MemoryError when Increasing Number of Variables

[LucasBoTang]

Description:

I encountered a MemoryError while solving a problem that involves a large number of decision variables and parameters. The error appears when the problem size exceeds 16 decision variables and parameters, which should be a resonable size. However, the code functions correctly for smaller problem sizes, such as 15 or fewer variables and parameters.

Code:

It is a multiple knapsack problem with parametric objective coefficients. You can find code within the link or blow.

"""
Parametric Multi-Dimensional Knapsack
"""

import numpy as np
import neuromancer as nm

def knapsack(var_keys, param_keys, weights, caps, penalty_weight=100):
    # mutable parameters
    params = {}
    for p in param_keys:
        params[p] = nm.constraint.variable(p)
    # decision variables
    vars = {}
    for v in var_keys:
        vars[v] = nm.constraint.variable(v)
    obj = [get_obj(vars, params)]
    constrs = get_constrs(vars, weights, caps, penalty_weight)
    return obj, constrs

def get_obj(vars, params):
    """
    Get neuroMANCER objective component
    """
    # get decision variables
    x, = vars.values()
    # get mutable parameters
    c = params.values()
    # objective function c^T x
    f = sum(- ci * xi for ci, xi in zip(c, x))
    obj = f.minimize(weight=1.0, name="obj")
    return obj

def get_constrs(vars, weights, caps, penalty_weight):
    """
    Get neuroMANCER constraint component
    """
    # problem size
    dim, num_var = weights.shape
    # get decision variables
    x, = vars.values()
    # constraints
    constraints = []
    for i in range(dim):
        g = sum(weights[i, j] * x[:, j] for j in range(num_var))
        con = penalty_weight * (g <= caps[i])
        con.name = "cap_{}".format(i)
        constraints.append(con)
    return constraints

if __name__ == "__main__":

    import torch
    from torch import nn

    # random seed
    np.random.seed(42)
    torch.manual_seed(42)

    # init
    num_var = 32      # number of variables
    #num_var = 15      # number of variables
    dim = 2           # dimension of constraints
    caps = [20] * dim # capacity
    num_data = 5000   # number of data
    test_size = 1000  # number of test size
    val_size = 1000   # number of validation size

    # data sample from PyEPO
    import pyepo
    # generate data
    weights, x, c = pyepo.data.knapsack.genData(num_data=num_data, num_features=5,
                                                num_items=num_var, dim=dim,
                                                deg=4, noise_width=0.5)
    c_samples = torch.FloatTensor(c)
    data = {"c":c_samples}
    # data split
    from src.utlis import data_split
    data_train, data_test, data_dev = data_split(data, test_size=test_size, val_size=val_size)
    # torch dataloaders
    from torch.utils.data import DataLoader
    loader_train = DataLoader(data_train, batch_size=32, num_workers=0,
                              collate_fn=data_train.collate_fn, shuffle=True)
    loader_test  = DataLoader(data_test, batch_size=32, num_workers=0,
                              collate_fn=data_test.collate_fn, shuffle=True)
    loader_dev   = DataLoader(data_dev, batch_size=32, num_workers=0,
                              collate_fn=data_dev.collate_fn, shuffle=True)

    # get objective function & constraints
    obj, constrs = knapsack(["x"], ["c"], weights=weights, caps=caps, penalty_weight=100)

    # define neural architecture for the solution map smap(c) -> x
    import neuromancer as nm
    func = nm.modules.blocks.MLP(insize=num_var, outsize=num_var, bias=True,
                                 linear_map=nm.slim.maps["linear"],
                                 nonlin=nn.ReLU, hsizes=[64]*2)
    components = [nm.system.Node(func, ["c"], ["x"], name="smap")]

    # build neuromancer problems
    loss = nm.loss.PenaltyLoss(obj, constrs)
    problem = nm.problem.Problem(components, loss)

    # training
    lr = 0.001    # step size for gradient descent
    epochs = 400  # number of training epochs
    warmup = 40   # number of epochs to wait before enacting early stopping policy
    patience = 40 # number of epochs with no improvement in eval metric to allow before early stopping
    # set adamW as optimizer
    optimizer = torch.optim.AdamW(problem.parameters(), lr=lr)
    # define trainer
    trainer = nm.trainer.Trainer(
        problem,
        loader_train,
        loader_dev,
        loader_test,
        optimizer,
        epochs=epochs,
        patience=patience,
        warmup=warmup)
    # train solution map
    best_model = trainer.train()
    # load best model dict
    problem.load_state_dict(best_model)

Error Message:

Traceback (most recent call last):
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\runpy.py", line 196, in _run_module_as_main
    return _run_code(code, main_globals, None,
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\runpy.py", line 86, in _run_code
    exec(code, run_globals)
  File "C:\OneDrive\Study\UofT\Research\DMIPL\src\problem\neuromancer\knapsack.py", line 116, in <module>
    trainer = nm.trainer.Trainer(
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\site-packages\neuromancer\trainer.py", line 227, in __init__
    self.best_model = deepcopy(self.model.state_dict())
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\site-packages\torch\nn\modules\module.py", line 1916, in state_dict
    module.state_dict(destination=destination, prefix=prefix + name + '.', keep_vars=keep_vars)
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\site-packages\torch\nn\modules\module.py", line 1916, in state_dict
    module.state_dict(destination=destination, prefix=prefix + name + '.', keep_vars=keep_vars)
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\site-packages\torch\nn\modules\module.py", line 1916, in state_dict
    module.state_dict(destination=destination, prefix=prefix + name + '.', keep_vars=keep_vars)
  [Previous line repeated 30 more times]
  File "C:\Users\lucas\anaconda3\envs\neuromancer\lib\site-packages\torch\nn\modules\module.py", line 1909, in state_dict
    destination._metadata[prefix[:-1]] = local_metadata
MemoryError

Environment:

• OS: Windows 11
• Python version: 3.10.13
• PyTorch version: 2.3.0
• NeuroMANCER version: 1.5.1

[drgona]

@LucasBoTang I can confirm that this code crashed when running in Colab. session crashed after using all available RAM.

[LucasBoTang]

When increasing the dimensions in a parametric Rosenbrock problem (with 32 decision variables and 17 parameters), the code runs for a very long time without producing any output, suggesting similar issues in handling higher dimensions.

Code for Rosenbrock Problem:

import numpy as np
import neuromancer as nm

def rosenbrock(var_keys, param_keys, steepness, num_blocks, penalty_weight=50):
    # mutable parameters
    params = {}
    for p in param_keys:
        params[p] = nm.constraint.variable(p)
    # decision variables
    vars = {}
    for v in var_keys:
        vars[v] = nm.constraint.variable(v)
    obj = [get_obj(vars, params, steepness, num_blocks)]
    constrs = get_constrs(vars, params, num_blocks, penalty_weight)
    return obj, constrs

def get_obj(vars, params, steepness, num_blocks):
    """
    Get neuroMANCER objective component
    """
    # get decision variables
    x, = vars.values()
    # get mutable parameters
    p, a = params.values()
    # objective function sum (a_i - x_2i)^2 + b * (x_2i+1 - x_2i^2)^2
    f = sum((a[:, i] - x[:, 2*i]) ** 2 + steepness * (x[:, 2*i+1] - x[:, 2*i] ** 2) ** 2
             for i in range(num_blocks))
    obj = f.minimize(weight=1.0, name="obj")
    return obj

def get_constrs(vars, params, num_blocks, penalty_weight):
    """
    Get neuroMANCER constraint component
    """
    # get decision variables
    x, = vars.values()
    # get mutable parameters
    p, a = params.values()
    # constraints
    constraints = []
    # inner
    g = sum(x[:, 2*i+1] for i in range(num_blocks))
    con = penalty_weight * (g >= num_blocks * p[:, 0] / 2)
    con.name = "c_inner"
    constraints.append(con)
    # outer
    g = sum(x[:, 2*i] ** 2 for i in range(num_blocks))
    con = penalty_weight * (g <= num_blocks * p[:, 0])
    con.name = "c_outer"
    constraints.append(con)
    return constraints

if __name__ == "__main__":

    import torch
    from torch import nn

    # random seed
    np.random.seed(42)
    torch.manual_seed(42)

    # init
    steepness = 30    # steepness factor
    num_blocks = 16   # number of expression blocks
    num_data = 5000   # number of data
    test_size = 1000  # number of test size
    val_size = 1000   # number of validation size

    # data sample from uniform distribution
    p_low, p_high = 1.0, 8.0
    a_low, a_high = 0.5, 4.5
    p_samples = torch.FloatTensor(num_data, 1).uniform_(p_low, p_high)
    a_samples = torch.FloatTensor(num_data, num_blocks).uniform_(a_low, a_high)
    data = {"p":p_samples, "a":a_samples}
    # data split
    from src.utlis import data_split
    data_train, data_test, data_dev = data_split(data, test_size=test_size, val_size=val_size)
    # torch dataloaders
    from torch.utils.data import DataLoader
    loader_train = DataLoader(data_train, batch_size=32, num_workers=0,
                              collate_fn=data_train.collate_fn, shuffle=True)
    loader_test  = DataLoader(data_test, batch_size=32, num_workers=0,
                              collate_fn=data_test.collate_fn, shuffle=True)
    loader_dev   = DataLoader(data_dev, batch_size=32, num_workers=0,
                              collate_fn=data_dev.collate_fn, shuffle=True)

    # get objective function & constraints
    obj, constrs = rosenbrock(["x"], ["p", "a"], steepness=steepness,
                              num_blocks=num_blocks, penalty_weight=100)

    # define neural architecture for the solution map smap(p, a) -> x
    import neuromancer as nm
    func = nm.modules.blocks.MLP(insize=num_blocks+1, outsize=2*num_blocks, bias=True,
                                 linear_map=nm.slim.maps["linear"],
                                 nonlin=nn.ReLU, hsizes=[32]*4)
    components = [nm.system.Node(func, ["p", "a"], ["x"], name="smap")]

    # build neuromancer problems
    loss = nm.loss.PenaltyLoss(obj, constrs)
    problem = nm.problem.Problem(components, loss)

    # training
    lr = 0.001    # step size for gradient descent
    epochs = 400  # number of training epochs
    warmup = 40   # number of epochs to wait before enacting early stopping policy
    patience = 40 # number of epochs with no improvement in eval metric to allow before early stopping
    # set adamW as optimizer
    optimizer = torch.optim.AdamW(problem.parameters(), lr=lr)
    # define trainer
    trainer = nm.trainer.Trainer(
        problem,
        loader_train,
        loader_dev,
        loader_test,
        optimizer,
        epochs=epochs,
        patience=patience,
        warmup=warmup)
    # train solution map
    best_model = trainer.train()
    # load best model dict
    problem.load_state_dict(best_model)
    print()

    # init mathmatic model
    from src.problem.math_solver.rosenbrock import rosenbrock
    model = rosenbrock(steepness=steepness, num_blocks=num_blocks)

    # test neuroMANCER
    from src.utlis import nm_test_solve
    p, a = data_train[0]["p"].tolist(), data_train[0]["a"].tolist()
    datapoint = {"p": torch.tensor([p], dtype=torch.float32),
                 "a": torch.tensor([a], dtype=torch.float32),
                 "name":"test"}
    model.set_param_val({"p":p, "a":a})
    print("neuroMANCER:")
    nm_test_solve(["x"], problem, datapoint, model)