Speedup for long benchmarks iteration

[lrzpellegrini]

A simple fix to speedup benchmark creation and stream iteration.

The impact is greatly noticeable in long benchmarks (such as SplitImagenet, n_experiences=1000), where the iteration time on the training stream drops from 1h to 1m. I also tried other optimizations regarding FlatData, but this seems to be the most straightforward solution.

[coveralls]

Pull Request Test Coverage Report for Build 5601711922

• 2 of 2 (100.0%) changed or added relevant lines in 1 file are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at 72.765%

---

Totals
Change from base Build 5601076566:	0.0%
Covered Lines:	16709
Relevant Lines:	22963

---

💛 - Coveralls

[AntonioCarta]

Thanks, this is great! does it impacts also smaller benchmarks such as CIFAR100?

[lrzpellegrini]

For SplitCIFAR100, n_exps=100, the following loop goes from 6.22 to 1.50 seconds (Ryzen 7 3700X):

start_time = time.time()
benchmark_instance = SplitCIFAR100(100)
for exp in benchmark_instance.train_stream:
    dataset = exp.dataset
end_time = time.time()
print(end_time-start_time)