"""Complete one function at a time, then run python check_learner.py.
Public tests assess YOUR functions. Unimplemented stages fail intentionally.
"""
import torch
from torch import nn
def predict_batch(x, w, b):
"""Stage 1: X[N,D], W[D,1], b[1] -> prediction[N,1]."""
raise NotImplementedError('Implement the affine map.')
def fit_standardizer(x_train):
"""Stage 2: return (mean[D], scale[D]); population std, clamp >= 1e-8."""
raise NotImplementedError('Fit using only the supplied training rows.')
def make_classifier():
"""Stage 3: 2 -> 16 -> 1 MLP with ReLU, returning RAW logits [N,1]."""
raise NotImplementedError('Build an nn.Sequential.')
def training_step(model, optimizer, xb, yb):
"""Stage 4: zero gradients, forward, BCE logits loss, backward, step; return float loss."""
raise NotImplementedError('Implement one optimizer update.')
def infer_probabilities(model, xb):
"""Stage 5: eval mode, no gradient recording, sigmoid of raw logits [N,1]."""
raise NotImplementedError('Keep inference free of autograd graphs.')
