Spaces:
Running
on
A100
Running
on
A100
Update app.py
Browse files
app.py
CHANGED
@@ -9,10 +9,10 @@ from transformers import AutoProcessor, AutoTokenizer, AutoModelForCausalLM, Aut
|
|
9 |
MODEL_ID = os.environ.get("MODEL_ID", "inference-net/ClipTagger-12b")
|
10 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
11 |
TEMP = 0.1
|
12 |
-
MAX_NEW_TOKENS = 768 #
|
13 |
DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32
|
14 |
|
15 |
-
# -------- Prompts
|
16 |
SYSTEM_PROMPT = (
|
17 |
"You are an image annotation API trained to analyze YouTube video keyframes. "
|
18 |
"You will be given instructions on the output format, what to caption, and how to perform your job. "
|
@@ -27,37 +27,33 @@ Your job is to extract detailed **factual elements directly visible** in the ima
|
|
27 |
Return JSON in this structure:
|
28 |
|
29 |
{
|
30 |
-
"description": "
|
31 |
-
"objects": ["
|
32 |
-
"actions": ["
|
33 |
-
"environment": "
|
34 |
-
"content_type": "
|
35 |
-
"specific_style": "
|
36 |
-
"production_quality": "
|
37 |
-
"summary": "
|
38 |
-
"logos": ["
|
39 |
}
|
40 |
|
41 |
Rules:
|
42 |
- Be specific and literal. Focus on what is explicitly visible.
|
43 |
- Do NOT include interpretations of emotion, mood, or narrative unless it's visually explicit.
|
44 |
- No artistic or cinematic analysis.
|
45 |
-
- Always include the language of any text in the image if present as an object
|
46 |
- Maximum 10 objects and 5 actions.
|
47 |
-
- Return
|
48 |
-
-
|
49 |
-
- Output **only the JSON**, no extra text or explanation.
|
50 |
"""
|
51 |
|
52 |
# -------- Utils --------
|
53 |
def extract_top_level_json(s: str):
|
54 |
-
"""Parse JSON; if extra text around it, extract the first balanced {...} block."""
|
55 |
-
# Fast path
|
56 |
try:
|
57 |
return json.loads(s)
|
58 |
except Exception:
|
59 |
pass
|
60 |
-
# Brace-stack extraction
|
61 |
start = None
|
62 |
depth = 0
|
63 |
for i, ch in enumerate(s):
|
@@ -73,11 +69,10 @@ def extract_top_level_json(s: str):
|
|
73 |
try:
|
74 |
return json.loads(chunk)
|
75 |
except Exception:
|
76 |
-
# continue scanning for the next candidate
|
77 |
start = None
|
78 |
return None
|
79 |
|
80 |
-
def build_messages(image
|
81 |
return [
|
82 |
{"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
|
83 |
{"role": "user", "content": [{"type": "image", "image": image},
|
@@ -85,8 +80,7 @@ def build_messages(image: Image.Image):
|
|
85 |
]
|
86 |
|
87 |
def downscale_if_huge(pil: Image.Image, max_side: int = 1792) -> Image.Image:
|
88 |
-
if pil is None:
|
89 |
-
return pil
|
90 |
w, h = pil.size
|
91 |
m = max(w, h)
|
92 |
if m <= max_side:
|
@@ -94,41 +88,26 @@ def downscale_if_huge(pil: Image.Image, max_side: int = 1792) -> Image.Image:
|
|
94 |
s = max_side / m
|
95 |
return pil.convert("RGB").resize((int(w*s), int(h*s)), Image.BICUBIC)
|
96 |
|
97 |
-
# -------- Load model
|
98 |
processor = tokenizer = model = None
|
99 |
LOAD_ERROR = None
|
100 |
-
|
101 |
try:
|
102 |
cfg = AutoConfig.from_pretrained(MODEL_ID, token=HF_TOKEN, trust_remote_code=True)
|
103 |
if "clip" in cfg.__class__.__name__.lower():
|
104 |
raise RuntimeError(f"MODEL_ID '{MODEL_ID}' is a CLIP/encoder repo; need a causal VLM.")
|
105 |
|
106 |
-
|
107 |
-
try:
|
108 |
-
processor = AutoProcessor.from_pretrained(
|
109 |
-
MODEL_ID, token=HF_TOKEN, trust_remote_code=True, use_fast=True
|
110 |
-
)
|
111 |
-
except TypeError:
|
112 |
-
processor = AutoProcessor.from_pretrained(
|
113 |
-
MODEL_ID, token=HF_TOKEN, trust_remote_code=True
|
114 |
-
)
|
115 |
-
|
116 |
-
print("[boot] loading model…", flush=True)
|
117 |
-
# Force full-precision path on A100 first; add quantized path later if desired
|
118 |
model = AutoModelForCausalLM.from_pretrained(
|
119 |
MODEL_ID,
|
120 |
token=HF_TOKEN,
|
121 |
device_map="auto",
|
122 |
torch_dtype=DTYPE,
|
123 |
trust_remote_code=True,
|
124 |
-
# quantization_config=None, #
|
125 |
)
|
126 |
-
|
127 |
tokenizer = getattr(processor, "tokenizer", None) or AutoTokenizer.from_pretrained(
|
128 |
MODEL_ID, token=HF_TOKEN, trust_remote_code=True, use_fast=True
|
129 |
)
|
130 |
-
print("[boot] ready.", flush=True)
|
131 |
-
|
132 |
except Exception as e:
|
133 |
LOAD_ERROR = f"{e}\n\n{traceback.format_exc()}"
|
134 |
|
@@ -141,82 +120,54 @@ def generate(image: Image.Image) -> Tuple[str, Dict[str, Any] | None, bool]:
|
|
141 |
|
142 |
image = downscale_if_huge(image)
|
143 |
|
144 |
-
# Build prompt
|
145 |
if hasattr(processor, "apply_chat_template"):
|
146 |
prompt = processor.apply_chat_template(build_messages(image), add_generation_prompt=True, tokenize=False)
|
147 |
else:
|
148 |
-
# fallback join (rare)
|
149 |
prompt = USER_PROMPT
|
150 |
|
151 |
-
# Tokenize with vision
|
152 |
inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
|
153 |
|
154 |
-
# Common gen kwargs
|
155 |
-
eos = getattr(model.config, "eos_token_id", None)
|
156 |
-
|
157 |
tried = []
|
158 |
-
|
159 |
-
# (1) Greedy, no sampling (most stable, no temperature arg)
|
160 |
try:
|
161 |
g = dict(do_sample=False, max_new_tokens=MAX_NEW_TOKENS)
|
|
|
162 |
if eos is not None:
|
163 |
g["eos_token_id"] = eos
|
164 |
with torch.inference_mode():
|
165 |
out = model.generate(**inputs, **g)
|
166 |
-
text =
|
167 |
-
if hasattr(processor, "decode")
|
168 |
-
else tokenizer.decode(out[0], skip_special_tokens=True))
|
169 |
parsed = extract_top_level_json(text)
|
170 |
if isinstance(parsed, dict):
|
171 |
return json.dumps(parsed, indent=2), parsed, True
|
172 |
-
tried.append(("greedy", "
|
173 |
except Exception as e:
|
174 |
tried.append(("greedy", f"err={e}"))
|
175 |
|
176 |
-
# (2) Sampling
|
177 |
try:
|
178 |
g = dict(do_sample=True, temperature=TEMP, max_new_tokens=MAX_NEW_TOKENS)
|
|
|
179 |
if eos is not None:
|
180 |
g["eos_token_id"] = eos
|
181 |
with torch.inference_mode():
|
182 |
out = model.generate(**inputs, **g)
|
183 |
-
text =
|
184 |
-
if hasattr(processor, "decode")
|
185 |
-
else tokenizer.decode(out[0], skip_special_tokens=True))
|
186 |
parsed = extract_top_level_json(text)
|
187 |
if isinstance(parsed, dict):
|
188 |
return json.dumps(parsed, indent=2), parsed, True
|
189 |
-
tried.append(("
|
190 |
except Exception as e:
|
191 |
-
tried.append(("
|
192 |
|
193 |
-
|
194 |
-
try:
|
195 |
-
g = dict(do_sample=False, max_new_tokens=min(512, MAX_NEW_TOKENS))
|
196 |
-
if eos is not None:
|
197 |
-
g["eos_token_id"] = eos
|
198 |
-
with torch.inference_mode():
|
199 |
-
out = model.generate(**inputs, **g)
|
200 |
-
text = (processor.decode(out[0], skip_special_tokens=True)
|
201 |
-
if hasattr(processor, "decode")
|
202 |
-
else tokenizer.decode(out[0], skip_special_tokens=True))
|
203 |
-
parsed = extract_top_level_json(text)
|
204 |
-
if isinstance(parsed, dict):
|
205 |
-
return json.dumps(parsed, indent=2), parsed, True
|
206 |
-
tried.append(("greedy_short", "parsed-failed"))
|
207 |
-
except Exception as e:
|
208 |
-
tried.append(("greedy_short", f"err={e}"))
|
209 |
-
|
210 |
-
# Debug info if all fail
|
211 |
-
return "Generation failed.\nTried: " + "\n".join([f"{t[0]} -> {t[1]}" for t in tried]), None, False
|
212 |
|
213 |
# -------- UI --------
|
214 |
-
with gr.Blocks(theme=gr.themes.Soft(), analytics_enabled=False, title="
|
215 |
-
gr.Markdown("#
|
216 |
if LOAD_ERROR:
|
217 |
with gr.Accordion("Startup Error Details", open=False):
|
218 |
gr.Markdown(f"```\n{LOAD_ERROR}\n```")
|
219 |
-
|
220 |
with gr.Row():
|
221 |
with gr.Column(scale=1):
|
222 |
image = gr.Image(type="pil", label="Upload Image", image_mode="RGB")
|
|
|
9 |
MODEL_ID = os.environ.get("MODEL_ID", "inference-net/ClipTagger-12b")
|
10 |
HF_TOKEN = os.environ.get("HF_TOKEN")
|
11 |
TEMP = 0.1
|
12 |
+
MAX_NEW_TOKENS = 768 # safe for demo; raise if needed
|
13 |
DTYPE = torch.bfloat16 if torch.cuda.is_available() else torch.float32
|
14 |
|
15 |
+
# -------- Prompts --------
|
16 |
SYSTEM_PROMPT = (
|
17 |
"You are an image annotation API trained to analyze YouTube video keyframes. "
|
18 |
"You will be given instructions on the output format, what to caption, and how to perform your job. "
|
|
|
27 |
Return JSON in this structure:
|
28 |
|
29 |
{
|
30 |
+
"description": "...",
|
31 |
+
"objects": ["..."],
|
32 |
+
"actions": ["..."],
|
33 |
+
"environment": "...",
|
34 |
+
"content_type": "...",
|
35 |
+
"specific_style": "...",
|
36 |
+
"production_quality": "...",
|
37 |
+
"summary": "...",
|
38 |
+
"logos": ["..."]
|
39 |
}
|
40 |
|
41 |
Rules:
|
42 |
- Be specific and literal. Focus on what is explicitly visible.
|
43 |
- Do NOT include interpretations of emotion, mood, or narrative unless it's visually explicit.
|
44 |
- No artistic or cinematic analysis.
|
45 |
+
- Always include the language of any text in the image if present as an object.
|
46 |
- Maximum 10 objects and 5 actions.
|
47 |
+
- Return [] for 'logos' if none are present.
|
48 |
+
- Strictly valid JSON only.
|
|
|
49 |
"""
|
50 |
|
51 |
# -------- Utils --------
|
52 |
def extract_top_level_json(s: str):
|
|
|
|
|
53 |
try:
|
54 |
return json.loads(s)
|
55 |
except Exception:
|
56 |
pass
|
|
|
57 |
start = None
|
58 |
depth = 0
|
59 |
for i, ch in enumerate(s):
|
|
|
69 |
try:
|
70 |
return json.loads(chunk)
|
71 |
except Exception:
|
|
|
72 |
start = None
|
73 |
return None
|
74 |
|
75 |
+
def build_messages(image):
|
76 |
return [
|
77 |
{"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
|
78 |
{"role": "user", "content": [{"type": "image", "image": image},
|
|
|
80 |
]
|
81 |
|
82 |
def downscale_if_huge(pil: Image.Image, max_side: int = 1792) -> Image.Image:
|
83 |
+
if pil is None: return pil
|
|
|
84 |
w, h = pil.size
|
85 |
m = max(w, h)
|
86 |
if m <= max_side:
|
|
|
88 |
s = max_side / m
|
89 |
return pil.convert("RGB").resize((int(w*s), int(h*s)), Image.BICUBIC)
|
90 |
|
91 |
+
# -------- Load model --------
|
92 |
processor = tokenizer = model = None
|
93 |
LOAD_ERROR = None
|
|
|
94 |
try:
|
95 |
cfg = AutoConfig.from_pretrained(MODEL_ID, token=HF_TOKEN, trust_remote_code=True)
|
96 |
if "clip" in cfg.__class__.__name__.lower():
|
97 |
raise RuntimeError(f"MODEL_ID '{MODEL_ID}' is a CLIP/encoder repo; need a causal VLM.")
|
98 |
|
99 |
+
processor = AutoProcessor.from_pretrained(MODEL_ID, token=HF_TOKEN, trust_remote_code=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
model = AutoModelForCausalLM.from_pretrained(
|
101 |
MODEL_ID,
|
102 |
token=HF_TOKEN,
|
103 |
device_map="auto",
|
104 |
torch_dtype=DTYPE,
|
105 |
trust_remote_code=True,
|
106 |
+
# quantization_config=None, # uncomment if you want to force full precision
|
107 |
)
|
|
|
108 |
tokenizer = getattr(processor, "tokenizer", None) or AutoTokenizer.from_pretrained(
|
109 |
MODEL_ID, token=HF_TOKEN, trust_remote_code=True, use_fast=True
|
110 |
)
|
|
|
|
|
111 |
except Exception as e:
|
112 |
LOAD_ERROR = f"{e}\n\n{traceback.format_exc()}"
|
113 |
|
|
|
120 |
|
121 |
image = downscale_if_huge(image)
|
122 |
|
|
|
123 |
if hasattr(processor, "apply_chat_template"):
|
124 |
prompt = processor.apply_chat_template(build_messages(image), add_generation_prompt=True, tokenize=False)
|
125 |
else:
|
|
|
126 |
prompt = USER_PROMPT
|
127 |
|
|
|
128 |
inputs = processor(text=prompt, images=image, return_tensors="pt").to(model.device)
|
129 |
|
|
|
|
|
|
|
130 |
tried = []
|
131 |
+
# (1) Greedy
|
|
|
132 |
try:
|
133 |
g = dict(do_sample=False, max_new_tokens=MAX_NEW_TOKENS)
|
134 |
+
eos = getattr(model.config, "eos_token_id", None)
|
135 |
if eos is not None:
|
136 |
g["eos_token_id"] = eos
|
137 |
with torch.inference_mode():
|
138 |
out = model.generate(**inputs, **g)
|
139 |
+
text = processor.decode(out[0], skip_special_tokens=True)
|
|
|
|
|
140 |
parsed = extract_top_level_json(text)
|
141 |
if isinstance(parsed, dict):
|
142 |
return json.dumps(parsed, indent=2), parsed, True
|
143 |
+
tried.append(("greedy", "parse-failed"))
|
144 |
except Exception as e:
|
145 |
tried.append(("greedy", f"err={e}"))
|
146 |
|
147 |
+
# (2) Sampling
|
148 |
try:
|
149 |
g = dict(do_sample=True, temperature=TEMP, max_new_tokens=MAX_NEW_TOKENS)
|
150 |
+
eos = getattr(model.config, "eos_token_id", None)
|
151 |
if eos is not None:
|
152 |
g["eos_token_id"] = eos
|
153 |
with torch.inference_mode():
|
154 |
out = model.generate(**inputs, **g)
|
155 |
+
text = processor.decode(out[0], skip_special_tokens=True)
|
|
|
|
|
156 |
parsed = extract_top_level_json(text)
|
157 |
if isinstance(parsed, dict):
|
158 |
return json.dumps(parsed, indent=2), parsed, True
|
159 |
+
tried.append(("sample", "parse-failed"))
|
160 |
except Exception as e:
|
161 |
+
tried.append(("sample", f"err={e}"))
|
162 |
|
163 |
+
return "Generation failed.\n" + str(tried), None, False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
|
165 |
# -------- UI --------
|
166 |
+
with gr.Blocks(theme=gr.themes.Soft(), analytics_enabled=False, title="ClipTagger (VLM)") as demo:
|
167 |
+
gr.Markdown("# ClipTagger\nUpload an image to get **strict JSON** annotations.")
|
168 |
if LOAD_ERROR:
|
169 |
with gr.Accordion("Startup Error Details", open=False):
|
170 |
gr.Markdown(f"```\n{LOAD_ERROR}\n```")
|
|
|
171 |
with gr.Row():
|
172 |
with gr.Column(scale=1):
|
173 |
image = gr.Image(type="pil", label="Upload Image", image_mode="RGB")
|