import torch
import numpy as np
import torch.nn as nn
import torch_npu
from torch_npu.testing.testcase import TestCase, run_tests
from torch_npu.testing.common_utils import create_common_tensor
class TestThnnConvDepthwise2d(TestCase):
weight_grad = []
input_grad = []
def get_weight_grad(self, grad):
self.weight_grad.append(grad.to("cpu"))
def get_input_grad(self, grad):
self.input_grad.append(grad.to("cpu"))
def op_exec_cpu(self, input1, weight, in_channels,
out_channels, kernel_size, padding=0, stride=1, dilation=1, bias=True, group=2):
weight1 = weight
input1.requires_grad = True
input1.register_hook(lambda grad: self.get_input_grad(grad))
bias1 = False
if bias is not None:
bias1 = True
m1 = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias=bias1, groups=group)
m1.weight.data = weight1
m1.weight.register_hook(lambda grad: self.get_weight_grad(grad))
cpuOutput = m1(input1)
cpuOutput = cpuOutput.requires_grad_()
tmp = torch.ones_like(cpuOutput)
cpuOutput.backward(tmp)
return cpuOutput
def op_exec_npu(self, input1, weight, in_channels,
out_channels, kernel_size, padding=0, stride=1, dilation=1, bias=True, group=2):
weight1 = weight
input1.requires_grad = True
input1.register_hook(lambda grad: self.get_input_grad(grad))
bias1 = False
if bias is not None:
bias1 = True
m1 = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias=bias1, groups=group)
m1.weight.data = weight1
m1.weight.register_hook(lambda grad: self.get_weight_grad(grad))
m1 = m1.to("npu")
npuOutput = m1(input1)
npuOutput = npuOutput.to("cpu")
npuOutput = npuOutput.requires_grad_()
tmp = torch.ones_like(npuOutput)
npuOutput.backward(tmp)
return npuOutput
def thnn_conv_depthwise2d_format(self, i):
shape_format = [
[[np.float32, 3, (64, 3, 32, 32)], [np.float32, -1, (3, 1, 3, 3)], 0, 1, (1, 1), True],
[[np.float16, 3, (128, 3, 64, 64)], [np.float16, -1, (3, 1, 3, 3)], 0, 1, 1, None],
[[np.float16, 3, (32, 3, 16, 16)], [np.float16, -1, (3, 1, 3, 3)], 0, 1, 1, None],
[[np.float16, 3, (32, 6, 32, 32)], [np.float16, -1, (6, 1, 3, 3)], 0, 1, 1, None],
[[np.float16, 3, (32, 6, 32, 32)], [np.float16, -1, (6, 1, 3, 3)], 0, 1, 1, None]
]
return shape_format[i]
def thnn_conv_depthwise2d_execute(self, item, group):
self.weight_grad.clear()
self.input_grad.clear()
input_cpu, input_npu = create_common_tensor(item[0], 0, 10)
if input_cpu.dtype == torch.float16:
input_cpu = input_cpu.to(torch.float32)
weight_cpu, weight_npu = create_common_tensor(item[1], 0, 10)
if weight_cpu.dtype == torch.float16:
weight_cpu = weight_cpu.to(torch.float32)
kernel_size = (item[1][2][2], item[1][2][3])
cpu_output = self.op_exec_cpu(input_cpu, weight_cpu, item[0][2][1], item[1][2][0], kernel_size=kernel_size,
padding=item[2], stride=item[3], dilation=item[4], bias=item[5], group=group)
weight_npu = weight_npu.to("cpu")
npu_output = self.op_exec_npu(input_npu, weight_npu, item[0][2][1], item[1][2][0], kernel_size=kernel_size,
padding=item[2], stride=item[3], dilation=item[4], bias=item[5], group=group)
cpu_output = cpu_output.to(npu_output.dtype)
if item[5] is True:
self.assertRtolEqual(cpu_output.detach().numpy(), npu_output.detach().numpy(), 0.005)
else:
self.assertRtolEqual(cpu_output.detach().numpy(), npu_output.detach().numpy())
def test_thnn_conv_depthwise2d_0(self):
item = self.thnn_conv_depthwise2d_format(0)
self.thnn_conv_depthwise2d_execute(item, 3)
def test_thnn_conv_depthwise2d_1(self):
item = self.thnn_conv_depthwise2d_format(1)
self.thnn_conv_depthwise2d_execute(item, 3)
def test_thnn_conv_depthwise2d_2(self):
item = self.thnn_conv_depthwise2d_format(2)
self.thnn_conv_depthwise2d_execute(item, 3)
def test_thnn_conv_depthwise2d_3(self):
item = self.thnn_conv_depthwise2d_format(3)
self.thnn_conv_depthwise2d_execute(item, 6)
def test_thnn_conv_depthwise2d_4(self):
item = self.thnn_conv_depthwise2d_format(4)
self.thnn_conv_depthwise2d_execute(item, 6)
if __name__ == "__main__":
np.random.seed(1234)
run_tests()