Fix a number of issues with the infrastructure, no major rework
1 unresolved thread
1 unresolved thread
Fix a number of issues present in the code:
- type issues
- README.md issues
- variable initializations
- model issues
- training loss / number of batches calculation issues
Merge request reports
Activity
Filter activity
requested review from @vlad_andrei.badoiu1
assigned to @agherghescu2411
added 1 commit
- 64302265 - Fix final training loss calculation, fix estimation interval
51 51 52 52 print(f"Done. Took {time.time() - start:.2f}s.") 53 53 54 self.num_batches = torch.cat(self._data, 55 dim=-1).shape[0] // self.bs // self.seq_len 54 # pre-calculate the number of batches in the dataset 55 56 # Note: there's a special case we need to be careful about; since the 57 # predictions are simply the inputs shifted to the right by one value; 58 # there's a case when the dataset ends before we can get these 59 # shifted-right predictions; this occurs iff `batch_len % seq_len == 0`; 60 # to avoid this, we have to be explicit about the available number of 61 # batches (by simply subtracting 1 from the total number of available 62 # batches) 63 dataset_stream_len = 0 64 for sample in self._data: mentioned in commit ba87b2ba
Please register or sign in to reply