Is it possible to have a costume mean squared error function with sample weight for each observation?
I am able to utilize the standard fastai training loop and I am able to implement this costume loss in PyTorch.
How to put that to fastai learner object on tabular data?
I know keras has this already implemented in the .fit method where sample_weight argument present.
def weighted_mse_loss(input, target, weight):
return torch.sum(weight * (input - target) ** 2)
from fastai.tabular.all import *
import seaborn as sns
df = sns.load_dataset('tips')
df = df.assign(sample_weight = np.random.normal(size = df.shape[0], loc = 10, scale = 2))
y = ['total_bill']
cont = ['tip']
cat = ['sex', 'smoker', 'day', 'time', 'size']
procs = [Normalize, Categorify]
df["Y"] = np.log(df[y] + 1)
MIN = df["Y"].min()
MAX = df["Y"].max()
splits = RandomSplitter(valid_pct=0.2)(range_of(df))
to = TabularPandas(
df,
procs=procs,
cat_names=cat,
cont_names=cont,
y_names="Y",
splits=splits,
y_block=RegressionBlock(n_out = 1),
)
dls = to.dataloaders(
bs=64, shuffle_train=True
)
config = tabular_config(
embed_p=0.05,
y_range=[0, MAX * 1.1],
bn_final=False,
ps=[0.05, 0.05, 0.05],
)
learner = tabular_learner(
dls,
layers=[1000, 500, 250],
config=config,
wd=0.2,
metrics=[rmse,],
)
learner.fit_one_cycle(40, lr_max = 0.01,
wd = 0.1)