aprender-core 0.29.3

pub(crate) use super::*;

#[test]
fn test_dropout_eval_mode() {
    let mut dropout = Dropout::new(0.5);
    dropout.eval();

    let x = Tensor::ones(&[10, 10]);
    let y = dropout.forward(&x);

    // In eval mode, output should equal input
    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropout_train_mode_zeros() {
    let dropout = Dropout::with_seed(0.5, 42);

    let x = Tensor::ones(&[100]);
    let y = dropout.forward(&x);

    // Should have some zeros
    let num_zeros = y.data().iter().filter(|&&v| v == 0.0).count();
    assert!(num_zeros > 0, "Expected some zeros in dropout output");
    assert!(num_zeros < 100, "Expected some non-zeros in dropout output");
}

#[test]
fn test_dropout_scaling() {
    let dropout = Dropout::with_seed(0.5, 42);

    let x = Tensor::ones(&[100]);
    let y = dropout.forward(&x);

    // Non-zero elements should be scaled by 2 (1 / (1 - 0.5))
    for &val in y.data() {
        assert!(val == 0.0 || (val - 2.0).abs() < 1e-5);
    }
}

#[test]
fn test_dropout_zero_probability() {
    let dropout = Dropout::new(0.0);

    let x = Tensor::ones(&[10, 10]);
    let y = dropout.forward(&x);

    // With p=0, output should equal input
    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropout_expected_value() {
    // With large samples, mean should be approximately preserved
    let dropout = Dropout::with_seed(0.3, 42);

    let x = Tensor::ones(&[10000]);
    let y = dropout.forward(&x);

    let mean: f32 = y.data().iter().sum::<f32>() / y.numel() as f32;

    // Expected value should be close to 1.0 (original value)
    assert!(
        (mean - 1.0).abs() < 0.1,
        "Mean {mean} should be close to 1.0"
    );
}

#[test]
fn test_dropout_reproducible() {
    let dropout1 = Dropout::with_seed(0.5, 42);
    let dropout2 = Dropout::with_seed(0.5, 42);

    let x = Tensor::ones(&[100]);
    let y1 = dropout1.forward(&x);
    let y2 = dropout2.forward(&x);

    assert_eq!(y1.data(), y2.data());
}

#[test]
fn test_dropout_train_eval_toggle() {
    let mut dropout = Dropout::new(0.5);

    assert!(dropout.training());

    dropout.eval();
    assert!(!dropout.training());

    dropout.train();
    assert!(dropout.training());
}

#[test]
#[should_panic(expected = "Dropout probability must be in [0, 1)")]
fn test_dropout_invalid_probability_high() {
    let _ = Dropout::new(1.0);
}

#[test]
#[should_panic(expected = "Dropout probability must be in [0, 1)")]
fn test_dropout_invalid_probability_negative() {
    let _ = Dropout::new(-0.1);
}

// Dropout2d tests

#[test]
fn test_dropout2d_eval_mode() {
    let mut dropout = Dropout2d::new(0.5);
    dropout.eval();

    let x = Tensor::ones(&[4, 64, 8, 8]);
    let y = dropout.forward(&x);

    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropout2d_shape() {
    let dropout = Dropout2d::with_seed(0.5, 42);

    let x = Tensor::ones(&[4, 64, 8, 8]);
    let y = dropout.forward(&x);

    assert_eq!(y.shape(), x.shape());
}

#[test]
fn test_dropout2d_drops_entire_channels() {
    let dropout = Dropout2d::with_seed(0.5, 42);

    let x = Tensor::ones(&[1, 16, 4, 4]); // 1 batch, 16 channels, 4x4
    let y = dropout.forward(&x);

    // Check that entire channels are either all zeros or all scaled
    let y_data = y.data();
    for c in 0..16 {
        let channel_start = c * 16;
        let channel_end = channel_start + 16;
        let channel_data = &y_data[channel_start..channel_end];

        // Either all zeros or all ~2.0 (scaled by 1/(1-0.5))
        let first_val = channel_data[0];
        for &val in channel_data {
            assert!(
                (val - first_val).abs() < 1e-5,
                "Channel should have uniform values, got {first_val} and {val}"
            );
        }
    }
}

#[test]
fn test_dropout2d_3d_input() {
    let dropout = Dropout2d::with_seed(0.5, 42);

    let x = Tensor::ones(&[4, 64, 100]); // 3D input (e.g., for Conv1d)
    let y = dropout.forward(&x);

    assert_eq!(y.shape(), &[4, 64, 100]);
}

#[test]
fn test_dropout2d_reproducible() {
    let dropout1 = Dropout2d::with_seed(0.5, 42);
    let dropout2 = Dropout2d::with_seed(0.5, 42);

    let x = Tensor::ones(&[4, 16, 8, 8]);
    let y1 = dropout1.forward(&x);
    let y2 = dropout2.forward(&x);

    assert_eq!(y1.data(), y2.data());
}

// AlphaDropout tests

#[test]
fn test_alpha_dropout_eval_mode() {
    let mut dropout = AlphaDropout::new(0.5);
    dropout.eval();

    let x = Tensor::ones(&[100]);
    let y = dropout.forward(&x);

    assert_eq!(y.data(), x.data());
}

#[test]
fn test_alpha_dropout_shape() {
    let dropout = AlphaDropout::with_seed(0.5, 42);

    let x = Tensor::ones(&[32, 64]);
    let y = dropout.forward(&x);

    assert_eq!(y.shape(), x.shape());
}

#[test]
fn test_alpha_dropout_not_zeros() {
    // AlphaDropout doesn't produce zeros, it uses negative saturation value
    let dropout = AlphaDropout::with_seed(0.5, 42);

    let x = Tensor::zeros(&[1000]);
    let y = dropout.forward(&x);

    // Some values should be non-zero (the dropped ones become negative saturation)
    let has_non_zero = y.data().iter().any(|&v| v != 0.0);
    assert!(
        has_non_zero,
        "AlphaDropout should produce non-zero dropped values"
    );
}

#[test]
fn test_alpha_dropout_train_eval_toggle() {
    let mut dropout = AlphaDropout::new(0.5);

    assert!(dropout.training());

    dropout.eval();
    assert!(!dropout.training());

    dropout.train();
    assert!(dropout.training());
}

// DropBlock tests

#[test]
fn test_dropblock_creation() {
    let db = DropBlock::new(3, 0.1);
    assert_eq!(db.block_size(), 3);
    assert_eq!(db.p(), 0.1);
}

#[test]
fn test_dropblock_eval_mode() {
    let mut db = DropBlock::new(3, 0.5);
    db.eval();

    let x = Tensor::ones(&[2, 4, 8, 8]);
    let y = db.forward(&x);

    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropblock_train_mode() {
    let db = DropBlock::with_seed(3, 0.3, 42);

    let x = Tensor::ones(&[1, 2, 8, 8]);
    let y = db.forward(&x);

    assert_eq!(y.shape(), x.shape());
    // Should have some zeros (blocks dropped)
    let num_zeros = y.data().iter().filter(|&&v| v == 0.0).count();
    assert!(num_zeros > 0);
}

#[test]
fn test_dropblock_train_eval_toggle() {
    let mut db = DropBlock::new(3, 0.1);

    assert!(db.training());
    db.eval();
    assert!(!db.training());
    db.train();
    assert!(db.training());
}

#[test]
fn test_dropblock_non_4d_fallback() {
    let db = DropBlock::with_seed(3, 0.3, 42);

    let x = Tensor::ones(&[10, 10]); // 2D, not 4D
    let y = db.forward(&x);

    assert_eq!(y.shape(), x.shape());
}

#[test]
fn test_dropblock_zero_prob() {
    let db = DropBlock::new(3, 0.0);

    let x = Tensor::ones(&[1, 2, 8, 8]);
    let y = db.forward(&x);

    assert_eq!(y.data(), x.data());
}

// DropConnect tests
#[test]
fn test_dropconnect_creation() {
    let dc = DropConnect::new(0.5);
    assert_eq!(dc.probability(), 0.5);
    assert!(dc.training());
}

#[test]
fn test_dropconnect_eval_mode() {
    let mut dc = DropConnect::new(0.5);
    dc.eval();

    let x = Tensor::ones(&[10, 10]);
    let y = dc.forward(&x);

    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropconnect_train_mode() {
    let dc = DropConnect::with_seed(0.5, 42);

    let x = Tensor::ones(&[100]);
    let y = dc.forward(&x);

    let num_zeros = y.data().iter().filter(|&&v| v == 0.0).count();
    assert!(num_zeros > 0);
    assert!(num_zeros < 100);
}

#[test]
fn test_dropconnect_apply_to_weights() {
    let dc = DropConnect::with_seed(0.5, 42);

    let weights = Tensor::ones(&[4, 4]);
    let masked = dc.apply_to_weights(&weights);

    assert_eq!(masked.shape(), weights.shape());
    let num_zeros = masked.data().iter().filter(|&&v| v == 0.0).count();
    assert!(num_zeros > 0);
}

#[test]
fn test_dropconnect_zero_prob() {
    let dc = DropConnect::new(0.0);

    let x = Tensor::ones(&[10]);
    let y = dc.forward(&x);

    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropconnect_train_eval_toggle() {
    let mut dc = DropConnect::new(0.5);

    assert!(dc.training());
    dc.eval();
    assert!(!dc.training());
    dc.train();
    assert!(dc.training());
}

// ==========================================================================
// Additional Coverage Tests
// ==========================================================================

#[test]
fn test_dropout_probability() {
    let dropout = Dropout::new(0.3);
    assert!((dropout.probability() - 0.3).abs() < 1e-6);
}

#[test]
fn test_dropout_debug() {
    let dropout = Dropout::new(0.5);
    let debug_str = format!("{:?}", dropout);
    assert!(debug_str.contains("Dropout"));
    assert!(debug_str.contains("0.5"));
}

#[test]
fn test_dropout2d_probability() {
    let dropout = Dropout2d::new(0.4);
    assert!((dropout.probability() - 0.4).abs() < 1e-6);
}

#[test]
fn test_dropout2d_zero_probability() {
    let dropout = Dropout2d::new(0.0);
    let x = Tensor::ones(&[2, 4, 8, 8]);
    let y = dropout.forward(&x);
    assert_eq!(y.data(), x.data());
}

#[test]
fn test_dropout2d_debug() {
    let dropout = Dropout2d::new(0.5);
    let debug_str = format!("{:?}", dropout);
    assert!(debug_str.contains("Dropout2d"));
}

#[test]
fn test_dropout2d_train_eval_toggle() {
    let mut dropout = Dropout2d::new(0.5);
    assert!(dropout.training());
    dropout.eval();
    assert!(!dropout.training());
    dropout.train();
    assert!(dropout.training());
}

#[test]
fn test_alpha_dropout_zero_probability() {
    let dropout = AlphaDropout::new(0.0);
    let x = Tensor::ones(&[100]);
    let y = dropout.forward(&x);
    assert_eq!(y.data(), x.data());
}

#[test]
fn test_alpha_dropout_debug() {
    let dropout = AlphaDropout::new(0.5);
    let debug_str = format!("{:?}", dropout);
    assert!(debug_str.contains("AlphaDropout"));
}

#[test]
fn test_dropblock_debug() {
    let db = DropBlock::new(3, 0.2);
    let debug_str = format!("{:?}", db);
    assert!(debug_str.contains("DropBlock"));
    assert!(debug_str.contains("block_size"));
}

#[test]
fn test_dropconnect_debug() {
    let dc = DropConnect::new(0.5);
    let debug_str = format!("{:?}", dc);
    assert!(debug_str.contains("DropConnect"));
}

#[test]
fn test_dropconnect_apply_to_weights_eval_mode() {
    let mut dc = DropConnect::new(0.5);
    dc.eval();

    let weights = Tensor::ones(&[4, 4]);
    let masked = dc.apply_to_weights(&weights);

    // In eval mode, should return unchanged
    assert_eq!(masked.data(), weights.data());
}

// =========================================================================
// FALSIFY-DO: dropout-v1.yaml contract (aprender Dropout)
//
// Five-Whys (PMAT-354):
//   Why 1: aprender had 20+ dropout unit tests but zero FALSIFY-DO-* tests
//   Why 2: unit tests verify train/eval modes, not mathematical invariants
//   Why 3: no mapping from dropout-v1.yaml to aprender test names
//   Why 4: aprender predates the provable-contracts YAML convention
//   Why 5: dropout was "obviously correct" (Bernoulli mask + scale)
//
// References:
//   - provable-contracts/contracts/dropout-v1.yaml
//   - Srivastava et al. (2014) "Dropout: A Simple Way to Prevent Overfitting"
// =========================================================================

/// FALSIFY-DO-001: Eval identity — dropout_eval(x) = x for all x
#[test]
fn falsify_do_001_eval_identity() {
    let mut dropout = Dropout::new(0.5);
    dropout.eval();

    let data = vec![1.0, -2.0, 3.5, 0.0, -0.001, 1e6, -1e6];
    let x = Tensor::new(&data, &[data.len()]);
    let y = dropout.forward(&x);

    for (i, (&orig, &out)) in data.iter().zip(y.data().iter()).enumerate() {
        assert_eq!(
            orig, out,
            "FALSIFIED DO-001: eval output[{i}] = {out}, expected {orig} (identity)"
        );
    }
}

/// FALSIFY-DO-002: Unbiased expectation — E[dropout_train(x, p)] ≈ x
///
/// Over many trials, the mean of dropout output should converge to input.
#[test]
fn falsify_do_002_unbiased_expectation() {
    let p = 0.3;
    let n_trials = 10_000;
    let input_val = 5.0;
    let input = Tensor::new(&[input_val; 8], &[8]);

    let mut sums = vec![0.0f64; 8];

    for trial in 0..n_trials {
        let dropout = Dropout::with_seed(p, trial);
        let y = dropout.forward(&input);
        for (i, &val) in y.data().iter().enumerate() {
            sums[i] += val as f64;
        }
    }

    for (i, &sum) in sums.iter().enumerate() {
        let mean = sum / n_trials as f64;
        let diff = (mean - input_val as f64).abs();
        assert!(
            diff < 0.15,
            "FALSIFIED DO-002: E[dropout(x)][{i}] = {mean:.4}, expected ≈ {input_val}, diff = {diff:.4}"
        );
    }
}

/// FALSIFY-DO-003: Shape preservation — shape(dropout(x)) = shape(x)
#[test]
fn falsify_do_003_shape_preservation() {
    for &shape in &[[1_usize, 4], [8, 16], [32, 64], [1, 1]] {
        let dropout = Dropout::with_seed(0.5, 42);
        let x = Tensor::ones(&shape);
        let y = dropout.forward(&x);
        assert_eq!(
            y.shape(),
            &shape,
            "FALSIFIED DO-003: output shape {:?}, expected {:?}",
            y.shape(),
            shape
        );
    }
}

/// FALSIFY-DO-004: Probability boundary — p=0 is identity, p=1.0 panics
#[test]
fn falsify_do_004_zero_probability_identity() {
    let dropout = Dropout::with_seed(0.0, 42);
    let data = vec![1.0, 2.0, 3.0, 4.0];
    let x = Tensor::new(&data, &[4]);
    let y = dropout.forward(&x);

    for (i, (&orig, &out)) in data.iter().zip(y.data().iter()).enumerate() {
        assert_eq!(
            orig, out,
            "FALSIFIED DO-004: p=0 output[{i}] = {out}, expected {orig}"
        );
    }
}

#[test]
#[should_panic(expected = "Dropout probability must be in [0, 1)")]
fn falsify_do_004_invalid_probability_panics() {
    let _dropout = Dropout::new(1.0);
}

#[path = "tests_dropconnect.rs"]
mod tests_dropconnect;