A working example using differentiable ray tracing.

[salbert83]

I was going through the tutorials and didn't find any examples differentiating through ray tracing (e.g., differentiating the SINR from the RadioMapSolver). Further, the notebooks in https://github.com/NVlabs/diff-rt/tree/main have deprecated functionality. Could someone point me to some sample code?

[salbert83]

I'm including some code which demonstrates the issue.

gradient_question.ipynb

[faycalaa]

Hi @salbert83 ,

Here is a working example demonstrating how to set up the optimizer and compute the gradients.
You can also have check out this developer guide.

map_size = np.array([600,600])
cell_size = np.array([5,5])

scene = load_scene()
scene.tx_array = PlanarArray(num_rows=1, num_cols=1, pattern="tr38901", polarization="V")
tx = Transmitter(name="tx", position=[0,0,5], orientation=[0,0,0])
scene.add(tx)

# Adam optimizer
opt = mi.ad.Adam(lr=1e-2)

opt['tx.orientation'] = mi.Point3f(0,0,0)
tx.orientation = opt['tx.orientation']

solver = RadioMapSolver()

# Switch the computation of field loop to "evaluated" mode to
# enable gradient backpropagation through the loop
solver.loop_mode = "evaluated"

rm = solver(scene, cell_size=cell_size, center=[0,0,0], size=map_size, orientation=[0,0,0])


loss = dr.sum(rm.path_gain)
dr.backward(loss) # Compute the gradients

print(loss)
print("Gradient: ", opt['tx.orientation'].grad)

Hope this helps.

[salbert83]

This was perfect. Thank you so much!

[nada-ak]

Hi @faycalaa! Thank you for providing this answer. I have been encountering an issue when adopting this approach whenever the number of transmitters is greater than one. I get this error RuntimeError: CustomOp::backward(): the reverse-mode derivative of a complex loop (with max_iterations != -1) is not yet implemented!

Here is a minimal code to reproduce the error:

scene = load_scene(sionna.rt.scene.etoile)

# Rectangle defining the target area
target_center = np.array([-220.,0.,1.5]) # Center
target_size = np.array([100.,50.]) # Size
target_orientation = np.array([0.1,0.1,0.1]) # Orientation: parallel to XY
cm_cell_size = np.array([2.,2.]) # Each cell is 2mx2m


scene.tx_array = PlanarArray(num_rows=1, num_cols=16,
                             vertical_spacing=0.5,
                             horizontal_spacing=0.5,
                             pattern="tr38901",
                             polarization="V",
                            )

tx = Transmitter(name = "tx", 
                 position = [-232, 130, 25.0],
                 orientation = [0.0,0.0,0.0]
                ) 
scene.add(tx)


tx_2 = Transmitter(name = "tx_2",
                   position = [48.149, 33.176, 25.0],
                   orientation = [0, 0, 0]
                  )
scene.add(tx_2)



# RMSProp optimizer
opt = mi.ad.RMSProp(lr=0.1) 

opt['tx.orientation'] = mi.Point3f(0,0,0)
tx.orientation = opt['tx.orientation']

rm_solver = RadioMapSolver()

# Switch the computation of field loop to "evaluated" mode to
# enable gradient backpropagation through the loop
rm_solver.loop_mode = "evaluated"

rm = rm_solver(scene = scene,
               max_depth = 3,
               cell_size = cm_cell_size,
               samples_per_tx = 10**6,
               center = target_center,
               orientation = target_orientation,
               size = target_size,
               seed = 1
              )

loss = dr.mean(dr.log(1 + rm.transmitter_radio_map("sinr")))
dr.backward(loss) # Compute the gradients

[faycalaa]

Thanks for sharing the code snippet, I was able to reproduce the issue. It appears to be related to the call to rm.transmitter_radio_map(). As a workaround, you can use rm.sinr instead:

loss = dr.mean(dr.log(1 + rm.sinr))

[nada-ak]

Hi @faycalaa, Thank you for your answer. It is true that rm.sinr could provide a workaround. However, when I try to do dr.max(rm.sinr, axis = 0) I still get the same error.