106062568_homework1 江愷笙

code/Alexnet.py

@@ -0,0 +1,133 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Tue Oct  3 14:31:28 2017
+
+@author: user
+"""
+import tensorflow as tf
+#import matplotlib.pyplot as plt 
+import numpy as np
+import input_data
+import os
+
+#train_dir = 'Lhand'
+# define some constants
+LR = 0.001                 # define learning rate
+batch_size = 16             # define batch size
+in_size = 200*200           # define input size
+out_size = 24               # define output size
+dropout = 0.5               # the probability to dropout
+epoch = 50
+train_data_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/data/frames/train/'
+label_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/'
+test_data_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/data/frames/test/'
+save_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/code/checkpoints/'
+image_list, label_list = input_data.read_data(train_data_dir, label_dir, train = True)
+test_image, test_label = input_data.read_data(test_data_dir, label_dir, train = False)
+N_sample = len(image_list)
+
+batch_xs, batch_ys = input_data.batch_generate(image_list, label_list, 200, 200, batch_size, 3000)
+test_batch_xs, test_batch_ys =input_data.batch_generate(test_image, test_label, 200, 200, batch_size, 3000)
+# define placeholder for inputs to network
+xs = tf.placeholder(tf.float32, [None, 200, 200, 3])
+ys = tf.placeholder(tf.float32, [None, out_size])
+keep_prob = tf.placeholder(tf.float32)
+
+# Alexnet structure
+# conv_layer1
+conv1 = tf.layers.conv2d(xs, filters = 64, kernel_size = 3, strides = 1, padding = 'same', activation = tf.nn.relu)
+MaxPool1 = tf.layers.max_pooling2d(conv1, pool_size = 2, strides = 2)
+norm1 = tf.nn.lrn(MaxPool1, 4, bias = 1.0, alpha = 0.001 / 9.0, beta = 0.75)
+
+# conv_layer2
+conv2 = tf.layers.conv2d(norm1, 256, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool2 = tf.layers.max_pooling2d(conv2, pool_size = 2, strides = 2)
+norm2 = tf.nn.lrn(MaxPool2, 4, bias = 1.0, alpha = 0.001 / 9.0, beta = 0.75)
+
+# conv_layer3
+conv3 = tf.layers.conv2d(norm2, 300, 3, 1, 'same', activation = tf.nn.relu)
+#MaxPool3 = tf.layers.max_pooling2d(conv3, pool_size = 2, strides = 2)
+
+# conv_layer4
+conv4 = tf.layers.conv2d(conv3, 300, 3, 1, 'same', activation = tf.nn.relu)
+#MaxPool4 = tf.layers.max_pooling2d(conv4, pool_size = 2, strides = 2)
+
+# conv_layer5
+conv5 = tf.layers.conv2d(conv4, 256, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool5 = tf.layers.max_pooling2d(conv5, pool_size = 2, strides = 2)
+#norm5 = tf.nn.lrn(MaxPool5, 4, bias = 1.0, alpha = 0.001 / 9.0, beta = 0.75)
+
+# flatened
+flat = tf.reshape(MaxPool5, [-1, 25*25*256])
+
+# fully connected layer6
+fc6 = tf.layers.dense(flat, 1024, activation = tf.nn.relu)
+drop6 = tf.nn.dropout(fc6, dropout)
+
+# fully connected layer7
+fc7 = tf.layers.dense(drop6, 1024, activation = tf.nn.relu)
+drop7 = tf.nn.dropout(fc7, dropout)
+
+# fully connected layer8
+output = tf.layers.dense(drop7, 24, activation = tf.nn.softmax)
+
+# define loss function
+loss = tf.losses.softmax_cross_entropy(onehot_labels = ys, logits = output)
+train = tf.train.AdamOptimizer(LR).minimize(loss)
+accuracy = tf.reduce_sum(tf.cast(tf.equal(tf.argmax(ys, axis=1), tf.argmax(output, axis=1)), tf.float32))
+t_accuracy = tf.reduce_sum(tf.cast(tf.equal(tf.argmax(ys, axis=1), tf.argmax(output, axis=1)), tf.float32))
+#acc_correct = 0
+saver = tf.train.Saver()
+
+# start a session
+sess = tf.Session()
+init = tf.global_variables_initializer()
+sess.run(init)                                          # initialize all global variables
+sess.run(tf.local_variables_initializer())              # initialize all local variables(such as acc_op)
+
+coord = tf.train.Coordinator()
+threads = tf.train.start_queue_runners(sess=sess, coord=coord)
+
+tot_count = 0
+count = 0
+compute_accuracy = 0
+for i in range(epoch):
+    for iteration in range(int(N_sample / batch_size)):
+        if coord.should_stop():
+            break
+        count += 1
+        tot_count += batch_size 
+        image_batch, label_batch = sess.run([batch_xs, batch_ys])
+        train_step = sess.run(train, feed_dict = {xs: image_batch, ys: label_batch})
+        acc_correct = sess.run(accuracy, feed_dict = {xs: image_batch, ys: label_batch})
+        compute_accuracy = acc_correct + compute_accuracy
+        if iteration % 50 == 0:
+            loss_step = sess.run(loss, feed_dict = {xs: image_batch, ys: label_batch})
+
+            tot_accuracy = compute_accuracy / tot_count
+            print('Step:', count, '| train loss: %.4f' % loss_step, '| train accuracy: %.2f' % tot_accuracy)
+        if (i == epoch - 1) and (iteration == int(N_sample / batch_size) - 1):
+            checkpoint_path = os.path.join(save_dir, 'model.ckpt')
+            saver.save(sess, checkpoint_path, global_step=count)
+
+
+
+tot_count = 0
+count = 0
+test_accuracy = 0
+tot_test_accuracy = 0
+for step in range(int(len(test_image) / batch_size)):
+    if coord.should_stop():
+        break
+    count += 1
+    tot_count += batch_size
+    test_batch_x, test_batch_y = sess.run([test_batch_xs, test_batch_ys])
+    acc_ = sess.run(t_accuracy, feed_dict = {xs: test_batch_x, ys: test_batch_y})
+    test_accuracy = acc_ + test_accuracy
+    tot_test_accuracy = test_accuracy / tot_count
+print('test accuracy: %.2f' % tot_test_accuracy)
+
+coord.request_stop()       
+coord.join(threads)
+sess.close()
+

code/VGG19.py

@@ -0,0 +1,147 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Tue Oct  3 14:31:28 2017
+
+@author: user
+"""
+import tensorflow as tf
+#import matplotlib.pyplot as plt 
+#import numpy as np
+import input_data
+import os
+
+#train_dir = 'Lhand'
+# define some constants
+LR = 0.001                 # define learning rate
+batch_size = 16             # define batch size
+in_size = 224*224           # define input size
+out_size = 24               # define output size
+dropout = 0.5               # the probability to dropout
+epoch = 20
+train_data_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/data/frames/train/'
+label_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/'
+test_data_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/data/frames/test/'
+save_dir = '/home/viplab/Desktop/petersci/CEDL_HW1/code/checkpoints_VGG/'
+image_list, label_list = input_data.read_data(train_data_dir, label_dir, train = True)
+test_image, test_label = input_data.read_data(test_data_dir, label_dir, train = False)
+N_sample = len(image_list)
+
+batch_xs, batch_ys = input_data.batch_generate(image_list, label_list, 224, 224, batch_size, 3000)
+test_batch_xs, test_batch_ys =input_data.batch_generate(test_image, test_label, 224, 224, batch_size, 3000)
+# define placeholder for inputs to network
+xs = tf.placeholder(tf.float32, [None, 224, 224, 3])
+ys = tf.placeholder(tf.float32, [None, out_size])
+keep_prob = tf.placeholder(tf.float32)
+
+# VGG structure
+# conv_layer1
+conv1_1 = tf.layers.conv2d(xs, filters = 32, kernel_size = 3, strides = 1, padding = 'same', activation = tf.nn.relu)
+conv1_2 = tf.layers.conv2d(conv1_1, 32, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool1 = tf.layers.max_pooling2d(conv1_2, pool_size = 2, strides = 2)
+#norm1 = tf.nn.lrn(MaxPool1, 4, bias = 1.0, alpha = 0.001 / 9.0, beta = 0.75)
+
+# conv_layer2
+conv2_1 = tf.layers.conv2d(MaxPool1, 64, 3, 1, 'same', activation = tf.nn.relu)
+conv2_2 = tf.layers.conv2d(conv2_1, 64, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool2 = tf.layers.max_pooling2d(conv2_2, pool_size = 2, strides = 2)
+#norm2 = tf.nn.lrn(MaxPool2, 4, bias = 1.0, alpha = 0.001 / 9.0, beta = 0.75)
+
+# conv_layer3
+conv3_1 = tf.layers.conv2d(MaxPool2, 128, 3, 1, 'same', activation = tf.nn.relu)
+conv3_2 = tf.layers.conv2d(conv3_1, 128, 3, 1, 'same', activation = tf.nn.relu)
+conv3_3 = tf.layers.conv2d(conv3_2, 128, 3, 1, 'same', activation = tf.nn.relu)
+conv3_4 = tf.layers.conv2d(conv3_3, 128, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool3 = tf.layers.max_pooling2d(conv3_4, pool_size = 2, strides = 2)
+
+# conv_layer4
+conv4_1 = tf.layers.conv2d(MaxPool3, 256, 3, 1, 'same', activation = tf.nn.relu)
+conv4_2 = tf.layers.conv2d(conv4_1, 256, 3, 1, 'same', activation = tf.nn.relu)
+conv4_3 = tf.layers.conv2d(conv4_2, 256, 3, 1, 'same', activation = tf.nn.relu)
+conv4_4 = tf.layers.conv2d(conv4_3, 256, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool4 = tf.layers.max_pooling2d(conv4_4, pool_size = 2, strides = 2)
+
+# conv_layer5
+conv5_1 = tf.layers.conv2d(MaxPool4, 256, 3, 1, 'same', activation = tf.nn.relu)
+conv5_2 = tf.layers.conv2d(conv5_1, 256, 3, 1, 'same', activation = tf.nn.relu)
+conv5_3 = tf.layers.conv2d(conv5_2, 256, 3, 1, 'same', activation = tf.nn.relu)
+conv5_4 = tf.layers.conv2d(conv5_3, 256, 3, 1, 'same', activation = tf.nn.relu)
+MaxPool5 = tf.layers.max_pooling2d(conv5_4, pool_size = 2, strides = 2)
+
+# flatened
+flat = tf.reshape(MaxPool5, [-1, 7*7*256])
+
+# fully connected layer6
+fc6 = tf.layers.dense(flat, 1024, activation = tf.nn.relu)
+drop6 = tf.nn.dropout(fc6, dropout)
+
+# fully connected layer7
+fc7 = tf.layers.dense(drop6, 1024, activation = tf.nn.relu)
+drop7 = tf.nn.dropout(fc7, dropout)
+
+# fully connected layer8
+output = tf.layers.dense(drop7, 24)
+
+# define loss function
+loss = tf.losses.softmax_cross_entropy(onehot_labels = ys, logits = output)
+train = tf.train.AdamOptimizer(LR).minimize(loss)
+accuracy = tf.reduce_sum(tf.cast(tf.equal(tf.argmax(ys, axis=1), tf.argmax(output, axis=1)), tf.float32))
+t_accuracy = tf.reduce_sum(tf.cast(tf.equal(tf.argmax(ys, axis=1), tf.argmax(output, axis=1)), tf.float32))
+#acc_correct = 0
+saver = tf.train.Saver()
+
+# start a session
+sess = tf.Session()
+init = tf.global_variables_initializer()
+sess.run(init)                                          # initialize all global variables
+sess.run(tf.local_variables_initializer())              # initialize all local variables(such as acc_op)
+
+coord = tf.train.Coordinator()
+threads = tf.train.start_queue_runners(sess=sess, coord=coord)
+
+tot_count = 0
+count = 0
+compute_accuracy = 0
+tot_accuracy = 0
+epoch_loss = 0
+tot_loss = 0
+for i in range(epoch):
+    print('[epoch %d]' % (i+1))
+    for iteration in range(int(N_sample / batch_size)):
+        if coord.should_stop():
+            break
+        count += 1
+        tot_count += batch_size 
+        image_batch, label_batch = sess.run([batch_xs, batch_ys])
+        loss_step = sess.run(loss, feed_dict = {xs: image_batch, ys: label_batch})
+        train_step = sess.run(train, feed_dict = {xs: image_batch, ys: label_batch})
+        acc_correct = sess.run(accuracy, feed_dict = {xs: image_batch, ys: label_batch})
+        compute_accuracy = acc_correct + compute_accuracy
+        tot_accuracy = compute_accuracy / tot_count
+        epoch_loss = loss_step + epoch_loss
+        tot_loss = epoch_loss / count
+        if iteration % 50 == 0:
+            print('Step:', count, '| loss: %.4f' % loss_step, '| train accuracy: %.4f' % tot_accuracy)
+        if (i == epoch - 1) and (iteration == int(N_sample / batch_size) - 1):
+            checkpoint_path = os.path.join(save_dir, 'model.ckpt')
+            saver.save(sess, checkpoint_path, global_step=count)
+    print('[epoch %d] ======> ' % (i+1), 'epoch loss: %.4f' % tot_loss, 'epoch accuracy: %.4f' % tot_accuracy)
+
+tot_count = 0
+count = 0
+test_accuracy = 0
+tot_test_accuracy = 0
+for step in range(int(len(test_image) / batch_size)):
+    if coord.should_stop():
+        break
+    count += 1
+    tot_count += batch_size
+    test_batch_x, test_batch_y = sess.run([test_batch_xs, test_batch_ys])
+    acc_ = sess.run(t_accuracy, feed_dict = {xs: test_batch_x, ys: test_batch_y})
+    test_accuracy = acc_ + test_accuracy
+    tot_test_accuracy = test_accuracy / tot_count
+print('test accuracy: %.4f' % tot_test_accuracy)
+
+coord.request_stop()       
+coord.join(threads)
+sess.close()
+

code/checkpoints_VGG/checkpoint

@@ -0,0 +1,2 @@
+model_checkpoint_path: "/home/viplab/Desktop/petersci/CEDL_HW1/code/checkpoints_VGG/model.ckpt-18740"
+all_model_checkpoint_paths: "/home/viplab/Desktop/petersci/CEDL_HW1/code/checkpoints_VGG/model.ckpt-18740"

code/input_data.py

@@ -0,0 +1,105 @@
+# -*- coding: utf-8 -*-
+"""
+Created on Wed Oct 11 20:23:41 2017
+
+@author: user
+"""
+
+import tensorflow as tf
+import numpy as np
+import os
+import re
+
+
+def read_data(file_dir, label_dir, train = True):
+
+    image_list = []
+    image_temp = []
+#    temp = []
+    for root, direc, files in os.walk(file_dir):
+        # image directories
+        for name in files:
+            image_temp.append(os.path.join(root, name))
+        image_list = image_list + sorted(image_temp, key = lambda x: int(re.sub('\D', '', x)))
+        image_temp = []
+
+    label_list = []  
+    for root, direc, files in os.walk(label_dir):
+#        print(root)
+        if train:   
+            if root == '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/lab':
+#                print("111")
+                for i in range(1, 5):
+                    #print('temp = ', temp)
+                    label_list.extend(np.load(os.path.join(root, 'obj_left'+ str(i) + '.npy')))
+                    #print('label_list = ', label_list)
+                for i in range(1, 5):
+                    label_list.extend(np.load(os.path.join(root, 'obj_right'+ str(i) + '.npy')))
+            if (root == '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/house') or (root == '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/office'):
+#                print("222")
+                for i in range(1, 4):
+                    label_list.extend(np.load(os.path.join(root, 'obj_left'+ str(i) + '.npy')))
+                for i in range(1, 4):
+                    label_list.extend(np.load(os.path.join(root, 'obj_right'+ str(i) + '.npy')))
+        else:
+            if root == '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/lab':
+                for i in range(5, 9):
+                    label_list.extend(np.load(os.path.join(root, 'obj_left'+ str(i) + '.npy')))
+                for i in range(5, 9):
+                    label_list.extend(np.load(os.path.join(root, 'obj_right'+ str(i) + '.npy')))
+            if(root == '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/house') or (root == '/home/viplab/Desktop/petersci/CEDL_HW1/data/labels/office'):
+                for i in range(4, 7):
+                    label_list.extend(np.load(os.path.join(root, 'obj_left'+ str(i) + '.npy')))
+                for i in range(4, 7):
+                    label_list.extend(np.load(os.path.join(root, 'obj_right'+ str(i) + '.npy')))
+
+#    image_list = []
+#    for file in os.listdir(file_dir):
+#        image_list.append(file_dir + file)
+#
+#    image_list = sorted(image_list, key = lambda x: int(re.sub('\D', '', x)))
+#    #print(image_list)
+#    label_list = []
+#    label_list = np.load(label_dir)
+
+    #print(len(label_list))
+#    print(label_list)
+    temp = np.array([image_list, label_list])
+    temp = temp.transpose()
+    np.random.shuffle(temp)
+
+    image_list = list(temp[:, 0])
+    label_list = list(temp[:, 1])
+    #label_list = tf.cast(label_list, tf.int32)
+    label_list = [int(float(i)) for i in label_list]
+
+    #print(label_list)
+
+    return image_list, label_list
+
+
+def batch_generate(image, label, image_W, image_H, batch_size, capacity):
+
+    image = tf.cast(image, tf.string)
+    label = tf.cast(label, tf.int32)
+
+    # make an input queue
+    input_queue = tf.train.slice_input_producer([image, label])
+
+    label = input_queue[1]
+    image_contents = tf.read_file(input_queue[0])
+    image = tf.image.decode_png(image_contents, channels=3)
+    image = tf.image.resize_images(image, [image_W, image_H])
+    image = tf.image.per_image_standardization(image)  
+    image_batch, label_batch = tf.train.batch([image, label],
+                                                batch_size= batch_size,
+                                                num_threads= 64, 
+                                                capacity = capacity)
+
+    label_batch = tf.one_hot(label_batch, 24)
+    label_batch = tf.cast(label_batch, tf.float32)
+    label_batch = tf.reshape(label_batch, [batch_size, 24])
+    image_batch = tf.cast(image_batch, tf.float32)
+
+    return image_batch, label_batch
+