[Bug Fix] fix the initialization bug of all zero (#955)

Author: zhtmike

examples/hunyuanvideo-i2v/hyvideo/modules/embed_layers.py

@@ -60,7 +60,7 @@ def __init__(
         # nn.init.zeros_(self.proj.bias)
         w = self.proj.weight
         w_flatted = w.reshape(w.shape[0], -1)
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
 
         self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()
 

examples/opensora_hpcai/opensora/models/stdit/stdit.py

@@ -449,7 +449,7 @@ def _basic_init(module):
         w_flatted = w.reshape(w.shape[0], -1)
         # FIXME: incompatible in optim parallel mode
         # FIXME: impl in torch can be incorrect. can be reshape order mismatch
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
 
         # Initialize timestep embedding MLP:
         normal_(self.t_embedder.mlp[0].weight, std=0.02)

examples/opensora_hpcai/opensora/models/stdit/stdit2.py

@@ -538,7 +538,7 @@ def _basic_init(module):
         w_flatted = w.reshape(w.shape[0], -1)
         # FIXME: incompatible in optim parallel mode
         # FIXME: impl in torch can be incorrect. can be reshape order mismatch
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
 
         # Initialize timestep embedding MLP:
         normal_(self.t_embedder.mlp[0].weight, std=0.02)

examples/pixart_sigma/pixart/modules/pixart/pixart.py

@@ -183,7 +183,7 @@ def _basic_init(module):
         # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
         w = self.x_embedder.proj.weight.data
         w_flatted = w.view(w.shape[0], -1)
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
         constant_(self.x_embedder.proj.bias, 0)
 
         # Initialize timestep embedding MLP:
@@ -358,7 +358,7 @@ def _basic_init(module):
         # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
         w = self.x_embedder.proj.weight.data
         w_flatted = w.view(w.shape[0], -1)
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
         constant_(self.x_embedder.proj.bias, 0)
 
         # Initialize timestep embedding MLP:

mindone/models/dit.py

@@ -531,7 +531,7 @@ def _basic_init(module):
         w = self.x_embedder.proj.weight
         # xavier_uniform_(w.view(w.shape[0], -1))
         w_flatted = w.view(w.shape[0], -1)
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
         constant_(self.x_embedder.proj.bias, 0)
 
         # Initialize label embedding table:

mindone/models/latte.py

@@ -123,7 +123,7 @@ def _basic_init(module):
         w = self.x_embedder.proj.weight
         # xavier_uniform_(w.view(w.shape[0], -1))
         w_flatted = w.view(w.shape[0], -1)
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
         constant_(self.x_embedder.proj.bias, 0)
 
         # Initialize label embedding table:

mindone/models/mmdit.py

@@ -350,7 +350,7 @@ def _basic_init(module):
         w = self.x_embedder.proj.weight
         # xavier_uniform_(w.view(w.shape[0], -1))
         w_flatted = w.view(w.shape[0], -1)
-        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).reshape(w.shape))
+        w.set_data(initializer(XavierUniform(), w_flatted.shape, w_flatted.dtype).init_data().reshape(w.shape))
         constant_(self.x_embedder.proj.bias, 0)
 
         # Initialize timestep embedding MLP: