GPflow: Predict call execution time degrades linearly with additional calls

• OS Platform and Distribution: Linux Ubuntu 17.04
• TensorFlow installed from (source or binary): binary
• Tensorflow version: tensorflow-gpu==1.4.0
• GPflow version: gpflow-1.0.0 (from github master source)
• CUDNN installed: 6
• GPU: GeForce GTX 1080 Ti
• Python Version: Python 3.5.3

Issue: Repeated calls to .predict (either .predict_f or .predict_y) have degrading execution time when called successively. The initial call might take perhaps 10ms, but later calls might take as long as several seconds each. This poses a problem when many calls to *.predict are desired. I did not observe this behaviour in versions of gpflow prior to 1.0.0. This occurs for at least the regression models (GPR, SGPR, SVGP), and possibly others.

import numpy as np
import matplotlib.pyplot as plt
import gpflow
import time

X = np.linspace(0,10, 100).reshape(-1,1)
y = np.sin(X) + np.random.normal(0, 0.25, size=100).reshape(-1,1)

k = gpflow.kernels.RBF(input_dim=1)
m = gpflow.models.gpr.GPR(X, y, kern=k)
m.compile()
gpflow.train.ScipyOptimizer().minimize(m)

print('Done optimize.')
predict_duration_log = []
for i in range(250):
    start = time.time()
    m.predict_y(X) 
    end = time.time()
    predict_duration_log.append(end-start)

plt.plot(predict_duration_log)
plt.show()