Parallel processing for testing ML model with pool.apply_async does not allow access to results

Question

I have a dataset of 2.7 million samples that I need to test my ML model on. I have 8 cores on my laptop and want to try parallelizing my testing code to save time. This is the test function :

def testMTGP(x_sample, y_sample, ind, model, likelihood):
    x_sample = x_sample.view(1, -1)
    y_sample = y_sample.view(1, -1)
    model.eval()
    likelihood.eval()
    with torch.no_grad():
        prediction = likelihood(model(x_sample))
        mean = (prediction.mean).detach().numpy()
        prewhiten_error = (y_sample.detach().numpy()) - mean
        cov_matrix = (prediction.covariance_matrix).detach().numpy()
        white_error, matcheck = Whiten(prewhiten_error, cov_matrix)
    return (
        ind,
        {
            "prediction": mean,
            "prewhiten_error": prewhiten_error,
            "white_error": white_error,
            "cov_matrix": cov_matrix,
            "matcheck": matcheck,
        },
    )

I return the index corresponding to the sample I tested and a dictionary of data related to the computations the model does for testing. The function Whiten(prewhiten_error, cov_matrix) is also defined by me and was imported at the beginning of the code file, so it is available globally. It simply takes the inputs, transforms cov_matrix and multiplies it with prewhiten_error and returns the answer, along with a variable that indicates some state information about the cov_matrix.

For the multiprocessing, the idea is to first divide the entire dataset into roughly equal sizes chunks; pick each chunk and send one sample to every core for processing. I am using pool.apply_async. This is the code:

test_X = torch.load(test_X_filename) #torch tensor of shape 2.7M x 3
test_Y = torch.load(test_Y_filename) #torch tensor of shape 2.7M x 3
cores = mp.cpu_count()
chunk_size = int(test_X.shape[0] / cores)
start_time = time.time()
parent_list = []
for start_ind in range(0, test_X.shape[0], chunk_size):
    pool = mp.Pool(processes=cores)
    proc_data_size = int(chunk_size / cores)
    stop_ind = min(test_X.shape[0], start_ind + chunk_size)
    results = [
        pool.apply_async(
            testMTGP, (test_X[i].detach(), test_Y[i].detach(), i, model, likelihood,)
        )
        for i in range(start_ind, stop_ind)
    ]
    for res in results:
        print("Length of results list= ", len(results))
        print("Data type of res is: ", type(res))
        res_dict = res.get()
        parent_list.append(res_dict)
    pool.close()

test_X[i] and test_Y[i] are both tensors with shape (3,). On executing the code I get:

Traceback (most recent call last):
File "multiproc_async.py", line 288, in
res_dict = res.get() # [1]
File "/home/aman/anaconda3/envs/thesis/lib/python3.8/multiprocessing/pool.py", line 771, in get
raise self._value
File "/home/aman/anaconda3/envs/thesis/lib/python3.8/multiprocessing/pool.py", line 537, in _handle_tasks
put(task)
File "/home/aman/anaconda3/envs/thesis/lib/python3.8/multiprocessing/connection.py", line 206, in send
self._send_bytes(_ForkingPickler.dumps(obj))
File "/home/aman/anaconda3/envs/thesis/lib/python3.8/multiprocessing/reduction.py", line 51, in dumps
cls(buf, protocol).dump(obj)
AttributeError: Can't pickle local object MultitaskGaussianLikelihood.__init__.<locals>.<lambda>

I am new to multiprocessing and googling for this error did not really help (some of it was not relevant and some was beyond my understanding). Can someone please help me understand what mistake I am making?

also, it is recommended to use a context manager and not pool, pool.close (`with multiprocessing.Pool(processes=3) as pool:`) — miquelvir, Mar 27 '21 at 22:27

score 1 · Answer 1 · answered Mar 27 '21 at 22:51

Well this issue is fairly complex and I've never used Torch, and I'm by no means an expert in multiprocessing. But I do have a decent grasp on the concepts here so I'll do my best to explain what is wrong but you will probably need to come up with the fix because it'll depend on your end goal.

Note: I notice you're just typing python. It looks like this is a Windows Store version of Ubuntu, which if that's the case you may want to run the program using python3. (If you've re-mapped the alias please ignore.)

So that final error in the stacktrace, Can't picke local object 'MultitaskGaussianLikelihood.__init__.<locals>.<lambda>'; This is refering to the library Pickle which is a serializer library. If you're unfamiliar with seralization, it's basically a standard format to rebuild something cross-system. For example, JSON is a very common serializer; it allows you to transfer multiple variables as an array across multiple programming languages. Pickle allows for the searlization of objects so they can be transfered to another program. I believe the reason res.get() is serializing here is due to the limited functionality in python with cores being able to talk to eachother which is evident throughout the multiprocessing documentation.

The problem is the Class MultitaskGaussianLikelihood appears to use a lambda as one of it's parameters, and according to that AttributeError, pickle is not capable of seralizing a lambda. Which means it can't serialize MultitaskGaussianLikelihood as it contains one. I don't have all the code here so I can't see where the MultitaskGaussianLikelihood object is in your return, but I would say you need to extract all the information you will need from that class and return that data instead of returning the class and extracting it after the fact.

Hope I explained that well!

miquelvir · Answer 2 · 2021-03-27T23:03:17.290

Let's simplify your problem to the root cause of it. We need a working example, for the multiprocessing part, otherwise we don't have a reproducible example to help you. Then you can patch in the actual training the model.

Let's use this dummy function:

def testMTGP(x_sample, y_sample, ind, model, likelihood):
    return (
        ind,
        {
            "prediction": 1,
            "prewhiten_error": 1,
            "white_error": 1,
            "cov_matrix": 1,
            "matcheck": 1,
        },
    )

Then a working and clean example is:

if __name__ == '__main__':
    cores = mp.cpu_count()
    args = [(None, None, i, None, None,) for i in range(0, 5)]

    start_time = time.time()
    with mp.Pool(processes=3) as pool:
        results = pool.starmap(testMTGP, args)
        
    end_time = time.time()
    
    print(results)
    print("it took %s" % (end_time-start_time))

Try with this and, little by little, bring in the actual logic you need for training the model. I suggest you start by passing the actual arguments you want each time, and at the end of it updating the testMTGP function (replacing the dummy one).

When you isolate what makes the code crash, and/or post the stack trace, I can help more.

Parallel processing for testing ML model with pool.apply_async does not allow access to results

2 Answers2