Open onlynewbie opened 2 months ago
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
Therefore, according to the code, the results of the paper cannot be reproduced, so how to modify the code in the warehouse, the accuracy rate is indeed too poor
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
Therefore, according to the code, the results of the paper cannot be reproduced, so how to modify the code in the warehouse, the accuracy rate is indeed too poor
According to the code, results can be reproduced. The code cannot ensure you can reproduce the results is because the code didn't take control of how you set up the dataset which does not provided in the repo, and you need to set it up. I cannot help if I don't have following info
what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
Therefore, according to the code, the results of the paper cannot be reproduced, so how to modify the code in the warehouse, the accuracy rate is indeed too poor
According to the code, results can be reproduced. The code cannot ensure you can reproduce the results is because the code didn't take control of how you set up the dataset which does not provided in the repo, and you need to set it up. I cannot help if I don't have following info
what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
I'm reading the code and see n-ways vs. novel_classes, is there any difference between the two of them
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
Therefore, according to the code, the results of the paper cannot be reproduced, so how to modify the code in the warehouse, the accuracy rate is indeed too poor
According to the code, results can be reproduced. The code cannot ensure you can reproduce the results is because the code didn't take control of how you set up the dataset which does not provided in the repo, and you need to set it up. I cannot help if I don't have following info what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
I'm reading the code and see n-ways vs. novel_classes, is there any difference between the two of them
Read the paper
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
Therefore, according to the code, the results of the paper cannot be reproduced, so how to modify the code in the warehouse, the accuracy rate is indeed too poor
According to the code, results can be reproduced. The code cannot ensure you can reproduce the results is because the code didn't take control of how you set up the dataset which does not provided in the repo, and you need to set it up. I cannot help if I don't have following info what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
I'm reading the code and see n-ways vs. novel_classes, is there any difference between the two of them
Read the paper
In Figure 7, the abscissa is the novel classes, and in the text it is the number of different ways
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?
Hi there, what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
tensorflow 2.12.In the pre-training stage, the code were exactly the same as the code in the repository,including shots and ways, and then I modified the code of finetune to make him run and the results were not very ideal, with an accuracy rate of only 15
Sorry, the environment refer to the environment of data, the result reported in the paper were produced by the CNN pretrained with significant dataset at lab environment with 5520 samples split into train and test
The dataset is SignFi, and the data processing and data partitioning work perfectly according to the repository code.Just now I tried the learning rate in the paper, and the accuracy rate of the pre-training is still 42%.
I guess you haven't look into the dataset yet,
Signfi dataset has different users and environments, run according to the code can't 100% ensure the setup, as you couldn't train with wrong dataset part.
Therefore, according to the code, the results of the paper cannot be reproduced, so how to modify the code in the warehouse, the accuracy rate is indeed too poor
According to the code, results can be reproduced. The code cannot ensure you can reproduce the results is because the code didn't take control of how you set up the dataset which does not provided in the repo, and you need to set it up. I cannot help if I don't have following info what was your training and testing environment and the number of shots and ways? How was the classification was done? By cosine similarity or fully connected layer?
I'm reading the code and see n-ways vs. novel_classes, is there any difference between the two of them
Read the paper
In Figure 7, the abscissa is the novel classes, and in the text it is the number of different ways
Terminology has been clearly defined in the paper, if you can open the PDF of the paper, and then search key word "way", you can easily find where in the paper the terminology "way" and "number of novel class" were defined (page number 5 up-left corner). Or you can use CHATGPT or CLAUDE 2, and throw the pdf in and entering your question, he will find the terminology definition for you.
In addition, I am not sure of the meaning of "abscissa". Let me know if you have any further questions about the paper.
I reproduced the pretraining signfi part but the accuracy rate is only 46%, and the modelFineTuning part doesn't seem to be called, can you tell me how to call it?