def plot_sched(self, keys=None, figsize=None):
assert hasattr(self, 'hps'), '[ERROR] You must run ParamSchedCallback to plot the sched.'
keys = self.hps.keys() if keys is None else listify(keys)
rows,cols = (len(keys)+1)//2, min(2, len(keys))
figsize = figsize or (6*cols,4*rows)
_, axs = plt.subplots(rows, cols, figsize=figsize)
axs = axs.flatten() if len(keys) > 1 else listify(axs)
for p,ax in zip(keys, axs):
ax.plot(self.hps[p])
ax.set_ylabel(p)
def fit_one_cycle(self, epochs, lr_max=None, div=25., div_final=1e5, pct_start=0.25, wd=None,
moms=None, cbs=None, reset_opt=False):
lr = lr_max
if lr is None:
lr = self.lr
if reset_opt or not self.opt:
self.create_opt()
set_hyper(self.opt, 'lr', lr)
lr_max = np.array([p['lr'] for p in self.opt.param_groups])
scheds = {
'lr': combined_cos(pct_start, lr_max/div, lr_max, lr_max/div_final),
'momentum': combined_cos(pct_start, *(self.moms if moms is None else moms))
}
self.fit(epochs, cbs=[ParamScheduler(scheds)] + listify(cbs), reset_opt=reset_opt, wd=wd)
def set_hyper(opt, key, val):
if isinstance(val, slice):
if val.start:
val = even_mults(val.start, val.stop, len(opt.param_groups))
else:
val = [val.stop / 10] * (len(opt.param_groups) - 1) + (val.stop)
vs = listify(val)
if len(vs) == 1:
vs = vs * len(opt.param_groups)
assert len(vs) == len(
opt.param_groups
), f"Trying to set {len(vs)} values for {n} but there are {len(opt.param_groups)} parameter groups."
for v, p in zip(vs, opt.param_groups):
p[key] = v
def __init__(self, model, data, loss_func,
opt_func, lr=1e-2, wd=None, moms=(0.95, 0.85, 0.95),
metrics=None, cbs=None, cb_funcs=None,
splitter=get_trainable_params,
path=Path('.'), model_dir='models'):
self.model, self.data, self.loss_func = model, data, loss_func
self.opt_func, self.lr, self.metrics = opt_func, lr, metrics
self.splitter = splitter
self.wd, self.moms = wd, moms
self.path = path
self.model_dir = self.path / model_dir
self.opt = None
self.cbs = []
self.in_train = False
self.epoch = 0
self.epochs = 1
self.loss = tensor(0.)
self.logger = print
self.add_cbs([cb() for cb in self._default_cbs])
self.add_cbs(cbs)
self.add_cbs([cbf() for cbf in listify(cb_funcs)])
def remove_cbs(self, cbs):
for cb in listify(cbs): self.cbs.remove(cb)
def add_cbs(self, cbs):
for cb in listify(cbs): self.add_cb(cb)
def __init__(self, metrics, in_train):
self.metrics, self.in_train = listify(metrics), in_train
self.reset()