{"id":108791,"date":"2021-01-20T15:42:28","date_gmt":"2021-01-20T14:42:28","guid":{"rendered":"https:\/\/blog.jetbrains.com\/?post_type=kotlin&p=108791"},"modified":"2021-01-20T15:42:28","modified_gmt":"2021-01-20T14:42:28","slug":"kotlindl-alpha","status":"publish","type":"kotlin","link":"https:\/\/blog.jetbrains.com\/zh-hans\/kotlin\/2021\/01\/kotlindl-alpha\/","title":{"rendered":"\u7528 Kotlin \u8fdb\u884c\u6df1\u5ea6\u5b66\u4e60\uff1aKotlinDL-alpha \u7b80\u4ecb"},"content":{"rendered":"

\u5927\u5bb6\u597d!
\n\u4eca\u5929\u6211\u4eec\u60f3\u548c\u5927\u5bb6\u5206\u4eab\u4e00\u4e0b KotlinDL<\/a> (v.0.1.0),\u7684\u7b2c\u4e00\u6b21\u9884\u89c8\uff0c\u8fd9\u662f\u4e00\u4e2a\u7528 Kotlin \u7f16\u5199\u7684\u9ad8\u7ea7\u6df1\u5ea6\u5b66\u4e60\u6846\u67b6\uff0c\u7075\u611f\u6765\u6e90\u4e8e Keras<\/a>\u3002
\n\u5b83\u4e3a\u5728 JVM \u73af\u5883\u4e2d\u6784\u5efa\u3001\u8bad\u7ec3\u548c\u90e8\u7f72\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u63d0\u4f9b\u4e86\u7b80\u5355\u7684 API\u3002 \u9ad8\u5c42\u6b21\u7684 API \u548c\u8bb8\u591a\u53c2\u6570\u7684\u5408\u7406\u9ed8\u8ba4\u503c\u4f7f\u5f97 KotlinDL \u5f88\u5bb9\u6613\u4e0a\u624b\u3002 \u60a8\u53ea\u9700\u8981\u51e0\u884c Kotlin \u4ee3\u7801\u5c31\u53ef\u4ee5\u521b\u5efa\u548c\u8bad\u7ec3\u7b2c\u4e00\u4e2a\u7b80\u5355\u7684\u795e\u7ecf\u7f51\u7edc\uff1a <\/p>\n

\r\nprivate val model = Sequential.of(\r\n    Input(28, 28, 1),\r\n    Flatten(),\r\n    Dense(300),\r\n    Dense(100),\r\n    Dense(10)\r\n)\r\n\r\nfun main() {\r\n    val (train, test) = Dataset.createTrainAndTestDatasets(\r\n        trainFeaturesPath = "datasets\/mnist\/train-images-idx3-ubyte.gz",\r\n        trainLabelsPath = "datasets\/mnist\/train-labels-idx1-ubyte.gz",\r\n        testFeaturesPath = "datasets\/mnist\/t10k-images-idx3-ubyte.gz",\r\n        testLabelsPath = "datasets\/mnist\/t10k-labels-idx1-ubyte.gz",\r\n        numClasses = 10,\r\n        ::extractImages,\r\n        ::extractLabels\r\n    )\r\n    val (newTrain, validation) = train.split(splitRatio = 0.95)\r\n\r\n    model.use {\r\n        it.compile(\r\n            optimizer = Adam(),\r\n            loss = Losses.SOFT_MAX_CROSS_ENTROPY_WITH_LOGITS,\r\n            metric = Metrics.ACCURACY\r\n        )\r\n\r\n        it.summary()\r\n\r\n        it.fit(\r\n            dataset = newTrain,\r\n            epochs = 10,\r\n            batchSize = 100,\r\n            verbose = false\r\n        )\r\n\r\n        val accuracy = it.evaluate(\r\n            dataset = validation,\r\n            batchSize = 100\r\n        ).metrics[Metrics.ACCURACY]\r\n\r\n        println("Accuracy: $accuracy")\r\n        it.save(File("src\/model\/my_model"))\r\n    }\r\n}\r\n<\/pre>\n
GPU \u652f\u6301<\/h2>\n
\u8bad\u7ec3\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\u5f88\u82b1\u8d39\u8d44\u6e90\uff0c\u800c\u60a8\u53ef\u80fd\u8003\u8651\u8fc7\u901a\u8fc7\u5728 GPU \u4e0a\u8fd0\u884c\u6765\u52a0\u901f\u8fd9\u4e2a\u8fc7\u7a0b\u3002 \u4f7f\u7528KotlinDL\u53ef\u4ee5\u8f7b\u677e\u5b9e\u73b0\u8fd9\u4e00\u70b9\uff01
\n\u53ea\u9700\u4e00\u4e2a\u989d\u5916\u7684\u4f9d\u8d56\uff0c\u4f60\u5c31\u53ef\u4ee5\u5728 NVIDIA GPU \u8bbe\u5907\u4e0a\u8fd0\u884c\u4e0a\u8ff0\u4ee3\u7801\u800c\u65e0\u9700\u4efb\u4f55\u4fee\u6539\u3002<\/p>\n
\u4e30\u5bcc\u7684 API<\/h2>\n
KotlinDL \u81ea\u5e26\u6240\u6709\u5fc5\u8981\u7684 API\uff0c\u7528\u4e8e\u6784\u5efa\u548c\u8bad\u7ec3\u524d\u9988\u795e\u7ecf\u7f51\u7edc\uff0c\u5305\u62ec\u5377\u79ef\u795e\u7ecf\u7f51\u7edc\u3002 \u5b83\u4e3a\u5927\u591a\u6570\u8d85\u53c2\u6570\u63d0\u4f9b\u4e86\u5408\u7406\u7684\u9ed8\u8ba4\u503c\uff0c\u5e76\u63d0\u4f9b\u4e86\u5e7f\u6cdb\u7684\u4f18\u5316\u5668\u3001weight \u521d\u59cb\u503c\u8bbe\u5b9a\u9879\u3001\u6fc0\u6d3b\u51fd\u6570\u4ee5\u53ca\u5176\u4ed6\u6240\u6709\u5fc5\u8981\u7684\u6760\u6746\uff0c\u4f9b\u60a8\u8c03\u6574\u6a21\u578b\u3002
\n\u901a\u8fc7 KotlinDL\uff0c\u60a8\u53ef\u4ee5\u4fdd\u5b58\u751f\u6210\u7684\u6a21\u578b\uff0c\u5e76\u5c06\u5176\u5bfc\u5165\u5230\u60a8\u7684 JVM \u540e\u7aef\u5e94\u7528\u7a0b\u5e8f\u4e2d\u8fdb\u884c\u63a8\u7406\u3002<\/p>\n
Keras \u6a21\u578b\u5bfc\u5165<\/h2>\n
\u5f00\u7bb1\u5373\u7528\uff0cKotlinDL \u63d0\u4f9b\u4e86 API\uff0c\u7528\u4e8e\u6784\u5efa\u3001\u8bad\u7ec3\u3001\u4fdd\u5b58\u6df1\u5ea6\u5b66\u4e60\u6a21\u578b\uff0c\u5e76\u52a0\u8f7d\u6a21\u578b\u8fd0\u884c\u63a8\u7406\u3002 \u5728\u5bfc\u5165\u6a21\u578b\u8fdb\u884c\u63a8\u7406\u65f6\uff0c\u60a8\u53ef\u4ee5\u4f7f\u7528\u7528 KotlinDL \u8bad\u7ec3\u7684\u6a21\u578b\uff0c\u4e5f\u53ef\u4ee5\u5bfc\u5165\u7528 Python \u548c Keras\uff082.* \u7248\u672c\uff09\u8bad\u7ec3\u7684\u6a21\u578b\u3002<\/p>\n
\u5bf9\u4e8e\u7528 KotlinDL \u6216 Keras \u8bad\u7ec3\u7684\u6a21\u578b\uff0cKotlinDL \u652f\u6301\u8f6c\u79fb\u5b66\u4e60\u65b9\u6cd5\uff0c\u5141\u8bb8\u60a8\u4f7f\u7528\u73b0\u6709\u7684\u9884\u8bad\u7ec3\u6a21\u578b\uff0c\u5e76\u6839\u636e\u60a8\u7684\u4efb\u52a1\u5bf9\u5176\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n
\u4e34\u65f6\u9650\u5236<\/h2>\n
\u5728\u8fd9\u4e2a\u7b2c\u4e00\u4e2a alpha \u7248\u672c\u4e2d\uff0c\u53ea\u6709\u6709\u9650\u7684\u51e0\u4e2a\u56fe\u5c42\u53ef\u7528\u3002\u5206\u522b\u662f: Input()<\/code>, Flatten()<\/code>, Dense()<\/code>, Dropout()<\/code>, Conv2D()<\/code>, MaxPool2D()<\/code>, \u548c AvgPool2D()<\/code>\u3002 \u8fd9\u4e2a\u9650\u5236\u610f\u5473\u7740\u76ee\u524d\u5e76\u4e0d\u662f\u6240\u6709\u7684 Keras \u6a21\u578b\u90fd\u88ab\u652f\u6301\u3002 \u60a8\u53ef\u4ee5\u5bfc\u5165\u5e76\u5fae\u8c03\u4e00\u4e2a\u9884\u5148\u8bad\u7ec3\u597d\u7684 VGG-16 \u6216 VGG-19 \u6a21\u578b\uff0c\u4f46\u4e0d\u80fd\u5bfc\u5165 ResNet50 \u6a21\u578b\u3002 \u6211\u4eec\u6b63\u5728\u52aa\u529b\u5728\u5373\u5c06\u53d1\u5e03\u7684\u7248\u672c\u4e2d\u4e3a\u60a8\u5e26\u6765\u66f4\u591a\u7684\u5c42\u6b21\u3002 <\/p>\n
\u53e6\u4e00\u4e2a\u6682\u65f6\u7684\u9650\u5236\u662f\u5173\u4e8e\u90e8\u7f72\u3002 \u60a8\u53ef\u4ee5\u5728\u670d\u52a1\u5668\u7aef JVM \u73af\u5883\u4e2d\u90e8\u7f72\u6a21\u578b\uff0c\u7136\u800c\uff0c\u76ee\u524d\u8fd8\u4e0d\u652f\u6301\u5728 Android \u8bbe\u5907\u4e0a\u8fdb\u884c\u63a8\u7406\uff0c\u4f46\u4f1a\u5728\u4ee5\u540e\u7684\u7248\u672c\u4e2d\u5b9e\u73b0\u3002<\/p>\n
\u540e\u53f0\u6709\u4ec0\u4e48\uff1f<\/h2>\n
KotlinDL \u6784\u5efa\u5728 TensorFlow Java API \u4e4b\u4e0a\uff0cTensorFlow Java API \u6b63\u5728\u88ab\u5f00\u6e90\u793e\u533a\u79ef\u6781\u5f00\u53d1\u3002<\/p>\n
\u8bd5\u8bd5\u5427!<\/h2>\n
\u6211\u4eec\u5df2\u7ecf\u51c6\u5907\u4e86\u4e00\u4e9b\u6559\u7a0b\u6765\u5e2e\u52a9\u60a8\u5f00\u59cb\u4f7f\u7528 KotlinDL\uff1a<\/p>\n