![\"\u3010\u6a5f\u68b0\u5b78\u7fd2\u3011\u6559\u4f60\u7684](\"https:\/\/pic.pimg.tw\/benjenq\/1532282161-2658330607.png\" "\\"\u3010\u6a5f\u68b0\u5b78\u7fd2\u3011\u6559\u4f60\u7684")
<\/p>\n<\/p>\n
\u524d\u60c5\u63d0\u8981\uff1a\u6559\u4f60\u7684 iPhone \u8a8d\u8b58 Gogoro \u63db\u96fb\u7ad9\uff08Part 1\uff09- \u7528\u9ed1\u860b\u679c\u96fb\u8166\u73a9\u8f49\u6700\u592f\u7684\u6a5f\u68b0\u5b78\u7fd2<\/a>\u3002<\/p>\n \u5728\u524d\u4e00\u7bc7\u6587\u7ae0<\/a>\u4e2d\uff0c\u7b2c\u4e00\u6b21\u5617\u8a66\u5f9e\u7121\u5230\u6709\u5b8c\u6210\u6574\u500b\u6a5f\u68b0\u5b78\u7fd2\u5f71\u50cf\u8fa8\u8b58\u7684\u7df4\u7fd2\u904e\u7a0b\uff0c\u6700\u7d42\u5f97\u5230\u4e00\u500b\u300c\u5fc5\u9808\u4ee5\u534a\u4f5c\u5f0a\u7684\u65b9\u5f0f\u5f97\u5230\u8c8c\u4f3c\u53ef\u7528\u7684\u8a13\u7df4\u6a21\u578b\u300d\u9019\u7a2e\u4e0d\u592a\u6eff\u610f\u7684\u7d50\u679c\uff0c\u5f8c\u7e8c\u4e5f\u7559\u4e0b\u8a31\u591a\u554f\u984c\u7559\u5f85\u5f8c\u7e8c\u63a2\u8a0e\u3002\u63a5\u4e0b\u4f86\u9019\u6bb5\u6642\u9593\u88e1\uff0c\u6211\u4e00\u76f4\u4e0d\u65b7\u5617\u8a66\u5404\u7a2e\u624b\u6bb5\uff0c\u589e\u52a0\u8fa8\u8b58\u6210\u529f\u7684\u6a5f\u7387\uff0c\u4e0d\u904e\u7d42\u7a76\u6548\u679c\u6709\u9650\uff0c\u8aa4\u5224\u7684\u539f\u56e0\u4e5f\u8d8a\u4f86\u8d8a\u96e3\u6df1\u7a76\uff0c\u4ee5\u81f3\u65bc\u6559\u5b78\u8ab2\u7a0b\u8a13\u7df4\u597d\u7684\u7bc4\u4f8b\u8cc7\u6599\u53ef\u7528\uff0c\u81ea\u5df1\u8a13\u7df4\u51fa\u4f86\u8cc7\u6599\u537b\u4e0d\u80fd\u7528\u3002\u624d\u6b63\u8981\u958b\u5fc3\u7684\u5411\u524d\u8e0f\u51fa\u4e00\u6b65\uff0c\u63a5\u8457\u53c8\u5361\u95dc\u4e86\uff0c\u5be6\u5728\u5f88\u7070\u5fc3\uff01<\/p>\n \u65bc\u662f\u53c8\u518d\u5ea6\u4e0a\u8c37\u6b4c\u5927\u6d77\u4e2d\u6f02\u6d41\uff0c\u53cd\u8986\u5617\u8a66\u641c\u5c0b\u81ea\u5df1\u53ef\u80fd\u907a\u6f0f\u4e86\u4ec0\u9ebc\u95dc\u9375\u5b57\uff0c\u4e00\u6b21\u53c8\u4e00\u6b21\u7684\u641c\u5c0b\uff0c\u4e00\u6b21\u53c8\u4e00\u6b21\u7684\u89f8\u78b0\u8271\u6df1\u96e3\u61c2\u4e5f\u59cb\u7d42\u641e\u4e0d\u61c2\u7684\u6f14\u7b97\u6cd5\u8b70\u984c\u3002\u5728\u67d0\u4e00\u6b21\u7684\u8cc7\u6599\u641c\u5c0b\u4e2d\uff0c\u641c\u5230\u4e00\u5f35\u5716\u7247\uff1a<\/p>\n <\/p>\n \u4e00\u76ee\u77ad\u7136…\u9019\u4e0d\u5c31\u662f\u6211\u60f3\u8981\u7684\u8fa8\u8b58\u7d50\u679c\u55ce\uff1f\u5728\u4e0a\u6b21\u7684\u64cd\u4f5c\u7df4\u7fd2\u4e2d\uff0c\u662f\u91dd\u5c0d\u6574\u5f35\u5716\u7684\u8fa8\u8b58\uff0c\u4e5f\u5c0e\u81f4\u300cGogoro \u63db\u96fb\u7ad9\u300d\u7684\u8a13\u7df4\u8cc7\u6599\u4e2d\uff0c\u5468\u906d\u4f34\u96a8\u4e00\u5806\u5730\u666f\u5730\u7269\u800c\u5e72\u64fe\u5b78\u7fd2\u7d50\u679c\uff0c\u6211\u7e3d\u7b97\u610f\u8b58\u5230\u81ea\u5df1\u5ffd\u7565\u7684\u5730\u65b9\u3002\u9806\u8457\u9019\u5f35\u5716\u7684\u7dda\u7d22\uff0c\u627e\u5230\u5716\u7247\u7684\u4f86\u6e90\uff1a<\/p>\n \u300cTuri Create – Object detection<\/a>\u300d<\/p>\n \u5728\u9019\u88e1\u51fa\u73fe\u4e86\u65b0\u7684\u95dc\u9375\u5b57\u300cObject Detection\u300d\uff08\u7269\u9ad4\u5075\u6e2c\uff09\uff0c\u6211\u7d42\u65bc\u641e\u61c2\u5361\u95dc\u7684\u539f\u56e0\u4e86\uff01\u539f\u4f86\u6211\u4e00\u76f4\u628a\u300c\u5f71\u50cf\u8fa8\u8b58\u300d\uff08Image Classification\uff09\u8207\u300c\u7269\u9ad4\u5075\u6e2c\u300d\uff08Object Detection\uff09\u5169\u8005\u6df7\u70ba\u4e00\u8ac7\u4e86\uff0c\u6a5f\u68b0\u5b78\u7fd2\u7684\u61c9\u7528\u65b9\u5f0f\u932f\u8aa4\uff0c\u6240\u4ee5\u59cb\u7d42\u64fa\u812b\u4e0d\u4e86\u8fa8\u8b58\u7d50\u679c\u906d\u8aa4\u5224\u7684\u547d\u904b\u3002\u4f86\u6e90\u7db2\u5740\u7d66\u51fa\u53e6\u4e00\u500b\u95dc\u9375\u5b57\u300cTuri Create\u300d\uff0c\u4e5f\u627e\u5230\u4e86\u300cTuri Create\u300d\u5c08\u6848\u7684 Github \u4e3b\u9801\uff1a<\/p>\n \u300cGitHub – apple\/turicreate<\/a>\u300d<\/p>\n \u86c7\u561b\uff1f\uff01\u8766\u6bc0\uff1f\u9019\u662f\u860b\u679c\u5728 GitHub \u4e0a\u7684\u5c08\u6848\uff1f\u4ec0\u9ebc\u6642\u5019\u860b\u679c\u7d66\u51fa\u9019\u73a9\u610f\u4f46\u6211\u4e0d\u77e5\u9053\uff0c\u6211\u610f\u8b58\u5230\u81ea\u5df1\u80af\u5b9a\u53c8\u5ffd\u7565\u4e86\u4ec0\u9ebc\uff0c\u518d\u5ea6\u56de\u5230\u860b\u679c\u7684\u958b\u767c\u8005\u7db2\u9801\uff0c\u91cd\u65b0\u700f\u89bd\u4e86\u4e00\u904d\uff0c\u624d\u767c\u73fe\u860b\u679c\u65e9\u5728\u300cMachine Learning – Working with Core ML Models<\/a>\u300d\u9019\u500b\u5c08\u984c\u9801\u9762\u4e2d\uff0c\u5c31\u63d0\u904e\u9019\u73a9\u610f\u5152\u3002\u9019\u7db2\u9801\u6211\u4e0d\u77e5\u5237\u4e86\u591a\u5c11\u56de\u90fd\u6c92\u6ce8\u610f\u5230\uff0c\u6b63\u6240\u8b02\u300c\u76f4\u89ba\u5f80\u5f80\u662f\u76f2\u9ede\u300d\uff0c\u4ee5\u70ba\u53c8\u662f\u96e3\u641e\u7684\u7b2c\u4e09\u65b9\u5b78\u7fd2\u5957\u4ef6\uff0c\u6240\u4ee5\u5c31\u81ea\u52d5\u5ffd\u7565\u3002<\/p>\n \u597d\u5427\uff01\u5c31\u627f\u8a8d\u81ea\u5df1\u6c92\u628a\u66f8\u8b80\u597d\uff0c\u7e5e\u4e86\u592a\u591a\u51a4\u6789\u8def\uff0c\u767c\u73fe\u932f\u8aa4\u4e5f\u662f\u5b78\u7fd2\u7684\u904e\u7a0b\u3002<\/p>\n \u65e2\u7136\u300cTuri Create\u300d\u662f\u860b\u679c\u5728 GitHub \u4e0a\u767c\u8868\u7684\u5c08\u6848\uff0c\u61c9\u8a72\u8981\u597d\u597d\u641e\u61c2\u300cTuri Create\u300d\u662f\u600e\u9ebc\u4e00\u56de\u4e8b\uff0c\u5e0c\u671b\u9019\u73a9\u610f\u5152\u5982\u540c\u860b\u679c\u81ea\u5df1\u5728\u5c08\u6848\u9801\u4e0a\u5ba3\u7a31\u300c You don\u2019t have to be a machine learning expert to…\u300d\uff08\u60a8\u4e0d\u9700\u5177\u5099\u6a5f\u68b0\u5b78\u7fd2\u5c08\u5bb6\u77e5\u8b58\u5c31\u80fd ooxx …\uff09\u3002<\/p>\n \u7d42\u65bc\uff0c\u900f\u904e Turi Create \u7684 \u7528\u6236\u6307\u5357\u7db2\u9801\uff08User Guide\uff09<\/a>\uff0c\u6211\u7e3d\u7b97\u5b8c\u6210\u7b2c\u4e00\u500b\uff0c\u5f9e\u7121\u5230\u6709\u7684\u8cc7\u6599\u6536\u96c6\u3001\u6574\u7406\uff0c\u6700\u7d42\u81ea\u5df1\u4e5f\u89ba\u5f97\u6eff\u610f\u7684\u8a13\u7df4\u6a21\u578b\uff0c\u8b93 iPhone \u80fd\u300c\u771f\u6b63\u8a8d\u8b58\u300d Gogoro \u63db\u96fb\u7ad9\u60f9\uff01<\/p>\n \u9019\u7bc7\u6587\u7ae0\u662f\u7c21\u55ae\u7d00\u9304\u6211\u7684\u300cTuri Create – Object Detection\u300d\u5b78\u7fd2\u904e\u7a0b\u3002\u5148\u8aaa\u7d50\u8ad6\uff1a\u8207\u524d\u4e00\u7bc7\u6587\u7ae0\u4e2d\u4f7f\u7528 Xcode \u5167\u5efa\u7684\u300cCreate ML\u300d\u7684\u7121\u8166\u8a13\u7df4\u65b9\u5f0f\u76f8\u6bd4\uff0c Turi Create \u9084\u662f\u6709\u4e00\u5b9a\u7a0b\u5ea6\u7684\u96e3\u5ea6\uff0c\u4e0d\u80fd\u50cf Create ML \u6ed1\u9f20\u62d6\u62c9\u7c21\u55ae\u64cd\u4f5c\u7684\u65b9\u5f0f\u3002\u96e3\u5ea6\u4e3b\u8981\u6709\u5169\u500b\uff1a<\/p>\n \u9996\u5148\uff0c\u7576\u7136\u662f\u5148\u628a Turi Create \u88dd\u8d77\u4f86\uff0c\u4e26\u9806\u8457 Turi Create \u7684 Object-Detection User Guide \u6587\u4ef6<\/a>\uff0c\u4e00\u6b65\u4e00\u6b65\u7684\u505a\u4e0b\u53bb…<\/p>\n \u6e96\u5099\u5de5\u4f5c\uff1a<\/strong><\/span><\/p>\n \u9019\u5169\u500b\u689d\u4ef6\u548c\u524d\u4e00\u7bc7\u6587\u7ae0<\/a>\u5e7e\u4e4e\u4e00\u6a23\uff0c\u5982\u4f55\u5b89\u88dd macOS 10.14 Mojave \u8207 Xcode\uff0c\u8acb\u53c3\u95b1\u524d\u4e00\u7bc7\u6587\u7ae0\u300c\u6559\u4f60\u7684 iPhone \u8a8d\u8b58 Gogoro \u63db\u96fb\u7ad9\uff08Part 1\uff09 – \u7528\u9ed1\u860b\u679c\u96fb\u8166\u73a9\u8f49\u6700\u592f\u7684\u6a5f\u68b0\u5b78\u7fd2<\/a>\u300d\u3002<\/p>\n \u53e6\u5916\uff0c\u5f37\u70c8\u63a8\u85a6\u5b89\u88dd\u4e00\u5f35 macOS \u539f\u751f\u652f\u63f4\u7684\u7368\u7acb\u986f\u5361\uff0c\u7d93\u5be6\u6e2c\u904b\u884c Object Detect \u6a21\u578b\u6240\u6d88\u8017\u7684\u6642\u9593\u8207\u8cc7\u6e90\uff0c\u6bd4 Image Classification \u9084\u8981\u9ad8\u51fa\u8a31\u591a\uff0c\u82e5\u53ea\u7528 CPU \u904b\u7b97\uff0c\u771f\u7684\u6703\u8dd1\u5230\u5730\u8001\u5929\u8352\u3002<\/p>\n \u5b89\u88dd Turi Create<\/strong><\/span><\/p>\n Turi Create \u662f Python \u7684\u6a21\u7d44\uff0c\u6240\u4ee5\u53ea\u9700\u628a Python \u8ddf pip \u88dd\u8d77\u4f86\u4e4b\u5f8c\uff0c\u7528 pip \u4f86\u5b89\u88dd\u5373\u53ef\u3002<\/p>\n macOS \u5b89\u88dd Python\uff1a<\/span><\/strong><\/p>\n \u96d6\u7136 macOS \u5df2\u7d93\u5167\u5efa Python\uff0c\u4f46\u5167\u5efa\u7684\u662f\u7d66 macOS \u7cfb\u7d71\u7528\u7684\uff0c\u4e5f\u4e0d\u80fd\u4efb\u610f\u66f4\u52d5\uff0c\u5426\u5247 macOS \u7cfb\u7d71\u6703\u51fa\u73fe\u4e00\u5806\u4e0d\u80fd\u57f7\u884c\u7684\u5167\u5efa\u7a0b\u5f0f\u3002\u7528\u6236\u4f7f\u7528\u7684 Python \u901a\u5e38\u6703\u53e6\u5916\u5b89\u88dd\uff0c\u5b89\u88dd\u65b9\u5f0f\u6709\u5169\u7a2e\uff1aPython \u5b98\u7db2\u8207 HomeBrew \u5b89\u88dd\uff0c\u64c7\u4e00\u5373\u53ef\uff0c\u6211\u7528\u7684\u662f\u5b98\u7db2\u5b89\u88dd\u65b9\u5f0f\u3002<\/p>\n Python Download<\/a><\/p>\n \u5b89\u88dd\u65b9\u5f0f\u5f88\u7c21\u55ae\uff0c\u4e0b\u8f09\u5f8c\u57f7\u884c\u5b83\u7684 .pkg \u4f9d\u6307\u793a\u4e0b\u4e00\u6b65\u5c31\u884c\u4e86\u3002\u5b89\u88dd 3.6.x \u6216 2.7 \u7686\u53ef\uff0c\u4e5f\u53ef\u5169\u7a2e\u7248\u672c\u90fd\u5b89\u88dd\u3002\u4e0d\u8981\u5b89\u88dd 3.7 \u7248\uff0c\u56e0\u70ba\u76ee\u524d Turi Create \u5c1a\u672a\u652f\u63f4\u3002<\/p>\n \u5b89\u88dd\u5b98\u7db2\u7248\u4e4b\u5f8c\uff0c\u9700\u518d\u624b\u52d5\u57f7\u884c Python \u5b89\u88dd\u76ee\u9304\u4e0b\u7684\u5169\u500b Command \uff0c\u514d\u5f97\u5f8c\u9762\u4e00\u5806\u554f\u984c\u3002<\/p>\n pip \u5b89\u88dd Turi Create<\/strong><\/p>\n \u7d42\u7aef\u6a5f\u8996\u7a97\uff0c\u4e00\u884c\u5c31\u641e\u5b9a\uff1a<\/p>\n $ pip install turicreate<\/p>\n \u4e0d\u904e\u76ee\u524d\u9019\u884c\u6307\u4ee4\u4e0b\u8f09\u7684\u662f 4 \u7248\uff0c\u4e26\u4e0d\u652f\u63f4 macOS 10.14 \u74b0\u5883\u4e0b\u7684 GPU \u52a0\u901f\u904b\u7b97\uff0c\u6240\u4ee5\u5efa\u8b70\u5b89\u88dd 5.0 beta \u7248\uff1a<\/p>\n $ pip install “turicreate==5.0b2”<\/p>\n \u7b2c\u4e00\u6b21\u5927\u6982\u8dd1\u500b\u5e7e\u5206\u9418\u4e4b\u5f8c\uff0cTuriCreate \u5c31\u5b89\u88dd\u5b8c\u6210\u4e86\u3002\u82e5\u662f\u4f7f\u7528 Python 3.6 \u7684\u8a71\uff0c\u8981\u7528 pip3 \u6307\u4ee4\uff1a<\/p>\n $ pip3 install “turicreate==5.0b2”<\/p>\n \u9019\u6a23\u5c31\u5b8c\u6210\u4e86\u5b89\u88dd\uff0c\u7c21\u55ae\u57f7\u884c\u4e00\u4e0b\u770b\u6709\u6c92\u6709\u5b89\u88dd\u5b8c\u6210\uff1a<\/p>\n \u6c92\u4ec0\u9ebc\u554f\u984c\u4e4b\u5f8c\uff0c\u9806\u8457\u300cObject Detection<\/a>\u300d\u7ae0\u7bc0\u7684\u6559\u5b78\u7bc4\u4f8b\u64cd\u4f5c\u4e00\u6b21\u3002<\/p>\n \u4f7f\u7528 Turi Create \u9032\u884c \u7269\u9ad4\u8b58\u5225\uff08Object Detection\uff09\u9805\u76ee<\/strong><\/span><\/p>\n (1) \u9996\u5148\uff0c\u7576\u7136\u662f\u4e0b\u8f09\u8a13\u7df4\u6a5f\u5668\u7684\u7bc4\u4f8b\u8cc7\u6599\u3002\u4f9d\u7167 Object Detect \u6559\u5b78\u6587\u4ef6\uff08\u7db2\u5740\uff09<\/a>\u4e2d\u300cData Prepare\u300d\u9019\u7bc0\u5167\u5bb9\u7684\u6307\u793a\uff0c\u4e0b\u8f09\u7bc4\u4f8b\u6a94<\/a>\uff1a<\/p>\n \u6559\u5b78\u6587\u4ef6\u4e2d\u8981\u6211\u5011\u9078\u64c7 ig02-v1.0-bikes.zip \u548c ig02-v1.0-cars \u5169\u500b\u5c31\u884c\u4e86\uff0c\u76ee\u7684\u5f88\u660e\u986f\u7684\u662f\u8981\u8a13\u7df4\u51fa\u80fd\u8b58\u5225\u300c\u8173\u8e0f\u8eca\uff08bikes\uff09\u300d\u548c\u300c\u6c7d\u8eca(cars)\u300d\u4e3b\u9ad4\u7684\u6a21\u578b\u3002\u4e0b\u8f09\u5f8c\u89e3\u58d3\u7e2e\uff0c\u5efa\u4e00\u500b\u76ee\u9304 ig02 \u5f8c\uff0c\u628a\u4e0b\u8f09\u6a94\u6848\u9023\u540c\u76ee\u9304\u653e\u9032\u53bb\uff0c\u6574\u7406\u5982\u4e0b\uff1a<\/p>\n (2) \u7522\u751f\u8a13\u7df4\u6a21\u578b<\/p>\n \u958b\u555f\u4e00\u7d42\u7aef\u6a5f\u8996\u7a97\uff0c\u5148\u79fb\u5230\u8207 ig02 \u540c\u4e00\u5c64\u7684\u76ee\u9304\u4e0b\uff0c\u518d\u57f7\u884c python\uff1a<\/p>\n \u63a5\u8457\u5c31\u4f9d\u7167\u6559\u5b78\u6587\u4ef6\u4e2d\u7684\u6307\u4ee4\u7167 Key \u5c31\u884c\u4e86\u3002\u8981\u7279\u5225\u6ce8\u610f\u6bcf\u4e00\u884c\u524d\u9762\u662f\u5426\u6709\u7a7a\u683c\uff0c\u9019\u500b\u662f\u64b0\u5beb python \u7684\u7279\u4f8b\uff0c\u7a7a\u683c\u4ee3\u8868\u7a0b\u5f0f\u53ef\u57f7\u884c\u7684\u5340\u6bb5\uff0c\u7a7a\u683c\u6c92\u5c0d\u9f4a\uff0c\u5f8c\u7e8c\u7a0b\u5f0f\u5c31\u6703\u51fa\u554f\u984c\u3002<\/p>\n \u7136\u5f8c\u53e6\u5916\u958b\u4e00\u500b\u7d42\u7aef\u6a5f\u8996\u7a97\uff0c\u4e00\u6a23\u662f\u79fb\u5230\u8207 ig02 \u540c\u4e00\u5c64\u7684\u76ee\u9304\uff0c\u57f7\u884c python\uff1a<\/p>\n \u958b\u59cb\u8a13\u7df4\u6a21\u578b\uff1a<\/p>\n \u9019\u6642\u5019\u53ef\u807d\u5230\u986f\u793a\u5361\u98a8\u6247\u8d77\u98db\u7684\u8072\u97f3\uff0cR9-280X \u7684\u6eab\u5ea6\u5927\u7d04\u5728 79~81 \u5ea6\u3002\u5982\u679c\u518d\u8dd1\u4e45\u4e00\u9ede\uff08max_iterations \u6578\u5b57\u52a0\u5927\uff09\uff0c\u9084\u53ef\u4ee5\u96b1\u7d04\u805e\u5230\u623f\u9593\u7a7a\u6c23\u4e2d\u4e00\u80a1\u6e05\u6de1\u7684\u5851\u81a0\u5473…<\/p>\n \u82e5 Loss \u7684\u503c\u96a8\u8457 Iteration \u9032\u5ea6\u800c\u8d8a\u4f86\u8d8a\u5c0f\uff0c\u8868\u793a\u8a13\u7df4\u6709\u6b63\u5e38\u9032\u884c\u3002\u82e5 Loss \u503c\u8d8a\u4f86\u8d8a\u5927\uff0c\u6216\u662f\u51fa\u73fe nan \u503c\uff0c\u5247\u8868\u793a\u8a13\u7df4\u5931\u6557\uff0c\u6709\u5f88\u591a\u7a2e\u53ef\u80fd\uff0c\u5f97\u4e0a\u7db2\u67e5\u4e00\u4e0b\u539f\u56e0\u3002\u9019\u6b21\u8a13\u7df4\u5728 Iterations = 500 \u7684\u53c3\u6578\u4e0b\uff0c\u8017\u6642\u5927\u7d04\u516d\u5206\u9418\uff08\u5982\u679c\u662f CPU \u5927\u6982\u8981 1~2 \u500b\u5c0f\u6642\uff0c\u4f60\u4e0d\u6703\u60f3\u7b49\u7684\uff09\uff0c\u6700\u5f8c\u6211\u5011\u5f97\u5230\u4e00\u500b\u300cMyCustomObjectDetector.mlmodel\u300d\u7684 Core ML \u6a21\u578b\u3002\u53ef\u4ee5\u5c07 python \u547d\u4ee4\u4e2d\u7684\u300cpredictions\u300d\u8207\u300cmetrics\u300d\u7684\u5167\u5bb9\u5370\u51fa\u4f86\u770b\u770b\uff0c\u8a55\u4f30\u9019\u7d44\u6a21\u578b\u7684\u53ef\u4fe1\u5ea6\u5982\u4f55\u3002<\/p>\n \u5982\u679c\u628a\u986f\u5361\u63db\u6210 GTX-780 \u9032\u884c\u8a13\u7df4\uff0c \u5f97\u5230\u5dee\u7570\u5f88\u5927\u7684\u7d50\u679c\uff1a<\/p>\n batch_size = 24<\/p>\n batch_size = 16<\/p>\n \u76f8\u540c\u8a13\u7df4\u8cc7\u6599\u8207\u53c3\u6578\u4e0b\uff08batch_size = 24, max_iterations = 500\uff09\uff0cGTX-780 \u4f7f\u7528\u4e86\u6bd4 R9-280X \u66f4\u591a\u7684\u6642\u9593\uff08557 \u79d2 vs 342 \u79d2\uff09\uff0c\u6a21\u578b\u4fe1\u4efb\u5ea6\u4e5f\u6bd4\u8f03\u4f4e\u3002\u5c07 batch_size \u4e0b\u4fee\u5230 16 \u6642\u8a13\u7df4\u6642\u9593\u6709\u52a0\u5feb\uff0c\u4f46\u8a13\u7df4\u51fa\u4f86\u7684\u4fe1\u4efb\u5ea6\u66f4\u4f4e\uff0c\u6a21\u578b\u53ef\u7528\u6027\u5c31\u5475\u5475\u4e86\u3002<\/p>\n \u5728 Turi Create – Object Detector \u7684\u6848\u4f8b\u4e2d\uff0cR9-280X \u986f\u5361\u6bd4 GTX-780 \u66f4\u9069\u5408\u62ff\u4f86\u8a13\u7df4\uff0c\u8207\u524d\u4e00\u7bc7 Xcode 10 \u7684\u5f71\u50cf\u8fa8\u8b58\u662f\u5b8c\u5168\u76f8\u53cd\u7684\u7d50\u679c\u3002\u81f3\u65bc\u7d14 CPU \u8a13\u7df4\u6240\u82b1\u7684\u6642\u9593\uff0c\u521d\u4f30\u7d04 R9-280X \u7684 9 ~ 10 \u500d\uff0c\u6c92\u6709\u7b49\u5b83\u8dd1\u5b8c\uff0c\u56e0\u70ba\u5be6\u5728\u592a\u4e45\u4e86\uff0c\u76f8\u540c\u7684\u8a13\u7df4\u53c3\u6578\u4e0b\u9810\u4f30\u8981\u4e00\u500b\u591a\u5c0f\u6642\u3002<\/p>\n 2019-01-26 \u88dc\u5145<\/span>\uff1aRadeon Vega 56 8G HMB2 \u7684\u8a13\u7df4\u6210\u7e3e<\/strong><\/p>\n \u53bb\u5e74\u5f9e\u8766\u76ae\u8cfc\u5165\u4e00\u5f35 Radeon Vega 56 \u7926\u5361\u5f8c\uff0c\u6eff\u5fc3\u671f\u5f85\u5b83\u7684\u8a13\u7df4\u8868\u73fe\uff0c\u6c92\u60f3\u5230 macOS Mojave 10.14 \u7cfb\u7d71\u7adf\u7136\u6709 BUG<\/a> \u5c0e\u81f4\u8a13\u7df4\u5931\u6557\uff0c\u76f4\u5230\u9019\u5169\u5929 10.14.4 Beta \u91cb\u51fa\u5f8c\u624d\u89e3\u6c7a\u9019\u500b\u554f\u984c\u3002Radeon Vega 56 \u8a13\u7df4\u7d50\u679c\u5982\u4e0b\uff1a<\/p>\n batch_size = 24<\/p>\n \u76f8\u540c\u7684\u8a13\u7df4\u96c6\u548c\u53c3\u6578\uff0cRadeon Vega 56 \u53ea\u82b1\u4e86 261 \u79d2\u89e3\u6c7a\uff0c\u6bd4 Radeon R9-280X \u5feb\u4e86\u7d04 30%\u3002\u6eab\u5ea6\u7684\u90e8\u5206\u5927\u7d04\u5728 68~78 \u5ea6\u4e4b\u9593\u5f98\u5f8a\u3002<\/p>\n \u6709\u95dc BUG \u8a73\u60c5\u8acb\u95b1\u8b80\u300c\u9ed1\u860b\u679c\u5b89\u88dd AMD Radeon RX Vega 56<\/a>\u300d\u6587\u7ae0\u3002<\/p>\n \u767c\u4f48\u81f3 iOS App<\/span><\/strong><\/p>\n \u63a5\u8457\u7528 iOS \u7a0b\u5f0f\u53bb\u5be6\u969b\u9a57\u8b49\u6a21\u578b\u7684\u4f7f\u7528\u65b9\u5f0f\u548c\u300c\u4fe1\u4efb\u5ea6\u300d\uff08confidence\uff09\u3002<\/p>\n![\"\u3010\u6a5f\u68b0\u5b78\u7fd2\u3011\u6559\u4f60\u7684](\"https:\/\/pic.pimg.tw\/benjenq\/1532224007-2201955033.jpg?v=1532224009\" "\\"\u3010\u6a5f\u68b0\u5b78\u7fd2\u3011\u6559\u4f60\u7684")
<\/p>\n\n
\n
![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532228711-1008245756.png\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532229365-659439415.png?v=1532229366\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532229817-3709245110.png\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532230129-1889867720.png\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532231363-593685888.png\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\nimport<\/span> turicreate as<\/span> tc\nimport<\/span> os\n# Change if applicable<\/span>\nig02_path = '~\/Desktop\/ig02'<\/span>\n# Load all images in random order<\/span>\nraw_sf = tc.image_analysis.load_images(ig02_path, recursive=True<\/span>,\nrandom_order=True<\/span>)\n# Split file names so that we can determine what kind of image each row is<\/span>\n# E.g. bike_005.mask.0.png -> ['bike_005', 'mask']<\/span>\ninfo = raw_sf['path'<\/span>].apply(lambda<\/span> path: os.path.basename(path).split('.'<\/span>)[:2<\/span>])\n# Rename columns to 'name' and 'type'<\/span>\ninfo = info.unpack().rename({'X.0'<\/span>: 'name'<\/span>, 'X.1'<\/span>: 'type'<\/span>})\n# Add to our main SFrame<\/span>\nraw_sf = raw_sf.add_columns(info)\n# Extract label (e.g. 'bike') from name (e.g. 'bike_003')<\/span>\nraw_sf['label'<\/span>] = raw_sf['name'<\/span>].apply(lambda<\/span> name: name.split('_'<\/span>)[0<\/span>])\n# Original path no longer needed<\/span>\ndel<\/span> raw_sf['path'<\/span>]\n# Split into images and masks<\/span>\nsf_images = raw_sf[raw_sf['type'<\/span>] == 'image'<\/span>]\nsf_masks = raw_sf[raw_sf['type'<\/span>] == 'mask'<\/span>]\ndef<\/span> mask_to_bbox_coordinates<\/span>(img)<\/span>:<\/span>\n\"\"\"\nTakes a tc.Image of a mask and returns a dictionary representing bounding\nbox coordinates: e.g. {'x': 100, 'y': 120, 'width': 80, 'height': 120}\n\"\"\"<\/span>\nimport<\/span> numpy as<\/span> np\nmask = img.pixel_data\nif<\/span> mask.max() == 0<\/span>:\nreturn<\/span> None<\/span>\n# Take max along both x and y axis, and find first and last non-zero value<\/span>\nx0, x1 = np.where(mask.max(0<\/span>))[0<\/span>][[0<\/span>, -1<\/span>]]\ny0, y1 = np.where(mask.max(1<\/span>))[0<\/span>][[0<\/span>, -1<\/span>]]\nreturn<\/span> {'x'<\/span>: (x0 + x1) \/ 2<\/span>, 'width'<\/span>: (x1 - x0),\n'y'<\/span>: (y0 + y1) \/ 2<\/span>, 'height'<\/span>: (y1 - y0)}\n# Convert masks to bounding boxes (drop masks that did not contain bounding box)<\/span>\nsf_masks['coordinates'<\/span>] = sf_masks['image'<\/span>].apply(mask_to_bbox_coordinates)\n# There can be empty masks (which returns None), so let's get rid of those<\/span>\nsf_masks = sf_masks.dropna('coordinates'<\/span>)\n# Combine label and coordinates into a bounding box dictionary<\/span>\nsf_masks = sf_masks.pack_columns(['label'<\/span>, 'coordinates'<\/span>],\nnew_column_name='bbox'<\/span>, dtype=dict)\n# Combine bounding boxes of the same 'name' into lists<\/span>\nsf_annotations = sf_masks.groupby('name'<\/span>,\n{'annotations'<\/span>: tc.aggregate.CONCAT('bbox'<\/span>)})\n# Join annotations with the images. Note, some images do not have annotations,<\/span>\n# but we still want to keep them in the dataset. This is why it is important to<\/span>\n# a LEFT join.<\/span>\nsf = sf_images.join(sf_annotations, on='name'<\/span>, how='left'<\/span>)\n# The LEFT join fills missing matches with None, so we replace these with empty<\/span>\n# lists instead using fillna.<\/span>\nsf['annotations'<\/span>] = sf['annotations'<\/span>].fillna([])\n# Remove unnecessary columns<\/span>\ndel<\/span> sf['type'<\/span>]\n# Save SFrame<\/span>\nsf.save('ig02.sframe'<\/span>)<\/code><\/pre>\n<\/p>\nimport<\/span> turicreate as<\/span> tc\n# Load the data<\/span>\ndata = tc.SFrame('ig02.sframe'<\/span>)\n# Make a train-test split<\/span>\ntrain_data, test_data = data.random_split(0.8<\/span>)\n# Create a model, \u9019\u88e1\u52a0\u4e0a max_iterations = 500 \u7e2e\u77ed\u8a13\u7df4\u6642\u9593\uff0c\u5148\u628a\u6d41\u7a0b\u8dd1\u5b8c\n# \u52a0\u4e0a _advanced_parameters={'batch_size':24} \u5247\u662f\u89e3\u6c7a GPU \u8a18\u61b6\u9ad4 < 4GB \u5c0e\u81f4 Loss \u51fa\u73fe nan \u7684\u554f\u984c\n<\/span>\nmodel = tc.object_detector.create(train_data,max_iterations = 500, _advanced_parameters={'batch_size':24})\n# Save predictions to an SArray<\/span>\npredictions = model.predict(test_data)\n# Evaluate the model and save the results into a dictionary<\/span>\nmetrics = model.evaluate(test_data)\n# Save the model for later use in Turi Create<\/span>\nmodel.save('mymodel.model'<\/span>)\n# Export for use in Core ML<\/span>\nmodel.export_coreml('MyCustomObjectDetector.mlmodel'<\/span>)<\/code><\/pre>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532248613-2803860502.png?v=1532248615\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532707221-595119497_l.png?v=1532707223\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1532708289-4103361024_l.png\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n![\"\u87a2\u5e55\u5feb\u7167](\"https:\/\/pic.pimg.tw\/benjenq\/1548498301-1790764359.png\" "\\"\u87a2\u5e55\u5feb\u7167")
<\/p>\n