def create_dfs(data_folder, checkpointfolder):
if os.path.isfile(checkpointfolder + "train_df.pkl"):
assert os.path.isfile(checkpointfolder + "test_df.pkl")
print("dfs already exist")
return
else:
print("Creating dfs")
#Data extraction and preparation
X_train_path = data_folder + "X_train.npy"
X_train = np.array(utils.load_npy(X_train_path))
X_test_path = data_folder + "X_test.npy"
X_test = np.array(utils.load_npy(X_test_path))
y_train_path = data_folder + "y_train.npy"
y_train = utils.load_npy(y_train_path)
y_train = np.array([utils.collapse_num_labels(y) for y in y_train])
y_test_path = data_folder + "y_test.npy"
y_test = utils.load_npy(y_test_path)
y_test = np.array([utils.collapse_num_labels(y) for y in y_test])
#SAVE MEMORY??
X_train, X_test = normalize_data(X_train, X_test, data_folder)
train_df = create_phon_df(X_train, y_train)
del (X_train, y_train)
train_df.to_pickle(checkpointfolder + "train_df.pkl")
test_df = create_phon_df(X_test, y_test)
del (X_test, y_test)
test_df.to_pickle(checkpointfolder + "test_df.pkl")
n_categories = 40 + 1
#Params
dropout = 0.5
n_feats = 600
n_hidden = 128
#hyper-params
lr = 0.001
epochs = 80
batchsize = 8
use_ctc = True
print("Load data")
X_test = np.array(
utils.load_npy(checkpointfolder_specific + "X_test_0.npy"))
y_test = np.array(
utils.load_npy(checkpointfolder_specific + "y_test_0.npy"))
# X_train, y_train = utils.get_mini_dataset(X_train, y_train, 7000)
X_test, y_test = utils.get_mini_dataset(X_test, y_test, 3000)
#LSTM create and feed
rnn = BLSTM(n_feats=n_feats,
n_hidden=n_hidden,
n_classes=n_categories,
logsdir=logdir)
rnn.load_weights(logdir)
if ngram == 3:
delta_err = p_err_mean - epoch_loss_train_mean
p_err_mean = epoch_loss_train_mean
self.learning_rate = self.lr_scheduler(learning_rate, e, delta_err)
if self.logsdir is not None:
self.save_weights()
return
if __name__ == "__main__":
epochs = 10
lr = 0.01
rnn = RNN_CTC_base(n_feats=600, n_classes=41, logsdir="./RNN_test/")
X_train = utils.load_npy('../../data/RBM_hann_v2/X_train.npy')
X_test = utils.load_npy('../../data/RBM_hann_v2/X_test.npy')
y_train = utils.load_npy('../../data/RBM_hann_v2/y_train.npy')
y_test = utils.load_npy('../../data/RBM_hann_v2/y_test.npy')
y_train = [utils.collapse_num_labels(y) for y in y_train]
y_test = [utils.collapse_num_labels(y) for y in y_test]
y_train = utils.to_ctc(y_train)
y_test = utils.to_ctc(y_test)
rnn.set_data(X_train, X_test, y_train, y_test)
del (X_train, y_train, X_test, y_test)
rnn.fit(n_epochs=epochs,
# l1 = 0.00
# dropout = 0.05
# logdir_feats = root_dir + w + "/v2/400_600/v1/600_300/300_200/relurelu_v1/200_100/relurelu_v1/"
i = 4
weights_dir = logdir_feats + model + "_" + str(n_visibles) + "_" + str(n_hidden) + "/"
out_feats_dir = weights_dir + model + "_v" + str(v) + "/"
lr = 0.001 #0.0005087695972171545
## """ TRAIN """
if i == 1:
from utils import plot_weights, plot_input_bias, plot_hidden_bias
n_epochs = int(200//dropout)
print("train for: ", str(n_epochs))
X_train = utils.load_npy(logdir_feats + "X_train.npy")
train_matrix = []
for x in X_train:
train_matrix.extend(x)
train_matrix = np.array(train_matrix)
del(X_train)
if normalize:
train_matrix = normalize_fn(train_matrix, logdir_feats)
# try:
# mean = utils.open_pickle(logdir_feats + "mean.pkl")
# except:
# mean = np.array(train_matrix).mean(0)
# utils.save_pickle(logdir_feats + "mean.pkl", mean)
# try:
# std = utils.open_pickle(logdir_feats + "std.pkl")
# folder_name = "./PreprocessedData/MFCCs_delta1/hann/"
# logdir = "./BLSTM/MFCC/Hann/"
folder_name = "./PreprocessedData/RBMs_longtrain/hann/v2/400_50/"
logdir = "./BLSTM/rbm_50_deltas_argmax/"
epochs = 8
lr = 0.001
batchsize = 8
dropout = 0.85
n_feats = 102
n_hidden = 128
n_classes = 39 + 1
X_train_path = folder_name + "X_train.npy"
X_train = np.array(utils.load_npy(X_train_path))
X_test_path = folder_name + "X_test.npy"
X_test = np.array(utils.load_npy(X_test_path))
y_train_path = folder_name + "y_train.npy"
y_train = np.array(utils.load_npy(y_train_path))
y_test_path = folder_name + "y_test.npy"
y_test = np.array(utils.load_npy(y_test_path))
# X_train,y_train = utils.get_mini_dataset(X_train,y_train,500)
# X_test, y_test = utils.get_mini_dataset(X_test, y_test,100)
X_train, X_test = normalize_data(X_train, X_test, folder_name)
y_train = utils.remove_q(
utils.to_ctc(np.array([utils.collapse_num_labels(y)