es.tell() Takes Forever to Run and Returns "Process finished with exit code 137 (interrupted by signal 9:SIGKILL)"

[deringezgin]

Hi everyone,

I am trying to optimize the weights of an LSTM using the CMA-ES. In my current code, I create the LSTM model, initialize random weights, and create the CMA-ES model.

Following this, I ask for solutions from the CMA-ES, and I get a fitness value for each solution. When I have all the possible solutions, I update the "cma.CMAEvolutionStrategy" object using tell.

During this process, the program uses excessive memory, around 80 GB. Moreover, when I come to the es.tell part, the program takes forever to respond and returns the exit code 137 error in the title.

This is a pseudo-code of what I am doing:

model = LSTM(
        input_size=INPUT_SIZE,
        hidden_size=128,
        output_size=OUTPUT_SIZE,
        num_lstm_layers=1,
        num_fc_layers=3,
        fc_hidden_size=64
    )

start_weights = model.get_weights()
es = cma.CMAEvolutionStrategy(start_weights, sigma)
for i in range(100):
       gen_fitness = []
       solutions = es.ask()
       for solution in solutions:
                 gen_fitness.append(get_fitness(solution))
       es.tell(solutions, gen_fitness)

I hope that this is enough information to explain the problem, and I hope that you can help me with it. My program crashes in the first iteration of es.tell(), so this is not a memory piling-up issue.

[nikohansen]

What is the size of start_weights?

[deringezgin]

Hi, thank you so much for your reply.

In the current configuration, the length of the LSTM start_weights is 102470.

I hope that this information helps. Have a nice day!

[nikohansen]

Looks like you are running out of memory which is quadratic in the above number (hence, if I am not mistaken, in the order of a few 100GB).

I suggest to replace

es = cma.CMAEvolutionStrategy(start_weights, sigma)

with

es = cma.CMAEvolutionStrategy(start_weights, sigma, {'CMA_diagonal': True})

and see how it goes.

[deringezgin]

Thank you for your answer.

I tried this and can see that I can run with these configurations.

What would be the downsides of optimizing such a model using a diagonal covariance matrix? I hope that you can clarify this for me. I would like to know what changes.

Thank you so much for your help.

[nikohansen]

The diagonal model is the middle in Figure 1 in the tutorial.

[nikohansen]

Otherwise, these sampler are alternatives for high dimension which do learn correlations too.

[deringezgin]

Thank you for your responses. I will check them out!