if __name__ == '__main__':
context.set_context(mode=context.GRAPH_MODE,
save_graphs=False,
device_target="Ascend")
device_id = int(os.getenv('DEVICE_ID'))
context.set_context(device_id=device_id)
if cfg.preprocess == 'true':
print("============== Starting Data Pre-processing ==============")
if os.path.exists(cfg.preprocess_path):
shutil.rmtree(cfg.preprocess_path)
os.mkdir(cfg.preprocess_path)
convert_to_mindrecord(cfg.embed_size, cfg.data_path,
cfg.preprocess_path, cfg.emb_path)
if cfg.cell == "vanilla":
print(
"============ Precision is lower than expected when using vanilla RNN architecture ==========="
)
embedding_table = np.loadtxt(
os.path.join(cfg.preprocess_path, "weight.txt")).astype(np.float32)
network = textrcnn(weight=Tensor(embedding_table),
vocab_size=embedding_table.shape[0],
cell=cfg.cell,
batch_size=cfg.batch_size)
ds_train = create_dataset(cfg.preprocess_path, cfg.batch_size, True)
help=
'the target device to run, support "GPU", "CPU". Default: "Ascend".')
args = parser.parse_args()
context.set_context(mode=context.GRAPH_MODE,
save_graphs=False,
device_target=args.device_target)
if args.device_target == 'Ascend':
cfg = lstm_cfg_ascend
else:
cfg = lstm_cfg
if args.preprocess == "true":
print("============== Starting Data Pre-processing ==============")
convert_to_mindrecord(cfg.embed_size, args.aclimdb_path,
args.preprocess_path, args.glove_path)
embedding_table = np.loadtxt(
os.path.join(args.preprocess_path, "weight.txt")).astype(np.float32)
# DynamicRNN in this network on Ascend platform only support the condition that the shape of input_size
# and hiddle_size is multiples of 16, this problem will be solved later.
if args.device_target == 'Ascend':
pad_num = int(np.ceil(cfg.embed_size / 16) * 16 - cfg.embed_size)
if pad_num > 0:
embedding_table = np.pad(embedding_table, [(0, 0), (0, pad_num)],
'constant')
cfg.embed_size = int(np.ceil(cfg.embed_size / 16) * 16)
network = SentimentNet(vocab_size=embedding_table.shape[0],
embed_size=cfg.embed_size,
num_hiddens=cfg.num_hiddens,
num_layers=cfg.num_layers,
mode=context.GRAPH_MODE,
save_graphs=False,
device_target=args.device_target)
# import moxing as mox
# mox.file.copy_parallel(src_url='s3://zhengnj-course/lstm/glove', dst_url=args.glove_path)
if args.device_target == 'Ascend':
cfg = lstm_cfg_ascend
else:
cfg = lstm_cfg
if args.preprocess == "true":
print("============== Starting Data Pre-processing ==============")
convert_to_mindrecord(cfg.embed_size, args.data_url, args.preprocess_path, args.glove_path)
embedding_table = np.loadtxt(os.path.join(args.preprocess_path, "weight.txt")).astype(np.float32)
# DynamicRNN in this network on Ascend platform only support the condition that the shape of input_size
# and hiddle_size is multiples of 16, this problem will be solved later.
if args.device_target == 'Ascend':
pad_num = int(np.ceil(cfg.embed_size / 16) * 16 - cfg.embed_size)
if pad_num > 0:
embedding_table = np.pad(embedding_table, [(0, 0), (0, pad_num)], 'constant')
cfg.embed_size = int(np.ceil(cfg.embed_size / 16) * 16)
network = SentimentNet(vocab_size=embedding_table.shape[0],
embed_size=cfg.embed_size,
num_hiddens=cfg.num_hiddens,
num_layers=cfg.num_layers,
bidirectional=cfg.bidirectional,
num_classes=cfg.num_classes,
