages/torch/nn/modules/module.py:1130, in Module._call_impl(self, *input, **kwargs)
   1126 # If we don't have any hooks, we want to skip the rest of the logic in
   1127 # this function, and just call forward.
   1128 if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1129         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1130     return forward_call(*input, **kwargs)
   1131 # Do not call functions when jit is used
   1132 full_backward_hooks, non_full_backward_hooks = [], []

File /usr/local/lib/python3.9/dist-packages/torchdrug/tasks/retrosynthesis.py:592, in SynthonCompletion.forward(self, batch)
    589 all_loss = torch.tensor(0, dtype=torch.float32, device=self.device)
    590 metric = {}
--> 592 pred, target = self.predict_and_target(batch, all_loss, metric)
    593 node_in_pred, node_out_pred, bond_pred, stop_pred = pred
    594 node_in_target, node_out_target, bond_target, stop_target, size = target

File /usr/local/lib/python3.9/dist-packages/torchdrug/tasks/retrosynthesis.py:984, in SynthonCompletion.predict_and_target(self, batch, all_loss, metric)
    981 with reactant.graph():
    982     reactant.reaction = batch["reaction"]
--> 984 graph1, node_in_target1, node_out_target1, bond_target1, stop_target1 = self.all_edge(reactant, synthon)
    985 graph2, node_in_target2, node_out_target2, bond_target2, stop_target2 = self.all_stop(reactant, synthon)
    987 graph = self._cat([graph1, graph2])

File /usr/local/lib/python3.9/dist-packages/torch/autograd/grad_mode.py:27, in _DecoratorContextManager.__call__.<locals>.decorate_context(*args, **kwargs)
     24 @functools.wraps(func)
     25 def decorate_context(*args, **kwargs):
     26     with self.clone():
---> 27         return func(*args, **kwargs)

File /usr/local/lib/python3.9/dist-packages/torchdrug/tasks/retrosynthesis.py:557, in SynthonCompletion.all_edge(self, reactant, synthon)
    555 assert (graph.num_edges % 2 == 0).all()
    556 # node / edge features may change because we mask some nodes / edges
--> 557 graph, feature_valid = self._update_molecule_feature(graph)
    559 return graph[feature_valid], node_in_target[feature_valid], node_out_target[feature_valid], \
    560        bond_target[feature_valid], stop_target[feature_valid]

File /usr/local/lib/python3.9/dist-packages/torchdrug/tasks/retrosynthesis.py:398, in SynthonCompletion._update_molecule_feature(self, graphs)
    395 bond_type[edge_mask] = new_graphs.bond_type.to(device=graphs.device)
    397 with graphs.node():
--> 398     graphs.node_feature = node_feature
    399 with graphs.edge():
    400     graphs.edge_feature = edge_feature

File /usr/local/lib/python3.9/dist-packages/torchdrug/data/graph.py:160, in Graph.__setattr__(self, key, value)
    158 if hasattr(self, "meta_dict"):
    159     self._check_attribute(key, value)
--> 160 super(Graph, self).__setattr__(key, value)

File /usr/local/lib/python3.9/dist-packages/torchdrug/core/core.py:84, in _MetaContainer.__setattr__(self, key, value)
     82     if types:
     83         self.meta_dict[key] = types.copy()
---> 84 self._setattr(key, value)

File /usr/local/lib/python3.9/dist-packages/torchdrug/core/core.py:93, in _MetaContainer._setattr(self, key, value)
     92 def _setattr(self, key, value):
---> 93     return super(_MetaContainer, self).__setattr__(key, value)

AttributeError: can't set attribute

Hi, thanks for sharing this repo!

I am wondering how I could input arbitrary target/product for retrosynthesis analysis? What target format would the model required besides SMILES? In the notebook, it's performing prediction on USPTO dataset. I am interested in knowing how I could apply this model to the target outside of USPTO.

Thanks!!

Hi torchdrug team, thank you for the awesome project! I am playing with molecule generation models, and am interested in trying to reproduce the benchmarks posted here: https://torchdrug.ai/docs/benchmark/generation.html

I am able to follow the tutorial for molecule generation: https://torchdrug.ai/docs/tutorials/generation.html

But I found that there was no mention of how we can evaluate models once they are fully trained. Is there any evaluator class or oracle that can be called to obtain the metrics as in your benchmark?

Additionally, do you have any advice on how to set the hyperparameters to fairly reproduce/compare to the GCPN or GraphAF papers?

Feature

This pull request consists of a redesign of the meter logger. There is a new abstract BaseLogger class in torchdrug.utils.loggers.base_logger. The update and log functions of the core.Meter class are now implemented as methods of this class and are called inside the core.Meter methods.

For a new logger, there are 2 abstract methods in BaseLogger: log and save_hyperparams. These two have to be defined for a new custom logger.

The ConsoleLogger is a child of the BaseLogger class and performs the logging as being done currently in TorchDrug. The ConsoleLogger is always on irrespective if another logger is being used or not.

Similarly, the WandbLogger is also a child of BaseLogger and logs all the metrics to the user's W&B account. The wandb package is not a dependency and if a user tries to use the WandbLogger without installing wandb, they are prompted to install it.

The constructor of core.Engine is updated to take two more optional arguments:

• metric_logger: This can be a str or an instance of the custom logger. The accepted str arguments currently are 'wandb' and 'console' (default value is 'console')
• project: The name of the W&B project the user wants to log to (default value is None)

Example:

engine = core.Engine(...,  metric_logger='wandb', project='PropertyPrediction')

or

 from torchdrug.utils.logger.wandb_logger import WandbLogger
 wandb_logger = WandbLogger(project="PropertyPrediction", name="Toxicity Prediction", save_dir="./ClinTox", log_interval=10)
 engine = core.Engine(..., metric_logger=wandb_logger)

A couple of runs for different tasks logged to W&B

• https://wandb.ai/manan-goel/TorchDrug-Generation
• https://wandb.ai/manan-goel/TorchDrug-Pretrain
• https://wandb.ai/manan-goel/TorchDrug-PropertyPrediction

I would like to use my custom data https://raw.githubusercontent.com/goga0001/graph/main/data.csv I prepared the data as CSV file and followed the implementation of existing datasets:

import os
from torchdrug import data, utils
from torchdrug.core import Registry as R
from collections import defaultdict
from torch.utils import data as torch_data
from torchdrug import data
from torchdrug.utils import doc


@R.register("datasets.Flavonoid2")
@doc.copy_args(data.MoleculeDataset.load_csv, ignore=("smiles_field", "target_fields"))
class Flavonoid2(data.MoleculeDataset):
    """
    Subset of Flavonoid compound database for virtual screening.

    Statistics:
        - #Molecule:  4806
        - #Regression task: 2

    Parameters:
        path (str): path to store the dataset
        verbose (int, optional): output verbose level
        **kwargs
    """

    csv_file = "/content/torchdrug/torchdrug/datasets/data.csv"
    target_fields = ["logP","qed"]

    def __init__(self, path, verbose=1, **kwargs):
        self.load_csv(self.csv_file, smiles_field="smiles", target_fields=self.target_fields,
                      verbose=verbose, **kwargs)

Molecules were constructed from smiles but I get assertion error:

from torch import nn, optim
optimizer = optim.Adam(task.parameters(), lr = 1e-3)
solver = core.Engine(task, dataset, None, None, optimizer,
                     gpus=(0,), batch_size=128, log_interval=10)

solver.train(num_epoch=1)
solver.save("graphaf_flavonoid_1epoch.pkl")

AssertionError                            Traceback (most recent call last)
[<ipython-input-23-a6a027c50b11>](https://localhost:8080/#) in <module>
      4                      gpus=(0,), batch_size=128, log_interval=10)
      5 
----> 6 solver.train(num_epoch=1)
      7 solver.save("graphaf_flavonoid_1epoch.pkl")

10 frames
[/content/torchdrug/torchdrug/layers/conv.py](https://localhost:8080/#) in message_and_aggregate(self, graph, input)
    414 
    415     def message_and_aggregate(self, graph, input):
--> 416         assert graph.num_relation == self.num_relation
    417 
    418         node_in, node_out, relation = graph.edge_list.t()

AssertionError:

Thank you! Looking forward to your reply!

there is an error when i tried to run the following test code.

from torchdrug import data

edge_list = [[0, 1], [1, 2], [2, 3], [3, 4], [4, 5], [5, 0]] graph = data.Graph(edge_list, num_node=6) graph = graph.cuda()

the subgraph induced by nodes 2, 3 & 4

subgraph = graph.subgraph([2, 3, 4])

the error is : TypeError: metaclass conflict: the metaclass of a derived class must be a (non-strict) subclass of the metaclasses of all its bases how can I fix it

Dear everyone,

I have install torchdrug correctly, and then follow the tutorial https://torchdrug.ai/docs/tutorials/retrosynthesis.html When I run the code as below:

from torchdrug import datasets

reaction_dataset = datasets.USPTO50k("D:/test/molecule-datasets/",
                                     node_feature="reaction_reaction_identification",
                                     kekulize=True)
synthon_dataset = datasets.USPTO50k("D:/test/molecule-dataset/", as_synthon=True,
                                    node_feature="synthon_completion",
                                    kekulize=True)

It happens error as follows:

Loading D:/test/molecule-datasets/data_processed.csv: 100%|██████████| 50017/50017 [00:00<00:00, 92358.37it/s]
Constructing molecules from SMILES:   0%|          | 0/50016 [00:00<?, ?it/s]
Traceback (most recent call last):
  File "F:/workdir/pycharm/Retrosynthesis/main.py", line 5, in <module>
    kekulize=True)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\decorator.py", line 232, in fun
    return caller(func, *(extras + args), **kw)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\torchdrug-0.1.0-py3.7.egg\torchdrug\core\core.py", line 282, in wrapper
    return init(self, *args, **kwargs)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\torchdrug-0.1.0-py3.7.egg\torchdrug\datasets\uspto50k.py", line 63, in __init__
    **kwargs)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\torchdrug-0.1.0-py3.7.egg\torchdrug\data\dataset.py", line 112, in load_csv
    self.load_smiles(smiles, targets, verbose=verbose, **kwargs)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\torchdrug-0.1.0-py3.7.egg\torchdrug\data\dataset.py", line 232, in load_smiles
    mol = data.Molecule.from_molecule(mol, **kwargs)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\torchdrug-0.1.0-py3.7.egg\torchdrug\data\molecule.py", line 183, in from_molecule
    func = R.get("features.atom.%s" % name)
  File "D:\soft\Anaconda3\envs\py37\lib\site-packages\torchdrug-0.1.0-py3.7.egg\torchdrug\core\core.py", line 208, in get
    raise KeyError("Can't find `%s` in `%s`" % (key, ".".join(keys[:i])))
KeyError: "Can't find `reaction_reaction_identification` in `features.atom`"

what is the problem? could you help me to solve it? Thanks.

Hey, I found a bug that when I load the related TorchDrug modules, I can't use the torch.optim.lr_scheduler. Look at this picture which comes from the TorchDrug Colab files(Property Prediction). I add one lr_schduler for the optimizer. and it throws an error.

However, When I don't load any TorchDrug modules, I can use the optimizer normally.

Trying to build a customized dataset as follows for the molecular generation task.

@R.register("datasets.Smol")

@doc.copy_args(data.MoleculeDataset.load_csv, ignore=("smiles_field", "target_fields"))

class Smol(data.MoleculeDataset):

  smiles_file = "/content/drive/MyDrive/molecule_design/resources/smiles_train.csv"
  target_fields = ["SPLIT"]

  def __init__(self, smiles_file, verbose=1, **kwargs):
    self.load_csv(self.smiles_file, smiles_field="smiles", target_fields=self.target_fields,lazy=True,
                      verbose=verbose, **kwargs)
    
  def split(self):
    indexes = defaultdict(list)
    for i, split in enumerate(self.targets["SPLIT"]):
        indexes[split].append(i)
    train_set = torch_data.Subset(self, indexes["train"])
    valid_set = torch_data.Subset(self, indexes["valid"])
    test_set = torch_data.Subset(self, indexes["test"])
    return train_set, valid_set, test_set

but get the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
[<ipython-input-29-ada51874abcc>](https://localhost:8080/#) in <module>()
      3 @doc.copy_args(data.MoleculeDataset.load_csv, ignore=("smiles_field", "target_fields"))
      4 
----> 5 class Smol(data.MoleculeDataset):
      6 
      7   smiles_file = "/content/drive/MyDrive/molecule_design/resources/smiles_train.csv"

[/usr/local/lib/python3.7/dist-packages/torchdrug/utils/doc.py](https://localhost:8080/#) in wrapper(obj)
     90         if len(docs) != len(parameters):
     91             raise ValueError("Fail to parse the docstring of `%s`. "
---> 92                              "Inconsistent number of parameters in signature and docstring." % obj.__name__)
     93         new_params = []
     94         new_docs = []

ValueError: Fail to parse the docstring of `Smol`. Inconsistent number of parameters in signature and docstring.

Did I miss something?

Hi,

In the _strict_negative method function of KnowledgeGraphCompletion, if 'A-->B', 'B-->C' (A and B are entities, --> is relation) are the samples of traning set (.i.e. self.fact_graph) while 'A-->C' is the sample of valiation set, then I think 'A-->C' will be regard as a negative sample in the traning stage. Is that a problem?

@torch.no_grad()
def _strict_negative(self, pos_h_index, pos_t_index, pos_r_index):
    batch_size = len(pos_h_index)
    any = -torch.ones_like(pos_h_index)

    pattern = torch.stack([pos_h_index, any, pos_r_index], dim=-1)
    pattern = pattern[:batch_size // 2]

    # ==================== Code I Talk About ======================
    edge_index, num_t_truth = self.fact_graph.match(pattern)
    t_truth_index = self.fact_graph.edge_list[edge_index, 1]
    pos_index = functional._size_to_index(num_t_truth)
    t_mask = torch.ones(len(pattern), self.num_entity, dtype=torch.bool, device=self.device)
    t_mask[pos_index, t_truth_index] = 0
    neg_t_candidate = t_mask.nonzero()[:, 1]
    num_t_candidate = t_mask.sum(dim=-1)
    neg_t_index = functional.variadic_sample(neg_t_candidate, num_t_candidate, self.num_negative)
    # =======================================================

    pattern = torch.stack([any, pos_t_index, pos_r_index], dim=-1)
    pattern = pattern[batch_size // 2:]
    edge_index, num_h_truth = self.fact_graph.match(pattern)
    h_truth_index = self.fact_graph.edge_list[edge_index, 0]
    pos_index = functional._size_to_index(num_h_truth)
    h_mask = torch.ones(len(pattern), self.num_entity, dtype=torch.bool, device=self.device)
    h_mask[pos_index, h_truth_index] = 0
    neg_h_candidate = h_mask.nonzero()[:, 1]
    num_h_candidate = h_mask.sum(dim=-1)
    neg_h_index = functional.variadic_sample(neg_h_candidate, num_h_candidate, self.num_negative)

    neg_index = torch.cat([neg_t_index, neg_h_index])

    return neg_index

I'm interested to understand why it is necessary to overwrite the default nn.Module of torch in patch.py:

https://github.com/DeepGraphLearning/torchdrug/blob/eeee19181572ef5b8a806b71bdd4d2d1a4e27f67/torchdrug/patch.py#L125

This seems to be a quite invasive thing since it alters the behavior of any torch.nn module after torchdrug has been imported.

For example, your implementation of register_buffer in patch.py lacks the keyword argument persistent which is present in native torch: https://github.com/pytorch/pytorch/blob/989b24855efe0a8287954040c89d679625dcabe1/torch/nn/modules/module.py#L277

I would greatly appreciate if you could please let me know how I can fall back to the native torch behavior after having imported torchdrug somewhere above in my code.

I came up against a weird obstacle: after running the same code for Retrosynthesis prediction task on gpu and cpu (perhaps only versions of certain libraries might have differed) I got significantly diffirent results... For gpu the accuracy is much larger. Do you maybe know the reason for this? Because as far as I understand even if results would differ this difference would be pretty small.

Hello,

I was wondering how can I use the generation model to generate specific molecules? For example, I have a small dataset of molecules I am interested in generating, should I use ZINC250k dataset to train GraphAF model on and then use property optimization to generate novel molecules with desired QED, logP properties or should I use my small dataset(around 4k) to train the GraphAF model?

Thank you kindly,

Looking forward for your reply

There was another issue on this that was closed but there wasn't really a resolution. The problem occurs when I use a custom dataset. The dataset loads correctly:

import torch

from torchdrug import datasets

​

dataset = datasets.flav("~/molecule-datasets/", kekulize=True,

                            atom_feature="symbol")

18:01:53   Downloading https://raw.githubusercontent.com/gdeol4/torchdrug/master/flav.csv to /root/molecule-datasets/flav.csv

Loading /root/molecule-datasets/flav.csv: 4807it [00:00, 70415.81it/s]            
Constructing molecules from SMILES: 100%|██████████| 4806/4806 [00:10<00:00, 473.69it/s]

However, when attempting to train a model, I encounter the assertion error:

solver.train(num_epoch=1)
solver.save("graphaf_data_1epoch.pkl")

File /notebooks/torchdrug/torchdrug/layers/conv.py:416, in RelationalGraphConv.message_and_aggregate(self, graph, input)
    415 def message_and_aggregate(self, graph, input):
--> 416     assert graph.num_relation == self.num_relation
    418     node_in, node_out, relation = graph.edge_list.t()
    419     node_out = node_out * self.num_relation + relation

AssertionError: 

There does seem to be a mismatch here:

dataset.num_bond_type

2

model.layers[0].num_relation

2

dataset[0]["graph"].num_relation

tensor(3)

My attempt at fixing it:

for data in dataset:
  data['graph'].num_relation = torch.tensor(2)

But the value remains unchanged

Hello, Thanks for sharing this library! The results of https://torchdrug.ai/docs/tutorials/retrosynthesis.html are different from the G2Gs. For reaction class is un-known, these are the reported results from the paper： top-1 accuracy: 0.489 top-3 accuracy: 0.676 top-5 accuracy: 0.725 top-10 accuracy: 0.755 These are the reported results from the https://torchdrug.ai/docs/tutorials/retrosynthesis.html: top-1 accuracy: 0.47541 top-3 accuracy: 0.741803 top-5 accuracy: 0.827869 top-10 accuracy: 0.879098 I cannot understand why the result of k>2 is higher than reported in the literature. Thank you very much

Below is my code

import torch
from torchdrug import core, datasets, tasks, models
from torchdrug.models import RotatE

import matplotlib
matplotlib.use('TkAgg')
import matplotlib.pyplot as plt

dataset = datasets.FB15k237("~/kg-datasets/")
train_set, valid_set, test_set = dataset.split()


model: RotatE = models.RotatE(num_entity=dataset.num_entity,
                      num_relation=dataset.num_relation,
                      embedding_dim=2048, max_score=9)

task = tasks.KnowledgeGraphCompletion(model, num_negative=256,
                                      adversarial_temperature=1)

optimizer = torch.optim.Adam(task.parameters(), lr=2e-5)
solver= core.Engine(task, train_set, valid_set, test_set, optimizer,
                     gpus=[0], batch_size=1024)
solver.train(num_epoch=100)
solver.evaluate("valid")

Below is the error:

Traceback (most recent call last):
  File "C:\Users\lenovo\PycharmProjects\pythonProject2\main.py", line 23, in <module>
    solver.train(num_epoch=100)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\core\engine.py", line 155, in train
    loss, metric = model(batch)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\tasks\reasoning.py", line 85, in forward
    pred = self.predict(batch, all_loss, metric)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\tasks\reasoning.py", line 160, in predict
    pred = self.model(self.fact_graph, h_index, t_index, r_index, all_loss=all_loss, metric=metric)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torch\nn\modules\module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\models\embedding.py", line 191, in forward
    score = functional.rotate_score(self.entity, self.relation * self.relation_scale,
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\layers\functional\embedding.py", line 266, in rotate_score
    score = RotatEFunction.apply(entity, relation, h_index, t_index, r_index)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\layers\functional\embedding.py", line 108, in forward
    forward = embedding.rotate_forward_cuda
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\utils\torch.py", line 27, in __getattr__
    return getattr(self.module, key)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\utils\decorator.py", line 102, in __get__
    result = self.func(obj)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torchdrug\utils\torch.py", line 31, in module
    return cpp_extension.load(self.name, self.sources, self.extra_cflags, self.extra_cuda_cflags,
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torch\utils\cpp_extension.py", line 1079, in load
    return _jit_compile(
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torch\utils\cpp_extension.py", line 1317, in _jit_compile
    return _import_module_from_library(name, build_directory, is_python_module)
  File "C:\Users\lenovo\.conda\envs\td2\lib\site-packages\torch\utils\cpp_extension.py", line 1699, in _import_module_from_library
    file, path, description = imp.find_module(module_name, [path])
  File "C:\Users\lenovo\.conda\evns\td2\lib\imp.py", line 296, in find_module
    raise ImportError(_ERR_MSG.format(name), name=name)
ImportError: No module named 'embedding'

I've did some research but couldn't figure out why, can anyone help me here?

python = 3.9 torch = 1.13 torchdrug = 0.2.0.post1 torchscatter = 2.1.0 torch cluster = 1.6.0

Code to reproduce error: " import torch import pickle from torchdrug import datasets from torchdrug import core, models, tasks

#dataset = datasets.QM9("~/molecule-datasets/",node_position=True) #with open("QM9.pkl", "wb") as fout:

pickle.dump(dataset, fout)

#exit() with open("QM9.pkl", "rb") as fin: dataset = pickle.load(fin)

lengths = [int(0.8 * len(dataset)), int(0.1 * len(dataset))] lengths += [len(dataset) - sum(lengths)] train_set, valid_set, test_set = torch.utils.data.random_split(dataset, lengths)

model = models.MPNN(input_dim=dataset.node_feature_dim, hidden_dim=256, edge_input_dim=dataset.edge_feature_dim, num_layer=1, num_gru_layer=1, num_mlp_layer=2, num_s2s_step=3, short_cut=False, batch_norm=False, activation='relu', concat_hidden=False)

task = tasks.PropertyPrediction(model, task=dataset.tasks)

optimizer = torch.optim.Adam(task.parameters(), lr=1e-3) solver = core.Engine(task, train_set, valid_set, test_set, optimizer, gpus=[0], batch_size=32) solver.train(num_epoch=1) #solver.evaluate("valid") "

Error: File "/home/nhattrup/deep_learning/final_proj/example.py", line 35, in solver.train(num_epoch=1) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/core/engine.py", line 155, in train loss, metric = model(batch) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl return forward_call(*input, **kwargs) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/tasks/property_prediction.py", line 96, in forward pred = self.predict(batch, all_loss, metric) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/tasks/property_prediction.py", line 134, in predict output = self.model(graph, graph.node_feature.float(), all_loss=all_loss, metric=metric) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl return forward_call(*input, **kwargs) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/models/mpnn.py", line 75, in forward x = self.layer(graph, layer_input) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1190, in _call_impl return forward_call(*input, **kwargs) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/layers/conv.py", line 92, in forward update = self.message_and_aggregate(graph, input) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/layers/conv.py", line 61, in message_and_aggregate message = self.message(graph, input) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torchdrug/layers/conv.py", line 650, in message message = torch.einsum("bed, bd -> be", transform, input[node_in]) File "/home/nhattrup/.conda/envs/dl/lib/python3.9/site-packages/torch/functional.py", line 378, in einsum return _VF.einsum(equation, operands) # type: ignore[attr-defined] RuntimeError: CUDA error: CUBLAS_STATUS_NOT_SUPPORTED when calling cublasSgemmStridedBatched( handle, opa, opb, m, n, k, &alpha, a, lda, stridea, b, ldb, strideb, &beta, c, ldc, stridec, num_batches)

One thing I should note is that I ran into similar issues to issue #95 so my pickle file only contains the molecules that were loaded properly (only < 100 couldn't be loaded properly so essentially the entire dataset). Thanks for any help on this.