Loss drops to 0 every 50 epochs on TOX21

Hello,

I’m currently using deepChem 2.3.0.

I have troubles when attempting to train on the TOX21 dataset. I tried multiple models (currently GraphConv, Weaver and MPNN), and for each one, my loss drop to 0 every 50 epochs.

Any idea what could cause this ?

Here is a little code to reproduce the problem with GraphConvModel.

import numpy as np
import deepchem as dc
from deepchem.molnet import load_tox21
from deepchem.models.graph_models import GraphConvModel
import matplotlib.pyplot as plt

model_dir = "/tmp/graph_conv"

tox21_tasks, tox21_datasets, transformers = load_tox21(featurizer='GraphConv')
train_dataset, valid_dataset, test_dataset = tox21_datasets

metric = dc.metrics.Metric(
    dc.metrics.roc_auc_score, np.mean, mode="classification")

batch_size = 50

model = GraphConvModel(
    len(tox21_tasks), batch_size=batch_size, mode='classification')

losses = []

for i in range(105):
    loss = model.fit(train_dataset, nb_epoch=1)
    losses.append(loss)

print(losses)
plt.plot(losses)
plt.show()

[chstem]

You were right, somehow I didn’t have the deepchem version pip was claiming I have. Now I can confirm the issues is fixed.

[rbharath]

Awesome 😃. Going to go ahead and close off this issue. Will reopen in case there are any more reports of this still being broken