# Remove unused variables
print "Removing Unused Variables..."
del TRAINING_SET,TRAINING_ID,TRAINING_FEATURES,LABEL_INDEXES,LABEL_FEATURES,\
BATCHED_TRAINING_FEATURES,BATCHED_LABEL_FEATURES
#pdb.set_trace()
########################
# Create Neural Net #
########################
print "Constructing Neural Network..."
DATA_LAYER = [ BATCHED_INPUT[0].shape[0] ]
LABEL_LAYER = [ BATCHED_OUTPUT[0].shape[0] ]
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
nn = DNN(LAYERS)
#pdb.set_trace()
########################
# pre-Train Neural Net #
########################
'''
print "Start pre-training. pretrain {0} epoches".format(PRETRAIN_EPOCH)
prop_input = data
for l,da in enumerate(DAs):
for epoch in xrange(PRETRAIN_EPOCH):
batch_cost = 0
tStart = time.time()
for i in xrange( (data.shape[1]-1)/PRETRAIN_BATCH_SIZE + 1):
begin = i * PRETRAIN_BATCH_SIZE
if (i+1)*PRETRAIN_BATCH_SIZE > data.shape[1]:
LABEL_LAYER = [ len( BATCHED_VECTORS[0][0] ) ]
# pdb.set_trace()
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
print "Data parsed!!!"
########################
# Create Neural Net #
########################
model_path = "./result/model/DATA_fbank_LABEL_phoneme48_HIDDEN_LAYERS_1024-1024_L_RATE_0.01_MOMENTUM_0.9_DROPOUT_0.1_EPOCH_500_at_100"
ll,ww,bb = pickle.load(open(model_path,'rb'))
nn = DNN(ll,
ww,
bb,
m_norm=MAX_NORM,
act=ACT_FUNC,
cost=COST_FUNC,
momentum_type=MOMENTUM_TYPE)
########################
# pre-Train Neural Net #
########################
'''
print "Start pre-training. pretrain {0} epoches".format(PRETRAIN_EPOCH)
prop_input = data
for l,da in enumerate(DAs):
for epoch in xrange(PRETRAIN_EPOCH):
batch_cost = 0
tStart = time.time()
for i in xrange( (data.shape[1]-1)/PRETRAIN_BATCH_SIZE + 1):
TEST_ROOT = './Data' TEST = '/fbank/test.ark' #TEST = './train_ant.ark' PREDICTION_ROOT ='./result/prediction/' PREDICTION = MODEL + '.csv' ######################## # load DNN open file # ######################## ACT_FUNC="leakyReLU" COST_FUNC ="CE" layers,Ws,bs = pickle.load(open(MODEL_ROOT+MODEL,'rb')) nn = DNN(layers,Ws,bs,act=ACT_FUNC,cost=COST_FUNC) # nn.rescale_params(0.9) # MODEL = "DATA_fbank_LABEL_phoneme48_HIDDEN_LAYERS_1024-1024-1024-1024_L_RATE_0.001_MOMENTUM_0.9_DROPOUT_0_EPOCH_100" TEST_DATA,VAL_DATA = readfile( TEST_ROOT+TEST,1 ) PRED_FILE = open( PREDICTION_ROOT + PREDICTION ,'wb') # Get Dictionaries Phone48 = load_list39to48() PhoneMap48to39 = load_dict_48to39() # For CSV HEADER = ["Id","Prediction"] ########################
pickList,val_data,val_label,val_IDs = parse_val_set(mem_data,mem_label,pickList,IDs,VAL_SET_RATIO)
# TRAINING_LABEL is a dict()
DATA_LAYER = [ mem_data.shape[0] ]
LABEL_LAYER = [ LABEL_VARIETY ]
# pdb.set_trace()
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
pdb.set_trace()
print "Data parsed!!!"
########################
# Create Neural Net #
########################
nn = DNN(LAYERS)
########################
# pre-Train Neural Net #
########################
'''
print "Start pre-training. pretrain {0} epoches".format(PRETRAIN_EPOCH)
prop_input = data
for l,da in enumerate(DAs):
for epoch in xrange(PRETRAIN_EPOCH):
batch_cost = 0
tStart = time.time()
for i in xrange( (data.shape[1]-1)/PRETRAIN_BATCH_SIZE + 1):
begin = i * PRETRAIN_BATCH_SIZE
if (i+1)*PRETRAIN_BATCH_SIZE > data.shape[1]:
end = data.shape[1]
MEM_DATA = './data.fbank.memmap'
MEM_TEST = './test.fbank.memmap'
PKL_ID = './ID.pkl'
PKL_ID_TEST = './ID_test.pkl'
MEM_DATA_shape = (621, 180406)
STATE_LENGTH = 1943
PHONE_LENGTH = 48
BATCH_SIZE = 2
########################
# load DNN open file #
########################
layers,Ws,bs = pickle.load(open(MODEL_ROOT+MODEL,'rb'))
nn = DNN(layers,Ws,bs,
act=ACT_FUNC,
cost=COST_FUNC)
# read Data #
mem_data = np.memmap(MEM_TEST,dtype='float32',mode='r',shape=MEM_DATA_shape)
IDs = readID(PKL_ID_TEST)
print "Data parsed"
########################
# Save posteriorgram #
########################
mem_shape = (len(IDs),PHONE_LENGTH)
posteriorgram = np.memmap(PGRAM_ROOT+PGRAM,
dtype='float32',
mode='w+',
MODEL = "DATA_fbank_LABEL_phonemeState_HIDDEN_LAYERS_2048-2048-2048_L_RATE_0.01_MOMENTUM_0.9_DROPOUT_0.1_EPOCH_200_at_150" TEST_ROOT = './Data/fbank/' TEST = 'test.ark' #TEST = './train_ant.ark' PREDICTION_ROOT ='./result/prediction/' PREDICTION = MODEL + '.csv' ######################## # load DNN open file # ######################## layers,Ws,bs = pickle.load(open(MODEL_ROOT+MODEL,'rb')) nn = DNN(layers,Ws,bs) TEST_DATA,VAL_DATA = readfile_( TEST_ROOT+TEST,1 ) PRED_FILE = open( PREDICTION_ROOT + PREDICTION ,'wb') # Get Dictionaries Phone48 = load_liststateto48() PhoneMap48to39 = load_dict_48to39() # For CSV HEADER = ["Id","Prediction"] ######################## # Predict # ########################
PGRAM_ROOT ='./result/posteriorgram/'
PGRAM = MODEL + '.pgram'
MEM_DATA = 'data.fbank.memmap'
PKL_ID = 'ID.pkl'
MEM_DATA_shape = (621,1124823)
STATE_LENGTH = 1943
BATCH_SIZE = 1847
########################
# load DNN open file #
########################
layers,Ws,bs = pickle.load(open(MODEL_ROOT+MODEL,'rb'))
nn = DNN(layers,Ws,bs,
act=ACT_FUNC,
cost=COST_FUNC)
# read Data #
mem_data = np.memmap(MEM_DATA,dtype='float32',mode='r',shape=MEM_DATA_shape)
IDs = readID(PKL_ID)
print "Data parsed"
########################
# Save posteriorgram #
########################
mem_shape = (STATE_LENGTH,len(IDs))
posteriorgram = np.memmap(PGRAM,dtype='float32',mode='w+',shape=mem_shape)
########################
for data in LABELED_VALIDATION_SET:
validationNlabel.append(data + [LABEL_DICT[data[0]]])
DATA_LAYER = [ len( BATCHED_TRAINING_SET[0][0] ) ]
LABEL_LAYER = [ len( BATCHED_VECTORS[0][0] ) ]
# pdb.set_trace()
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
########################
# Create Neural Net #
########################
print "Initializing Neural Network..."
nn = DNN(LAYERS)
########################
# pre-Train Neural Net #
########################
'''
prop_input = _DATA_BATCH_
for l in xrange(1,len(nn.layers)):
for epoch in PRE_TRAIN_EPOCH:
nn.pretrainer[l].train()
# get the hiiden vector of current dA and
# pass into the next dA as data.
prop_input = nn.pretrainer[l].get_hidden(prop_input)
'''
DATA_LAYER = [ len( BATCHED_TRAINING_SET[0][0] ) ]
LABEL_LAYER = [ len( BATCHED_VECTORS[0][0] ) ]
del BATCHED_TRAINING_SET
del BATCHED_VECTORS
# pdb.set_trace()
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
print "Data parsed!!!"
########################
# Create Neural Net #
########################
nn = DNN(LAYERS)
########################
# pre-Train Neural Net #
########################
'''
print "Start pre-training. pretrain {0} epoches".format(PRETRAIN_EPOCH)
prop_input = data
for l,da in enumerate(DAs):
for epoch in xrange(PRETRAIN_EPOCH):
batch_cost = 0
tStart = time.time()
for i in xrange( (data.shape[1]-1)/PRETRAIN_BATCH_SIZE + 1):
begin = i * PRETRAIN_BATCH_SIZE
if (i+1)*PRETRAIN_BATCH_SIZE > data.shape[1]:
end = data.shape[1]
LABEL_LAYER = [len(BATCHED_VECTORS[0][0])]
# pdb.set_trace()
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
print "Data parsed!!!"
########################
# Create Neural Net #
########################
model_path = "./result/model/DATA_fbank_LABEL_phoneme48_HIDDEN_LAYERS_1024-1024_L_RATE_0.01_MOMENTUM_0.9_DROPOUT_0.1_EPOCH_500_at_100"
ll, ww, bb = pickle.load(open(model_path, 'rb'))
nn = DNN(ll,
ww,
bb,
m_norm=MAX_NORM,
act=ACT_FUNC,
cost=COST_FUNC,
momentum_type=MOMENTUM_TYPE)
########################
# pre-Train Neural Net #
########################
'''
print "Start pre-training. pretrain {0} epoches".format(PRETRAIN_EPOCH)
prop_input = data
for l,da in enumerate(DAs):
for epoch in xrange(PRETRAIN_EPOCH):
batch_cost = 0
tStart = time.time()
for i in xrange( (data.shape[1]-1)/PRETRAIN_BATCH_SIZE + 1):
for data in LABELED_VALIDATION_SET:
validationNlabel.append(data + [LABEL_DICT[data[0]]])
DATA_LAYER = [len(BATCHED_TRAINING_SET[0][0])]
LABEL_LAYER = [len(BATCHED_VECTORS[0][0])]
# pdb.set_trace()
LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER
########################
# Create Neural Net #
########################
print "Initializing Neural Network..."
nn = DNN(LAYERS)
########################
# pre-Train Neural Net #
########################
'''
prop_input = _DATA_BATCH_
for l in xrange(1,len(nn.layers)):
for epoch in PRE_TRAIN_EPOCH:
nn.pretrainer[l].train()
# get the hiiden vector of current dA and
# pass into the next dA as data.
prop_input = nn.pretrainer[l].get_hidden(prop_input)
'''
TEST_ROOT = './Data' TEST = '/fbank/test.ark' #TEST = './train_ant.ark' PREDICTION_ROOT = './result/prediction/' PREDICTION = MODEL + '.csv' ######################## # load DNN open file # ######################## ACT_FUNC = "leakyReLU" COST_FUNC = "CE" layers, Ws, bs = pickle.load(open(MODEL_ROOT + MODEL, 'rb')) nn = DNN(layers, Ws, bs, act=ACT_FUNC, cost=COST_FUNC) # nn.rescale_params(0.9) # MODEL = "DATA_fbank_LABEL_phoneme48_HIDDEN_LAYERS_1024-1024-1024-1024_L_RATE_0.001_MOMENTUM_0.9_DROPOUT_0_EPOCH_100" TEST_DATA, VAL_DATA = readfile(TEST_ROOT + TEST, 1) PRED_FILE = open(PREDICTION_ROOT + PREDICTION, 'wb') # Get Dictionaries Phone48 = load_list39to48() PhoneMap48to39 = load_dict_48to39() # For CSV HEADER = ["Id", "Prediction"] ########################