Why does the first epoch take way longer than the following epochs?

[easypickings]

I ran your example with the sample config in README (except I do not add memory limits; from other issues apparently 8gb is not enough to prevent the program getting killed). The epoch time of the first epoch is twice as those of the rest epochs. How is that happening?

My result:

Running Epoch 0...
Epoch 00, Loss: 1.8456, Approx. Train: 0.4851
Epoch time: 948457.4916 ms

==============================
Running Epoch 1...
Epoch 01, Loss: 1.6139, Approx. Train: 0.5369
Epoch time: 487353.0626 ms

==============================
Running Epoch 2...
Epoch 02, Loss: 1.5796, Approx. Train: 0.5456
Epoch time: 460697.0980 ms

==============================
Running Epoch 3...
Epoch 03, Loss: 1.5638, Approx. Train: 0.5505
Epoch time: 448294.0218 ms

==============================
Running Epoch 4...
Epoch 04, Loss: 1.5548, Approx. Train: 0.5532
Epoch time: 441110.5313 ms

==============================
Running Epoch 5...
Epoch 05, Loss: 1.5468, Approx. Train: 0.5557
Epoch time: 441491.4703 ms

==============================
Running Epoch 6...
Epoch 06, Loss: 1.5430, Approx. Train: 0.5565
Epoch time: 435827.5142 ms

==============================
Running Epoch 7...
Epoch 07, Loss: 1.5392, Approx. Train: 0.5582
Epoch time: 434762.4381 ms

==============================
Running Epoch 8...
Epoch 08, Loss: 1.5366, Approx. Train: 0.5586
Epoch time: 434168.7288 ms

==============================
Running Epoch 9...
Epoch 09, Loss: 1.5342, Approx. Train: 0.5603
Epoch time: 432678.5955 ms

Also, the epoch time seems to be decreasing by epochs (I observed it in all my three runs). Is there any explanation for this too?