def download_model(model_name, logger=None):
dir_path = os.path.dirname(os.path.realpath(__file__))
model_path = os.path.join(dir_path, 'model')
if logger is not None:
logger.info('Downloading model %s... into path %s' %
(model_name, model_path))
return modelzoo.download_model(args.model, os.path.join(dir_path, 'model'))
def load_pretrained_resnext_to_unext101_64_4d(
ctx,
need_download=False,
fine_tune=False,
migrate_input_norm=True,
migration_list='layer_migration_list.csv'):
# Note: fine-tune here is wheather tune pre-trained part
if need_download:
zoo.download_model('imagenet1k-resnext-101-64x4d')
unext_param_list = []
resnext_params_list = []
with open(migration_list, newline='') as csvfile:
list_reader = csv.reader(csvfile, delimiter=',')
for row in list_reader:
unext_param_list.append(row[0])
resnext_params_list.append(row[1])
if not migrate_input_norm:
unext_param_list = unext_param_list[4:len(unext_param_list)]
resnext_params_list = resnext_params_list[4:len(resnext_params_list)]
migration_dict = dict(zip(unext_param_list, resnext_params_list))
#print(unext_param_list)
#print(resnext_params_list)
#print(migration_dict)
#sym, arg_params, aux_params = mx.model.load_checkpoint('imagenet1k-resnext-101-64x4d', 0)
model = unext.unext101_64x4d()
model_params = model.collect_params()
migration_params = mx.ndarray.load(
'imagenet1k-resnext-101-64x4d-0000.params')
for key in migration_dict:
model_params[key]._load_init(migration_params[migration_dict[key]],
ctx)
if not fine_tune:
for param in unext_param_list:
model_params[param].grad_req = 'null'
# for param in model_params.values():
# param.grad_req = 'null'
return model
def load_pretrained_resnext101_64_4d(ctx,
need_download = True):
if need_download:
zoo.download_model('imagenet1k-resnext-101-64x4d')
symbol_path = 'imagenet1k-resnext-101-64x4d-symbol.json'
param_path = 'imagenet1k-resnext-101-64x4d-0000.params'
sym = mx.sym.load(symbol_path)
inter = sym.get_internals()
# print(inter)
new_sym = inter['flatten0_output']
inputsym = mx.sym.var('data', dtype=mx.base.mx_real_t)
model = mx.gluon.nn.SymbolBlock(outputs=new_sym, inputs=inputsym)
model.collect_params().load(param_path, ctx=ctx, allow_missing=False, ignore_extra=True)
return model