ValueError: Input contains NaN, infinity or a value too large for dtype('float32') when training on the 'burgers_delta' dataset

[psaegert]

Hi,

I am trying to reproduce your results using the 'burgers_delta' dataset. However, I often get the following error when the training begins:

Using device: cuda
Dataset is using device:  cuda
2021-08-17 13:22:26.965021: I tensorflow/stream_executor/platform/default/dso_loader.cc:53] Successfully opened dynamic library libcudart.so.11.0
     0  MSE:      nan  Reg: 1.52e-05  L1: 1.83e+01 Traceback (most recent call last):
  File "replicate_burgers_troubleshoot.py", line 96, in <module>
    train(
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/deepymod/training/training.py", line 100, in train
    _ = model.sparse_estimator(
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/deepymod/model/deepmod.py", line 141, in forward
    self.coeff_vectors = [
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/deepymod/model/deepmod.py", line 142, in <listcomp>
    self.fit(theta, time_deriv.squeeze())[:, None]
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/deepymod/model/sparse_estimators.py", line 77, in fit
    coeffs = self.estimator.fit(X, y).coef_
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/sklearn/linear_model/_coordinate_descent.py", line 1216, in fit
    X, y = self._validate_data(X, y,
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/sklearn/base.py", line 430, in _validate_data
    X = check_array(X, **check_X_params)
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/sklearn/utils/validation.py", line 63, in inner_f
    return f(*args, **kwargs)
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/sklearn/utils/validation.py", line 720, in check_array
    _assert_all_finite(array,
  File "/home/paulsaegert/.local/lib/python3.8/site-packages/sklearn/utils/validation.py", line 103, in _assert_all_finite
    raise ValueError(
ValueError: Input contains NaN, infinity or a value too large for dtype('float32').

This is the python script I used:

import os
import torch

from deepymod import DeepMoD
from deepymod.data import Dataset, get_train_test_loader
from deepymod.data.burgers import burgers_delta
from deepymod.data.samples import Subsample_random
from deepymod.model.func_approx import NN
from deepymod.model.library import Library1D
from deepymod.model.constraint import LeastSquares
from deepymod.model.sparse_estimators import Threshold
from deepymod.training import train
from deepymod.training.sparsity_scheduler import TrainTestPeriodic
import time
import shutil

if torch.cuda.is_available() and True:
    device = "cuda"
else:
    device = "cpu"
print(f'Using device: {device}')

temp_log_dir = 'temp_log'

if os.path.exists(temp_log_dir): shutil.rmtree(temp_log_dir)
if os.path.exists(f'{temp_log_dir}checkpoint.pt'): os.remove(f'{temp_log_dir}checkpoint.pt')
if os.path.exists(f'{temp_log_dir}model.pt'): os.remove(f'{temp_log_dir}model.pt')

result = {
    'true_coefficients': (0.1, 1.0)
}

parameters = {
    'x_min': -8,
    'x_max': 8,
    't_min': 0,
    't_max': 10,
    'n_samples': 100,
    'x_resolution': 256,
    't_resolution': 101,
    'noise_level': 0.01,
    'train_size': None,
    'network_topology': (2, [50, 50, 50, 50], 1),
    'lib_poly_order': 2,
    'lib_diff_order': 3,
    'sparse_threshold': 0.1,
    'sparsity_scheduler_periodicity': 50,
    'sparsity_scheduler_patience': 200,
    'sparsity_scheduler_delta': 1e-5,
    'write_iterations': 25,
    'max_iterations': 50_000,
    'converge_patience': 200,
    'converge_delta': 1e-4,
    'optimizer_learning_rate': 2e-3,
    'optimizer_beta1': 0.99,
    'optimizer_beta2': 0.99,
    'optimizer_amsgrad': True,
}

v, A = result['true_coefficients']

x = torch.linspace(parameters['x_min'], parameters['x_max'], parameters['x_resolution'])
t = torch.linspace(parameters['t_min'], parameters['t_max'], parameters['t_resolution'])
load_kwargs = {"x": x, "t": t, "v": v, "A": A}
preprocess_kwargs = {"noise_level": parameters['noise_level']}

dataset = Dataset(
    burgers_delta,
    load_kwargs=load_kwargs,
    preprocess_kwargs=preprocess_kwargs,
    subsampler=Subsample_random,
    subsampler_kwargs={"number_of_samples": parameters['n_samples']},
    device=device,
)

train_dataloader, test_dataloader = get_train_test_loader(
    dataset, train_test_split=0.8
)

parameters['train_size'] = len(train_dataloader[0][0])

network = NN(*parameters['network_topology'])

library = Library1D(poly_order=parameters['lib_poly_order'], diff_order=parameters['lib_diff_order'])
sparsity_estimator = Threshold(parameters['sparse_threshold']) 
sparsity_scheduler = TrainTestPeriodic(periodicity=parameters['sparsity_scheduler_periodicity'], patience=parameters['sparsity_scheduler_patience'], delta=parameters['sparsity_scheduler_delta']) 
constraint = LeastSquares()

model = DeepMoD(network, library, sparsity_estimator, constraint).to(device)
optimizer = torch.optim.Adam(model.parameters(), betas=(parameters['optimizer_beta1'], parameters['optimizer_beta2']), amsgrad=parameters['optimizer_amsgrad'], lr=parameters['optimizer_learning_rate']) 

start_time = time.time()

train(
    model,
    train_dataloader,
    test_dataloader,
    optimizer,
    sparsity_scheduler,
    exp_ID="0.0",
    log_dir=temp_log_dir,
    write_iterations=parameters['write_iterations'],
    max_iterations=parameters['max_iterations'],
    delta=parameters['converge_delta'],
    patience=parameters['converge_patience'],
)

shutil.rmtree(temp_log_dir)

It seems like setting x_min and x_max to values closer to 0 helps but I would like to specifically test -8 to 8 as in your paper.

[GJBoth]

The issue here isn't in the x, but in the t. The initial condition is a delta peak, so the solution goes to infinity at t=0 - setting t_min=0.1 or something similar should fix your issues. Let me know if it doesn't!

[psaegert]

Setting t_min=0.1 fixed the issue. Thank you very much!