Spaces:
Running
Running
update configs
Browse files
configs/remodiffuse/remodiffuse_kit.py
CHANGED
|
@@ -31,37 +31,6 @@ text_latent_dim = 256
|
|
| 31 |
ff_size = 1024
|
| 32 |
num_heads = 8
|
| 33 |
dropout = 0
|
| 34 |
-
|
| 35 |
-
def scale_func(timestep):
|
| 36 |
-
import random
|
| 37 |
-
w = (1 - (1000 - timestep) / 1000) * 4.0 + 1
|
| 38 |
-
if timestep > 100:
|
| 39 |
-
if random.randint(0, 1) == 0:
|
| 40 |
-
output = {
|
| 41 |
-
'both_coef': w,
|
| 42 |
-
'text_coef': 0,
|
| 43 |
-
'retr_coef': 1 - w,
|
| 44 |
-
'none_coef': 0
|
| 45 |
-
}
|
| 46 |
-
else:
|
| 47 |
-
output = {
|
| 48 |
-
'both_coef': 0,
|
| 49 |
-
'text_coef': w,
|
| 50 |
-
'retr_coef': 0,
|
| 51 |
-
'none_coef': 1 - w
|
| 52 |
-
}
|
| 53 |
-
else:
|
| 54 |
-
both_coef = 0.78123
|
| 55 |
-
text_coef = 0.39284
|
| 56 |
-
retr_coef = -0.12475
|
| 57 |
-
none_coef = 1 - both_coef - text_coef - retr_coef
|
| 58 |
-
output = {
|
| 59 |
-
'both_coef': both_coef,
|
| 60 |
-
'text_coef': text_coef,
|
| 61 |
-
'retr_coef': retr_coef,
|
| 62 |
-
'none_coef': none_coef
|
| 63 |
-
}
|
| 64 |
-
return output
|
| 65 |
|
| 66 |
# model settings
|
| 67 |
model = dict(
|
|
@@ -121,7 +90,12 @@ model = dict(
|
|
| 121 |
dropout=dropout
|
| 122 |
),
|
| 123 |
),
|
| 124 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
),
|
| 126 |
loss_recon=dict(type='MSELoss', loss_weight=1, reduction='none'),
|
| 127 |
diffusion_train=dict(
|
|
|
|
| 31 |
ff_size = 1024
|
| 32 |
num_heads = 8
|
| 33 |
dropout = 0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
# model settings
|
| 36 |
model = dict(
|
|
|
|
| 90 |
dropout=dropout
|
| 91 |
),
|
| 92 |
),
|
| 93 |
+
scale_func_cfg=dict(
|
| 94 |
+
coarse_scale=4.0,
|
| 95 |
+
both_coef=0.78123,
|
| 96 |
+
text_coef=0.39284,
|
| 97 |
+
retr_coef=-0.12475
|
| 98 |
+
)
|
| 99 |
),
|
| 100 |
loss_recon=dict(type='MSELoss', loss_weight=1, reduction='none'),
|
| 101 |
diffusion_train=dict(
|
configs/remodiffuse/remodiffuse_t2m.py
CHANGED
|
@@ -32,37 +32,6 @@ ff_size = 1024
|
|
| 32 |
num_heads = 8
|
| 33 |
dropout = 0
|
| 34 |
|
| 35 |
-
def scale_func(timestep):
|
| 36 |
-
import random
|
| 37 |
-
w = (1 - (1000 - timestep) / 1000) * 6.5 + 1
|
| 38 |
-
if timestep > 100:
|
| 39 |
-
if random.randint(0, 1) == 0:
|
| 40 |
-
output = {
|
| 41 |
-
'both_coef': w,
|
| 42 |
-
'text_coef': 0,
|
| 43 |
-
'retr_coef': 1 - w,
|
| 44 |
-
'none_coef': 0
|
| 45 |
-
}
|
| 46 |
-
else:
|
| 47 |
-
output = {
|
| 48 |
-
'both_coef': 0,
|
| 49 |
-
'text_coef': w,
|
| 50 |
-
'retr_coef': 0,
|
| 51 |
-
'none_coef': 1 - w
|
| 52 |
-
}
|
| 53 |
-
else:
|
| 54 |
-
both_coef = 0.52351
|
| 55 |
-
text_coef = -0.28419
|
| 56 |
-
retr_coef = 2.39872
|
| 57 |
-
none_coef = 1 - both_coef - text_coef - retr_coef
|
| 58 |
-
output = {
|
| 59 |
-
'both_coef': both_coef,
|
| 60 |
-
'text_coef': text_coef,
|
| 61 |
-
'retr_coef': retr_coef,
|
| 62 |
-
'none_coef': none_coef
|
| 63 |
-
}
|
| 64 |
-
return output
|
| 65 |
-
|
| 66 |
# model settings
|
| 67 |
model = dict(
|
| 68 |
type='MotionDiffusion',
|
|
@@ -121,7 +90,12 @@ model = dict(
|
|
| 121 |
dropout=dropout
|
| 122 |
),
|
| 123 |
),
|
| 124 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 125 |
),
|
| 126 |
loss_recon=dict(type='MSELoss', loss_weight=1, reduction='none'),
|
| 127 |
diffusion_train=dict(
|
|
|
|
| 32 |
num_heads = 8
|
| 33 |
dropout = 0
|
| 34 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
# model settings
|
| 36 |
model = dict(
|
| 37 |
type='MotionDiffusion',
|
|
|
|
| 90 |
dropout=dropout
|
| 91 |
),
|
| 92 |
),
|
| 93 |
+
scale_func_cfg=dict(
|
| 94 |
+
coarse_scale=6.5,
|
| 95 |
+
both_coef=0.52351,
|
| 96 |
+
text_coef=-0.28419,
|
| 97 |
+
retr_coef=2.39872
|
| 98 |
+
)
|
| 99 |
),
|
| 100 |
loss_recon=dict(type='MSELoss', loss_weight=1, reduction='none'),
|
| 101 |
diffusion_train=dict(
|