vqdang
CONSEP - Getting dice_0=0 Classification during training
when training with CoNSeP i am getting 0 at type 1 dice after nearly every 20-23 epochs of stage 2 training.
This behavior also happens with the PanNuke Dataset. I have used the same settings as described in the paper i.e. LR=1e-4 for first 25 epochs and 1e-5 for the last 25 of both the stages, the same parameters as described in the paper, but still encountering these 0 values when training for classification of consep and pannuke.
Hi, I am sorry for the late reply. The above can happen when the number of instances of type 1 for training is small. I think it should be the Miscellaneous type. You should try to split your training set (or double-check it) so that it contains more of that type.
yes i figured much, that it's because of the miscellaneous type. I am using whole of CoNSeP training data just like in the paper, so I am confused why is it going 0? am i doing something wrong? did you guys split the training data w.r.t the misc type to get the results? Actually, i am confused about, why go to 0, okay a low score is possible but why 0? When it doesn't do such a behavior in the paper? is there any specific setting that i am missing? how did you guys cater this in the paper?
Miscellaneous is a very noisy class. Sometimes you will get some unexpected behaviour. Do you have the dice over time during training? You should have this in your tensorboard output.
am i missing something in the classification pipeline?? @simongraham More insight onto the training details are as follows:
- extract patches: 540x540_80x80, all of train, and test as validation, original mode, act_shape = [270, 270], nr_type = 5
- stage 1: epochs=50, msge weight = 2 , all others are 1. Lr =1e-4 and decays by 0.1 every 25 epochs. The "pretrained" model= ImageNet-ResNet50-Preact_pytorch.tar. batch_size 8
- stage2: epochs=50, msge weight = 2, all the others are 1. Lr=1e-4 and decays by 0.1 every 25 epochs. batch_size is 4
Hello! I have met a similarly question with you. like flow: ------valid-np_acc : 0.92795 ------valid-np_dice : 0.73184 ------valid-tp_dice_0 : 0.94518 ------valid-tp_dice_1 : 0.02475 ------valid-tp_dice_2 : 0.49475 ------valid-tp_dice_3 : 0.00000 ------valid-tp_dice_4 : 0.61169 ------valid-hv_mse : 0.06671 the dice_3 is always 0. I use the Consep dataset,and set the nr_types are 5.
