Hi, this is a good project.
I tried it, and the overall installation and training were very simple and straightforward. I experimented with FPENet, but the final result was larger than the original one. specifically:
I modified the hyperparameters to the original:
model = FPENet
dataset = cityscapes
input_size = 512, 1024
classes = 19
train_type = train
max_epochs = 400
lr_schedule = poly
The loss used CrossEntropy2d:
That is, line 133 of train: criteria = CrossEntropyLoss2dLabelSmooth (weight = weight, ignore_label = ignore_label), changed to: CrossEntropyLoss2d function, while other settings remain unchanged, using an RTX2080Ti GPU, training for 9h
But in the end, mIOU is only 46% on the val set, and the original effect is 60–70% on the test set, but I feel that there should not be such a big gap between val and test. I checked some output and noticed that the model parameters printed by train.py were 0.12M, but the original model was 0.4M. At first, I thought the model was wrong, but after checking the paper, I felt that your implementation was correct. Then I used torchsummary in the model to see that the model was 0.44M, so I didn't know what went wrong.
Maybe FPENet itself is difficult to reproduce? (Although this is common in AI papers).
Has anyone used this project to reproduce and roughly achieve the effect of an original model? Can you discuss and share the parameters and strategies set?
Hi, this is a good project.
I tried it, and the overall installation and training were very simple and straightforward. I experimented with FPENet, but the final result was larger than the original one. specifically:
I modified the hyperparameters to the original:
model = FPENet
dataset = cityscapes
input_size = 512, 1024
classes = 19
train_type = train
max_epochs = 400
lr_schedule = poly
The loss used CrossEntropy2d:
That is, line 133 of train: criteria = CrossEntropyLoss2dLabelSmooth (weight = weight, ignore_label = ignore_label), changed to: CrossEntropyLoss2d function, while other settings remain unchanged, using an RTX2080Ti GPU, training for 9h
But in the end, mIOU is only 46% on the val set, and the original effect is 60–70% on the test set, but I feel that there should not be such a big gap between val and test. I checked some output and noticed that the model parameters printed by train.py were 0.12M, but the original model was 0.4M. At first, I thought the model was wrong, but after checking the paper, I felt that your implementation was correct. Then I used torchsummary in the model to see that the model was 0.44M, so I didn't know what went wrong.
Maybe FPENet itself is difficult to reproduce? (Although this is common in AI papers).
Has anyone used this project to reproduce and roughly achieve the effect of an original model? Can you discuss and share the parameters and strategies set?