摘要:摘要本文對膠囊網(wǎng)絡(luò)進(jìn)行了非技術(shù)性的簡要概括,分析了其兩個(gè)重要屬性,之后針對手寫體數(shù)據(jù)集上驗(yàn)證多層感知機(jī)卷積神經(jīng)網(wǎng)絡(luò)以及膠囊網(wǎng)絡(luò)的性能。這是一個(gè)非結(jié)構(gòu)化的數(shù)字圖像識(shí)別問題,使用深度學(xué)習(xí)算法能夠獲得最佳性能。作者信息,數(shù)據(jù)科學(xué),深度學(xué)習(xí)初學(xué)者。
摘要: 本文對膠囊網(wǎng)絡(luò)進(jìn)行了非技術(shù)性的簡要概括,分析了其兩個(gè)重要屬性,之后針對MNIST手寫體數(shù)據(jù)集上驗(yàn)證多層感知機(jī)、卷積神經(jīng)網(wǎng)絡(luò)以及膠囊網(wǎng)絡(luò)的性能。
神經(jīng)網(wǎng)絡(luò)于上世紀(jì)50年代提出,直到最近十年里才得以發(fā)展迅速,正改變著我們世界的方方面面。從圖像分類到自然語言處理,研究人員正在對不同領(lǐng)域建立深層神經(jīng)網(wǎng)絡(luò)模型并取得相關(guān)的突破性成果。但是隨著深度學(xué)習(xí)的進(jìn)一步發(fā)展,又面臨著新的瓶頸——只對成熟網(wǎng)絡(luò)模型進(jìn)行加深加寬操作。直到最近,Hinton老爺子提出了新的概念——膠囊網(wǎng)絡(luò)(Capsule Networks),它提高了傳統(tǒng)方法的有效性和可理解性。
本文將講解膠囊網(wǎng)絡(luò)受歡迎的原因以及通過實(shí)際代碼來加強(qiáng)和鞏固對該概念的理解。
為什么膠囊網(wǎng)絡(luò)受到這么多的關(guān)注?
對于每種網(wǎng)絡(luò)結(jié)構(gòu)而言,一般用MINST手寫體數(shù)據(jù)集驗(yàn)證其性能。對于識(shí)別數(shù)字手寫體問題,即給定一個(gè)簡單的灰度圖,用戶需要預(yù)測它所顯示的數(shù)字。這是一個(gè)非結(jié)構(gòu)化的數(shù)字圖像識(shí)別問題,使用深度學(xué)習(xí)算法能夠獲得最佳性能。本文將以這個(gè)數(shù)據(jù)集測試三個(gè)深度學(xué)習(xí)模型,即:多層感知機(jī)(MLP)、卷積神經(jīng)網(wǎng)絡(luò)(CNN)以及膠囊網(wǎng)絡(luò)(Capsule Networks)。
多層感知機(jī)(MLP)使用Keras建立多層感知機(jī)模型,代碼如下:
# define variables input_num_units = 784 hidden_num_units = 50 output_num_units = 10 epochs = 15 batch_size = 128 # create model model = Sequential([ Dense(units=hidden_num_units, input_dim=input_num_units, activation="relu"), Dense(units=output_num_units, input_dim=hidden_num_units, activation="softmax"), ]) # compile the model with necessary attributes model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
打印模型參數(shù)概要:
在經(jīng)過15次迭代訓(xùn)練后,結(jié)果如下:
Epoch 14/15 34300/34300 [==============================] - 1s 41us/step - loss: 0.0597 - acc: 0.9834 - val_loss: 0.1227 - val_acc: 0.9635 Epoch 15/15 34300/34300 [==============================] - 1s 41us/step - loss: 0.0553 - acc: 0.9842 - val_loss: 0.1245 - val_acc: 0.9637
可以看到,該模型實(shí)在是簡單!
卷積神經(jīng)網(wǎng)絡(luò)(CNN)卷積神經(jīng)網(wǎng)絡(luò)在深度學(xué)習(xí)領(lǐng)域應(yīng)用十分廣泛,表現(xiàn)優(yōu)異。下面構(gòu)建卷積神經(jīng)網(wǎng)絡(luò)模型,代碼如下:
# define variables input_shape = (28, 28, 1) hidden_num_units = 50 output_num_units = 10 batch_size = 128 model = Sequential([ InputLayer(input_shape=input_reshape), Convolution2D(25, 5, 5, activation="relu"), MaxPooling2D(pool_size=pool_size), Convolution2D(25, 5, 5, activation="relu"), MaxPooling2D(pool_size=pool_size), Convolution2D(25, 4, 4, activation="relu"), Flatten(), Dense(output_dim=hidden_num_units, activation="relu"), Dense(output_dim=output_num_units, input_dim=hidden_num_units, activation="softmax"), ]) model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
打印模型參數(shù)概要:
從上圖可以發(fā)現(xiàn),CNN比MLP模型更加復(fù)雜,下面看看其性能:
Epoch 14/15 34/34 [==============================] - 4s 108ms/step - loss: 0.1278 - acc: 0.9604 - val_loss: 0.0820 - val_acc: 0.9757 Epoch 15/15 34/34 [==============================] - 4s 110ms/step - loss: 0.1256 - acc: 0.9626 - val_loss: 0.0827 - val_acc: 0.9746
可以發(fā)現(xiàn),CNN訓(xùn)練耗費(fèi)的時(shí)間比較長,但其性能優(yōu)異。
膠囊網(wǎng)絡(luò)(Capsule Network)膠囊網(wǎng)絡(luò)的結(jié)構(gòu)比CNN網(wǎng)絡(luò)更加復(fù)雜,下面構(gòu)建膠囊網(wǎng)絡(luò)模型,代碼如下:
def CapsNet(input_shape, n_class, routings): x = layers.Input(shape=input_shape) # Layer 1: Just a conventional Conv2D layer conv1 = layers.Conv2D(filters=256, kernel_size=9, strides=1, padding="valid", activation="relu", name="conv1")(x) # Layer 2: Conv2D layer with `squash` activation, then reshape to [None, num_capsule, dim_capsule] primarycaps = PrimaryCap(conv1, dim_capsule=8, n_channels=32, kernel_size=9, strides=2, padding="valid") # Layer 3: Capsule layer. Routing algorithm works here. digitcaps = CapsuleLayer(num_capsule=n_class, dim_capsule=16, routings=routings, name="digitcaps")(primarycaps) # Layer 4: This is an auxiliary layer to replace each capsule with its length. Just to match the true label"s shape. # If using tensorflow, this will not be necessary. :) out_caps = Length(name="capsnet")(digitcaps) # Decoder network. y = layers.Input(shape=(n_class,)) masked_by_y = Mask()([digitcaps, y]) # The true label is used to mask the output of capsule layer. For training masked = Mask()(digitcaps) # Mask using the capsule with maximal length. For prediction # Shared Decoder model in training and prediction decoder = models.Sequential(name="decoder") decoder.add(layers.Dense(512, activation="relu", input_dim=16*n_class)) decoder.add(layers.Dense(1024, activation="relu")) decoder.add(layers.Dense(np.prod(input_shape), activation="sigmoid")) decoder.add(layers.Reshape(target_shape=input_shape, name="out_recon")) # Models for training and evaluation (prediction) train_model = models.Model([x, y], [out_caps, decoder(masked_by_y)]) eval_model = models.Model(x, [out_caps, decoder(masked)]) # manipulate model noise = layers.Input(shape=(n_class, 16)) noised_digitcaps = layers.Add()([digitcaps, noise]) masked_noised_y = Mask()([noised_digitcaps, y]) manipulate_model = models.Model([x, y, noise], decoder(masked_noised_y)) return train_model, eval_model, manipulate_model
打印模型參數(shù)概要:
該模型耗費(fèi)時(shí)間比較長,訓(xùn)練一段時(shí)間后,得到如下結(jié)果:
Epoch 14/15 34/34 [==============================] - 108s 3s/step - loss: 0.0445 - capsnet_loss: 0.0218 - decoder_loss: 0.0579 - capsnet_acc: 0.9846 - val_loss: 0.0364 - val_capsnet_loss: 0.0159 - val_decoder_loss: 0.0522 - val_capsnet_acc: 0.9887 Epoch 15/15 34/34 [==============================] - 107s 3s/step - loss: 0.0423 - capsnet_loss: 0.0201 - decoder_loss: 0.0567 - capsnet_acc: 0.9859 - val_loss: 0.0362 - val_capsnet_loss: 0.0162 - val_decoder_loss: 0.0510 - val_capsnet_acc: 0.9880
可以發(fā)現(xiàn),該網(wǎng)絡(luò)比之前傳統(tǒng)的網(wǎng)絡(luò)模型效果更好,下圖總結(jié)了三個(gè)實(shí)驗(yàn)結(jié)果:
這個(gè)實(shí)驗(yàn)也證明了膠囊網(wǎng)絡(luò)值得我們深入的研究和討論。
膠囊網(wǎng)絡(luò)背后的概念為了理解膠囊網(wǎng)絡(luò)的概念,本文將以貓的圖片為例來說明膠囊網(wǎng)絡(luò)的潛力,首先從一個(gè)問題開始——下圖中的動(dòng)物是什么?
它是一只貓,你肯定猜對了吧!但是你是如何知道它是一只貓的呢?現(xiàn)在將這張圖片進(jìn)行分解:
情況1——簡單圖像
你是如何知道它是一只貓的呢?可能的方法是將其分解為多帶帶的特征,如眼睛、鼻子、耳朵等。如下圖所示:
因此,本質(zhì)上是把高層次的特征分解為低層次的特征。比如定義為:
P(臉) = P(鼻子) & ( 2 x P(胡須) ) & P(嘴巴) & ( 2 x P(眼睛) ) & ( 2 x P(耳朵) )
其中,P(臉) 定義為圖像中貓臉的存在。通過迭代,可以定義更多的低級別特性,如形狀和邊緣,以簡化過程。
情況2——旋轉(zhuǎn)圖像
將圖像旋轉(zhuǎn)30度,如下圖所示:
如果還是按照之前定義的相同特征,那么將無法識(shí)別出它是貓。這是因?yàn)榈讓犹卣鞯姆较虬l(fā)生了改變,導(dǎo)致先前定義的特征也將發(fā)生變化。
綜上,貓識(shí)別器可能看起來像這樣:
更具體一點(diǎn),表示為:
P(臉) = ( P(鼻子) & ( 2 x P(胡須) ) & P(嘴巴) & ( 2 x P(眼睛) ) & ( 2 x P(耳朵) ) ) OR
( P(rotated_鼻子) & ( 2 x P(rotated_胡須) ) & P(rotated_嘴巴) & ( 2 x P(rotated_眼睛) ) & ( 2 x P(rotated_耳朵) ) )
情況3——翻轉(zhuǎn)圖像
為了增加復(fù)雜性,下面是一個(gè)完全翻轉(zhuǎn)的圖像:
可能想到的方法是靠蠻力搜索低級別特征所有可能的旋轉(zhuǎn),但這種方法耗時(shí)耗力。因此,研究人員提出,包含低級別特征本身的附加屬性,比如旋轉(zhuǎn)角度。這樣不僅可以檢測特征是否存在,還可以檢測其旋轉(zhuǎn)是否存在,如下圖所示:
更具體一點(diǎn),表示為:
P(臉) = [ P(鼻子), R(鼻子) ] & [ P(胡須_1), R(胡須_1) ] & [ P(胡須_2), R(胡須_2) ] & [ P(嘴巴), R(嘴巴) ] & …
其中,旋轉(zhuǎn)特征用R()表示,這一特性也被稱作旋轉(zhuǎn)等價(jià)性。
從上述情況中可以看到,擴(kuò)大想法之后能夠捕捉更多低層次的特征,如尺度、厚度等,這將有助于我們更清楚地理解一個(gè)物體的形象。這就是膠囊網(wǎng)絡(luò)在設(shè)計(jì)時(shí)設(shè)想的工作方式。
膠囊網(wǎng)絡(luò)另外一個(gè)特點(diǎn)是動(dòng)態(tài)路由,下面以貓狗分類問題講解這個(gè)特點(diǎn)。
上面兩只動(dòng)物看起來非常相似,但存在一些差異。你可以從中發(fā)現(xiàn)哪只是狗嗎?
正如之前所做的那樣,將定義圖像中的特征以找出其中的差異。
如圖所示,定義非常低級的面部特征,比如眼睛、耳朵等,并將其結(jié)合以找到一個(gè)臉。之后,將面部和身體特征結(jié)合來完成相應(yīng)的任務(wù)——判斷它是一只貓或狗。
現(xiàn)在假設(shè)有一個(gè)新的圖像,以及提取的低層特征,需要根據(jù)以上信息判斷出其類別。我們從中隨機(jī)選取一個(gè)特征,比如眼睛,可以只根據(jù)它來判斷其類別嗎?
答案是否定的,因?yàn)檠劬Σ⒉皇且粋€(gè)區(qū)分因素。下一步是分析更多的特征,比如隨機(jī)挑選的下一個(gè)特征是鼻子。
只有眼睛和鼻子特征并不能夠完成分類任務(wù),下一步獲取所有特征,并將其結(jié)合以判斷所屬類別。如下圖所示,通過組合眼睛、鼻子、耳朵和胡須這四個(gè)特征就能夠判斷其所屬類別?;谝陨线^程,將在每個(gè)特征級別迭代地執(zhí)行這一步驟,就可以將正確的信息路由到需要分類信息的特征檢測器。
在膠囊構(gòu)件中,當(dāng)更高級的膠囊同意較低級的膠囊輸入時(shí),較低級的膠囊將其輸入到更高級膠囊中,這就是動(dòng)態(tài)路由算法的精髓。
膠囊網(wǎng)絡(luò)相對于傳統(tǒng)深度學(xué)習(xí)架構(gòu)而言,在對數(shù)據(jù)方向和角度方面更魯棒,甚至可以在相對較少的數(shù)據(jù)點(diǎn)上進(jìn)行訓(xùn)練。膠囊網(wǎng)絡(luò)存在的缺點(diǎn)是需要更多的訓(xùn)練時(shí)間和資源。
膠囊網(wǎng)絡(luò)在MNIST數(shù)據(jù)集上的代碼詳解
首先從識(shí)別數(shù)字手寫體項(xiàng)目下載數(shù)據(jù)集,數(shù)字手寫體識(shí)別問題主要是將給定的28x28大小的圖片識(shí)別出其顯示的數(shù)字。在開始運(yùn)行代碼之前,確保安裝好Keras。
下面打開Jupyter Notebook軟件,輸入以下代碼。首先導(dǎo)入所需的模塊:
然后進(jìn)行隨機(jī)初始化:
# To stop potential randomness seed = 128 rng = np.random.RandomState(seed)
下一步設(shè)置目錄路徑:
root_dir = os.path.abspath(".") data_dir = os.path.join(root_dir, "data")
下面加載數(shù)據(jù)集,數(shù)據(jù)集是“.CSV”格式。
train = pd.read_csv(os.path.join(data_dir, "train.csv")) test = pd.read_csv(os.path.join(data_dir, "test.csv")) train.head()
展示數(shù)據(jù)表示的數(shù)字:
img_name = rng.choice(train.filename) filepath = os.path.join(data_dir, "train", img_name) img = imread(filepath, flatten=True) pylab.imshow(img, cmap="gray") pylab.axis("off") pylab.show()
現(xiàn)在將所有圖像保存為Numpy數(shù)組:
temp = [] for img_name in train.filename: image_path = os.path.join(data_dir, "train", img_name) img = imread(image_path, flatten=True) img = img.astype("float32") temp.append(img) train_x = np.stack(temp) train_x /= 255.0 train_x = train_x.reshape(-1, 784).astype("float32") temp = [] for img_name in test.filename: image_path = os.path.join(data_dir, "test", img_name) img = imread(image_path, flatten=True) img = img.astype("float32") temp.append(img) test_x = np.stack(temp) test_x /= 255.0 test_x = test_x.reshape(-1, 784).astype("float32") train_y = keras.utils.np_utils.to_categorical(train.label.values)
這是一個(gè)典型的機(jī)器學(xué)習(xí)問題,將數(shù)據(jù)集分成7:3。其中70%作為訓(xùn)練集,30%作為驗(yàn)證集。
split_size = int(train_x.shape[0]*0.7) train_x, val_x = train_x[:split_size], train_x[split_size:] train_y, val_y = train_y[:split_size], train_y[split_size:]
下面將分析三個(gè)不同深度學(xué)習(xí)模型對該數(shù)據(jù)的性能,分別是多層感知機(jī)、卷積神經(jīng)網(wǎng)絡(luò)以及膠囊網(wǎng)絡(luò)。
1.多層感知機(jī)
定義一個(gè)三層神經(jīng)網(wǎng)絡(luò),一個(gè)輸入層、一個(gè)隱藏層以及一個(gè)輸出層。輸入和輸出神經(jīng)元的數(shù)目是固定的,輸入為28x28圖像,輸出是代表類的10x1向量,隱層設(shè)置為50個(gè)神經(jīng)元,并使用梯度下降算法訓(xùn)練。
# define vars input_num_units = 784 hidden_num_units = 50 output_num_units = 10 epochs = 15 batch_size = 128 # import keras modules from keras.models import Sequential from keras.layers import InputLayer, Convolution2D, MaxPooling2D, Flatten, Dense # create model model = Sequential([ Dense(units=hidden_num_units, input_dim=input_num_units, activation="relu"), Dense(units=output_num_units, input_dim=hidden_num_units, activation="softmax"), ]) # compile the model with necessary attributes model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
打印模型參數(shù)概要:
trained_model = model.fit(train_x, train_y, nb_epoch=epochs, batch_size=batch_size, validation_data=(val_x, val_y))
在迭代15次之后,結(jié)果如下:
Epoch 14/15 34300/34300 [==============================] - 1s 41us/step - loss: 0.0597 - acc: 0.9834 - val_loss: 0.1227 - val_acc: 0.9635 Epoch 15/15 34300/34300 [==============================] - 1s 41us/step - loss: 0.0553 - acc: 0.9842 - val_loss: 0.1245 - val_acc: 0.9637
結(jié)果不錯(cuò),但可以繼續(xù)改進(jìn)。
2.卷積神經(jīng)網(wǎng)絡(luò)
把圖像轉(zhuǎn)換成灰度圖(2D),然后將其輸入到CNN模型中:
# reshape data train_x_temp = train_x.reshape(-1, 28, 28, 1) val_x_temp = val_x.reshape(-1, 28, 28, 1) # define vars input_shape = (784,) input_reshape = (28, 28, 1) pool_size = (2, 2) hidden_num_units = 50 output_num_units = 10 batch_size = 128
下面定義CNN模型:下面定義CNN模型:
model = Sequential([ InputLayer(input_shape=input_reshape), Convolution2D(25, 5, 5, activation="relu"), MaxPooling2D(pool_size=pool_size), Convolution2D(25, 5, 5, activation="relu"), MaxPooling2D(pool_size=pool_size), Convolution2D(25, 4, 4, activation="relu"), Flatten(), Dense(output_dim=hidden_num_units, activation="relu"), Dense(output_dim=output_num_units, input_dim=hidden_num_units, activation="softmax"), ]) model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"]) #trained_model_conv = model.fit(train_x_temp, train_y, nb_epoch=epochs, batch_size=batch_size, validation_data=(val_x_temp, val_y)) model.summary()
打印模型參數(shù)概要:
通過增加數(shù)據(jù)來調(diào)整進(jìn)程:
# Begin: Training with data augmentation ---------------------------------------------------------------------# def train_generator(x, y, batch_size, shift_fraction=0.1): train_datagen = ImageDataGenerator(width_shift_range=shift_fraction, height_shift_range=shift_fraction) # shift up to 2 pixel for MNIST generator = train_datagen.flow(x, y, batch_size=batch_size) while 1: x_batch, y_batch = generator.next() yield ([x_batch, y_batch]) # Training with data augmentation. If shift_fraction=0., also no augmentation. trained_model2 = model.fit_generator(generator=train_generator(train_x_temp, train_y, 1000, 0.1), steps_per_epoch=int(train_y.shape[0] / 1000), epochs=epochs, validation_data=[val_x_temp, val_y]) # End: Training with data augmentation -----------------------------------------------------------------------#
CNN模型的結(jié)果:
Epoch 14/15 34/34 [==============================] - 4s 108ms/step - loss: 0.1278 - acc: 0.9604 - val_loss: 0.0820 - val_acc: 0.9757 Epoch 15/15 34/34 [==============================] - 4s 110ms/step - loss: 0.1256 - acc: 0.9626 - val_loss: 0.0827 - val_acc: 0.9746
3.膠囊網(wǎng)絡(luò)
建立膠囊網(wǎng)絡(luò)模型,結(jié)構(gòu)如圖所示:
下面建立該模型,代碼如下:
def CapsNet(input_shape, n_class, routings): """ A Capsule Network on MNIST. :param input_shape: data shape, 3d, [width, height, channels] :param n_class: number of classes :param routings: number of routing iterations :return: Two Keras Models, the first one used for training, and the second one for evaluation. `eval_model` can also be used for training. """ x = layers.Input(shape=input_shape) # Layer 1: Just a conventional Conv2D layer conv1 = layers.Conv2D(filters=256, kernel_size=9, strides=1, padding="valid", activation="relu", name="conv1")(x) # Layer 2: Conv2D layer with `squash` activation, then reshape to [None, num_capsule, dim_capsule] primarycaps = PrimaryCap(conv1, dim_capsule=8, n_channels=32, kernel_size=9, strides=2, padding="valid") # Layer 3: Capsule layer. Routing algorithm works here. digitcaps = CapsuleLayer(num_capsule=n_class, dim_capsule=16, routings=routings, name="digitcaps")(primarycaps) # Layer 4: This is an auxiliary layer to replace each capsule with its length. Just to match the true label"s shape. # If using tensorflow, this will not be necessary. :) out_caps = Length(name="capsnet")(digitcaps) # Decoder network. y = layers.Input(shape=(n_class,)) masked_by_y = Mask()([digitcaps, y]) # The true label is used to mask the output of capsule layer. For training masked = Mask()(digitcaps) # Mask using the capsule with maximal length. For prediction # Shared Decoder model in training and prediction decoder = models.Sequential(name="decoder") decoder.add(layers.Dense(512, activation="relu", input_dim=16*n_class)) decoder.add(layers.Dense(1024, activation="relu")) decoder.add(layers.Dense(np.prod(input_shape), activation="sigmoid")) decoder.add(layers.Reshape(target_shape=input_shape, name="out_recon")) # Models for training and evaluation (prediction) train_model = models.Model([x, y], [out_caps, decoder(masked_by_y)]) eval_model = models.Model(x, [out_caps, decoder(masked)]) # manipulate model noise = layers.Input(shape=(n_class, 16)) noised_digitcaps = layers.Add()([digitcaps, noise]) masked_noised_y = Mask()([noised_digitcaps, y]) manipulate_model = models.Model([x, y, noise], decoder(masked_noised_y)) return train_model, eval_model, manipulate_model def margin_loss(y_true, y_pred): """ Margin loss for Eq.(4). When y_true[i, :] contains not just one `1`, this loss should work too. Not test it. :param y_true: [None, n_classes] :param y_pred: [None, num_capsule] :return: a scalar loss value. """ L = y_true * K.square(K.maximum(0., 0.9 - y_pred)) + 0.5 * (1 - y_true) * K.square(K.maximum(0., y_pred - 0.1)) return K.mean(K.sum(L, 1)) model, eval_model, manipulate_model = CapsNet(input_shape=train_x_temp.shape[1:], n_class=len(np.unique(np.argmax(train_y, 1))), routings=3) # compile the model model.compile(optimizer=optimizers.Adam(lr=0.001), loss=[margin_loss, "mse"], loss_weights=[1., 0.392], metrics={"capsnet": "accuracy"}) model.summary()
打印模型參數(shù)概要:
膠囊模型的結(jié)果:
Epoch 14/15 34/34 [==============================] - 108s 3s/step - loss: 0.0445 - capsnet_loss: 0.0218 - decoder_loss: 0.0579 - capsnet_acc: 0.9846 - val_loss: 0.0364 - val_capsnet_loss: 0.0159 - val_decoder_loss: 0.0522 - val_capsnet_acc: 0.9887 Epoch 15/15 34/34 [==============================] - 107s 3s/step - loss: 0.0423 - capsnet_loss: 0.0201 - decoder_loss: 0.0567 - capsnet_acc: 0.9859 - val_loss: 0.0362 - val_capsnet_loss: 0.0162 - val_decoder_loss: 0.0510 - val_capsnet_acc: 0.9880
為了便于總結(jié)分析,將以上三個(gè)實(shí)驗(yàn)的結(jié)構(gòu)繪制出測試精度圖:
plt.figure(figsize=(10, 8)) plt.plot(trained_model.history["val_acc"], "r", trained_model2.history["val_acc"], "b", trained_model3.history["val_capsnet_acc"], "g") plt.legend(("MLP", "CNN", "CapsNet"), loc="lower right", fontsize="large") plt.title("Validation Accuracies") plt.show()
從結(jié)果中可以看出,膠囊網(wǎng)絡(luò)的精度優(yōu)于CNN和MLP。
總結(jié)
本文對膠囊網(wǎng)絡(luò)進(jìn)行了非技術(shù)性的簡要概括,分析了其兩個(gè)重要屬性,之后針對MNIST手寫體數(shù)據(jù)集上驗(yàn)證多層感知機(jī)、卷積神經(jīng)網(wǎng)絡(luò)以及膠囊網(wǎng)絡(luò)的性能。
作者信息
Faizan Shaikh,數(shù)據(jù)科學(xué),深度學(xué)習(xí)初學(xué)者。
文章原標(biāo)題《Essentials of Deep Learning: Getting to know CapsuleNets (with Python codes)》,作者:Faizan Shaikh
詳情請閱讀原文
文章版權(quán)歸作者所有,未經(jīng)允許請勿轉(zhuǎn)載,若此文章存在違規(guī)行為,您可以聯(lián)系管理員刪除。
轉(zhuǎn)載請注明本文地址:http://systransis.cn/yun/41551.html
摘要:值得一提的是每篇文章都是我用心整理的,編者一貫堅(jiān)持使用通俗形象的語言給我的讀者朋友們講解機(jī)器學(xué)習(xí)深度學(xué)習(xí)的各個(gè)知識(shí)點(diǎn)。今天,紅色石頭特此將以前所有的原創(chuàng)文章整理出來,組成一個(gè)比較合理完整的機(jī)器學(xué)習(xí)深度學(xué)習(xí)的學(xué)習(xí)路線圖,希望能夠幫助到大家。 一年多來,公眾號(hào)【AI有道】已經(jīng)發(fā)布了 140+ 的原創(chuàng)文章了。內(nèi)容涉及林軒田機(jī)器學(xué)習(xí)課程筆記、吳恩達(dá) deeplearning.ai 課程筆記、機(jī)...
閱讀 3286·2021-11-24 09:39
閱讀 3042·2021-11-23 09:51
閱讀 918·2021-11-18 10:07
閱讀 3574·2021-10-11 10:57
閱讀 2792·2021-10-08 10:04
閱讀 3039·2021-09-26 10:11
閱讀 1100·2021-09-23 11:21
閱讀 2873·2019-08-29 17:28