Inital commit

data_preparation.py

@@ -0,0 +1,191 @@
+#!/usr/bin/env python
+# coding: utf-8
+
+# In[ ]:
+
+
+## Importing Libraries
+
+
+# In[1]:
+
+
+import os
+import numpy as np
+import matplotlib.pyplot as plt
+import scipy.io.wavfile
+import math
+import librosa
+import librosa.display
+import IPython.display as ipd
+import cv2
+
+
+# # Extracting Features
+
+# ## Dataset
+
+# The dataset contains 6 singers:
+# 
+# 1. Arijit singh
+# 2. Avril lavigne
+# 3. Ed sheeran
+# 4. Lata mangeshkar
+# 5. Sonu nigam
+# 6. Taylor swift
+# 
+# Each singer folder has 40 songs. Total dataset= 240 songs.
+
+# In[2]:
+
+
+root_dir= os.getcwd()
+root_dir
+
+
+# In[3]:
+
+
+src_dir= root_dir + '/songs_dataset'
+src_dir
+
+
+# In[4]:
+
+
+dest_dir= root_dir + '/split'
+dest_dir
+
+
+# # Extracting Clips
+# 
+# ## 5 clips of 20 sec from each song
+
+# In[5]:
+
+
+folders=os.listdir(src_dir)
+
+
+# In[6]:
+
+
+print(folders)
+folders=os.listdir(dest_dir)
+print(folders)
+#same on src_dir and dest_dir
+
+
+# In[8]:
+
+
+for folder in folders:
+    #folder='taylor_swift'
+    ctr=0
+    songs=os.listdir(src_dir + '/' + folder )
+    for song in songs:
+        rate, X = scipy.io.wavfile.read(src_dir+'/'+folder+'/'+song)
+        length=np.shape(X)[0]/float(rate)
+
+        if length>=160:
+            clips=[]
+
+            clips.append(X[rate*35:rate*55])
+            clips.append(X[rate*55:rate*75])
+            clips.append(X[rate*75:rate*95])
+            clips.append(X[rate*120:rate*140])
+            clips.append(X[rate*140:rate*160])
+
+        for clip in clips:
+            # No augmentation
+            ctr+=1
+
+            clip = scipy.mean(clip, axis=1)
+
+            plt.axis('off')
+            plt.specgram(clip, Fs=rate, cmap=plt.get_cmap("viridis"))
+            #plt.show()
+            print(dest_dir+'/'+folder+'/'+"_clip_"+str(ctr)+".png")
+            plt.axis('off')
+            plt.savefig(dest_dir+'/'+folder+'/'+"_clip_"+str(ctr)+".png", bbox_inches='tight') 
+            plt.close()
+
+            #time_shift (Augmentation)
+            ctr+=1      
+            start_ = int(np.random.uniform(-4800,4800))
+            if start_ >= 0:
+                wav_time_shift = np.r_[clip[start_:], np.random.uniform(-0.001,0.001, start_)]
+            else:
+                wav_time_shift = np.r_[np.random.uniform(-0.001,0.001, -start_), clip[:start_]]
+
+
+            plt.specgram(wav_time_shift, Fs=rate, cmap=plt.get_cmap("viridis"))
+            #plt.show()
+            print(dest_dir+'/'+folder+'/'+"_clip_"+str(ctr)+".png")
+            plt.axis('off')
+            plt.savefig(dest_dir+'/'+folder+'/'+"_clip_"+str(ctr)+".png", bbox_inches='tight') 
+            plt.close()
+
+
+            #speed tuning (Augmentation)
+            ctr+=1   
+            speed_rate = np.random.uniform(0.9,1.1)
+            wav_speed_tune = cv2.resize(clip, (1, int(len(clip) * speed_rate))).squeeze()
+               #     print(wav_speed_tune.shape)
+              #      print('speed rate: %.3f' % speed_rate, '(lower is faster)')
+            if len(wav_speed_tune) < 441000:
+                pad_len = 441000 - len(wav_speed_tune)
+                wav_speed_tune = np.r_[np.random.uniform(-0.001,0.001,int(pad_len/2)),
+                                               wav_speed_tune,
+                                               np.random.uniform(-0.001,0.001,int(np.ceil(pad_len/2)))]
+            else: 
+                cut_len = len(wav_speed_tune) - 441000
+                wav_speed_tune = wav_speed_tune[int(cut_len/2):int(cut_len/2)+441000]
+                #    print('wav length: ', wav_speed_tune.shape[0])
+
+
+            plt.axis('off')
+            plt.specgram(wav_speed_tune, Fs=rate, cmap=plt.get_cmap("viridis"))
+            #plt.show()
+            print(dest_dir+'/'+folder+'/'+"_clip_"+str(ctr)+".png")
+            plt.axis('off')
+            plt.savefig(dest_dir+'/'+folder+'/'+"_clip_"+str(ctr)+".png", bbox_inches='tight') 
+            plt.close()
+
+
+
+
+
+
+
+
+
+
+# ## Resizing the spectrograms to 225x150x3
+
+# In[9]:
+
+
+for folder in folders:
+    count=0
+    specs=os.listdir(dest_dir+'/'+folder)
+    for spec in specs:
+        count+=1
+        q = dest_dir + '/' + folder + '/' + spec
+        pixels=cv2.imread(q)
+        pixels=cv2.resize(pixels, (225,150))
+        cv2.imwrite(q, pixels)
+        print(q)
+
+
+# In[9]:
+
+
+count
+
+
+# In[ ]:
+
+
+
+