Python\u306e\u5f37\u529b\u306a\u6a5f\u68b0\u5b66\u7fd2\u30e9\u30a4\u30d6\u30e9\u30ea\u3067\u3042\u308btheano\u306b\u3064\u3044\u3066\u3001\u305d\u306e\u57fa\u672c\u7684\u306a\u4f7f\u3044\u65b9\u304b\u3089TensorFlow\u3068\u306e\u9055\u3044\u3001CNN\u30fbRNN\u306e\u5b9f\u88c5\u4f8b\u3001\u7814\u7a76\u3084\u5b9f\u52d9\u306e\u30d9\u30b9\u30c8\u30d7\u30e9\u30af\u30c6\u30a3\u30b9\u307e\u3067\u3001\u5305\u62ec\u7684\u306b\u89e3\u8aac\u3057\u307e\u3059\u3002\u521d\u5fc3\u8005\u304b\u3089\u30d9\u30c6\u30e9\u30f3\u307e\u3067\u3001theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u305f\u3081\u306e\u30dd\u30a4\u30f3\u30c8\u304c\u6e80\u8f09\u3067\u3059\u3002<\/p>\n\n\n\n
\u76ee\u6b21 <\/p>\n
theano\u306f\u3001\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u69cb\u7bc9\u3068\u5b66\u7fd2\u306b\u7279\u5316\u3057\u305fPython\u306e\u30e9\u30a4\u30d6\u30e9\u30ea\u3067\u3059\u3002\u30b7\u30f3\u30dc\u30ea\u30c3\u30af\u8a08\u7b97\u306b\u3088\u308a\u6570\u5f0f\u3092\u30b3\u30f3\u30d1\u30a4\u30eb\u3057\u3001CPU\u3084GPU\u3067\u52b9\u7387\u7684\u306a\u6570\u5024\u8a08\u7b97\u3092\u53ef\u80fd\u306b\u3059\u308b\u3053\u3068\u304c\u5927\u304d\u306a\u7279\u5fb4\u3068\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
theano\u3067\u306f\u3001\u30c6\u30f3\u30bd\u30eb\uff08\u591a\u6b21\u5143\u914d\u5217\uff09\u3068\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u3092\u7528\u3044\u3066\u6570\u5f0f\u3092\u5b9a\u7fa9\u3057\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306a\u30b3\u30fc\u30c9\u3067\u7c21\u5358\u306a\u6570\u5f0f\u3092\u8868\u73fe\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
import theano.tensor as T\n\nx = T.dscalar('x')\ny = T.dscalar('y')\nz = x + y<\/pre>\n\n\n\n\u3053\u3053\u3067\u306f\u3001x<\/code>\u3068y<\/code>\u3068\u3044\u3046\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u3092\u5b9a\u7fa9\u3057\u3001\u305d\u308c\u3089\u306e\u548c\u3092z<\/code>\u306b\u4ee3\u5165\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u306e\u3088\u3046\u306b\u30b7\u30f3\u30dc\u30ea\u30c3\u30af\u306b\u6570\u5f0f\u3092\u8a18\u8ff0\u3057\u305f\u3042\u3068\u3001theano.function<\/code>\u3092\u4f7f\u3063\u3066\u30b3\u30f3\u30d1\u30a4\u30eb\u3059\u308b\u3053\u3068\u3067\u3001\u5b9f\u969b\u306e\u6570\u5024\u8a08\u7b97\u3092\u9ad8\u901f\u306b\u5b9f\u884c\u3067\u304d\u308b\u3088\u3046\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\nimport theano\n\nf = theano.function([x, y], z)\nprint(f(2, 3)) # 5.0<\/pre>\n\n\n\n\u30b3\u30f3\u30d1\u30a4\u30eb\u3055\u308c\u305f\u95a2\u6570f<\/code>\u306b\u5f15\u6570\u3068\u3057\u3066\u5177\u4f53\u7684\u306a\u5024\u3092\u4e0e\u3048\u308b\u3068\u3001\u8a08\u7b97\u7d50\u679c\u304c\u8fd4\u3063\u3066\u304d\u307e\u3059\u3002<\/p>\n\n\n\ntheano\u306e\u3082\u3046\u4e00\u3064\u306e\u5927\u304d\u306a\u30e1\u30ea\u30c3\u30c8\u306f\u3001\u81ea\u52d5\u5fae\u5206\u6a5f\u80fd\u3092\u5099\u3048\u3066\u3044\u308b\u3053\u3068\u3067\u3059\u3002\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b66\u7fd2\u3067\u306f\u3001\u8aa4\u5dee\u9006\u4f1d\u64ad\u6cd5\uff08\u30d0\u30c3\u30af\u30d7\u30ed\u30d1\u30b2\u30fc\u30b7\u30e7\u30f3\uff09\u306b\u3088\u3063\u3066\u30d1\u30e9\u30e1\u30fc\u30bf\u3092\u66f4\u65b0\u3059\u308b\u305f\u3081\u306b\u3001\u640d\u5931\u95a2\u6570\u306e\u5404\u30d1\u30e9\u30e1\u30fc\u30bf\u306b\u95a2\u3059\u308b\u52fe\u914d\u3092\u6c42\u3081\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002theano\u3067\u306f\u3001\u3053\u308c\u3092\u81ea\u52d5\u3067\u8a08\u7b97\u3057\u3066\u304f\u308c\u308b\u305f\u3081\u3001\u30e6\u30fc\u30b6\u30fc\u306f\u52fe\u914d\u306e\u8a08\u7b97\u3092\u610f\u8b58\u305b\u305a\u306b\u3001\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b66\u7fd2\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u7c21\u6f54\u306b\u8a18\u8ff0\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
\u307e\u305f\u3001theano\u306fGPU\u306b\u3082\u5bfe\u5fdc\u3057\u3066\u3044\u308b\u305f\u3081\u3001\u5927\u898f\u6a21\u306a\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b66\u7fd2\u3092\u9ad8\u901f\u306b\u9032\u3081\u308b\u3053\u3068\u304c\u53ef\u80fd\u3067\u3059\u3002\u7279\u306b\u3001\u7573\u307f\u8fbc\u307f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\uff08CNN\uff09\u3084\u518d\u5e30\u578b\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\uff08RNN\uff09\u306a\u3069\u3001\u5927\u91cf\u306e\u884c\u5217\u6f14\u7b97\u3092\u5fc5\u8981\u3068\u3059\u308b\u30e2\u30c7\u30eb\u306e\u5b66\u7fd2\u306b\u306f\u5a01\u529b\u3092\u767a\u63ee\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
theano\u3092\u4f7f\u3048\u3070\u3001\u6570\u5f0f\u3092\u67d4\u8edf\u306b\u8a18\u8ff0\u3057\u306a\u304c\u3089\u3001\u305d\u306e\u5b9f\u884c\u3092\u6700\u9069\u5316\u3059\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002\u521d\u5fc3\u8005\u3067\u3082\u6bd4\u8f03\u7684\u7c21\u5358\u306b\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b9f\u88c5\u306b\u30c1\u30e3\u30ec\u30f3\u30b8\u3067\u304d\u308b\u3067\u3057\u3087\u3046\u3002\u3082\u3061\u308d\u3093\u3001\u672c\u683c\u7684\u306b\u5229\u7528\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u57fa\u672c\u6982\u5ff5\u3084\u6a5f\u68b0\u5b66\u7fd2\u306e\u4e00\u822c\u7684\u306a\u30d7\u30ed\u30bb\u30b9\u306b\u3064\u3044\u3066\u306e\u7406\u89e3\u304c\u5fc5\u8981\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u6b21\u7bc0\u3067\u306f\u3001\u5b9f\u969b\u306btheano\u3092\u4f7f\u3063\u3066\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u69cb\u7bc9\u30fb\u5b66\u7fd2\u3059\u308b\u65b9\u6cd5\u3092\u898b\u3066\u3044\u304d\u307e\u3059\u3002\u6570\u5f0f\u306e\u5b9a\u7fa9\u304b\u3089\u30e2\u30c7\u30eb\u306e\u5b9f\u88c5\u307e\u3067\u3001\u5177\u4f53\u7684\u306a\u30b9\u30c6\u30c3\u30d7\u3092\u8ffd\u3063\u3066\u89e3\u8aac\u3057\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
theano\u306e\u57fa\u672c\u7684\u306a\u4f7f\u3044\u65b9 \u2013 \u6570\u5f0f\u304b\u3089\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b9f\u88c5\u307e\u3067<\/h2>\n\n\n\n
theano\u3092\u4f7f\u3063\u3066\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u5b9f\u88c5\u3059\u308b\u306b\u306f\u3001\u5927\u304d\u304f\u5206\u3051\u3066\u4ee5\u4e0b\u306e3\u3064\u306e\u30b9\u30c6\u30c3\u30d7\u304c\u5fc5\u8981\u3067\u3059\u3002<\/p>\n\n\n\n
\n- \u30b7\u30f3\u30dc\u30eb\u5909\u6570\u306e\u5b9a\u7fa9<\/li>\n\n\n\n
- \u6570\u5f0f\u306e\u5b9a\u7fa9\uff08\u30e2\u30c7\u30eb\u306e\u69cb\u7bc9\uff09<\/li>\n\n\n\n
- \u95a2\u6570\u306e\u30b3\u30f3\u30d1\u30a4\u30eb\u3068\u5b9f\u884c<\/li>\n<\/ol>\n\n\n\n
\u307e\u305a\u3001\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u3092\u5b9a\u7fa9\u3057\u307e\u3059\u3002\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u306f\u3001\u5165\u529b\u30c7\u30fc\u30bf\u3001\u30d1\u30e9\u30e1\u30fc\u30bf\u3001\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306a\u3069\u3092\u8868\u73fe\u3059\u308b\u305f\u3081\u306b\u4f7f\u7528\u3057\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001\u5165\u529b\u30c7\u30fc\u30bf\u3092\u8868\u3059\u5909\u6570x<\/code>\u3068\u3001\u30e9\u30d9\u30eb\u3092\u8868\u3059\u5909\u6570y<\/code>\u306f\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u5b9a\u7fa9\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\nimport theano\nimport theano.tensor as T\n\nx = T.matrix('x') # \u5165\u529b\u30c7\u30fc\u30bf\ny = T.ivector('y') # \u30e9\u30d9\u30eb<\/pre>\n\n\n\n\u6b21\u306b\u3001\u6570\u5f0f\u3092\u5b9a\u7fa9\u3057\u3066\u30e2\u30c7\u30eb\u3092\u69cb\u7bc9\u3057\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u3092\u7528\u3044\u3066\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5404\u5c64\u306e\u8a08\u7b97\u3092\u8a18\u8ff0\u3057\u3066\u3044\u304d\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001\u5358\u7d14\u306a\u591a\u5c64\u30d1\u30fc\u30bb\u30d7\u30c8\u30ed\u30f3\u306e\u5834\u5408\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u5404\u5c64\u306e\u8a08\u7b97\u3092\u5b9a\u7fa9\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
W1 = theano.shared(np.random.randn(100, 50), name='W1') # \u7b2c1\u5c64\u306e\u91cd\u307f\nb1 = theano.shared(np.zeros(50), name='b1') # \u7b2c1\u5c64\u306e\u30d0\u30a4\u30a2\u30b9\nW2 = theano.shared(np.random.randn(50, 10), name='W2') # \u7b2c2\u5c64\u306e\u91cd\u307f\nb2 = theano.shared(np.zeros(10), name='b2') # \u7b2c2\u5c64\u306e\u30d0\u30a4\u30a2\u30b9\n\nhidden = T.nnet.sigmoid(T.dot(x, W1) + b1) # \u96a0\u308c\u5c64\u306e\u8a08\u7b97\noutput = T.nnet.softmax(T.dot(hidden, W2) + b2) # \u51fa\u529b\u5c64\u306e\u8a08\u7b97<\/pre>\n\n\n\n\u3053\u3053\u3067\u306f\u3001W1<\/code>\u3001b1<\/code>\u3001W2<\/code>\u3001b2<\/code>\u304c\u305d\u308c\u305e\u308c\u7b2c1\u5c64\u3068\u7b2c2\u5c64\u306e\u91cd\u307f\u3068\u30d0\u30a4\u30a2\u30b9\u3092\u8868\u3059\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u3067\u3059\u3002T.dot<\/code>\u306f\u884c\u5217\u306e\u7a4d\u3092\u8a08\u7b97\u3057\u3001T.nnet.sigmoid<\/code>\u306f\u30b7\u30b0\u30e2\u30a4\u30c9\u95a2\u6570\u3092\u9069\u7528\u3057\u307e\u3059\u3002\u96a0\u308c\u5c64\u306e\u51fa\u529b\u306b\u5bfe\u3057\u3066\u3001\u518d\u5ea6\u91cd\u307f\u3068\u30d0\u30a4\u30a2\u30b9\u3092\u9069\u7528\u3057\u3001\u30bd\u30d5\u30c8\u30de\u30c3\u30af\u30b9\u95a2\u6570\u3067\u78ba\u7387\u3092\u8a08\u7b97\u3059\u308b\u3053\u3068\u3067\u3001\u51fa\u529b\u5c64\u306e\u8a08\u7b97\u3092\u884c\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n\u30e2\u30c7\u30eb\u306e\u69cb\u7bc9\u5f8c\u306f\u3001\u640d\u5931\u95a2\u6570\u3092\u5b9a\u7fa9\u3057\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u591a\u30af\u30e9\u30b9\u5206\u985e\u306e\u305f\u3081\u306e\u30af\u30ed\u30b9\u30a8\u30f3\u30c8\u30ed\u30d4\u30fc\u95a2\u6570\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
loss = T.mean(T.nnet.categorical_crossentropy(output, y))<\/pre>\n\n\n\n\u6700\u5f8c\u306b\u3001theano.function\u3092\u4f7f\u3063\u3066\u95a2\u6570\u3092\u30b3\u30f3\u30d1\u30a4\u30eb\u3057\u3001\u5b9f\u884c\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
train = theano.function(inputs=[x, y], outputs=loss, updates=[(W1, W1 - 0.01 * T.grad(loss, W1)), (b1, b1 - 0.01 * T.grad(loss, b1)), (W2, W2 - 0.01 * T.grad(loss, W2)), (b2, b2 - 0.01 * T.grad(loss, b2))])\npredict = theano.function(inputs=[x], outputs=T.argmax(output, axis=1))<\/pre>\n\n\n\ntrain<\/code>\u95a2\u6570\u3067\u306f\u3001\u5165\u529b\u30c7\u30fc\u30bfx<\/code>\u3068\u30e9\u30d9\u30eby<\/code>\u3092\u53d7\u3051\u53d6\u308a\u3001\u640d\u5931\u95a2\u6570\u306e\u5024\u3092\u8a08\u7b97\u3057\u307e\u3059\u3002\u540c\u6642\u306b\u3001updates<\/code>\u5f15\u6570\u3092\u4f7f\u3063\u3066\u52fe\u914d\u964d\u4e0b\u6cd5\u306b\u3088\u308b\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u66f4\u65b0\u3092\u884c\u3044\u307e\u3059\u3002T.grad<\/code>\u306f\u3001\u640d\u5931\u95a2\u6570\u306e\u5404\u30d1\u30e9\u30e1\u30fc\u30bf\u306b\u95a2\u3059\u308b\u52fe\u914d\u3092\u81ea\u52d5\u3067\u8a08\u7b97\u3057\u3066\u304f\u308c\u307e\u3059\u3002<\/p>\n\n\n\npredict<\/code>\u95a2\u6570\u3067\u306f\u3001\u5165\u529b\u30c7\u30fc\u30bfx<\/code>\u306b\u5bfe\u3059\u308b\u4e88\u6e2c\u7d50\u679c\u3092\u8fd4\u3057\u307e\u3059\u3002T.argmax<\/code>\u306f\u3001\u51fa\u529b\u5c64\u306e\u78ba\u7387\u5206\u5e03\u304b\u3089\u6700\u3082\u78ba\u7387\u306e\u9ad8\u3044\u30af\u30e9\u30b9\u306e\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u3092\u53d6\u5f97\u3057\u307e\u3059\u3002<\/p>\n\n\n\n\u5b9f\u969b\u306b\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u5b66\u7fd2\u3055\u305b\u308b\u306b\u306f\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u30c7\u30fc\u30bf\u3092\u6e96\u5099\u3057\u3001train<\/code>\u95a2\u6570\u3068predict<\/code>\u95a2\u6570\u3092\u547c\u3073\u51fa\u3057\u307e\u3059\u3002<\/p>\n\n\n\n# \u30c7\u30fc\u30bf\u306e\u6e96\u5099\ntrain_x, train_y = ..., ... # \u5b66\u7fd2\u7528\u306e\u5165\u529b\u30c7\u30fc\u30bf\u3068\u30e9\u30d9\u30eb\ntest_x, test_y = ..., ... # \u30c6\u30b9\u30c8\u7528\u306e\u5165\u529b\u30c7\u30fc\u30bf\u3068\u30e9\u30d9\u30eb\n\n# \u5b66\u7fd2\u306e\u5b9f\u884c\nfor i in range(1000):\n loss_value = train(train_x, train_y)\n if i % 100 == 0:\n print(f\"Epoch {i}: loss = {loss_value}\")\n\n# \u4e88\u6e2c\u306e\u5b9f\u884c\npredictions = predict(test_x)\naccuracy = np.mean(predictions == test_y)\nprint(f\"Accuracy: {accuracy}\")<\/pre>\n\n\n\n\u4ee5\u4e0a\u304c\u3001theano\u3092\u4f7f\u3063\u305f\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b9f\u88c5\u624b\u9806\u306e\u57fa\u672c\u7684\u306a\u6d41\u308c\u3067\u3059\u3002\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u3092\u7528\u3044\u3066\u6570\u5f0f\u3092\u5b9a\u7fa9\u3057\u3001\u95a2\u6570\u3092\u30b3\u30f3\u30d1\u30a4\u30eb\u3057\u3066\u5b9f\u884c\u3059\u308b\u3068\u3044\u3046\u4e00\u9023\u306e\u30d7\u30ed\u30bb\u30b9\u3092\u7406\u89e3\u3059\u308b\u3053\u3068\u304c\u91cd\u8981\u3067\u3059\u3002<\/p>\n\n\n\n
theano\u306e\u7279\u5fb4\u3067\u3042\u308b\u81ea\u52d5\u5fae\u5206\u306b\u3088\u308a\u3001\u30d0\u30c3\u30af\u30d7\u30ed\u30d1\u30b2\u30fc\u30b7\u30e7\u30f3\u306e\u305f\u3081\u306e\u52fe\u914d\u8a08\u7b97\u3092\u7c21\u6f54\u306b\u8a18\u8ff0\u3067\u304d\u308b\u306e\u306f\u5927\u304d\u306a\u30e1\u30ea\u30c3\u30c8\u3067\u3059\u3002\u307e\u305f\u3001GPU\u3092\u5229\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u5927\u898f\u6a21\u306a\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b66\u7fd2\u3092\u9ad8\u901f\u306b\u884c\u3048\u307e\u3059\u3002<\/p>\n\n\n\n
\u6b21\u7bc0\u3067\u306f\u3001theano\u3092\u4f7f\u3063\u305f\u5177\u4f53\u7684\u306a\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u5b9f\u88c5\u4f8b\u3092\u898b\u3066\u3044\u304d\u307e\u3057\u3087\u3046\u3002\u591a\u5c64\u30d1\u30fc\u30bb\u30d7\u30c8\u30ed\u30f3\u306b\u3088\u308b\u5206\u985e\u554f\u984c\u3092\u984c\u6750\u306b\u3001\u30b3\u30fc\u30c9\u3092\u8a73\u3057\u304f\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
theano\u3068TensorFlow\u306e\u9055\u3044\u3092\u6bd4\u8f03\uff01\u4f7f\u3044\u5206\u3051\u65b9\u3082\u89e3\u8aac<\/h2>\n\n\n\n
theano\u3068TensorFlow\u306f\u3001\u3068\u3082\u306b\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u69cb\u7bc9\u3068\u5b66\u7fd2\u306b\u7528\u3044\u3089\u308c\u308b\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3067\u3059\u304c\u3001\u3044\u304f\u3064\u304b\u306e\u91cd\u8981\u306a\u9055\u3044\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u3053\u3067\u306f\u3001\u4e21\u8005\u306e\u7279\u5fb4\u3092\u6bd4\u8f03\u3057\u306a\u304c\u3089\u3001\u4f7f\u3044\u5206\u3051\u306e\u30dd\u30a4\u30f3\u30c8\u3092\u89e3\u8aac\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
\u307e\u305a\u3001\u30e6\u30fc\u30b6\u30fc\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u306e\u898f\u6a21\u3068\u6d3b\u767a\u3055\u306b\u5927\u304d\u306a\u5dee\u304c\u3042\u308a\u307e\u3059\u3002TensorFlow\u306fGoogle\u304c\u958b\u767a\u3057\u3066\u304a\u308a\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u304c\u975e\u5e38\u306b\u6d3b\u767a\u306b\u6d3b\u52d5\u3057\u3066\u3044\u307e\u3059\u3002\u4e00\u65b9\u3001theano\u306f\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u898f\u6a21\u304c\u76f8\u5bfe\u7684\u306b\u5c0f\u3055\u304f\u3001\u958b\u767a\u30da\u30fc\u30b9\u3082\u3084\u3084\u7de9\u3084\u304b\u3067\u3059\u3002<\/p>\n\n\n\n
\u958b\u767a\u306e\u7d4c\u7def\u3092\u898b\u308b\u3068\u3001theano\u306f2007\u5e74\u306b\u958b\u767a\u304c\u59cb\u307e\u308a\u3001\u9577\u5e74\u306b\u308f\u305f\u3063\u3066\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u306e\u7814\u7a76\u3067\u4f7f\u308f\u308c\u3066\u304d\u307e\u3057\u305f\u3002\u5bfe\u3059\u308bTensorFlow\u306f2015\u5e74\u306b\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u5316\u3055\u308c\u3001\u6025\u901f\u306b\u666e\u53ca\u304c\u9032\u3093\u3067\u3044\u307e\u3059\u3002\u73fe\u5728\u306fTensorFlow\u306e\u65b9\u304c\u5229\u7528\u8005\u304c\u591a\u304f\u3001\u6d3b\u767a\u306b\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u3055\u308c\u3066\u3044\u308b\u72b6\u6cc1\u3067\u3059\u3002<\/p>\n\n\n\n
\u6587\u6cd5\u306e\u9055\u3044\u3067\u306f\u3001theano\u306f\u6570\u5f0f\u306b\u8fd1\u3044\u5f62\u3067\u76f4\u611f\u7684\u306b\u8a18\u8ff0\u3067\u304d\u308b\u306e\u306b\u5bfe\u3057\u3001TensorFlow\u306f\u30c7\u30fc\u30bf\u30d5\u30ed\u30fc\u30b0\u30e9\u30d5\u3092\u69cb\u7bc9\u3059\u308b\u5f62\u5f0f\u3067\u3001\u3084\u3084\u5197\u9577\u306a\u8a18\u8ff0\u306b\u306a\u308b\u50be\u5411\u304c\u3042\u308a\u307e\u3059\u3002\u305f\u3060\u3057\u3001TensorFlow\u306f\u5236\u5fa1\u30d5\u30ed\u30fc\u306e\u8a18\u8ff0\u304c\u5f97\u610f\u3067\u3001\u6761\u4ef6\u5206\u5c90\u3084\u30eb\u30fc\u30d7\u3092\u67d4\u8edf\u306b\u6271\u3048\u308b\u3068\u3044\u3046\u30e1\u30ea\u30c3\u30c8\u3082\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u8a08\u7b97\u30b0\u30e9\u30d5\u306e\u6271\u3044\u65b9\u306b\u3082\u9055\u3044\u304c\u3042\u308a\u307e\u3059\u3002theano\u306f\u52d5\u7684\u306a\u8a08\u7b97\u30b0\u30e9\u30d5\u3092\u63a1\u7528\u3057\u3066\u304a\u308a\u3001\u5b9f\u884c\u6642\u306b\u8a08\u7b97\u30b0\u30e9\u30d5\u304c\u69cb\u7bc9\u3055\u308c\u307e\u3059\u3002\u4e00\u65b9\u3001TensorFlow\u306f\u9759\u7684\u306a\u8a08\u7b97\u30b0\u30e9\u30d5\u3092\u63a1\u7528\u3057\u3001\u5b9f\u884c\u524d\u306b\u30b0\u30e9\u30d5\u69cb\u9020\u3092\u5b9a\u7fa9\u3059\u308b\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002\u3053\u306e\u9055\u3044\u306b\u3088\u308a\u3001TensorFlow\u306e\u65b9\u304c\u30c7\u30d0\u30c3\u30b0\u3084\u30a8\u30e9\u30fc\u51e6\u7406\u304c\u3057\u3084\u3059\u3044\u3068\u8a00\u308f\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u6a5f\u80fd\u9762\u3067\u306f\u3001TensorFlow\u306f\u30c7\u30fc\u30bf\u306e\u53ef\u8996\u5316\u3084\u4fdd\u5b58\u30fb\u8aad\u307f\u8fbc\u307f\u6a5f\u80fd\u304c\u5145\u5b9f\u3057\u3066\u304a\u308a\u3001\u30e2\u30d0\u30a4\u30eb\u3084\u30d6\u30e9\u30a6\u30b6\u3067\u306e\u5b9f\u884c\u306b\u3082\u5bfe\u5fdc\u3057\u3066\u3044\u307e\u3059\u3002theano\u306b\u306fTensorFlow\u306b\u306f\u306a\u3044\u3001\u30b7\u30f3\u30dc\u30ea\u30c3\u30af\u5fae\u5206\u306a\u3069\u306e\u6a5f\u80fd\u3082\u3042\u308a\u307e\u3059\u304c\u3001\u5168\u4f53\u7684\u306b\u306fTensorFlow\u306e\u65b9\u304c\u5229\u4fbf\u6027\u304c\u9ad8\u3044\u3067\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u6027\u80fd\u9762\u3067\u306e\u6bd4\u8f03\u3067\u306f\u3001\u5358\u7d14\u306a\u30d9\u30f3\u30c1\u30de\u30fc\u30af\u306b\u304a\u3044\u3066theano\u306e\u65b9\u304c\u82e5\u5e72\u9ad8\u901f\u3068\u306e\u7d50\u679c\u3082\u5831\u544a\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u305f\u3060\u3057\u3001\u73fe\u5728\u306eTensorFlow\u306f\u6027\u80fd\u9762\u3067\u3082\u5341\u5206\u306b\u9ad8\u901f\u5316\u3055\u308c\u3066\u304a\u308a\u3001\u5b9f\u7528\u4e0a\u306f\u5927\u304d\u306a\u5dee\u306f\u306a\u3044\u3068\u8003\u3048\u3089\u308c\u307e\u3059\u3002<\/p>\n\n\n\n
\u3067\u306f\u3001\u5b9f\u969b\u306e\u4f7f\u3044\u5206\u3051\u3067\u306f\u3069\u306e\u3088\u3046\u306a\u70b9\u306b\u7559\u610f\u3059\u3079\u304d\u3067\u3057\u3087\u3046\u304b\u3002\u307e\u305a\u3001\u65b0\u3057\u304f\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u3092\u5b66\u3076\u5834\u5408\u306f\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u306e\u6d3b\u767a\u3055\u3084\u60c5\u5831\u306e\u8c4a\u5bcc\u3055\u304b\u3089\u3001TensorFlow\u304b\u3089\u59cb\u3081\u308b\u306e\u304c\u7121\u96e3\u3067\u3057\u3087\u3046\u3002\u4e00\u65b9\u3001\u7814\u7a76\u76ee\u7684\u3067\u67d4\u8edf\u6027\u304c\u5fc5\u8981\u306a\u5834\u5408\u306f\u3001theano\u3092\u9078\u3076\u306e\u3082\u4e00\u6848\u3067\u3059\u3002\u305f\u3060\u3057\u3001\u5b9f\u904b\u7528\u3092\u898b\u636e\u3048\u305f\u958b\u767a\u3067\u306f\u3001\u591a\u304f\u306e\u30b1\u30fc\u30b9\u3067TensorFlow\u304c\u9078\u3070\u308c\u308b\u50be\u5411\u306b\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u3068\u306f\u3044\u3048\u3001\u4e21\u8005\u306e\u5f79\u5272\u306f\u5f90\u3005\u306b\u63a5\u8fd1\u3057\u3066\u304a\u308a\u3001\u73fe\u5728\u3067\u306f\u3069\u3061\u3089\u3092\u9078\u3093\u3067\u3082\u5927\u304d\u306a\u5dee\u306f\u3042\u308a\u307e\u305b\u3093\u3002\u81ea\u5206\u306e\u30d7\u30ed\u30b8\u30a7\u30af\u30c8\u306e\u8981\u4ef6\u3084\u76ee\u7684\u306b\u5408\u308f\u305b\u3066\u3001\u9069\u5207\u306a\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3092\u9078\u3076\u3053\u3068\u304c\u91cd\u8981\u3067\u3059\u3002<\/p>\n\n\n\n
\u30b3\u30fc\u30c9\u306e\u66f8\u304d\u65b9\u306e\u5177\u4f53\u4f8b\u3092\u898b\u3066\u307f\u307e\u3057\u3087\u3046\u3002\u4ee5\u4e0b\u306f\u3001\u591a\u5c64\u30d1\u30fc\u30bb\u30d7\u30c8\u30ed\u30f3\u306b\u3088\u308b\u4e8c\u5024\u5206\u985e\u3092\u884c\u3046\u30b3\u30fc\u30c9\u306etheano\u7248\u3068TensorFlow\u7248\u3067\u3059\u3002<\/p>\n\n\n\n
theano\u7248:<\/p>\n\n\n\n
import theano\nimport theano.tensor as T\n\nx = T.matrix('x')\ny = T.vector('y')\nW1 = theano.shared(np.random.randn(10, 5), name='W1')\nb1 = theano.shared(np.zeros(5), name='b1')\nW2 = theano.shared(np.random.randn(5, 1), name='W2')\nb2 = theano.shared(np.zeros(1), name='b2')\n\nz1 = T.dot(x, W1) + b1\na1 = T.nnet.sigmoid(z1)\nz2 = T.dot(a1, W2) + b2\ny_hat = T.nnet.sigmoid(z2)\n\nloss = T.mean(T.nnet.binary_crossentropy(y_hat, y))<\/pre>\n\n\n\nTensorFlow\u7248:<\/p>\n\n\n\n
import tensorflow as tf\n\nx = tf.placeholder(tf.float32, shape=[None, 10])\ny = tf.placeholder(tf.float32, shape=[None, 1])\n\nW1 = tf.Variable(tf.random_normal([10, 5]))\nb1 = tf.Variable(tf.zeros([5]))\nW2 = tf.Variable(tf.random_normal([5, 1]))\nb2 = tf.Variable(tf.zeros([1]))\n\nz1 = tf.matmul(x, W1) + b1\na1 = tf.nn.sigmoid(z1)\nz2 = tf.matmul(a1, W2) + b2\ny_hat = tf.nn.sigmoid(z2)\n\nloss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=y_hat))<\/pre>\n\n\n\ntheano\u7248\u3067\u306f\u3001\u30b7\u30f3\u30dc\u30eb\u5909\u6570\u306b\u5bfe\u3057\u3066\u6570\u5f0f\u3063\u307d\u304f\u6f14\u7b97\u5b50\u3092\u9069\u7528\u3057\u3066\u3044\u304f\u30b9\u30bf\u30a4\u30eb\u306a\u306e\u306b\u5bfe\u3057\u3001TensorFlow\u7248\u3067\u306f\u30d7\u30ec\u30fc\u30b9\u30db\u30eb\u30c0\u30fc\u3068\u5909\u6570\u3092\u4f7f\u3063\u3066\u30b0\u30e9\u30d5\u3092\u69cb\u7bc9\u3057\u3066\u3044\u304f\u6d41\u308c\u306b\u306a\u3063\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u3053\u306e\u3088\u3046\u306b\u3001\u6587\u6cd5\u9762\u3067\u306f\u591a\u5c11\u306e\u9055\u3044\u304c\u3042\u308b\u3082\u306e\u306e\u3001\u4e21\u8005\u3068\u3082\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u306e\u69cb\u7bc9\u3068\u5b66\u7fd2\u3092\u884c\u3046\u3068\u3044\u3046\u672c\u8cea\u7684\u306a\u6a5f\u80fd\u306f\u540c\u3058\u3067\u3059\u3002\u81ea\u5206\u306b\u3068\u3063\u3066\u6271\u3044\u3084\u3059\u3044\u65b9\u3092\u9078\u3076\u3001\u3042\u308b\u3044\u306f\u4e21\u65b9\u3092\u5b66\u3093\u3067\u307f\u308b\u306e\u3082\u3088\u3044\u3067\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u6b21\u7bc0\u3067\u306f\u3001theano\u3092\u4f7f\u3063\u305f\u753b\u50cf\u8a8d\u8b58\u3084\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u306e\u5b9f\u8df5\u7684\u306a\u4f8b\u3092\u898b\u3066\u3044\u304d\u307e\u3059\u3002CNN\u3084RNN\u306e\u5b9f\u88c5\u65b9\u6cd5\u3092\u3001\u30b3\u30fc\u30c9\u3092\u4ea4\u3048\u3066\u89e3\u8aac\u3057\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
theano\u3067\u753b\u50cf\u8a8d\u8b58\u3084\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u306b\u6311\u6226\uff01CNN\u3068RNN\u306e\u5b9f\u88c5\u4f8b<\/h2>\n\n\n\n
\u3053\u3053\u307e\u3067\u3001theano\u306e\u57fa\u672c\u7684\u306a\u4f7f\u3044\u65b9\u3084TensorFlow\u3068\u306e\u6bd4\u8f03\u3092\u898b\u3066\u304d\u307e\u3057\u305f\u3002\u3053\u306e\u7bc0\u3067\u306f\u3001theano\u3092\u4f7f\u3063\u3066\u753b\u50cf\u8a8d\u8b58\u3084\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u306b\u30c1\u30e3\u30ec\u30f3\u30b8\u3057\u3066\u307f\u307e\u3057\u3087\u3046\u3002CNN\u3068RNN\u306e\u5b9f\u88c5\u4f8b\u3092\u901a\u3057\u3066\u3001\u3088\u308a\u5b9f\u8df5\u7684\u306atheano\u306e\u6d3b\u7528\u65b9\u6cd5\u3092\u5b66\u3073\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u306f\u3058\u3081\u306b\u3001\u753b\u50cf\u8a8d\u8b58\u3067\u5e83\u304f\u4f7f\u308f\u308c\u3066\u3044\u308bCNN\u306e\u5b9f\u88c5\u4f8b\u3092\u898b\u3066\u3044\u304d\u307e\u3059\u3002CNN\u306f\u3001\u7573\u307f\u8fbc\u307f\u5c64\u3001\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u3001\u5168\u7d50\u5408\u5c64\u3092\u7d44\u307f\u5408\u308f\u305b\u305f\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u3001\u753b\u50cf\u304b\u3089\u7279\u5fb4\u91cf\u3092\u62bd\u51fa\u3057\u3001\u5206\u985e\u306a\u3069\u306e\u30bf\u30b9\u30af\u3092\u884c\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u4ee5\u4e0b\u306f\u3001theano\u3092\u4f7f\u3063\u3066MNIST\u30c7\u30fc\u30bf\u30bb\u30c3\u30c8\u306e\u624b\u66f8\u304d\u6570\u5b57\u3092\u8a8d\u8b58\u3059\u308bCNN\u306e\u5b9f\u88c5\u4f8b\u3067\u3059\u3002<\/p>\n\n\n\n
import theano\nimport theano.tensor as T\nfrom theano.tensor.nnet import conv2d\nfrom theano.tensor.signal import pool\n\n# \u5165\u529b\u30c7\u30fc\u30bf\u3068\u30e9\u30d9\u30eb\u306e\u30b7\u30f3\u30dc\u30eb\u5909\u6570\ninput = T.tensor4(name='input')\ntarget = T.ivector(name='target')\n\n# \u7573\u307f\u8fbc\u307f\u5c64\u3068\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u306e\u5b9a\u7fa9\nconv1 = conv2d(input, W_conv1, border_mode='valid')\npool1 = pool.pool_2d(conv1, (2, 2), ignore_border=True)\nconv2 = conv2d(pool1, W_conv2, border_mode='valid')\npool2 = pool.pool_2d(conv2, (2, 2), ignore_border=True)\n\n# \u5168\u7d50\u5408\u5c64\u306e\u5b9a\u7fa9\nfc1 = T.dot(pool2.flatten(2), W_fc1) + b_fc1\nfc2 = T.dot(fc1, W_fc2) + b_fc2\n\n# \u640d\u5931\u95a2\u6570\u3068\u7cbe\u5ea6\u306e\u5b9a\u7fa9\nloss = T.mean(T.nnet.categorical_crossentropy(T.nnet.softmax(fc2), target))\naccuracy = T.mean(T.eq(T.argmax(fc2, axis=1), target))<\/pre>\n\n\n\n\u3053\u3053\u3067\u306f\u3001conv2d<\/code>\u95a2\u6570\u3092\u4f7f\u3063\u3066\u7573\u307f\u8fbc\u307f\u5c64\u3092\u3001pool_2d<\/code>\u95a2\u6570\u3092\u4f7f\u3063\u3066\u30d7\u30fc\u30ea\u30f3\u30b0\u5c64\u3092\u5b9a\u7fa9\u3057\u3066\u3044\u307e\u3059\u3002\u5168\u7d50\u5408\u5c64\u306f\u3001T.dot<\/code>\u3092\u4f7f\u3063\u3066\u884c\u5217\u306e\u7a4d\u3092\u8a08\u7b97\u3059\u308b\u3053\u3068\u3067\u5b9f\u73fe\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n\u640d\u5931\u95a2\u6570\u306b\u306f\u3001\u30ab\u30c6\u30b4\u30ea\u30ab\u30eb\u30af\u30ed\u30b9\u30a8\u30f3\u30c8\u30ed\u30d4\u30fc\u3092\u4f7f\u7528\u3057\u3001T.nnet.categorical_crossentropy<\/code>\u3067\u8a08\u7b97\u3057\u3066\u3044\u307e\u3059\u3002\u7cbe\u5ea6\u306f\u3001\u4e88\u6e2c\u30e9\u30d9\u30eb\u3068\u771f\u306e\u30e9\u30d9\u30eb\u304c\u4e00\u81f4\u3059\u308b\u5272\u5408\u3092T.eq<\/code>\u3068T.mean<\/code>\u3092\u4f7f\u3063\u3066\u6c42\u3081\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n\u6b21\u306b\u3001\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u3067\u3088\u304f\u4f7f\u308f\u308c\u308bRNN\u306e\u5b9f\u88c5\u4f8b\u3092\u898b\u3066\u307f\u307e\u3057\u3087\u3046\u3002RNN\u306f\u3001\u6642\u7cfb\u5217\u30c7\u30fc\u30bf\u3092\u6271\u3046\u306e\u306b\u9069\u3057\u305f\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3067\u3001\u96a0\u308c\u72b6\u614b\u3092\u6b21\u306e\u6642\u523b\u306b\u4f1d\u64ad\u3059\u308b\u3053\u3068\u3067\u3001\u6587\u8108\u3092\u8003\u616e\u3057\u305f\u51e6\u7406\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
\u4ee5\u4e0b\u306f\u3001theano\u3092\u4f7f\u3063\u3066\u6587\u5b57\u30ec\u30d9\u30eb\u306e\u8a00\u8a9e\u30e2\u30c7\u30eb\u3092\u5b66\u7fd2\u3059\u308bRNN\u306e\u5b9f\u88c5\u4f8b\u3067\u3059\u3002<\/p>\n\n\n\n
import theano\nimport theano.tensor as T\nfrom theano.tensor.nnet import sigmoid\nfrom theano.tensor.nnet import softmax\n\n# \u5165\u529b\u30c7\u30fc\u30bf\u3068\u30e9\u30d9\u30eb\u306e\u30b7\u30f3\u30dc\u30eb\u5909\u6570\ninput = T.matrix(name='input')\ntarget = T.matrix(name='target')\n\n# \u57cb\u3081\u8fbc\u307f\u5c64\u306e\u5b9a\u7fa9\nemb = T.dot(input, W_emb)\n\n# RNN\u306e\u5b9a\u7fa9\uff08\u7c21\u5358\u306e\u305f\u3081\u30011\u3064\u306e\u96a0\u308c\u5c64\u306e\u307f\uff09\ndef step(x_t, h_tm1):\n h_t = sigmoid(T.dot(x_t, W_xh) + T.dot(h_tm1, W_hh) + b_h)\n y_t = softmax(T.dot(h_t, W_hy) + b_y)\n return h_t, y_t\n\n[h, y_pred], _ = theano.scan(step, sequences=emb, outputs_info=[h0, None])\n\n# \u640d\u5931\u95a2\u6570\u306e\u5b9a\u7fa9\nloss = T.mean(T.nnet.categorical_crossentropy(y_pred, target))<\/pre>\n\n\n\n\u3053\u3053\u3067\u306f\u3001theano.scan<\/code>\u3092\u4f7f\u3063\u3066RNN\u3092\u5b9f\u88c5\u3057\u3066\u3044\u307e\u3059\u3002step<\/code>\u95a2\u6570\u304c\u3001\u5404\u6642\u523b\u306e\u51e6\u7406\u3092\u5b9a\u7fa9\u3057\u3066\u304a\u308a\u3001\u5165\u529bx_t<\/code>\u30681\u3064\u524d\u306e\u96a0\u308c\u72b6\u614bh_tm1<\/code>\u304b\u3089\u3001\u73fe\u5728\u306e\u96a0\u308c\u72b6\u614bh_t<\/code>\u3068\u51fa\u529by_t<\/code>\u3092\u8a08\u7b97\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n\u57cb\u3081\u8fbc\u307f\u5c64\u3067\u306f\u3001\u5165\u529b\u306e\u5358\u8a9e\u3092\u30d9\u30af\u30c8\u30eb\u306b\u5909\u63db\u3057\u3066\u3044\u307e\u3059\u3002\u3053\u308c\u306b\u3088\u308a\u3001\u5358\u8a9e\u306e\u610f\u5473\u3092\u9023\u7d9a\u7684\u306a\u7a7a\u9593\u3067\u8868\u73fe\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
\u640d\u5931\u95a2\u6570\u306f\u3001CNN\u306e\u5834\u5408\u3068\u540c\u69d8\u306b\u30ab\u30c6\u30b4\u30ea\u30ab\u30eb\u30af\u30ed\u30b9\u30a8\u30f3\u30c8\u30ed\u30d4\u30fc\u3092\u4f7f\u7528\u3057\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u3053\u308c\u3089\u306e\u5b9f\u88c5\u4f8b\u3092\u8e0f\u307e\u3048\u3066\u3001\u5b9f\u969b\u306e\u30bf\u30b9\u30af\u306b\u5fdc\u7528\u3057\u3066\u307f\u3066\u304f\u3060\u3055\u3044\u3002\u753b\u50cf\u8a8d\u8b58\u3067\u306f\u3001\u7269\u4f53\u691c\u51fa\u3084\u9854\u8a8d\u8b58\u3001\u533b\u7642\u753b\u50cf\u306e\u5206\u6790\u306a\u3069\u3001\u69d8\u3005\u306a\u5834\u9762\u3067\u6d3b\u7528\u3067\u304d\u307e\u3059\u3002\u81ea\u7136\u8a00\u8a9e\u51e6\u7406\u3067\u306f\u3001\u611f\u60c5\u5206\u6790\u3001\u6a5f\u68b0\u7ffb\u8a33\u3001\u6587\u7ae0\u8981\u7d04\u3001\u30c1\u30e3\u30c3\u30c8\u30dc\u30c3\u30c8\u306a\u3069\u3001\u5e45\u5e83\u3044\u30a2\u30d7\u30ea\u30b1\u30fc\u30b7\u30e7\u30f3\u304c\u8003\u3048\u3089\u308c\u307e\u3059\u3002<\/p>\n\n\n\n
theano\u3092\u4f7f\u3048\u3070\u3001CNN\u3084RNN\u3092\u6bd4\u8f03\u7684\u7c21\u6f54\u306b\u8a18\u8ff0\u3067\u304d\u3001GPU\u3092\u6d3b\u7528\u3057\u305f\u9ad8\u901f\u306a\u5b66\u7fd2\u3082\u53ef\u80fd\u3067\u3059\u3002\u305f\u3060\u3057\u3001\u5927\u898f\u6a21\u306a\u30bf\u30b9\u30af\u306b\u53d6\u308a\u7d44\u3080\u5834\u5408\u306f\u3001TensorFlow\u3084PyTorch\u306a\u3069\u3001\u3088\u308a\u9ad8\u6c34\u6e96\u306a\u30d5\u30ec\u30fc\u30e0\u30ef\u30fc\u30af\u3092\u691c\u8a0e\u3059\u308b\u3053\u3068\u3092\u304a\u3059\u3059\u3081\u3057\u307e\u3059\u3002<\/p>\n\n\n\n
\u6b21\u7bc0\u3067\u306f\u3001theano\u3092\u4f7f\u3063\u305f\u7814\u7a76\u3084\u5b9f\u52d9\u306e\u30d9\u30b9\u30c8\u30d7\u30e9\u30af\u30c6\u30a3\u30b9\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u305f\u3081\u306e\u30b3\u30c4\u3084\u30dd\u30a4\u30f3\u30c8\u3092\u3001\u7d4c\u9a13\u8005\u306e\u77e5\u898b\u3092\u4ea4\u3048\u3066\u89e3\u8aac\u3057\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
theano\u3092\u4f7f\u3063\u305f\u7814\u7a76\u3084\u5b9f\u52d9\u306e\u30d9\u30b9\u30c8\u30d7\u30e9\u30af\u30c6\u30a3\u30b9<\/h2>\n\n\n\n
\u3053\u3053\u307e\u3067\u3001theano\u306e\u57fa\u672c\u7684\u306a\u4f7f\u3044\u65b9\u304b\u3089\u3001\u5fdc\u7528\u7684\u306a\u4e8b\u4f8b\u307e\u3067\u3092\u898b\u3066\u304d\u307e\u3057\u305f\u3002\u3053\u306e\u7bc0\u3067\u306f\u3001theano\u3092\u4f7f\u3063\u305f\u7814\u7a76\u3084\u5b9f\u52d9\u306e\u30d9\u30b9\u30c8\u30d7\u30e9\u30af\u30c6\u30a3\u30b9\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002\u521d\u5fc3\u8005\u304b\u3089\u30d9\u30c6\u30e9\u30f3\u307e\u3067\u3001theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u305f\u3081\u306e\u30b3\u30c4\u3084\u30dd\u30a4\u30f3\u30c8\u3092\u62bc\u3055\u3048\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u307e\u305a\u3001\u5927\u898f\u6a21\u306a\u30c7\u30fc\u30bf\u3092\u6271\u3046\u969b\u306f\u3001\u52b9\u7387\u7684\u306a\u30c7\u30fc\u30bf\u306e\u8aad\u307f\u8fbc\u307f\u3068\u524d\u51e6\u7406\u304c\u91cd\u8981\u3067\u3059\u3002theano\u3067\u306f\u3001\u30e1\u30e2\u30ea\u306b\u8f09\u308b\u7a0b\u5ea6\u306e\u91cf\u306b\u5206\u5272\u3057\u3001\u30d0\u30c3\u30c1\u51e6\u7406\u3092\u884c\u3046\u3053\u3068\u304c\u63a8\u5968\u3055\u308c\u307e\u3059\u3002\u30c7\u30fc\u30bf\u306e\u524d\u51e6\u7406\uff08\u6b63\u898f\u5316\u3001\u7279\u5fb4\u91cf\u62bd\u51fa\u306a\u3069\uff09\u3082\u30d0\u30c3\u30c1\u5358\u4f4d\u3067\u884c\u3046\u3068\u52b9\u7387\u7684\u3067\u3059\u3002\u4ee5\u4e0b\u306f\u3001\u30d0\u30c3\u30c1\u51e6\u7406\u306e\u7c21\u5358\u306a\u4f8b\u3067\u3059\u3002<\/p>\n\n\n\n
import theano\nimport theano.tensor as T\n\ndef batch_iterator(data, batch_size):\n num_batches = len(data) \/\/ batch_size\n for i in range(num_batches):\n yield data[i*batch_size:(i+1)*batch_size]\n\n# \u30b7\u30f3\u30dc\u30eb\u5909\u6570\u306e\u5b9a\u7fa9\nx = T.matrix('x')\ny = T.ivector('y')\n\n# \u30e2\u30c7\u30eb\u306e\u5b9a\u7fa9\n...\n\n# \u5b66\u7fd2\u30eb\u30fc\u30d7\nfor epoch in range(num_epochs):\n for batch_x, batch_y in zip(batch_iterator(train_x, batch_size),\n batch_iterator(train_y, batch_size)):\n # \u30df\u30cb\u30d0\u30c3\u30c1\u5b66\u7fd2\u306e\u5b9f\u884c\n train(batch_x, batch_y)<\/pre>\n\n\n\n\u6b21\u306b\u3001\u30e1\u30e2\u30ea\u4f7f\u7528\u91cf\u306e\u6700\u9069\u5316\u3082\u91cd\u8981\u306a\u30dd\u30a4\u30f3\u30c8\u3067\u3059\u3002theano\u306f\u3001\u4e2d\u9593\u306e\u8a08\u7b97\u7d50\u679c\u3092\u30e1\u30e2\u30ea\u4e0a\u306b\u4fdd\u6301\u3059\u308b\u305f\u3081\u3001\u5927\u898f\u6a21\u306a\u30e2\u30c7\u30eb\u3067\u306f\u30e1\u30e2\u30ea\u4e0d\u8db3\u306b\u306a\u308b\u3053\u3068\u3082\u3042\u308a\u307e\u3059\u3002\u4f7f\u7528\u6e08\u307f\u306e\u5909\u6570\u306f\u660e\u793a\u7684\u306b\u524a\u9664\u3057\u3001\u30e1\u30e2\u30ea\u3092\u89e3\u653e\u3059\u308b\u3088\u3046\u306b\u3057\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u307e\u305f\u3001theano\u306fGPU\u3092\u7528\u3044\u305f\u9ad8\u901f\u306a\u8a08\u7b97\u304c\u53ef\u80fd\u3067\u3059\u3002GPU\u3092\u4f7f\u3046\u5834\u5408\u306f\u3001THEANO_FLAGS<\/code>\u74b0\u5883\u5909\u6570\u3067\u8a2d\u5b9a\u3092\u884c\u3044\u307e\u3059\u3002\u4f8b\u3048\u3070\u3001\u4ee5\u4e0b\u306e\u3088\u3046\u306b\u8a2d\u5b9a\u3059\u308b\u3053\u3068\u3067\u3001GPU\u3092\u4f7f\u7528\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\nexport THEANO_FLAGS=\"device=cuda,floatX=float32\"<\/pre>\n\n\n\n\u3055\u3089\u306b\u3001cuDNN\u306a\u3069\u306eGPU\u7528\u30e9\u30a4\u30d6\u30e9\u30ea\u3092\u6d3b\u7528\u3059\u308b\u3068\u3001\u3088\u308a\u9ad8\u901f\u5316\u304c\u671f\u5f85\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
theano\u3067\u306f\u3001\u8a08\u7b97\u30b0\u30e9\u30d5\u3092\u6700\u9069\u5316\u3059\u308b\u3053\u3068\u3067\u3001\u7121\u99c4\u306a\u8a08\u7b97\u3092\u524a\u6e1b\u3057\u3001\u9ad8\u901f\u5316\u3092\u56f3\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002theano.function<\/code>\u306emode<\/code>\u5f15\u6570\u3067\u3001\u6700\u9069\u5316\u306e\u30ec\u30d9\u30eb\u3092\u6307\u5b9a\u3067\u304d\u307e\u3059\u3002\u305f\u3060\u3057\u3001\u8907\u96d1\u306a\u30e2\u30c7\u30eb\u3067\u306f\u3001\u6700\u9069\u5316\u304c\u3046\u307e\u304f\u3044\u304b\u306a\u3044\u3053\u3068\u3082\u3042\u308b\u305f\u3081\u3001\u6ce8\u610f\u304c\u5fc5\u8981\u3067\u3059\u3002<\/p>\n\n\n\n\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u8abf\u6574\u3082\u3001\u6a5f\u68b0\u5b66\u7fd2\u30e2\u30c7\u30eb\u306e\u6027\u80fd\u3092\u5de6\u53f3\u3059\u308b\u91cd\u8981\u306a\u8981\u7d20\u3067\u3059\u3002\u5b66\u7fd2\u7387\u3001\u30d0\u30c3\u30c1\u30b5\u30a4\u30ba\u3001\u6b63\u5247\u5316\u306e\u5f37\u3055\u306a\u3069\u3001\u6700\u9069\u306a\u7d44\u307f\u5408\u308f\u305b\u3092\u30b0\u30ea\u30c3\u30c9\u30b5\u30fc\u30c1\u3084\u30e9\u30f3\u30c0\u30e0\u30b5\u30fc\u30c1\u3067\u63a2\u7d22\u3059\u308b\u3068\u826f\u3044\u3067\u3057\u3087\u3046\u3002early stopping\u3084\u5b66\u7fd2\u7387\u306e\u30b9\u30b1\u30b8\u30e5\u30fc\u30ea\u30f3\u30b0\u306a\u3069\u3082\u3001\u904e\u5b66\u7fd2\u3092\u9632\u3050\u305f\u3081\u306b\u6709\u52b9\u3067\u3059\u3002<\/p>\n\n\n\n
\u30c7\u30d0\u30c3\u30b0\u3084\u30ed\u30b0\u51fa\u529b\u3082\u3001theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u4e0a\u3067\u6b20\u304b\u305b\u307e\u305b\u3093\u3002\u8a08\u7b97\u30b0\u30e9\u30d5\u304c\u8907\u96d1\u306b\u306a\u308b\u3068\u3001\u30c7\u30d0\u30c3\u30b0\u304c\u96e3\u3057\u304f\u306a\u308b\u3053\u3068\u3082\u3042\u308a\u307e\u3059\u3002theano.printing.debugprint<\/code>\u3092\u4f7f\u3063\u3066\u3001\u8a08\u7b97\u30b0\u30e9\u30d5\u306e\u8a73\u7d30\u3092\u51fa\u529b\u3059\u308b\u3068\u3001\u554f\u984c\u306e\u7279\u5b9a\u306b\u5f79\u7acb\u3061\u307e\u3059\u3002\u307e\u305f\u3001\u5b66\u7fd2\u306e\u9032\u884c\u72b6\u6cc1\u3084\u3001\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u5909\u5316\u3092\u30ed\u30b0\u51fa\u529b\u3057\u3001\u76e3\u8996\u3059\u308b\u3053\u3068\u3082\u5927\u5207\u3067\u3059\u3002<\/p>\n\n\n\n\u6700\u5f8c\u306b\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u306e\u4ea4\u6d41\u3082\u5fd8\u308c\u3066\u306f\u3044\u3051\u307e\u305b\u3093\u3002theano\u306b\u306f\u3001GitHub\u4e0a\u3067\u6d3b\u767a\u306a\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u304c\u3042\u308a\u307e\u3059\u3002\u4e0d\u660e\u70b9\u304c\u3042\u308c\u3070\u3001Issue\u3092\u691c\u7d22\u3057\u305f\u308a\u3001\u8cea\u554f\u3092\u6295\u7a3f\u3057\u305f\u308a\u3057\u3066\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u306e\u77e5\u6075\u3092\u501f\u308a\u307e\u3057\u3087\u3046\u3002\u81ea\u5206\u304c\u958b\u767a\u3057\u305f\u30b3\u30fc\u30c9\u3084\u30c6\u30af\u30cb\u30c3\u30af\u3082\u3001\u7a4d\u6975\u7684\u306b\u30b7\u30a7\u30a2\u3059\u308b\u3053\u3068\u3067\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u306b\u8ca2\u732e\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
\u4ee5\u4e0a\u3001theano\u3092\u4f7f\u3063\u305f\u7814\u7a76\u3084\u5b9f\u52d9\u306e\u30d9\u30b9\u30c8\u30d7\u30e9\u30af\u30c6\u30a3\u30b9\u3092\u7d39\u4ecb\u3057\u307e\u3057\u305f\u3002\u52b9\u7387\u7684\u306a\u30c7\u30fc\u30bf\u51e6\u7406\u3001\u30e1\u30e2\u30ea\u6700\u9069\u5316\u3001GPU\u6d3b\u7528\u306a\u3069\u3001\u5b9f\u8df5\u7684\u306a\u30c6\u30af\u30cb\u30c3\u30af\u3092\u8eab\u306b\u3064\u3051\u308b\u3053\u3068\u3067\u3001theano\u3092\u3088\u308a\u52b9\u679c\u7684\u306b\u4f7f\u3044\u3053\u306a\u305b\u308b\u3088\u3046\u306b\u306a\u308b\u3067\u3057\u3087\u3046\u3002\u3064\u307e\u305a\u3044\u305f\u6642\u306f\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3092\u983c\u308b\u3053\u3068\u3082\u5927\u5207\u3067\u3059\u3002theano\u3092\u4f7f\u3044\u3053\u306a\u3057\u3066\u3001\u6a5f\u68b0\u5b66\u7fd2\u30a8\u30f3\u30b8\u30cb\u30a2\u3068\u3057\u3066\u306e\u30b9\u30ad\u30eb\u30a2\u30c3\u30d7\u3092\u76ee\u6307\u3057\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u305f\u3081\u306e\u5b66\u7fd2\u65b9\u6cd5\u3068\u30ad\u30e3\u30ea\u30a2\u30d1\u30b9<\/h3>\n\n\n\n
theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u306b\u306f\u3001\u6570\u5b66\u7684\u306a\u77e5\u8b58\u3068Python\u306e\u30d7\u30ed\u30b0\u30e9\u30df\u30f3\u30b0\u30b9\u30ad\u30eb\u304c\u5fc5\u8981\u4e0d\u53ef\u6b20\u3067\u3059\u3002\u307e\u305a\u306f\u3001\u7dda\u5f62\u4ee3\u6570\u3001\u5fae\u5206\u3001\u78ba\u7387\u30fb\u7d71\u8a08\u3068\u3044\u3063\u305f\u6570\u5b66\u306e\u57fa\u790e\u3092\u56fa\u3081\u307e\u3057\u3087\u3046\u3002\u305d\u3057\u3066\u3001Python\u306e\u6587\u6cd5\u3084\u30e9\u30a4\u30d6\u30e9\u30ea\u306e\u4f7f\u3044\u65b9\u306b\u7fd2\u719f\u3059\u308b\u305f\u3081\u3001\u7a4d\u6975\u7684\u306b\u30b3\u30fc\u30c7\u30a3\u30f3\u30b0\u306e\u7df4\u7fd2\u3092\u91cd\u306d\u308b\u3053\u3068\u304c\u5927\u5207\u3067\u3059\u3002<\/p>\n\n\n\n
theano\u306e\u5b66\u7fd2\u306b\u306f\u3001\u516c\u5f0f\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3084\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u304c\u5f79\u7acb\u3061\u307e\u3059\u3002\u30b5\u30f3\u30d7\u30eb\u30b3\u30fc\u30c9\u3092\u5b9f\u969b\u306b\u52d5\u304b\u3057\u306a\u304c\u3089\u3001theano\u306e\u57fa\u672c\u7684\u306a\u4f7f\u3044\u65b9\u3092\u8eab\u306b\u3064\u3051\u3066\u3044\u304d\u307e\u3057\u3087\u3046\u3002\u307e\u305f\u3001\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u306e\u7406\u8ad6\u306b\u3064\u3044\u3066\u3082\u4e26\u884c\u3057\u3066\u5b66\u3076\u3053\u3068\u3067\u3001theano\u3092\u4f7f\u3063\u305f\u30e2\u30c7\u30eb\u69cb\u7bc9\u306e\u7406\u89e3\u304c\u6df1\u307e\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u5b9f\u8df5\u7684\u306a\u30b9\u30ad\u30eb\u3092\u8eab\u306b\u3064\u3051\u308b\u306b\u306f\u3001Kaggle\u306a\u3069\u306e\u30c7\u30fc\u30bf\u5206\u6790\u30b3\u30f3\u30da\u30c6\u30a3\u30b7\u30e7\u30f3\u306b\u53c2\u52a0\u3059\u308b\u306e\u3082\u826f\u3044\u65b9\u6cd5\u3067\u3059\u3002\u5b9f\u30c7\u30fc\u30bf\u3092\u4f7f\u3063\u305f\u554f\u984c\u89e3\u6c7a\u306e\u7d4c\u9a13\u306f\u3001\u7814\u7a76\u3084\u5b9f\u52d9\u3067\u5f79\u7acb\u3064\u3053\u3068\u9593\u9055\u3044\u3042\u308a\u307e\u305b\u3093\u3002<\/p>\n\n\n\n
\u30ad\u30e3\u30ea\u30a2\u30d1\u30b9\u3068\u3057\u3066\u306f\u3001\u7814\u7a76\u8077\u3084\u6a5f\u68b0\u5b66\u7fd2\u30a8\u30f3\u30b8\u30cb\u30a2\u3092\u76ee\u6307\u3059\u306e\u304c\u4e00\u822c\u7684\u3067\u3057\u3087\u3046\u3002\u7814\u7a76\u8077\u3067\u3042\u308c\u3070\u3001theano\u3092\u4f7f\u3063\u3066\u65b0\u3057\u3044\u30e2\u30c7\u30eb\u3084\u30a2\u30eb\u30b4\u30ea\u30ba\u30e0\u3092\u958b\u767a\u3057\u3001\u5b66\u4f1a\u3067\u767a\u8868\u3059\u308b\u3053\u3068\u304c\u6c42\u3081\u3089\u308c\u307e\u3059\u3002\u6a5f\u68b0\u5b66\u7fd2\u30a8\u30f3\u30b8\u30cb\u30a2\u3067\u3042\u308c\u3070\u3001theano\u3092\u4f7f\u3063\u3066\u5b9f\u30c7\u30fc\u30bf\u304b\u3089\u4fa1\u5024\u3092\u5f15\u304d\u51fa\u3059\u3053\u3068\u304c\u91cd\u8981\u3067\u3059\u3002<\/p>\n\n\n\n
\u3069\u3061\u3089\u306e\u30ad\u30e3\u30ea\u30a2\u30d1\u30b9\u3092\u9078\u3076\u306b\u3057\u3066\u3082\u3001\u5e38\u306b\u65b0\u3057\u3044\u6280\u8853\u3084\u624b\u6cd5\u3092\u5b66\u3073\u7d9a\u3051\u308b\u59ff\u52e2\u304c\u5927\u5207\u3067\u3059\u3002theano\u306f\u65e5\u3005\u9032\u5316\u3057\u3066\u304a\u308a\u3001\u6700\u65b0\u306e\u52d5\u5411\u3092\u30d5\u30a9\u30ed\u30fc\u3057\u3001\u81ea\u5206\u306e\u77e5\u8b58\u3084\u30b9\u30ad\u30eb\u3092\u5e38\u306b\u30a2\u30c3\u30d7\u30c7\u30fc\u30c8\u3057\u3066\u3044\u304f\u5fc5\u8981\u304c\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
\u5927\u898f\u6a21\u306a\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u5b9f\u88c5\u6642\u306e\u30c6\u30af\u30cb\u30c3\u30af\u3068TIPS<\/h3>\n\n\n\n
theano\u3092\u4f7f\u3063\u3066\u5927\u898f\u6a21\u306a\u30cb\u30e5\u30fc\u30e9\u30eb\u30cd\u30c3\u30c8\u30ef\u30fc\u30af\u3092\u5b9f\u88c5\u3059\u308b\u969b\u306f\u3001\u69d8\u3005\u306a\u30c6\u30af\u30cb\u30c3\u30af\u3084TIPS\u3092\u6d3b\u7528\u3059\u308b\u3053\u3068\u3067\u3001\u52b9\u7387\u7684\u306b\u958b\u767a\u3092\u9032\u3081\u3089\u308c\u307e\u3059\u3002<\/p>\n\n\n\n
\u307e\u305a\u3001\u8a08\u7b97\u30b0\u30e9\u30d5\u306e\u30b3\u30f3\u30d1\u30a4\u30eb\u306b\u306f\u6642\u9593\u304c\u304b\u304b\u308b\u305f\u3081\u3001\u53ef\u80fd\u306a\u9650\u308a\u8a08\u7b97\u30b0\u30e9\u30d5\u3092\u518d\u5229\u7528\u3059\u308b\u3053\u3068\u304c\u91cd\u8981\u3067\u3059\u3002\u307e\u305f\u3001\u8a08\u7b97\u30b0\u30e9\u30d5\u306e\u30c7\u30d0\u30c3\u30b0\u306b\u306f\u3001theano.printing.pydotprint<\/code>\u3092\u4f7f\u3063\u3066\u53ef\u8996\u5316\u3059\u308b\u3068\u4fbf\u5229\u3067\u3059\u3002<\/p>\n\n\n\n\u30e1\u30e2\u30ea\u4f7f\u7528\u91cf\u3092\u524a\u6e1b\u3059\u308b\u305f\u3081\u306b\u306f\u3001\u30df\u30cb\u30d0\u30c3\u30c1\u5b66\u7fd2\u3092\u884c\u3046\u3053\u3068\u304c\u52b9\u679c\u7684\u3067\u3059\u3002\u30df\u30cb\u30d0\u30c3\u30c1\u306e\u30b5\u30a4\u30ba\u306f\u3001\u30e1\u30e2\u30ea\u91cf\u3068GPU\u306e\u7279\u6027\u3092\u8003\u616e\u3057\u3066\u9069\u5207\u306b\u8a2d\u5b9a\u3057\u307e\u3057\u3087\u3046\u3002\u307e\u305f\u3001theano.gpuarray.dnn<\/code>\u30e2\u30b8\u30e5\u30fc\u30eb\u3092\u4f7f\u3048\u3070\u3001cuDNN\u30e9\u30a4\u30d6\u30e9\u30ea\u306b\u3088\u308b\u9ad8\u901f\u5316\u304c\u53ef\u80fd\u3067\u3059\u3002<\/p>\n\n\n\n\u30e2\u30c7\u30eb\u306e\u4fdd\u5b58\u3068\u30ed\u30fc\u30c9\u306b\u306f\u3001pickle<\/code>\u30e2\u30b8\u30e5\u30fc\u30eb\u3092\u4f7f\u3046\u3068\u7c21\u5358\u306b\u5b9f\u88c5\u3067\u304d\u307e\u3059\u3002\u305f\u3060\u3057\u3001\u30d0\u30fc\u30b8\u30e7\u30f3\u9593\u306e\u4e92\u63db\u6027\u306b\u306f\u6ce8\u610f\u304c\u5fc5\u8981\u3067\u3059\u3002<\/p>\n\n\n\n\u30cf\u30a4\u30d1\u30fc\u30d1\u30e9\u30e1\u30fc\u30bf\u306e\u63a2\u7d22\u306b\u306f\u3001Hyperopt\u306a\u3069\u306e\u30e9\u30a4\u30d6\u30e9\u30ea\u3092\u6d3b\u7528\u3059\u308b\u3068\u4fbf\u5229\u3067\u3059\u3002\u30d9\u30a4\u30ba\u6700\u9069\u5316\u306b\u3088\u308b\u52b9\u7387\u7684\u306a\u63a2\u7d22\u304c\u53ef\u80fd\u306b\u306a\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
theano\u306b\u306f\u3001theano.sandbox<\/code>\u30e2\u30b8\u30e5\u30fc\u30eb\u306b\u5b9f\u9a13\u7684\u306a\u6a5f\u80fd\u304c\u542b\u307e\u308c\u3066\u3044\u307e\u3059\u3002\u65b0\u3057\u3044\u30c6\u30af\u30cb\u30c3\u30af\u3092\u8a66\u3059\u969b\u306f\u3001\u7a4d\u6975\u7684\u306b\u6d3b\u7528\u3057\u3066\u307f\u308b\u3068\u826f\u3044\u3067\u3057\u3087\u3046\u3002<\/p>\n\n\n\ntheano\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u304b\u3089\u5b66\u3076 \u2013 \u60c5\u5831\u53ce\u96c6\u3068\u4ea4\u6d41\u306e\u65b9\u6cd5<\/h3>\n\n\n\n
\u6700\u65b0\u306e\u30c7\u30a3\u30fc\u30d7\u30e9\u30fc\u30cb\u30f3\u30b0\u6280\u8853\u3092\u5b66\u3076\u306b\u306f\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u304b\u3089\u60c5\u5831\u3092\u53ce\u96c6\u3057\u3001\u4ea4\u6d41\u3059\u308b\u3053\u3068\u304c\u6b20\u304b\u305b\u307e\u305b\u3093\u3002theano\u306e\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3067\u306f\u3001\u69d8\u3005\u306a\u5834\u6240\u3067\u60c5\u5831\u4ea4\u63db\u304c\u884c\u308f\u308c\u3066\u3044\u307e\u3059\u3002<\/p>\n\n\n\n
\u307e\u305a\u3001theano\u306e\u516c\u5f0f\u30c9\u30ad\u30e5\u30e1\u30f3\u30c8\u3084\u30c1\u30e5\u30fc\u30c8\u30ea\u30a2\u30eb\u306f\u3001\u4fe1\u983c\u3067\u304d\u308b\u60c5\u5831\u6e90\u3068\u3057\u3066\u91cd\u8981\u3067\u3059\u3002\u307e\u305f\u3001GitHub\u306eIssue\u3084Pull Request\u3092\u78ba\u8a8d\u3059\u308c\u3070\u3001\u6700\u65b0\u306e\u958b\u767a\u72b6\u6cc1\u3084\u30d0\u30b0\u60c5\u5831\u3092\u77e5\u308b\u3053\u3068\u304c\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n
Stack Overflow\u3067\u306f\u3001theano\u306b\u95a2\u3059\u308b\u8cea\u554f\u3068\u56de\u7b54\u304c\u6570\u591a\u304f\u6295\u7a3f\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u56f0\u3063\u305f\u3053\u3068\u304c\u3042\u308c\u3070\u3001\u307e\u305a\u306f\u95a2\u9023\u3059\u308b\u8cea\u554f\u3092\u691c\u7d22\u3057\u3066\u307f\u307e\u3057\u3087\u3046\u3002\u81ea\u5206\u3067\u8cea\u554f\u3092\u6295\u7a3f\u3059\u308b\u3053\u3068\u3067\u3001\u4e16\u754c\u4e2d\u306e\u30a8\u30ad\u30b9\u30d1\u30fc\u30c8\u304b\u3089\u56de\u7b54\u304c\u5f97\u3089\u308c\u308b\u3053\u3068\u3082\u3042\u308a\u307e\u3059\u3002<\/p>\n\n\n\n
Twitter\u4e0a\u3067\u306f\u3001theano\u306b\u95a2\u3059\u308b\u6700\u65b0\u60c5\u5831\u304c\u30ea\u30a2\u30eb\u30bf\u30a4\u30e0\u3067\u5171\u6709\u3055\u308c\u3066\u3044\u307e\u3059\u3002#theano<\/code>\u3084#deeplearning<\/code>\u306a\u3069\u306e\u30cf\u30c3\u30b7\u30e5\u30bf\u30b0\u3092\u4f7f\u3063\u3066\u3001\u95a2\u9023\u3059\u308b\u30c4\u30a4\u30fc\u30c8\u3092\u691c\u7d22\u3059\u308b\u3068\u826f\u3044\u3067\u3057\u3087\u3046\u3002\u8457\u540d\u306a\u7814\u7a76\u8005\u3084\u30a8\u30f3\u30b8\u30cb\u30a2\u3092\u30d5\u30a9\u30ed\u30fc\u3059\u308b\u3053\u3068\u3067\u3001\u6700\u5148\u7aef\u306e\u52d5\u5411\u3092\u30ad\u30e3\u30c3\u30c1\u3067\u304d\u307e\u3059\u3002<\/p>\n\n\n\n\u30aa\u30d5\u30e9\u30a4\u30f3\u3067\u306f\u3001\u52c9\u5f37\u4f1a\u3084\u30ab\u30f3\u30d5\u30a1\u30ec\u30f3\u30b9\u306b\u53c2\u52a0\u3059\u308b\u306e\u304c\u304a\u3059\u3059\u3081\u3067\u3059\u3002\u767a\u8868\u3092\u805e\u304f\u3060\u3051\u3067\u306a\u304f\u3001\u53c2\u52a0\u8005\u3068\u76f4\u63a5\u8b70\u8ad6\u3059\u308b\u3053\u3068\u3067\u3001\u6df1\u3044\u7406\u89e3\u304c\u5f97\u3089\u308c\u307e\u3059\u3002\u81ea\u5206\u3067\u767a\u8868\u3059\u308b\u6a5f\u4f1a\u304c\u3042\u308c\u3070\u3001\u7a4d\u6975\u7684\u306b\u30c1\u30e3\u30ec\u30f3\u30b8\u3057\u3066\u307f\u307e\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3067\u306f\u3001\u30aa\u30fc\u30d7\u30f3\u30bd\u30fc\u30b9\u7cbe\u795e\u306b\u57fa\u3065\u3044\u305f\u77e5\u8b58\u306e\u5171\u6709\u304c\u91cd\u8981\u8996\u3055\u308c\u3066\u3044\u307e\u3059\u3002\u81ea\u5206\u304c\u5b66\u3093\u3060\u3053\u3068\u3084\u958b\u767a\u3057\u305f\u30b3\u30fc\u30c9\u306f\u3001\u3067\u304d\u308b\u9650\u308a\u30aa\u30fc\u30d7\u30f3\u306b\u3057\u3001\u30d5\u30a3\u30fc\u30c9\u30d0\u30c3\u30af\u3092\u6c42\u3081\u308b\u3053\u3068\u304c\u5927\u5207\u3067\u3059\u3002\u4ed6\u8005\u306e\u77e5\u898b\u3092\u53d6\u308a\u5165\u308c\u3064\u3064\u3001\u81ea\u3089\u306e\u77e5\u898b\u3092\u5171\u6709\u3059\u308b\u59ff\u52e2\u3092\u6301\u3064\u3053\u3068\u3067\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u5171\u306b\u6210\u9577\u3067\u304d\u308b\u3067\u3057\u3087\u3046\u3002<\/p>\n\n\n\n
\u4ee5\u4e0a\u3001theano\u3092\u4f7f\u3044\u3053\u306a\u3059\u305f\u3081\u306e\u30dd\u30a4\u30f3\u30c8\u3092\u3001\u5b66\u7fd2\u65b9\u6cd5\u304b\u3089\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u306e\u4ea4\u6d41\u307e\u3067\u5e45\u5e83\u304f\u7d39\u4ecb\u3057\u307e\u3057\u305f\u3002theano\u3092\u6d3b\u7528\u3057\u3066\u6a5f\u68b0\u5b66\u7fd2\u30a8\u30f3\u30b8\u30cb\u30a2\u3068\u3057\u3066\u6d3b\u8e8d\u3059\u308b\u306b\u306f\u3001\u6280\u8853\u7684\u306a\u30b9\u30ad\u30eb\u3060\u3051\u3067\u306a\u304f\u3001\u30b3\u30df\u30e5\u30cb\u30c6\u30a3\u3068\u306e\u3064\u306a\u304c\u308a\u3082\u91cd\u8981\u3060\u3068\u3044\u3046\u3053\u3068\u3092\u5fd8\u308c\u306a\u3044\u3067\u304f\u3060\u3055\u3044\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
Python\u306e\u5f37\u529b\u306a\u6a5f\u68b0\u5b66\u7fd2\u30e9\u30a4\u30d6\u30e9\u30ea\u3067\u3042\u308btheano\u306b\u3064\u3044\u3066\u3001\u305d\u306e\u57fa\u672c\u7684\u306a\u4f7f\u3044\u65b9\u304b\u3089TensorFlow\u3068\u306e\u9055\u3044\u3001CNN\u30fbRNN\u306e\u5b9f\u88c5\u4f8b\u3001\u7814\u7a76\u3084\u5b9f\u52d9\u306e\u30d9\u30b9\u30c8\u30d7\u30e9\u30af\u30c6\u30a3\u30b9\u307e\u3067\u3001\u5305\u62ec\u7684\u306b\u89e3\u8aac\u3057\u307e\u3059\u3002\u521d\u5fc3\u8005\u304b\u3089\u30d9\u30c6\u30e9\u30f3\u307e\u3067 … <\/p>\n","protected":false},"author":1,"featured_media":485,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[],"class_list":{"0":"post-185","1":"post","2":"type-post","3":"status-publish","4":"format-standard","5":"has-post-thumbnail","7":"category-python"},"_links":{"self":[{"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/posts\/185","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/chocottopro.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=185"}],"version-history":[{"count":3,"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/posts\/185\/revisions"}],"predecessor-version":[{"id":429,"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/posts\/185\/revisions\/429"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/chocottopro.com\/index.php?rest_route=\/wp\/v2\/media\/485"}],"wp:attachment":[{"href":"https:\/\/chocottopro.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=185"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/chocottopro.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=185"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/chocottopro.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=185"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}