Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -12,7 +12,6 @@ import spaces
|
|
12 |
MODEL_ID = "Qwen/Qwen-Audio-Chat"
|
13 |
|
14 |
def load_model():
|
15 |
-
print("Loading model and tokenizer...")
|
16 |
model = AutoModelForCausalLM.from_pretrained(
|
17 |
MODEL_ID,
|
18 |
torch_dtype=torch.float16,
|
@@ -21,7 +20,6 @@ def load_model():
|
|
21 |
)
|
22 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
23 |
|
24 |
-
# Define a custom chat template
|
25 |
chat_template = """<s>[INST] <<SYS>>
|
26 |
You are a helpful assistant.
|
27 |
<</SYS>>
|
@@ -29,67 +27,39 @@ You are a helpful assistant.
|
|
29 |
{{ message['role'] }}: {{ message['content'] }}
|
30 |
{% endfor %}[/INST]"""
|
31 |
|
32 |
-
# Assign the custom chat template to the tokenizer
|
33 |
tokenizer.chat_template = chat_template
|
34 |
-
|
35 |
-
print("Model and tokenizer loaded successfully")
|
36 |
return model, tokenizer
|
37 |
|
38 |
def process_audio(audio_path):
|
39 |
-
"""Process audio file for the model."""
|
40 |
try:
|
41 |
-
print(f"Processing audio file: {audio_path}")
|
42 |
-
# Read audio file
|
43 |
audio_data, sample_rate = sf.read(audio_path)
|
44 |
|
45 |
-
# Convert to mono if stereo
|
46 |
if len(audio_data.shape) > 1:
|
47 |
audio_data = audio_data.mean(axis=1)
|
48 |
|
49 |
-
# Ensure float32 format
|
50 |
audio_data = audio_data.astype(np.float32)
|
51 |
|
52 |
-
# Create in-memory buffer
|
53 |
audio_buffer = BytesIO()
|
54 |
-
|
55 |
-
# Write audio to buffer in WAV format
|
56 |
sf.write(audio_buffer, audio_data, sample_rate, format='WAV')
|
57 |
|
58 |
-
# Get the buffer content and encode to base64
|
59 |
audio_buffer.seek(0)
|
60 |
audio_base64 = base64.b64encode(audio_buffer.read()).decode('utf-8')
|
61 |
|
62 |
-
print(f"Audio processed successfully. Sample rate: {sample_rate}, Shape: {audio_data.shape}")
|
63 |
return {
|
64 |
"audio": audio_base64,
|
65 |
"sampling_rate": sample_rate
|
66 |
}
|
67 |
-
except Exception
|
68 |
-
print(f"Error processing audio: {e}")
|
69 |
-
import traceback
|
70 |
-
traceback.print_exc()
|
71 |
return None
|
72 |
|
73 |
@spaces.GPU
|
74 |
def analyze_audio(audio_path: str, question: str = None) -> str:
|
75 |
-
"""
|
76 |
-
Main function for audio analysis that will be exposed as a tool.
|
77 |
-
Args:
|
78 |
-
audio_path: Path to the audio file
|
79 |
-
question: Optional question about the audio
|
80 |
-
Returns:
|
81 |
-
str: Model's response about the audio
|
82 |
-
"""
|
83 |
-
print(f"\nStarting analysis with audio_path: {audio_path}, question: {question}")
|
84 |
-
|
85 |
-
# Input validation
|
86 |
if audio_path is None or not isinstance(audio_path, str):
|
87 |
return "Please provide a valid audio file."
|
88 |
|
89 |
if not os.path.exists(audio_path):
|
90 |
return f"Audio file not found: {audio_path}"
|
91 |
|
92 |
-
# Process audio
|
93 |
audio_data = process_audio(audio_path)
|
94 |
if not audio_data or "audio" not in audio_data or "sampling_rate" not in audio_data:
|
95 |
return "Failed to process the audio file. Please ensure it's a valid audio format."
|
@@ -98,7 +68,6 @@ def analyze_audio(audio_path: str, question: str = None) -> str:
|
|
98 |
model, tokenizer = load_model()
|
99 |
query = question if question else "Please describe what you hear in this audio clip."
|
100 |
|
101 |
-
print("Preparing messages...")
|
102 |
messages = [
|
103 |
{
|
104 |
"role": "user",
|
@@ -106,7 +75,6 @@ def analyze_audio(audio_path: str, question: str = None) -> str:
|
|
106 |
}
|
107 |
]
|
108 |
|
109 |
-
print("Applying chat template...")
|
110 |
if tokenizer.chat_template:
|
111 |
text = tokenizer.apply_chat_template(
|
112 |
messages,
|
@@ -116,12 +84,8 @@ def analyze_audio(audio_path: str, question: str = None) -> str:
|
|
116 |
else:
|
117 |
raise ValueError("Tokenizer chat_template is not set.")
|
118 |
|
119 |
-
print(f"Generated prompt text: {text[:200]}...")
|
120 |
-
|
121 |
-
print("Tokenizing input...")
|
122 |
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
123 |
|
124 |
-
print("Generating response...")
|
125 |
with torch.no_grad():
|
126 |
outputs = model.generate(
|
127 |
**model_inputs,
|
@@ -134,24 +98,14 @@ def analyze_audio(audio_path: str, question: str = None) -> str:
|
|
134 |
)
|
135 |
|
136 |
if outputs is None or len(outputs) == 0:
|
137 |
-
print("Model generated None or empty output")
|
138 |
return "The model failed to generate a response. Please try again."
|
139 |
|
140 |
-
print(f"Output shape: {outputs.shape}")
|
141 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
142 |
-
print(f"Generated response: {response[:200]}...")
|
143 |
return response
|
144 |
|
145 |
-
except
|
146 |
-
|
147 |
-
return "An error occurred with the data processing. Please check the inputs."
|
148 |
-
except Exception as e:
|
149 |
-
print(f"Error during processing: {str(e)}")
|
150 |
-
import traceback
|
151 |
-
traceback.print_exc()
|
152 |
-
return f"An error occurred while processing: {str(e)}"
|
153 |
|
154 |
-
# Create Gradio interface with clear input/output specifications
|
155 |
demo = gr.Interface(
|
156 |
fn=analyze_audio,
|
157 |
inputs=[
|
@@ -159,7 +113,7 @@ demo = gr.Interface(
|
|
159 |
type="filepath",
|
160 |
label="Audio Input",
|
161 |
sources=["upload", "microphone"],
|
162 |
-
format="mp3"
|
163 |
),
|
164 |
gr.Textbox(
|
165 |
label="Question",
|
|
|
12 |
MODEL_ID = "Qwen/Qwen-Audio-Chat"
|
13 |
|
14 |
def load_model():
|
|
|
15 |
model = AutoModelForCausalLM.from_pretrained(
|
16 |
MODEL_ID,
|
17 |
torch_dtype=torch.float16,
|
|
|
20 |
)
|
21 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True)
|
22 |
|
|
|
23 |
chat_template = """<s>[INST] <<SYS>>
|
24 |
You are a helpful assistant.
|
25 |
<</SYS>>
|
|
|
27 |
{{ message['role'] }}: {{ message['content'] }}
|
28 |
{% endfor %}[/INST]"""
|
29 |
|
|
|
30 |
tokenizer.chat_template = chat_template
|
|
|
|
|
31 |
return model, tokenizer
|
32 |
|
33 |
def process_audio(audio_path):
|
|
|
34 |
try:
|
|
|
|
|
35 |
audio_data, sample_rate = sf.read(audio_path)
|
36 |
|
|
|
37 |
if len(audio_data.shape) > 1:
|
38 |
audio_data = audio_data.mean(axis=1)
|
39 |
|
|
|
40 |
audio_data = audio_data.astype(np.float32)
|
41 |
|
|
|
42 |
audio_buffer = BytesIO()
|
|
|
|
|
43 |
sf.write(audio_buffer, audio_data, sample_rate, format='WAV')
|
44 |
|
|
|
45 |
audio_buffer.seek(0)
|
46 |
audio_base64 = base64.b64encode(audio_buffer.read()).decode('utf-8')
|
47 |
|
|
|
48 |
return {
|
49 |
"audio": audio_base64,
|
50 |
"sampling_rate": sample_rate
|
51 |
}
|
52 |
+
except Exception:
|
|
|
|
|
|
|
53 |
return None
|
54 |
|
55 |
@spaces.GPU
|
56 |
def analyze_audio(audio_path: str, question: str = None) -> str:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
if audio_path is None or not isinstance(audio_path, str):
|
58 |
return "Please provide a valid audio file."
|
59 |
|
60 |
if not os.path.exists(audio_path):
|
61 |
return f"Audio file not found: {audio_path}"
|
62 |
|
|
|
63 |
audio_data = process_audio(audio_path)
|
64 |
if not audio_data or "audio" not in audio_data or "sampling_rate" not in audio_data:
|
65 |
return "Failed to process the audio file. Please ensure it's a valid audio format."
|
|
|
68 |
model, tokenizer = load_model()
|
69 |
query = question if question else "Please describe what you hear in this audio clip."
|
70 |
|
|
|
71 |
messages = [
|
72 |
{
|
73 |
"role": "user",
|
|
|
75 |
}
|
76 |
]
|
77 |
|
|
|
78 |
if tokenizer.chat_template:
|
79 |
text = tokenizer.apply_chat_template(
|
80 |
messages,
|
|
|
84 |
else:
|
85 |
raise ValueError("Tokenizer chat_template is not set.")
|
86 |
|
|
|
|
|
|
|
87 |
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
|
88 |
|
|
|
89 |
with torch.no_grad():
|
90 |
outputs = model.generate(
|
91 |
**model_inputs,
|
|
|
98 |
)
|
99 |
|
100 |
if outputs is None or len(outputs) == 0:
|
|
|
101 |
return "The model failed to generate a response. Please try again."
|
102 |
|
|
|
103 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
|
|
104 |
return response
|
105 |
|
106 |
+
except Exception:
|
107 |
+
return "An error occurred while processing. Please check your inputs and try again."
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
|
|
|
109 |
demo = gr.Interface(
|
110 |
fn=analyze_audio,
|
111 |
inputs=[
|
|
|
113 |
type="filepath",
|
114 |
label="Audio Input",
|
115 |
sources=["upload", "microphone"],
|
116 |
+
format="mp3"
|
117 |
),
|
118 |
gr.Textbox(
|
119 |
label="Question",
|