3.4示例
首先，需要导入库
>>> import numpy as np
>>> import tensorflow as tf
>>> import deepchem as dc
>>>
>>> # Run before every test for reproducibility
>>> def seed_all():
... np.random.seed(123)
... tf.random.set_seed(123)
3.4.1 Delaney (ESOL)
使用MoleculeNet中的Dalaney(ESOL)数据集训练模型的示例。
我们使用smiles字段来训练模型，实验性的预测溶解能（expt).
MultitaskRegressor
首先，我们用load_delaney()方法加载数据集，并拟合MultitaskRegressor。
>>> seed_all()
>>> # Load dataset with default 'scaffold' splitting
>>> tasks, datasets, transformers = dc.molnet.load_delaney()
>>> tasks
['measured log solubility in mols per litre']
>>> train_dataset, valid_dataset, test_dataset = datasets
>>>
>>> # We want to know the pearson R squared score, averaged across tasks
>>> avg_pearson_r2 = dc.metrics.Metric(dc.metrics.pearson_r2_score, np.mean)
>>>
>>> # We'll train a multitask regressor (fully connected network)
>>> model = dc.models.MultitaskRegressor(
... len(tasks),
... n_features=1024,
... layer_sizes=[500])
>>>
>>> model.fit(train_dataset)
0...
>>>
>>> # We now evaluate our fitted model on our training and validation sets
>>> train_scores = model.evaluate(train_dataset, [avg_pearson_r2], transformers)
>>> assert train_scores['mean-pearson_r2_score'] > 0.7, train_scores
>>>
>>> valid_scores = model.evaluate(valid_dataset, [avg_pearson_r2], transformers)
>>> assert valid_scores['mean-pearson_r2_score'] > 0.3, valid_scores
GraphConvModel
Delaney黙认的特征化器是ECFP， “Extended-connectivity fingerprints”的缩写。对于GraphConvModel，我们将用featurizer='GraphConv'重载数据集。
>>> seed_all()
>>> tasks, datasets, transformers = dc.molnet.load_delaney(featurizer='GraphConv')
>>> train_dataset, valid_dataset, test_dataset = datasets
>>>
>>> model = dc.models.GraphConvModel(len(tasks), mode='regression', dropout=0.5)
>>>
>>> model.fit(train_dataset, nb_epoch=30)
0...
>>>
>>> # We now evaluate our fitted model on our training and validation sets
>>> train_scores = model.evaluate(train_dataset, [avg_pearson_r2], transformers)
>>> assert train_scores['mean-pearson_r2_score'] > 0.5, train_scores
>>>
>>> valid_scores = model.evaluate(valid_dataset, [avg_pearson_r2], transformers)
>>> assert valid_scores['mean-pearson_r2_score'] > 0.3, valid_scores
3.4.2ChEMBL
用MoleculeNet中ChEMBL数据集训练模型的例子。
ChEMBL是手工组织的有类药物特性的生物活性分子数据库。它集成化学，生物活性，基因数据以帮助将基因信息转化为有效的新药。
MultitaskRegressor
>>> seed_all()
>>> # Load ChEMBL 5thresh dataset with random splitting
>>> chembl_tasks, datasets, transformers = dc.molnet.load_chembl(
... shard_size=2000, featurizer="ECFP", set="5thresh", split="random")
>>> train_dataset, valid_dataset, test_dataset = datasets
>>> len(chembl_tasks)
691
>>> f'Compound train/valid/test split: {len(train_dataset)}/{len(valid_dataset)}/
?→{len(test_dataset)}'
'Compound train/valid/test split: 19096/2387/2388'
>>>
>>> # We want to know the RMS, averaged across tasks
>>> avg_rms = dc.metrics.Metric(dc.metrics.rms_score, np.mean)
>>>
>>> # Create our model
>>> n_layers = 3
>>> model = dc.models.MultitaskRegressor(
... len(chembl_tasks),
... n_features=1024,
... layer_sizes=[1000] * n_layers,
... dropouts=[.25] * n_layers,
... weight_init_stddevs=[.02] * n_layers,
... bias_init_consts=[1.] * n_layers,
... learning_rate=.0003,
... weight_decay_penalty=.0001,
... batch_size=100)
>>>
>>> model.fit(train_dataset, nb_epoch=5)
0...
>>>
>>> # We now evaluate our fitted model on our training and validation sets
>>> train_scores = model.evaluate(train_dataset, [avg_rms], transformers)
>>> assert train_scores['mean-rms_score'] < 10.00
>>>
>>> valid_scores = model.evaluate(valid_dataset, [avg_rms], transformers)
>>> assert valid_scores['mean-rms_score'] < 10.00
GraphConvModel
>>> # Load ChEMBL dataset
>>> chembl_tasks, datasets, transformers = dc.molnet.load_chembl(
... shard_size=2000, featurizer="GraphConv", set="5thresh", split="random")
>>> train_dataset, valid_dataset, test_dataset = datasets
>>>
>>> # RMS, averaged across tasks
>>> avg_rms = dc.metrics.Metric(dc.metrics.rms_score, np.mean)
>>>
>>> model = dc.models.GraphConvModel(
... len(chembl_tasks), batch_size=128, mode='regression')
>>>
>>> # Fit trained model
>>> model.fit(train_dataset, nb_epoch=5)
0...
>>>
>>> # We now evaluate our fitted model on our training and validation sets
>>> train_scores = model.evaluate(train_dataset, [avg_rms], transformers)
>>> assert train_scores['mean-rms_score'] < 10.00
>>>
>>> valid_scores = model.evaluate(valid_dataset, [avg_rms], transformers)
>>> assert valid_scores['mean-rms_score'] < 10.00