Learn.fit_fc() won't complete the given total epochs without any errors

I’ve got something that will work in the meantime @MichaelScofield

def FlatCosAnnealSchedulerWOverSample(learn, lr:float=4e-3, tot_epochs:int=1, moms:Floats=(0.95,0.999),
                          start_pct:float=0.72, curve='cosine', weights:torch.Tensor=None):
  "Manage FCFit trainnig as found in the ImageNette experiments"
  ds, dl = learn.data.train_ds, learn.data.train_dl
  labels = ds.y.items
  assert np.issubdtype(labels.dtype, np.integer), "Can only oversample integer values"
  _,label_counts = np.unique(labels,return_counts=True)
  if weights is None: weights = torch.DoubleTensor((1/label_counts)[labels])
  total_len_oversample = int(learn.data.c*np.max(label_counts))
  sampler = WeightedRandomSampler(weights, total_len_oversample)
  learn.data.train_dl = dl.new(shuffle=False, sampler=sampler)
  n = len(learn.data.train_dl)
  anneal_start = int(n * tot_epochs * start_pct)
  batch_finish = ((n * tot_epochs) - anneal_start)
  if curve=="cosine":        curve_type=annealing_cos
  elif curve=="linear":      curve_type=annealing_linear
  elif curve=="exponential": curve_type=annealing_exp
  else: raiseValueError(f"annealing type not supported {curve}")
  phase0 = TrainingPhase(anneal_start).schedule_hp('lr', lr).schedule_hp('mom', moms[0])
  phase1 = TrainingPhase(batch_finish).schedule_hp('lr', lr, anneal=curve_type).schedule_hp('mom', moms[1])
  phases = [phase0, phase1]
  return GeneralScheduler(learn, phases)

 def fit_fcO(learn:Learner, tot_epochs:int=1, lr:float=defaults.lr,  moms:Tuple[float,float]=(0.95,0.85), start_pct:float=0.72,
                  wd:float=None, callbacks:Optional[CallbackList]=None)->None:
    "Fit a model with Flat Cosine Annealing"
    max_lr = learn.lr_range(lr)
    callbacks = listify(callbacks)
    callbacks.append(FlatCosAnnealSchedulerWOverSample(learn, lr, moms=moms, start_pct=start_pct, tot_epochs=tot_epochs))
    learn.fit(tot_epochs, max_lr, wd=wd, callbacks=callbacks)

To use:

fit_fcO(learn, n_epochs, lr)

I’m unsure quite where it needs to fall in that order bit, I need to do more work on that but this is what I can fit in for tonight :slight_smile: Thanks for the assist @ilovescience


I will run tests with the new one and let you two know if it’s the answer we are looking for. Thank you.

1 Like

@muellerzr sorry to say this but fit_fcO still ran into the bug in my latest test…