{"id":10006879,"date":"2022-11-21T14:59:23","date_gmt":"2022-11-21T06:59:23","guid":{"rendered":"https:\/\/ityet.com\/?p=10006879"},"modified":"2022-11-21T14:59:23","modified_gmt":"2022-11-21T06:59:23","slug":"%e4%bd%bf%e7%94%a8easycv-mask2former%e8%bd%bb%e6%9d%be%e5%ae%9e%e7%8e%b0%e5%9b%be%e5%83%8f%e5%88%86%e5%89%b2","status":"publish","type":"post","link":"https:\/\/itmax.cn\/?p=10006879","title":{"rendered":"\u4f7f\u7528EasyCV Mask2Former\u8f7b\u677e\u5b9e\u73b0\u56fe\u50cf\u5206\u5272"},"content":{"rendered":"

\u7b80\u4ecb\uff1a\u00a0EasyCV\u53ef\u4ee5\u8f7b\u677e\u9884\u6d4b\u56fe\u50cf\u7684\u5206\u5272\u8c31\u4ee5\u53ca\u8bad\u7ec3\u5b9a\u5236\u5316\u7684\u5206\u5272\u6a21\u578b\u3002\u672c\u6587\u4e3b\u8981\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528EasyCV\u5b9e\u73b0\u5b9e\u4f8b\u5206\u5272\u3001\u5168\u666f\u5206\u5272\u548c\u8bed\u4e49\u5206\u5272\uff0c\u53ca\u76f8\u5173\u7b97\u6cd5\u601d\u60f3\u3002\u4f5c\u8005\uff1a\u8d3a\u5f18 \u8c26\u8a00 \u4e34\u5728\u5bfc\u8a00\u56fe\u50cf\u5206\u5272(Image Segmentation)\u662f\u6307\u5bf9\u56fe\u7247\u8fdb\u884c\u50cf\u7d20\u7ea7\u7684\u5206\u7c7b\uff0c\u6839\u636e\u5206\u7c7b\u7c92\u5ea6\u7684\u4e0d\u540c\u53ef\u4ee5\u5206\u4e3a\u8bed\u4e49\u5206\u5272(Semantic Segmentation)\u3001\u5b9e\u4f8b\u5206\u5272(Instance Segmentation)\u3001\u5168\u666f\u5206\u5272(Panoptic Segmentation)\u4e09\u7c7b\u3002\u56fe\u50cf\u5206\u5272\u662f\u8ba1\u7b97\u673a\u89c6\u89c9\u4e2d\u7684\u4e3b\u8981\u7814\u7a76\u65b9\u5411\u4e4b\u4e00\uff0c\u5728\u533b\u5b66\u56fe\u50cf\u5206\u6790\u3001\u81ea\u52a8\u9a7e\u9a76\u3001\u89c6\u9891\u76d1\u63a7\u3001\u589e\u5f3a\u73b0\u5b9e\u3001\u56fe\u50cf\u538b\u7f29\u7b49\u9886\u57df\u6709\u91cd\u8981\u7684\u5e94\u7528\u4ef7\u503c\u3002\u6211\u4eec\u5728EasyCV\u6846\u67b6\u4e2d\u5bf9\u8fd9\u4e09\u7c7b\u5206\u5272SOTA\u7b97\u6cd5\u8fdb\u884c\u4e86\u96c6\u6210\uff0c\u5e76\u63d0\u4f9b\u4e86\u76f8\u5173\u6a21\u578b\u6743\u91cd\u3002\u901a\u8fc7EasyCV\u53ef\u4ee5\u8f7b\u677e\u9884\u6d4b\u56fe\u50cf\u7684\u5206\u5272\u8c31\u4ee5\u53ca\u8bad\u7ec3\u5b9a\u5236\u5316\u7684\u5206\u5272\u6a21\u578b\u3002\u672c\u6587\u4e3b\u8981\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528EasyCV\u5b9e\u73b0\u5b9e\u4f8b\u5206\u5272\u3001\u5168\u666f\u5206\u5272\u548c\u8bed\u4e49\u5206\u5272\uff0c\u53ca\u76f8\u5173\u7b97\u6cd5\u601d\u60f3\u3002\u4f7f\u7528EasyCV\u9884\u6d4b\u5206\u5272\u56feEasyCV\u63d0\u4f9b\u4e86\u5728coco\u6570\u636e\u96c6\u4e0a\u8bad\u7ec3\u7684\u5b9e\u4f8b\u5206\u5272\u6a21\u578b\u548c\u5168\u666f\u5206\u5272\u6a21\u578b\u4ee5\u53ca\u5728ADE20K\u4e0a\u8bad\u7ec3\u7684\u8bed\u4e49\u5206\u5272\u6a21\u578b\uff0c\u53c2\u8003EasyCV quick start\uff08https:\/\/github.com\/alibaba\/Ea…<\/a>\uff09\u5b8c\u6210\u4f9d\u8d56\u73af\u5883\u7684\u914d\u7f6e\u540e\uff0c\u53ef\u4ee5\u76f4\u63a5\u4f7f\u7528\u8fd9\u4e9b\u6a21\u578b\u5b8c\u6210\u5bf9\u56fe\u50cf\u7684\u5206\u5272\u8c31\u9884\u6d4b\uff0c\u76f8\u5173\u6a21\u578b\u94fe\u63a5\u5728reference\u4e2d\u7ed9\u51fa\u3002\u5b9e\u4f8b\u5206\u5272\u9884\u6d4b\u7531\u4e8e\u8be5\u793a\u4f8b\u4e2d\u7684mask2fromer\u7b97\u6cd5\u4f7f\u7528\u4e86Deformable attention (\u5728DETR\u7cfb\u5217\u7b97\u6cd5\u4e2d\u4f7f\u7528\u8be5\u7b97\u5b50\u53ef\u4ee5\u6709\u6548\u63d0\u5347\u7b97\u6cd5\u6536\u655b\u901f\u5ea6\u548c\u8ba1\u7b97\u6548\u7387)\uff0c\u9700\u8981\u989d\u5916\u5bf9\u8be5\u7b97\u5b50\u8fdb\u884c\u7f16\u8bd1cd thirdparty\/deformable_attention
python setup.py build install\u901a\u8fc7Mask2formerPredictor\u9884\u6d4b\u56fe\u50cf\u5b9e\u4f8b\u5206\u5272\u56feimport cv2
from easycv.predictors.segmentation import Mask2formerPredictor<\/p>\n

predictor = Mask2formerPredictor(model_path=’mask2former_instance_export.pth’,task_mode=’instance’)
img = cv2.imread(‘000000123213.jpg’)
predict_out = predictor([‘000000123213.jpg’])
instance_img = predictor.show_instance(img, **predict_out[0])
cv2.imwrite(‘instance_out.jpg’,instance_img)\u8f93\u51fa\u7ed3\u679c\u5982\u4e0b\u56fe\uff1a
\"\u56fe\u7247\"<\/p>\n

\"\u56fe\u7247\"
\u5168\u666f\u5206\u5272\u9884\u6d4b\u901a\u8fc7Mask2formerPredictor\u9884\u6d4b\u56fe\u50cf\u5168\u666f\u5206\u5272\u56feimport cv2
from easycv.predictors.segmentation import Mask2formerPredictor<\/p>\n

predictor = Mask2formerPredictor(model_path=’mask2former_pan_export.pth’,task_mode=’panoptic’)
img = cv2.imread(‘000000123213.jpg’)
predict_out = predictor([‘000000123213.jpg’])
pan_img = predictor.show_panoptic(img, **predict_out[0])
cv2.imwrite(‘pan_out.jpg’,pan_img)\u8f93\u51fa\u7ed3\u679c\u5982\u4e0b\u56fe\uff1a
\"\u56fe\u7247\"<\/p>\n

\"\u56fe\u7247\"
\u8bed\u4e49\u5206\u5272\u9884\u6d4b\u901a\u8fc7Mask2formerPredictor\u9884\u6d4b\u56fe\u50cf\u8bed\u4e49\u5206\u5272\u56feimport cv2
from easycv.predictors.segmentation import Mask2formerPredictor<\/p>\n

predictor = Mask2formerPredictor(model_path=’mask2former_semantic_export.pth’,task_mode=’semantic’)
img = cv2.imread(‘000000123213.jpg’)
predict_out = predictor([‘000000123213.jpg’])
semantic_img = predictor.show_panoptic(img, **predict_out[0])
cv2.imwrite(‘semantic_out.jpg’,semantic_img)
\"\u56fe\u7247\"<\/p>\n

\"\u56fe\u7247\"
\u793a\u4f8b\u56fe\u7247\u6765\u6e90\uff1acocodataset\u5728\u963f\u91cc\u4e91\u673a\u5668\u5b66\u4e60\u5e73\u53f0PAI\u4e0a\u4f7f\u7528Mask2Former\u6a21\u578bPAI-DSW\uff08Data Science Workshop\uff09\u662f\u963f\u91cc\u4e91\u673a\u5668\u5b66\u4e60\u5e73\u53f0PAI\u5f00\u53d1\u7684\u4e91\u4e0aIDE\uff0c\u9762\u5411\u5404\u7c7b\u5f00\u53d1\u8005\uff0c\u63d0\u4f9b\u4e86\u4ea4\u4e92\u5f0f\u7684\u7f16\u7a0b\u73af\u5883\u3002\u5728DSW Gallery\u4e2d(\u94fe\u63a5)\uff0c\u63d0\u4f9b\u4e86\u5404\u79cdNotebook\u793a\u4f8b\uff0c\u65b9\u4fbf\u7528\u6237\u8f7b\u677e\u4e0a\u624bDSW\uff0c\u642d\u5efa\u5404\u79cd\u673a\u5668\u5b66\u4e60\u5e94\u7528\u3002\u6211\u4eec\u4e5f\u5728DSW Gallery\u4e2d\u4e0a\u67b6\u4e86Mask2Former\u8fdb\u884c\u56fe\u50cf\u5206\u5272\u7684Sample Notebook\uff08\u89c1\u4e0b\u56fe\uff09\uff0c\u6b22\u8fce\u5927\u5bb6\u4f53\u9a8c\uff01
\"\u56fe\u7247\"
Mask2Former\u7b97\u6cd5\u89e3\u8bfb\u4e0a\u8ff0\u4f8b\u5b50\u4e2d\u91c7\u7528\u7684\u6a21\u578b\u662f\u57fa\u4e8eMask2former\u5b9e\u73b0\u7684\uff0cMask2former\u662f\u4e00\u4e2a\u7edf\u4e00\u7684\u5206\u5272\u67b6\u6784\uff0c\u80fd\u591f\u540c\u65f6\u8fdb\u884c\u8bed\u4e49\u5206\u5272\u3001\u5b9e\u4f8b\u5206\u5272\u4ee5\u53ca\u5168\u666f\u5206\u5272\uff0c\u5e76\u4e14\u53d6\u5f97SOTA\u7684\u7ed3\u679c\uff0c\u5728COCO\u6570\u636e\u96c6\u4e0a\u5168\u666f\u5206\u5272\u7cbe\u5ea657.8 PQ\uff0c\u5b9e\u4f8b\u5206\u5272\u7cbe\u5ea6\u8fbe50.1 AP\uff0c\u5728ADE20K\u6570\u636e\u96c6\u4e0a\u8bed\u4e49\u5206\u5272\u7cbe\u5ea6\u8fbe57.7 mIoU\u3002
\"\u56fe\u7247\"
\u6838\u5fc3\u601d\u60f3Mask2Former\u91c7\u7528mask classification\u7684\u5f62\u5f0f\u6765\u8fdb\u884c\u5206\u5272\uff0c\u5373\u901a\u8fc7\u6a21\u578b\u53bb\u9884\u6d4b\u4e00\u7ec4\u4e8c\u503cmask\u518d\u7ec4\u5408\u6210\u6700\u7ec8\u7684\u5206\u5272\u56fe\u3002\u6bcf\u4e2a\u4e8c\u503cmask\u53ef\u4ee5\u4ee3\u8868\u7c7b\u522b\u6216\u5b9e\u4f8b\uff0c\u5c31\u53ef\u4ee5\u5b9e\u73b0\u8bed\u4e49\u5206\u5272\u3001\u5b9e\u4f8b\u5206\u5272\u7b49\u4e0d\u540c\u7684\u5206\u5272\u4efb\u52a1\u3002\u5728mask classsification\u4efb\u52a1\u4e2d\uff0c\u4e00\u4e2a\u6bd4\u8f83\u6838\u5fc3\u7684\u95ee\u9898\u662f\u5982\u4f55\u53bb\u627e\u5230\u4e00\u4e2a\u597d\u7684\u5f62\u5f0f\u5b66\u4e60\u4e8c\u503cMask\u3002\u5982\u5148\u524d\u7684\u5de5\u4f5c Mask R-CNN\u901a\u8fc7bounding boxes\u6765\u9650\u5236\u7279\u5f81\u533a\u57df\uff0c\u5728\u533a\u57df\u5185\u9884\u6d4b\u5404\u81ea\u7684\u5206\u5272\u8c31\u3002\u8fd9\u79cd\u65b9\u5f0f\u4e5f\u5bfc\u81f4Mask R-CNN\u53ea\u80fd\u8fdb\u884c\u5b9e\u4f8b\u5206\u5272\u3002Mask2Former\u53c2\u8003DETR\u7684\u65b9\u5f0f\uff0c\u901a\u8fc7\u4e00\u7ec4\u56fa\u5b9a\u6570\u91cf\u7684\u7279\u5f81\u5411\u91cf(object query)\u53bb\u8868\u793a\u4e8c\u503cMask\uff0c\u901a\u8fc7Transformer Decoder\u8fdb\u884c\u89e3\u7801\u53bb\u9884\u6d4b\u8fd9\u4e00\u7ec4Mask\u3002\uff08ps\uff1a\u5173\u4e8eDETR\u7684\u89e3\u8bfb\u53ef\u4ee5\u53c2\u8003\uff1a\u57fa\u4e8eEasyCV\u590d\u73b0DETR\u548cDAB-DETR\uff0cObject Query\u7684\u6b63\u786e\u6253\u5f00\u65b9\u5f0f\uff09\u5728DETR\u7cfb\u5217\u7684\u7b97\u6cd5\u4e2d\uff0c\u6709\u4e00\u4e2a\u6bd4\u8f83\u91cd\u8981\u7684\u7f3a\u9677\u662f\u5728Transformer Decoder\u4e2d\u7684cross attention\u4e2d\u4f1a\u5bf9\u5168\u5c40\u7684\u7279\u5f81\u8fdb\u884c\u5904\u7406\uff0c\u5bfc\u81f4\u6a21\u578b\u5f88\u96be\u5173\u6ce8\u5230\u771f\u6b63\u60f3\u8981\u5173\u6ce8\u7684\u533a\u57df\uff0c\u4f1a\u964d\u4f4e\u6a21\u578b\u7684\u6536\u655b\u901f\u5ea6\u548c\u6700\u7ec8\u7684\u7b97\u6cd5\u7cbe\u5ea6\u3002\u5bf9\u4e8e\u8fd9\u4e2a\u95ee\u9898Mask2former\u63d0\u51fa\u4e86Transformer Decoder with mask attention\uff0c\u6bcf\u4e2aTransformer Decoder block \u4f1a\u53bb\u9884\u6d4b\u4e00\u4e2aattention mask\u5e76\u4ee50.5\u4e3a\u9608\u503c\u8fdb\u884c\u4e8c\u503c\u5316\uff0c\u7136\u540e\u5c06\u8fd9\u4e2aattentino mask\u4f5c\u4e3a\u4e0b\u4e00\u4e2ablock\u7684\u8f93\u5165\uff0c\u8ba9attention\u6a21\u5757\u8ba1\u7b97\u65f6\u53ea\u5173\u6ce8\u5728mask\u7684\u524d\u666f\u90e8\u5206\u3002\u6a21\u578b\u7ed3\u6784
\"\u56fe\u7247\"
Mask2Former\u7531\u4e09\u4e2a\u90e8\u5206\u7ec4\u6210\uff1aBackbone\uff08ResNet\u3001Swin Transformer\uff09\u4ece\u56fe\u7247\u4e2d\u62bd\u53d6\u4f4e\u5206\u8fa8\u7387\u7279\u5f81Pixel Decoder \u4ece\u4f4e\u5206\u8fa9\u7387\u7279\u5f81\u4e2d\u9010\u6b65\u8fdb\u884c\u4e0a\u91c7\u6837\u89e3\u7801\uff0c\u83b7\u5f97\u4ece\u4f4e\u5206\u8fa8\u7387\u5230\u9ad8\u5206\u8fa8\u7387\u7684\u7279\u5f81\u91d1\u5b57\u5854\uff0c\u5faa\u73af\u7684\u4f5c\u4e3aTransformer Decoder\u4e2dV\u3001K\u7684\u8f93\u5165\u3002\u901a\u8fc7\u591a\u5c3a\u5ea6\u7684\u7279\u5f81\u6765\u4fdd\u8bc1\u6a21\u578b\u5bf9\u4e0d\u540c\u5c3a\u5ea6\u7684\u76ee\u6807\u7684\u9884\u6d4b\u7cbe\u5ea6\u3002\u5176\u4e2d\u4e00\u5c42\u7684Trasformer\u4ee3\u7801\u5982\u4e0b\u6240\u793a\uff08ps\uff1a\u4e3a\u4e86\u8fdb\u4e00\u6b65\u52a0\u901f\u6a21\u578b\u7684\u6536\u655b\u901f\u5ea6\uff0c\u5728Pixel Decoder\u4e2d\u91c7\u7528\u4e86Deformable attention\u6a21\u5757\uff09\uff1aclass MSDeformAttnTransformerEncoderLayer(nn.Module):<\/p>\n

def __init__(self,\n             d_model=256,\n             d_ffn=1024,\n             dropout=0.1,\n             activation='relu',\n             n_levels=4,\n             n_heads=8,\n             n_points=4):\n                 super().__init__()\n\n                 # self attention\n                 self.self_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)\n                 self.dropout1 = nn.Dropout(dropout)\n                 self.norm1 = nn.LayerNorm(d_model)\n\n                 # ffn\n                 self.linear1 = nn.Linear(d_model, d_ffn)\n                 self.activation = _get_activation_fn(activation)\n                 self.dropout2 = nn.Dropout(dropout)\n                 self.linear2 = nn.Linear(d_ffn, d_model)\n                 self.dropout3 = nn.Dropout(dropout)\n                 self.norm2 = nn.LayerNorm(d_model)\n\n@staticmethod\ndef with_pos_embed(tensor, pos):\n    return tensor if pos is None else tensor + pos\n\ndef forward_ffn(self, src):\n    src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))\n    src = src + self.dropout3(src2)\n    src = self.norm2(src)\n    return src\n\ndef forward(self,\n            src,\n            pos,\n            reference_points,\n            spatial_shapes,\n            level_start_index,\n            padding_mask=None):\n                # self attention\n                src2 = self.self_attn(\n                    self.with_pos_embed(src, pos), reference_points, src,\n                    spatial_shapes, level_start_index, padding_mask)\n                src = src + self.dropout1(src2)\n                src = self.norm1(src)\n\n                # ffn\n                src = self.forward_ffn(src)\n\n                return srcTransformer Decoder with mask attention \u901a\u8fc7Object query\u548cPixel Decoder\u4e2d\u5f97\u5230\u7684Multi-scale feature\u53bb\u9010\u5c42\u53bbrefine\u4e8c\u503cmask\u56fe\uff0c\u5f97\u5230\u6700\u7ec8\u7684\u7ed3\u679c\u3002\u5176\u4e2d\u6838\u5fc3\u7684mask cross attention\uff0c\u4f1a\u5c06\u524d\u4e00\u5c42\u7684\u9884\u6d4b\u7684mask\u4f5c\u4e3aMultiheadAttention\u7684atten_mask\u8f93\u5165\uff0c\u4ee5\u6b64\u6765\u5c06\u6ce8\u610f\u529b\u7684\u8ba1\u7b97\u9650\u5236\u5728\u8fd9\u4e2aquery\u5173\u6ce8\u7684\u524d\u666f\u4e2d\u3002\u5177\u4f53\u5b9e\u73b0\u4ee3\u7801\u5982\u4e0b\uff1aclass CrossAttentionLayer(nn.Module):\n\ndef __init__(self,\n             d_model,\n             nhead,\n             dropout=0.0,\n             activation='relu',\n             normalize_before=False):\n    super().__init__()\n    self.multihead_attn = nn.MultiheadAttention(\n        d_model, nhead, dropout=dropout)\n\n    self.norm = nn.LayerNorm(d_model)\n    self.dropout = nn.Dropout(dropout)\n\n    self.activation = _get_activation_fn(activation)\n    self.normalize_before = normalize_before\n\n    self._reset_parameters()\n\ndef _reset_parameters(self):\n    for p in self.parameters():\n        if p.dim() > 1:\n            nn.init.xavier_uniform_(p)\n\ndef with_pos_embed(self, tensor, pos: Optional[Tensor]):\n    return tensor if pos is None else tensor + pos\n\ndef forward_post(self,\n                 tgt,\n                 memory,\n                 memory_mask: Optional[Tensor] = None,\n                 memory_key_padding_mask: Optional[Tensor] = None,\n                 pos: Optional[Tensor] = None,\n                 query_pos: Optional[Tensor] = None):\n    tgt2 = self.multihead_attn(\n        query=self.with_pos_embed(tgt, query_pos),\n        key=self.with_pos_embed(memory, pos),\n        value=memory,\n        attn_mask=memory_mask,\n        key_padding_mask=memory_key_padding_mask)[0]\n    tgt = tgt + self.dropout(tgt2)\n    tgt = self.norm(tgt)\n\n    return tgt\n\ndef forward_pre(self,\n                tgt,\n                memory,\n                memory_mask: Optional[Tensor] = None,\n                memory_key_padding_mask: Optional[Tensor] = None,\n                pos: Optional[Tensor] = None,\n                query_pos: Optional[Tensor] = None):\n    tgt2 = self.norm(tgt)\n    tgt2 = self.multihead_attn(\n        query=self.with_pos_embed(tgt2, query_pos),\n        key=self.with_pos_embed(memory, pos),\n        value=memory,\n        attn_mask=memory_mask,\n        key_padding_mask=memory_key_padding_mask)[0]\n    tgt = tgt + self.dropout(tgt2)\n\n    return tgt\n\ndef forward(self,\n            tgt,\n            memory,\n            memory_mask: Optional[Tensor] = None,\n            memory_key_padding_mask: Optional[Tensor] = None,\n            pos: Optional[Tensor] = None,\n            query_pos: Optional[Tensor] = None):\n    if self.normalize_before:\n        return self.forward_pre(tgt, memory, memory_mask,\n                                memory_key_padding_mask, pos, query_pos)\n    return self.forward_post(tgt, memory, memory_mask,\n                             memory_key_padding_mask, pos, query_pos)Tricks1.efficient multi-scale strategy\u5728pixel decoder\u4e2d\u4f1a\u89e3\u7801\u5f97\u5230\u5c3a\u5ea6\u4e3a\u539f\u56fe1\/32\u30011\/16\u30011\/8\u7684\u7279\u5f81\u91d1\u5b57\u5854\u4f9d\u6b21\u4f5c\u4e3a\u5bf9\u5e94transformer decoder block\u7684K\u3001V\u7684\u8f93\u5165\u3002\u53c2\u7167deformable detr\u7684\u505a\u6cd5\uff0c\u5bf9\u6bcf\u4e2a\u8f93\u5165\u90fd\u52a0\u4e0a\u4e86sinusoidal positional embedding\u548clearnable scale-level embedding\u3002\u6309\u5206\u8fa8\u7387\u4ece\u4f4e\u5230\u9ad8\u7684\u5faa\u5e8f\u4f9d\u6b21\u8f93\u5165\uff0c\u5e76\u5faa\u73afL\u6b21\u30022.PointRend\u901a\u8fc7PointRend\u7684\u65b9\u5f0f\u6765\u8282\u7701\u8bad\u7ec3\u8fc7\u7a0b\u4e2d\u7684\u5185\u5b58\u6d88\u8017\uff0c\u4e3b\u8981\u4f53\u73b0\u5728\u4e24\u4e2a\u90e8\u5206a.\u5728\u4f7f\u7528\u5308\u7259\u5229\u7b97\u6cd5\u5339\u914d\u9884\u6d4bmask\u548c\u771f\u503c\u6807\u7b7e\u65f6\uff0c\u901a\u8fc7\u5747\u5300\u91c7\u6837\u7684K\u4e2a\u70b9\u96c6\u4ee3\u66ff\u5b8c\u6574\u7684mask\u56fe\u6765\u8ba1\u7b97match cost b.\u5728\u8ba1\u7b97\u635f\u5931\u65f6\u6309\u7167importance sampling\u7b56\u7565\u91c7\u6837\u7684K\u4e2a\u70b9\u96c6\u4ee3\u66ff\u5b8c\u6574\u7684mask\u56fe\u6765\u8ba1\u7b97loss\uff08ps\u5b9e\u9a8c\u8bc1\u660e\u57fa\u4e8epointreind\u65b9\u5f0f\u6765\u8ba1\u7b97\u635f\u5931\u80fd\u591f\u6709\u6548\u63d0\u5347\u6a21\u578b\u7cbe\u5ea6\uff093.Optimization improvements\u66f4\u6362\u4e86self-attention\u548ccross-attention\u7684\u987a\u5e8f\u3002self-attention->cross-attention\u53d8\u6210cross-attention->self-attention\u3002\u8ba9query\u53d8\u6210\u53ef\u5b66\u4e60\u7684\u53c2\u6570\u3002\u8ba9query\u8fdb\u884c\u76d1\u7763\u5b66\u4e60\u53ef\u4ee5\u8d77\u5230\u7c7b\u4f3cregion proposal\u7684\u4f5c\u7528\u3002\u901a\u8fc7\u5b9e\u9a8c\u53ef\u4ee5\u8bc1\u660e\u53ef\u5b66\u4e60\u7684query\u53ef\u4ee5\u4ea7\u751fmask proposal\u3002\u53bb\u6389\u4e86transformer deocder\u4e2d\u7684dropout\u64cd\u4f5c\u3002\u901a\u8fc7\u5b9e\u9a8c\u53d1\u73b0\u8fd9\u4e2a\u64cd\u4f5c\u4f1a\u964d\u4f4e\u7cbe\u5ea6\u3002\u590d\u73b0\u7cbe\u5ea6\u5b9e\u4f8b\u5206\u5272\u53ca\u5168\u666f\u5206\u5272\u5728COCO\u4e0a\u7684\u590d\u73b0\u7cbe\u5ea6\uff0c\u5b9e\u9a8c\u5728\u5355\u673a8\u5361A100\u73af\u5883\u4e0b\u8fdb\u884c(ps :\u5173\u4e8e\u5b9e\u4f8b\u5206\u5272\u590d\u73b0\u7cbe\u5ea6\u95ee\u9898\u5728\u5b98\u65b9repo issue 46\u4e2d\u6709\u63d0\u53ca)ModelPQBox mAPMask mAPmemorytrain_timemask2former_r50_instance_official\u00a0\u00a043.7\u00a0\u00a0mask2former_r50_8xb2_epoch50_instance\u00a046.0943.2613G3day2hmask2former_r50_panoptic_official51.9\u00a041.7\u00a0\u00a0mask2former_r50_8xb2_epoch50_panoptic51.6444.8141.8813G3day4h\u00a0\u8bed\u4e49\u5206\u5272\u5728ADE20K\u6570\u636e\u96c6\u4e0a\u8fdb\u884c\u590d\u73b0ModelmIoUtrain memorytrain_timemask2former_r50_semantic_official47.2\u00a0\u00a0mask2former_r50_8xb2_e127_samantic47.035.6G15h35m\u00a0\u4f7f\u7528EasyCV\u8bad\u7ec3\u5206\u5272\u6a21\u578b\u5bf9\u4e8e\u7279\u5b9a\u573a\u666f\u7684\u5206\u5272\uff0c\u53ef\u4ee5\u4f7f\u7528EasyCV\u6846\u67b6\u548c\u76f8\u5e94\u6570\u636e\u8bad\u7ec3\u5b9a\u5236\u5316\u7684\u5206\u5272\u6a21\u578b\u3002\u8fd9\u91cc\u4ee5\u5b9e\u4f8b\u5206\u5272\u4e3a\u4f8b\u5b50\uff0c\u4ecb\u7ecd\u8bad\u7ec3\u6d41\u7a0b\u3002\u4e00\u3001\u6570\u636e\u51c6\u5907\u76ee\u524dEasyCV\u652f\u6301COCO\u5f62\u5f0f\u7684\u6570\u636e\u683c\u5f0f\uff0c\u6211\u4eec\u63d0\u4f9b\u4e86\u793a\u4f8bCOCO\u6570\u636e\u7528\u4e8e\u5feb\u901f\u8d70\u901a\u6d41\u7a0b\u3002wget http:\/\/pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com\/data\/small_coco_demo\/small_coco_demo.tar.gz && tar -zxf small_coco_demo.tar.gz\n<\/code><\/pre>\n
mkdir -p data\/  && mv small_coco_demo data\/coco\u4e8c\u3001\u6a21\u578b\u8bad\u7ec3\u5728EasyCV\u7684config\u6587\u4ef6\u5939\u4e0b\uff0c\u6211\u4eec\u63d0\u4f9b\u4e86mask2former\u7684\u6570\u636e\u5904\u7406\u548c\u6a21\u578b\u8bad\u7ec3\u53ca\u9a8c\u8bc1\u7684\u914d\u7f6e\u6587\u4ef6(configs\/segmentation\/mask2former\/mask2former_r50_8xb2_e50_instance.py),\u6839\u636e\u9700\u8981\u4fee\u6539\u9884\u6d4b\u7684\u7c7b\u522b\u3001\u6570\u636e\u8def\u5f84\u3002\u6267\u884c\u8bad\u7ec3\u547d\u4ee4\uff0c\u5982\u4e0b\u6240\u793a\uff1a#\u5355\u673a\u516b\u5361
python -m torch.distributed.launch –nproc_per_node=8 –master_port 11111 tools\/train.py \/<\/p>\n
                                    configs\/segmentation\/mask2former\/mask2former_r50_8xb2_e50_instance.py \/\n                                    --launcher pytorch \/\n                                    --work_dir experiments\/mask2former_instance \/\n                                    --fp16 \u6a21\u578b\u5bfc\u51fa\uff0c\u5c06config\u6587\u4ef6\u4fdd\u5b58\u5230\u6a21\u578b\u4e2d\uff0c\u4ee5\u4fbf\u5728predictor\u4e2d\u5f97\u5230\u6a21\u578b\u548c\u6570\u636e\u5904\u7406\u7684\u914d\u7f6e\uff0c\u5bfc\u51fa\u540e\u7684\u6a21\u578b\u5c31\u53ef\u76f4\u63a5\u7528\u4e8e\u5206\u5272\u56fe\u7684\u9884\u6d4b\u3002python tools\/export.py configs\/segmentation\/mask2former\/mask2former_r50_8xb2_e50_instance.py epoch_50.pth mask2former_instance_export.pthReference\u5b9e\u4f8b\u5206\u5272\u6a21\u578b\uff1ahttp:\/\/pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com\/EasyCV\/modelzoo\/segmentation\/mask2former_r50_instance\/mask2former_instance_export.pth\u5168\u666f\u5206\u5272\u6a21\u578b\uff1ahttp:\/\/pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com\/EasyCV\/modelzoo\/segmentation\/mask2former_r50_panoptic\/mask2former_pan_export.pth\u8bed\u4e49\u5206\u5272\u6a21\u578b\uff1ahttp:\/\/pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com\/EasyCV\/modelzoo\/segmentation\/mask2former_r50_semantic\/mask2former_semantic_export.pthEasyCV\u5f80\u671f\u5206\u4eabEasyCV\u5f00\u6e90\u5730\u5740\uff1ahttps:\/\/github.com\/alibaba\/EasyCVEasyCV DataHub \u63d0\u4f9b\u591a\u9886\u57df\u89c6\u89c9\u6570\u636e\u96c6\u4e0b\u8f7d\uff0c\u52a9\u529b\u6a21\u578b\u751f\u4ea7\u00a0https:\/\/zhuanlan.zhihu.com\/p\/572593950EasyCV\u5e26\u4f60\u590d\u73b0\u66f4\u597d\u66f4\u5feb\u7684\u81ea\u76d1\u7763\u7b97\u6cd5-FastConvMAE\u00a0https:\/\/zhuanlan.zhihu.com\/p\/566988235\u57fa\u4e8eEasyCV\u590d\u73b0DETR\u548cDAB-DETR\uff0cObject Query\u7684\u6b63\u786e\u6253\u5f00\u65b9\u5f0f\u00a0https:\/\/zhuanlan.zhihu.com\/p\/543129581\u57fa\u4e8eEasyCV\u590d\u73b0ViTDet\uff1a\u5355\u5c42\u7279\u5f81\u8d85\u8d8aFPN\u00a0https:\/\/zhuanlan.zhihu.com\/p\/528733299MAE\u81ea\u76d1\u7763\u7b97\u6cd5\u4ecb\u7ecd\u548c\u57fa\u4e8eEasyCV\u7684\u590d\u73b0\u00a0https:\/\/zhuanlan.zhihu.com\/p\/515859470EasyCV\u5f00\u6e90\uff5c\u5f00\u7bb1\u5373\u7528\u7684\u89c6\u89c9\u81ea\u76d1\u7763+Transformer\u7b97\u6cd5\u5e93\u00a0https:\/\/zhuanlan.zhihu.com\/p\/50521999ENDEasyCV\u4f1a\u6301\u7eed\u8fdb\u884cSOTA\u8bba\u6587\u590d\u73b0\u8fdb\u884c\u7cfb\u5217\u7684\u5de5\u4f5c\u4ecb\u7ecd\uff0c\u6b22\u8fce\u5927\u5bb6\u5173\u6ce8\u548c\u4f7f\u7528\uff0c\u6b22\u8fce\u5927\u5bb6\u5404\u79cd\u7ef4\u5ea6\u7684\u53cd\u9988\u548c\u6539\u8fdb\u5efa\u8bae\u4ee5\u53ca\u6280\u672f\u8ba8\u8bba\uff0c\u540c\u65f6\u6211\u4eec\u5341\u5206\u6b22\u8fce\u548c\u671f\u5f85\u5bf9\u5f00\u6e90\u793e\u533a\u5efa\u8bbe\u611f\u5174\u8da3\u7684\u540c\u884c\u4e00\u8d77\u53c2\u4e0e\u5171\u5efa\u3002\u539f\u6587\u94fe\u63a5\uff1ahttps:\/\/click.aliyun.com\/m\/1000364647\/\u672c\u6587\u4e3a\u963f\u91cc\u4e91\u539f\u521b\u5185\u5bb9\uff0c\u672a\u7ecf\u5141\u8bb8\u4e0d\u5f97\u8f6c\u8f7d\u3002\n\n<\/code><\/pre>\n
https:\/\/segmentfault.com\/a\/1190000042856363<\/a><\/p>\n","protected":false},"excerpt":{"rendered":"
\u7b80\u4ecb\uff1a\u00a0EasyCV\u53ef\u4ee5\u8f7b\u677e\u9884\u6d4b\u56fe\u50cf\u7684\u5206\u5272\u8c31\u4ee5\u53ca\u8bad\u7ec3\u5b9a\u5236\u5316\u7684\u5206\u5272\u6a21\u578b\u3002\u672c\u6587\u4e3b\u8981\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528EasyCV\u5b9e\u73b0\u5b9e\u4f8b […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_jetpack_memberships_contains_paid_content":false,"footnotes":""},"categories":[4],"tags":[],"class_list":["post-10006879","post","type-post","status-publish","format-standard","hentry","category-share"],"jetpack_sharing_enabled":true,"jetpack_featured_media_url":"","_links":{"self":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006879","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=10006879"}],"version-history":[{"count":0,"href":"https:\/\/itmax.cn\/index.php?rest_route=\/wp\/v2\/posts\/10006879\/revisions"}],"wp:attachment":[{"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=10006879"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=10006879"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/itmax.cn\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=10006879"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}