{"id":15356,"date":"2023-07-15T14:53:56","date_gmt":"2023-07-15T06:53:56","guid":{"rendered":"http:\/\/139.9.1.231\/?p=15356"},"modified":"2023-07-15T14:58:11","modified_gmt":"2023-07-15T06:58:11","slug":"denoising-diffusion-implicit-modelsddim","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2023\/07\/15\/denoising-diffusion-implicit-modelsddim\/","title":{"rendered":"\u53bb\u566a\u6269\u6563\u9690\u5f0f\u6a21\u578b\uff08Denoising Diffusion Implicit Models,DDIM\uff09"},"content":{"rendered":"\n

Paper: https:\/\/arxiv.org\/abs\/2010.02502<\/a><\/strong><\/p>\n\n\n\n

Code: https:\/\/github.com\/ermongroup\/ddim<\/a><\/strong><\/p>\n\n\n\n

\u6458\u81ea\uff1a\u6269\u6563\u6a21\u578b\u4e4bDDIM<\/a><\/p>\n\n\n\n

\u5728 DDPM \u4e2d\uff0c\u751f\u6210\u8fc7\u7a0b\u88ab\u5b9a\u4e49\u4e3a\u9a6c\u5c14\u53ef\u592b\u6269\u6563\u8fc7\u7a0b\u7684\u53cd\u5411\u8fc7\u7a0b\uff0c\u5728\u9006\u5411\u91c7\u6837\u8fc7\u7a0b\u7684\u6bcf\u4e00\u6b65\uff0c\u6a21\u578b\u9884\u6d4b\u566a\u58f0<\/p>\n\n\n\n

DDIM \u7684\u4f5c\u8005\u53d1\u73b0\uff0c\u6269\u6563\u8fc7\u7a0b\u5e76\u4e0d\u662f\u5fc5\u987b\u9075\u5faa\u9a6c\u5c14\u79d1\u592b\u94fe\uff0c \u5728\u4e4b\u540e\u7684\u57fa\u4e8e\u5206\u6570\u7684\u6269\u6563\u6a21\u578b\u4ee5\u53ca\u57fa\u4e8e\u968f\u673a\u5fae\u5206\u7b49\u5f0f\u7684\u7406\u8bba\u90fd\u6709\u76f8\u540c\u7684\u7ed3\u8bba\u3002 \u57fa\u4e8e\u6b64\uff0cDDIM \u7684\u4f5c\u8005\u91cd\u65b0\u5b9a\u4e49\u4e86\u6269\u6563\u8fc7\u7a0b\u548c\u9006\u8fc7\u7a0b\uff0c\u5e76\u63d0\u51fa\u4e86\u4e00\u79cd\u65b0\u7684\u91c7\u6837\u6280\u5de7\uff0c \u53ef\u4ee5\u5927\u5e45\u51cf\u5c11\u91c7\u6837\u7684\u6b65\u9aa4\uff0c\u6781\u5927\u7684\u63d0\u9ad8\u4e86\u56fe\u50cf\u751f\u6210\u7684\u6548\u7387\uff0c\u4ee3\u4ef7\u662f\u727a\u7272\u4e86\u4e00\u5b9a\u7684\u591a\u6837\u6027\uff0c \u56fe\u50cf\u8d28\u91cf\u7565\u5fae\u4e0b\u964d\uff0c\u4f46\u5728\u53ef\u63a5\u53d7\u7684\u8303\u56f4\u5185\u3002<\/p>\n\n\n\n

\u5bf9\u4e8e\u6269\u6563\u6a21\u578b\u6765\u8bf4\uff0c\u4e00\u4e2a\u6700\u5927\u7684\u7f3a\u70b9\u662f\u9700\u8981\u8bbe\u7f6e\u8f83\u957f\u7684\u6269\u6563\u6b65\u6570\u624d\u80fd\u5f97\u5230\u597d\u7684\u6548\u679c\uff0c\u8fd9\u5bfc\u81f4\u4e86\u751f\u6210\u6837\u672c\u7684\u901f\u5ea6\u8f83\u6162<\/strong>\uff0c\u6bd4\u5982\u6269\u6563\u6b65\u6570\u4e3a1000\u7684\u8bdd\uff0c\u90a3\u4e48\u751f\u6210\u4e00\u4e2a\u6837\u672c\u5c31\u8981\u6a21\u578b\u63a8\u74061000\u6b21\u3002\u8fd9\u7bc7\u6587\u7ae0\u6211\u4eec\u5c06\u4ecb\u7ecd\u53e6\u5916\u4e00\u79cd\u6269\u6563\u6a21\u578bDDIM<\/strong>\uff08Denoising Diffusion Implicit Models<\/a>\uff09\uff0cDDIM\u548cDDPM\u6709\u76f8\u540c\u7684\u8bad\u7ec3\u76ee\u6807<\/strong>\uff0c\u4f46\u662f\u5b83\u4e0d\u518d\u9650\u5236\u6269\u6563\u8fc7\u7a0b\u5fc5\u987b\u662f\u4e00\u4e2a\u9a6c\u5c14\u5361\u592b\u94fe\uff0c\u8fd9\u4f7f\u5f97DDIM\u53ef\u4ee5\u91c7\u7528\u66f4\u5c0f\u7684\u91c7\u6837\u6b65\u6570\u6765\u52a0\u901f\u751f\u6210\u8fc7\u7a0b<\/strong>\uff0cDDIM\u7684\u53e6\u5916\u662f\u4e00\u4e2a\u7279\u70b9\u662f\u4ece\u4e00\u4e2a\u968f\u673a\u566a\u97f3\u751f\u6210\u6837\u672c\u7684\u8fc7\u7a0b\u662f\u4e00\u4e2a\u786e\u5b9a\u7684\u8fc7\u7a0b<\/strong>\uff08\u4e2d\u95f4\u6ca1\u6709\u52a0\u5165\u968f\u673a\u566a\u97f3\uff09\u3002<\/p>\n\n\n\n

\u524d\u63d0\u6761\u4ef6\uff1a1.\u9a6c\u5c14\u53ef\u592b\u8fc7\u7a0b\u30022.\u5fae\u5c0f\u566a\u58f0\u53d8\u5316\u3002<\/strong><\/p>\n\n\n\n

\u6b65\u9aa4\u4e00\uff1a<\/strong>\u5728DDPM\u4e2d\u6211\u4eec\u57fa\u4e8e\u521d\u59cb\u56fe\u50cf\u72b6\u6001\u4ee5\u53ca\u6700\u7ec8\u9ad8\u65af\u566a\u58f0\u72b6\u6001\uff0c\u901a\u8fc7\u8d1d\u53f6\u65af\u516c\u5f0f\u4ee5\u53ca\u591a\u5143\u9ad8\u65af\u5206\u5e03\u7684\u6563\u5ea6\u516c\u5f0f\uff0c\u53ef\u4ee5\u8ba1\u7b97\u51fa\u6bcf\u4e00\u6b65\u9aa4\u7684\u9006\u5411\u5206\u5e03\u3002\u4e4b\u540e\u7ee7\u7eed\u91cd\u590d\u4e0a\u8ff0\u5bf9\u9006\u5411\u5206\u5e03\u7684\u6c42\u89e3\u6b65\u9aa4\uff0c\u6700\u7ec8\u5b9e\u73b0\u4ece\u7eaf\u9ad8\u65af\u566a\u58f0\uff0c\u6062\u590d\u5230\u539f\u59cb\u56fe\u7247\u7684\u6b65\u9aa4\u3002<\/p>\n\n\n\n

\u6b65\u9aa4\u4e8c\uff1a<\/strong>\u6a21\u578b\u4f18\u5316\u90e8\u5206\u901a\u8fc7\u6700\u5c0f\u5316\u5206\u5e03\u7684\u4ea4\u53c9\u71b5\uff0c\u9884\u6d4b\u51fa\u6a21\u578b\u9006\u5411\u5206\u5e03\u7684\u5747\u503c\u548c\u65b9\u5dee\uff0c\u5c06\u5176\u5e26\u5165\u6b65\u9aa4\u4e00\u4e2d\u7684\u63a8\u7406\u8fc7\u7a0b\u5373\u53ef\u3002<\/p>\n\n\n\n

\u6587\u7ae0\u4e2d\u5b58\u5728\u7684\u4e00\u4e2a\u6838\u5fc3\u95ee\u9898\u662f\uff1a\u7531\u4e8e1.\u6bcf\u4e2a\u6b65\u9aa4\u90fd\u662f\u9a6c\u5c14\u53ef\u592b\u94fe\u30022.\u6bcf\u6b21\u52a0\u7279\u5f81\u7684\u5747\u503c\u548c\u65b9\u5dee\u90fd\u9700\u8981\u63a7\u5236\u5728\u5f88\u5c0f\u7684\u8303\u56f4\u4e0b\u3002\u56e0\u6b64\u6211\u4eec\u4e0d\u5f97\u4e0d\u6bcf\u4e00\u6b65\u90fd\u8fdb\u884c\u9006\u5411\u7684\u63a8\u7406\u548c\u8fd0\u7b97\uff0c\u5bfc\u81f4\u6a21\u578b\u6574\u4f53\u8017\u65f6\u5f88\u957f\u3002\u672c\u6587\u6838\u5fc3\u9488\u5bf9\u8017\u65f6\u95ee\u9898\u8fdb\u884c\u4f18\u5316\uff0c\u4e00\u53e5\u8bdd\u603b\u7ed3\uff1a\u5728\u6ee1\u8db3DDPM\u4e2d\u9006\u5411\u63a8\u7406\u7684\u6761\u4ef6\u4e0b\uff0c\u627e\u5230\u4e00\u79cd\u7528<\/u><\/strong> xt \u548c<\/u><\/strong> x0 \u8868\u8fbe<\/u><\/strong> xt\u22121 \u4e14\u80fd\u80fd\u5927\u5e45\u51cf\u5c11\u8ba1\u7b97\u91cf\u7684\u63a8\u7406\u65b9\u5f0f\u3002<\/u><\/strong><\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n

\u4ee3\u7801\u5b9e\u73b0\uff1a<\/h3>\n\n\n\n

DDIM\u548cDDPM\u7684\u8bad\u7ec3\u8fc7\u7a0b\u4e00\u6837\uff0c\u6240\u4ee5\u53ef\u4ee5\u76f4\u63a5\u5728DDPM\u7684\u57fa\u7840\u4e0a\u52a0\u4e00\u4e2a\u65b0\u7684\u751f\u6210\u65b9\u6cd5\uff08\u8fd9\u91cc\u4e3b\u8981\u53c2\u8003\u4e86DDIM\u5b98\u65b9\u4ee3\u7801<\/a>\u4ee5\u53cadiffusers\u5e93<\/a>\uff09\uff0c\u5177\u4f53\u4ee3\u7801\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n

class GaussianDiffusion:\n    def __init__(self, timesteps=1000, beta_schedule='linear'):\n     pass\n\n    # ...<\/em>\n        \n # use ddim to sample<\/em>\n    @torch.no_grad()\n    def ddim_sample(\n        self,\n        model,\n        image_size,\n        batch_size=8,\n        channels=3,\n        ddim_timesteps=50,\n        ddim_discr_method=\"uniform\",\n        ddim_eta=0.0,\n        clip_denoised=True):\n        # make ddim timestep sequence<\/em>\n        if ddim_discr_method == 'uniform':\n            c = self.timesteps \/\/ ddim_timesteps\n            ddim_timestep_seq = np.asarray(list(range(0, self.timesteps, c)))\n        elif ddim_discr_method == 'quad':\n            ddim_timestep_seq = (\n                (np.linspace(0, np.sqrt(self.timesteps * .8), ddim_timesteps)) ** 2\n            ).astype(int)\n        else:\n            raise NotImplementedError(f'There is no ddim discretization method called \"{ddim_discr_method}\"')\n        # add one to get the final alpha values right (the ones from first scale to data during sampling)<\/em>\n        ddim_timestep_seq = ddim_timestep_seq + 1\n        # previous sequence<\/em>\n        ddim_timestep_prev_seq = np.append(np.array([0]), ddim_timestep_seq[:-1])\n        \n        device = next(model.parameters()).device\n        # start from pure noise (for each example in the batch)<\/em>\n        sample_img = torch.randn((batch_size, channels, image_size, image_size), device=device)\n        for i in tqdm(reversed(range(0, ddim_timesteps)), desc='sampling loop time step', total=ddim_timesteps):\n            t = torch.full((batch_size,), ddim_timestep_seq[i], device=device, dtype=torch.long)\n            prev_t = torch.full((batch_size,), ddim_timestep_prev_seq[i], device=device, dtype=torch.long)\n            \n            # 1. get current and previous alpha_cumprod<\/em>\n            alpha_cumprod_t = self._extract(self.alphas_cumprod, t, sample_img.shape)\n            alpha_cumprod_t_prev = self._extract(self.alphas_cumprod, prev_t, sample_img.shape)\n    \n            # 2. predict noise using model<\/em>\n            pred_noise = model(sample_img, t)\n            \n            # 3. get the predicted x_0<\/em>\n            pred_x0 = (sample_img - torch.sqrt((1. - alpha_cumprod_t)) * pred_noise) \/ torch.sqrt(alpha_cumprod_t)\n            if clip_denoised:\n                pred_x0 = torch.clamp(pred_x0, min=-1., max=1.)\n            \n            # 4. compute variance: \"sigma_t(\u03b7)\" -> see formula (16)<\/em>\n            # \u03c3_t = sqrt((1 \u2212 \u03b1_t\u22121)\/(1 \u2212 \u03b1_t)) * sqrt(1 \u2212 \u03b1_t\/\u03b1_t\u22121)<\/em>\n            sigmas_t = ddim_eta * torch.sqrt(\n                (1 - alpha_cumprod_t_prev) \/ (1 - alpha_cumprod_t) * (1 - alpha_cumprod_t \/ alpha_cumprod_t_prev))\n            \n            # 5. compute \"direction pointing to x_t\" of formula (12)<\/em>\n            pred_dir_xt = torch.sqrt(1 - alpha_cumprod_t_prev - sigmas_t**2) * pred_noise\n            \n            # 6. compute x_{t-1} of formula (12)<\/em>\n            x_prev = torch.sqrt(alpha_cumprod_t_prev) * pred_x0 + pred_dir_xt + sigmas_t * torch.randn_like(sample_img)\n\n            sample_img = x_prev\n            \n        return sample_img.cpu().numpy()\n<\/code><\/pre>\n\n\n\n
\u8fd9\u91cc\u4ee5MNIST\u6570\u636e\u96c6\u4e3a\u4f8b\uff0c\u8bad\u7ec3\u7684\u6269\u6563\u6b65\u6570\u4e3a500\uff0c\u76f4\u63a5\u91c7\u7528DDPM\uff08\u5373\u63a8\u7406500\u6b21\uff09\u751f\u6210\u7684\u6837\u672c\u5982\u4e0b\u6240\u793a\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u540c\u6837\u7684\u6a21\u578b\uff0c\u6211\u4eec\u91c7\u7528DDIM\u6765\u52a0\u901f\u751f\u6210\u8fc7\u7a0b\uff0c\u8fd9\u91ccDDIM\u7684\u91c7\u6837\u6b65\u6570\u4e3a50\uff0c\u5176\u751f\u6210\u7684\u6837\u672c\u8d28\u91cf\u548c500\u6b65\u7684DDPM\u76f8\u5f53\uff1a<\/p>\n\n\n\n
\"\"\/<\/figure>\n\n\n\n
\u5b8c\u6574\u7684\u4ee3\u7801\u793a\u4f8b\u89c1https:\/\/github.com\/xiaohu2015\/nngen<\/a>\u3002<\/p>\n\n\n\n
\u5176\u5b83\uff1a\u91cd\u5efa\u548c\u63d2\u503c<\/strong><\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\"\"<\/figure>\n\n\n\n
         \u5982\u679c\u4ece\u76f4\u89c2\u4e0a\u770b\uff0cDDIM\u7684\u52a0\u901f\u65b9\u5f0f\u975e\u5e38\u7b80\u5355\uff0c\u76f4\u63a5\u91c7\u6837\u4e00\u4e2a\u5b50\u5e8f\u5217\uff0c\u5176\u5b9e\u8bba\u6587DDPM+<\/a>\u4e5f\u91c7\u7528\u4e86\u7c7b\u4f3c\u7684\u65b9\u5f0f\u6765\u52a0\u901f\u3002\u53e6\u5916DDIM\u548c\u5176\u5b83\u6269\u6563\u6a21\u578b\u7684\u4e00\u4e2a\u8f83\u5927\u7684\u533a\u522b\u662f\u5176\u751f\u6210\u8fc7\u7a0b\u662f\u786e\u5b9a\u6027\u7684\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
Paper: https:\/\/arxiv.org\/abs\/2010.02502 Code: https:\/\/g … \u7ee7\u7eed\u9605\u8bfb\u53bb\u566a\u6269\u6563\u9690\u5f0f\u6a21\u578b\uff08Denoising Diffusion Implicit Models,DDIM\uff09<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[37,29,9],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15356"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=15356"}],"version-history":[{"count":11,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15356\/revisions"}],"predecessor-version":[{"id":15381,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/15356\/revisions\/15381"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=15356"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=15356"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=15356"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}