Accuracy using pretrained weights is 0.69

[zhanxw]

The frozen_embeddings_classification.ipynb reported that the accuracy is 0.86.
I rerun this notebook on two servers, and both have accuracy around 0.69. (notebook attached).

Has anyone else noticed a similar decrease in accuracy?

replicate.ipynb.zip

[mtinti]

if I use:
test_df = pd.read_csv('../../data/bace/test.csv')
insted of:
test_df = pd.read_csv('../../data/bace/test.csv').sample(frac=0.1)

I get ROC 0.8374908290535583