About the ImageNet pre-training model

[ZC0015wqj]

Hello, thank you for your help to me before, but I still want to ask how to pre-train on the ImageNet-1K dataset. It is the same as the kitti_data dataset training method, modify the file path in eigen_zhou, and then follow Is it ok to train with the command given. I have downloaded the ImageNet-1K dataset now, thank you very much for your help

[noahzn]

Hi,

No, this repo does not contain the code to pre-train a model on ImageNet. Please prepare your own code to pre-train the encoder. PyTorch official has a minimal example code. Please feel free to have a look at it.

[ZC0015wqj]

Thank you very much for your help, may I take the liberty to ask you, can you share this part of the training code

[noahzn]

Hi, the pre-training code needs some cleaning up, but I have no time to do it. I am not going to include it in this repo as it is not that relevant. I can give you the code in a few days if you send me an email.

[ZC0015wqj]

Thank you very much, can you give me your email

[ZC0015wqj]

I sent an email, I'm not sure if you received it

[noahzn]

Yes, I received your email. I will send you the code in a few days.

[ZC0015wqj]

Thank you very much!!!!!!!!!!

[ZC0015wqj]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-training model?

[noahzn]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-training model?

If you do not want to use the ImageNet pre-training you can simply not set --mypretrain.

[ZC0015wqj]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-trained model? At the same time, I still have a problem. When I use the command 'tensorboard --log_dir ./tmp/mytrain' to check the log, I found that in a certain part, it has achieved higher accuracy and lower accuracy than the final training result. Error, but when I evaluate all the generated weights, I find that the accuracy of these models is far worse than what I see in the log, can you please tell me why?

[ZC0015wqj]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-training model?

If you do not want to use the ImageNet pre-training you can simply not set --mypretrain.

This means that I don't need to modify other parameters, just don't add this option, right?

[noahzn]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-training model?

If you do not want to use the ImageNet pre-training you can simply not set --mypretrain.

This means that I don't need to modify other parameters, just don't add this option, right?

yes.

[noahzn]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-trained model? At the same time, I still have a problem. When I use the command 'tensorboard --log_dir ./tmp/mytrain' to check the log, I found that in a certain part, it has achieved higher accuracy and lower accuracy than the final training result. Error, but when I evaluate all the generated weights, I find that the accuracy of these models is far worse than what I see in the log, can you please tell me why?

Please see the comments here.

[ZC0015wqj]

Hello, how can I set the parameters to train the optimal model without the ImageNet pre-trained model? At the same time, I still have a problem. When I use the command 'tensorboard --log_dir ./tmp/mytrain' to check the log, I found that in a certain part, it has achieved higher accuracy and lower accuracy than the final training result. Error, but when I evaluate all the generated weights, I find that the accuracy of these models is far worse than what I see in the log, can you please tell me why?

Please see the comments here.

Well, when I found out that there is a big difference between them, I thought I could get a good result, now my result is not very ideal

[noahzn]

"The final training result", do you mean the last epoch? If so, please check the results of all the epochs, you might get the best result at earlier epoch.

[ZC0015wqj]

Yes, it's strange that when I checked all epochs, I didn't find a better value

[noahzn]

I sent an email, I'm not sure if you received it

I have sent you the code.

[ZC0015wqj]

I sent an email, I'm not sure if you received it

I have sent you the code.

Thank you very much！

[ZC0015wqj]

Could you please provide your complete training code for reference? When I used the code you sent to the email for training, I encountered a problem similar to the one in the picture. I don’t know how to solve it, so I would like to take the liberty of asking a copy of your complete training code is ok

[noahzn]

Hi, the output of the encoder should be a single tensor (the classification result). I already gave you a hint in the code return self.norm(x.mean([-2, -1])) # Global average pooling, (N, C, H, W) -> (N, C) please check if the dimension of your output is NxC.

[ZC0015wqj]

can you help me solve this problem Traceback (most recent call last): File "nain.py", line 502, in main(args) File "main.py" , line 412, in mainmodel_flops = flops.total() File " /hone/pc405/anaconda3/envs/nd/lb/python3.8/site-packages/fvcore/n/jit_analysis.py",line 248，in totalstats = self._analyze() File " /hone/pc405/anaconda3/envs/nd/lib/python3.8/site-packages/fvcore/n/jit_analysis.py", line 551,in analyzegraph = _get_scoped_trace_graph(self._nodel, self._inputs, self._aliases) File "/hone/pc4O5/ anaconda3/envs /nd/l.b/python3.8/stite-packages/fvcore/n/jit_analysis.py",line 176,in _get scoped_trace graphgraph,- =_get_trace_graph( module, inputs) File "/hone/pc405/anaconda3/envs/nd/l.b/python3.8/site-packages/torch/jit/l_trace.py",line 1175， in _get_trace_graphouts ' : ONWXTracedNodule(f, strict,_force_outplace,return_inputs,_return_inputs_states)(*args，kwargs) File "/hone/pc405/anaconda3/envs/nd/lib/python3.8/site-packages/torch/nn/nodules/Rodule. py", line 1130， in _call_implreturn forward_call( *input,kwargs) File " /hone/pc405/anaconda3/envs/nd/lib/python3.8/site-packages/torch/jit/l_trace.py",line 127， in forwardgraph, out = torch._c._create_graph_by_tracing( File "/hone/pc405/anaconda3/envs/nd/lib/python3.8/site-packages/torch/jit/_trace.py", line 118，in wrapperouts.append(self.inner(trace_inputs)) File "/hone/pc405/anaconda3/envs/nd/lib/python3.8/site-packages/torch/n/nodules /nodule.py",line 1148, in _call_inplresult = forward call(input,kwargs) File "/hore/pc405/anaconda3/envs/nd/lib/python3.8/site-packages/torch/nn/nodules/nodule.py", line 1118,in _slow_forward result = self.forward(*input,kwargs) File " /hove/pc405/deeplearning/zc/lite123/lite-nono-pretrain-code (1)/nodels/depth_encoder .py",line 481，in forwardx = self.forward_features(x) File "/hone/pc405/deeplearning/zc/lite123/lite-nono-pretrain-code (1)/nodels/depth_encoder.py",line 467，in forward_featiresx =torch.cat(tmpx,dim=1) RuntineError: Tensors must have sane nunber of dimensions: got 4 and 2

[noahzn]

Hi, I'm not able to help you because you modified the code and I don't know what forward_featiresx is. It seems to be a simple feature channel misalignment problem.

[ZC0015wqj]

您好，我无法帮助您，因为您修改了代码，但我不知道是什么forward_featiresx。这似乎是一个简单的特征通道错位问题。

When I added depth_encoder to the lite_mono_pretrain code for training, it prompted me that the depth_encoder output is a 3D list type when calculating the loss, but the given target is a tensor type with a size of 128. May I ask how you unify the output of depth_encoder with the target?

[noahzn]

can you print the shapes of the output and the tensor before self.norm(x.mean([-2, -1])) of your depth_encoder?

[ZC0015wqj]

(N, C, H, W) -> (N, C)

I don't really understand that the output of the encoding layer should be a list structure as shown in the figure. We don't seem to be able to change its dimension directly through torch.mean (N, C, H, W) -> (N, C), do you need to convert the list to Tensor type?

[noahzn]

You should keep the output as a tensor, not a list. If the output of your network is x = (N,C,H,W), here the C is 1000 because you have 1000 classes. Then you use self.norm(x.mean([-2, -1])), and you can get a tensor of dimension (N, C). It's not a list!

[ZC0015wqj]

网络的输出不是三个不同阶段的特征图吗

您应该将输出保留为张量，而不是列表。如果网络的输出为x = (N,C,H,W)，则此处 C 为 1000，因为您有 1000 个类。然后使用self.norm(x.mean([-2, -1]))，就可以得到一个维度的张量(N, C)。这不是一个清单！

[noahzn]

不是的，当中阶段的特征是用于之后级联到decoder上面的，但是你这里是为了预训练一个分类encoder，因此你只应该输出最后一个结果。

[ZC0015wqj]

Can such two structures directly make losses? If we do it directly, we will encounter problems. Is this correct?

[noahzn]

No. You cannot compute the loss with nan.

[ZC0015wqj]

When the sample passes the model, its type becomes this

[noahzn]

There must be some errors in your network. My pre-training code has been shared with several people and they have no such problems.

[ZC0015wqj]

This is the shape when the encoding layer is output, but there is a problem during pre-training

[ZC0015wqj]

There must be some errors in your network. My pre-training code has been shared with several people and they have no such problems.

这个跟输入的H,W有关吗

[noahzn]

No. The main.py defines the image size for training. parser.add_argument('--input_size', default=256, type=int, help='image input size')

[ZC0015wqj]

否。它main.py定义了训练的图像大小。parser.add_argument('--input_size', default=256, type=int, help='image input size') When calculating theloss, the value of NAN appeared again, and now the output of the encoding layer is normal

[noahzn]

Do you have this layer after getting your x?
self.head = nn.Linear(dims[-2], num_classes) x = self.head(x)

[ZC0015wqj]

拿到手后有这一层吗x？ self.head = nn.Linear(dims[-2], num_classes) `x = self.head(x) Yes, this structure exists

[noahzn]

Have you solved this problem?

[noahzn]

I'm now closing this issue because I haven't heard back from you. Please reopen it or create a new one if you have further questions.