Fix

app.py

@@ -27,8 +27,8 @@ def parse_option():
     parser.add_argument("--guidance_scale", type=float,  default=5, help="")
     parser.add_argument("--alpha_scale", type=float,  default=1, help="scale tanh(alpha). If 0, the behaviour is same as original model")
     parser.add_argument("--load-text-box-generation", type=arg_bool, default=True, help="Load text-box generation pipeline.")
-    parser.add_argument("--load-text-box-inpainting", type=arg_bool, default=False, help="Load text-box inpainting pipeline.")
-    parser.add_argument("--load-text-image-box-generation", type=arg_bool, default=False, help="Load text-image-box generation pipeline.")
+    parser.add_argument("--load-text-box-inpainting", type=arg_bool, default=True, help="Load text-box inpainting pipeline.")
+    parser.add_argument("--load-text-image-box-generation", type=arg_bool, default=True, help="Load text-image-box generation pipeline.")
     args = parser.parse_args()
     return args
 args = parse_option()

dataset/tsv_dataset.py

@@ -190,7 +190,7 @@ class TSVDataset(BaseDataset):
         self.which_layer_image = which_layer[1]
 
         #self.projection_matrix = torch.load(os.path.join(os.path.dirname(__file__), 'projection_matrix')  )
-        self.projection_matrix = torch.load('projection_matrix')
+        self.projection_matrix = torch.load('projection_matrix.pth')
 
         # Load tsv data
         self.tsv_file = TSVFile(self.tsv_path)

gligen/projection_matrix.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:819d51fde084e16e5960323c8bafba07fa8ee727e5403e5e4bdced4333c68faa
+size 2360043

gligen/task_grounded_generation.py

@@ -107,7 +107,7 @@ def get_clip_feature(model, processor, input, is_image=False):
         if feature_type[1] == 'after_renorm':
             feature = feature*28.7
         if feature_type[1] == 'after_reproject':
-            feature = project( feature, torch.load('gligen/projection_matrix').cuda().T ).squeeze(0)
+            feature = project( feature, torch.load('gligen/projection_matrix.pth').cuda().T ).squeeze(0)
             feature = ( feature / feature.norm() )  * 28.7 
             feature = feature.unsqueeze(0)
     else:
@@ -249,16 +249,9 @@ def grounded_generation_box(loaded_model_list, instruction, *args, **kwargs):
 
 
     # ------------- other logistics ------------- #
-    os.makedirs( os.path.join(save_folder, 'images'), exist_ok=True)
-    os.makedirs( os.path.join(save_folder, 'layout'), exist_ok=True)
-    os.makedirs( os.path.join(save_folder, 'overlay'), exist_ok=True)
-
-    start = len(  os.listdir(os.path.join(save_folder, 'images')) )
-    image_ids = list(range(start,start+batch_size))
-    print(image_ids)
 
     sample_list = []
-    for image_id, sample in zip(image_ids, samples_fake):
+    for sample in samples_fake:
         sample = torch.clamp(sample, min=-1, max=1) * 0.5 + 0.5
         sample = sample.cpu().numpy().transpose(1,2,0) * 255 
         sample = Image.fromarray(sample.astype(np.uint8))