Spaces:
Runtime error
Runtime error
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,351 +1,351 @@
|
|
| 1 |
-
import os
|
| 2 |
-
import torch
|
| 3 |
-
import numpy as np
|
| 4 |
-
import gradio as gr
|
| 5 |
-
import zipfile
|
| 6 |
-
import json
|
| 7 |
-
import requests
|
| 8 |
-
import subprocess
|
| 9 |
-
import shutil
|
| 10 |
-
from transformers import BlipProcessor, BlipForConditionalGeneration
|
| 11 |
-
|
| 12 |
-
title = "# 🗜️ CLaMP 3 - Multimodal & Multilingual Semantic Music Search"
|
| 13 |
-
|
| 14 |
-
badges = """
|
| 15 |
-
<div style="text-align: center;">
|
| 16 |
-
<a href="https://sanderwood.github.io/clamp3/">
|
| 17 |
-
<img src="https://img.shields.io/badge/CLaMP%203%20Homepage-GitHub-181717?style=for-the-badge&logo=home-assistant" alt="Homepage">
|
| 18 |
-
</a>
|
| 19 |
-
<a href="https://arxiv.org/abs/2502.10362">
|
| 20 |
-
<img src="https://img.shields.io/badge/CLaMP%203%20Paper-Arxiv-red?style=for-the-badge&logo=arxiv" alt="Paper">
|
| 21 |
-
</a>
|
| 22 |
-
<a href="https://github.com/sanderwood/clamp3">
|
| 23 |
-
<img src="https://img.shields.io/badge/CLaMP%203%20Code-GitHub-181717?style=for-the-badge&logo=github" alt="GitHub">
|
| 24 |
-
</a>
|
| 25 |
-
<a href="https://huggingface.co/spaces/sander-wood/clamp3">
|
| 26 |
-
<img src="https://img.shields.io/badge/CLaMP%203%20Demo-Gradio-green?style=for-the-badge&logo=gradio" alt="Demo">
|
| 27 |
-
</a>
|
| 28 |
-
<a href="https://huggingface.co/sander-wood/clamp3/tree/main">
|
| 29 |
-
<img src="https://img.shields.io/badge/Model%20Weights-Hugging%20Face-ffcc00?style=for-the-badge&logo=huggingface" alt="Model Weights">
|
| 30 |
-
</a>
|
| 31 |
-
<a href="https://huggingface.co/datasets/sander-wood/m4-rag">
|
| 32 |
-
<img src="https://img.shields.io/badge/M4--RAG%20Dataset-Hugging%20Face-ffcc00?style=for-the-badge&logo=huggingface" alt="Dataset">
|
| 33 |
-
</a>
|
| 34 |
-
<a href="https://huggingface.co/datasets/sander-wood/wikimt-x">
|
| 35 |
-
<img src="https://img.shields.io/badge/WikiMT--X%20Benchmark-Hugging%20Face-ffcc00?style=for-the-badge&logo=huggingface" alt="Benchmark">
|
| 36 |
-
</a>
|
| 37 |
-
</div>
|
| 38 |
-
|
| 39 |
-
<style>
|
| 40 |
-
div a {
|
| 41 |
-
display: inline-block;
|
| 42 |
-
margin: 5px;
|
| 43 |
-
}
|
| 44 |
-
div a img {
|
| 45 |
-
height: 30px;
|
| 46 |
-
}
|
| 47 |
-
</style>
|
| 48 |
-
"""
|
| 49 |
-
|
| 50 |
-
description = """CLaMP 3 is a **multimodal and multilingual** music information retrieval (MIR) framework, supporting **sheet music, audio, and performance signals** in **100 languages**. Using **contrastive learning**, it aligns these modalities in a shared space for **cross-modal retrieval**.
|
| 51 |
-
|
| 52 |
-
### 🔍 **How This Demo Works**
|
| 53 |
-
- You can **retrieve music using any text input (in any language) or an image** (`.png`, `.jpg`).
|
| 54 |
-
- When using an image, **BLIP** generates a caption, which is then used for retrieval.
|
| 55 |
-
- Since CLaMP 3's training data includes **rich visual descriptions of musical scenes**, it can **match images to semantically relevant music**.
|
| 56 |
-
|
| 57 |
-
### ⚠️ **Limitations**
|
| 58 |
-
- This demo retrieves music **only from the WikiMT-X benchmark (1,000 pieces)**.
|
| 59 |
-
- These pieces are **mainly from the U.S. and Western Europe (especially the U.S.)** and **mostly from the 20th century**.
|
| 60 |
-
- Thus, retrieval results are **mostly limited to Western 20th-century music**, so you **won’t** find music from **other regions or historical periods**.
|
| 61 |
-
|
| 62 |
-
🔧 **Need retrieval for a different music collection?** Deploy **[CLaMP 3](https://github.com/sanderwood/clamp3)** on your own dataset.
|
| 63 |
-
Generally, the larger and more diverse the reference music dataset, the better the retrieval quality, increasing the likelihood of finding relevant and accurately matched music.
|
| 64 |
-
|
| 65 |
-
**Note: This project is for research use only.**
|
| 66 |
-
"""
|
| 67 |
-
|
| 68 |
-
# Load BLIP image captioning model and processor
|
| 69 |
-
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 70 |
-
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 71 |
-
|
| 72 |
-
# Download weight file if it does not exist
|
| 73 |
-
weights_url = "https://huggingface.co/sander-wood/clamp3/resolve/main/weights_clamp3_saas_h_size_768_t_model_FacebookAI_xlm-roberta-base_t_length_128_a_size_768_a_layers_12_a_length_128_s_size_768_s_layers_12_p_size_64_p_length_512.pth"
|
| 74 |
-
weights_filename = "weights_clamp3_saas_h_size_768_t_model_FacebookAI_xlm-roberta-base_t_length_128_a_size_768_a_layers_12_a_length_128_s_size_768_s_layers_12_p_size_64_p_length_512.pth"
|
| 75 |
-
|
| 76 |
-
if not os.path.exists(weights_filename):
|
| 77 |
-
print("Downloading weights file...")
|
| 78 |
-
response = requests.get(weights_url, stream=True)
|
| 79 |
-
response.raise_for_status()
|
| 80 |
-
with open(weights_filename, "wb") as f:
|
| 81 |
-
for chunk in response.iter_content(chunk_size=8192):
|
| 82 |
-
if chunk:
|
| 83 |
-
f.write(chunk)
|
| 84 |
-
print("Weights file downloaded.")
|
| 85 |
-
|
| 86 |
-
ZIP_PATH = "features.zip"
|
| 87 |
-
if os.path.exists(ZIP_PATH):
|
| 88 |
-
print(f"Extracting {ZIP_PATH}...")
|
| 89 |
-
with zipfile.ZipFile(ZIP_PATH, "r") as zip_ref:
|
| 90 |
-
zip_ref.extractall(".")
|
| 91 |
-
print("Extraction complete.")
|
| 92 |
-
|
| 93 |
-
# Load metadata
|
| 94 |
-
metadata_map = {}
|
| 95 |
-
METADATA_FILE = "wikimt-x-public.jsonl"
|
| 96 |
-
if os.path.exists(METADATA_FILE):
|
| 97 |
-
with open(METADATA_FILE, "r", encoding="utf-8") as f:
|
| 98 |
-
for line in f:
|
| 99 |
-
data = json.loads(line)
|
| 100 |
-
metadata_map[data["id"]] = data
|
| 101 |
-
else:
|
| 102 |
-
print(f"Warning: {METADATA_FILE} not found.")
|
| 103 |
-
|
| 104 |
-
features_cache = {}
|
| 105 |
-
|
| 106 |
-
def get_info(folder_path):
|
| 107 |
-
"""
|
| 108 |
-
Load all .npy files from the specified folder and return a dictionary
|
| 109 |
-
with the file names (without extension) as keys.
|
| 110 |
-
"""
|
| 111 |
-
if folder_path in features_cache:
|
| 112 |
-
return features_cache[folder_path]
|
| 113 |
-
if not os.path.exists(folder_path):
|
| 114 |
-
return {}
|
| 115 |
-
files = sorted(os.listdir(folder_path))
|
| 116 |
-
features = {}
|
| 117 |
-
for file in files:
|
| 118 |
-
if file.endswith(".npy"):
|
| 119 |
-
key = file.split(".")[0]
|
| 120 |
-
try:
|
| 121 |
-
features[key] = np.load(os.path.join(folder_path, file))[0]
|
| 122 |
-
except Exception as e:
|
| 123 |
-
print(f"Error loading {file}: {e}")
|
| 124 |
-
features_cache[folder_path] = features
|
| 125 |
-
return features
|
| 126 |
-
|
| 127 |
-
def find_top_similar(query_file, reference_folder):
|
| 128 |
-
"""
|
| 129 |
-
Compare the query feature with all reference features in the specified folder
|
| 130 |
-
using cosine similarity and return the top 10 candidate results in the format:
|
| 131 |
-
Title | Artists | sim: SimilarityScore.
|
| 132 |
-
"""
|
| 133 |
-
top_k = 10
|
| 134 |
-
try:
|
| 135 |
-
query_feature = np.load(query_file.name)[0]
|
| 136 |
-
except Exception as e:
|
| 137 |
-
return [], f"Error loading query feature: {e}"
|
| 138 |
-
query_tensor = torch.tensor(query_feature, dtype=torch.float32).unsqueeze(dim=0)
|
| 139 |
-
key_features = get_info(reference_folder)
|
| 140 |
-
if not key_features:
|
| 141 |
-
return [], f"No reference features found in {reference_folder}."
|
| 142 |
-
ref_keys = list(key_features.keys())
|
| 143 |
-
ref_array = np.array([key_features[k] for k in ref_keys])
|
| 144 |
-
key_feats_tensor = torch.tensor(ref_array, dtype=torch.float32)
|
| 145 |
-
query_tensor_expanded = query_tensor.expand(key_feats_tensor.size(0), -1)
|
| 146 |
-
similarities = torch.cosine_similarity(query_tensor_expanded, key_feats_tensor, dim=1)
|
| 147 |
-
ranked_indices = torch.argsort(similarities, descending=True)
|
| 148 |
-
candidate_ids = []
|
| 149 |
-
candidate_display = []
|
| 150 |
-
for i in range(top_k):
|
| 151 |
-
if i < len(ref_keys):
|
| 152 |
-
candidate_idx = ranked_indices[i].item()
|
| 153 |
-
candidate_id = ref_keys[candidate_idx]
|
| 154 |
-
sim = round(similarities[candidate_idx].item(), 4)
|
| 155 |
-
meta = metadata_map.get(candidate_id, {})
|
| 156 |
-
title = meta.get("title", candidate_id)
|
| 157 |
-
artists = meta.get("artists", "Unknown")
|
| 158 |
-
if isinstance(artists, list):
|
| 159 |
-
artists = ", ".join(artists)
|
| 160 |
-
candidate_ids.append(candidate_id)
|
| 161 |
-
candidate_display.append(f"{title} | {artists} | sim: {sim}")
|
| 162 |
-
else:
|
| 163 |
-
candidate_ids.append("N/A")
|
| 164 |
-
candidate_display.append("N/A")
|
| 165 |
-
return candidate_ids, candidate_display
|
| 166 |
-
|
| 167 |
-
def show_details(selected_id):
|
| 168 |
-
"""
|
| 169 |
-
Return detailed metadata and embedded YouTube video HTML based on the candidate ID.
|
| 170 |
-
"""
|
| 171 |
-
if selected_id == "N/A":
|
| 172 |
-
return ("", "", "", "", "", "", "", "")
|
| 173 |
-
data = metadata_map.get(selected_id, {})
|
| 174 |
-
if not data:
|
| 175 |
-
return ("No details found", "", "", "", "", "", "", "")
|
| 176 |
-
title = data.get("title", "")
|
| 177 |
-
artists = data.get("artists", "")
|
| 178 |
-
if isinstance(artists, list):
|
| 179 |
-
artists = ", ".join(artists)
|
| 180 |
-
genre = data.get("genre", "")
|
| 181 |
-
background = data.get("background", "")
|
| 182 |
-
analysis = data.get("analysis", "")
|
| 183 |
-
description = data.get("description", "")
|
| 184 |
-
scene = data.get("scene", "")
|
| 185 |
-
youtube_html = (
|
| 186 |
-
f'<iframe width="560" height="315" src="https://www.youtube.com/embed/{selected_id}" '
|
| 187 |
-
f'frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; '
|
| 188 |
-
f'gyroscope; picture-in-picture" allowfullscreen></iframe>'
|
| 189 |
-
)
|
| 190 |
-
return title, artists, genre, background, analysis, description, scene, youtube_html
|
| 191 |
-
|
| 192 |
-
def extract_features_from_text(text):
|
| 193 |
-
"""
|
| 194 |
-
Save the input text to a file, call the CLaMP 3 feature extraction script,
|
| 195 |
-
and return the generated feature file path.
|
| 196 |
-
"""
|
| 197 |
-
input_dir = "input_dir"
|
| 198 |
-
output_dir = "output_dir"
|
| 199 |
-
os.makedirs(input_dir, exist_ok=True)
|
| 200 |
-
os.makedirs(output_dir, exist_ok=True)
|
| 201 |
-
# Clear input_dir and output_dir
|
| 202 |
-
for d in [input_dir, output_dir]:
|
| 203 |
-
for filename in os.listdir(d):
|
| 204 |
-
file_path = os.path.join(d, filename)
|
| 205 |
-
if os.path.isfile(file_path) or os.path.islink(file_path):
|
| 206 |
-
os.unlink(file_path)
|
| 207 |
-
elif os.path.isdir(file_path):
|
| 208 |
-
shutil.rmtree(file_path)
|
| 209 |
-
input_file = os.path.join(input_dir, "input.txt")
|
| 210 |
-
print("Text input:", text)
|
| 211 |
-
with open(input_file, "w", encoding="utf-8") as f:
|
| 212 |
-
f.write(text)
|
| 213 |
-
command = ["python", "extract_clamp3.py", input_dir, output_dir, "--get_global"]
|
| 214 |
-
subprocess.run(command, check=True)
|
| 215 |
-
output_file = os.path.join(output_dir, "input.npy")
|
| 216 |
-
return output_file
|
| 217 |
-
|
| 218 |
-
def generate_caption(image):
|
| 219 |
-
"""
|
| 220 |
-
Use the BLIP model to generate a descriptive caption for the given image.
|
| 221 |
-
"""
|
| 222 |
-
inputs = processor(image, return_tensors="pt")
|
| 223 |
-
outputs = blip_model.generate(**inputs)
|
| 224 |
-
caption = processor.decode(outputs[0], skip_special_tokens=True)
|
| 225 |
-
return caption
|
| 226 |
-
|
| 227 |
-
class FileWrapper:
|
| 228 |
-
"""
|
| 229 |
-
Simulate a file object with a .name attribute.
|
| 230 |
-
"""
|
| 231 |
-
def __init__(self, path):
|
| 232 |
-
self.name = path
|
| 233 |
-
|
| 234 |
-
def search_wrapper(search_mode, text_input, image_input):
|
| 235 |
-
"""
|
| 236 |
-
Perform retrieval based on the selected input mode:
|
| 237 |
-
- If search_mode is "Image", use the uploaded image to generate a caption, then extract features
|
| 238 |
-
and search in the "image/" folder.
|
| 239 |
-
- If search_mode is "Text", use the provided text to extract features and search in the "image/" folder.
|
| 240 |
-
"""
|
| 241 |
-
if search_mode == "Image":
|
| 242 |
-
if image_input is None:
|
| 243 |
-
return text_input, gr.update(choices=[]), "Please upload an image.", "", "", "", "", "", "", ""
|
| 244 |
-
caption = generate_caption(image_input)
|
| 245 |
-
text_to_use = caption
|
| 246 |
-
reference_folder = "image/"
|
| 247 |
-
elif search_mode == "Text":
|
| 248 |
-
if not text_input or text_input.strip() == "":
|
| 249 |
-
return "Describe the music you're looking for (in any language)", gr.update(choices=[]), "Please enter text for retrieval.", "", "", "", "", "", "", ""
|
| 250 |
-
text_to_use = text_input
|
| 251 |
-
reference_folder = "text/"
|
| 252 |
-
else:
|
| 253 |
-
return "Describe the music you're looking for (in any language)", gr.update(choices=[]), "Invalid search mode selected.", "", "", "", "", "", "", ""
|
| 254 |
-
|
| 255 |
-
try:
|
| 256 |
-
output_file = extract_features_from_text(text_to_use)
|
| 257 |
-
query_file = FileWrapper(output_file)
|
| 258 |
-
except Exception as e:
|
| 259 |
-
return text_to_use, gr.update(choices=[]), f"Error during feature extraction: {e}", "", "", "", "", "", "", ""
|
| 260 |
-
candidate_ids, candidate_display = find_top_similar(query_file, reference_folder)
|
| 261 |
-
if not candidate_ids:
|
| 262 |
-
return text_to_use, gr.update(choices=[]), "", "", "", "", "", "", "", ""
|
| 263 |
-
choices = [(f"{i+1}. {disp}", cid) for i, (cid, disp) in enumerate(zip(candidate_ids, candidate_display))]
|
| 264 |
-
top_candidate = candidate_ids[0]
|
| 265 |
-
details = show_details(top_candidate)
|
| 266 |
-
return text_to_use, gr.update(choices=choices), *details
|
| 267 |
-
|
| 268 |
-
# 定义示例数据(示例数据放在组件定义之后也可以正常运行)
|
| 269 |
-
examples = [
|
| 270 |
-
["Image", None, "V4EauuhVEw4.jpg"],
|
| 271 |
-
["Image", None, "Kw-_Ew5bVxs.jpg"],
|
| 272 |
-
["Image", None, "BuYf0taXoNw.webp"],
|
| 273 |
-
["Image", None, "4tDYMayp6Dk.jpg"],
|
| 274 |
-
["Text", "classic rock, British, 1960s, upbeat", None],
|
| 275 |
-
["Text", "A Latin jazz piece with rhythmic percussion and brass", None],
|
| 276 |
-
["Text", "big band, major key, swing, brass-heavy, syncopation, baritone vocal", None],
|
| 277 |
-
["Text", "Heartfelt and nostalgic, with a bittersweet, melancholic feel", None],
|
| 278 |
-
["Text", "Melodía instrumental en re mayor con progresión armónica repetitiva y fluida", None],
|
| 279 |
-
["Text", "D大调四四拍的爱尔兰舞曲", None],
|
| 280 |
-
["Text", "Ιερή μουσική με πνευματική ατμόσφαιρα", None],
|
| 281 |
-
["Text", "የፍቅር ሙዚቃ ሞቅ እና ስሜታማ ከሆነ ነገር ግን ድንቅ እና አስደሳች ቃላት ያካትታል", None],
|
| 282 |
-
]
|
| 283 |
-
|
| 284 |
-
with gr.Blocks() as demo:
|
| 285 |
-
gr.Markdown(title)
|
| 286 |
-
gr.HTML(badges)
|
| 287 |
-
gr.Markdown(description)
|
| 288 |
-
|
| 289 |
-
with gr.Row():
|
| 290 |
-
with gr.Column():
|
| 291 |
-
search_mode = gr.Radio(
|
| 292 |
-
choices=["Text", "Image"],
|
| 293 |
-
label="Select Search Mode",
|
| 294 |
-
value="Text",
|
| 295 |
-
interactive=True,
|
| 296 |
-
elem_classes=["vertical-radio"]
|
| 297 |
-
)
|
| 298 |
-
text_input = gr.Textbox(
|
| 299 |
-
placeholder="Describe the music you're looking for (in any language)",
|
| 300 |
-
lines=4
|
| 301 |
-
)
|
| 302 |
-
image_input = gr.Image(
|
| 303 |
-
label="Or upload an image (PNG, JPG)",
|
| 304 |
-
type="pil"
|
| 305 |
-
)
|
| 306 |
-
search_button = gr.Button("Search")
|
| 307 |
-
candidate_radio = gr.Radio(choices=[], label="Select Retrieval Result", interactive=True, elem_classes=["vertical-radio"])
|
| 308 |
-
with gr.Column():
|
| 309 |
-
gr.Markdown("### YouTube Video")
|
| 310 |
-
youtube_box = gr.HTML(label="YouTube Video")
|
| 311 |
-
gr.Markdown("### Metadata")
|
| 312 |
-
title_box = gr.Textbox(label="Title", interactive=False)
|
| 313 |
-
artists_box = gr.Textbox(label="Artists", interactive=False)
|
| 314 |
-
genre_box = gr.Textbox(label="Genre", interactive=False)
|
| 315 |
-
background_box = gr.Textbox(label="Background", interactive=False)
|
| 316 |
-
analysis_box = gr.Textbox(label="Analysis", interactive=False)
|
| 317 |
-
description_box = gr.Textbox(label="Description", interactive=False)
|
| 318 |
-
scene_box = gr.Textbox(label="Scene", interactive=False)
|
| 319 |
-
|
| 320 |
-
gr.HTML(
|
| 321 |
-
"""
|
| 322 |
-
<style>
|
| 323 |
-
.vertical-radio .gradio-radio label {
|
| 324 |
-
display: block !important;
|
| 325 |
-
margin-bottom: 5px;
|
| 326 |
-
}
|
| 327 |
-
</style>
|
| 328 |
-
"""
|
| 329 |
-
)
|
| 330 |
-
|
| 331 |
-
gr.Examples(
|
| 332 |
-
examples=examples,
|
| 333 |
-
inputs=[search_mode, text_input, image_input],
|
| 334 |
-
outputs=[text_input, candidate_radio, title_box, artists_box, genre_box, background_box, analysis_box, description_box, scene_box, youtube_box],
|
| 335 |
-
fn=search_wrapper,
|
| 336 |
-
cache_examples=False,
|
| 337 |
-
)
|
| 338 |
-
|
| 339 |
-
search_button.click(
|
| 340 |
-
fn=search_wrapper,
|
| 341 |
-
inputs=[search_mode, text_input, image_input],
|
| 342 |
-
outputs=[text_input, candidate_radio, title_box, artists_box, genre_box, background_box, analysis_box, description_box, scene_box, youtube_box]
|
| 343 |
-
)
|
| 344 |
-
|
| 345 |
-
candidate_radio.change(
|
| 346 |
-
fn=show_details,
|
| 347 |
-
inputs=candidate_radio,
|
| 348 |
-
outputs=[title_box, artists_box, genre_box, background_box, analysis_box, description_box, scene_box, youtube_box]
|
| 349 |
-
)
|
| 350 |
-
|
| 351 |
-
demo.launch()
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import torch
|
| 3 |
+
import numpy as np
|
| 4 |
+
import gradio as gr
|
| 5 |
+
import zipfile
|
| 6 |
+
import json
|
| 7 |
+
import requests
|
| 8 |
+
import subprocess
|
| 9 |
+
import shutil
|
| 10 |
+
from transformers import BlipProcessor, BlipForConditionalGeneration
|
| 11 |
+
|
| 12 |
+
title = "# 🗜️ CLaMP 3 - Multimodal & Multilingual Semantic Music Search"
|
| 13 |
+
|
| 14 |
+
badges = """
|
| 15 |
+
<div style="text-align: center;">
|
| 16 |
+
<a href="https://sanderwood.github.io/clamp3/">
|
| 17 |
+
<img src="https://img.shields.io/badge/CLaMP%203%20Homepage-GitHub-181717?style=for-the-badge&logo=home-assistant" alt="Homepage">
|
| 18 |
+
</a>
|
| 19 |
+
<a href="https://arxiv.org/abs/2502.10362">
|
| 20 |
+
<img src="https://img.shields.io/badge/CLaMP%203%20Paper-Arxiv-red?style=for-the-badge&logo=arxiv" alt="Paper">
|
| 21 |
+
</a>
|
| 22 |
+
<a href="https://github.com/sanderwood/clamp3">
|
| 23 |
+
<img src="https://img.shields.io/badge/CLaMP%203%20Code-GitHub-181717?style=for-the-badge&logo=github" alt="GitHub">
|
| 24 |
+
</a>
|
| 25 |
+
<a href="https://huggingface.co/spaces/sander-wood/clamp3">
|
| 26 |
+
<img src="https://img.shields.io/badge/CLaMP%203%20Demo-Gradio-green?style=for-the-badge&logo=gradio" alt="Demo">
|
| 27 |
+
</a>
|
| 28 |
+
<a href="https://huggingface.co/sander-wood/clamp3/tree/main">
|
| 29 |
+
<img src="https://img.shields.io/badge/Model%20Weights-Hugging%20Face-ffcc00?style=for-the-badge&logo=huggingface" alt="Model Weights">
|
| 30 |
+
</a>
|
| 31 |
+
<a href="https://huggingface.co/datasets/sander-wood/m4-rag">
|
| 32 |
+
<img src="https://img.shields.io/badge/M4--RAG%20Dataset-Hugging%20Face-ffcc00?style=for-the-badge&logo=huggingface" alt="Dataset">
|
| 33 |
+
</a>
|
| 34 |
+
<a href="https://huggingface.co/datasets/sander-wood/wikimt-x">
|
| 35 |
+
<img src="https://img.shields.io/badge/WikiMT--X%20Benchmark-Hugging%20Face-ffcc00?style=for-the-badge&logo=huggingface" alt="Benchmark">
|
| 36 |
+
</a>
|
| 37 |
+
</div>
|
| 38 |
+
|
| 39 |
+
<style>
|
| 40 |
+
div a {
|
| 41 |
+
display: inline-block;
|
| 42 |
+
margin: 5px;
|
| 43 |
+
}
|
| 44 |
+
div a img {
|
| 45 |
+
height: 30px;
|
| 46 |
+
}
|
| 47 |
+
</style>
|
| 48 |
+
"""
|
| 49 |
+
|
| 50 |
+
description = """CLaMP 3 is a **multimodal and multilingual** music information retrieval (MIR) framework, supporting **sheet music, audio, and performance signals** in **100 languages**. Using **contrastive learning**, it aligns these modalities in a shared space for **cross-modal retrieval**.
|
| 51 |
+
|
| 52 |
+
### 🔍 **How This Demo Works**
|
| 53 |
+
- You can **retrieve music using any text input (in any language) or an image** (`.png`, `.jpg`).
|
| 54 |
+
- When using an image, **BLIP** generates a caption, which is then used for retrieval.
|
| 55 |
+
- Since CLaMP 3's training data includes **rich visual descriptions of musical scenes**, it can **match images to semantically relevant music**.
|
| 56 |
+
|
| 57 |
+
### ⚠️ **Limitations**
|
| 58 |
+
- This demo retrieves music **only from the WikiMT-X benchmark (1,000 pieces)**.
|
| 59 |
+
- These pieces are **mainly from the U.S. and Western Europe (especially the U.S.)** and **mostly from the 20th century**.
|
| 60 |
+
- Thus, retrieval results are **mostly limited to Western 20th-century music**, so you **won’t** find music from **other regions or historical periods**.
|
| 61 |
+
|
| 62 |
+
🔧 **Need retrieval for a different music collection?** Deploy **[CLaMP 3](https://github.com/sanderwood/clamp3)** on your own dataset.
|
| 63 |
+
Generally, the larger and more diverse the reference music dataset, the better the retrieval quality, increasing the likelihood of finding relevant and accurately matched music.
|
| 64 |
+
|
| 65 |
+
**Note: This project is for research use only.**
|
| 66 |
+
"""
|
| 67 |
+
|
| 68 |
+
# Load BLIP image captioning model and processor
|
| 69 |
+
processor = BlipProcessor.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 70 |
+
blip_model = BlipForConditionalGeneration.from_pretrained("Salesforce/blip-image-captioning-base")
|
| 71 |
+
|
| 72 |
+
# Download weight file if it does not exist
|
| 73 |
+
weights_url = "https://huggingface.co/sander-wood/clamp3/resolve/main/weights_clamp3_saas_h_size_768_t_model_FacebookAI_xlm-roberta-base_t_length_128_a_size_768_a_layers_12_a_length_128_s_size_768_s_layers_12_p_size_64_p_length_512.pth"
|
| 74 |
+
weights_filename = "weights_clamp3_saas_h_size_768_t_model_FacebookAI_xlm-roberta-base_t_length_128_a_size_768_a_layers_12_a_length_128_s_size_768_s_layers_12_p_size_64_p_length_512.pth"
|
| 75 |
+
|
| 76 |
+
if not os.path.exists(weights_filename):
|
| 77 |
+
print("Downloading weights file...")
|
| 78 |
+
response = requests.get(weights_url, stream=True)
|
| 79 |
+
response.raise_for_status()
|
| 80 |
+
with open(weights_filename, "wb") as f:
|
| 81 |
+
for chunk in response.iter_content(chunk_size=8192):
|
| 82 |
+
if chunk:
|
| 83 |
+
f.write(chunk)
|
| 84 |
+
print("Weights file downloaded.")
|
| 85 |
+
|
| 86 |
+
ZIP_PATH = "features.zip"
|
| 87 |
+
if os.path.exists(ZIP_PATH):
|
| 88 |
+
print(f"Extracting {ZIP_PATH}...")
|
| 89 |
+
with zipfile.ZipFile(ZIP_PATH, "r") as zip_ref:
|
| 90 |
+
zip_ref.extractall(".")
|
| 91 |
+
print("Extraction complete.")
|
| 92 |
+
|
| 93 |
+
# Load metadata
|
| 94 |
+
metadata_map = {}
|
| 95 |
+
METADATA_FILE = "wikimt-x-public.jsonl"
|
| 96 |
+
if os.path.exists(METADATA_FILE):
|
| 97 |
+
with open(METADATA_FILE, "r", encoding="utf-8") as f:
|
| 98 |
+
for line in f:
|
| 99 |
+
data = json.loads(line)
|
| 100 |
+
metadata_map[data["id"]] = data
|
| 101 |
+
else:
|
| 102 |
+
print(f"Warning: {METADATA_FILE} not found.")
|
| 103 |
+
|
| 104 |
+
features_cache = {}
|
| 105 |
+
|
| 106 |
+
def get_info(folder_path):
|
| 107 |
+
"""
|
| 108 |
+
Load all .npy files from the specified folder and return a dictionary
|
| 109 |
+
with the file names (without extension) as keys.
|
| 110 |
+
"""
|
| 111 |
+
if folder_path in features_cache:
|
| 112 |
+
return features_cache[folder_path]
|
| 113 |
+
if not os.path.exists(folder_path):
|
| 114 |
+
return {}
|
| 115 |
+
files = sorted(os.listdir(folder_path))
|
| 116 |
+
features = {}
|
| 117 |
+
for file in files:
|
| 118 |
+
if file.endswith(".npy"):
|
| 119 |
+
key = file.split(".")[0]
|
| 120 |
+
try:
|
| 121 |
+
features[key] = np.load(os.path.join(folder_path, file))[0]
|
| 122 |
+
except Exception as e:
|
| 123 |
+
print(f"Error loading {file}: {e}")
|
| 124 |
+
features_cache[folder_path] = features
|
| 125 |
+
return features
|
| 126 |
+
|
| 127 |
+
def find_top_similar(query_file, reference_folder):
|
| 128 |
+
"""
|
| 129 |
+
Compare the query feature with all reference features in the specified folder
|
| 130 |
+
using cosine similarity and return the top 10 candidate results in the format:
|
| 131 |
+
Title | Artists | sim: SimilarityScore.
|
| 132 |
+
"""
|
| 133 |
+
top_k = 10
|
| 134 |
+
try:
|
| 135 |
+
query_feature = np.load(query_file.name)[0]
|
| 136 |
+
except Exception as e:
|
| 137 |
+
return [], f"Error loading query feature: {e}"
|
| 138 |
+
query_tensor = torch.tensor(query_feature, dtype=torch.float32).unsqueeze(dim=0)
|
| 139 |
+
key_features = get_info(reference_folder)
|
| 140 |
+
if not key_features:
|
| 141 |
+
return [], f"No reference features found in {reference_folder}."
|
| 142 |
+
ref_keys = list(key_features.keys())
|
| 143 |
+
ref_array = np.array([key_features[k] for k in ref_keys])
|
| 144 |
+
key_feats_tensor = torch.tensor(ref_array, dtype=torch.float32)
|
| 145 |
+
query_tensor_expanded = query_tensor.expand(key_feats_tensor.size(0), -1)
|
| 146 |
+
similarities = torch.cosine_similarity(query_tensor_expanded, key_feats_tensor, dim=1)
|
| 147 |
+
ranked_indices = torch.argsort(similarities, descending=True)
|
| 148 |
+
candidate_ids = []
|
| 149 |
+
candidate_display = []
|
| 150 |
+
for i in range(top_k):
|
| 151 |
+
if i < len(ref_keys):
|
| 152 |
+
candidate_idx = ranked_indices[i].item()
|
| 153 |
+
candidate_id = ref_keys[candidate_idx]
|
| 154 |
+
sim = round(similarities[candidate_idx].item(), 4)
|
| 155 |
+
meta = metadata_map.get(candidate_id, {})
|
| 156 |
+
title = meta.get("title", candidate_id)
|
| 157 |
+
artists = meta.get("artists", "Unknown")
|
| 158 |
+
if isinstance(artists, list):
|
| 159 |
+
artists = ", ".join(artists)
|
| 160 |
+
candidate_ids.append(candidate_id)
|
| 161 |
+
candidate_display.append(f"{title} | {artists} | sim: {sim}")
|
| 162 |
+
else:
|
| 163 |
+
candidate_ids.append("N/A")
|
| 164 |
+
candidate_display.append("N/A")
|
| 165 |
+
return candidate_ids, candidate_display
|
| 166 |
+
|
| 167 |
+
def show_details(selected_id):
|
| 168 |
+
"""
|
| 169 |
+
Return detailed metadata and embedded YouTube video HTML based on the candidate ID.
|
| 170 |
+
"""
|
| 171 |
+
if selected_id == "N/A":
|
| 172 |
+
return ("", "", "", "", "", "", "", "")
|
| 173 |
+
data = metadata_map.get(selected_id, {})
|
| 174 |
+
if not data:
|
| 175 |
+
return ("No details found", "", "", "", "", "", "", "")
|
| 176 |
+
title = data.get("title", "")
|
| 177 |
+
artists = data.get("artists", "")
|
| 178 |
+
if isinstance(artists, list):
|
| 179 |
+
artists = ", ".join(artists)
|
| 180 |
+
genre = data.get("genre", "")
|
| 181 |
+
background = data.get("background", "")
|
| 182 |
+
analysis = data.get("analysis", "")
|
| 183 |
+
description = data.get("description", "")
|
| 184 |
+
scene = data.get("scene", "")
|
| 185 |
+
youtube_html = (
|
| 186 |
+
f'<iframe width="560" height="315" src="https://www.youtube.com/embed/{selected_id}" '
|
| 187 |
+
f'frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; '
|
| 188 |
+
f'gyroscope; picture-in-picture" allowfullscreen></iframe>'
|
| 189 |
+
)
|
| 190 |
+
return title, artists, genre, background, analysis, description, scene, youtube_html
|
| 191 |
+
|
| 192 |
+
def extract_features_from_text(text):
|
| 193 |
+
"""
|
| 194 |
+
Save the input text to a file, call the CLaMP 3 feature extraction script,
|
| 195 |
+
and return the generated feature file path.
|
| 196 |
+
"""
|
| 197 |
+
input_dir = "input_dir"
|
| 198 |
+
output_dir = "output_dir"
|
| 199 |
+
os.makedirs(input_dir, exist_ok=True)
|
| 200 |
+
os.makedirs(output_dir, exist_ok=True)
|
| 201 |
+
# Clear input_dir and output_dir
|
| 202 |
+
for d in [input_dir, output_dir]:
|
| 203 |
+
for filename in os.listdir(d):
|
| 204 |
+
file_path = os.path.join(d, filename)
|
| 205 |
+
if os.path.isfile(file_path) or os.path.islink(file_path):
|
| 206 |
+
os.unlink(file_path)
|
| 207 |
+
elif os.path.isdir(file_path):
|
| 208 |
+
shutil.rmtree(file_path)
|
| 209 |
+
input_file = os.path.join(input_dir, "input.txt")
|
| 210 |
+
print("Text input:", text)
|
| 211 |
+
with open(input_file, "w", encoding="utf-8") as f:
|
| 212 |
+
f.write(text)
|
| 213 |
+
command = ["python", "extract_clamp3.py", input_dir, output_dir, "--get_global"]
|
| 214 |
+
subprocess.run(command, check=True)
|
| 215 |
+
output_file = os.path.join(output_dir, "input.npy")
|
| 216 |
+
return output_file
|
| 217 |
+
|
| 218 |
+
def generate_caption(image):
|
| 219 |
+
"""
|
| 220 |
+
Use the BLIP model to generate a descriptive caption for the given image.
|
| 221 |
+
"""
|
| 222 |
+
inputs = processor(image, return_tensors="pt")
|
| 223 |
+
outputs = blip_model.generate(**inputs)
|
| 224 |
+
caption = processor.decode(outputs[0], skip_special_tokens=True)
|
| 225 |
+
return caption
|
| 226 |
+
|
| 227 |
+
class FileWrapper:
|
| 228 |
+
"""
|
| 229 |
+
Simulate a file object with a .name attribute.
|
| 230 |
+
"""
|
| 231 |
+
def __init__(self, path):
|
| 232 |
+
self.name = path
|
| 233 |
+
|
| 234 |
+
def search_wrapper(search_mode, text_input, image_input):
|
| 235 |
+
"""
|
| 236 |
+
Perform retrieval based on the selected input mode:
|
| 237 |
+
- If search_mode is "Image", use the uploaded image to generate a caption, then extract features
|
| 238 |
+
and search in the "image/" folder.
|
| 239 |
+
- If search_mode is "Text", use the provided text to extract features and search in the "image/" folder.
|
| 240 |
+
"""
|
| 241 |
+
if search_mode == "Image":
|
| 242 |
+
if image_input is None:
|
| 243 |
+
return text_input, gr.update(choices=[]), "Please upload an image.", "", "", "", "", "", "", ""
|
| 244 |
+
caption = generate_caption(image_input)
|
| 245 |
+
text_to_use = caption
|
| 246 |
+
reference_folder = "image/"
|
| 247 |
+
elif search_mode == "Text":
|
| 248 |
+
if not text_input or text_input.strip() == "":
|
| 249 |
+
return "Describe the music you're looking for (in any language)", gr.update(choices=[]), "Please enter text for retrieval.", "", "", "", "", "", "", ""
|
| 250 |
+
text_to_use = text_input
|
| 251 |
+
reference_folder = "text/"
|
| 252 |
+
else:
|
| 253 |
+
return "Describe the music you're looking for (in any language)", gr.update(choices=[]), "Invalid search mode selected.", "", "", "", "", "", "", ""
|
| 254 |
+
|
| 255 |
+
try:
|
| 256 |
+
output_file = extract_features_from_text(text_to_use)
|
| 257 |
+
query_file = FileWrapper(output_file)
|
| 258 |
+
except Exception as e:
|
| 259 |
+
return text_to_use, gr.update(choices=[]), f"Error during feature extraction: {e}", "", "", "", "", "", "", ""
|
| 260 |
+
candidate_ids, candidate_display = find_top_similar(query_file, reference_folder)
|
| 261 |
+
if not candidate_ids:
|
| 262 |
+
return text_to_use, gr.update(choices=[]), "", "", "", "", "", "", "", ""
|
| 263 |
+
choices = [(f"{i+1}. {disp}", cid) for i, (cid, disp) in enumerate(zip(candidate_ids, candidate_display))]
|
| 264 |
+
top_candidate = candidate_ids[0]
|
| 265 |
+
details = show_details(top_candidate)
|
| 266 |
+
return text_to_use, gr.update(choices=choices), *details
|
| 267 |
+
|
| 268 |
+
# 定义示例数据(示例数据放在组件定义之后也可以正常运行)
|
| 269 |
+
examples = [
|
| 270 |
+
["Image", None, "V4EauuhVEw4.jpg"],
|
| 271 |
+
["Image", None, "Kw-_Ew5bVxs.jpg"],
|
| 272 |
+
["Image", None, "BuYf0taXoNw.webp"],
|
| 273 |
+
["Image", None, "4tDYMayp6Dk.jpg"],
|
| 274 |
+
["Text", "classic rock, British, 1960s, upbeat", None],
|
| 275 |
+
["Text", "A Latin jazz piece with rhythmic percussion and brass", None],
|
| 276 |
+
["Text", "big band, major key, swing, brass-heavy, syncopation, baritone vocal", None],
|
| 277 |
+
["Text", "Heartfelt and nostalgic, with a bittersweet, melancholic feel", None],
|
| 278 |
+
["Text", "Melodía instrumental en re mayor con progresión armónica repetitiva y fluida", None],
|
| 279 |
+
["Text", "D大调四四拍的爱尔兰舞曲", None],
|
| 280 |
+
["Text", "Ιερή μουσική με πνευματική ατμόσφαιρα", None],
|
| 281 |
+
["Text", "የፍቅር ሙዚቃ ሞቅ እና ስሜታማ ከሆነ ነገር ግን ድንቅ እና አስደሳች ቃላት ያካትታል", None],
|
| 282 |
+
]
|
| 283 |
+
|
| 284 |
+
with gr.Blocks() as demo:
|
| 285 |
+
gr.Markdown(title)
|
| 286 |
+
gr.HTML(badges)
|
| 287 |
+
gr.Markdown(description)
|
| 288 |
+
|
| 289 |
+
with gr.Row():
|
| 290 |
+
with gr.Column():
|
| 291 |
+
search_mode = gr.Radio(
|
| 292 |
+
choices=["Text", "Image"],
|
| 293 |
+
label="Select Search Mode",
|
| 294 |
+
value="Text",
|
| 295 |
+
interactive=True,
|
| 296 |
+
elem_classes=["vertical-radio"]
|
| 297 |
+
)
|
| 298 |
+
text_input = gr.Textbox(
|
| 299 |
+
placeholder="Describe the music you're looking for (in any language)",
|
| 300 |
+
lines=4
|
| 301 |
+
)
|
| 302 |
+
image_input = gr.Image(
|
| 303 |
+
label="Or upload an image (PNG, JPG)",
|
| 304 |
+
type="pil"
|
| 305 |
+
)
|
| 306 |
+
search_button = gr.Button("Search from 1,000 Western 20th-century music in WikiMT-X")
|
| 307 |
+
candidate_radio = gr.Radio(choices=[], label="Select Retrieval Result", interactive=True, elem_classes=["vertical-radio"])
|
| 308 |
+
with gr.Column():
|
| 309 |
+
gr.Markdown("### YouTube Video")
|
| 310 |
+
youtube_box = gr.HTML(label="YouTube Video")
|
| 311 |
+
gr.Markdown("### Metadata")
|
| 312 |
+
title_box = gr.Textbox(label="Title", interactive=False)
|
| 313 |
+
artists_box = gr.Textbox(label="Artists", interactive=False)
|
| 314 |
+
genre_box = gr.Textbox(label="Genre", interactive=False)
|
| 315 |
+
background_box = gr.Textbox(label="Background", interactive=False)
|
| 316 |
+
analysis_box = gr.Textbox(label="Analysis", interactive=False)
|
| 317 |
+
description_box = gr.Textbox(label="Description", interactive=False)
|
| 318 |
+
scene_box = gr.Textbox(label="Scene", interactive=False)
|
| 319 |
+
|
| 320 |
+
gr.HTML(
|
| 321 |
+
"""
|
| 322 |
+
<style>
|
| 323 |
+
.vertical-radio .gradio-radio label {
|
| 324 |
+
display: block !important;
|
| 325 |
+
margin-bottom: 5px;
|
| 326 |
+
}
|
| 327 |
+
</style>
|
| 328 |
+
"""
|
| 329 |
+
)
|
| 330 |
+
|
| 331 |
+
gr.Examples(
|
| 332 |
+
examples=examples,
|
| 333 |
+
inputs=[search_mode, text_input, image_input],
|
| 334 |
+
outputs=[text_input, candidate_radio, title_box, artists_box, genre_box, background_box, analysis_box, description_box, scene_box, youtube_box],
|
| 335 |
+
fn=search_wrapper,
|
| 336 |
+
cache_examples=False,
|
| 337 |
+
)
|
| 338 |
+
|
| 339 |
+
search_button.click(
|
| 340 |
+
fn=search_wrapper,
|
| 341 |
+
inputs=[search_mode, text_input, image_input],
|
| 342 |
+
outputs=[text_input, candidate_radio, title_box, artists_box, genre_box, background_box, analysis_box, description_box, scene_box, youtube_box]
|
| 343 |
+
)
|
| 344 |
+
|
| 345 |
+
candidate_radio.change(
|
| 346 |
+
fn=show_details,
|
| 347 |
+
inputs=candidate_radio,
|
| 348 |
+
outputs=[title_box, artists_box, genre_box, background_box, analysis_box, description_box, scene_box, youtube_box]
|
| 349 |
+
)
|
| 350 |
+
|
| 351 |
+
demo.launch()
|