ask for help about "TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first."

[yanyuechuixue]

Hey everyone, I am a beginner of ML and I am trying to use deepxde to solve a partial differential equation (PDE) for my machine learning project.

I was able to run the code with the tensorflow backend, but when I tried running it with pytorch, I encountered an error.

I have searched through various issues and discussions related to this problem and I think this issue on GitHub (#1046) might be related to my problem, but I am not sure because I am a very beginner in machine learning and may not be able to fully understand it.

I apologize for any inconvenience I may have caused to the community by asking such a simple and specific question.

my code is

import deepxde as dde
import matplotlib.pyplot as plt
# Import tf if using backend tensorflow.compat.v1 or tensorflow
# Import torch if using backend pytorch
import torch
# Import paddle if using backend paddle
# import paddle
sqrt = torch.sqrt
device = torch.device('cuda')
def pde(x, u):
    u_xx = dde.grad.hessian(u, x, i=0, j=0)
    u_yy = dde.grad.hessian(u, x, i=1, j=1)
    u_zz = dde.grad.hessian(u, x, i=2, j=2)
    r = torch.sqrt(torch.tensor(x[:, 0:1]**2+x[:, 1:2]**2+x[:, 2:3]**2, dtype=torch.float,device = device))
    mask = r < 0.5
    rhs = torch.as_tensor(torch.where(mask, 1, 0), dtype=u_xx.dtype)
    return -(u_xx + u_yy + u_zz) - rhs

def boundary(x, on_boundary):
    return on_boundary

def func(x):
    r = torch.sqrt(torch.tensor(x[:, 0:1]**2+x[:, 1:2]**2+x[:, 2:3]**2, dtype=torch.float,device = device))
    mask = r < 0.5
    y1 = -r**2 / 6 + 1/8
    y2 = 1/(24*r)
    return torch.where(mask, y1, y2)

spatial_domain = dde.geometry.Cuboid(xmin=[-1, -1, -1], xmax=[1, 1, 1])
bc = dde.icbc.DirichletBC(spatial_domain, func, boundary)
data = dde.data.PDE(spatial_domain, pde, bc, 8192, 256,solution=func, num_test=343)
layer_size = [3] + [256] * 3 + [1] 
# activation = "tanh"
activation = "sin"
initializer = "Glorot uniform"
net = dde.nn.FNN(layer_size, activation, initializer)

model = dde.Model(data, net)
#model.compile("adam", lr=0.001,decay = ("inverse time", 1000, 0.3),metrics=["l2 relative error"])
model.compile("adam", lr=0.0001, metrics=["l2 relative error"])

resampler = dde.callbacks.PDEPointResampler(period=100)

losshistory, train_state = model.train(
    iterations=100000,callbacks=[resampler])

# losshistory, train_state = model.train(
#     iterations=10000
# )
model.save('''toymodel.h5''')

and the error is :

Using backend: pytorch

Compiling model...
'compile' took 0.000257 s

Training model...

/tmp/ipykernel_5289/81008028.py:14: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).
  r = torch.sqrt(torch.tensor(x[:, 0:1]**2+x[:, 1:2]**2+x[:, 2:3]**2, dtype=torch.float,device = device))
Output exceeds the [size limit](command:workbench.action.openSettings?[). Open the full output data [in a text editor](command:workbench.action.openLargeOutput?fffc11ab-256e-4cb5-b04c-5b42e0396df7)
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Cell In [1], line 44
     40 model.compile("adam", lr=0.0001, metrics=["l2 relative error"])
     42 resampler = dde.callbacks.PDEPointResampler(period=100)
---> 44 losshistory, train_state = model.train(
     45     iterations=100000,callbacks=[resampler])
     47 # losshistory, train_state = model.train(
     48 #     iterations=10000
     49 # )
     50 model.save('''toymodel.h5''')

File /home1/1T/zhaozc/miniconda3/envs/deepxde/lib/python3.9/site-packages/deepxde/utils/internal.py:22, in timing.<locals>.wrapper(*args, **kwargs)
     19 @wraps(f)
     20 def wrapper(*args, **kwargs):
     21     ts = timeit.default_timer()
---> 22     result = f(*args, **kwargs)
     23     te = timeit.default_timer()
     24     print("%r took %f s\n" % (f.__name__, te - ts))

File /home1/1T/zhaozc/miniconda3/envs/deepxde/lib/python3.9/site-packages/deepxde/model.py:584, in Model.train(self, iterations, batch_size, display_every, disregard_previous_best, callbacks, model_restore_path, model_save_path, epochs)
    582 self.train_state.set_data_train(*self.data.train_next_batch(self.batch_size))
    583 self.train_state.set_data_test(*self.data.test())
--> 584 self._test()
    585 self.callbacks.on_train_begin()
...
--> 955     return self.numpy()
    956 else:
    957     return self.numpy().astype(dtype, copy=False)

TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

By the way, I found the cpu utilization is 100% (but just one core) and the GPU utilization is only 7%. Is that means any error in my code?

Thanks all of you !

[elliot-hicks]

@yanyuechuixue can you try this without:

• Masking
• Setting the device explicitly when defining tensors (deepxde will run fine without these lines, I can move between CPU and GPU without adding any new code)

P.S. you dont have to apologise for asking - the problem looks cool :)