How to solve this internal error problem？

[BAIGE666]

When I set reverse_derection=True, the code will error, but when I cancel it, the code will not error. Is this a bug?

[faycalaa]

Hi @BAIGE666,

I think I have identified the origin of this bug, but to be sure, could you share a minimal code snippet that reproduces the issue?

[BAIGE666]

generate_cfr.zip
This is a Jupyter notebook file, hoping to help you fix this bug.

[faycalaa]

Unfortunately, we cannot download and uncompress zip files.

Could you please provide a minimal reproducer inlined in the GitHub message? It could use one of the built-in scenes from Sionna.

[BAIGE666]

import math
import os
import drjit as dr
import mitsuba as mi
import sionna
import numpy as np
import tensorflow as tf
import matplotlib.pyplot as plt
from copy import deepcopy
from sionna.rt import load_scene, PlanarArray, Transmitter, Receiver, Camera,\
                      PathSolver, RadioMapSolver, subcarrier_frequencies
from sionna.phy import config
config.seed = 123
from sionna.rt.utils import r_hat, subcarrier_frequencies
#假设中心点位于[4.6,-31],且三个基站为一个等边三角形
a = 100
height = 30
middle_point = [4.6,-31,0]
tx1_position = mi.Point3f([middle_point[0],middle_point[1]+a/math.sqrt(3),height])
tx2_position = mi.Point3f([middle_point[0]+a/2,middle_point[1]-a/(2*math.sqrt(3)),height])
tx3_position = mi.Point3f([middle_point[0]-a/2,middle_point[1]-a/(2*math.sqrt(3)),height])
Batch_size = 64

M = 124
N = 12
delta_f = 30e+3
c0=3e8
T = 1/delta_f
fc=2.4e+9
lambda_wave=c0/fc

# Configure which GPU
if os.getenv("CUDA_VISIBLE_DEVICES") is None:
    gpu_num = 0 # Use "" to use the CPU
    os.environ["CUDA_VISIBLE_DEVICES"] = f"{gpu_num}"
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'

gpus = tf.config.list_physical_devices('GPU')
if gpus:
    for i in range(len(gpus)):
        try:
            tf.config.experimental.set_memory_growth(gpus[i], True)
        except RuntimeError as e:
            print(e)
tf.get_logger().setLevel('ERROR')
def position_generate(num):
    random_points = []  # 用于存储随机点的列表

    for _ in range(num):
        # 随机生成两个重心坐标u, v
        u, v = np.random.rand(2)  # 生成两个随机数
        w = 1 - u - v  # 根据u和v计算w

        # 确保所有权重非负
        if w < 0:
            u, v = 1 - u, 1 - v  # 调整u和v
            w = 1 - u - v  # 重新计算w

        # 计算随机点的坐标
        point = u * tx1_position + v * tx2_position + w * tx3_position
        random_points.append(point)

    # 将结果转换为numpy.array
    pos_ue = np.array(random_points)
    return pos_ue

def generate_channel_impulse_responses(scene, map_name, num_cirs, batch_size_cir, num_tx_ant, num_rx_ant):
    max_depth = 5 #Sionna的计算深度
    # current_num = 0
    #选择不同的场景配置不同的tx和rx，但是这里的tx和rx都是用来演示使用，之后会重新生成
    scene.synthetic_array = False  # Emulate multiple antennas to reduce ray tracing complexity
    solver = PathSolver()
    scene.tx_array = PlanarArray(num_rows=1,
                                 num_cols=num_tx_ant,
                                 # We want to transmitter to be equiped with the 16 rx antennas
                                 vertical_spacing=0.5,
                                 horizontal_spacing=0.5,
                                 pattern="iso",
                                 polarization="V")
    # 创建三个基站，基站为接收端
    tx1 = Transmitter(name="tx1",
                     position=tx1_position,
                     look_at=middle_point)  # optional, defines view direction
    tx2 = Transmitter(name="tx2",
                     position=tx2_position,
                     look_at=middle_point)  # optional, defines view direction
    tx3 = Transmitter(name="tx3",
                     position=tx3_position,
                     look_at=middle_point)  # optional, defines view direction
    scene.add(tx1)
    scene.add(tx2)
    scene.add(tx3)

    # while current_num < num_cirs:


    scene.rx_array = PlanarArray(num_rows=1,
                                 num_cols=num_rx_ant,  # We want to transmitter to be equiped with the 16 rx antennas
                                 vertical_spacing=0.5,
                                 horizontal_spacing=0.5,
                                 pattern="iso",
                                 polarization="V")  # Single antenna

    print('Generating batches of CIRs (a, tau): ...')
    h = None
    num_runs = int(np.ceil(num_cirs / batch_size_cir))
    pos_ue = position_generate(batch_size_cir)
    for idx in range(batch_size_cir):
        scene.add(Receiver(f"rx-{idx}", mi.Point3f([pos_ue[idx][0],pos_ue[idx][1],1.5])))
    paths = solver(
        scene,
        max_depth=max_depth,
        specular_reflection = True,
        los = True,
        diffuse_reflection = True
    )  # shared between all tx in a scene
    frequencies = subcarrier_frequencies(M, delta_f)
    # print(frequencies)
    h_ = paths.cfr(frequencies=frequencies,
                  sampling_frequency=1 / T,
                  num_time_steps=N,
                  normalize_delays=False,
                  normalize=True,
                  out_type="numpy")
    
    # loop for creating random batch of a and tau
    for idx in range(num_runs):
        print('Progress: {}/{}'.format(idx, num_runs), end='\r')
        # Sample random user positions
        pos_ue = position_generate(batch_size_cir)

        # Update all receiver positions
        for idx in range(batch_size_cir):
            scene.receivers[f"rx-{idx}"].position = mi.Point3f([pos_ue[idx][0],pos_ue[idx][1],1.5])

        # Simulate CIR
        paths = solver(
            scene,
            max_depth=max_depth,
            specular_reflection = True,
            los = True,
            diffuse_reflection = True
        )  # shared between all tx in a scene
        # Transform paths into channel impulse responses
        frequencies = subcarrier_frequencies(M, delta_f)
        print(frequencies)
        h_ = paths.cfr(frequencies=frequencies,
                      sampling_frequency=1 / T,
                      num_time_steps=N,
                      normalize_delays=False,
                      normalize=True,
                      out_type="numpy",
                      reverse_direction = True)

        # We fix here the maximum number of paths to 75 which ensures
        # that we can simply concatenate different channel impulse reponses
        del paths  # Free memory

        if h is None:
            h = h_.numpy()
        else:
            # Concatenate along the num_tx dimension
            h = np.concatenate([h, h_], axis=2)

        # Exchange the num_tx and batchsize dimensions
    h = np.transpose(h, [2, 0, 1, 3, 4, 5])  # [3, 1, 2, 0, 4, 5, 6]

    # Remove CIRs that have no active link (i.e., a is all-zero)
    p_link = np.sum(np.abs(h) ** 2, axis=(1, 2, 3, 4, 5))
    h = h[p_link > 0., ...]

    # Remove CIRs that have invalid shapes
    h_list = [h[i] for i in range(h.shape[0])]

    # Filter out elements with invalid shape
    valid_a_list = [arr for arr in h_list if len(arr.shape) == 5]

    # Convert back to a numpy array if needed
    valid_a_array = np.array(valid_a_list)

    np.save('data/cfrdata/a_dataset_{}.npy'.format(map_name), valid_a_array)

    return h
scene = load_scene(sionna.rt.scene.florence)
generate_channel_impulse_responses(scene, "florence", 6000, 100, 16, 1)

[faycalaa]

Thank you for reporting this bug. The fix will be available in the next release.

If you want to fix it locally, simply change this line to:
num_rx, num_tx, num_paths = tau_.shape