[Bug]: Shape mistmatch error at Paddle frontend with Scalars

[akshatvishu]

Bug Explanation

When trying to test Paddle functions with scalar inputs, a shape mismatch error is encountered due to the conversion of scalars (0-D Tensors) to NumPy arrays. This issue arises because Paddle will be supporting 0-D tensors starting from version 2.6, and converting scalars to NumPy arrays cast them automatically to 1D at present.

Also, paddle started throwing this warning since version 2.5 when we try to do this conversion:

eager_method.cc:140] Warning:: 0D Tensor cannot be used as 'Tensor.numpy()[0]' .  In order to avoid this problem, 0D Tensor will be changed to 1D numpy currently, but it's not correct and will be removed in release 2.6. For Tensor contain only one element, Please modify  'Tensor.numpy()[0]' to 'float(Tensor)' as soon as possible, otherwise 'Tensor.numpy()[0]' will raise error in release 2.6.

This probably happens because we convert the result from our current backend v/s the native function result for the same input during our value_test to numpy as part of the ivy frontend test suite for our frontend implementation:

    value_test(
        ret_np_flat=ret_np_flat,
        ret_np_from_gt_flat=frontend_ret_np_flat,
        rtol=rtol,
        atol=atol,
        backend=backend_to_test,
        ground_truth_backend=frontend,
    )

refs: https://github.com/unifyai/ivy/blob/master/ivy_tests/test_ivy/helpers/function_testing.py#L833

Steps to Reproduce Bug

import numpy as np
import paddle
import ivy
import ivy.functional.frontends.paddle as ivy_paddle

# Ivy backend "NumPy"
ivy.set_backend("numpy")
x_ivy = ivy_paddle.to_tensor([False], dtype="bool")
y_ivy = ivy_paddle.any(x_ivy, axis=0)
print(f"Ivy Output: {y_ivy}, Output Shape: {y_ivy.shape}\n") 

# Ivy backend "Paddle"
ivy.set_backend("paddle")
x_paddle_ivy = ivy_paddle.to_tensor([False], dtype="bool")
y_paddle_ivy = ivy_paddle.any(x_paddle_ivy, axis=0)
print(f"PaddlePaddle (via Ivy) Output: {y_paddle_ivy}, Output Shape: {y_paddle_ivy.shape}\n") #

# PaddlePaddle 
x_paddle = paddle.to_tensor([False], dtype=paddle.bool)
y_paddle = paddle.any(x_paddle, axis=0)
y_numpy = y_paddle.numpy()
print(f"PaddlePaddle Output: {y_paddle}, Output Shape: {y_paddle.shape}\n")
print(f"PaddleNumpy Output: {y_numpy}, Output Shape: {y_numpy.shape}\n")

# NumPy: 
x_np = np.array([False])
y_np = np.any(x_np, axis=0)
print(f"NumPy Output: {y_np}, Output Shape: {y_np.shape}\n")
"""
workaround for paddle converting 0-D tensors
to 1-D NumPy Array
"""
# PaddlePaddle 
x_paddle = paddle.to_tensor([False], dtype=paddle.bool)
y_paddle = paddle.any(x_paddle, axis=0)
y_numpy = y_paddle.numpy().reshape(())
print(f"PaddlePaddle Output: {y_paddle}, Output Shape: {y_paddle.shape}\n")
print(f"PaddleNumpy Output: {y_numpy}, Output Shape: {y_numpy.shape}\n")

# Print
"""

Ivy Output: ivy.frontends.paddle.Tensor(False), Output Shape: ivy.Shape()

PaddlePaddle (via Ivy) Output: ivy.frontends.paddle.Tensor(False), Output Shape: ivy.Shape()

I0727 18:09:04.001052 621063 eager_method.cc:140] Warning:: 0D Tensor cannot be used as 'Tensor.numpy()[0]' . In order to avoid this problem, 0D Tensor will be changed to 1D numpy currently, but it's not correct and will be removed in release 2.6. For Tensor contain only one element, Please modify  'Tensor.numpy()[0]' to 'float(Tensor)' as soon as possible, otherwise 'Tensor.numpy()[0]' will raise error in release 2.6.
PaddlePaddle Output: Tensor(shape=[], dtype=bool, place=Place(cpu), stop_gradient=True,
       False), Output Shape: []

PaddleNumpy Output: [False], Output Shape: (1,)

NumPy Output: False, Output Shape: ()

I0727 18:09:04.001331 621063 eager_method.cc:140] Warning:: 0D Tensor cannot be used as 'Tensor.numpy()[0]' . In order to avoid this problem, 0D Tensor will be changed to 1D numpy currently, but it's not correct and will be removed in release 2.6. For Tensor contain only one element, Please modify  'Tensor.numpy()[0]' to 'float(Tensor)' as soon as possible, otherwise 'Tensor.numpy()[0]' will raise error in release 2.6.
PaddlePaddle Output: Tensor(shape=[], dtype=bool, place=Place(cpu), stop_gradient=True,
       False), Output Shape: []

PaddleNumpy Output: False, Output Shape: ()
"""

Also, paddle says to use float(input_tensor) if it's a scalar (0-D) tensor instead of .numpy() to access the single scalar value but we can also use the .item() for the same !

Additionally, only Paddle has this issue ; other frameworks handle this correctly:

import numpy as np
import tensorflow as tf
import torch
import paddle

x = np.array(1)
print(x.shape) # ()
y = paddle.to_tensor(1)
print(y.shape) # []
z = y.numpy()
print(z.shape) # (1,)
tf_1 = tf.constant(1)
print(tf_1.shape) # ()
z1 = tf_1.numpy()
print(z1.shape) # ()
t_orch = torch.tensor(1)
z2 = t_orch.numpy()
print(z2.shape) # ()

Paddle Frontend PR's encountering this issue:

• anyadded any to paddle math #19553
• rankRank paddle frontend #20185
• roll add roll paddle frontend manipulation #19258
• clonePaddle creation clone #19288
• kronEHN: adding kron in paddle math..py #18280
• unstackadd unstack paddle frontend manipulation #19242

Environment

Linux, VsCode + docker

Ivy Version

1.1.9

Backend

• NumPy
• TensorFlow
• PyTorch
• JAX

Device

CPU

[NripeshN]

HI @akshatvishu,
I faced this issue in #19242 too

[akshatvishu]

HI @akshatvishu, I faced this issue in #19242 too

added it to the list!

[CatB1t]

Thanks @akshatvishu, fixed in fabadb6

[NripeshN]

Hi @CatB1t, I am still facing this issue in #19242.

[akshatvishu]

Thanks @akshatvishu, fixed in fabadb6