Merge pull request #26 from 1eedaegon/master

update

PROJECT2/audio_conv2d_features.py

@@ -0,0 +1,169 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Apr 29 05:53:22 2018
+@author: itwill03
+"""
+#전처리 code : extract_features.py
+
+import numpy as np
+import os
+import glob
+import tensorflow as tf
+import pandas as pd
+
+tf.set_random_seed(777) 
+
+train_info = pd.read_csv("C:/data/sound/train.csv",delimiter=',')
+train_data = np.genfromtxt("C:/data/sound/feature_train.csv", delimiter=',')
+
+
+#label set : 라벨 값을 0~41로 바꿈
+
+def labels2Num(labels):
+    df_label = pd.DataFrame(labels)
+    l = train_info['label'].unique()
+    for i in range(len(l)):
+        df_label[df_label==l[i]] = i
+    return df_label
+
+labels = train_info['label']
+df_label = labels2Num(labels)    
+
+#train data set    
+train_data = pd.DataFrame(train_data)
+train_data['label']=df_label
+train_data = train_data.astype(np.float32)
+
+#훈련세트, validation세트 나누기
+from sklearn.model_selection import train_test_split
+train_set, validate_set = train_test_split(train_data, test_size = 0.3)
+trainData = train_set.values[:,0:193]  
+trainLabel = train_set.values[:,-1]
+validateData = validate_set.values[:,0:193]
+validataLabel = validate_set.values[:,-1]
+
+
+# 텐서플로우 모델 생성
+tf.reset_default_graph()
+
+n_dim = 193
+n_classes = 41
+training_epochs = 100
+learning_rate = 0.001
+batch_size = 100
+steps_for_validate = 5
+keep_prob = tf.placeholder(tf.float32)
+
+X = tf.placeholder(tf.float32, [None, n_dim])
+Y = tf.placeholder(tf.int32, [None, 1])
+Y_onehot=tf.reshape(tf.one_hot(Y, 41), [-1, 41])
+p_keep_conv = tf.placeholder(tf.float32, name='p_keep_conv')
+p_keep_hidden = tf.placeholder(tf.float32, name='p_keep_hidden')
+
+#convolution layer 1
+c1 = tf.layers.conv2d(tf.reshape(X, [-1, 1, n_dim, 1]), 32, kernel_size=[2, 1], strides=(1, 1), padding='same', 
+                      activation=tf.nn.elu, name="c1")  
+n1 = tf.layers.batch_normalization(c1)  
+p1 = tf.layers.max_pooling2d(inputs=n1, pool_size=[1, 2], strides=2) 
+p1 = tf.nn.dropout(p1, p_keep_conv)
+
+#convolution layer 2
+    # shape=(?, 1, 96, 32)
+c2 = tf.layers.conv2d(p1, 64, kernel_size=[2, 1], strides=(1, 1), padding='same', 
+                      activation=tf.nn.elu, name="c2")
+n2 = tf.layers.batch_normalization(c2)  
+p2 = tf.layers.max_pooling2d(inputs=n2, pool_size=[1, 2], strides=2) 
+p2 = tf.nn.dropout(p2, p_keep_conv)
+
+#convolution layer 3
+    # shape=(?, 1, 48, 64)
+c3 = tf.layers.conv2d(p2, 128, kernel_size=[2, 1], strides=(1, 1), padding='same', 
+                      activation=tf.nn.elu, name="c3")
+n3 = tf.layers.batch_normalization(c3)  
+p3 = tf.layers.max_pooling2d(inputs=n3, pool_size=[1, 2], strides=2) 
+p3 = tf.nn.dropout(p3, p_keep_conv)
+
+
+# shape=(?, 1, 24, 128)
+L4_flat = tf.reshape(p3, shape=[-1, 1*24*128]) 
+W1 = tf.get_variable("W1", shape=[1*24*128, 624], initializer=tf.contrib.layers.xavier_initializer())
+L5 = tf.nn.relu(tf.matmul(L4_flat, W1))
+n4 = tf.layers.batch_normalization(L5)
+L5 = tf.nn.dropout(n4, p_keep_hidden)
+
+#shape=(?, 624)
+W2 = tf.get_variable("W2", shape=[624,41],initializer=tf.contrib.layers.xavier_initializer())
+b = tf.Variable(tf.random_normal([41]))
+logits = tf.matmul(L5, W2) + b
+
+# define cost/loss & optimizer
+cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits,labels= Y_onehot))
+optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost) # 아담버젼
+predict_op = tf.argmax(logits, 1, name="pred")
+
+# initialize
+sess = tf.Session()
+sess.run(tf.global_variables_initializer())
+saver = tf.train.Saver()
+
+
+# train my model
+
+print('Learning started. It takes sometime.')
+for epoch in range(training_epochs):
+    avg_cost = 0
+    total_batch = int(len(trainData) / batch_size)
+    for i in range(total_batch):
+        batch_xs = trainData[i*batch_size:(i+1)*batch_size]
+        batch_ys = trainLabel[i*batch_size:(i+1)*batch_size].reshape(-1, 1)
+        feed_dict = {X: batch_xs, Y: batch_ys, p_keep_conv: .7, p_keep_hidden: .5}
+        c, _ = sess.run([cost, optimizer], feed_dict=feed_dict)
+        avg_cost += c / total_batch
+    print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.9f}'.format(avg_cost))
+    if epoch % steps_for_validate == steps_for_validate-1:
+        correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(Y_onehot, 1))
+        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
+        print('Accuracy:', sess.run(accuracy, feed_dict={
+                X: validateData, Y: validataLabel.reshape(-1, 1), p_keep_conv: 1, p_keep_hidden: 1}))       
+print('Finished!')
+
+tf.reset_default_graph() 
+"""
+#1
+training_epochs = 500
+learning_rate = 0.001
+sigmoid
+conv2d : kenel size = 1,2 , strides=(1, 1)
+pool_size=[1, 2], strides=2
+flat = 1000
+adam
+cost = 1.203582104
+ccuracy: 0.476073
+
+#2  'sigmoid -> relu'
+training_epochs = 500
+learning_rate = 0.001
+relu
+conv2d kenel size = 1,2
+flat = 1000
+adam
+cost = 0.021049510
+Accuracy: 0.626319
+
+#3 'kanel size = 1,2 -> 1,5'
+training_epochs = 500
+learning_rate = 0.001
+relu
+conv2d kenel size = 1,5
+flat = 1000
+adam
+cost = 0.026283483
+Accuracy: 0.626671
+
+#4 relu -> elu
+Accuracy: 0.64145
+#5 conv2d layer +1, drop out
+cost = 0.300675828
+Accuracy: 0.632301
+"""

PROJECT2/audio_conv2d_mel.py

@@ -0,0 +1,187 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+"""
+Created on Sun Apr 29 05:53:22 2018
+@author: itwill03
+"""
+
+# 전처리 코드 : mel_extract_128.py
+
+
+import numpy as np
+import os
+import glob
+import tensorflow as tf
+import pandas as pd
+
+tf.set_random_seed(777) 
+
+train_info = pd.read_csv("/home/itwill03/sound/train.csv",delimiter=',')
+train_data = np.genfromtxt("/home/itwill03/sound/mel_train2.csv", delimiter=',')
+#train_info = pd.read_csv("C:\data\sound/train.csv",delimiter=',')
+#train_data = np.genfromtxt("C:\data\sound\mel_train2.csv", delimiter=',')
+
+
+#labels to number : 라벨은 0~41의 숫자로 전환
+def labels2Num(labels):
+    df_label = pd.DataFrame(labels)
+    l = train_info['label'].unique()
+    for i in range(len(l)):
+        df_label[df_label==l[i]] = i
+    return df_label
+
+labels = train_info['label']
+df_label = labels2Num(labels)    
+
+#train data set    
+train_data = pd.DataFrame(train_data)
+train_data['label']=df_label
+train_data = train_data.astype(np.float32)
+
+#np.savetxt("c:/data/sound/train_data.csv",train_data, delimiter=",")
+
+#훈련세트, validation세트 나누기
+from sklearn.model_selection import train_test_split
+train_set, validate_set = train_test_split(train_data, test_size = 0.3)
+trainData = train_set.values[:,0:16384]  
+trainLabel = train_set.values[:,-1]
+validateData = validate_set.values[:,0:16384]
+validataLabel = validate_set.values[:,-1]
+
+# graph reset
+tf.reset_default_graph()
+
+# 텐서플로우 모델 생성
+n_dim = 16384
+n_classes = 41
+training_epochs = 700
+learning_rate = 0.0004
+batch_size = 100
+steps_for_validate = 5
+keep_prob = tf.placeholder(tf.float32)
+
+X = tf.placeholder(tf.float32, [None, n_dim])
+Y = tf.placeholder(tf.int32, [None, 1])
+Y_onehot=tf.reshape(tf.one_hot(Y, 41), [-1, 41])
+p_keep_conv = tf.placeholder(tf.float32, name='p_keep_conv')
+p_keep_hidden = tf.placeholder(tf.float32, name='p_keep_hidden')
+
+
+#convolution layer1
+    # img shape = (?, 128, 128, 1)
+c1 = tf.layers.conv2d(tf.reshape(X, [-1, 128, 128, 1]), 32, kernel_size=[5, 3], strides=(1, 1), 
+                      padding='same', activation=tf.nn.relu, name="c1")
+n1 = tf.layers.batch_normalization(c1)  
+p1 = tf.layers.max_pooling2d(inputs=n1, pool_size=[4, 4], strides=4) 
+p1 = tf.nn.dropout(p1, p_keep_conv)
+
+#convolution layer2
+    # img shape = (?, 32, 32, 32)
+c2 = tf.layers.conv2d(tf.reshape(p1, [-1, 32, 32, 32]), 64, kernel_size=[5, 3], strides=(1, 1), 
+                      padding='same', activation=tf.nn.relu, name="c2")
+n2 = tf.layers.batch_normalization(c2) 
+p2 = tf.layers.max_pooling2d(inputs=n2, pool_size=[3, 3], strides=2) #shape = [?, 1, 48, 100]
+p2 = tf.nn.dropout(p2, p_keep_conv)
+
+#convolution layer3
+    # img shape = (?, 15, 15, 64)
+c3 = tf.layers.conv2d(tf.reshape(p2, [-1, 15, 15, 64]), 128, kernel_size=[5, 3], strides=(1, 1), 
+                      padding='same', activation=tf.nn.relu, name="c3")
+n3 = tf.layers.batch_normalization(c3) 
+p3 = tf.layers.max_pooling2d(inputs=n3, pool_size=[3, 3], strides=2) #shape = [?, 1, 24, 200]
+p3 = tf.nn.dropout(p3, p_keep_conv)
+
+#flating layer, hidenlayer  
+L4_flat = tf.reshape(p3, shape=[-1, 7*7*128]) 
+W1 = tf.get_variable("W1", shape=[7*7*128, 640], initializer=tf.contrib.layers.xavier_initializer())
+L5 = tf.nn.relu(tf.matmul(L4_flat, W1))
+n5 = tf.layers.batch_normalization(L5) 
+L5 = tf.nn.dropout(n5, p_keep_hidden)
+
+W2 = tf.get_variable("W2", shape=[640,41],initializer=tf.contrib.layers.xavier_initializer())
+b = tf.Variable(tf.random_normal([41]))
+logits = tf.matmul(L5, W2) + b
+
+# define cost/loss & optimizer
+cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits,labels= Y_onehot))
+optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost) # 아담버젼
+predict_op = tf.argmax(logits, 1, name="pred")
+
+# initialize
+sess = tf.Session()
+sess.run(tf.global_variables_initializer())
+saver = tf.train.Saver()
+
+
+# train my model
+print('Learning started. It takes sometime.')
+for epoch in range(training_epochs):
+    avg_cost = 0
+    total_batch = int(len(trainData) / batch_size)
+    for i in range(total_batch):
+        batch_xs = trainData[i*batch_size:(i+1)*batch_size]
+        batch_ys = trainLabel[i*batch_size:(i+1)*batch_size].reshape(-1, 1)
+        feed_dict = {X: batch_xs, Y: batch_ys, p_keep_conv: .7, p_keep_hidden: .5}
+        c, _ = sess.run([cost, optimizer], feed_dict=feed_dict)
+        avg_cost += c / total_batch
+    print('Epoch:', '%04d' % (epoch + 1), 'cost =', '{:.9f}'.format(avg_cost))
+    if epoch % steps_for_validate == steps_for_validate-1:
+        correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(Y_onehot, 1))
+        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
+        x=np.random.choice(validataLabel.shape[0], 500, replace=False)
+        print('Accuracy:', sess.run(accuracy, feed_dict={
+            X: validateData[x], Y: validataLabel[x].reshape(-1, 1), p_keep_conv: 1, p_keep_hidden: 1}))
+        #save_path = saver.save(sess, '/home/paperspace/Downloads/optx/optx')
+print('Finished!')
+
+"""
+#1.
+training_epochs = 700
+learning_rate = 0.0004
+activation = elu
+conv = kernel_size=[3, 3], strides=(1, 1)
+       kernel_size=[3, 3], strides=(1, 1)
+       kernel_size=[3, 3], strides=(1, 1)
+pool = pool_size=[2, 2], strides=2
+        pool_size=[2, 2], strides=2
+        pool_size=[2, 2], strides=2
+flat = 16*16*128 -> 64
+p_keep_conv: .7, p_keep_hidden: .5
+cost : 0.06~0.08에서 계속 떨어지는중
+Accuracy: 59~63% 
+전처리시 정규화 고려.. librosa.util.normalize
+윈도우 14*2 고려..
+러닝레이트 조절
+
+#2.
+training_epochs = 700
+learning_rate = 0.0001
+activation = relu
+conv = kernel_size=[5, 3], strides=(1, 1)
+        kernel_size=[5, 3], strides=(1, 1)
+        kernel_size=[5, 3], strides=(1, 1)
+pool = pool_size=[4, 4], strides=4
+        pool_size=[3, 3], strides=2
+        pool_size=[3, 3], strides=2
+flat = 7*7*128, 640
+p_keep_conv: .7, p_keep_hidden: .5
+cost = 0.122217526
+Accuracy: 60~67% 
+
+#3.
+training_epochs = 700
+learning_rate = 0.0004
+activation = relu
+conv = kernel_size=[5, 3], strides=(1, 1)
+	kernel_size=[5, 3], strides=(1, 1)
+	kernel_size=[5, 3], strides=(1, 1)
+pool = pool_size=[12, 4], strides=4
+	pool_size=[12, 4], strides=2
+	pool_size=[2, 2], strides=2
+flat = 5*7*128, 420
+p_keep_conv: .8, p_keep_hidden: .7
+Accuracy: 57~64%
+
+
+
+"""