-
Notifications
You must be signed in to change notification settings - Fork 23
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Validation accuracy on aircraft dataset and Model for Stanford Cars #5
Comments
hi,sorry to bother you.i got some trouble when i tried to run 'summit_val.py'.thr erorr are shown below. Traceback (most recent call last): looking for your replyment.thanks. |
@gyhui14 hi,sorry to bother you.i got some trouble when i tried to run 'summit_val.py'.thr erorr are shown below. Traceback (most recent call last): looking for your replyment.thanks. |
Hi Dr Guo
|
Hi Dr Guo
Sorry to bother you again. After the last chat with you, I split the validation set from training set. So, I train only on the training set and evaluate on the validation set. However, after several times training, I can only get an validation accuracy around 52% on aircraft dataset. I am using the default settings and hyperparameters. And the weight decay = 0.0005 is default as well. I only load the aircraft dataset and only train and test on this dataset. Could you please give me any explanation or guidance why the validation accuracy is much lower than the figure shown in the paper? Should I change any of the hyperparameters? Or I made mistake on anywhere? The settings and hyperparameters are shown below. Thanks.
parser.add_argument('--nb_epochs', default=110, type=int, help='nb epochs')
parser.add_argument('--lr', default=0.1, type=float, help='initial learning rate of net')
parser.add_argument('--lr_agent', default=0.01, type=float, help='initial learning rate of agent')
parser.add_argument('--datadir', default='./data/decathlon-1.0/', help='folder containing data folder')
parser.add_argument('--imdbdir', default='./data/decathlon-1.0/annotations/', help='annotation folder')
parser.add_argument('--ckpdir', default='./cv/', help='folder saving checkpoint')
parser.add_argument('--seed', default=0, type=int, help='seed')
parser.add_argument('--step1', default=40, type=int, help='nb epochs before first lr decrease')
parser.add_argument('--step2', default=60, type=int, help='nb epochs before second lr decrease')
parser.add_argument('--step3', default=80, type=int, help='nb epochs before third lr decrease')
Also, I am going to use Stanford Cars as my dataset. I noticed that in the paper it is stated that the Stanfords Car dataset is trained on the standard Pytorch ResNet50 model that is pretrained on ImageNet. However, the code for this model and training is not provided in this GitHub page. So, I have no idea how you combine the policy network with the default Pytorch ResNet50 Model. I tried to use the ResNet26 as the model to train on Stanford Cars, but the validation accuracy in the first epochs are not quite good (less than 1% after 5 epochs). Is it possible to provide this part of code of how you train on the Stanford Cars dataset by using ResNet50? Thank you very much!
The text was updated successfully, but these errors were encountered: