def predict_nn(trndata,
epochs=300,
test_error=0.0147,
weight_decay=0.0001,
momentum=0.15):
"""
FF neural net
"""
fnn = buildNetwork(trndata.indim,
trndata.indim / 4,
trndata.outdim,
outclass=SoftmaxLayer)
trainer = BackpropTrainer(fnn,
trndata,
momentum=momentum,
weightdecay=weight_decay)
epoch_delta = 1
stop = False
totEpochs = np.array([])
trnResults = np.array([])
trnLogLoss = np.array([])
f, (ax1, ax2) = plt.subplots(1, 2, figsize=(15, 8))
while not stop:
trainer.trainEpochs(epoch_delta)
trnresult = percentError(trainer.testOnClassData(), trndata['class'])
trnLogLoss = append_to_arr(trnLogLoss, nn_log_loss(fnn, trndata))
print("epoch: %4d" % trainer.totalepochs, \
" train error: %5.2f%%" % trnresult, \
" train logloss: %2.4f" % trnLogLoss[-1])
trnResults = append_to_arr(trnResults, trnresult)
totEpochs = append_to_arr(totEpochs, trainer.totalepochs)
plt.sca(ax1)
plt.cla()
ax1.plot(totEpochs, trnResults, label='Train')
plt.sca(ax2)
plt.cla()
ax2.plot(totEpochs, trnLogLoss, label='Train')
ax1.legend()
ax2.legend()
plt.draw()
time.sleep(0.1)
plt.pause(0.0001)
stop = (trnLogLoss[-1] <= test_error or trainer.totalepochs >= epochs)
return fnn
def get_1dlbp_features(neighborhood):
tf = TrainFiles(inp_path, floor = neighborhood * 2 + 1)
inputs = tf.get_training_inputs()
start = timer()
hist = np.array([])
outs = np.array([])
i = 0
writeBatch = 2000
prep_out_path(out_path)
p = 1 << np.array(range(0, 2 * neighborhood), dtype='int32')
d_powers = cuda.to_device(p)
for inp in inputs:
data_file = path.join(inp_path, inp)
out_file = path.join(out_path, path.splitext(inp)[0] + ext)
arr = np.fromfile(data_file, dtype = 'uint8')
##GPU##
file_hist = extract_1dlbp_gpu(arr, neighborhood, d_powers)
##CPU##
#file_hist = extract_1dlbp_cpu(arr, neighborhood, p)
#file_hist = file_histogram(file_hist, neighborhood)
i += 1
hist = append_to_arr(hist, file_hist)
outs = append_to_arr(outs, out_file)
if i == writeBatch:
i = 0
first = True
print "Writing....."
for j in range(0, outs.shape[0]):
hist[j].tofile(outs[j])
hist = np.array([])
outs = np.array([])
print "==============Done==================="
print "Elapsed: ", timer() - start
print "Writing......."
for i in range(0, outs.shape[0]):
hist[i].tofile(outs[i])
print "==============Done==================="
def predict_nn(trndata, epochs = 300, test_error = 0.0147, weight_decay = 0.0001, momentum = 0.15):
"""
FF neural net
"""
fnn = buildNetwork(trndata.indim, trndata.indim / 4, trndata.outdim, outclass = SoftmaxLayer)
trainer = BackpropTrainer(fnn, trndata, momentum = momentum, weightdecay = weight_decay)
epoch_delta = 1
stop = False
totEpochs = np.array([])
trnResults = np.array([])
trnLogLoss = np.array([])
f, (ax1, ax2) = plt.subplots(1, 2, figsize=(15,8))
while not stop:
trainer.trainEpochs(epoch_delta)
trnresult = percentError( trainer.testOnClassData(),
trndata['class'] )
trnLogLoss = append_to_arr(trnLogLoss, nn_log_loss(fnn, trndata))
print "epoch: %4d" % trainer.totalepochs, \
" train error: %5.2f%%" % trnresult, \
" train logloss: %2.4f" % trnLogLoss[-1]
trnResults = append_to_arr(trnResults, trnresult)
totEpochs = append_to_arr(totEpochs, trainer.totalepochs)
plt.sca(ax1)
plt.cla()
ax1.plot(totEpochs, trnResults, label = 'Train')
plt.sca(ax2)
plt.cla()
ax2.plot(totEpochs, trnLogLoss, label = 'Train')
ax1.legend()
ax2.legend()
plt.draw()
time.sleep(0.1)
plt.pause(0.0001)
stop = (trnLogLoss[-1] <= test_error or trainer.totalepochs >= epochs)
return fnn
def _connect_labeled_data(self, inp_path, inputs, training):
if self.labels == None:
self.labels = self.get_labels_csv()
X = np.array([])
Y = np.array([])
zip = False
if self.isZip:
zip = ZipFile(inp_path)
for inp in inputs:
inp_file = path.join(inp_path, inp) if not zip else inp
x = np.fromfile(inp_file,
dtype='int') if not zip else np.frombuffer(
zip.read(inp_file), dtype='int')
x = x.astype('float')
X = append_to_arr(X, x)
if training or self.validate:
label_name = path.splitext(path.split(inp_file)[1])[0]
label = filter(lambda x: x[0] == label_name, self.labels)[0][1]
Y = np.append(Y, label)
if self.debug:
print "Processed: " + inp_file
return X, Y
def connect_labeled_data(self, training, load_from_file):
inputs, inp_path = self._get_inp_input_path(training)
self.labels = self.get_labels_csv()
X = np.array([])
# TODO: this is probably not the best way of doing it
zip = False
if self.isZip:
zip = ZipFile(inp_path)
for inp in inputs:
inp_file = path.join(inp_path, inp) if not zip else inp
x = load_from_file(inp_file)
x = np.r_[[path.splitext(inp)[0]], x]
X = append_to_arr(X, x)
if self.debug:
print "Processed: " + inp_file
id = self.labels.columns[0]
dfX = pd.DataFrame(X)
dfX.columns=np.r_[[id], range(X.shape[1] - 1)]
dfX = dfX.merge(self.labels, how='inner', on=id)
return np.array(dfX[range(1, X.shape[1])]).astype(float), np.array(dfX[self.labels.columns[1]]).astype(int)
def connect_labeled_data(self, training, load_from_file):
inputs, inp_path = self._get_inp_input_path(training)
self.labels = self.get_labels_csv()
X = np.array([])
# TODO: this is probably not the best way of doing it
zip = False
if self.isZip:
zip = ZipFile(inp_path)
for inp in inputs:
inp_file = path.join(inp_path, inp) if not zip else inp
x = load_from_file(inp_file)
x = np.r_[[path.splitext(inp)[0]], x]
X = append_to_arr(X, x)
if self.debug:
print "Processed: " + inp_file
id = self.labels.columns[0]
dfX = pd.DataFrame(X)
dfX.columns = np.r_[[id], range(X.shape[1] - 1)]
dfX = dfX.merge(self.labels, how='inner', on=id)
return np.array(dfX[range(1, X.shape[1])]).astype(float), np.array(
dfX[self.labels.columns[1]]).astype(int)
def _connect_labeled_data_csv(self, inp_path, inputs, training):
if self.labels == None:
self.labels = self.get_labels_csv()
X = np.array([])
Y = np.array([])
for inp in inputs:
inp_file = path.join(inp_path, inp)
x = np.loadtxt(inp_file)
X = append_to_arr(X, x)
if training or self.validate:
label_name = path.splitext(inp)[0]
label = filter(lambda x: x[0] == label_name, self.labels)[0][1]
Y = np.append(Y, label)
if self.debug:
print "Processed: " + inp_file
return X, Y
def _connect_labeled_data_csv(self, inp_path, inputs, training):
if self.labels == None:
self.labels = self.get_labels_csv()
X = np.array([])
Y = np.array([])
for inp in inputs:
inp_file = path.join(inp_path, inp)
x = np.loadtxt(inp_file)
X = append_to_arr(X, x)
if training or self.validate:
label_name = path.splitext(inp)[0]
label = filter(lambda x: x[0] == label_name, self.labels)[0][1]
Y = np.append(Y, label)
if self.debug:
print "Processed: " + inp_file
return X, Y
def _connect_labeled_data(self, inp_path, inputs, training):
if self.labels == None:
self.labels = self.get_labels_csv()
X = np.array([])
Y = np.array([])
zip = False
if self.isZip:
zip = ZipFile(inp_path)
for inp in inputs:
inp_file = path.join(inp_path, inp) if not zip else inp
x = np.fromfile(inp_file, dtype='int') if not zip else np.frombuffer(zip.read(inp_file), dtype='int')
x = x.astype('float')
X = append_to_arr(X, x)
if training or self.validate:
label_name = path.splitext(path.split(inp_file)[1])[0]
label = filter(lambda x: x[0] == label_name, self.labels)[0][1]
Y = np.append(Y, label)
if self.debug:
print "Processed: " + inp_file
return X, Y
