jadechoghari
commited on
Commit
•
80a6dc1
1
Parent(s):
1df7b61
add safetensors
Browse files- pipeline.py +10 -10
pipeline.py
CHANGED
@@ -6,8 +6,7 @@ import sys
|
|
6 |
from huggingface_hub import hf_hub_download
|
7 |
from safetensors.torch import load_file
|
8 |
import os
|
9 |
-
from
|
10 |
-
from PIL import Image
|
11 |
from .vae import AutoencoderKL
|
12 |
from .mar import mar_base, mar_large, mar_huge
|
13 |
|
@@ -46,20 +45,22 @@ class MARModel(DiffusionPipeline):
|
|
46 |
if model_type == "mar_base":
|
47 |
diffloss_d = 6
|
48 |
diffloss_w = 1024
|
|
|
49 |
elif model_type == "mar_large":
|
50 |
diffloss_d = 8
|
51 |
diffloss_w = 1280
|
|
|
52 |
elif model_type == "mar_huge":
|
53 |
diffloss_d = 12
|
54 |
diffloss_w = 1536
|
|
|
55 |
else:
|
56 |
raise NotImplementedError
|
57 |
# download and load the model weights (.safetensors or .pth)
|
58 |
model_checkpoint_path = hf_hub_download(
|
59 |
repo_id=kwargs.get("repo_id", "jadechoghari/mar"),
|
60 |
-
filename=kwargs.get("model_filename",
|
61 |
)
|
62 |
-
model_checkpoint_path = kwargs.get("model_checkpoint_path", "./mar/checkpoint-last.pth")
|
63 |
|
64 |
model_fn = model_mapping[model_type]
|
65 |
|
@@ -70,7 +71,8 @@ class MARModel(DiffusionPipeline):
|
|
70 |
num_sampling_steps=str(num_sampling_steps_diffloss)
|
71 |
).cuda()
|
72 |
|
73 |
-
|
|
|
74 |
model.load_state_dict(state_dict)
|
75 |
model.eval()
|
76 |
|
@@ -85,7 +87,7 @@ class MARModel(DiffusionPipeline):
|
|
85 |
vae = vae.to(device).eval()
|
86 |
|
87 |
# set up user-specified or default values for generation
|
88 |
-
seed = kwargs.get("seed",
|
89 |
torch.manual_seed(seed)
|
90 |
np.random.seed(seed)
|
91 |
|
@@ -93,9 +95,7 @@ class MARModel(DiffusionPipeline):
|
|
93 |
cfg_scale = kwargs.get("cfg_scale", 4)
|
94 |
cfg_schedule = kwargs.get("cfg_schedule", "constant")
|
95 |
temperature = kwargs.get("temperature", 1.0)
|
96 |
-
|
97 |
-
class_labels = 207, 360, 388, 113, 355, 980, 323, 979
|
98 |
-
print("the labels", class_labels)
|
99 |
|
100 |
# generate the tokens and images
|
101 |
with torch.cuda.amp.autocast():
|
@@ -113,7 +113,7 @@ class MARModel(DiffusionPipeline):
|
|
113 |
|
114 |
# save the images
|
115 |
image_path = os.path.join(output_dir, "sampled_image.png")
|
116 |
-
samples_per_row = kwargs.get("samples_per_row",
|
117 |
|
118 |
save_image(
|
119 |
sampled_images, image_path, nrow=int(samples_per_row), normalize=True, value_range=(-1, 1)
|
|
|
6 |
from huggingface_hub import hf_hub_download
|
7 |
from safetensors.torch import load_file
|
8 |
import os
|
9 |
+
from safetensors.torch import load_file
|
|
|
10 |
from .vae import AutoencoderKL
|
11 |
from .mar import mar_base, mar_large, mar_huge
|
12 |
|
|
|
45 |
if model_type == "mar_base":
|
46 |
diffloss_d = 6
|
47 |
diffloss_w = 1024
|
48 |
+
model_path = "mar-base.safetensors"
|
49 |
elif model_type == "mar_large":
|
50 |
diffloss_d = 8
|
51 |
diffloss_w = 1280
|
52 |
+
model_path = "mar-large.safetensors"
|
53 |
elif model_type == "mar_huge":
|
54 |
diffloss_d = 12
|
55 |
diffloss_w = 1536
|
56 |
+
model_path = "mar-huge.safetensors"
|
57 |
else:
|
58 |
raise NotImplementedError
|
59 |
# download and load the model weights (.safetensors or .pth)
|
60 |
model_checkpoint_path = hf_hub_download(
|
61 |
repo_id=kwargs.get("repo_id", "jadechoghari/mar"),
|
62 |
+
filename=kwargs.get("model_filename", model_path)
|
63 |
)
|
|
|
64 |
|
65 |
model_fn = model_mapping[model_type]
|
66 |
|
|
|
71 |
num_sampling_steps=str(num_sampling_steps_diffloss)
|
72 |
).cuda()
|
73 |
|
74 |
+
# use safetensors
|
75 |
+
state_dict = load_file(safetensors_path)
|
76 |
model.load_state_dict(state_dict)
|
77 |
model.eval()
|
78 |
|
|
|
87 |
vae = vae.to(device).eval()
|
88 |
|
89 |
# set up user-specified or default values for generation
|
90 |
+
seed = kwargs.get("seed", 6)
|
91 |
torch.manual_seed(seed)
|
92 |
np.random.seed(seed)
|
93 |
|
|
|
95 |
cfg_scale = kwargs.get("cfg_scale", 4)
|
96 |
cfg_schedule = kwargs.get("cfg_schedule", "constant")
|
97 |
temperature = kwargs.get("temperature", 1.0)
|
98 |
+
class_labels = kwargs.get("class_labels", 207, 360, 388, 113, 355, 980, 323, 979)
|
|
|
|
|
99 |
|
100 |
# generate the tokens and images
|
101 |
with torch.cuda.amp.autocast():
|
|
|
113 |
|
114 |
# save the images
|
115 |
image_path = os.path.join(output_dir, "sampled_image.png")
|
116 |
+
samples_per_row = kwargs.get("samples_per_row", 4)
|
117 |
|
118 |
save_image(
|
119 |
sampled_images, image_path, nrow=int(samples_per_row), normalize=True, value_range=(-1, 1)
|