name = 'new_baseline_hengs_LessAugs_211_Mu10_Wd0_fit160epochs'
logger = CSVLogger(learn, name)
# learn.unfreeze()
# In[ ]:
# learn.fit_one_cycle(
# 64,
# max_lr=.01,
# wd=0.,
# pct_start=0.0,
# div_factor=100,
# callbacks=[logger, SaveModelCallback(learn, monitor='metric_tot', mode='max', name=name), MixUpCallback(learn, alpha=1.)]
# )
learn.fit(
160,
lr=.05,
wd=0.,
callbacks=[
logger,
SaveModelCallback(learn, monitor='metric_tot', mode='max', name=name),
# MixUpCallback(learn, alpha=1.),
ReduceLROnPlateauCallback(learn, patience=5, factor=0.5, min_lr=.0001)
]
)
# ==================================================================
classifier = Seresnext50MishFrac()
logging_name = 'new_baseline_seresnext50_Mish_Frac'
learn = Learner(
data_bunch,
classifier,
loss_func=Loss_combine_weighted_v2(),
opt_func=Over9000,
metrics=[Metric_grapheme(), Metric_vowel(), Metric_consonant(), Metric_tot()]
)
logger = CSVLogger(learn, logging_name)
learn.clip_grad = 1.0
# ==================================================================
learn.fit_one_cycle(
120,
max_lr=1e-2,
wd=0.,
pct_start=0.0,
div_factor=100,
callbacks=[logger, SaveModelCallback(learn, monitor='metric_tot', mode='max', name=logging_name), MixUpCallback(learn, alpha=1)]
)
# =========================================================================================
classifier = mdl_ResDenHybrid()
class SGD_m5(SGD):
def __init__(self, *args, **kwargs):
super().__init__(momentum=0.5, *args, **kwargs)
learn = Learner(
data_bunch,
classifier,
loss_func=Loss_combine_weighted_v2(),
opt_func=SGD_m5,
metrics=[Metric_grapheme(), Metric_vowel(), Metric_consonant(), Metric_tot()]
)
name = 'mdl_ResDenHybrid_sgd_lessaugs_mucm_fixed_raw_onecycle_fld1of5'
logger = CSVLogger(learn, name)
# =========================================================================================
learn.fit_one_cycle(
160,
max_lr=0.05,
wd=0.0,
pct_start=0.0,
div_factor=50.,
final_div=100.,
callbacks=[logger, SaveModelCallback(learn, monitor='metric_tot', mode='max', name=name), MuCmCallback(learn)]
)
learn = Learner(
data_bunch,
classifier,
#loss_func=Loss_single(),
loss_func=AdvancedLoss_Single(),
opt_func=Over9000,
metrics=[Metric_grapheme()]
)
logger = CSVLogger(learn, logging_name)
learn.clip_grad = 1.0
# learn.split([classifier.cls])
learn.unfreeze()
# In[9]:
learn.fit_one_cycle(
64,
# max_lr=slice(0.2e-2, 1e-2),
max_lr=1e-2,
wd=0.,
pct_start=0.0,
div_factor=100,
callbacks=[logger, SaveModelCallback(learn, monitor='metric_idx', mode='max', name=logging_name)]
)
