Covolution<\/h2>\n\n\n\n
Instead of learning full matrix, only a couple of weights of the kernel is needed to be learned.<\/p>\n\n\n\n
Weights are repeated (the same).<\/p>\n\n\n\n It looks like a feed forward network.<\/p>\n\n\n\n With sparse connectivity inputs are only influences the activation of the layers its connected to in the next layer. This means chaning the input, might not yield influential changes in the next layer, in comparinson to using the entire matrix. Although it will have larger influence in future layers. <\/p>\n\n\n\n Detpth of the output colume is a hyerparameter. The set of neurons that are in the same reguin of the input is defined as input depth.<\/p>\n\n\n\n Multiple kernels works in multiplication as defined (w^t)*x+h. The images do not need to match any sizes for multiple kernels to work. It can however get enificient.<\/p>\n\n\n\n Image –> Low level features –> Mid-level features –> High-level features — Lineraly seprable classifier –> …<\/p>\n\n\n\n Absence of equvariance<\/p>\n\n\n\n To have argumented data, images need to be rotated, zoomed and manipulated manually.<\/p>\n\n\n\n With pooling we empathise the need for the input image for low level features, and that it is no longer needed when working with high-level features.<\/p>\n\n\n\n Pooling modifes the output to a smaller dimension.<\/p>\n\n\n\n Stages in deep learning:<\/p>\n\n\n\n A classic aritecure:<\/p>\n\n\n\n The types of pooling functions:<\/p>\n\n\n\n Pooling should\/could be used to:<\/p>\n\n\n\n VGG is a convolutional network model. Diverstions to the model are common.<\/p>\n\n\n\n Vgg is used for very deep convolutional networks for large scale image recognition.<\/p>\n\n\n\n From keras it can be chosen if convolutional netork, all or both should be used.<\/p>\n\n\n\n It’s a small kernel. It’s more intresting to look at previous levels that already modified the input.<\/p>\n\n\n\n For images comparable to constructing adversarial examples can be created. This is done by using any grey scale image and calculate the overall activation of the filter of interest. If maximization of the peak of this particular filter is performed computation of the gradient with respect to the input and adaptation of the out\/input image to maximize overall activation of the filter can be performed.<\/p>\n\n\n\n The more layers are used, the more complex and colourfully diverse the filters will be.<\/p>\n","protected":false},"excerpt":{"rendered":" Last time: Key-ideas to exploited Convolution Visual explanations of conolution In convolution the entire input is used and manipulated by a kernel in chunks From a input string, an matrix is created 2d image manipulation example Covolution Instead of learning full matrix, only a couple of weights of the kernel is needed to be learned….<\/p>\n","protected":false},"author":1,"featured_media":715,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[11],"tags":[33,30],"class_list":["post-1464","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-noter","tag-deep-learning","tag-sdu"],"_links":{"self":[{"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/posts\/1464","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/comments?post=1464"}],"version-history":[{"count":6,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/posts\/1464\/revisions"}],"predecessor-version":[{"id":1495,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/posts\/1464\/revisions\/1495"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/media\/715"}],"wp:attachment":[{"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/media?parent=1464"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/categories?post=1464"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/maxfest.dk\/port\/wp-json\/wp\/v2\/tags?post=1464"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-1.png\")
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image.png\")
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-6.png\")
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-2.png\")
Parameter sharing:<\/h3>\n\n\n\n
Depths of the outcome volume<\/h3>\n\n\n\n
Multiple kernels<\/h3>\n\n\n\n
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-7-1024x482.png\")
The method is not equivariant.<\/h3>\n\n\n\n
Pooling<\/h2>\n\n\n\n
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-3.png\")
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-4-1024x722.png\")
VVG \/ VVG Net<\/h3>\n\n\n\n
![\"\"](\"https:\/\/maxfest.dk\/port\/wp-content\/uploads\/2021\/10\/image-5.png\")
Variants of Convolution<\/h2>\n\n\n\n
Initialiazation<\/h2>\n\n\n\n
Convolution in Keras<\/h2>\n\n\n\n
Defining Convolutions in Keras <\/h3>\n\n\n\n
# Defining Convolutions in Keras\nfrom keras import layers\nfrom keras import models\n\nmodel = models.Sequential()\nmodel.add(layers.Conv2D(32, (3, 3), activation='relu',\ninput_shape=(28, 28, 1)))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2, 2)))\nmodel.add(layers.Conv2D(64, (3, 3), activation='relu\u2019))\n\nmodel.add(layers.Flatten())\nmodel.add(layers.Dense(64, activation='relu'))\nmodel.add(layers.Dense(10, activation='softmax'))<\/pre>\n\n\n\n
The Convolutional Layer in Keras <\/h3>\n\n\n\n
\n# The Convolutional Layer in Keras\nkeras.layers.Conv2D(filters, kernel_size, #Number of filters, kernel dimensions\n strides=(1, 1), # Using a stride?\n padding='valid\u2019, # \u2018valid\u2019 means only valid positions are used.\n \u2018same\u2019 means output has same dimension as input\n data_format=None, # \u2018channels_last\u2019 or \u2018channels_first\u2019\n dilation_rate=(1, 1), # Increase the field of \u201cvision\u201d\n activation=None, # Rest of the parameters same as dense layer\n use_bias=True,\n kernel_initializer='glorot_uniform\u2019,\n bias_initializer='zeros\u2019,\n kernel_regularizer=None,\n bias_regularizer=None,\n activity_regularizer=None,\n kernel_constraint=None,\n bias_constraint=None\n)<\/pre>\n\n\n\n
The Pooling Layer in Keras<\/h3>\n\n\n\n
# The Pooling Layer in Keras\nkeras.layers.MaxPooling2D(\n pool_size=(2, 2), # Size of the pooling\n strides=None, # Integer, tuple of 2 integers, or None. Strides\n values. If None, it will default to pool_size.\n padding='valid\u2019, # One of \"valid\" or \"same\"\n data_format=None # Same as with Conv2D\n)<\/pre>\n\n\n\n
Training your Model<\/h3>\n\n\n\n
# Training your Model\nhistory = model.fit_generator(\n train_generator,\n steps_per_epoch=100,\n epochs=30,\n validation_data=validation_generator,\n validation_steps=50\n)<\/pre>\n\n\n\n
Automatically Distort Images <\/h3>\n\n\n\n
# Automatically Distort Images\ndatagen = ImageDataGenerator(\n rotation_range=40,\n width_shift_range=0.2,\n height_shift_range=0.2,\n shear_range=0.2,\n zoom_range=0.2,\n horizontal_flip=True,\n fill_mode='nearest\u2019\n)<\/pre>\n\n\n\n
Visualize Activations<\/h3>\n\n\n\n
from keras.models import Model\nlayer_outputs = [layer.output for layer in model.layers]\nactivation_model = Model(inputs=model.input, outputs=layer_outputs)\nactivations = activation_model.predict(X_train[10].reshape(1,28,28,1))\n\ndef display_activation(activations, col_s, row_s, act_index):\n activation = activations[act_index]\n activation_index=0\n fig, ax = plt.subplots(row_size, col_size, figsize=(row_s2.5,col_s<\/em>1.5))\n for row in range(0,row_size):\n for col in range(0,col_size):\n ax[row][col].imshow(activation[0, :, :, activation_index], cmap='gray')\n activation_index += 1<\/pre>\n\n\n\n
Activation filters<\/h2>\n\n\n\n