Impact of Iteration Discrepancy Between Supervised and Unimatch Models on Optimization and Results
Opened this issue · 0 comments
yusirhhh commented
In the Unimatch code, I've observed that the 'self.ids' of labeled data are extended to match the length of the unlabeled data. Consequently, within the same epoch, the number of iterations for the supervised model and Unimatch differ significantly. This leads to discrepancies in the optimizer's step count for the model. What would the results be if the iterations were identical?