def test(md, csv_file, model):
# load name of wavs to be classified
with open(csv_file, 'rb') as f:
reader = csv.reader(f)
lis = list(reader)
# do classification for each file
y_pred = []
te_y = []
for li in lis:
na = li[1]
#audio evaluation name
fe_path = eva_fd + '/' + feature + '/' + na + '.f'
info_path = label_csv + '/' + na + '.csv'
with open(info_path, 'rb') as g:
reader2 = csv.reader(g)
lis2 = list(reader2)
tags = lis2[-2][1]
y = np.zeros(num_classes)
for ch in tags:
y[lb_to_id[ch]] = 1
te_y.append(y)
X0 = cPickle.load(open(fe_path, 'rb'))
X0 = aud_utils.mat_2d_to_3d(X0, agg_num, hop)
X0 = aud_utils.mat_3d_to_nd(model, X0)
# predict
p_y_pred = md.predict(X0)
p_y_pred = np.mean(p_y_pred, axis=0) # shape:(n_label)
y_pred.append(p_y_pred)
return np.array(te_y), np.array(y_pred)
def get_test_predictions(self, md):
model = self.model
feature = self.feature
fe_fd = cfg.eva_fd + '/' + feature
csv_file = cfg.txt_eva_path
agg_num = self.agg_num
hop = self.hop
eva_file = cfg.eva_file
# load name of wavs to be classified
with open(csv_file, 'rb') as f:
reader = csv.reader(f)
lis = list(reader)
# do classification for each file
names = []
pred_lbs = []
for li in lis:
names.append(li[0])
na = li[0][6:-4]
#audio evaluation name
fe_path = fe_fd + '/' + na + '.f'
X0 = cPickle.load(open(fe_path, 'rb'))
X0 = aud_utils.mat_2d_to_3d(X0, agg_num, hop)
X0 = aud_utils.mat_3d_to_nd(model, X0)
# predict
p_y_preds = md.predict(X0) # probability, size: (n_block,label)
preds = np.argmax(p_y_preds, axis=-1) # size: (n_block)
b = scipy.stats.mode(preds)
pred = int(b[0])
pred_lbs.append(cfg.id_to_lb[pred])
pred = []
# write out result
for i1 in xrange(len(names)):
fname = names[i1] + '\t' + pred_lbs[i1] + '\n'
pred.append(fname)
print 'write out finished!'
truth = open(eva_file, 'r').readlines()
pred = [i.split('\t')[1].split('\n')[0] for i in pred]
truth = [i.split('\t')[1] for i in truth]
pred.sort()
truth.sort()
return truth, pred
def get_test_predictions(self, md):
model = self.model
feature = self.feature
fe_fd = cfg.eva_fd + '/' + feature
csv_file = cfg.meta_test_csv
agg_num = self.agg_num
hop = self.hop
# load name of wavs to be classified
with open(csv_file, 'rb') as f:
reader = csv.reader(f)
lis = list(reader)
# do classification for each file
y_pred = []
te_y = []
for li in lis:
na = li[1]
#audio evaluation name
fe_path = fe_fd + '/' + na + '.f'
info_path = cfg.label_csv + '/' + na + '.csv'
with open(info_path, 'rb') as g:
reader2 = csv.reader(g)
lis2 = list(reader2)
tags = lis2[-2][1]
y = np.zeros(len(cfg.labels))
for ch in tags:
y[cfg.lb_to_id[ch]] = 1
te_y.append(y)
X0 = cPickle.load(open(fe_path, 'rb'))
X0 = aud_utils.mat_2d_to_3d(X0, agg_num, hop)
X0 = aud_utils.mat_3d_to_nd(model, X0)
# predict
p_y_pred = md.predict(X0)
p_y_pred = np.mean(p_y_pred, axis=0) # shape:(n_label)
y_pred.append(p_y_pred)
return np.array(te_y), np.array(y_pred)
tr_X = list(np.zeros(len(feature),dtype='int'))
dimy = list(np.zeros(len(feature),dtype='int'))
for i in range(len(feature)):
tr_X[i], tr_y = GetAllData( dev_fd+'/'+feature[i], label_csv, agg_num, hop )
print(tr_X[i].shape)
print(tr_y.shape)
for i in range(len(feature)):
dimy[i]=tr_X[i].shape[-1]
#aud_utils.check_dimension(feature[i],dimy[i],'defaults.yaml')
#tr_X=aud_utils.equalise(tr_X)
for i in range(len(feature)):
tr_X[i]=aud_utils.mat_3d_to_nd(model,tr_X[i])
print(tr_X[i].shape)
dimx=tr_X[0].shape[-2]
if prep=='dev':
cross_validation=True
else:
cross_validation=False
miz=aud_model.Functional_Model(input_neurons=input_neurons,cross_validation=cross_validation,dropout1=dropout1,
act1=act1,act2=act2,act3=act3,nb_filter = nb_filter, filter_length=filter_length,
num_classes=num_classes,
model=model,dimx=dimx,dimy=dimy)
np.random.seed(68)
if cross_validation:
file_logger = FileLogger('out_{}.tsv'.format(model),
['step', 'tr_loss', 'te_loss', 'tr_acc', 'te_acc'])
############# EXTRACT FEATURES #####################
dropout, act1, act2,act3, input_neurons, epochs, batchsize, num_classes, _, _,\
loss,nb_filter, filter_length,pool_size, optimizer=seth.get_parameters()
############# LOAD DATA ###########################
tr_X, tr_y = seth.get_train_data('logmel_trainx2.cpkl')
v_X, v_y = seth.get_val_data('logmel_testx2.cpkl')
dimx = tr_X.shape[-2]
dimy = tr_X.shape[-1]
tr_X = aud_utils.mat_3d_to_nd(model, tr_X)
v_X = aud_utils.mat_3d_to_nd(model, v_X)
print("Functional_model {}".format(model))
print("Activation 1 {} 2 {} 3 {}".format(act1, act2, act3))
print("Dropout {}".format(dropout))
print("Kernels {} Size {} Poolsize {}".format(nb_filter, filter_length,
pool_size))
print("Loss {} Optimizer {}".format(loss, optimizer))
#train_x=np.array(tr_X)
#train_y=np.array(tr_y)
#test_x=np.array(v_X)
#test_y=np.array(v_y)
##train_x, test_x, train_y, test_y = train_test_split( tr_X, tr_y, test_size=0.2, random_state=42)
#print("Evaluation mode")
