ryanzhangfan
commited on
Commit
•
66ecdd5
1
Parent(s):
0f8e8b9
Update app.py
Browse files
app.py
CHANGED
@@ -39,15 +39,6 @@ gen_model = AutoModelForCausalLM.from_pretrained(
|
|
39 |
trust_remote_code=True,
|
40 |
)
|
41 |
|
42 |
-
gen_tokenizer = AutoTokenizer.from_pretrained(EMU_GEN_HUB, trust_remote_code=True)
|
43 |
-
gen_image_processor = AutoImageProcessor.from_pretrained(
|
44 |
-
VQ_HUB, trust_remote_code=True
|
45 |
-
)
|
46 |
-
gen_image_tokenizer = AutoModel.from_pretrained(
|
47 |
-
VQ_HUB, device_map="cuda:0", trust_remote_code=True
|
48 |
-
).eval()
|
49 |
-
gen_processor = Emu3Processor(gen_image_processor, gen_image_tokenizer, gen_tokenizer)
|
50 |
-
|
51 |
# Emu3-Chat model and processor
|
52 |
chat_model = AutoModelForCausalLM.from_pretrained(
|
53 |
EMU_CHAT_HUB,
|
@@ -57,18 +48,18 @@ chat_model = AutoModelForCausalLM.from_pretrained(
|
|
57 |
trust_remote_code=True,
|
58 |
)
|
59 |
|
60 |
-
|
61 |
-
|
62 |
VQ_HUB, trust_remote_code=True
|
63 |
)
|
64 |
-
|
65 |
VQ_HUB, device_map="cuda:0", trust_remote_code=True
|
66 |
).eval()
|
67 |
-
|
68 |
-
|
69 |
)
|
70 |
|
71 |
-
@spaces.GPU(duration=
|
72 |
def generate_image(prompt):
|
73 |
POSITIVE_PROMPT = " masterpiece, film grained, best quality."
|
74 |
NEGATIVE_PROMPT = (
|
@@ -86,8 +77,8 @@ def generate_image(prompt):
|
|
86 |
image_area=gen_model.config.image_area,
|
87 |
return_tensors="pt",
|
88 |
)
|
89 |
-
pos_inputs =
|
90 |
-
neg_inputs =
|
91 |
|
92 |
# Prepare hyperparameters
|
93 |
GENERATION_CONFIG = GenerationConfig(
|
@@ -100,7 +91,7 @@ def generate_image(prompt):
|
|
100 |
)
|
101 |
|
102 |
h, w = pos_inputs.image_size[0]
|
103 |
-
constrained_fn =
|
104 |
logits_processor = LogitsProcessorList(
|
105 |
[
|
106 |
UnbatchedClassifierFreeGuidanceLogitsProcessor(
|
@@ -122,14 +113,14 @@ def generate_image(prompt):
|
|
122 |
logits_processor=logits_processor,
|
123 |
)
|
124 |
|
125 |
-
mm_list =
|
126 |
for idx, im in enumerate(mm_list):
|
127 |
if isinstance(im, Image.Image):
|
128 |
return im
|
129 |
return None
|
130 |
|
131 |
def vision_language_understanding(image, text):
|
132 |
-
inputs =
|
133 |
text=text,
|
134 |
image=image,
|
135 |
mode="U",
|
@@ -154,7 +145,7 @@ def vision_language_understanding(image, text):
|
|
154 |
)
|
155 |
|
156 |
outputs = outputs[:, inputs.input_ids.shape[-1] :]
|
157 |
-
response =
|
158 |
return response
|
159 |
|
160 |
def chat(history, user_input, user_image):
|
|
|
39 |
trust_remote_code=True,
|
40 |
)
|
41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
# Emu3-Chat model and processor
|
43 |
chat_model = AutoModelForCausalLM.from_pretrained(
|
44 |
EMU_CHAT_HUB,
|
|
|
48 |
trust_remote_code=True,
|
49 |
)
|
50 |
|
51 |
+
tokenizer = AutoTokenizer.from_pretrained(EMU_CHAT_HUB, trust_remote_code=True)
|
52 |
+
image_processor = AutoImageProcessor.from_pretrained(
|
53 |
VQ_HUB, trust_remote_code=True
|
54 |
)
|
55 |
+
image_tokenizer = AutoModel.from_pretrained(
|
56 |
VQ_HUB, device_map="cuda:0", trust_remote_code=True
|
57 |
).eval()
|
58 |
+
processor = Emu3Processor(
|
59 |
+
image_processor, image_tokenizer, tokenizer
|
60 |
)
|
61 |
|
62 |
+
@spaces.GPU(duration=300)
|
63 |
def generate_image(prompt):
|
64 |
POSITIVE_PROMPT = " masterpiece, film grained, best quality."
|
65 |
NEGATIVE_PROMPT = (
|
|
|
77 |
image_area=gen_model.config.image_area,
|
78 |
return_tensors="pt",
|
79 |
)
|
80 |
+
pos_inputs = processor(text=full_prompt, **kwargs)
|
81 |
+
neg_inputs = processor(text=NEGATIVE_PROMPT, **kwargs)
|
82 |
|
83 |
# Prepare hyperparameters
|
84 |
GENERATION_CONFIG = GenerationConfig(
|
|
|
91 |
)
|
92 |
|
93 |
h, w = pos_inputs.image_size[0]
|
94 |
+
constrained_fn = processor.build_prefix_constrained_fn(h, w)
|
95 |
logits_processor = LogitsProcessorList(
|
96 |
[
|
97 |
UnbatchedClassifierFreeGuidanceLogitsProcessor(
|
|
|
113 |
logits_processor=logits_processor,
|
114 |
)
|
115 |
|
116 |
+
mm_list = processor.decode(outputs[0])
|
117 |
for idx, im in enumerate(mm_list):
|
118 |
if isinstance(im, Image.Image):
|
119 |
return im
|
120 |
return None
|
121 |
|
122 |
def vision_language_understanding(image, text):
|
123 |
+
inputs = processor(
|
124 |
text=text,
|
125 |
image=image,
|
126 |
mode="U",
|
|
|
145 |
)
|
146 |
|
147 |
outputs = outputs[:, inputs.input_ids.shape[-1] :]
|
148 |
+
response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
|
149 |
return response
|
150 |
|
151 |
def chat(history, user_input, user_image):
|