Ajout solutions

solutions/classifier_pretrained_model.py

@@ -0,0 +1,21 @@
+model_VGG_fcm = km.Sequential()
+model_VGG_fcm.add(kl.Flatten(input_shape=features_train.shape[1:]))
+model_VGG_fcm.add(kl.Dense(64, activation='relu'))
+model_VGG_fcm.add(kl.Dropout(0.5))
+model_VGG_fcm.add(kl.Dense(1, activation='sigmoid'))
+
+model_VGG_fcm.compile(optimizer='rmsprop',
+              loss='binary_crossentropy',
+              metrics=['accuracy'])
+
+model_VGG_fcm.summary()
+
+
+train_labels = np.array([0] * int((N_train/2)) + [1] * int((N_train/2)))
+validation_labels = np.array([0] * int((N_val/2)) + [1] * int((N_val/2)))
+
+model_VGG_fcm.fit(features_train, train_labels,
+          epochs=epochs,
+          batch_size=batch_size,
+          validation_data=(features_validation, validation_labels))
+t_learning_VGG_fcm = te-ts

solutions/max_pooling.py

@@ -0,0 +1,13 @@
+conv_mp = km.Sequential([ kl.MaxPool2D(pool_size=(2,2))])
+
+img_in = np.expand_dims(x, 0)
+img_out = conv_mp.predict(img_in)
+
+fig, (ax0, ax1) = plt.subplots(ncols=2, figsize=(10, 5))
+ax0.imshow(img_in[0,:,:,0].astype(np.uint8),
+           cmap="binary");
+ax0.grid(False)
+
+ax1.imshow(img_out[0,:,:,0].astype(np.uint8),
+           cmap="binary");
+ax1.grid(False)

solutions/mnist_conv_architecture.py

@@ -0,0 +1,23 @@
+model = km.Sequential()
+model.add(kl.Conv2D(32, kernel_size=(3, 3), activation='relu', input_shape=(28,28, 1), data_format="channels_last"))
+model.add(kl.Conv2D(64, (3, 3), activation='relu'))
+model.add(kl.MaxPooling2D(pool_size=(2, 2)))
+model.add(kl.Dropout(0.25))
+model.add(kl.Flatten())
+model.add(kl.Dense(128, activation='relu'))
+model.add(kl.Dropout(0.5))
+model.add(kl.Dense(N_classes, activation='softmax'))
+# Résumé
+model.summary()
+# Apprentissage
+model.compile(loss="sparse_categorical_crossentropy",
+              optimizer=ko.Adadelta(),
+              metrics=['accuracy'])
+ts=time.time()
+model.fit(X_train_conv, Y_train,
+          batch_size=batch_size,
+          epochs=epochs,
+          verbose=1,
+          validation_data=(X_test_conv, Y_test))
+te=time.time()
+t_train_conv = te-ts

solutions/test_kaggle.py

@@ -0,0 +1,29 @@
+data_dir_test = data_dir+'test/'
+N_test = len(os.listdir(data_dir_test+"/test"))
+
+test_datagen = kpi.ImageDataGenerator(rescale=1. / 255)
+
+test_generator = test_datagen.flow_from_directory(
+    data_dir_test,
+    #data_dir_sub+"/train/",
+    target_size=(img_height, img_width),
+    batch_size=batch_size,
+    class_mode=None,
+    shuffle=False)
+
+test_prediction = model_VGG_LastConv_fcm.predict_generator(test_generator, N_test // batch_size)
+
+images_test = [data_dir_test+"/test/"+k for k in os.listdir(data_dir_test+"/test")][:9]
+x_test  = [kpi.img_to_array(kpi.load_img(image_test))/255 for image_test in images_test]  # this is a PIL image
+
+fig = plt.figure(figsize=(10,10))
+for k in range(9):
+    ax = fig.add_subplot(3,3,k+1)
+    ax.imshow(x_test[k], interpolation='nearest')
+    pred = test_prediction[k]
+    if pred >0.5:
+        title = "Probabiliy for dog : %.1f" %(pred*100)
+    else:
+        title = "Probabiliy for cat : %.1f" %((1-pred)*100)
+    ax.set_title(title)
+plt.show()