def __init__(self, args=DEFAULT_ARGS):
args = fill_undefined_args(args, DEFAULT_ARGS, to_bunch=True)
super().__init__()
self.args = args
if not hasattr(self.args, 'sim_x_seed'):
self.args.sim_x_seed = np.random.randint(1e6)
self.x_seed = self.args.sim_x_seed
self.activation = torch.tanh
self.tau_x = 10
# weights in the recurrent net
self.J = nn.Linear(self.args.N_sim, self.args.N_sim, bias=False)
self.S_u = nn.Linear(self.args.L + self.args.D,
self.args.N_sim,
bias=False)
self.S_ro = nn.Linear(self.args.N_sim, self.args.L)
# dealing with latent output
self.latency = self.args.sim_latency
self.latent_idx = -1
self.latent_arr = [None] * self.latency
self.latent_decay = self.args.latent_decay
self.latent_out = 0
def __init__(self, args=DEFAULT_ARGS):
args = fill_undefined_args(args, DEFAULT_ARGS, to_bunch=True)
super().__init__()
self.args = args
self.W_sim = nn.Sequential(
nn.Linear(self.args.L + self.args.D, 16, bias=self.args.bias),
nn.Tanh(), nn.Linear(16, self.args.L, bias=self.args.bias))
if args.sim_path is not None:
self.load_state_dict(torch.load(args.sim_path))
def __init__(self, args=DEFAULT_ARGS):
super().__init__()
args = fill_undefined_args(args, DEFAULT_ARGS, to_bunch=True)
self.args = args
if not hasattr(self.args, 'network_seed'):
self.args.network_seed = random.randrange(1e6)
self._init_vars()
if self.args.model_path is not None:
self.load_state_dict(torch.load(args.model_path))
self.latent = Latent(self.args.h_latency, self.args.latent_decay)
self.reset()
def __init__(self, args=DEFAULT_ARGS):
super().__init__()
self.args = fill_undefined_args(args, DEFAULT_ARGS, to_bunch=True)
if not hasattr(self.args, 'res_seed'):
self.args.res_seed = random.randrange(1e6)
if not hasattr(self.args, 'res_x_seed'):
self.args.res_x_seed = np.random.randint(1e6)
self.tau_x = 10
self._init_vars()
self.reset()
def __init__(self, args=DEFAULT_ARGS):
super().__init__()
args = fill_undefined_args(args, DEFAULT_ARGS, to_bunch=True)
self.args = args
if not hasattr(self.args, 'network_seed'):
self.args.network_seed = random.randrange(1e6)
self._init_vars()
self.log_h_yes = Latent(1, 0.95)
self.log_s_yes = Latent(1, 0.95)
self.log_conf = Latent(1, 0.95)
self.switch = False
self.reset()
def __init__(self, args=DEFAULT_ARGS):
super().__init__()
args = fill_undefined_args(args, DEFAULT_ARGS, to_bunch=True)
self.args = args
if not hasattr(self.args, 'network_seed'):
self.args.network_seed = random.randrange(1e6)
self._init_vars()
if self.args.model_path is not None:
self.load_state_dict(torch.load(args.model_path))
self.out_act = get_output_activation(self.args)
self.network_delay = args.network_delay
self.reset()
# assuming config is in the same folder as the model
config = get_config(args.model)
if args.noise != 0:
J = model['W_f.weight']
v = J.std()
shp = J.shape
model['W_f.weight'] += torch.normal(0, v * .5, shp)
J = model['W_ro.weight']
v = J.std()
shp = J.shape
model['W_ro.weight'] += torch.normal(0, v * .5, shp)
config = fill_undefined_args(args, config, overwrite_none=True)
net = load_model_path(args.model, config=config)
if args.test_all:
_, loss2 = test_model(net, config)
print('avg summed loss (all):', loss2)
if not args.no_plot:
data, loss = test_model(net, config, n_tests=6)
print('avg summed loss (plotted):', loss)
run_id = '/'.join(args.model.split('/')[-3:-1])
fig, ax = plt.subplots(2, 3, sharex=True, sharey=True, figsize=(12, 7))
def adjust_args(args):
# don't use logging.info before we initialize the logger!! or else stuff is gonna fail
# dealing with slurm. do this first!! before anything else other than seed setting, which we want to override
if args.slurm_id is not None:
from parameters import apply_parameters
args = apply_parameters(args.param_path, args)
# in case we are loading from a model
# if we don't use this we might end up with an error when loading model
if args.model_path is not None:
config = get_config(args.model_path)
args = fill_undefined_args(args, config, overwrite_none=True)
enforce_same = [
'N', 'D', 'L', 'Z', 'T', 'net', 'bias', 'use_reservoir'
]
for v in enforce_same:
if v in config and args.__dict__[v] != config[v]:
print(
f'Warning: based on config, changed {v} from {args.__dict__[v]} -> {config[v]}'
)
args.__dict__[v] = config[v]
# shortcut for specifying train everything including reservoir
if args.train_parts == ['all']:
args.train_parts = ['']
# # output activation depends on the task / dataset used
# if args.out_act is None:
# args.out_act = 'none'
# set the dataset
if 'goals' in args.dataset:
args.dset_type = 'goals'
elif 'copy' in args.dataset:
args.dset_type = 'copy'
else:
args.dset_type = 'unknown'
# use custom goals loss for goals dataset, override default loss fn
if args.dset_type == 'goals':
args.loss = 'goals'
args.argv = sys.argv
# setting seeds
if args.seed is None:
args.seed = random.randrange(1e6)
if args.network_seed is None:
args.network_seed = random.randrange(1e6)
if args.res_seed is None:
args.res_seed = random.randrange(1e6)
torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
# initializing logging
# do this last, because we will be logging previous parameters into the config file
if not args.no_log:
if args.slurm_id is not None:
log = log_this(args,
'logs',
os.path.join(
args.name.split('_')[0],
args.name.split('_')[1]),
checkpoints=args.log_checkpoint_models)
else:
log = log_this(args,
'logs',
args.name,
checkpoints=args.log_checkpoint_models)
logging.basicConfig(format='%(message)s',
filename=log.run_log,
level=logging.DEBUG)
console = logging.StreamHandler()
console.setLevel(logging.DEBUG)
logging.getLogger('').addHandler(console)
args.log = log
else:
logging.basicConfig(format='%(message)s', level=logging.DEBUG)
logging.info('NOT LOGGING THIS RUN.')
# logging, when loading models from paths
if args.model_path is not None:
logging.info(f'Using model path {args.model_path}')
logging.info(
f'Seeds:\n general: {args.seed}\n network: {args.network_seed}')
return args