prepare large-scale clip encoder training

configs/stable-diffusion/txt2img-1p4B-multinode-clip-encoder.yaml

@@ -74,8 +74,9 @@ data:
   target: ldm.data.laion.WebDataModuleFromConfig
   params:
     tar_base: "pipe:aws s3 cp s3://s-datasets/laion5b/laion2B-data/"
-    batch_size: 12
+    batch_size: 56
     num_workers: 4
+    multinode: True
     train:
       shards: '{000000..231317}.tar -'
       shuffle: 10000
@@ -89,19 +90,19 @@ data:
         params:
           size: 256
 
-    # NOTE use enough shards to avoid empty validation loops in workers
+#    # NOTE use enough shards to avoid empty validation loops in workers
-    validation:
+#    validation:
-      shards: '{231318..231349}.tar -'
+#      shards: '{231318..231349}.tar -'
-      shuffle: 0
+#      shuffle: 0
-      image_key: jpg
+#      image_key: jpg
-      image_transforms:
+#      image_transforms:
-      - target: torchvision.transforms.Resize
+#      - target: torchvision.transforms.Resize
-        params:
+#        params:
-          size: 256
+#          size: 256
-          interpolation: 3
+#          interpolation: 3
-      - target: torchvision.transforms.CenterCrop
+#      - target: torchvision.transforms.CenterCrop
-        params:
+#        params:
-          size: 256
+#          size: 256
 
 
 lightning:
@@ -118,8 +119,6 @@ lightning:
   trainer:
     #replace_sampler_ddp: False
     benchmark: True
-    val_check_interval: 50000
+    #val_check_interval: 50000
     num_sanity_val_steps: 0
-
+    accumulate_grad_batches: 2
-
-