Merge pull request #87 from aganeshLab41/prod

made changes to nmf and plotting

Author: karllab41

Dockerfile

@@ -14,21 +14,5 @@ WORKDIR /magnoliaWork
 RUN pip install .
 WORKDIR /magnoliaWork/src/demo
 
-RUN mkdir /magnoliaWork/src/demo/app/static/models && \
-    echo "Now getting Deep Clustering code" && \
-    wget "https://www.dropbox.com/s/01f1jrss5leqarq/deep_clustering.ckpt.data-00000-of-00001?dl=0" -O \
-         /magnoliaWork/src/demo/app/static/models/deep_clustering.ckpt.data-00000-of-00001 && \
-    wget "https://www.dropbox.com/s/jeyjsgo4eg1e7uz/deep_clustering.ckpt.index?dl=0" -O \
-         /magnoliaWork/src/demo/app/static/models/deep_clustering.ckpt.index && \
-    wget "https://www.dropbox.com/s/w56skx9k5plz316/deep_clustering.ckpt.meta?dl=0" -O \
-         /magnoliaWork/src/demo/app/static/models/deep_clustering.ckpt.meta && \
-    echo "Now getting Lab41 Model" && \
-    wget "https://www.dropbox.com/s/yrct5fmtxh9xisv/lab41_nonorm-final.ckpt.data-00000-of-00001?dl=0" -O \
-         /magnoliaWork/src/demo/app/static/models/lab41_nonorm-final.ckpt.data-00000-of-00001 && \
-    wget "https://www.dropbox.com/s/oa05j8mrfenmohf/lab41_nonorm-final.ckpt.index?dl=0" -O \
-         /magnoliaWork/src/demo/app/static/models/lab41_nonorm-final.ckpt.index && \
-    wget "https://www.dropbox.com/s/xka5emn0fbfzvq7/lab41_nonorm-final.ckpt.meta?dl=0" -O \
-         /magnoliaWork/src/demo/app/static/models/lab41_nonorm-final.ckpt.meta
 
-RUN mkdir /magnoliaWork/src/demo/app/resources
 

Dockerfile_copy

@@ -0,0 +1,34 @@
+FROM continuumio/anaconda3:4.2.0
+
+RUN conda install -y flask
+RUN pip install --upgrade pip
+RUN apt-get install -y libsndfile1 && \
+    pip install h5py
+
+EXPOSE 5000
+RUN pip install --upgrade matplotlib && pip install https://storage.googleapis.com/tensorflow/linux/cpu/tensorflow-1.1.0-cp35-cp35m-linux_x86_64.whl
+RUN pip install --upgrade scipy
+
+ADD . /magnoliaWork
+WORKDIR /magnoliaWork
+RUN pip install .
+WORKDIR /magnoliaWork/src/demo
+
+RUN mkdir /magnoliaWork/src/demo/app/static/models && \
+    echo "Now getting Deep Clustering code" && \
+    wget "https://www.dropbox.com/s/01f1jrss5leqarq/deep_clustering.ckpt.data-00000-of-00001?dl=0" -O \
+         /magnoliaWork/src/demo/app/static/models/deep_clustering.ckpt.data-00000-of-00001 && \
+    wget "https://www.dropbox.com/s/jeyjsgo4eg1e7uz/deep_clustering.ckpt.index?dl=0" -O \
+         /magnoliaWork/src/demo/app/static/models/deep_clustering.ckpt.index && \
+    wget "https://www.dropbox.com/s/w56skx9k5plz316/deep_clustering.ckpt.meta?dl=0" -O \
+         /magnoliaWork/src/demo/app/static/models/deep_clustering.ckpt.meta && \
+    echo "Now getting Lab41 Model" && \
+    wget "https://www.dropbox.com/s/yrct5fmtxh9xisv/lab41_nonorm-final.ckpt.data-00000-of-00001?dl=0" -O \
+         /magnoliaWork/src/demo/app/static/models/lab41_nonorm-final.ckpt.data-00000-of-00001 && \
+    wget "https://www.dropbox.com/s/oa05j8mrfenmohf/lab41_nonorm-final.ckpt.index?dl=0" -O \
+         /magnoliaWork/src/demo/app/static/models/lab41_nonorm-final.ckpt.index && \
+    wget "https://www.dropbox.com/s/xka5emn0fbfzvq7/lab41_nonorm-final.ckpt.meta?dl=0" -O \
+         /magnoliaWork/src/demo/app/static/models/lab41_nonorm-final.ckpt.meta
+
+RUN mkdir /magnoliaWork/src/demo/app/resources
+

src/demo/app/tflow_functions.py

@@ -9,70 +9,117 @@
 #from .model_functions import separate_sources
 
 import soundfile as sf
-
+import numpy as np
 from magnolia.dnnseparate.L41model import L41Model
 from magnolia.dnnseparate.deep_clustering_model import DeepClusteringModel
 #from .deep_clustering_models import DeepClusteringModel
 #from .clustering_utils import clustering_separate
-from magnolia.utils.clustering_utils import clustering_separate
+from magnolia.utils.clustering_utils import clustering_separate,preprocess_signal
 #from .l41_models import L41Model
-
+from magnolia.factorization.nmf import easy_nmf_separate
+from magnolia.features.preprocessing import undo_preemphasis,make_stft_features
+from magnolia.features.spectral_features import istft
+from magnolia.utils.postprocessing import reconstruct
 '''
 Input: noisy signal path
 Output: list of separated speakers (in numpy form)
 '''
 
 '''def tflow_separate(input_path):
-	
-	rate = 10000
+    
+    rate = 10000
 
-	wav_list = []
+    wav_list = []
 
-	model = Conv1DModel([None,None,251,1],[None,None,251,2],20,600,50)
-	model.load(app.root_path + '/static/models/better_cnn-mask-model.ckpt')
+    model = Conv1DModel([None,None,251,1],[None,None,251,2],20,600,50)
+    model.load(app.root_path + '/static/models/better_cnn-mask-model.ckpt')
 
-	outputs = separate_sources(input_path,model)
-	for row in outputs:
-		wav_list.append(row)
+    outputs = separate_sources(input_path,model)
+    for row in outputs:
+        wav_list.append(row)
 
-	return wav_list	
+    return wav_list 
 '''
 '''
 Input: noisy signal path
 Output: list of separated speakers (in numpy form)
 '''
 def deep_cluster_separate(input_path):
 
-	
-	wav_list = []
-
-	input_signal,sample_rate = sf.read(input_path)
+    
+    wav_list = []
 
-	model  = DeepClusteringModel()
-	model.load(app.root_path + '/static/models/deep_clustering.ckpt')
+    input_signal,sample_rate = sf.read(input_path)
 
-	outputs = clustering_separate(input_signal,sample_rate,model,2)
-	for row in outputs:
-		wav_list.append(row)
+    model  = DeepClusteringModel()
+    model.load(app.root_path + '/static/models/deep_clustering.ckpt')
 
-	return wav_list	
+    outputs = clustering_separate(input_signal,sample_rate,model,2)
+    for row in outputs:
+        wav_list.append(row)
 
+    return wav_list 
 
 
 '''
 Input: noisy signal path
 Output: list of separated speakers (in numpy form)
 '''
 def l41_separate(input_path):
-	wav_list = []
+    wav_list = []
+
+    input_signal,sample_rate = sf.read(input_path)
+
+    model  = L41Model(nonlinearity='tanh', normalize=False)
+    model.load(app.root_path + '/static/models/lab41_nonorm-final.ckpt')
+
+    outputs = clustering_separate(input_signal,sample_rate,model,2)
+    for row in outputs:
+        wav_list.append(row)
+
+    return wav_list 
+
+def nmf_sep(input_path):
+
+    wav_list = []
+
+    input_signal,sample_rate = sf.read(input_path)
+
+    # Preprocess the signal into an input feature
+    #spectrogram, X_in = preprocess_signal(input_signal, sample_rate)
+    spectrogram = make_stft_features(input_signal,sample_rate)
+
+    print("Shape of spect", spectrogram.shape)
+
+    easy_sep = easy_nmf_separate(spectrogram)
+    print("easy sep", easy_sep[:10])
+
+
+    separated_speakers = np.square(easy_sep)
+    phases = np.unwrap(np.angle(spectrogram))
+
+    print("Shape of sep speakers", separated_speakers.shape)
+
+    # Invert the STFT to recover the output waveforms, remembering to undo the
+    # preemphasis
+    print("Lil phase's", phases[:10])
+    '''waveforms = []
+    for i in range(2):
+        waveform = istft(separated_speakers[i]*np.exp(phases*1.0j), 1e4, None, 0.0256, two_sided=False,fft_size=512)
+        unemphasized = undo_preemphasis(waveform)
+        waveforms.append(unemphasized)
+
+    sources = np.stack(waveforms)
 
-	input_signal,sample_rate = sf.read(input_path)
+    for row in sources:
+        wav_list.append(row)
+    '''
+    for i in range(2):
+        recon = reconstruct(separated_speakers[i],phases,fs=10000, window_size=None, step_size=0.0256, square=True, preemphasis=0)
+        print("Type of stuffs",type(recon))
+        print("Recon shape",recon.shape)    
+        wav_list.append(recon)
 
-	model  = L41Model(nonlinearity='tanh', normalize=False)
-	model.load(app.root_path + '/static/models/lab41_nonorm-final.ckpt')
+    return wav_list 
 
-	outputs = clustering_separate(input_signal,sample_rate,model,2)
-	for row in outputs:
-		wav_list.append(row)
 
-	return wav_list	

src/demo/app/views.py

@@ -13,7 +13,8 @@
 import pylab
 import collections
 import os
-from .tflow_functions import deep_cluster_separate,l41_separate
+from .tflow_functions import deep_cluster_separate,l41_separate,nmf_sep
+from magnolia.utils.clustering_utils import clustering_separate,preprocess_signal
 import soundfile as sf
 #from .keras_functions import keras_separate
 #from .keras_functions import keras_spec
@@ -60,17 +61,22 @@ def index():
 
         state['wav_list'][:] = [] 
 
-        if(input_signal_filename == 'mixed_signal'):
-            state['wav_list'].append('b_pit_mf_source_1.wav')
-            state['wav_list'].append('b_pit_mf_source_0.wav')
+        #if(input_signal_filename == 'mixed_signal'):
+        #    state['wav_list'].append('b_pit_mf_source_1.wav')
+        #    state['wav_list'].append('b_pit_mf_source_0.wav')
 
 
     elif request.method == 'POST' and request.form['btn'] == 'NMF' and state['input_signal_url'] != None:
         state['wav_list'][:] = [] 
 
-        if(input_signal_filename == 'mixed_signal'):
-            state['wav_list'].append('nmf_mf_source_0.wav')
-            state['wav_list'].append('nmf_mf_source_1.wav')
+        signals = nmf_sep(project_root + state['input_signal_url'])
+        for index,speaker in enumerate(signals): 
+            sf.write(project_root + '/resources/' + input_signal_filename + 'nmfsplit'+str(index)+'.wav',speaker/speaker.max(),10000) 
+            state['wav_list'].append(input_signal_filename + 'nmfsplit'+str(index)+'.wav') 
+
+        #if(input_signal_filename == 'mixed_signal'):
+        #    state['wav_list'].append('nmf_mf_source_0.wav')
+        #    state['wav_list'].append('nmf_mf_source_1.wav')
 
 
     return render_template('index.html',
@@ -107,9 +113,10 @@ def allowed_file(filename):
         #Plot spectogram with uploaded input file
         input_signal_filename = os.path.splitext(os.path.basename(state['input_signal_url']))[0] 
 
-        f,t,Sxx = calc_spec(project_root + state['input_signal_url'])
+        #f,t,Sxx = calc_spec(project_root + state['input_signal_url'])
 
-        plot_spectogram(f,t,Sxx,project_root + '/resources/spec_'+ input_signal_filename + '.png')
+        #plot_spectogram(f,t,Sxx,project_root + '/resources/spec_'+ input_signal_filename + '.png')
+        plot_spectogram(project_root + state['input_signal_url'],project_root + '/resources/spec_'+ input_signal_filename + '.png')
 
         state['spec_file'] = 'spec_' + input_signal_filename + '.png'
 
@@ -127,8 +134,13 @@ def resources(file_name):
     app.logger.info('In the resources')
     return send_file(project_root + '/resources/'+ file_name)
 
+def plot_spectogram(input_path,output_path):
+    sample_rate,signal = wav.read(input_path)
+    spectrogram, _ = preprocess_signal(signal, sample_rate)
+    plt.matshow(np.sqrt(np.abs(spectrogram)).T, origin='lower', cmap='bone')
+    plt.savefig(output_path)
 
-def plot_spectogram(f,t,Sxx,file_path):
+'''def plot_spectogram(f,t,Sxx,file_path):
     plt.clf()
     plt.pcolormesh(t, f, Sxx, cmap='bone_r')
     plt.ylabel('Frequency [Hz]')
@@ -147,6 +159,6 @@ def calc_spec(signal_path):
 
     fs,noisy_signal = wav.read(signal_path)
 
-    f,t,Sxx = signal.spectrogram(noisy_signal,fs)
+    f,t,Sxx = signal.spectrogram(noisy_signal,fs,nperseg=nfft,noverlap=256)
     return [f,t,np.log(Sxx)]
-
+'''

src/utils/postprocessing.py

@@ -20,7 +20,7 @@ def reconstruct(spec_mag, spec_full, fs, window_size, step_size, square=False,
         preemphasis: if not None, float: undo preemphasis at this level
     '''
     mag = np.absolute(spec_mag)
-    if spec_phase is None:
+    if spec_full is None:
         phase = np.random.randn(*spec_mag.shape)
     else:
         phase = np.exp(1.0j*np.unwrap(np.angle(spec_full)))

src/utils/tf_utils.py

@@ -17,7 +17,7 @@ def scope_decorator(function):
     @functools.wraps(function)
     def decorator(self):
         if not hasattr(self,attribute):
-            with tf.device("/gpu:0"):
+            with tf.device("/cpu:0"):
                 with tf.variable_scope(name):
                     setattr(self,attribute,function(self))
         return getattr(self,attribute)