Hi here,
I got stuck with using MPNNModel with MolGraphConvFeaturizer which is known to be acceptable featurizer for MPNNModel. If I do model.fit(), then I get AttributeError: 'GraphData' object has no attribute 'num_atoms' . I used the following codes:
featurizer = dc.feat.MolGraphConvFeaturizer()
_feature = featurizer.featurize(input_pd['smiles'].to_list())
train_dataset = dc.data.NumpyDataset(X=_feature, y=input_pd['property'].to_list())
model = dc.models.MPNNModel(n_tasks=1, mode = 'regression')
model.fit(train_dataset, nb_epoch=50)And it's error as below:
1174 start = 0
1175 for im, mol in enumerate(X_b):
-> 1176 n_atoms = mol.get_num_atoms()
1177 # number of atoms in each molecule
1178 atom_split.extend([im] * n_atoms)
AttributeError: 'GraphData' object has no attribute 'get_num_atoms'MolGraphConvModel but the default TF one uses WeaveModel https://github.com/deepchem/deepchem/blob/master/deepchem/models/tests/test_graph_models.py#L258
Modeling_Protein_Ligand_Interactions_With_Atomic_Convolutions.ipynb notebook but when running acm.fit(train, nb_epoch=max_epochs, max_checkpoints_to_keep=1, callbacks=[val_cb]) I get ValueError: could not broadcast input array from shape (154,) into shape (1,) (I tested on colab and in a local containter, same error in both). Is there any change of the imported data size?
I am planning to contribute to deepchem for the first time. I am trying to port LSTMStep to PyTorch as first step. Precommit hooks like flake8 are failing due to changes which I haven't made (as below). Is there any general procedure in such scenarios? Should I fix these checks or is there any way to ignore these and proceed with commit and push
deepchem/models/tests/test_layers.py:8:3: F401 'tensorflow.python.framework.test_util' imported but unused
deepchem/models/tests/test_layers.py:124:3: F841 local variable 'out_channels' is assigned to but never used
deepchem/models/tests/test_layers.py:169:3: F841 local variable 'out_channels' is assigned to but never used
deepchem/models/tests/test_layers.py:211:3: F401 'tensorflow as tf' imported but unused
deepchem/models/tests/test_layers.py:213:3: F841 local variable 'out_channels' is assigned to but never used
deepchem/models/tests/test_layers.py:232:3: E265 block comment should start with '# '
deepchem/models/tests/test_layers.py:239:3: E265 block comment should start with '# '
deepchem/models/tests/test_layers.py:290:3: F841 local variable 'n_atoms' is assigned to but never used
deepchem/models/tests/test_layers.py:310:3: F841 local variable 'max_depth' is assigned to but never used
deepchem/models/tests/test_layers.py:501:3: F841 local variable 'result' is assigned to but never used
deepchem/models/tests/test_layers.py:574:3: E265 block comment should start with '# '
deepchem/models/tests/test_layers.py:579:3: F841 local variable 'outputs' is assigned to but never used
deepchem/models/tests/test_layers.py:585:7: E265 block comment should start with '# '
deepchem/models/tests/test_layers.py:587:3: E266 too many leading '#' for block comment
deepchem/models/tests/test_layers.py:588:3: E266 too many leading '#' for block comment
deepchem/models/tests/test_layers.py:608:3: F841 local variable 'outputs' is assigned to but never used
There I want to ask some questions about how to better use it.
I used a graph model created by deepchem.models.GraphConvModel class, and I set the parameters n_tasks=1, dropout=0.5, optimizer=dc.models.optimizers.Adam(learning_rate=0.001), batch_size=128, graph_conv_layers=[32, 32], dense_layer_size=64, number_atom_features=75, mode="regression" to train the model. The first question is how I can visualize the model to see what actually the layers' structure is, just like the torch.model.summary.
After training the first model, I also want to use another dataset to train the model using transfer learning technology. So the second question is how I can change the model's structure, such as how can I add another dense layer to the model at the end. The last question is how to freeze the weights of certain layers, you know when using the transfer learning technology, sometimes we need to freeze some model weights.
Anyone who answers the questions will be appreciated.
Thanks in advance!
Thanks for the wonderful resource. I'm having a few issues with Sklearn models. Most of them work fine but I get the following error only with GaussianProcessClassifier, KNeighborsClassifier, MLPClassifier and QuadraticDiscriminantAnalysis (i.e., RandomForestClassifier works fine and gives the expected output, for example)
TypeError Traceback (most recent call last)
/tmp/ipykernel_29886/3380587234.py in <cell line: 2>()
1 model = dc.models.SklearnModel(GaussianProcessClassifier())
----> 2 model.fit(train_dataset)
3 model.predict(test_dataset)
~/anaconda3/envs/tensorflow2_p38/lib/python3.8/site-packages/deepchem/models/sklearn_models/sklearn_model.py in fit(self, dataset)
109 # Some scikit-learn models don't use weights.
110 if self.use_weights:
--> 111 self.model.fit(X, y, w)
112 return
113 self.model.fit(X, y)
TypeError: fit() takes 3 positional arguments but 4 were givenAny ideas what's going wrong here?
3 replies
from tokenizers import Encoding
Does anyone here have any tips?
