bonus: semi-supervised

hw3/semi.py

@@ -0,0 +1,189 @@
+#!/usr/bin/env python
+import sys, os
+import numpy as np
+from keras.utils import to_categorical
+from keras.preprocessing.image import ImageDataGenerator
+from keras.models import Sequential
+from keras.layers import Dense, Activation, Flatten, Dropout, LeakyReLU
+from keras.layers import Conv2D, MaxPooling2D, AveragePooling2D
+from keras.layers import BatchNormalization
+from keras import losses
+from keras import optimizers
+from keras.callbacks import CSVLogger, ModelCheckpoint
+from keras.models import load_model
+
+# Parameter
+height = width = 48
+num_classes = 7
+input_shape = (height, width, 1)
+batch_size = 128
+epochs_1 = 30
+epochs_2 = 50
+zoom_range = 0.2
+model_name = 'semi.h5'
+isValid = 1
+
+def construct_model(model):
+ model.add(Conv2D(64, (3, 3), padding='same', input_shape=input_shape))
+ model.add(LeakyReLU(alpha=0.03))
+ model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))
+ model.add(Dropout(0.2))
+
+ model.add(Conv2D(128, (3, 3), padding='same'))
+ model.add(LeakyReLU(alpha=0.03))
+ model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))
+ model.add(Dropout(0.25))
+
+ model.add(Conv2D(256, (3, 3), padding='same'))
+ model.add(LeakyReLU(alpha=0.03))
+ model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))
+ model.add(Dropout(0.3))
+
+ model.add(Conv2D(512, (3, 3), padding='same'))
+ model.add(LeakyReLU(alpha=0.03))
+ model.add(BatchNormalization())
+ model.add(MaxPooling2D(pool_size=(2, 2), padding='same'))
+ model.add(Dropout(0.4))
+
+ model.add(Flatten())
+
+ model.add(Dense(256, activation='relu'))
+ model.add(BatchNormalization())
+ model.add(Dropout(0.5))
+ model.add(Dense(512, activation='relu'))
+ model.add(BatchNormalization())
+ model.add(Dropout(0.5))
+ model.add(Dense(num_classes))
+ model.add(Activation('softmax'))
+
+ model.summary()
+
+def compile_model(model):
+ model.compile(loss='categorical_crossentropy',
+ optimizer='adam',
+ metrics=['accuracy'])
+
+
+# Read the train data
+try:
+ X = np.load('X.npy')
+ Y = np.load('Y.npy')
+except:
+ with open(sys.argv[1], "r+") as f:
+ line = f.read().strip().replace(',', ' ').split('\n')[1:]
+ raw_data = ' '.join(line)
+ length = width*height+1 #1 is for label
+ data = np.array(raw_data.split()).astype('float').reshape(-1, length)
+ X = data[:, 1:]
+ Y = data[:, 0]
+ # Change data into CNN format
+ X = X.reshape(X.shape[0], height, width, 1)
+ Y = Y.reshape(Y.shape[0], 1)
+ print('Saving X.npy & Y.npy')
+ np.save('X.npy', X) # (-1, height, width, 1)
+ np.save('Y.npy', Y) # (-1, 1)
+
+X /= 255
+Y = to_categorical(Y, num_classes)
+
+
+# Split the data
+if isValid:
+ valid_num = 3000
+ X_train, Y_train = X[:-valid_num], Y[:-valid_num]
+ X_valid, Y_valid = X[-valid_num:], Y[-valid_num:]
+
+else:
+ X_train, Y_train = X, Y
+
+print(X_train.shape)
+print(Y_train.shape)
+print(X_valid.shape)
+print(Y_valid.shape)
+
+## split the label && unlabel
+split_point = 20000
+X_label, Y_label = X_train[:-split_point], Y_train[:-split_point]
+X_unlabel = X_train[-split_point:]
+print(X_label.shape)
+print(Y_label.shape)
+print(X_unlabel.shape)
+
+
+# Construct the model
+model = Sequential()
+construct_model(model)
+compile_model(model)
+
+# Image PreProcessing
+train_gen = ImageDataGenerator(rotation_range=25,
+ width_shift_range=0.1,
+ height_shift_range=0.1,
+ shear_range=0.1,
+ zoom_range=[1-zoom_range, 1+zoom_range],
+ horizontal_flip=True)
+train_gen.fit(X_label)
+
+# Callbacks
+callbacks = []
+modelcheckpoint = ModelCheckpoint('semi1_ckpt/weights.{epoch:03d}-{val_acc:.5f}.h5', monitor='val_acc', save_best_only=True)
+callbacks.append(modelcheckpoint)
+csv_logger = CSVLogger('_semi_log.csv', separator=',', append=False)
+callbacks.append(csv_logger)
+
+# Fit the model
+if isValid:
+ model.fit_generator(train_gen.flow(X_label, Y_label, batch_size=batch_size),
+ steps_per_epoch=10*X_label.shape[0]//batch_size,
+ epochs=epochs_1,
+ callbacks=callbacks,
+ validation_data=(X_valid, Y_valid))
+else:
+ model.fit_generator(train_gen.flow(X_label, Y_label, batch_size=batch_size),
+ steps_per_epoch=10*X_label.shape[0]//batch_size,
+ epochs=epochs_1,
+ callbacks=callbacks)
+
+# Save model
+model.save('_{}'.format(model_name))
+# model = load_model(model_name)
+
+# Predict on unlabel data
+Y_unlabel = model.predict_classes(X_unlabel)
+Y_unlabel = to_categorical(Y_unlabel, num_classes)
+
+new_X = np.concatenate((X_label, X_unlabel), axis=0)
+new_Y = np.concatenate((Y_label, Y_unlabel), axis=0)
+
+# Training 2
+model2 = Sequential()
+construct_model(model2)
+compile_model(model2)
+
+train_gen.fit(new_X)
+
+# Callbacks
+callbacks = []
+modelcheckpoint = ModelCheckpoint('semi2_ckpt/weights.{epoch:03d}-{val_acc:.5f}.h5', monitor='val_acc', save_best_only=True)
+callbacks.append(modelcheckpoint)
+csv_logger = CSVLogger('semi_log.csv', separator=',', append=True)
+callbacks.append(csv_logger)
+
+# Fit the model
+if isValid:
+ model2.fit_generator(train_gen.flow(new_X, new_Y, batch_size=batch_size),
+ steps_per_epoch=10*new_X.shape[0]//batch_size,
+ epochs=epochs_2,
+ callbacks=callbacks,
+ validation_data=(X_valid, Y_valid))
+else:
+ model2.fit_generator(train_gen.flow(new_X, new_Y, batch_size=batch_size),
+ steps_per_epoch=10*new_X.shape[0]//batch_size,
+ epochs=epochs_2,
+ callbacks=callbacks)
+
+# Save model
+model.save(model_name)