I have 2 branches of output and a snip of the program is given below:
def forward(self, samples: NestedTensor):
# get the backbone features
features = self.backbone(samples)
# forward the feature pyramid
features_fpn = self.fpn([features[1], features[2], features[3]])
batch_size = features[0].shape[0]
# run the regression and classification branch
regression = self.regression(features_fpn[1]) * 100 # 8x
classification = self.classification(features_fpn[1])
anchor_points = self.anchor_points(samples)
# decode the points as prediction
output_coord = regression + anchor_points
output_class = classification
out = torch.cat((output_class, output_coord))
return out
When I try to compile this it shows me this error:
[UNILOG][FATAL][XCOM_SIZE_UNMATCH][The object's size is not not matching the requirement.] xir::Op{name = P2PNet__P2PNet_3645, type = concat}'s axis is error. It's 0
I updated the code to this:
def forward(self, samples: NestedTensor):
get the backbone features
features = self.backbone(samples)
# forward the feature pyramid
features_fpn = self.fpn([features[1], features[2], features[3]])
batch_size = features[0].shape[0]
# run the regression and classification branch
regression = self.regression(features_fpn[1]) * 100 # 8x
classification = self.classification(features_fpn[1])
anchor_points = self.anchor_points(samples)
# decode the points as prediction
output_coord = regression + anchor_points
output_class = classification
outt = torch.cat((output_class, output_coord), 1)
print(outt)
return outt
When I try to compile this it shows me this warning:
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3872, type = eltwise-fix} has been assigned to CPU: [DPU only supports positive "input_channel"(0)].
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3875, type = concat-fix} has been assigned to CPU: [Input xir::Op{name = P2PNet__P2PNet_3872, type = eltwise-fix} is not in DPU subgraph. And output dimension is not 4.].
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3870_new, type = const-fix} has been assigned to CPU: [Has no fanout or at least one fanout is out of DPU subgraph.].
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3875, type = concat-fix} has been assigned to CPU: [Input xir::Op{name = P2PNet__P2PNet_3872, type = eltwise-fix} is not in DPU subgraph. And output dimension is not 4.].
The xmodel is created but i am not getting the result:
I have 2 branches of output and a snip of the program is given below:
When I try to compile this it shows me this error:
[UNILOG][FATAL][XCOM_SIZE_UNMATCH][The object's size is not not matching the requirement.] xir::Op{name = P2PNet__P2PNet_3645, type = concat}'s axis is error. It's 0
I updated the code to this:
def forward(self, samples: NestedTensor):
get the backbone features
When I try to compile this it shows me this warning:
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3872, type = eltwise-fix} has been assigned to CPU: [DPU only supports positive "input_channel"(0)].
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3875, type = concat-fix} has been assigned to CPU: [Input xir::Op{name = P2PNet__P2PNet_3872, type = eltwise-fix} is not in DPU subgraph. And output dimension is not 4.].
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3870_new, type = const-fix} has been assigned to CPU: [Has no fanout or at least one fanout is out of DPU subgraph.].
[UNILOG][WARNING] xir::Op{name = P2PNet__P2PNet_3875, type = concat-fix} has been assigned to CPU: [Input xir::Op{name = P2PNet__P2PNet_3872, type = eltwise-fix} is not in DPU subgraph. And output dimension is not 4.].
The xmodel is created but i am not getting the result:
My cpu result for the code is:
output dimension: (1, 102400, 2)
tensor([[[ 4.3518, -6.3351], [ 5.4877, -5.5600], [ 4.8831, -5.5953], ..., [1278.6753, 631.6872], [1271.6644, 639.0435], [1277.7587, 632.8604]]], grad_fn=)
but in DPU I am getting:
output_ndim: (1, 51200, 2) [array([[ 1.5 , -1.5 ], [-8.5 , -3. ], [-0.75, -1.25], ..., [ 1.25, -2.25], [-2.25, 0.75], [-0.5 , -5. ]], dtype=float32)]
Can you tell me how to fix this?
To get some understanding of what I am doing please visit. The work is similar but not the same:
https://blog.csdn.net/wjytbest/article/details/124188661