AttributeError: 'float' object has no attribute 'dtype' when using extra arguments in jacrev()

autograd
1

Hi, I’m getting an error due to passing an argument (I think its the float k*r_trans) when using extra arguments to the function jacrev(). I tried declaring this as a tensor and then converting it back to a float inside the function pressureSpatialCompRealFunc but then it says the new variable doesnt have storage. Is there a way of passing this variable? The code is below.

Also, when using jacfwd() on the output of jacrev(), do I need the extra parameters?

Thank you,
Alex

import numpy as np
from scipy.special import jv
import torch

import time


#Autograd class declaration of function J_v(sqrt(x)) for pytorch
#This implementation does not have the machine precision error due to /sqrt(x) when x is or is close to 0
class modJv(torch.autograd.Function):
    @staticmethod
    def forward(ctx, x, v, a):
        ctx.save_for_backward(x)
        ctx._v = v
        ctx._a = a
        return torch.from_numpy(jv(v, np.array(a*np.sqrt(x.detach().numpy()))))

    @staticmethod
    def backward(ctx, grad_out):
        x, = ctx.saved_tensors
        v = ctx._v
        a = ctx._a
        
        Jv   = modJv.apply(x,   v,a)
        Jvm2 = modJv.apply(x, v-2,a)
        Jv2  = modJv.apply(x, v+2,a)
        return grad_out*a*a/8.0*((Jv+Jvm2)/(v-1) - (Jv2+Jv)/(v+1)), None, None


#Function to compute parts of sound field that are spatialy dependent and don't have complex exponents
def pressureSpatialCompFunc(point, transLocation,transNormal, kr_trans):
    dirV = 1 #Nu parameter for Bessel function in original form of transducer field directivity function
    
    distVec   = point-transLocation
    dist2     = distVec.pow(2).sum().unsqueeze(dim=0)
    sinTheta2 = torch.cross(distVec,transNormal).pow(2).sum().unsqueeze(dim=0)/dist2
    
    directivity = 0.25*(modJv.apply(sinTheta2,dirV+1,kr_trans) + modJv.apply(sinTheta2,dirV-1,kr_trans))
    
    dist = dist2.sqrt()
    return directivity/dist, dist



#All values in SI units
c_m     = 332.0  #Speed of sound in air (m/s)
f_trans = 40.0e3 #Transducer frequency (40 kHz)

k = 2*np.pi*f_trans/c_m #Sound wavenumber

r_trans = 5.0e-3 #Transducer radius (5 mm)


#Transducer location
transLocation = torch.tensor([4.0e-3, 2.0e-3,  1.0e-3], dtype=torch.double, requires_grad=False)
#Transducer unit normal vector
transNormal   = torch.tensor([0.0,    0.0,     1.0   ], dtype=torch.double, requires_grad=False)
#Point where sound field is measured
point         = torch.tensor([4.0e-3, 2.0e-3, 20.0e-3], dtype=torch.double, requires_grad=True)


#Compute real and imaginary parts of spatialy depedent component of transducer sound field
def pressureSpatialCompRealFunc(point, transLocation,transNormal, kr_trans):
    p_spatial_comp, tpDist = pressureSpatialCompFunc(point, transLocation,transNormal, kr_trans)
    p_spatial_comp_real = p_spatial_comp*torch.sin(k*tpDist)


start_time = time.time()
#p_spatial_comp, tpDist = pressureSpatialCompFunc(point, transLocation,transNormal, kr_trans)
#p_spatial_comp_real = p_spatial_comp*torch.sin(k*tpDist)
#jac_f = torch.autograd.grad(p_spatial_comp_real, point, create_graph=True)
jac_f = torch.func.jacrev(pressureSpatialCompRealFunc,argnums=3)(point, transLocation,transNormal,k*r_trans)
jac_time   = time.time() - start_time
print("--- %s seconds ---" % jac_time)

start_time = time.time()
hess_f     = torch.func.jacfwd(jac_f,argnums=0)(point)
hess_time  = time.time() - start_time
print("--- %s seconds ---" % hess_time)
print("--- %s seconds ---" % (jac_time + hess_time))


print(jac_f)
print(hess_f)
2

That’s strange, as I see this error:

RuntimeError: In order to use an autograd.Function with functorch transforms (vmap, grad, jvp, jacrev, ...), it must override the setup_context staticmethod. For more details, please see https://pytorch.org/docs/master/notes/extending.func.html

after passing torch.tensor(k*r_trans) to the call, which points to these docs.
Are you seeing an internal error instead?

3

Hi ptrblock, when I try passing the variable like you did, as shown below, I get the following error:
RuntimeError: Cannot access data pointer of Tensor that doesn't have storage

def pressureSpatialCompRealFunc(point, transLocation,transNormal, kr_trans):
    kr_trans = kr_trans.detach().numpy()
    p_spatial_comp, tpDist = pressureSpatialCompFunc(point, transLocation,transNormal, kr_trans)
    p_spatial_comp_real = p_spatial_comp*torch.sin(k*tpDist)

kr_trans = torch.tensor(k*r_trans)
start_time = time.time()
jac_f = torch.func.jacrev(pressureSpatialCompRealFunc,argnums=3)(point, transLocation,transNormal,kr_trans)
jac_time   = time.time() - start_time
print("--- %s seconds ---" % jac_time)
4

You can’t call .numpy() on tensors (outside of forward of custom autograd Function) when performing using torch.func transforms like jacrev. If you cannot avoid that numpy call should move it to be done inside custom Function.

(this is expected, but the error definitely needs to be improved)

5

Actually I’ve been using argnums wrong; from the documentation in jacrev, the code below should work. I want to do the jacobian w.r.t point, and I want the function value as well, so I’m using has_aux=True. However when I run this now I get this error: AttributeError: 'float' object has no attribute 'dtype'

c_m     = 332.0
f_trans = 40.0e3
k       = 2*np.pi*f_trans/c_m 
r_trans = 5.0e-3

transLocation = torch.tensor([4.0e-3, 2.0e-3,  1.0e-3], dtype=torch.double, requires_grad=False)
transNormal   = torch.tensor([0.0,    0.0,     1.0   ], dtype=torch.double, requires_grad=False)
point         = torch.tensor([4.0e-3, 2.0e-3, 20.0e-3], dtype=torch.double, requires_grad=True)

def pressureSpatialCompRealFunc(point, transLocation, transNormal, kr_trans):
    p_spatial_comp, tpDist = pressureSpatialCompFunc(point, transLocation, transNormal, kr_trans)
    p_spatial_comp_real = p_spatial_comp*torch.sin(k*tpDist)

jac_f, f_x = torch.func.jacrev(pressureSpatialCompRealFunc, argnums=0, has_aux=True) \
(point, transLocation, transNormal, k*r_trans)