frame_max = max(IDs_utter, key=lambda x: x[1]) train_data_length = len(pickList) * VAL_SET_RATIO DATA_LAYER = [mem_pgram.shape[1]] LABEL_LAYER = DATA_LAYER LAYERS = DATA_LAYER + HIDDEN_LAYERS + LABEL_LAYER ######################## # Create lstm # ######################## print "Creating lstm..." nn = LSTM_net(LAYERS, batch_size=BATCH_SIZE, momentum_type=MOMENTUM_TYPE, act_type=ACT_FUNC, cost_type=COST_FUNC) ######################## # Train lstm # ######################## val_label_vec = None StateToVec = get_PhoneStateVec() PhoneState = load_liststateto48() PhoneIdx = load_dict_IdxPh48() prev_err = float('inf') prev_2 = float('inf') prev_3 = float('inf') cal_dev = 3
MEM_PGRAM = PGRAM_ROOT + DNN_MODEL + '_test.pgram' MEM_PGRAM_shape = (180406, 48) ######################## # load lstm open file # ######################## print "Loading lstm..." layers, W, Wi, Wf, Wo, b, bi, bf, bo = pickle.load( open(MODEL_ROOT + MODEL, 'rb')) nn = LSTM_net(layers, W, Wi, Wf, Wo, b, bi, bf, bo, batch_size=BATCH_SIZE, momentum_type=MOMENTUM_TYPE, act_type=ACT_FUNC, cost_type=COST_FUNC) #IDs,TEST_DATA,VAL_DATA = readfile_for_test( TEST_ROOT+TEST,1 ) print "Reading data..." mem_pgram = np.memmap(MEM_PGRAM, dtype='float32', mode='r', shape=MEM_PGRAM_shape) IDs = readID(PKL_ID)