def train():
# prepare data
tr_X, tr_y, _, te_X, te_y, te_na_list = pp_dev_data.GetAllData(
fe_fd, agg_num, hop, fold)
[batch_num, n_time, n_freq] = tr_X.shape
print tr_X.shape, tr_y.shape
print te_X.shape, te_y.shape
# build model
seq = Sequential()
seq.add(InputLayer((n_time, n_freq)))
seq.add(Flatten()) # flatten to 2d: (n_time, n_freq) to 1d:(n_time*n_freq)
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_out, act='sigmoid'))
md = seq.compile()
md.summary()
# optimizer
optimizer = Adam(1e-4)
# callbacks
# tr_err, te_err are frame based. To get event based err, run recognize.py
validation = Validation(tr_x=tr_X,
tr_y=tr_y,
va_x=None,
va_y=None,
te_x=te_X,
te_y=te_y,
batch_size=2000,
metrics=['binary_crossentropy'],
call_freq=1,
dump_path=None)
# save model
pp_dev_data.CreateFolder(cfg.dev_md_fd)
save_model = SaveModel(dump_fd=cfg.dev_md_fd, call_freq=10)
# callbacks
callbacks = [validation, save_model]
# fit model
md.fit(x=tr_X,
y=tr_y,
batch_size=2000,
n_epochs=100,
loss_func='binary_crossentropy',
optimizer=optimizer,
callbacks=callbacks,
verbose=1)
def recognize():
# prepare data
_, _, _, te_X, te_y, _ = pp_dev_data.GetAllData(fe_fd, agg_num, hop, fold)
# do recognize and evaluation
thres = 0.4 # thres, tune to prec=recall
n_labels = len(cfg.labels)
gt_roll = []
pred_roll = []
with open(cfg.dev_cv_csv_path, 'rb') as f:
reader = csv.reader(f)
lis = list(reader)
# read one line
for li in lis:
na = li[1]
curr_fold = int(li[2])
if fold == curr_fold:
# get features, tags
fe_path = fe_fd + '/' + na + '.f'
info_path = cfg.dev_wav_fd + '/' + na + '.csv'
tags = pp_dev_data.GetTags(info_path)
y = pp_dev_data.TagsToCategory(tags)
X = cPickle.load(open(fe_path, 'rb'))
# aggregate data
X3d = mat_2d_to_3d(X, agg_num, hop)
p_y_pred = md.predict(X3d)
p_y_pred = np.mean(p_y_pred, axis=0) # shape:(n_label)
pred = np.zeros(n_labels)
pred[np.where(p_y_pred > thres)] = 1
pred_roll.append(pred)
gt_roll.append(y)
pred_roll = np.array(pred_roll)
gt_roll = np.array(gt_roll)
# calculate prec, recall, fvalue
prec, recall, fvalue = prec_recall_fvalue(pred_roll, gt_roll, thres)
print prec, recall, fvalue
from hat.optimizers import SGD, Rmsprop
import hat.backend as K
import config as cfg
import prepare_dev_data as pp_dev_data
# hyper-params
agg_num = 11 # concatenate frames
hop = 5 # step_len
act = 'relu'
n_hid = 500
fold = 1
n_out = len(cfg.labels)
# prepare data
tr_X, tr_y, _ = pp_dev_data.GetAllData(cfg.dev_fe_mel_fd,
agg_num,
hop,
fold=None)
[batch_num, n_time, n_freq] = tr_X.shape
print tr_X.shape, tr_y.shape
# build model
seq = Sequential()
seq.add(InputLayer((n_time, n_freq)))
seq.add(Flatten()) # flatten to 2d: (n_time, n_freq) to 1d:(n_time*n_freq)
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
seq.add(Dense(n_hid, act=act))
seq.add(Dropout(0.1))
from hat.layers.core import InputLayer, Dense, Dropout, Lambda, Flatten from hat.layers.cnn import Convolution2D, Convolution1D from hat.layers.pool import Pool2D, GlobalMaxPool from hat.callbacks import Validation, SaveModel from hat.preprocessing import pad_trunc_seqs, sparse_to_categorical, reshape_3d_to_4d from hat.optimizers import Rmsprop import prepare_dev_data as pp_dev_data import config as cfg # hyper-params tr_fe_fd = cfg.dev_tr_fe_mel_fd max_len = 100 n_out = len(cfg.labels) # prepare data tr_X, tr_y = pp_dev_data.GetAllData(tr_fe_fd, max_len) tr_y = sparse_to_categorical(tr_y, n_out) print tr_X.shape, tr_y.shape (_, n_time, n_freq) = tr_X.shape # build model seq = Sequential() seq.add(InputLayer((n_time, n_freq))) seq.add(Flatten()) seq.add(Dropout(0.1)) seq.add(Dense(500, 'relu')) seq.add(Dropout(0.1)) seq.add(Dense(500, 'relu')) seq.add(Dropout(0.1)) seq.add(Dense(500, 'relu'))