reference = js["nodes_reference"]
base_nodes = map(int, js["nodes_base"].split(','))
stack_nodes = map(int, js["nodes_stack"].split(','))
final_nodes = [] #map(int, js["nodes_final"].split(','))
num_att = 4
num_output = 3
renetFile = None
if 'retrain' in js:
renetFile = HOME + 'NNs/' + js['retrain'] + '.p'
if mode == 'train':
tr = DataSet(tr_data, js['block'], feature_len)
# tr.set_t_scale(t_scale)
tr.set_num_output(num_output)
te = DataSet(te_data, js['block'], feature_len)
# te.set_t_scale(t_scale)
te.set_num_output(num_output)
else:
if mode == 'te':
te = DataSet(te_data, js['block'], feature_len)
else:
te = DataSet(tr_data, js['block'], feature_len)
te.set_num_output(num_output)
sz_in = te.sz
loop = js['loop']
print "input shape", sz_in, "LR", lr, 'feature', feature_len
if renetFile is not None:
iterations = 10000
js = Utils.load_json_file(config_file)
dtype = torch.float
device = torch.device("cpu")
# device = torch.device("cuda:0") # Uncomment this to run on GPU
cfg = Config(config_file)
tr = DataSet(cfg.tr_data, cfg.memory_size, cfg.feature_len)
te = DataSet(cfg.te_data, cfg.memory_size, cfg.feature_len)
tr.set_net_type(cfg.net_type)
te.set_net_type(cfg.net_type)
tr.set_t_scale(cfg.t_scale)
te.set_t_scale(cfg.t_scale)
tr.set_num_output(cfg.num_output)
te.set_num_output(cfg.num_output)
att = te.sz[1]
D_in = cfg.feature_len * att
D_out = cfg.num_output
tr_pre_data = tr.prepare(multi=1)
for b in tr_pre_data:
d = torch.from_numpy(b[0]).type(torch.FloatTensor)
t = torch.from_numpy(b[1]).type(torch.FloatTensor)
te_pre_data = te.prepare(multi=1)
for b in te_pre_data:
de = torch.from_numpy(b[0]).type(torch.FloatTensor)
te = torch.from_numpy(b[1]).type(torch.FloatTensor)