{"id":16985,"date":"2025-01-02T18:04:31","date_gmt":"2025-01-02T10:04:31","guid":{"rendered":"https:\/\/www.aisharenet.com\/?p=16985"},"modified":"2025-01-02T18:04:31","modified_gmt":"2025-01-02T10:04:31","slug":"tangoflux","status":"publish","type":"post","link":"https:\/\/www.kdjingpai.com\/en\/tangoflux\/","title":{"rendered":"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891"},"content":{"rendered":"<p>TangoFlux \u662f\u4e00\u4e2a\u7531 DeCLaRe Lab \u5f00\u53d1\u7684\u9ad8\u6548\u6587\u672c\u5230\u97f3\u9891\uff08TTA\uff09\u751f\u6210\u6a21\u578b\u3002\u8be5\u6a21\u578b\u80fd\u591f\u5728\u77ed\u77ed 3.7 \u79d2\u5185\u751f\u6210\u957f\u8fbe 30 \u79d2\u7684 44.1kHz \u7acb\u4f53\u58f0\u97f3\u9891\u3002TangoFlux \u91c7\u7528\u6d41\u5339\u914d\u548c Clap-Ranked Preference Optimization\uff08CRPO\uff09\u6280\u672f\uff0c\u901a\u8fc7\u751f\u6210\u548c\u4f18\u5316\u504f\u597d\u6570\u636e\u6765\u589e\u5f3a TTA \u5bf9\u9f50\u3002\u8be5\u6a21\u578b\u5728\u5ba2\u89c2\u548c\u4e3b\u89c2\u57fa\u51c6\u6d4b\u8bd5\u4e2d\u5747\u8868\u73b0\u51fa\u8272\uff0c\u5e76\u4e14\u6240\u6709\u4ee3\u7801\u548c\u6a21\u578b\u5747\u5df2\u5f00\u6e90\uff0c\u652f\u6301\u8fdb\u4e00\u6b65\u7684 TTA \u751f\u6210\u7814\u7a76\u3002<\/p>\n<div style=\"width: 1255px\" class=\"wp-caption alignnone\"><img loading=\"lazy\" decoding=\"async\" title=\"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/3aa92881a2cef9e.png\" alt=\"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891-1\" width=\"1245\" height=\"587\" \/><p class=\"wp-caption-text\">\u4f53\u9a8c\u5730\u5740\uff1ahttps:\/\/huggingface.co\/spaces\/declare-lab\/TangoFlux<\/p><\/div>\n<p>&nbsp;<\/p>\n<p>\u65b0\u52a0\u5761\u79d1\u6280\u8bbe\u8ba1\u5927\u5b66\uff08SUTD\uff09\u548cNVIDIA\u8054\u5408\u53d1\u5e03\u4e86\u4e00\u79cd\u9ad8\u6548\u7684\u6587\u672c\u5230\u97f3\u9891\u751f\u6210\u6a21\u578b\u00a0\uff08TTA\uff09\u2014\u2014TangoFlux\u00a0\u3002\u8be5\u6a21\u578b\u62e5\u6709\u7ea65.15\u4ebf\u4e2a\u53c2\u6570\uff0c\u80fd\u591f\u5728\u5355\u4e2aA40 GPU\u4e0a\u4ec5\u75283.7\u79d2\u751f\u6210\u957f\u8fbe30\u79d2\u768444.1kHz\u97f3\u9891\u3002TangoFlux\u4e0d\u4ec5\u5177\u6709\u8d85\u5feb\u7684\u751f\u6210\u901f\u5ea6\uff0c\u540c\u65f6\u5728\u97f3\u9891\u8d28\u91cf\u4e0a\u4e5f\u597d\u4e8eStable Audio\u7b49\u5f00\u6e90\u97f3\u9891\u6a21\u578b\u3002<\/p>\n<p>\u5c06TANGoFLux\u4e0e\u5176\u4ed6\u6700\u5148\u8fdb\u7684\u5f00\u6e90\u6587\u672c\u5230\u97f3\u9891\u751f\u6210\u6a21\u578b\u8fdb\u884c\u6bd4\u8f83\uff1aTANGoFLux\u4e0d\u4ec5\u751f\u6210\u901f\u5ea6\u6bd4\u6700\u5feb\u7684\u6a21\u578b\u5feb\u5927\u7ea62\u500d\uff0c\u800c\u4e14\u5b9e\u73b0\u4e86\u66f4\u597d\u7684\u97f3\u9891\u8d28\u91cf\uff08\u901a\u8fc7CLAP\u548cFD\u5206\u6570\u8861\u91cf\uff09\uff0c\u6240\u6709\u8fd9\u4e9b\u8fd8\u5177\u6709\u66f4\u5c11\u7684\u53ef\u8bad\u7ec3\u53c2\u6570\u3002<\/p>\n<p><img decoding=\"async\" title=\"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/e1775aab1c5b3e1.png\" alt=\"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891-1\" \/><\/p>\n<p>&nbsp;<\/p>\n<p>TangoFlux\u00a0\u8bba\u6587\u540d\u4e3a\u201c\u901a\u8fc7\u6d41\u5339\u914d\u548c\u00a0Clap-Ranked\u00a0\u504f\u597d\u4f18\u5316\u5b9e\u73b0\u8d85\u5feb\u901f\u3001\u5fe0\u5b9e\u7684\u6587\u672c\u5230\u97f3\u9891\u751f\u6210\u201d\uff0c\u5176\u7531\u00a0FluxTransformer\u00a0\u5757\u7ec4\u6210\uff0c\u8fd9\u4e9b\u5757\u662f\u6269\u6563\u53d8\u6362\u5668\u00a0(DiT)\u00a0\u548c\u591a\u6a21\u6001\u6269\u6563\u53d8\u6362\u5668\u00a0(MMDiT)\uff0c\u5b83\u4eec\u4ee5\u6587\u672c\u63d0\u793a\u548c\u6301\u7eed\u65f6\u95f4\u5d4c\u5165\u4e3a\u6761\u4ef6\uff0c\u4ee5\u751f\u6210\u6700\u957f\u00a030\u00a0\u79d2\u7684\u00a044.1kHz\u00a0\u97f3\u9891\u3002TangoFlux\u00a0\u5b66\u4e60\u7531\u53d8\u5206\u81ea\u52a8\u7f16\u7801\u5668\u00a0(VAE)\u00a0\u7f16\u7801\u7684\u97f3\u9891\u6f5c\u5728\u8868\u793a\u7684\u6574\u6d41\u6d41\u8f68\u8ff9\u3002TangoFlux\u00a0\u8bad\u7ec3\u7ba1\u9053\u5305\u62ec\u4e09\u4e2a\u9636\u6bb5\uff1a\u9884\u8bad\u7ec3\u3001\u5fae\u8c03\u548c\u4f7f\u7528\u00a0CRPO\u00a0\u8fdb\u884c\u504f\u597d\u4f18\u5316\u3002\u5177\u4f53\u6765\u8bf4\uff0cCRPO\u00a0\u4f1a\u8fed\u4ee3\u751f\u6210\u65b0\u7684\u5408\u6210\u6570\u636e\uff0c\u5e76\u4f7f\u7528\u00a0DPO\u00a0\u635f\u5931\u6784\u5efa\u504f\u597d\u5bf9\u4ee5\u8fdb\u884c\u504f\u597d\u4f18\u5316\uff0c\u4ece\u800c\u5b9e\u73b0\u6d41\u5339\u914d\u3002<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"\" title=\"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891-1\" src=\"https:\/\/www.kdjingpai.com\/wp-content\/uploads\/2025\/01\/349e31eff2ffe56.png\" alt=\"TangoFlux\uff1a\u5feb\u901f\u6587\u672c\u5230\u914d\u97f3\u8f6c\u6362\u5de5\u5177\uff0c3\u79d2\u949f\u751f\u621030\u79d2\u957f\u97f3\u9891-1\" width=\"1067\" height=\"441\" \/><\/p>\n<p>&nbsp;<\/p>\n<h2>\u529f\u80fd\u5217\u8868<\/h2>\n<ul>\n<li><strong>\u5feb\u901f\u97f3\u9891\u751f\u6210<\/strong>\uff1a\u5728 3.7 \u79d2\u5185\u751f\u6210\u957f\u8fbe 30 \u79d2\u7684\u9ad8\u8d28\u91cf\u97f3\u9891\u3002<\/li>\n<li><strong>\u6d41\u5339\u914d\u6280\u672f<\/strong>\uff1a\u4f7f\u7528 FluxTransformer \u548c Multimodal Diffusion Transformers \u8fdb\u884c\u97f3\u9891\u751f\u6210\u3002<\/li>\n<li><strong>CRPO \u4f18\u5316<\/strong>\uff1a\u901a\u8fc7\u751f\u6210\u548c\u4f18\u5316\u504f\u597d\u6570\u636e\u6765\u63d0\u9ad8\u97f3\u9891\u751f\u6210\u8d28\u91cf\u3002<\/li>\n<li><strong>\u591a\u9636\u6bb5\u8bad\u7ec3<\/strong>\uff1a\u5305\u62ec\u9884\u8bad\u7ec3\u3001\u5fae\u8c03\u548c\u504f\u597d\u4f18\u5316\u4e09\u4e2a\u9636\u6bb5\u3002<\/li>\n<li><strong>\u5f00\u6e90\u4ee3\u7801<\/strong>\uff1a\u6240\u6709\u4ee3\u7801\u548c\u6a21\u578b\u5747\u5df2\u5f00\u6e90\uff0c\u652f\u6301\u8fdb\u4e00\u6b65\u7814\u7a76\u3002<\/li>\n<\/ul>\n<p>&nbsp;<\/p>\n<h2>\u4f7f\u7528\u5e2e\u52a9<\/h2>\n<h3>\u5b89\u88c5\u6d41\u7a0b<\/h3>\n<ol>\n<li><strong>\u73af\u5883\u914d\u7f6e<\/strong>\uff1a\u786e\u4fdd\u5df2\u5b89\u88c5 Python 3.7 \u53ca\u4ee5\u4e0a\u7248\u672c\uff0c\u5e76\u5b89\u88c5\u5fc5\u8981\u7684\u4f9d\u8d56\u5e93\u3002<\/li>\n<li><strong>\u514b\u9686\u4ed3\u5e93<\/strong>\uff1a\u5728\u7ec8\u7aef\u4e2d\u8fd0\u884c <code>git clone https:\/\/github.com\/declare-lab\/TangoFlux.git<\/code> \u514b\u9686\u4ed3\u5e93\u3002<\/li>\n<li><strong>\u5b89\u88c5\u4f9d\u8d56<\/strong>\uff1a\u8fdb\u5165\u9879\u76ee\u76ee\u5f55\uff0c\u8fd0\u884c <code>pip install -r requirements.txt<\/code> \u5b89\u88c5\u6240\u6709\u4f9d\u8d56\u3002<\/li>\n<\/ol>\n<h3>\u4f7f\u7528\u6d41\u7a0b<\/h3>\n<ol>\n<li><strong>\u6a21\u578b\u8bad\u7ec3<\/strong>\uff1a\n<ul>\n<li>\u914d\u7f6e\u52a0\u901f\u5668\uff1a\u8fd0\u884c <code>accelerate config<\/code> \u5e76\u6839\u636e\u63d0\u793a\u914d\u7f6e\u8fd0\u884c\u73af\u5883\u3002<\/li>\n<li>\u914d\u7f6e\u8bad\u7ec3\u6587\u4ef6\u8def\u5f84\uff1a\u5728 <code>configs\/tangoflux_config.yaml<\/code> \u4e2d\u6307\u5b9a\u8bad\u7ec3\u6587\u4ef6\u8def\u5f84\u548c\u6a21\u578b\u8d85\u53c2\u6570\u3002<\/li>\n<li>\u8fd0\u884c\u8bad\u7ec3\u811a\u672c\uff1a\u4f7f\u7528\u4ee5\u4e0b\u547d\u4ee4\u542f\u52a8\u8bad\u7ec3\uff1a<\/li>\n<\/ul>\n<pre><code> CUDA_VISIBLE_DEVICES=0,1 accelerate launch --config_file='configs\/accelerator_config.yaml' src\/train.py --checkpointing_steps=\"best\" --save_every=5 --config='configs\/tangoflux_config.yaml'\r\n<\/code><\/pre>\n<ul>\n<li>DPO \u8bad\u7ec3\uff1a\u4fee\u6539\u8bad\u7ec3\u6587\u4ef6\u4ee5\u5305\u542b &#8220;chosen&#8221;, &#8220;reject&#8221;, &#8220;caption&#8221; \u548c &#8220;duration&#8221; \u5b57\u6bb5\uff0c\u5e76\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4\uff1a<\/li>\n<\/ul>\n<pre><code> CUDA_VISIBLE_DEVICES=0,1 accelerate launch --config_file='configs\/accelerator_config.yaml' src\/train_dpo.py --checkpointing_steps=\"best\" --save_every=5 --config='configs\/tangoflux_config.yaml'\r\n<\/code><\/pre>\n<\/li>\n<li><strong>\u6a21\u578b\u63a8\u7406<\/strong>\uff1a\n<ul>\n<li>\u4e0b\u8f7d\u6a21\u578b\uff1a\u786e\u4fdd\u5df2\u4e0b\u8f7d TangoFlux \u6a21\u578b\u3002<\/li>\n<li>\u751f\u6210\u97f3\u9891\uff1a\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u4ece\u6587\u672c\u63d0\u793a\u751f\u6210\u97f3\u9891\uff1a<\/li>\n<\/ul>\n<pre><code> import torchaudio\r\nfrom tangoflux import TangoFluxInference\r\nfrom IPython.display import Audio\r\nmodel = TangoFluxInference(name='declare-lab\/TangoFlux')\r\naudio = model.generate(\"\u751f\u6210\u97f3\u9891\u7684\u6587\u672c\u63d0\u793a\", duration=10)\r\nAudio(audio, rate=44100)\r\n<\/code><\/pre>\n<\/li>\n<\/ol>\n<h3>\u8be6\u7ec6\u529f\u80fd\u64cd\u4f5c<\/h3>\n<ul>\n<li><strong>\u6587\u672c\u5230\u97f3\u9891\u751f\u6210<\/strong>\uff1a\u8f93\u5165\u6587\u672c\u63d0\u793a\uff0c\u8bbe\u7f6e\u751f\u6210\u97f3\u9891\u7684\u6301\u7eed\u65f6\u95f4\uff081 \u5230 30 \u79d2\uff09\uff0c\u6a21\u578b\u5c06\u751f\u6210\u76f8\u5e94\u7684\u9ad8\u8d28\u91cf\u97f3\u9891\u3002<\/li>\n<li><strong>\u504f\u597d\u4f18\u5316<\/strong>\uff1a\u901a\u8fc7 CRPO \u6280\u672f\uff0c\u6a21\u578b\u80fd\u591f\u751f\u6210\u66f4\u7b26\u5408\u7528\u6237\u504f\u597d\u7684\u97f3\u9891\u3002<\/li>\n<li><strong>\u591a\u9636\u6bb5\u8bad\u7ec3<\/strong>\uff1a\u5305\u62ec\u9884\u8bad\u7ec3\u3001\u5fae\u8c03\u548c\u504f\u597d\u4f18\u5316\u4e09\u4e2a\u9636\u6bb5\uff0c\u786e\u4fdd\u6a21\u578b\u751f\u6210\u7684\u97f3\u9891\u8d28\u91cf\u548c\u4e00\u81f4\u6027\u3002<\/li>\n<\/ul>\n<h3>\u6ce8\u610f\u4e8b\u9879<\/h3>\n<ul>\n<li><strong>\u786c\u4ef6\u8981\u6c42<\/strong>\uff1a\u5efa\u8bae\u4f7f\u7528\u5177\u6709\u8f83\u9ad8\u8ba1\u7b97\u80fd\u529b\u7684 GPU\uff08\u5982 A40\uff09\u4ee5\u83b7\u5f97\u6700\u4f73\u6027\u80fd\u3002<\/li>\n<li><strong>\u6570\u636e\u51c6\u5907<\/strong>\uff1a\u786e\u4fdd\u8bad\u7ec3\u6570\u636e\u7684\u591a\u6837\u6027\u548c\u8d28\u91cf\uff0c\u4ee5\u63d0\u9ad8\u6a21\u578b\u7684\u751f\u6210\u6548\u679c\u3002<\/li>\n<\/ul>\n<p>\u901a\u8fc7\u4ee5\u4e0a\u6b65\u9aa4\uff0c\u7528\u6237\u53ef\u4ee5\u5feb\u901f\u4e0a\u624b\u4f7f\u7528 TangoFlux \u8fdb\u884c\u9ad8\u8d28\u91cf\u7684\u6587\u672c\u5230\u97f3\u9891\u8f6c\u6362\u3002\u8be6\u7ec6\u7684\u5b89\u88c5\u548c\u4f7f\u7528\u8bf4\u660e\u786e\u4fdd\u7528\u6237\u80fd\u591f\u987a\u5229\u5b8c\u6210\u6a21\u578b\u7684\u8bad\u7ec3\u548c\u63a8\u7406\u8fc7\u7a0b\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"<p>TangoFlux \u662f\u4e00\u4e2a\u7531 DeCLaRe Lab \u5f00\u53d1\u7684\u9ad8\u6548\u6587\u672c\u5230\u97f3\u9891\uff08TTA\uff09\u751f\u6210\u6a21\u578b\u3002\u8be5\u6a21\u578b\u80fd\u591f\u5728\u77ed\u77ed 3.7 \u79d2\u5185\u751f\u6210\u957f\u8fbe 30 \u79d2\u7684 44.1kHz \u7acb\u4f53\u58f0\u97f3\u9891\u3002TangoFlux \u91c7\u7528\u6d41\u5339\u914d\u548c Clap-Ranked Pref&#8230;<\/p>\n","protected":false},"author":1,"featured_media":61548,"comment_status":"open","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[230,206],"class_list":["post-16985","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tool","tag-aikaiyuanxiangmu","tag-aiyinle"],"_links":{"self":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/16985","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/comments?post=16985"}],"version-history":[{"count":0,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/posts\/16985\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media\/61548"}],"wp:attachment":[{"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/media?parent=16985"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/categories?post=16985"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.kdjingpai.com\/en\/wp-json\/wp\/v2\/tags?post=16985"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}