Merge pull request #1 from Shashank-Tripathi-07/mcdc-01-core-tensor-autograd

test(tinytorch): regression coverage for core tensor/layers/autograd/optimizer gaps

Author: Shashank-Tripathi-07

tinytorch/tests/01_tensor/test_tensor_core.py

@@ -568,6 +568,42 @@ def test_tensor_from_tensor_list(self):
         assert np.array_equal(stacked.data[0], t1.data)
         assert np.array_equal(stacked.data[1], t2.data)
 
+    def test_tensor_from_empty_list(self):
+        """Tensor([]) constructs successfully with shape (0,)."""
+        t = Tensor([])
+        assert t.shape == (0,)
+
+    def test_matmul_1d_self_2d_other(self):
+        """matmul with a 1D self tensor and a 2D other tensor works."""
+        t1 = Tensor([1, 2, 3])
+        t2 = Tensor(rng.standard_normal((3, 4)))
+        result = t1.matmul(t2)
+        assert result.shape == (4,)
+
+    def test_reshape_scalar_arg(self):
+        """reshape accepts a single scalar argument, not just a tuple."""
+        t = Tensor(np.arange(6))
+        reshaped = t.reshape(6)
+        assert reshaped.shape == (6,)
+
+    def test_transpose_only_dim0_raises(self):
+        """transpose() with only dim0 specified (dim1 omitted) raises ValueError."""
+        t = Tensor(np.arange(6).reshape(2, 3))
+        with pytest.raises(ValueError):
+            t.transpose(dim0=0)
+
+    def test_reshape_multiple_negative_one_raises(self):
+        """reshape() with more than one -1 dimension raises ValueError."""
+        t = Tensor(np.arange(6).reshape(2, 3))
+        with pytest.raises(ValueError):
+            t.reshape(-1, -1)
+
+    def test_reshape_indivisible_negative_one_raises(self):
+        """reshape() raises ValueError when size isn't divisible for -1 inference."""
+        t = Tensor(np.arange(7))
+        with pytest.raises(ValueError):
+            t.reshape(2, -1)
+
 
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])

tinytorch/tests/03_layers/test_layers_core.py

@@ -28,7 +28,7 @@
 
 sys.path.insert(0, str(Path(__file__).parent.parent.parent))
 
-from tinytorch.core.layers import Layer
+from tinytorch.core.layers import Layer, Sequential, Linear, ReLU, Dropout
 from tinytorch.core.tensor import Tensor
 
 
@@ -488,5 +488,74 @@ def output_shape(self, input_shape):
             )
 
 
+class TestSequentialRealImplementation:
+    """
+    Test the real tinytorch.core.layers.Sequential container.
+
+    CONCEPT: Sequential accepts layers either as separate positional
+    arguments (Sequential(l1, l2)) or as a single list (Sequential([l1, l2])).
+    Both forms must produce a working, correctly shaped forward pass.
+    """
+
+    def test_sequential_positional_args_construction_and_shape(self):
+        """
+        WHAT: Sequential(Linear(2, 2), ReLU()) built from positional args.
+
+        WHY: Students commonly write Sequential(layer1, layer2, ...) the
+        same way they would in PyTorch, without wrapping layers in a list.
+
+        STUDENT LEARNING: The real Sequential supports both call styles.
+        """
+        model = Sequential(Linear(2, 2), ReLU())
+        assert len(model.layers) == 2
+
+        x = Tensor(np.array([[1.0, -1.0]]))
+        output = model(x)
+        assert output.shape == (1, 2), (
+            f"Sequential(Linear(2,2), ReLU()) forward shape wrong.\n"
+            f"  Expected: (1, 2)\n"
+            f"  Got: {output.shape}"
+        )
+
+
+class TestDropoutLayer:
+    """
+    Test the Dropout layer's validation and training/inference behavior.
+
+    CONCEPT: Dropout only zeros elements during training and only when
+    p > 0. It must also reject invalid probabilities at construction.
+    """
+
+    def test_dropout_valid_construction(self):
+        """Dropout(0.5) constructs successfully and stores p."""
+        dropout = Dropout(0.5)
+        assert dropout.p == 0.5
+
+    def test_dropout_negative_p_raises(self):
+        """Dropout(-0.1) raises ValueError."""
+        with pytest.raises(ValueError):
+            Dropout(-0.1)
+
+    def test_dropout_p_above_one_raises(self):
+        """Dropout(1.1) raises ValueError."""
+        with pytest.raises(ValueError):
+            Dropout(1.1)
+
+    def test_should_apply_dropout_training_and_p_positive(self):
+        """_should_apply_dropout is True when training=True and p > 0."""
+        dropout = Dropout(0.5)
+        assert dropout._should_apply_dropout(training=True) is True
+
+    def test_should_apply_dropout_false_when_not_training(self):
+        """_should_apply_dropout is False when training=False."""
+        dropout = Dropout(0.5)
+        assert dropout._should_apply_dropout(training=False) is False
+
+    def test_should_apply_dropout_false_when_p_zero(self):
+        """_should_apply_dropout is False when p=0, even during training."""
+        dropout = Dropout(0.0)
+        assert dropout._should_apply_dropout(training=True) is False
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])

tinytorch/tests/06_autograd/test_autograd_gradient_flow.py

@@ -14,7 +14,7 @@
 sys.path.insert(0, str(Path(__file__).parent.parent.parent))
 
 from tinytorch.core.tensor import Tensor
-from tinytorch.core.autograd import enable_autograd
+from tinytorch.core.autograd import enable_autograd, Function
 from tinytorch.core.activations import GELU
 # Try to import transformer for mean/sqrt monkey-patches (Module 13)
 # This is optional - tests will skip if not available
@@ -148,6 +148,63 @@ def test_reshape_gradient_flow():
     print("✅ Reshape gradient flow works correctly")
 
 
+def test_sum_axis0_backward():
+    """Test that sum(axis=0).backward(gradient) propagates the gradient to every row."""
+    print("Testing sum(axis=0) backward pass...")
+
+    x = Tensor(np.random.randn(3, 4), requires_grad=True)
+    y = x.sum(axis=0)
+    assert y.shape == (4,)
+
+    y.backward(np.ones(4))
+
+    assert x.grad.shape == (3, 4), "Gradient shape should match input shape"
+    assert np.allclose(x.grad, 1.0), "Every element's gradient should be 1.0"
+
+    print("✅ sum(axis=0) backward pass correct")
+
+
+def test_sum_axis1_backward():
+    """Test that sum(axis=1).backward(gradient) propagates the gradient to every column."""
+    print("Testing sum(axis=1) backward pass...")
+
+    x = Tensor(np.random.randn(3, 4), requires_grad=True)
+    y = x.sum(axis=1)
+    assert y.shape == (3,)
+
+    y.backward(np.ones(3))
+
+    assert x.grad.shape == (3, 4), "Gradient shape should match input shape"
+    assert np.allclose(x.grad, 1.0), "Every element's gradient should be 1.0"
+
+    print("✅ sum(axis=1) backward pass correct")
+
+
+def test_function_apply_not_implemented():
+    """Test that the base Function.apply() raises NotImplementedError."""
+    print("Testing bare Function.apply()...")
+
+    x = Tensor(np.array([1.0, 2.0]))
+    fn = Function(x)
+
+    with pytest.raises(NotImplementedError):
+        fn.apply(1.0)
+
+    print("✅ Function.apply() raises NotImplementedError as expected")
+
+
+def test_backward_without_gradient_on_nonscalar_raises():
+    """Test that backward() with no gradient argument on a non-scalar tensor raises ValueError."""
+    print("Testing backward() without gradient on non-scalar tensor...")
+
+    x = Tensor(np.random.randn(3, 4), requires_grad=True)
+
+    with pytest.raises(ValueError):
+        x.backward()
+
+    print("✅ backward() without gradient on non-scalar tensor raises ValueError")
+
+
 if __name__ == "__main__":
     print("\n" + "="*70)
     print("GRADIENT FLOW TEST SUITE")

tinytorch/tests/07_optimizers/test_optimizer_core.py

@@ -283,5 +283,58 @@ def test_optimizer_updates_all_parameters(self):
             )
 
 
+class TestSGDMomentumState:
+    """
+    Test SGD momentum state introspection and checkpointing.
+
+    CONCEPT: has_momentum(), get_momentum_state(), and set_momentum_state()
+    give checkpointing code (Module 08) a safe API to save and restore
+    momentum buffers without using hasattr() checks.
+    """
+
+    def test_has_momentum_false_without_momentum(self):
+        """has_momentum() returns False when momentum=0.0."""
+        param = Tensor([1.0, 2.0, 3.0], requires_grad=True)
+        optimizer = SGD([param], lr=0.1, momentum=0.0)
+        assert optimizer.has_momentum() is False
+
+    def test_has_momentum_true_with_momentum(self):
+        """has_momentum() returns True when momentum > 0."""
+        param = Tensor([1.0, 2.0, 3.0], requires_grad=True)
+        optimizer = SGD([param], lr=0.1, momentum=0.9)
+        assert optimizer.has_momentum() is True
+
+    def test_momentum_state_round_trip(self):
+        """
+        WHAT: Save momentum state from one SGD optimizer and restore it
+        into a fresh SGD optimizer built on equivalent parameters.
+
+        WHY: Training checkpoints must be able to resume momentum exactly,
+        otherwise resumed training diverges from an uninterrupted run.
+        """
+        param = Tensor([1.0, 2.0, 3.0], requires_grad=True)
+        optimizer = SGD([param], lr=0.1, momentum=0.9)
+
+        param.grad = np.array([1.0, 1.0, 1.0])
+        optimizer.step()
+
+        saved_state = optimizer.get_momentum_state()
+
+        # Fresh optimizer with the same params structure (new tensors, same shape)
+        fresh_param = Tensor([1.0, 2.0, 3.0], requires_grad=True)
+        fresh_optimizer = SGD([fresh_param], lr=0.1, momentum=0.9)
+        fresh_optimizer.set_momentum_state(saved_state)
+
+        restored_state = fresh_optimizer.get_momentum_state()
+
+        assert len(restored_state) == len(saved_state)
+        for original_buf, restored_buf in zip(saved_state, restored_state):
+            assert np.array_equal(original_buf, restored_buf), (
+                f"Restored momentum buffer does not match saved state.\n"
+                f"  Saved: {original_buf}\n"
+                f"  Restored: {restored_buf}"
+            )
+
+
 if __name__ == "__main__":
     pytest.main([__file__, "-v"])