estimator inference files

ArspecExtract.py

@@ -0,0 +1,246 @@
+from __future__ import absolute_import
+from __future__ import print_function
+import numpy as np
+import wave
+import os
+import math
+from scipy.fftpack.realtransforms import dct
+from copy import deepcopy
+from scipy.fftpack import fft, ifft
+from scikits.talkbox.linpred import lpc
+np.random.seed(1337)
+epsilon = 0.0000000001
+
+
+def build_data(wav, begin=None, end=None):
+    wav_in_file = wave.Wave_read(wav)
+    wav_in_num_samples = wav_in_file.getnframes()
+    N = wav_in_file.getnframes()
+    dstr = wav_in_file.readframes(N)
+    data = np.fromstring(dstr, np.int16)
+    return data
+
+
+def periodogram(x, nfft=None, fs=1):
+    """Compute the periodogram of the given signal, with the given fft size.
+
+    Parameters
+    ----------
+    x : array-like
+        input signal
+    nfft : int
+        size of the fft to compute the periodogram. If None (default), the
+        length of the signal is used. if nfft > n, the signal is 0 padded.
+    fs : float
+        Sampling rate. By default, is 1 (normalized frequency. e.g. 0.5 is the
+        Nyquist limit).
+
+    Returns
+    -------
+    pxx : array-like
+        The psd estimate.
+    fgrid : array-like
+        Frequency grid over which the periodogram was estimated.
+
+    Examples
+    --------
+    Generate a signal with two sinusoids, and compute its periodogram:
+
+    >>> fs = 1000
+    >>> x = np.sin(2 * np.pi  * 0.1 * fs * np.linspace(0, 0.5, 0.5*fs))
+    >>> x += np.sin(2 * np.pi  * 0.2 * fs * np.linspace(0, 0.5, 0.5*fs))
+    >>> px, fx = periodogram(x, 512, fs)
+
+    Notes
+    -----
+    Only real signals supported for now.
+
+    Returns the one-sided version of the periodogram.
+
+    Discrepency with matlab: matlab compute the psd in unit of power / radian /
+    sample, and we compute the psd in unit of power / sample: to get the same
+    result as matlab, just multiply the result from talkbox by 2pi"""
+    x = np.atleast_1d(x)
+    n = x.size
+
+    if x.ndim > 1:
+        raise ValueError("Only rank 1 input supported for now.")
+    if not np.isrealobj(x):
+        raise ValueError("Only real input supported for now.")
+    if not nfft:
+        nfft = n
+    if nfft < n:
+        raise ValueError("nfft < signal size not supported yet")
+
+    pxx = np.abs(fft(x, nfft)) ** 2
+    if nfft % 2 == 0:
+        pn = nfft / 2 + 1
+    else:
+        pn = (nfft + 1) / 2
+
+    fgrid = np.linspace(0, fs * 0.5, pn)
+    return pxx[:pn] / (n * fs), fgrid
+
+
+def arspec(x, order, nfft=None, fs=1):
+    """Compute the spectral density using an AR model.
+
+    An AR model of the signal is estimated through the Yule-Walker equations;
+    the estimated AR coefficient are then used to compute the spectrum, which
+    can be computed explicitely for AR models.
+
+    Parameters
+    ----------
+    x : array-like
+        input signal
+    order : int
+        Order of the LPC computation.
+    nfft : int
+        size of the fft to compute the periodogram. If None (default), the
+        length of the signal is used. if nfft > n, the signal is 0 padded.
+    fs : float
+        Sampling rate. By default, is 1 (normalized frequency. e.g. 0.5 is the
+        Nyquist limit).
+
+    Returns
+    -------
+    pxx : array-like
+        The psd estimate.
+    fgrid : array-like
+        Frequency grid over which the periodogram was estimated.
+    """
+
+    x = np.atleast_1d(x)
+    n = x.size
+
+    if x.ndim > 1:
+        raise ValueError("Only rank 1 input supported for now.")
+    if not np.isrealobj(x):
+        raise ValueError("Only real input supported for now.")
+    if not nfft:
+        nfft = n
+    a, e, k = lpc(x, order)
+
+    # This is not enough to deal correctly with even/odd size
+    if nfft % 2 == 0:
+        pn = nfft / 2 + 1
+    else:
+        pn = (nfft + 1) / 2
+
+    px = 1 / np.fft.fft(a, nfft)[:pn]
+    pxx = np.real(np.conj(px) * px)
+    pxx /= fs / e
+    fx = np.linspace(0, fs * 0.5, pxx.size)
+    return pxx, fx
+
+
+def arspecs(input_wav, order, Atal=False):
+    epsilon = 0.0000000001
+    data = input_wav
+    ar = []
+    ars = arspec(data, order, 4096)
+    for k, l in zip(ars[0], ars[1]):
+        ar.append(math.log(math.sqrt((k ** 2) + (l ** 2))))
+    for val in range(0, len(ar)):
+        if ar[val] == 0.0:
+            ar[val] = deepcopy(epsilon)
+    mspec1 = np.log10(ar)
+    # Use the DCT to 'compress' the coefficients (spectrum -> cepstrum domain)
+    ar = dct(mspec1, type=2, norm='ortho', axis=-1)
+    return ar[:30]
+
+
+def specPS(input_wav, pitch):
+    N = len(input_wav)
+    samps = N / pitch
+    if samps == 0:
+        samps = 1
+    frames = N / samps
+    data = input_wav[0:frames]
+    specs = periodogram(data, nfft=4096)
+    for i in range(1, int(samps)):
+        data = input_wav[frames * i:frames * (i + 1)]
+        peri = periodogram(data, nfft=4096)
+        for sp in range(len(peri[0])):
+            specs[0][sp] += peri[0][sp]
+    for s in range(len(specs[0])):
+        specs[0][s] /= float(samps)
+    peri = []
+    for k, l in zip(specs[0], specs[1]):
+        if k == 0 and l == 0:
+            peri.append(epsilon)
+        else:
+            peri.append(math.log(math.sqrt((k ** 2) + (l ** 2))))
+    # Filter the spectrum through the triangle filterbank
+    mspec = np.log10(peri)
+    # Use the DCT to 'compress' the coefficients (spectrum -> cepstrum domain)
+    ceps = dct(mspec, type=2, norm='ortho', axis=-1)
+    return ceps[:50]
+
+
+def build_single_feature_row(data):
+    lpcs = [8, 9, 10, 11, 12, 13, 14, 15, 16, 17]
+    arr = []
+    periodo = specPS(data, 50)
+    arr.extend(periodo)
+    for j in lpcs:
+        ars = arspecs(data, j)
+        arr.extend(ars)
+    for i in range(len(arr)):
+        if np.isnan(np.float(arr[i])):
+            arr[i] = 0.0
+    return arr
+
+
+def get_y():
+    data = np.load('timit.npy')
+    timit = []
+    for row in data:
+        y = open('Y/' + str(row[0]).replace("timit", "VTRFormants") + ".y").readline().split()
+        arr = []
+        arr.append(float(y[0]))
+        arr.append(float(y[1]))
+        arr.append(float(y[2]))
+        arr.append(float(y[3]))
+        arr.extend(row)
+        timit.append(arr)
+    nump = np.asarray(timit)
+    np.save('timit_train_arspec',nump)
+    return
+
+
+def build_timit_data():
+    arcep_mat = []
+    path = 'X_test/'
+    for file in [f for f in os.listdir(path) if f.endswith('.wav')]:
+        name = file.replace('.wav', '')
+        y = open('Y_test' + '/' + str(name).replace("timit", "VTRFormants") + ".y").readline().split()
+        X = build_data(path + file)
+        arr = [name]
+        arr.append(float(y[0]))
+        arr.append(float(y[1]))
+        arr.append(float(y[2]))
+        arr.append(float(y[3]))
+        arr.extend(build_single_feature_row(X))
+        arcep_mat.append(arr)
+    nump = np.asarray(arcep_mat)
+    np.save('timitTest',nump)
+
+    arcep_mat = []
+    path = 'X/'
+    for file in [f for f in os.listdir(path) if f.endswith('.wav')]:
+        name = file.replace('.wav', '')
+        y = open('Y/' + str(name).replace("timit", "VTRFormants") + ".y").readline().split()
+        X = build_data(path + file)
+        arr = [name]
+        arr.append(float(y[0]))
+        arr.append(float(y[1]))
+        arr.append(float(y[2]))
+        arr.append(float(y[3]))
+        arr.extend(build_single_feature_row(X))
+        arcep_mat.append(arr)
+    nump = np.asarray(arcep_mat)
+    np.save('timitTrain',nump)
+    return
+
+build_timit_data()

LPC_Estimate_Class.py

@@ -0,0 +1,135 @@
+from __future__ import print_function, division
+import torch
+import torch.nn as nn
+from torch.autograd import Variable
+import torch.nn.functional as F
+from torch import optim
+import numpy as np
+
+train_data = np.load("timitTrain.npy")
+test_data = np.load("timitTest.npy")
+Xtrain = train_data[:,5:].astype(np.float32)
+Ytrain = train_data[:,1:5].astype(np.float32)
+Xtest = test_data[:,5:].astype(np.float32)
+Ytest = test_data[:,1:5].astype(np.float32)
+
+use_cuda = torch.cuda.is_available()
+device = torch.device("cuda" if use_cuda else "cpu")
+_, D = Xtrain.shape
+K = len(Ytrain)
+
+print(D, K)
+
+class Net(nn.Module):
+
+    def __init__(self):
+        super(Net, self).__init__()
+        self.Dense1 = nn.Linear(D, 1024)
+        self.Dense2 = nn.Linear(1024, 512)
+        self.Dense3 = nn.Linear(512, 256)
+        self.out = nn.Linear(256, 4)
+
+    def forward(self, x):
+        x = torch.sigmoid(self.Dense1(x))
+        x = torch.sigmoid(self.Dense2(x))
+        x = torch.sigmoid(self.Dense3(x))
+        return self.out(x)
+
+
+
+
+loss = nn.L1Loss()
+
+def train(model, loss, optimizer, inputs, labels):
+    inputs = Variable(inputs.to(device))
+    labels = Variable(labels.to(device))
+    optimizer.zero_grad()
+
+    logits = model.forward(inputs)
+    output = loss.forward(logits, labels)
+    output.backward()
+    optimizer.step()
+
+    return output.item()
+
+
+def predict(model, inputs):
+    inputs = Variable(inputs)
+    logits = model.forward(inputs.to(device))
+    return logits.data.cpu().numpy()
+
+
+torch.manual_seed(0)
+
+Xtrain = torch.from_numpy(Xtrain).float().to(device)
+Ytrain = torch.from_numpy(Ytrain).float().to(device)
+Xtest = torch.from_numpy(Xtest).float().to(device)
+Ytest = torch.from_numpy(Ytest).float().to(device)
+
+model = Net().to(device)
+
+
+optimizer = optim.Adagrad(model.parameters(), lr=0.01)
+
+epochs = 80
+batchSize = 20
+n_batches = Xtrain.size()[0]
+
+costs = []
+test_accuracies = []
+print("Starting training ")
+for i in range(epochs):
+    cost = 0.0
+    for j in range(n_batches):
+        Xbatch = Xtrain[j*batchSize:(j+1)*batchSize]
+        Ybatch = Ytrain[j*batchSize:(j+1)*batchSize]
+        cost += train(model, loss, optimizer, Xbatch, Ybatch)
+
+    loss1 = 0.0
+    loss2 = 0.0
+    loss3 = 0.0
+    loss4 = 0.0
+    max_1 = 0.0
+    max_2 = 0.0
+    max_3 = 0.0
+    max_4 = 0.0
+    list_1 = []
+    list_2 = []
+    list_3 = []
+    list_4 = []
+    print('predicting...')
+    Ypred = predict(model, Xtest)
+    for k in range(0, len(Ytest)):
+        # print(y_hat[i])
+        l1 = np.abs(float(Ytest[k, 0]) - Ypred[k, 0])
+        l2 = np.abs(float(Ytest[k, 1]) - Ypred[k, 1])
+        l3 = np.abs(float(Ytest[k, 2]) - Ypred[k, 2])
+        l4 = np.abs(float(Ytest[k, 3]) - Ypred[k, 3])
+        list_1.append(l1)
+        list_2.append(l2)
+        list_3.append(l3)
+        list_4.append(l4)
+        max_1 = max(max_1, l1)
+        max_2 = max(max_2, l2)
+        max_3 = max(max_3, l3)
+        max_4 = max(max_4, l4)
+        loss1 += l1
+        loss2 += l2
+        loss3 += l3
+        loss4 += l4
+    loss1 /= len(Ytest)
+    loss2 /= len(Ytest)
+    loss3 /= len(Ytest)
+    loss4 /= len(Ytest)
+    total_loss = loss1 + loss2 + loss3 + loss4
+    total_loss /= 4.0
+    print('median: %.3f %.3f %.3f %.3f' % (np.median(list_1), np.median(list_2), np.median(list_3), np.median(list_4)))
+    print('max loss: %.3f %.3f %.3f %.3f' % (max_1, max_2, max_3, max_4))
+    print('Real test score: %.3f %.3f %.3f %.3f' % (loss1, loss2, loss3, loss4))
+    print("Epoch: %d, acc: %.3f" % (i, total_loss))
+
+    costs.append(cost/n_batches)
+    test_accuracies.append(round(total_loss, 3))
+    torch.save(model.state_dict(), "LPC_NN.pt")
+
+print(test_accuracies)

LPC_Model_inference.py

@@ -0,0 +1,114 @@
+from __future__ import print_function, division
+import torch
+import torch.nn as nn
+from torch.autograd import Variable
+import torch.nn.functional as F
+from torch import optim
+import numpy as np
+
+test_data = np.load("timitTest.npy")
+Xtest = test_data[:,5:].astype(np.float32)
+Ytest = test_data[:,1:5].astype(np.float32)
+
+use_cuda = torch.cuda.is_available()
+device = torch.device("cuda" if use_cuda else "cpu")
+_, D = Xtest.shape
+print(D)
+
+class Net(nn.Module):
+
+    def __init__(self):
+        super(Net, self).__init__()
+        self.Dense1 = nn.Linear(D, 1024)
+        self.Dense2 = nn.Linear(1024, 512)
+        self.Dense3 = nn.Linear(512, 256)
+        self.out = nn.Linear(256, 4)
+
+    def forward(self, x):
+        x = torch.sigmoid(self.Dense1(x))
+        x = torch.sigmoid(self.Dense2(x))
+        x = torch.sigmoid(self.Dense3(x))
+        return self.out(x)
+
+def scaledLoss(output, target):
+	scale = torch.tensor([2.0, 1.0, .5, .1]).to(device)
+	loss = torch.abs(output - target)
+	scaled = loss*scale
+	return torch.mean(scaled)
+
+#loss = nn.L1Loss()
+
+def train(model, optimizer, inputs, labels):
+    inputs = Variable(inputs.to(device))
+    labels = Variable(labels.to(device))
+    optimizer.zero_grad()
+
+    logits = model.forward(inputs)
+    output = scaledLoss(logits, labels)
+    output.backward()
+    optimizer.step()
+
+    return output.item()
+
+
+def predict(model, inputs):
+    inputs = Variable(inputs)
+    logits = model.forward(inputs.to(device))
+    return logits.data.cpu().numpy()
+
+
+torch.manual_seed(0)
+
+Xtest = torch.from_numpy(Xtest).float().to(device)
+Ytest = torch.from_numpy(Ytest).float().to(device)
+
+model = Net().to(device)
+
+
+optimizer = optim.Adagrad(model.parameters(), lr=0.01)
+
+model.load_state_dict(torch.load("LPC_NN_scaledLoss.pt"))
+model.eval()
+loss1 = 0.0
+loss2 = 0.0
+loss3 = 0.0
+loss4 = 0.0
+max_1 = 0.0
+max_2 = 0.0
+max_3 = 0.0
+max_4 = 0.0
+list_1 = []
+list_2 = []
+list_3 = []
+list_4 = []
+print('predicting...')
+Ypred = predict(model, Xtest)
+for k in range(0, len(Ytest)):
+	# print(y_hat[i])
+	l1 = np.abs(float(Ytest[k, 0]) - Ypred[k, 0])
+	l2 = np.abs(float(Ytest[k, 1]) - Ypred[k, 1])
+	l3 = np.abs(float(Ytest[k, 2]) - Ypred[k, 2])
+	l4 = np.abs(float(Ytest[k, 3]) - Ypred[k, 3])
+	list_1.append(l1)
+	list_2.append(l2)
+	list_3.append(l3)
+	list_4.append(l4)
+	max_1 = max(max_1, l1)
+	max_2 = max(max_2, l2)
+	max_3 = max(max_3, l3)
+	max_4 = max(max_4, l4)
+	loss1 += l1
+	loss2 += l2
+	loss3 += l3
+	loss4 += l4
+loss1 /= len(Ytest)
+loss2 /= len(Ytest)
+loss3 /= len(Ytest)
+loss4 /= len(Ytest)
+total_loss = loss1 + loss2 + loss3 + loss4
+total_loss /= 4.0
+print('median: %.3f %.3f %.3f %.3f' % (np.median(list_1), np.median(list_2), np.median(list_3), np.median(list_4)))
+print('max loss: %.3f %.3f %.3f %.3f' % (max_1, max_2, max_3, max_4))
+print('Real test score: %.3f %.3f %.3f %.3f' % (loss1, loss2, loss3, loss4))
+print("acc: %.3f" % (total_loss))
+

Models/models

@@ -0,0 +1 @@
+

README.md

@@ -40,7 +40,7 @@ cd ~/torch; bash install-deps;
 luarocks install rnn
 ```
 The Estimation model can be downloaded here and because of size constraints the Tracking model can be abtained by download from this link 
-[tracking_model.mat] (https://drive.google.com/open?id=0Bxkc5_D0JjpiZWx4eTU1d0hsVXc)
+[tracking_model.dat.gz](https://drive.google.com/open?id=1-BwlbbHykIV52c-SL1ofcppxZ5pTTXai)
 
 ## How to use:
 

VTR_Results.py

@@ -0,0 +1,142 @@
+from keras.models import model_from_json
+import numpy as np
+import csv
+import math
+
+model = model_from_json(open('model.json').read())
+model.load_weights('weights.h5')
+data_dir = ""
+X_test = np.load(data_dir+'VTR_test_X.npy')
+Y = np.load(data_dir+'VTR_test_Y.npy')
+
+names = Y[:, :1]
+Y_test = Y[:,1:]
+predictions = []
+
+loss1 = 0.0
+loss2 = 0.0
+loss3 = 0.0
+loss4 = 0.0
+max_1 = 0.0
+max_2 = 0.0
+max_3 = 0.0
+max_4 = 0.0
+list_1 = []
+list_2 = []
+list_3 = []
+list_4 = []
+male = [0.0, 0.0, 0.0, 0.0, 0.0, [], [], [], []]
+female = [0.0, 0.0, 0.0, 0.0, 0.0, [], [], [], []]
+karma_list = [0, 0.0, 0.0, 0.0, 0.0]
+AVG_list = [0, 0.0, 0.0, 0.0, 0.0]
+
+y_hat = model.predict(X_test)
+for i in range(0,len(Y_test)):
+    l1 = np.abs(float(Y_test[i, 0]) - y_hat[i, 0])
+    l2 = np.abs(float(Y_test[i, 1]) - y_hat[i, 1])
+    l3 = np.abs(float(Y_test[i, 2]) - y_hat[i, 2])
+    l4 = np.abs(float(Y_test[i, 3]) - y_hat[i, 3])
+    pred = [names[i][0], float(Y_test[i, 0]), float(Y_test[i, 1]), float(Y_test[i, 2]), float(Y_test[i, 3])]
+
+    AVG_list[0] += 1
+    AVG_list[1] += float(Y_test[i, 0]) - y_hat[i, 0]
+    AVG_list[2] += float(Y_test[i, 1]) - y_hat[i, 1]
+    AVG_list[3] += float(Y_test[i, 2]) - y_hat[i, 2]
+    AVG_list[4] += float(Y_test[i, 3]) - y_hat[i, 3]
+
+    pred.extend([y_hat[i, 0], y_hat[i, 1], y_hat[i, 2], y_hat[i, 3]])
+
+    if names[i][0].split('_')[3][0] == 'f':
+        female[0] += 1
+        female[1] += l1
+        female[2] += l2
+        female[3] += l3
+        female[4] += l4
+        female[5].append(l1)
+        female[6].append(l2)
+        female[7].append(l3)
+        female[8].append(l4)
+    elif names[i][0].split('_')[3][0] == 'm':
+        male[0] += 1
+        male[1] += l1
+        male[2] += l2
+        male[3] += l3
+        male[4] += l4
+        male[5].append(l1)
+        male[6].append(l2)
+        male[7].append(l3)
+        male[8].append(l4)
+
+    predictions.append(pred)
+
+    list_1.append(l1)
+    list_2.append(l2)
+    list_3.append(l3)
+    list_4.append(l4)
+    max_1 = max(max_1,l1)
+    max_2 = max(max_2,l2)
+    max_3 = max(max_3,l3)
+    max_4 = max(max_4,l4)
+
+    loss1 += l1
+    loss2 += l2
+    loss3 += l3
+    loss4 += l4
+
+    karma_list[0] += 1
+    karma_list[1] += l1 * l1
+    karma_list[2] += l2 * l2
+    karma_list[3] += l3 * l3
+    karma_list[4] += l4 * l4
+loss1 /= len(Y_test)
+loss2 /= len(Y_test)
+loss3 /= len(Y_test)
+loss4 /= len(Y_test)
+total_loss = loss1+loss2+loss3+loss4
+total_loss /= 4.0
+print('standard deviation', round(np.std(list_1)*1000, 2), round(np.std(list_2)*1000, 2), round(np.std(list_3)*1000, 2), round(np.std(list_4)*1000, 2))
+print('median', round(np.median(list_1)*1000, 2), round(np.median(list_2)*1000, 2), round(np.median(list_3)*1000, 2), round(np.median(list_4)*1000, 2))
+print('max loss ', round(max_1*1000, 2), round(max_2*1000, 2), round(max_3*1000, 2), round(max_4*1000, 2))
+print('total loss ', round(total_loss*1000, 2))
+print('Real test score:', round(loss1*1000, 2), round(loss2*1000, 2), round(loss3*1000, 2), round(loss4*1000, 2))
+
+female[1] = round((female[1] / female[0])*1000, 2)
+female[2] = round((female[2] / female[0])*1000, 2)
+female[3] = round((female[3] / female[0])*1000, 2)
+female[4] = round((female[4] / female[0])*1000, 2)
+female[5] = round(np.std(female[5])*1000, 2)
+female[6] = round(np.std(female[6])*1000, 2)
+female[7] = round(np.std(female[7])*1000, 2)
+female[8] = round(np.std(female[8])*1000, 2)
+
+male[1] = round((male[1] / male[0])*1000, 2)
+male[2] = round((male[2] / male[0])*1000, 2)
+male[3] = round((male[3] / male[0])*1000, 2)
+male[4] = round((male[4] / male[0])*1000, 2)
+male[5] = round(np.std(male[5])*1000, 2)
+male[6] = round(np.std(male[6])*1000, 2)
+male[7] = round(np.std(male[7])*1000, 2)
+male[8] = round(np.std(male[8])*1000, 2)
+
+print("male: ", male)
+print("female: ", female)
+
+# karma
+
+karma_list[1] /= karma_list[0]
+karma_list[2] /= karma_list[0]
+karma_list[3] /= karma_list[0]
+karma_list[4] /= karma_list[0]
+print('root mean squared error ', round(math.sqrt(karma_list[1]) * 1000, 2), round(math.sqrt(karma_list[2]) * 1000, 2),
+                            round(math.sqrt(karma_list[3]) * 1000, 2), round(math.sqrt(karma_list[4]) * 1000, 2))
+
+AVG_list[1] /= AVG_list[0]
+AVG_list[2] /= AVG_list[0]
+AVG_list[3] /= AVG_list[0]
+AVG_list[4] /= AVG_list[0]
+print('AVG ', round(AVG_list[1] * 1000, 2), round(AVG_list[2] * 1000, 2), round(AVG_list[3] * 1000, 2), round(AVG_list[4] * 1000, 2))
+
+
+with open("results/VTR.csv", "wb") as f:
+    writer = csv.writer(f)
+    writer.writerows(predictions)

extract_features.py

@@ -26,7 +26,8 @@ def build_data(wav,begin=None,end=None):
     dstr = wav_in_file.readframes(N)
     data = np.fromstring(dstr, np.int16)
     if begin is not None and end is not None:
-        return data[begin*16000:end*16000]
+        #return data[begin*16000:end*16000] #numpy 1.11.0
+        return data[np.int(begin*16000):np.int(end*16000)] #numpy 1.14.0
     X = []
     l = len(data)
     for i in range(0, l-100, 160):

spectrogramEstimate.py

@@ -0,0 +1,141 @@
+from __future__ import print_function, division
+import torch
+import torch.nn as nn
+from torch.autograd import Variable
+import torch.nn.functional as F
+from torch import optim
+import numpy as np
+torch.manual_seed(1)
+
+trainY = np.load("norm_cnn_timit_train_Y.npy")
+testY = np.load("norm_cnn_timit_test_Y.npy")
+Xtrain = np.load("norm_cnn_timit_train_X.npy").astype(np.float32)
+Ytrain = trainY[:,1:5].astype(np.float32)
+Xtest = np.load("norm_cnn_timit_test_X.npy").astype(np.float32)
+Ytest = testY[:,1:5].astype(np.float32)
+
+use_cuda = torch.cuda.is_available()
+device = torch.device("cuda" if use_cuda else "cpu")
+D = Xtrain.shape[1]
+K = len(Ytrain)
+
+print(D, K)
+
+class Net(nn.Module):
+
+    def __init__(self):
+        super(Net, self).__init__()
+        self.Conv1 = nn.Conv2d(in_channels=1, out_channels=96, kernel_size=(3, 3), stride=1, padding=0)
+        self.Conv2 = nn.Conv2d(in_channels=96, out_channels=32, kernel_size=(3, 3), stride=1, padding=0)
+        self.Conv3 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3), stride=1, padding=0)
+        self.Conv4 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(5, 5), stride=1, padding=0)
+        self.Dense5 = nn.Linear(43*38*64, 512)
+        self.out = nn.Linear(512, 4)
+
+    def forward(self, x):
+        in_size = x.size(0)
+        x = F.relu(self.Conv1(x))
+        x = F.relu(self.Conv2(x))
+        x = F.max_pool2d(x, kernel_size=2, stride=1)
+        x = F.relu(self.Conv3(x))
+        x = F.relu(self.Conv4(x))
+        x = F.max_pool2d(x, kernel_size=2, stride=1)
+        #print(in_size)
+        x = x.view(x.size(0), -1)
+        x = F.relu(self.Dense5(x))
+        return self.out(x)
+
+
+def train(model, loss, optimizer, inputs, labels):
+    inputs = Variable(inputs.to(device))
+    labels = Variable(labels.to(device))
+    optimizer.zero_grad()
+
+    logits = model.forward(inputs)
+    output = loss.forward(logits, labels)
+    output.backward()
+    optimizer.step()
+
+    return output.item()
+
+
+def predict(model, inputs):
+    inputs = Variable(inputs)
+    with torch.no_grad():
+        logits = model.forward(inputs.to(device))
+    return logits.data.cpu().numpy()
+
+
+Xtrain = torch.from_numpy(Xtrain).float().to(device)
+Ytrain = torch.from_numpy(Ytrain).float().to(device)
+Xtest = torch.from_numpy(Xtest).float().to(device)
+Ytest = torch.from_numpy(Ytest).float().to(device)
+
+
+model = Net().to(device)
+loss = nn.L1Loss()
+optimizer = optim.Adagrad(model.parameters())
+
+epochs = 80
+batchSize = 32
+n_batches = int(np.floor(Xtrain.size()[0]/batchSize))
+
+costs = []
+test_accuracies = []
+print("Starting training ")
+for i in range(epochs):
+    cost = 0.0
+    for j in range(n_batches):
+        #print(j, '/', n_batches)
+        Xbatch = Xtrain[j*batchSize:(j+1)*batchSize]
+        Ybatch = Ytrain[j*batchSize:(j+1)*batchSize]
+        cost += train(model, loss, optimizer, Xbatch, Ybatch)
+
+    loss1 = 0.0
+    loss2 = 0.0
+    loss3 = 0.0
+    loss4 = 0.0
+    max_1 = 0.0
+    max_2 = 0.0
+    max_3 = 0.0
+    max_4 = 0.0
+    list_1 = []
+    list_2 = []
+    list_3 = []
+    list_4 = []
+    print('predicting...')
+    Ypred = predict(model, Xtest)
+    for k in range(0, len(Ytest)):
+        # print(y_hat[i])
+        l1 = np.abs(float(Ytest[k, 0]) - Ypred[k, 0])
+        l2 = np.abs(float(Ytest[k, 1]) - Ypred[k, 1])
+        l3 = np.abs(float(Ytest[k, 2]) - Ypred[k, 2])
+        l4 = np.abs(float(Ytest[k, 3]) - Ypred[k, 3])
+        list_1.append(l1)
+        list_2.append(l2)
+        list_3.append(l3)
+        list_4.append(l4)
+        max_1 = max(max_1, l1)
+        max_2 = max(max_2, l2)
+        max_3 = max(max_3, l3)
+        max_4 = max(max_4, l4)
+        loss1 += l1
+        loss2 += l2
+        loss3 += l3
+        loss4 += l4
+    loss1 /= len(Ytest)
+    loss2 /= len(Ytest)
+    loss3 /= len(Ytest)
+    loss4 /= len(Ytest)
+    total_loss = loss1 + loss2 + loss3 + loss4
+    total_loss /= 4.0
+    print('median: %.3f %.3f %.3f %.3f' % (np.median(list_1), np.median(list_2), np.median(list_3), np.median(list_4)))
+    print('max loss: %.3f %.3f %.3f %.3f' % (max_1, max_2, max_3, max_4))
+    print('Real test score: %.3f %.3f %.3f %.3f' % (loss1, loss2, loss3, loss4))
+    print("Epoch: %d, acc: %.3f" % (i, total_loss))
+
+    costs.append(cost/n_batches)
+    test_accuracies.append(round(total_loss, 3))
+    torch.save(model.state_dict(), "CNN_estimate.pt")
+
+print(test_accuracies)

spectrogramEstimate_inference.py

@@ -0,0 +1,121 @@
+from __future__ import print_function, division
+import torch
+import torch.nn as nn
+from torch.autograd import Variable
+import torch.nn.functional as F
+from torch import optim
+import numpy as np
+torch.manual_seed(1)
+
+testY = np.load("norm_cnn_timit_test_Y.npy")
+Xtest = np.load("norm_cnn_timit_test_X.npy").astype(np.float32)
+Ytest = testY[:,1:5].astype(np.float32)
+
+use_cuda = torch.cuda.is_available()
+device = torch.device("cuda" if use_cuda else "cpu")
+D = Xtest.shape
+print(D)
+
+print(Xtest.shape[1], len(Ytest))
+
+ 
+class Net(nn.Module):
+
+    def __init__(self):
+        super(Net, self).__init__()
+        self.Conv1 = nn.Conv2d(in_channels=1, out_channels=96, kernel_size=(3, 3), stride=1, padding=0)
+        self.Conv2 = nn.Conv2d(in_channels=96, out_channels=32, kernel_size=(3, 3), stride=1, padding=0)
+        self.Conv3 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=(3, 3), stride=1, padding=0)
+        self.Conv4 = nn.Conv2d(in_channels=64, out_channels=64, kernel_size=(5, 5), stride=1, padding=0)
+        self.Dense5 = nn.Linear(43*38*64, 512)
+        self.out = nn.Linear(512, 4)
+
+    def forward(self, x):
+        in_size = x.size(0)
+        x = F.relu(self.Conv1(x))
+        x = F.relu(self.Conv2(x))
+        x = F.max_pool2d(x, kernel_size=2, stride=1)
+        x = F.relu(self.Conv3(x))
+        x = F.relu(self.Conv4(x))
+        x = F.max_pool2d(x, kernel_size=2, stride=1)
+        #print(in_size)
+        x = x.view(x.size(0), -1)
+        x = F.relu(self.Dense5(x))
+        return self.out(x)
+
+
+def train(model, loss, optimizer, inputs, labels):
+    inputs = Variable(inputs.to(device))
+    labels = Variable(labels.to(device))
+    optimizer.zero_grad()
+
+    logits = model.forward(inputs)
+    output = loss.forward(logits, labels)
+    output.backward()
+    optimizer.step()
+
+    return output.item()
+
+
+def predict(model, inputs):
+    inputs = Variable(inputs)
+    with torch.no_grad():
+        logits = model.forward(inputs.to(device))
+    return logits.data.cpu().numpy()
+
+Xtest = torch.from_numpy(Xtest).float().to(device)
+Ytest = torch.from_numpy(Ytest).float().to(device)
+
+
+model = Net().to(device)
+loss = nn.L1Loss()
+optimizer = optim.Adagrad(model.parameters())
+
+model.load_state_dict(torch.load("CNN_estimate.pt"))
+model.eval()
+loss1 = 0.0
+loss2 = 0.0
+loss3 = 0.0
+loss4 = 0.0
+max_1 = 0.0
+max_2 = 0.0
+max_3 = 0.0
+max_4 = 0.0
+list_1 = []
+list_2 = []
+list_3 = []
+list_4 = []
+print('predicting...')
+Ypred1 = predict(model, Xtest[:1000])
+Ypred2 = predict(model, Xtest[1000:2000])
+Ypred3 = predict(model, Xtest[2000:])
+Ypred = np.concatenate((Ypred1, Ypred2, Ypred3))
+for k in range(0, len(Ytest)):
+	# print(y_hat[i])
+	l1 = np.abs(float(Ytest[k, 0]) - Ypred[k, 0])
+	l2 = np.abs(float(Ytest[k, 1]) - Ypred[k, 1])
+	l3 = np.abs(float(Ytest[k, 2]) - Ypred[k, 2])
+	l4 = np.abs(float(Ytest[k, 3]) - Ypred[k, 3])
+	list_1.append(l1)
+	list_2.append(l2)
+	list_3.append(l3)
+	list_4.append(l4)
+	max_1 = max(max_1, l1)
+	max_2 = max(max_2, l2)
+	max_3 = max(max_3, l3)
+	max_4 = max(max_4, l4)
+	loss1 += l1
+	loss2 += l2
+	loss3 += l3
+	loss4 += l4
+loss1 /= len(Ytest)
+loss2 /= len(Ytest)
+loss3 /= len(Ytest)
+loss4 /= len(Ytest)
+total_loss = loss1 + loss2 + loss3 + loss4
+total_loss /= 4.0
+print('median: %.3f %.3f %.3f %.3f' % (np.median(list_1), np.median(list_2), np.median(list_3), np.median(list_4)))
+print('max loss: %.3f %.3f %.3f %.3f' % (max_1, max_2, max_3, max_4))
+print('Real test score: %.3f %.3f %.3f %.3f' % (loss1, loss2, loss3, loss4))
+print("acc: %.3f" % (total_loss))
+

training_for_git/batch.lua

@@ -1,104 +0,0 @@
-require 'torch'   -- torch
-require 'optim'
-require 'nn'      -- provides a normalization operator
-local train_file_path = 'train.th7' 
-local test_file_path = 'test.th7'
-local train_data = torch.load(train_file_path)
-local test_data = torch.load(test_file_path)
-local Y = train_data[{{},{2,5}}]
-local X = train_data[{{},{6,-1}}]
-local test_labels = test_data[{{},{2,5}}]
-local test_X = test_data[{{},{6,-1}}]
-local batch_size = 30
-epochs = 3
-
-model = nn.Sequential()                 -- define the container
-ninputs = 350; noutputs = 4 ; nhiddens1 = 1024; nhiddens2 = 512; nhiddens3 = 256
-model:add(nn.Linear(ninputs,nhiddens1))
-model:add(nn.Sigmoid())
-model:add(nn.Linear(nhiddens1,nhiddens2))
-model:add(nn.Sigmoid())
-model:add(nn.Linear(nhiddens2,nhiddens3))
-model:add(nn.Sigmoid())
-model:add(nn.Linear(nhiddens3,noutputs))
-criterion = nn.AbsCriterion()--MSECriterion()
-x, dl_dx = model:getParameters()
-sgd_params = {
-   learningRate = 0.01,
-   learningRateDecay = 1e-08,
-   weightDecay = 0,
-   momentum = 0
-}
-
-function train(X,Y)
-	   
-   current_loss = 0
-   for batch = 1,(#train_data)[1], batch_size do
-
-      local inputs = {}
-      local targets = {} 
-      local x_start = batch
-      local x_end = math.min(batch + batch_size-1, (#train_data)[1])
-      for i = x_start,x_end do    
-      local target = Y[i]    
-      local input = X[i]  
-      table.insert(inputs, input)
-      table.insert(targets, target)
-      end
-	local feval = function(x_new)
-	   if x ~= x_new then
-	      x:copy(x_new)
-	   end
-	   dl_dx:zero()
-           local f=0
-	for i = 1, #inputs do
-	   local loss_x = criterion:forward(model:forward(inputs[i]), targets[i])
-	   model:backward(inputs[i], criterion:backward(model.output, targets[i]))
-	f = f+loss_x
-	end
-	   return f/#inputs, dl_dx:div(#inputs)
-	end
-      _,fs = optim.adagrad(feval,x,sgd_params)
-      current_loss = current_loss + fs[1]
-   end
-   current_loss = current_loss/( (#train_data)[1]/batch_size)
-   print('train loss = ' .. current_loss)
-   return current_loss
-end
-
-time = sys.clock()
-local cumm_loss = 0.
-for j = 1, epochs do
-    print(j)
-    cumm_loss = train( X, Y )
-print( 'Final loss = ' .. cumm_loss )
-if j%10 == 0 then
-print('id  approx   text')
-local loss1 = 0.0
-local loss2 = 0.0
-local loss3 = 0.0
-local loss4 = 0.0
-for i = 1,(#test_data)[1] do
-   local myPrediction = model:forward(test_X[i])
-   loss1 = loss1+math.abs(myPrediction[1] - test_labels[i][1])
-   loss2 = loss2+math.abs(myPrediction[2] - test_labels[i][2])
-   loss3 = loss3+math.abs(myPrediction[3] - test_labels[i][3])
-   loss4 = loss4+math.abs(myPrediction[4] - test_labels[i][4])
-end
-
-loss1 = loss1/(#test_data)[1]
-loss2 = loss2/(#test_data)[1]
-loss3 = loss3/(#test_data)[1]
-loss4 = loss4/(#test_data)[1]
-end
-end
-
-
--- time taken
-time = sys.clock() - time
-print( "Time per epoch = " .. (time / epochs) .. '[s]')
-
-
-
-print(loss1,loss2,loss3,loss4)
-torch.save('estimation_model.dat',model)

training_for_git/gpu_rec.lua

@@ -1,130 +0,0 @@
-require 'rnn'
-require 'optim'
-
-batchSize = 30
-rho = 10
-hiddenSize = 512
-hiddenSize1 = 256
-inputSize = 400
-outputSize = 3
-epochs = 100
-xStart = 6
-yStart = 2
-yEnd = 4
-
-
-local train_file_path = 'recurrent_train.th7' 
-local train_data = torch.load(train_file_path)
-local Y = train_data[{{},{yStart,yEnd}}]
-local X = train_data[{{},{xStart,-1}}]
-seriesSize = (#train_data)[1]
-print(seriesSize)
-local test_file_path = 'recurrent_test.th7'
-local test_data = torch.load(test_file_path)
-local test_labels = test_data[{{},{yStart,yEnd}}]
-local test_X = test_data[{{},{xStart,-1}}]
-
-model = nn.Sequential()
-model:add(nn.Sequencer(nn.FastLSTM(inputSize, hiddenSize, rho)))
-model:add(nn.Sequencer(nn.FastLSTM(hiddenSize, hiddenSize1, rho)))
-model:add(nn.Sequencer(nn.Linear(hiddenSize1, outputSize)))
-
-criterion = nn.SequencerCriterion(nn.AbsCriterion())
-
--- dummy dataset (task predict the next item)
---dataset = torch.randn(seriesSize, inputSize)
-
--- define the index of the batch elements
-offsets = {}
-for i= 1, batchSize do
-   table.insert(offsets, i)--math.ceil(math.random() * batchSize))
-end
-offsets = torch.LongTensor(offsets)
-
-function nextBatch()
-   local inputs, targets = {}, {}
-   for step = 1, rho do
-      --get a batch of inputs
-      table.insert(inputs, X:index(1, offsets))
-      -- shift of one batch indexes
-      offsets:add(1)
-      for j=1,batchSize do
-         if offsets[j] > seriesSize then
-            offsets[j] = 1
-         end
-      end
-      -- a batch of targets
-      table.insert(targets, Y[{{},{1,3}}]:index(1,offsets))
-   end
-   return inputs, targets
-end
-
--- get weights and loss wrt weights from the model
-x, dl_dx = model:getParameters()
-
-feval = function(x_new)
-	-- copy the weight if are changed
-	if x ~= x_new then
-		x:copy(x_new)
-	end
-
-	-- select a training batch
-	local inputs, targets = nextBatch()
-
-	-- reset gradients (gradients are always accumulated, to accommodate
-	-- batch methods)
-	dl_dx:zero()
-
-	-- evaluate the loss function and its derivative wrt x, given mini batch
-	local prediction = model:forward(inputs)
-	local loss_x = criterion:forward(prediction, targets)
-	model:backward(inputs, criterion:backward(prediction, targets))
-	
-	return loss_x, dl_dx
-end
-
-sgd_params = {
-   learningRate = 0.01,
-   learningRateDecay = 1e-08,
-   weightDecay = 0,
-   momentum = 0
-}
-
-time = sys.clock()
-for j = 1, epochs do
-   -- train a mini_batch of batchSize in parallel
-	 _, fs = optim.adagrad(feval,x, sgd_params)
-	print('error for iteration ' .. sgd_params.evalCounter  .. ' is ' .. fs[1])
-
-end
-
-
-print('id  approx   text')
-local loss1 = 0.0
-local loss2 = 0.0
-local loss3 = 0.0
-local loss4 = 0.0
-for i = 1,(#test_data)[1], 1 do
-   local inputs = {}
-   for step = 1, 1 do
-      --get a batch of inputs
-      table.insert(inputs, test_X[i])
-   end
-   local myPrediction = model:forward(inputs)
-   loss1 = loss1+math.abs(myPrediction[1][1] - test_labels[i][1])
-   loss2 = loss2+math.abs(myPrediction[1][2] - test_labels[i][2])
-   loss3 = loss3+math.abs(myPrediction[1][3] - test_labels[i][3])
-   --loss4 = loss4+math.abs(myPrediction[4] - test_labels[i][4])
-end
-
-loss1 = loss1/(#test_data)[1]
-loss2 = loss2/(#test_data)[1]
-loss3 = loss3/(#test_data)[1]
---loss4 = loss4/(#test_data)[1]
-
--- time taken
-time = sys.clock() - time
-print( "Time per epoch = " .. (time / epochs) .. '[s]')
-
-print(loss1,loss2,loss3,loss4)
-torch.save('recurrent.dat',model)

training_for_git/no_validation.lua

@@ -1,129 +0,0 @@
-require 'torch'   -- torch
-require 'optim'
-require 'nn'      -- provides a normalization operator
-local train_file_path = 'train.th7' 
-local test_file_path = 'test.th7'
-local train_data = torch.load(train_file_path)
-local test_data = torch.load(test_file_path)
-local train_labels = train_data[{{},{2,5}}]
-local train_X = train_data[{{},{6,-1}}]
-local test_labels = test_data[{{},{2,5}}]
-local test_X = test_data[{{},{6,-1}}]
-local batch_size = 30
-model = nn.Sequential()                 -- define the container
-ninputs = 350; noutputs = 4 ; nhiddens1 = 1024; nhiddens2 = 512; nhiddens3 = 256
---model:add(nn.Linear(ninputs, noutputs)) -- define the only module
-model:add(nn.Linear(ninputs,nhiddens1))
-model:add(nn.Sigmoid())
-model:add(nn.Linear(nhiddens1,nhiddens2))
-model:add(nn.Sigmoid())
-model:add(nn.Linear(nhiddens2,nhiddens3))
-model:add(nn.Sigmoid())
-model:add(nn.Linear(nhiddens3,noutputs))
-criterion = nn.AbsCriterion()--MSECriterion()
-x, dl_dx = model:getParameters()
-
-feval = function(x_new)
-   if x ~= x_new then
-      x:copy(x_new)
-   end
-   -- select a new training sample
-   _nidx_ = (_nidx_ or 0) + 1
-   if _nidx_ > (#train_data)[1] then _nidx_ = 1 end
-   --local sample = data[_nidx_]
-   local target = train_labels[_nidx_]      -- this funny looking syntax allows
-   local inputs = train_X[_nidx_]    -- slicing of arrays.
-   -- reset gradients (gradients are always accumulated, to accommodate 
-   -- batch methods)
-   dl_dx:zero()
-   -- evaluate the loss function and its derivative wrt x, for that sample
-   --print(inputs)
-   --print(target)
-   for i=1, 350 do
-  if type(inputs[i]) ~= 'number' then
-  print(i)
-  print(inputs[i])
-  print(type(inputs[i])) end
-  end
-   --io.write("continue with this operation (y/n)?")
-   --answer=io.read()
-   local loss_x = criterion:forward(model:forward(inputs), target)
-   model:backward(inputs, criterion:backward(model.output, target))
-   -- return loss(x) and dloss/dx
-   return loss_x, dl_dx
-end
--- Given the function above, we can now easily train the model using SGD.
--- For that, we need to define four key parameters:
---   + a learning rate: the size of the step taken at each stochastic 
---     estimate of the gradient
---   + a weight decay, to regularize the solution (L2 regularization)
---   + a momentum term, to average steps over time
---   + a learning rate decay, to let the algorithm converge more precisely
-sgd_params = {
-   learningRate = 0.01,
-   learningRateDecay = 1e-08,
-   weightDecay = 0,
-   momentum = 0
-}
--- We're now good to go... all we have left to do is run over the dataset
--- for a certain number of iterations, and perform a stochastic update 
--- at each iteration. The number of iterations is found empirically here,
--- but should typically be determinined using cross-validation.
--- we cycle 1e4 times over our training data
-for i = 1,1 do
-   print(i)
-   -- this variable is used to estimate the average loss
-   current_loss = 0
-   -- an epoch is a full loop over our training data
-   for i = 1,(#train_data)[1] do
-      -- optim contains several optimization algorithms. 
-      -- All of these algorithms assume the same parameters:
-      --   + a closure that computes the loss, and its gradient wrt to x, 
-      --     given a point x
-      --   + a point x
-      --   + some parameters, which are algorithm-specific      
-      _,fs = optim.adagrad(feval,x,sgd_params)
-      -- Functions in optim all return two things:
-      --   + the new x, found by the optimization method (here SGD)
-      --   + the value of the loss functions at all points that were used by
-      --     the algorithm. SGD only estimates the function once, so
-      --     that list just contains one value.
-      current_loss = current_loss + fs[1]
-   end
-   -- report average error on epoch
-   current_loss = current_loss / (#train_data)[1]
-   print('train loss = ' .. current_loss)
-   
-end
-----------------------------------------------------------------------
--- 5. Test the trained model.
-
--- Now that the model is trained, one can test it by evaluating it
--- on new samples.
-
--- The text solves the model exactly using matrix techniques and determines
--- that 
---   corn = 31.98 + 0.65 * fertilizer + 1.11 * insecticides
-
--- We compare our approximate results with the text's results.
-
-print('id  approx   text')
-local loss1 = 0.0
-local loss2 = 0.0
-local loss3 = 0.0
-local loss4 = 0.0
-for i = 1,(#test_data)[1] do
-   local myPrediction = model:forward(test_X[i])
-   loss1 = loss1+math.abs(myPrediction[1] - test_labels[i][1])
-   loss2 = loss2+math.abs(myPrediction[2] - test_labels[i][2])
-   loss3 = loss3+math.abs(myPrediction[3] - test_labels[i][3])
-   loss4 = loss4+math.abs(myPrediction[4] - test_labels[i][4])
-end
-
-loss1 = loss1/(#test_data)[1]
-loss2 = loss2/(#test_data)[1]
-loss3 = loss3/(#test_data)[1]
-loss4 = loss4/(#test_data)[1]
-
-print(loss1,loss2,loss3,loss4)
-torch.save('save.dat',model)

training_for_git/recurrent.lua

@@ -1,109 +0,0 @@
-require 'rnn'
-require 'optim'
-
-function range(from, to, step)
-  step = step or 1
-  return function(_, lastvalue)
-    local nextvalue = lastvalue + step
-    if step > 0 and nextvalue <= to or step < 0 and nextvalue >= to or
-       step == 0
-    then
-      return nextvalue
-    end
-  end, nil, from - step
-end
-
-local train_file_path = 'recurrent_train.th7' 
-local test_file_path = 'recurrent_test.th7'
-local train_data = torch.load(train_file_path)
-local test_data = torch.load(test_file_path)
-local Y = train_data[{{},{2,5}}]
-local X = train_data[{{},{6,-1}}]
-local test_labels = test_data[{{},{2,5}}]
-local test_X = test_data[{{},{6,-1}}]
-
-batchSize = 5
-rho = 10
-hiddenSize1 = 1024
-hiddenSize2 = 512
-hiddenSize3 = 256
-inputSize = 1
-outputSize = 1
-seriesSize = 100
-
-model = nn.Sequential()
-model:add(nn.Sequencer(nn.FastLSTM(inputSize, hiddenSize2, rho)))
-model:add(nn.Sequencer(nn.FastLSTM(hiddenSize2, hiddenSize3, rho)))
---model:add(nn.Sequencer(nn.Linear(hiddenSize2, hiddenSize3, rho)))
---model:add(nn.Sequencer(nn.Sigmoid()))
-model:add(nn.Sequencer(nn.Linear(hiddenSize3, outputSize)))
-
-criterion = nn.SequencerCriterion(nn.MSECriterion())
-
--- dummy dataset (task predict the next item)
---dataset = torch.randn(seriesSize, inputSize)
-
--- define the index of the batch elements
-offsets = {}
-for i= 1, batchSize do
-   table.insert(offsets,i)
-end
-offsets = torch.LongTensor(offsets)
-print(offsets)
-function nextBatch()
-   local inputs, targets = {}, {}
-   for step = 1, rho do
-      --get a batch of inputs
-      table.insert(inputs, X:index(1, offsets))
-      -- shift of one batch indexes
-      offsets:add(1)
-      for j=1,batchSize do
-         if offsets[j] > seriesSize then
-            offsets[j] = 1
-         end
-      end
-      -- a batch of targets
-      table.insert(targets, Y:index(1,offsets))
-   end
-   return inputs, targets
-end
-
--- get weights and loss wrt weights from the model
-x, dl_dx = model:getParameters()
-
-feval = function(x_new)
-  -- copy the weight if are changed
-  if x ~= x_new then
-    x:copy(x_new)
-  end
-
-  -- select a training batch
-  local inputs, targets = nextBatch()
-
-  -- reset gradients (gradients are always accumulated, to accommodate
-  -- batch methods)
-  dl_dx:zero()
-
-  -- evaluate the loss function and its derivative wrt x, given mini batch
-  local prediction = model:forward(inputs)
-  local loss_x = criterion:forward(prediction, targets)
-  model:backward(inputs, criterion:backward(prediction, targets))
-
-  return loss_x, dl_dx
-end
-
-sgd_params = {
-   learningRate = 0.01,
-   learningRateDecay = 1e-08,
-   weightDecay = 0,
-   momentum = 0
-}
-
-for i = 1, 2 do
-   -- train a mini_batch of batchSize in parallel
-   _, fs = optim.adagrad(feval,x, sgd_params)
-
-  if sgd_params.evalCounter % 100 == 0 then
-    print('error for iteration ' .. sgd_params.evalCounter  .. ' is ' .. fs[1] / rho)
-  end
-end

training_for_git/recurrent_training.lua

@@ -1,144 +0,0 @@
-require 'rnn'
-require 'optim'
-
-batchSize = 30
-rho = 20
-hiddenSize = 512
-hiddenSize1 = 256
-inputSize = 400
-outputSize = 4
-epochs = 10000
-xStart = 6
-yStart = 2
-yEnd = 5
-
-
-local train_file_path = 'recurrent_train.th7' 
-local train_data = torch.load(train_file_path)
-local Y = train_data[{{},{yStart,yEnd}}]
-local X = train_data[{{},{xStart,-1}}]
-local place = train_data[{{},{1}}]
-seriesSize = (#train_data)[1]
-print(seriesSize)
-local test_file_path = 'recurrent_test.th7'
-local test_data = torch.load(test_file_path)
-local test_labels = test_data[{{},{yStart,yEnd}}]
-local test_X = test_data[{{},{xStart,-1}}]
-
-model = nn.Sequential()
-model:add(nn.Sequencer(nn.FastLSTM(inputSize, hiddenSize, rho)))
-model:add(nn.Sequencer(nn.FastLSTM(hiddenSize, hiddenSize1, rho)))
-model:add(nn.Sequencer(nn.Linear(hiddenSize1, outputSize)))
-
-criterion = nn.SequencerCriterion(nn.AbsCriterion())
---local method = 'xavier'
---local model_new = require('weight-init')(model, method)
-
--- define the index of the batch elements
-offsets = {}
-function offset_(seed)
-offsets = {}
-math.randomseed(seed)
-for i= 1, batchSize do
-   table.insert(offsets, math.ceil(math.random() * batchSize))
-end
-offsets = torch.LongTensor(offsets)
-end
-function nextBatch()
-   local inputs, targets = {}, {}
-   local nums = {}
-   for step = 1, rho do
-      --get a batch of inputs
-      table.insert(inputs, X:index(1, offsets))
-      -- shift of one batch indexes
-      offsets:add(1)
-      for j=1,batchSize do
-         if offsets[j] > seriesSize then
-            offsets[j] = 1
-         end
-      end
-      -- a batch of targets
-      table.insert(targets, Y[{{},{1,4}}]:index(1,offsets))
-      table.insert(nums,place:index(1,offsets))
-   end
-   return inputs, targets
-end
-
--- get weights and loss wrt weights from the model
-x, dl_dx = model:getParameters()
-
-feval = function(x_new)
-	-- copy the weight if are changed
-	if x ~= x_new then
-		x:copy(x_new)
-	end
-
-	-- select a training batch
-	local inputs, targets = nextBatch()
-
-	-- reset gradients (gradients are always accumulated, to accommodate
-	-- batch methods)
-	dl_dx:zero()
-
-	-- evaluate the loss function and its derivative wrt x, given mini batch
-	local prediction = model:forward(inputs)
-	local loss_x = criterion:forward(prediction, targets)
-	model:backward(inputs, criterion:backward(prediction, targets))
-	
-	return loss_x, dl_dx
-end
-
-adagrad_params = {
-   learningRate = 0.01,
-   learningRateDecay = 1e-08,
-   weightDecay = 0,
-   momentum = 0
-}
-seed = 1
-offset_(seed)
-time = sys.clock()
-for j = 1, epochs do
-	if j%1000 == 0 then 
-		seed = seed + 1
-		offset_(seed)
-	end
-   -- train a mini_batch of batchSize in parallel
-	 _, fs = optim.adagrad(feval,x, adagrad_params)
-	print('error for iteration ' .. adagrad_params.evalCounter  .. ' is ' .. fs[1]/rho)
-
-end
-
-
-print('id  approx   text')
-local loss1 = 0.0
-local loss2 = 0.0
-local loss3 = 0.0
-local loss4 = 0.0
-predict_batch = 100
-for i = 1,(#test_data)[1], predict_batch do
-   local inputs = {}
-   for step = 0, predict_batch-1 do
-      --get a batch of inputs
-      table.insert(inputs, test_X[i+step])
-   end
-   local myPrediction = model:forward(inputs)
-   for step = 1, predict_batch do
-   loss1 = loss1+math.abs(myPrediction[step][1] - test_labels[i+step-1][1])
-   loss2 = loss2+math.abs(myPrediction[step][2] - test_labels[i+step-1][2])
-   loss3 = loss3+math.abs(myPrediction[step][3] - test_labels[i+step-1][3])
-   loss4 = loss4+math.abs(myPrediction[4] - test_labels[i][4])
-   end
-
-end
-
-loss1 = loss1/(#test_data)[1]
-loss2 = loss2/(#test_data)[1]
-loss3 = loss3/(#test_data)[1]
-loss4 = loss4/(#test_data)[1]
-
--- time taken
-time = sys.clock() - time
-print( "Time per epoch = " .. (time / epochs) .. '[s]')
-
-print(loss1,loss2,loss3,loss4)
-torch.save('recurrent3.dat',model)