change_loading_options

app.py

@@ -35,7 +35,19 @@ from mogen.datasets.paramUtil import (
 from mogen.datasets.utils import recover_from_ric
 from mogen.datasets.pipelines import RetargetSkeleton
 
-
+import requests
+
+def load_large_files(relative_path):
+    # URL to file in the Hugging Face Hub
+    url = "https://huggingface.co/mingyuan/data_hf/" + relative_path
+    # Temporary download in Space
+    file_path = "/tmp/" + relative_path.split('/')[-1]
+    
+    response = requests.get(url)
+    with open(file_path, "wb") as f:
+        f.write(response.content)
+    return file_path
+    
 def motion_temporal_filter(motion, sigma=1):
     motion = motion.reshape(motion.shape[0], -1)
     for i in range(motion.shape[1]):
@@ -57,7 +69,7 @@ def plot_tomato(data, kinematic_chain, result_path, npy_path, fps, sigma=None):
 
 def create_lmm():
     config_path = "configs/lmm/lmm_small_demo.py"
-    ckpt_path = "data/motionverse/pretrained/lmm_small_demo.pth"
+    ckpt_path = load_file_list["lmm"]
     cfg = mmcv.Config.fromfile(config_path)
     model = build_architecture(cfg.model)
     load_checkpoint(model, ckpt_path, map_location='cpu')
@@ -70,15 +82,37 @@ def create_lmm():
 
 # device = 'cpu'
 device = 'cuda'
-# os.environ["NO_PROXY"]    = os.environ["no_proxy"]    = "localhost, 127.0.0.1:7860"
+type = "hf"
+
+if type == "local":
+    load_file_list = {
+        "mean": "../data_hf/data/motionverse/statistics/mean.npy", 
+        "std": "../data_hf/data/motionverse/statistics/std.npy", 
+        "skeleton": "../data_hf/data/motionverse/statistics/skeleton.npy", 
+        "lmm": "../data_hf/data/motionverse/pretrained/lmm_small_demo.pth", 
+        "imagebind": "../data_hf/data/motionverse/pretrained/imagebind_huge.pth"
+    }
+    os.environ["NO_PROXY"] = os.environ["no_proxy"] = "localhost, 127.0.0.1:7860"
+else:
+    src_file_list = {
+        "mean": "data/motionverse/statistics/mean.npy", 
+        "std": "data/motionverse/statistics/std.npy", 
+        "skeleton": "data/motionverse/statistics/skeleton.npy", 
+        "lmm": "data/motionverse/pretrained/lmm_small_demo.pth", 
+        "imagebind": "data/motionverse/pretrained/imagebind_huge.pth"
+    }
+    load_file_list = {}
+    for key in src_file_list.keys():
+        load_file_list[key] = load_large_files(src_file_list[key])
+
 model_lmm = create_lmm()
-model_imagebind = imagebind_huge(pretrained=True)
+model_imagebind = imagebind_huge(pretrained=True, ckpt_path=load_file_list["imagebind"])
 model_imagebind.eval()
 model_imagebind.to(device)
-rtg_skl = RetargetSkeleton(tgt_skel_file='data/motionverse/statistics/skeleton.npy')
+rtg_skl = RetargetSkeleton(tgt_skel_file=load_file_list["skeleton"])
 
-mean_path = "data/motionverse/statistics/mean.npy"
-std_path = "data/motionverse/statistics/std.npy"
+mean_path = load_file_list["mean"]
+std_path = load_file_list["std"]
 mean = np.load(mean_path)
 std = np.load(std_path)
 

mogen/models/utils/imagebind_wrapper.py

@@ -42,7 +42,7 @@ class FeatureExtractor(imagebind_model.ImageBindModel):
         return word_feat, seq_feat
     
 
-def imagebind_huge(pretrained=False):
+def imagebind_huge(pretrained=False, ckpt_path=None):
     model = FeatureExtractor(
         vision_embed_dim=1280,
         vision_num_blocks=32,
@@ -56,9 +56,9 @@ def imagebind_huge(pretrained=False):
     )
 
     if pretrained:
-        file_path = os.path.abspath(os.path.dirname(__file__))
-        ckpt_dir = os.path.join(file_path, '../../../data/motionverse/pretrained')
-        ckpt_path = os.path.join(ckpt_dir, 'imagebind_huge.pth')
+        # file_path = os.path.abspath(os.path.dirname(__file__))
+        # ckpt_dir = os.path.join(file_path, '../../../data/motionverse/pretrained')
+        # ckpt_path = os.path.join(ckpt_dir, 'imagebind_huge.pth')
         if not os.path.exists(ckpt_path):
             print(
                 "Downloading imagebind weights to motionverse/pretrained/imagebind_huge.pth ..."