This is a U-Net based model to predict noise _^_0_^_.
\nU-Net is a gets it's name from the U shape in the model diagram. It processes a given image by progressively lowering (halving) the feature map resolution and then increasing the resolution. There are pass-through connection at each resolution.
\n_^_1_^_
\nThis implementation contains a bunch of modifications to original U-Net (residual blocks, multi-head attention) and also adds time-step embeddings _^_2_^_.
\n": "\u8fd9\u662f\u4e00\u4e2a\u57fa\u4e8e U-Net \u7684\u6a21\u578b\uff0c\u7528\u4e8e\u9884\u6d4b\u566a\u58f0_^_0_^_\u3002
\nU-Net \u662f\u4ece\u6a21\u578b\u56fe\u4e2d\u7684 U \u5f62\u4e2d\u83b7\u53d6\u5b83\u7684\u540d\u5b57\u3002\u5b83\u901a\u8fc7\u9010\u6b65\u964d\u4f4e\uff08\u51cf\u534a\uff09\u8981\u7d20\u56fe\u5206\u8fa8\u7387\uff0c\u7136\u540e\u63d0\u9ad8\u5206\u8fa8\u7387\u6765\u5904\u7406\u7ed9\u5b9a\u7684\u56fe\u50cf\u3002\u6bcf\u79cd\u5206\u8fa8\u7387\u90fd\u6709\u76f4\u901a\u8fde\u63a5\u3002
\n_^_1_^_
\n\u6b64\u5b9e\u73b0\u5305\u542b\u5bf9\u539f\u59cb U-Net\uff08\u6b8b\u5dee\u5757\u3001\u591a\u5934\u6ce8\u610f\uff09\u7684\u5927\u91cf\u4fee\u6539\uff0c\u8fd8\u6dfb\u52a0\u4e86\u65f6\u95f4\u6b65\u957f\u5d4c\u5165_^_2_^_\u3002
\n", "This is similar to transformer multi-head attention.
\n": "\u8fd9\u7c7b\u4f3c\u4e8e\u53d8\u538b\u5668\u591a\u5934\u7684\u5173\u6ce8\u3002
\n", "This combines _^_0_^_ and _^_1_^_. These are used in the first half of U-Net at each resolution.
\n": "\u8fd9\u7ed3\u5408\u4e86_^_0_^_\u548c_^_1_^_.\u8fd9\u4e9b\u5728U-Net\u7684\u524d\u534a\u90e8\u5206\u4ee5\u6bcf\u79cd\u5206\u8fa8\u7387\u4f7f\u7528\u3002
\n", "It combines a _^_0_^_, _^_1_^_, followed by another _^_2_^_. This block is applied at the lowest resolution of the U-Net.
\n": "\u5b83\u7ed3\u5408\u4e86_^_0_^_\u3001_^_1_^_\u3001\u540e\u8ddf\u53e6\u4e00\u4e2a_^_2_^_\u3002\u6b64\u5757\u5e94\u7528\u4e8e U-Net \u7684\u6700\u4f4e\u5206\u8fa8\u7387\u3002
\n", "A residual block has two convolution layers with group normalization. Each resolution is processed with two residual blocks.
\n": "\u6b8b\u5dee\u5757\u5177\u6709\u4e24\u4e2a\u5177\u6709\u7ec4\u5f52\u4e00\u5316\u7684\u5377\u79ef\u5c42\u3002\u6bcf\u4e2a\u5206\u8fa8\u7387\u90fd\u4f7f\u7528\u4e24\u4e2a\u6b8b\u5dee\u5757\u8fdb\u884c\u5904\u7406\u3002
\n", "_^_0_^_
\n": "_^_0_^_
\n", "This combines _^_0_^_ and _^_1_^_. These are used in the second half of U-Net at each resolution.
\n": "\u8fd9\u7ed3\u5408\u4e86_^_0_^_\u548c_^_1_^_.\u8fd9\u4e9b\u5728U-Net\u7684\u540e\u534a\u90e8\u5206\u4ee5\u6bcf\u79cd\u5206\u8fa8\u7387\u4f7f\u7528\u3002
\n", "\n": "\n", "
_^_0_^_ at the same resolution
\n": "_^_0_^_\u4ee5\u76f8\u540c\u7684\u5206\u8fa8\u7387
\n", "_^_0_^_ is not used, but it's kept in the arguments because for the attention layer function signature to match with _^_1_^_.
\n": "_^_0_^_\u672a\u4f7f\u7528\uff0c\u4f46\u5b83\u4fdd\u7559\u5728\u53c2\u6570\u4e2d\uff0c\u56e0\u4e3a\u8981\u4e0e\u6ce8\u610f\u5c42\u51fd\u6570\u7b7e\u540d\u5339\u914d_^_1_^_\u3002
\n", "_^_0_^_ will store outputs at each resolution for skip connection
\n": "_^_0_^_\u5c06\u4ee5\u6bcf\u79cd\u5206\u8fa8\u7387\u5b58\u50a8\u8f93\u51fa\u4ee5\u8fdb\u884c\u8df3\u8fc7\u8fde\u63a5
\n", "Activation
\n": "\u6fc0\u6d3b
\n", "Add _^_0_^_
\n": "\u6dfb\u52a0_^_0_^_
\n", "Add skip connection
\n": "\u6dfb\u52a0\u8df3\u8fc7\u8fde\u63a5
\n", "Add the shortcut connection and return
\n": "\u6dfb\u52a0\u5feb\u6377\u65b9\u5f0f\u8fde\u63a5\u5e76\u8fd4\u56de
\n", "Add time embeddings
\n": "\u6dfb\u52a0\u65f6\u95f4\u5d4c\u5165
\n", "Calculate scaled dot-product _^_0_^_
\n": "\u8ba1\u7b97\u7f29\u653e\u7684\u70b9\u79ef_^_0_^_
\n", "Change _^_0_^_ to shape _^_1_^_
\n": "\u6539_^_0_^_\u6210\u5f62\u72b6_^_1_^_
\n", "Change to shape _^_0_^_
\n": "\u6539\u6210\u5f62\u72b6_^_0_^_
\n", "Combine the set of modules
\n": "\u7ec4\u5408\u8fd9\u7ec4\u6a21\u5757
\n", "Create sinusoidal position embeddings same as those from the transformer
\n_^_0_^_where _^_1_^_ is _^_2_^_
\n": "\u521b\u5efa\u4e0e\u53d8\u538b\u5668\u76f8\u540c\u7684\u6b63\u5f26\u4f4d\u7f6e\u5d4c\u5165
\n_^_0_^_\u5728\u54ea_^_1_^_\u91cc_^_2_^_
\n", "Default _^_0_^_
\n": "\u9ed8\u8ba4_^_0_^_
\n", "Down sample at all resolutions except the last
\n": "\u9664\u6700\u540e\u4e00\u4e2a\u5206\u8fa8\u7387\u4e4b\u5916\u7684\u6240\u6709\u5206\u8fa8\u7387\u90fd\u5411\u4e0b\u91c7\u6837
\n", "Final block to reduce the number of channels
\n": "\u51cf\u5c11\u4fe1\u9053\u6570\u91cf\u7684\u6700\u7ec8\u533a\u5757
\n", "Final normalization and convolution
\n": "\u6700\u7ec8\u5f52\u4e00\u5316\u548c\u5377\u79ef
\n", "Final normalization and convolution layer
\n": "\u6700\u7ec8\u5f52\u4e00\u5316\u548c\u5377\u79ef\u5c42
\n", "First convolution layer
\n": "\u7b2c\u4e00\u4e2a\u5377\u79ef\u5c42
\n", "First half of U-Net
\n": "U-Net \u7684\u4e0a\u534a\u5e74
\n", "First linear layer
\n": "\u7b2c\u4e00\u4e2a\u7ebf\u6027\u5c42
\n", "For each resolution
\n": "\u5bf9\u4e8e\u6bcf\u79cd\u5206\u8fa8\u7387
\n", "Get image projection
\n": "\u83b7\u53d6\u56fe\u50cf\u6295\u5f71
\n", "Get query, key, and values (concatenated) and shape it to _^_0_^_
\n": "\u83b7\u53d6\u67e5\u8be2\u3001\u952e\u548c\u503c\uff08\u4e32\u8054\uff09\u5e76\u5c06\u5176\u8c03\u6574\u4e3a_^_0_^_
\n", "Get shape
\n": "\u5851\u9020\u8eab\u6750
\n", "Get the skip connection from first half of U-Net and concatenate
\n": "\u4ece U-Net \u7684\u524d\u534a\u90e8\u5206\u83b7\u53d6\u8df3\u8fc7\u8fde\u63a5\u5e76\u8fde\u63a5
\n", "Get time-step embeddings
\n": "\u83b7\u53d6\u65f6\u95f4\u6b65\u957f\u5d4c\u5165
\n", "Group normalization and the first convolution layer
\n": "\u7ec4\u5f52\u4e00\u5316\u548c\u7b2c\u4e00\u4e2a\u5377\u79ef\u5c42
\n", "Group normalization and the second convolution layer
\n": "\u7ec4\u5f52\u4e00\u5316\u548c\u7b2c\u4e8c\u4e2a\u5377\u79ef\u5c42
\n", "If the number of input channels is not equal to the number of output channels we have to project the shortcut connection
\n": "\u5982\u679c\u8f93\u5165\u901a\u9053\u7684\u6570\u91cf\u4e0d\u7b49\u4e8e\u8f93\u51fa\u901a\u9053\u7684\u6570\u91cf\uff0c\u6211\u4eec\u5fc5\u987b\u6295\u5f71\u5feb\u6377\u65b9\u5f0f\u8fde\u63a5
\n", "Linear layer for final transformation
\n": "\u7528\u4e8e\u6700\u7ec8\u53d8\u6362\u7684\u7ebf\u6027\u5c42
\n", "Linear layer for time embeddings
\n": "\u7528\u4e8e\u65f6\u95f4\u5d4c\u5165\u7684\u7ebf\u6027\u5c42
\n", "Middle (bottom)
\n": "\u4e2d\u95f4\uff08\u5e95\u90e8\uff09
\n", "Middle block
\n": "\u4e2d\u95f4\u65b9\u5757
\n", "Multiply by values
\n": "\u4e58\u4ee5\u503c
\n", "Normalization layer
\n": "\u5f52\u4e00\u5316\u5c42
\n", "Number of channels
\n": "\u9891\u9053\u6570\u91cf
\n", "Number of output channels at this resolution
\n": "\u6b64\u5206\u8fa8\u7387\u4e0b\u7684\u8f93\u51fa\u58f0\u9053\u6570
\n", "Number of resolutions
\n": "\u5206\u8fa8\u7387\u6570\u91cf
\n", "Project image into feature map
\n": "\u5c06\u56fe\u50cf\u6295\u5f71\u5230\u8981\u7d20\u5730\u56fe\u4e2d
\n", "Projections for query, key and values
\n": "\u67e5\u8be2\u3001\u952e\u548c\u503c\u7684\u6295\u5f71
\n", "Reshape to _^_0_^_
\n": "\u91cd\u5851\u4e3a_^_0_^_
\n", "Scale for dot-product attention
\n": "\u7f29\u653e\u70b9\u4ea7\u54c1\u6ce8\u610f\u529b
\n", "Second convolution layer
\n": "\u7b2c\u4e8c\u4e2a\u5377\u79ef\u5c42
\n", "Second half of U-Net
\n": "U-Net \u7684\u4e0b\u534a\u573a
\n", "Second linear layer
\n": "\u7b2c\u4e8c\u4e2a\u7ebf\u6027\u5c42
\n", "Softmax along the sequence dimension _^_0_^_
\n": "\u987a\u5e8f\u7ef4\u5ea6\u4e0a\u7684 Softmax_^_0_^_
\n", "Split query, key, and values. Each of them will have shape _^_0_^_
\n": "\u62c6\u5206\u67e5\u8be2\u3001\u952e\u548c\u503c\u3002\u4ed6\u4eec\u6bcf\u4e2a\u4eba\u90fd\u4f1a\u6709\u5f62\u72b6_^_0_^_
\n", "The input has _^_0_^_ because we concatenate the output of the same resolution from the first half of the U-Net
\n": "\u8f93\u5165\u4e4b_^_0_^_\u6240\u4ee5\u6709\uff0c\u662f\u56e0\u4e3a\u6211\u4eec\u5c06 U-Net \u524d\u534a\u90e8\u5206\u76f8\u540c\u5206\u8fa8\u7387\u7684\u8f93\u51fa\u8fde\u63a5\u8d77\u6765
\n", "Time embedding layer. Time embedding has _^_0_^_ channels
\n": "\u65f6\u95f4\u5d4c\u5165\u5c42\u3002\u65f6\u95f4\u5d4c\u5165\u6709_^_0_^_\u9891\u9053
\n", "Transform to _^_0_^_
\n": "\u53d8\u6362\u4e3a_^_0_^_
\n", "Transform with the MLP
\n": "\u4f7f\u7528 MLP \u8fdb\u884c\u8f6c\u578b
\n", "Up sample at all resolutions except last
\n": "\u9664\u6700\u540e\u4e00\u4e2a\u4ee5\u5916\u7684\u6240\u6709\u5206\u8fa8\u7387\u5411\u4e0a\u91c7\u6837
\n", "