Quickfix WDS (lucidrains#357)

robvanvolt · web-flow · commit 55f21ad7eb8d · 2021-09-12T18:05:46.000-07:00
Deprecated length attribute got removed in the latest WebDataset version, so only check the length of the dataset if not using WDS.
Thanks to ARKseal for pointing out to that bug.
diff --git a/train_dalle.py b/train_dalle.py
@@ -382,8 +382,8 @@ def tokenize(s):
         tokenizer=tokenizer,
         shuffle=is_shuffle,
     )
+    assert len(ds) > 0, 'dataset is empty'
 
-assert len(ds) > 0, 'dataset is empty'
 if distr_backend.is_root_worker():
     if not ENABLE_WEBDATASET:
         print(f'{len(ds)} image-text pairs found for training')

Original file line number	Diff line number	Diff line change
`@@ -382,8 +382,8 @@ def tokenize(s):`
`382`	`382`	`tokenizer=tokenizer,`
`383`	`383`	`shuffle=is_shuffle,`
`384`	`384`	`)`
	`385`	`+ assert len(ds) > 0, 'dataset is empty'`
`385`	`386`
`386`		`-assert len(ds) > 0, 'dataset is empty'`
`387`	`387`	`if distr_backend.is_root_worker():`
`388`	`388`	`if not ENABLE_WEBDATASET:`
`389`	`389`	`print(f'{len(ds)} image-text pairs found for training')`