- When I implemented Unet network on VTA, I encountered the following problems: due to the upsample sampling layer in Unet network, I encountered an error in the graph_pack process. But if you set stop_name before upsample, you can successfully graph_pack. Here are the Settings in my code:
Upsample in Unet is set as follows:
class Up(nn.Module):
"""Upscaling then double conv"""
def __init__(self, in_channels, out_channels, bilinear=True):
super().__init__()
# if bilinear, use the normal convolutions to reduce the number of channels
if bilinear:
self.up = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
self.conv = DoubleConv(in_channels, out_channels, in_channels // 2)
else:
self.up = nn.ConvTranspose2d(in_channels, in_channels // 2, kernel_size=2, stride=2)
self.conv = DoubleConv(in_channels, out_channels)
def forward(self, x):
x = self.up(x)
x = torch.cat([x, x], dim=1)
return self.conv(x)
After importing the model, its corresponding in Realy is:
%227 = image.resize2d(%226, size=[64, 64], coordinate_transformation_mode="align_corners", rounding_method="", cubic_alpha=-0.75f) /* ty=Tensor[(1, 128, 64, 64), float32] */;
%228 = (%227, %227);
%229 = concatenate(%228, axis=1) /* ty=Tensor[(1, 256, 64, 64), float32] */;
%230 = multiply(%229, 16f /* ty=float32 */) /* ty=Tensor[(1, 256, 64, 64), float32] */;
%231 = round(%230) /* ty=Tensor[(1, 256, 64, 64), float32] */;
%232 = clip(%231, a_min=-127f, a_max=127f) /* ty=Tensor[(1, 256, 64, 64), float32] */;
%233 = cast(%232, dtype="int8") /* ty=Tensor[(1, 256, 64, 64), int8] */;
- But before I run Deploy Pretrained Vision Detection Model from Darknet on VTA，I found that I could compile the following layer:
%123 = nn.upsampling(%122, scale_h=2f, scale_w=2f) /* ty=Tensor[(1, 128, 26, 26), float32] */;
%124 = (%123, %63);
%125 = concatenate(%124, axis=1) /* ty=Tensor[(1, 384, 26, 26), float32] */;
%126 = multiply(%125, 5.56522f /* ty=float32 */) /* ty=Tensor[(1, 384, 26, 26), float32] */;
%127 = round(%126) /* ty=Tensor[(1, 384, 26, 26), float32] */;
%128 = clip(%127, a_min=-127f, a_max=127f) /* ty=Tensor[(1, 384, 26, 26), float32] */;
%129 = cast(%128, dtype="int8") /* ty=Tensor[(1, 384, 26, 26), int8] */;
It seems that Upsample can be implemented
Therefore, NOW I am confused. Can Upsample be implemented on VTA? Is there something wrong with the Upsample and cat set in my network? Who has encountered this problem, I hope you can get the answer. Thank you