{"cells":[{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import numpy as np # linear algebra\nimport pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)\nimport os\nfrom skimage import data,io,filters\nfrom sklearn import preprocessing\n\nfrom keras import Sequential\nfrom keras.layers.core import Dense,Flatten,Dropout\nfrom keras.layers.convolutional import Conv2D, MaxPooling2D, UpSampling2D\nfrom keras.optimizers import SGD\nfrom keras.layers import Input,Reshape\nfrom keras.models import Model\n\nimport cv2","execution_count":null,"outputs":[]},{"metadata":{"_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","trusted":true,"collapsed":true},"cell_type":"code","source":"df = pd.read_csv(\"../input/train.csv\")\ny = df['label'].values\nX = (1 - df.iloc[:,1:].values/255).reshape(-1,28,28)\nX = np.expand_dims(X, axis=3)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"01bc1fc2239c87091a68b622e226e647c25bd88e"},"cell_type":"code","source":"input_img = Input(shape=(28, 28, 1))  # adapt this if using `channels_first` image data format\nprint(input_img.get_shape)\nx = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)\nprint(x.get_shape)\nx = MaxPooling2D((2, 2), padding='same')(x)\nprint(x.get_shape)\nx = Conv2D(8, (3, 3), activation='relu', padding='same')(x)\nprint(x.get_shape)\nx = MaxPooling2D((2, 2), padding='same')(x)\nprint(x.get_shape)\nx = Conv2D(8, (3, 3), activation='relu', padding='same')(x)\nprint(x.get_shape)\nencoded = MaxPooling2D((2, 2), padding='same')(x)\n\n# at this point the representation is (4, 4, 8) i.e. 128-dimensional\n\nx = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)\nprint(x.get_shape)\nx = UpSampling2D((2, 2))(x)\nprint(x.get_shape)\nx = Conv2D(8, (3, 3), activation='relu', padding='same')(x)\nprint(x.get_shape)\nx = UpSampling2D((2, 2))(x)\nprint(x.get_shape)\nx = Conv2D(16, (3, 3), activation='relu')(x)\nprint(x.get_shape)\nx = UpSampling2D((2, 2))(x)\nprint(x.get_shape)\ndecoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)\n\nprint(decoded.get_shape)\nautoencoder = Model(input_img,decoded)\nautoencoder.compile(optimizer='adadelta', loss='binary_crossentropy', metrics=['accuracy'])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"3578e378369e29433685f5acbcd7ba32660879bf"},"cell_type":"code","source":"from keras.callbacks import TensorBoard\n\nautoencoder.fit(X, X,\n                epochs=10,\n                batch_size=128,\n                shuffle=True,\n                callbacks=[TensorBoard(log_dir='/tmp/autoencoder')])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"debb76e301650007aeca7aa7aaa32a6fdc1e8cce"},"cell_type":"code","source":"import matplotlib.pyplot as plt\n\ndecoded_imgs = autoencoder.predict(X[0:10])\n\nn = 10\nplt.figure(figsize=(20, 4))\nfor i in range(1,n):\n    # display original\n    ax = plt.subplot(2, n, i)\n    plt.imshow(X[i].reshape(28, 28))\n    plt.gray()\n    ax.get_xaxis().set_visible(False)\n    ax.get_yaxis().set_visible(False)\n\n    # display reconstruction\n    ax = plt.subplot(2, n, i + n)\n    plt.imshow(decoded_imgs[i].reshape(28, 28))\n    plt.gray()\n    ax.get_xaxis().set_visible(False)\n    ax.get_yaxis().set_visible(False)\nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"dbdcca75d8c13f5dd0bdc4f8faf11e303c49d3d6"},"cell_type":"code","source":"encoder = Sequential()\nfor layer in autoencoder.layers[0:7]:\n    encoder.add(layer)\n\ndecoder = Sequential()\ndecoder.add(Reshape(input_shape=(4,4,8), target_shape=(4,4,8)))\nfor layer in autoencoder.layers[7:]:\n    print(layer.output_shape)\n    decoder.add(layer)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"bff3a5bfb78c6fb01f557af08adfaef07ee395b6"},"cell_type":"code","source":"x_0 = X[10]\nprint(y[10])\nplt.figure()\nplt.subplot(131)\nplt.imshow(x_0.reshape(28,28))\nx_0 = np.expand_dims(x_0, axis=0)\nx_enc = encoder.predict(x_0)\nnew_x = decoder.predict(x_enc)\nplt.subplot(132)\nplt.imshow(x_enc.reshape(16, 8))\nplt.subplot(133)\nplt.imshow(new_x.reshape(28, 28))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"ea1697d2d935a83a194faaf772a33f0be84a082d","collapsed":true},"cell_type":"code","source":"from sklearn.decomposition import PCA\n\nnumbers_avg = np.zeros((10,128))\nnumbers_occ = np.zeros(10)\n\nX_enc = encoder.predict(X).reshape(-1,128)\n\nPCA_ = PCA(n_components = 128).fit(X_enc)\nX_enc_pca = PCA_.transform(X_enc)\nfor i, number in enumerate(y):\n    numbers_occ[number] += 1\n    numbers_avg[number] += X_enc_pca[i]\n\nfor i in range(0,10):\n    numbers_avg[i] /= numbers_occ[i]","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"0f7589de7653db94de582b390b6f4459d96d8503"},"cell_type":"code","source":"n = 11\nplt.figure(figsize=(20, 4))\nfor i in range(1,n):\n    # display original\n    ax = plt.subplot(1, n, i)\n    # display reconstruction\n    plt.imshow(decoder.predict(PCA_.inverse_transform(numbers_avg[i-1]).reshape(1,4,4,8)).reshape(28, 28))\n    plt.gray()\n    ax.get_xaxis().set_visible(False)\n    ax.get_yaxis().set_visible(False)\nplt.show()\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"ac263c76abf4d766a18951996ab6b21d34501f44"},"cell_type":"code","source":"from sklearn.manifold import TSNE\n\ntsne = TSNE()\nx_tsne = tsne.fit_transform(X_enc_pca[0:2000])\nclasses = set(y)\ncolors = [plt.cm.tab10(float(i)/max(classes)) for i in classes]\nfor i, c in enumerate(classes):\n    xi = [x_tsne[j,0] for j  in range(len(x_tsne)) if y[j] == c]\n    yi = [x_tsne[j,1] for j  in range(len(x_tsne)) if y[j] == c]\n    plt.scatter(xi, yi, c=colors[i], label=str(c))\nplt.legend()\nplt.show()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"_uuid":"a28f1c21d429a5f071dd1e3efb63c81506f0abe3","scrolled":true},"cell_type":"code","source":"mutant = np.zeros(128)\nn = 11\nplt.figure(figsize=(20, 20))\nfor i in range(1,n):\n    mutant[0] = 10*(i-1)\n    for j in range(1,n):\n        # display original\n        ax = plt.subplot(n, n, (i-1)*n+j)\n        # display reconstruction\n        mutant[1] = 10*(j-1)\n        plt.imshow(decoder.predict(PCA_.inverse_transform(mutant).reshape(1,4,4,8)).reshape(28, 28))\n        plt.gray()\n        ax.get_xaxis().set_visible(False)\n        ax.get_yaxis().set_visible(False)\nplt.show()\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true,"collapsed":true,"_uuid":"df69312430b4f3e5bda7062123b6ded6e62d08fa"},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"display_name":"Python 3","language":"python","name":"python3"},"language_info":{"name":"python","version":"3.6.4","mimetype":"text/x-python","codemirror_mode":{"name":"ipython","version":3},"pygments_lexer":"ipython3","nbconvert_exporter":"python","file_extension":".py"}},"nbformat":4,"nbformat_minor":1}