soochan-lee / cn-dpm Goto Github PK

Thank you very much for the code which is well written and easy to launch for training. I could train the generative model version however if I have troubles understanding how to run inference on external dataset from the saved model, could you provide some script to run testing?
For instance for reconstruction, do you always use the last expert, let say at the end of split-mnist if I want to retest on 0/1?

When I change the ckpt_step from 1000000000 to 10000 in cifar10-cndpm.yaml, it ave the following error:

Saving checkpoint... Traceback (most recent call last):
  File "main.py", line 92, in <module>
    main()
  File "main.py", line 88, in main
    train_model(config, model, data_scheduler, writer)
  File "./CN-DPM/train.py", line 76, in train_model
    pickle.dump(model, f)
AttributeError: Can't pickle local object 'CnnSharingVae.__init__.<locals>.<lambda>'

pip install -r requirements.txt gives the following error:

ERROR: Could not find a version that satisfies the requirement pytorch==1.0.1 (from -r requirements.txt (line 9)) (from versions: 0.1.2, 1.0.2)
ERROR: No matching distribution found for pytorch==1.0.1 (from -r requirements.txt (line 9))

It seems conda install pytorch==1.0.1 torchvision==0.2.2 -c pytorch works.

Firstly thank you so much for sharing the codebase. In the function nll(self, x, y, step=None), you calculate the chilled_log_softmax. I am wondering what's the meaning of it and why do you want to use it. Thank you in advance.

        # Classifier chilling
        chilled_log_softmax = F.log_softmax(
            log_softmax / self.config['classifier_chill'], dim=1)
        chilled_loss_pred = self.ce_loss(chilled_log_softmax, y)

        # Value with chill & gradient without chill
        loss_pred = loss_pred - loss_pred.detach() \
            + chilled_loss_pred.detach()

The original code works well on cifar100, however after I add normalization to cifar100 it gets very low accuracy. Is there any way to make it work on normalized dataset?

Hi,

We are trying to apply CN-DPM on another dataset called CORE50 and the result is quite bad. Since CNDPM has so many hyper-parameters, except for log-alpha and classifier-chill, what hyper-parameters do you think are crucial to tune? Thank you so much and I look forward to hearing back from you.