Update inference_lora.py
Browse files- inference_lora.py +4 -17
inference_lora.py
CHANGED
|
@@ -51,24 +51,11 @@ from safetensors.torch import load_file
|
|
| 51 |
from diffsynth.pipelines.wan_video import WanVideoPipeline
|
| 52 |
from diffsynth.core import ModelConfig
|
| 53 |
from diffsynth.models.wan_video_dit import TemporalAttentionAdapter, FramewiseCrossAttention
|
|
|
|
|
|
|
|
|
|
| 54 |
|
| 55 |
-
|
| 56 |
-
# ============================================================================
|
| 57 |
-
# ActionFFNEncoder (must match training definition)
|
| 58 |
-
# ============================================================================
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
class ActionFFNEncoder(nn.Module):
|
| 62 |
-
def __init__(self, action_dim: int, embed_dim: int, num_layers: int = 2):
|
| 63 |
-
super().__init__()
|
| 64 |
-
layers: list[nn.Module] = [nn.Linear(action_dim, embed_dim), nn.GELU()]
|
| 65 |
-
for _ in range(max(0, num_layers - 2)):
|
| 66 |
-
layers += [nn.Linear(embed_dim, embed_dim), nn.GELU()]
|
| 67 |
-
self.mlp = nn.Sequential(*layers)
|
| 68 |
-
self.norm = nn.LayerNorm(embed_dim)
|
| 69 |
-
|
| 70 |
-
def forward(self, actions: torch.Tensor) -> torch.Tensor:
|
| 71 |
-
return self.norm(self.mlp(actions))
|
| 72 |
|
| 73 |
|
| 74 |
# ============================================================================
|
|
|
|
| 51 |
from diffsynth.pipelines.wan_video import WanVideoPipeline
|
| 52 |
from diffsynth.core import ModelConfig
|
| 53 |
from diffsynth.models.wan_video_dit import TemporalAttentionAdapter, FramewiseCrossAttention
|
| 54 |
+
_MT_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), "examples", "wanvideo", "model_training")
|
| 55 |
+
if _MT_DIR not in sys.path:
|
| 56 |
+
sys.path.insert(0, _MT_DIR)
|
| 57 |
|
| 58 |
+
from train_acwm import ActionFFNEncoder
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
|
| 60 |
|
| 61 |
# ============================================================================
|