dataset_tool.py docs update, tf32 disable for calc_metrics.py

- Add a justification for using uncompressed zip and uncompressed png (#22)
- Clarify class label usage with dataset_tool.py (#18)
- Disable tf32 in calc_metrics as this has not been tested.

Author: jannehellsten

calc_metrics.py

@@ -47,6 +47,8 @@ def subprocess_fn(rank, args, temp_dir):
     # Print network summary.
     device = torch.device('cuda', rank)
     torch.backends.cudnn.benchmark = True
+    torch.backends.cuda.matmul.allow_tf32 = False
+    torch.backends.cudnn.allow_tf32 = False
     G = copy.deepcopy(args.G).eval().requires_grad_(False).to(device)
     if rank == 0 and args.verbose:
         z = torch.empty([1, G.z_dim], device=device)

dataset_tool.py

@@ -331,14 +331,34 @@ def convert_dataset(
     --source path/                      Recursively load all images from path/
     --source dataset.zip                Recursively load all images from dataset.zip
 
-    The output dataset format can be either an image folder or a zip archive.
     Specifying the output format and path:
 
     \b
     --dest /path/to/dir                 Save output files under /path/to/dir
     --dest /path/to/dataset.zip         Save output files into /path/to/dataset.zip
 
+    The output dataset format can be either an image folder or an uncompressed zip archive.
+    Zip archives makes it easier to move datasets around file servers and clusters, and may
+    offer better training performance on network file systems.
+
     Images within the dataset archive will be stored as uncompressed PNG.
+    Uncompresed PNGs can be efficiently decoded in the training loop.
+
+    Class labels are stored in a file called 'dataset.json' that is stored at the
+    dataset root folder.  This file has the following structure:
+
+    \b
+    {
+        "labels": [
+            ["00000/img00000000.png",6],
+            ["00000/img00000001.png",9],
+            ... repeated for every image in the datase
+            ["00049/img00049999.png",1]
+        ]
+    }
+
+    If the 'dataset.json' file cannot be found, the dataset is interpreted as
+    not containing class labels.
 
     Image scale/crop and resolution requirements: