brainstate 0.1.10__py2.py3-none-any.whl → 0.2.0__py2.py3-none-any.whl

brainstate/nn/_conv_test.py

@@ -1,238 +1,849 @@
+# Copyright 2024 BrainX Ecosystem Limited. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
 # -*- coding: utf-8 -*-
 
+import unittest
+
+import jax
 import jax.numpy as jnp
-import pytest
-from absl.testing import absltest
-from absl.testing import parameterized
 
 import brainstate
 
 
-class TestConv(parameterized.TestCase):
-    def test_Conv2D_img(self):
-        img = jnp.zeros((2, 200, 198, 4))
-        for k in range(4):
-            x = 30 + 60 * k
-            y = 20 + 60 * k
-            img = img.at[0, x:x + 10, y:y + 10, k].set(1.0)
-            img = img.at[1, x:x + 20, y:y + 20, k].set(3.0)
-
-        net = brainstate.nn.Conv2d((200, 198, 4), out_channels=32, kernel_size=(3, 3),
-                                   stride=(2, 1), padding='VALID', groups=4)
-        out = net(img)
-        print("out shape: ", out.shape)
-        self.assertEqual(out.shape, (2, 99, 196, 32))
-        # print("First output channel:")
-        # plt.figure(figsize=(10, 10))
-        # plt.imshow(np.array(img)[0, :, :, 0])
-        # plt.show()
-
-    def test_conv1D(self):
-        model = brainstate.nn.Conv1d((5, 3), out_channels=32, kernel_size=(3,))
-        input = jnp.ones((2, 5, 3))
-        out = model(input)
-        print("out shape: ", out.shape)
-        self.assertEqual(out.shape, (2, 5, 32))
-        # print("First output channel:")
-        # plt.figure(figsize=(10, 10))
-        # plt.imshow(np.array(out)[0, :, :])
-        # plt.show()
-
-    def test_conv2D(self):
-        model = brainstate.nn.Conv2d((5, 5, 3), out_channels=32, kernel_size=(3, 3))
-        input = jnp.ones((2, 5, 5, 3))
-
-        out = model(input)
-        print("out shape: ", out.shape)
-        self.assertEqual(out.shape, (2, 5, 5, 32))
-
-    def test_conv3D(self):
-        model = brainstate.nn.Conv3d((5, 5, 5, 3), out_channels=32, kernel_size=(3, 3, 3))
-        input = jnp.ones((2, 5, 5, 5, 3))
-        out = model(input)
-        print("out shape: ", out.shape)
-        self.assertEqual(out.shape, (2, 5, 5, 5, 32))
-
-
-@pytest.mark.skip(reason="not implemented yet")
-class TestConvTranspose1d(parameterized.TestCase):
-    def test_conv_transpose(self):
-
-        x = jnp.ones((1, 8, 3))
-        for use_bias in [True, False]:
-            conv_transpose_module = brainstate.nn.ConvTranspose1d(
-                in_channels=3,
-                out_channels=4,
-                kernel_size=(3,),
-                padding='VALID',
-                w_initializer=brainstate.init.Constant(1.),
-                b_initializer=brainstate.init.Constant(1.) if use_bias else None,
-            )
-            self.assertEqual(conv_transpose_module.w.shape, (3, 3, 4))
-            y = conv_transpose_module(x)
-            print(y.shape)
-            correct_ans = jnp.array([[[4., 4., 4., 4.],
-                                      [7., 7., 7., 7.],
-                                      [10., 10., 10., 10.],
-                                      [10., 10., 10., 10.],
-                                      [10., 10., 10., 10.],
-                                      [10., 10., 10., 10.],
-                                      [10., 10., 10., 10.],
-                                      [10., 10., 10., 10.],
-                                      [7., 7., 7., 7.],
-                                      [4., 4., 4., 4.]]])
-            if not use_bias:
-                correct_ans -= 1.
-            self.assertTrue(jnp.allclose(y, correct_ans))
-
-    def test_single_input_masked_conv_transpose(self):
-
-        x = jnp.ones((1, 8, 3))
-        m = jnp.tril(jnp.ones((3, 3, 4)))
-        conv_transpose_module = brainstate.nn.ConvTranspose1d(
-            in_channels=3,
-            out_channels=4,
-            kernel_size=(3,),
-            padding='VALID',
-            mask=m,
-            w_initializer=brainstate.init.Constant(),
-            b_initializer=brainstate.init.Constant(),
-        )
-        self.assertEqual(conv_transpose_module.w.shape, (3, 3, 4))
-        y = conv_transpose_module(x)
-        print(y.shape)
-        correct_ans = jnp.array([[[4., 3., 2., 1.],
-                                  [7., 5., 3., 1.],
-                                  [10., 7., 4., 1.],
-                                  [10., 7., 4., 1.],
-                                  [10., 7., 4., 1.],
-                                  [10., 7., 4., 1.],
-                                  [10., 7., 4., 1.],
-                                  [10., 7., 4., 1.],
-                                  [7., 5., 3., 1.],
-                                  [4., 3., 2., 1.]]])
-        self.assertTrue(jnp.allclose(y, correct_ans))
-
-    def test_computation_padding_same(self):
-
-        data = jnp.ones([1, 3, 1])
-        for use_bias in [True, False]:
-            net = brainstate.nn.ConvTranspose1d(
-                in_channels=1,
-                out_channels=1,
+class TestConv1d(unittest.TestCase):
+    """Test cases for 1D convolution."""
+
+    def test_basic_channels_last(self):
+        """Test basic Conv1d with channels-last format."""
+        conv = brainstate.nn.Conv1d(in_size=(100, 16), out_channels=32, kernel_size=5)
+        x = jnp.ones((4, 100, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 100, 32))
+        self.assertEqual(conv.in_channels, 16)
+        self.assertEqual(conv.out_channels, 32)
+        self.assertFalse(conv.channel_first)
+
+    def test_basic_channels_first(self):
+        """Test basic Conv1d with channels-first format."""
+        conv = brainstate.nn.Conv1d(in_size=(16, 100), out_channels=32, kernel_size=5, channel_first=True)
+        x = jnp.ones((4, 16, 100))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 32, 100))
+        self.assertEqual(conv.in_channels, 16)
+        self.assertEqual(conv.out_channels, 32)
+        self.assertTrue(conv.channel_first)
+
+    def test_without_batch(self):
+        """Test Conv1d without batch dimension."""
+        conv = brainstate.nn.Conv1d(in_size=(50, 8), out_channels=16, kernel_size=3)
+        x = jnp.ones((50, 8))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (50, 16))
+
+    def test_stride(self):
+        """Test Conv1d with stride."""
+        conv = brainstate.nn.Conv1d(in_size=(100, 8), out_channels=16, kernel_size=3, stride=2, padding='VALID')
+        x = jnp.ones((2, 100, 8))
+        y = conv(x)
+
+        # VALID padding: output = (100 - 3 + 1) / 2 = 49
+        self.assertEqual(y.shape, (2, 49, 16))
+
+    def test_dilation(self):
+        """Test Conv1d with dilated convolution."""
+        conv = brainstate.nn.Conv1d(in_size=(100, 8), out_channels=16, kernel_size=3, rhs_dilation=2)
+        x = jnp.ones((2, 100, 8))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 100, 16))
+
+    def test_groups(self):
+        """Test Conv1d with grouped convolution."""
+        conv = brainstate.nn.Conv1d(in_size=(100, 16), out_channels=32, kernel_size=3, groups=4)
+        x = jnp.ones((2, 100, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 100, 32))
+        self.assertEqual(conv.groups, 4)
+
+    def test_with_bias(self):
+        """Test Conv1d with bias."""
+        conv = brainstate.nn.Conv1d(in_size=(50, 8), out_channels=16, kernel_size=3,
+                                    b_init=brainstate.init.Constant(0.0))
+        x = jnp.ones((2, 50, 8))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 50, 16))
+        self.assertIn('bias', conv.weight.value)
+
+
+class TestConv2d(unittest.TestCase):
+    """Test cases for 2D convolution."""
+
+    def test_basic_channels_last(self):
+        """Test basic Conv2d with channels-last format."""
+        conv = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=64, kernel_size=3)
+        x = jnp.ones((8, 32, 32, 3))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (8, 32, 32, 64))
+        self.assertEqual(conv.in_channels, 3)
+        self.assertEqual(conv.out_channels, 64)
+        self.assertFalse(conv.channel_first)
+
+    def test_basic_channels_first(self):
+        """Test basic Conv2d with channels-first format."""
+        conv = brainstate.nn.Conv2d(in_size=(3, 32, 32), out_channels=64, kernel_size=3, channel_first=True)
+        x = jnp.ones((8, 3, 32, 32))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (8, 64, 32, 32))
+        self.assertEqual(conv.in_channels, 3)
+        self.assertEqual(conv.out_channels, 64)
+        self.assertTrue(conv.channel_first)
+
+    def test_rectangular_kernel(self):
+        """Test Conv2d with rectangular kernel."""
+        conv = brainstate.nn.Conv2d(in_size=(64, 64, 16), out_channels=32, kernel_size=(3, 5))
+        x = jnp.ones((4, 64, 64, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 64, 64, 32))
+        self.assertEqual(conv.kernel_size, (3, 5))
+
+    def test_stride_2d(self):
+        """Test Conv2d with different strides."""
+        conv = brainstate.nn.Conv2d(in_size=(64, 64, 3), out_channels=32, kernel_size=3, stride=(2, 2), padding='VALID')
+        x = jnp.ones((4, 64, 64, 3))
+        y = conv(x)
+
+        # VALID padding: output = (64 - 3 + 1) / 2 = 31
+        self.assertEqual(y.shape, (4, 31, 31, 32))
+
+    def test_depthwise_convolution(self):
+        """Test depthwise convolution (groups = in_channels)."""
+        conv = brainstate.nn.Conv2d(in_size=(32, 32, 16), out_channels=16, kernel_size=3, groups=16)
+        x = jnp.ones((4, 32, 32, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 32, 32, 16))
+        self.assertEqual(conv.groups, 16)
+
+    def test_padding_same_vs_valid(self):
+        """Test different padding modes."""
+        conv_same = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=5, padding='SAME')
+        conv_valid = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=5, padding='VALID')
+
+        x = jnp.ones((2, 32, 32, 3))
+        y_same = conv_same(x)
+        y_valid = conv_valid(x)
+
+        self.assertEqual(y_same.shape, (2, 32, 32, 16))  # SAME preserves size
+        self.assertEqual(y_valid.shape, (2, 28, 28, 16))  # VALID reduces size
+
+
+class TestConv3d(unittest.TestCase):
+    """Test cases for 3D convolution."""
+
+    def test_basic_channels_last(self):
+        """Test basic Conv3d with channels-last format."""
+        conv = brainstate.nn.Conv3d(in_size=(16, 16, 16, 1), out_channels=32, kernel_size=3)
+        x = jnp.ones((2, 16, 16, 16, 1))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 16, 16, 16, 32))
+        self.assertEqual(conv.in_channels, 1)
+        self.assertEqual(conv.out_channels, 32)
+
+    def test_basic_channels_first(self):
+        """Test basic Conv3d with channels-first format."""
+        conv = brainstate.nn.Conv3d(in_size=(1, 16, 16, 16), out_channels=32, kernel_size=3, channel_first=True)
+        x = jnp.ones((2, 1, 16, 16, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 32, 16, 16, 16))
+        self.assertEqual(conv.in_channels, 1)
+        self.assertEqual(conv.out_channels, 32)
+
+    def test_video_data(self):
+        """Test Conv3d for video data."""
+        conv = brainstate.nn.Conv3d(in_size=(8, 32, 32, 3), out_channels=64, kernel_size=(3, 3, 3))
+        x = jnp.ones((4, 8, 32, 32, 3))  # batch, frames, height, width, channels
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 8, 32, 32, 64))
+
+
+class TestScaledWSConv1d(unittest.TestCase):
+    """Test cases for 1D convolution with weight standardization."""
+
+    def test_basic(self):
+        """Test basic ScaledWSConv1d."""
+        conv = brainstate.nn.ScaledWSConv1d(in_size=(100, 16), out_channels=32, kernel_size=5)
+        x = jnp.ones((4, 100, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 100, 32))
+        self.assertIsNotNone(conv.eps)
+
+    def test_with_gain(self):
+        """Test ScaledWSConv1d with gain parameter."""
+        conv = brainstate.nn.ScaledWSConv1d(in_size=(100, 16), out_channels=32, kernel_size=5, ws_gain=True)
+        x = jnp.ones((4, 100, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 100, 32))
+        self.assertIn('gain', conv.weight.value)
+
+    def test_without_gain(self):
+        """Test ScaledWSConv1d without gain parameter."""
+        conv = brainstate.nn.ScaledWSConv1d(in_size=(100, 16), out_channels=32, kernel_size=5, ws_gain=False)
+        x = jnp.ones((4, 100, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 100, 32))
+        self.assertNotIn('gain', conv.weight.value)
+
+    def test_custom_eps(self):
+        """Test ScaledWSConv1d with custom epsilon."""
+        conv = brainstate.nn.ScaledWSConv1d(in_size=(100, 16), out_channels=32, kernel_size=5, eps=1e-5)
+        self.assertEqual(conv.eps, 1e-5)
+
+
+class TestScaledWSConv2d(unittest.TestCase):
+    """Test cases for 2D convolution with weight standardization."""
+
+    def test_basic_channels_last(self):
+        """Test basic ScaledWSConv2d with channels-last format."""
+        conv = brainstate.nn.ScaledWSConv2d(in_size=(64, 64, 3), out_channels=32, kernel_size=3)
+        x = jnp.ones((8, 64, 64, 3))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (8, 64, 64, 32))
+
+    def test_basic_channels_first(self):
+        """Test basic ScaledWSConv2d with channels-first format."""
+        conv = brainstate.nn.ScaledWSConv2d(in_size=(3, 64, 64), out_channels=32, kernel_size=3, channel_first=True)
+        x = jnp.ones((8, 3, 64, 64))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (8, 32, 64, 64))
+
+    def test_with_group_norm_style(self):
+        """Test ScaledWSConv2d for use with group normalization."""
+        conv = brainstate.nn.ScaledWSConv2d(
+            in_size=(32, 32, 16),
+            out_channels=32,
+            kernel_size=3,
+            ws_gain=True,
+            groups=1
+        )
+        x = jnp.ones((4, 32, 32, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (4, 32, 32, 32))
+
+
+class TestScaledWSConv3d(unittest.TestCase):
+    """Test cases for 3D convolution with weight standardization."""
+
+    def test_basic(self):
+        """Test basic ScaledWSConv3d."""
+        conv = brainstate.nn.ScaledWSConv3d(in_size=(8, 16, 16, 3), out_channels=32, kernel_size=3)
+        x = jnp.ones((2, 8, 16, 16, 3))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 8, 16, 16, 32))
+
+    def test_channels_first(self):
+        """Test ScaledWSConv3d with channels-first format."""
+        conv = brainstate.nn.ScaledWSConv3d(in_size=(3, 8, 16, 16), out_channels=32, kernel_size=3, channel_first=True)
+        x = jnp.ones((2, 3, 8, 16, 16))
+        y = conv(x)
+
+        self.assertEqual(y.shape, (2, 32, 8, 16, 16))
+
+
+class TestErrorHandling(unittest.TestCase):
+    """Test error handling and edge cases."""
+
+    def test_invalid_input_shape(self):
+        """Test that invalid input shapes raise appropriate errors."""
+        conv = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=3)
+        x_wrong = jnp.ones((8, 32, 32, 16))  # Wrong number of channels
+
+        with self.assertRaises(ValueError):
+            conv(x_wrong)
+
+    def test_invalid_groups(self):
+        """Test that invalid group configurations raise errors."""
+        with self.assertRaises(AssertionError):
+            # out_channels not divisible by groups
+            conv = brainstate.nn.Conv2d(in_size=(32, 32, 16), out_channels=30, kernel_size=3, groups=4)
+
+    def test_dimension_mismatch(self):
+        """Test dimension mismatch detection."""
+        conv = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=3)
+        x_1d = jnp.ones((8, 32, 3))  # 1D instead of 2D
+
+        with self.assertRaises(ValueError):
+            conv(x_1d)
+
+
+class TestOutputShapes(unittest.TestCase):
+    """Test output shape calculations."""
+
+    def test_same_padding_preserves_size(self):
+        """Test that SAME padding preserves spatial dimensions when stride=1."""
+        for kernel_size in [3, 5, 7]:
+            conv = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=kernel_size, padding='SAME')
+            x = jnp.ones((4, 32, 32, 3))
+            y = conv(x)
+            self.assertEqual(y.shape, (4, 32, 32, 16), f"Failed for kernel_size={kernel_size}")
+
+    def test_valid_padding_reduces_size(self):
+        """Test that VALID padding reduces spatial dimensions."""
+        conv = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=5, padding='VALID')
+        x = jnp.ones((4, 32, 32, 3))
+        y = conv(x)
+        # 32 - 5 + 1 = 28
+        self.assertEqual(y.shape, (4, 28, 28, 16))
+
+    def test_output_size_attribute(self):
+        """Test that out_size attribute is correctly computed."""
+        conv_cl = brainstate.nn.Conv2d(in_size=(64, 64, 3), out_channels=32, kernel_size=3, channel_first=False)
+        conv_cf = brainstate.nn.Conv2d(in_size=(3, 64, 64), out_channels=32, kernel_size=3, channel_first=True)
+
+        self.assertEqual(conv_cl.out_size, (64, 64, 32))
+        self.assertEqual(conv_cf.out_size, (32, 64, 64))
+
+
+class TestChannelFormatConsistency(unittest.TestCase):
+    """Test consistency between channels-first and channels-last formats."""
+
+    def test_conv1d_output_channels(self):
+        """Test that output channels are in correct position for both formats."""
+        conv_cl = brainstate.nn.Conv1d(in_size=(100, 16), out_channels=32, kernel_size=3)
+        conv_cf = brainstate.nn.Conv1d(in_size=(16, 100), out_channels=32, kernel_size=3, channel_first=True)
+
+        x_cl = jnp.ones((4, 100, 16))
+        x_cf = jnp.ones((4, 16, 100))
+
+        y_cl = conv_cl(x_cl)
+        y_cf = conv_cf(x_cf)
+
+        # Channels-last: channels in last dimension
+        self.assertEqual(y_cl.shape[-1], 32)
+        # Channels-first: channels in first dimension (after batch)
+        self.assertEqual(y_cf.shape[1], 32)
+
+    def test_conv2d_output_channels(self):
+        """Test 2D output channel positions."""
+        conv_cl = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=64, kernel_size=3)
+        conv_cf = brainstate.nn.Conv2d(in_size=(3, 32, 32), out_channels=64, kernel_size=3, channel_first=True)
+
+        x_cl = jnp.ones((4, 32, 32, 3))
+        x_cf = jnp.ones((4, 3, 32, 32))
+
+        y_cl = conv_cl(x_cl)
+        y_cf = conv_cf(x_cf)
+
+        self.assertEqual(y_cl.shape[-1], 64)
+        self.assertEqual(y_cf.shape[1], 64)
+
+
+class TestReproducibility(unittest.TestCase):
+    """Test reproducibility with fixed seeds."""
+
+    def test_deterministic_output(self):
+        """Test that same seed produces same output."""
+        key = jax.random.PRNGKey(42)
+
+        conv1 = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=3)
+        conv2 = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=16, kernel_size=3)
+
+        # Use same random key for input
+        x = jax.random.normal(key, (4, 32, 32, 3))
+
+        # Note: outputs will differ due to different weight initialization
+        # This test just ensures no crashes with random inputs
+        y1 = conv1(x)
+        y2 = conv2(x)
+
+        self.assertEqual(y1.shape, y2.shape)
+
+
+class TestRepr(unittest.TestCase):
+    """Test string representations."""
+
+    def test_conv_repr_channels_last(self):
+        """Test __repr__ for channels-last format."""
+        conv = brainstate.nn.Conv2d(in_size=(32, 32, 3), out_channels=64, kernel_size=3)
+        repr_str = repr(conv)
+
+        self.assertIn('Conv2d', repr_str)
+        self.assertIn('channel_first=False', repr_str)
+        self.assertIn('in_channels=3', repr_str)
+        self.assertIn('out_channels=64', repr_str)
+
+    def test_conv_repr_channels_first(self):
+        """Test __repr__ for channels-first format."""
+        conv = brainstate.nn.Conv2d(in_size=(3, 32, 32), out_channels=64, kernel_size=3, channel_first=True)
+        repr_str = repr(conv)
+
+        self.assertIn('Conv2d', repr_str)
+        self.assertIn('channel_first=True', repr_str)
+
+
+class TestConvTranspose1d(unittest.TestCase):
+    """Test cases for ConvTranspose1d layer."""
+
+    def setUp(self):
+        """Set up test fixtures."""
+        self.in_size = (28, 16)
+        self.out_channels = 8
+        self.kernel_size = 4
+
+    def test_basic_channels_last(self):
+        """Test basic ConvTranspose1d with channels-last format."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=self.in_size,
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size,
+            stride=1
+        )
+        x = jnp.ones((2, 28, 16))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 3)
+        self.assertEqual(y.shape[0], 2)  # batch size
+        self.assertEqual(y.shape[-1], self.out_channels)
+        self.assertEqual(conv_t.in_channels, 16)
+        self.assertEqual(conv_t.out_channels, 8)
+        self.assertFalse(conv_t.channel_first)
+
+    def test_basic_channels_first(self):
+        """Test basic ConvTranspose1d with channels-first format."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(16, 28),  # (C, L) for channels-first
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size,
+            stride=1,
+            channel_first=True
+        )
+        x = jnp.ones((2, 16, 28))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 3)
+        self.assertEqual(y.shape[0], 2)  # batch size
+        self.assertEqual(y.shape[1], self.out_channels)  # channels first
+        self.assertEqual(conv_t.in_channels, 16)
+        self.assertTrue(conv_t.channel_first)
+
+    def test_stride_upsampling(self):
+        """Test transposed convolution with stride for upsampling."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=8,
+            kernel_size=4,
+            stride=2,
+            padding='SAME'
+        )
+        x = jnp.ones((2, 28, 16))
+        y = conv_t(x)
+
+        # With stride=2, output should be approximately 2x larger
+        self.assertGreater(y.shape[1], x.shape[1])
+
+    def test_with_bias(self):
+        """Test ConvTranspose1d with bias."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(50, 8),
+            out_channels=16,
+            kernel_size=3,
+            b_init=brainstate.init.Constant(0.0)
+        )
+        x = jnp.ones((4, 50, 8))
+        y = conv_t(x)
+
+        self.assertTrue('bias' in conv_t.weight.value)
+        self.assertEqual(y.shape[-1], 16)
+
+    def test_without_batch(self):
+        """Test ConvTranspose1d without batch dimension."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=8,
+            kernel_size=4
+        )
+        x = jnp.ones((28, 16))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 2)
+        self.assertEqual(y.shape[-1], 8)
+
+    def test_groups(self):
+        """Test grouped transposed convolution."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=16,
+            kernel_size=3,
+            groups=4
+        )
+        x = jnp.ones((2, 28, 16))
+        y = conv_t(x)
+
+        self.assertEqual(y.shape[-1], 16)
+
+
+class TestConvTranspose2d(unittest.TestCase):
+    """Test cases for ConvTranspose2d layer."""
+
+    def setUp(self):
+        """Set up test fixtures."""
+        self.in_size = (16, 16, 32)
+        self.out_channels = 16
+        self.kernel_size = 4
+
+    def test_basic_channels_last(self):
+        """Test basic ConvTranspose2d with channels-last format."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=self.in_size,
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size
+        )
+        x = jnp.ones((4, 16, 16, 32))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 4)
+        self.assertEqual(y.shape[0], 4)  # batch size
+        self.assertEqual(y.shape[-1], self.out_channels)
+        self.assertEqual(conv_t.in_channels, 32)
+        self.assertFalse(conv_t.channel_first)
+
+    def test_basic_channels_first(self):
+        """Test basic ConvTranspose2d with channels-first format."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(32, 16, 16),  # (C, H, W) for channels-first
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size,
+            channel_first=True
+        )
+        x = jnp.ones((4, 32, 16, 16))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 4)
+        self.assertEqual(y.shape[1], self.out_channels)  # channels first
+        self.assertTrue(conv_t.channel_first)
+
+    def test_stride_upsampling(self):
+        """Test 2x upsampling with stride=2."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=4,
+            stride=2,
+            padding='SAME'
+        )
+        x = jnp.ones((4, 16, 16, 32))
+        y = conv_t(x)
+
+        # With stride=2, output should be approximately 2x larger in each spatial dimension
+        self.assertGreater(y.shape[1], x.shape[1])
+        self.assertGreater(y.shape[2], x.shape[2])
+
+    def test_rectangular_kernel(self):
+        """Test ConvTranspose2d with rectangular kernel."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=(3, 5),
+            stride=1
+        )
+        x = jnp.ones((2, 16, 16, 32))
+        y = conv_t(x)
+
+        self.assertEqual(conv_t.kernel_size, (3, 5))
+        self.assertEqual(y.shape[-1], 16)
+
+    def test_padding_valid(self):
+        """Test ConvTranspose2d with VALID padding."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=4,
+            stride=2,
+            padding='VALID'
+        )
+        x = jnp.ones((2, 16, 16, 32))
+        y = conv_t(x)
+
+        # VALID padding means no padding, output computed by formula:
+        # out = (in - 1) * stride + kernel
+        # out = (16 - 1) * 2 + 4 = 34 (but JAX may compute it slightly differently)
+        # At minimum, it should upsample
+        self.assertGreater(y.shape[1], 16)
+
+    def test_groups(self):
+        """Test grouped transposed convolution."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=32,
+            kernel_size=3,
+            groups=4
+        )
+        x = jnp.ones((2, 16, 16, 32))
+        y = conv_t(x)
+
+        self.assertEqual(y.shape[-1], 32)
+
+
+class TestConvTranspose3d(unittest.TestCase):
+    """Test cases for ConvTranspose3d layer."""
+
+    def setUp(self):
+        """Set up test fixtures."""
+        self.in_size = (8, 8, 8, 16)
+        self.out_channels = 8
+        self.kernel_size = 4
+
+    def test_basic_channels_last(self):
+        """Test basic ConvTranspose3d with channels-last format."""
+        conv_t = brainstate.nn.ConvTranspose3d(
+            in_size=self.in_size,
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size
+        )
+        x = jnp.ones((2, 8, 8, 8, 16))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 5)
+        self.assertEqual(y.shape[0], 2)  # batch size
+        self.assertEqual(y.shape[-1], self.out_channels)
+        self.assertEqual(conv_t.in_channels, 16)
+
+    def test_basic_channels_first(self):
+        """Test basic ConvTranspose3d with channels-first format."""
+        conv_t = brainstate.nn.ConvTranspose3d(
+            in_size=(16, 8, 8, 8),  # (C, H, W, D) for channels-first
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size,
+            channel_first=True
+        )
+        x = jnp.ones((2, 16, 8, 8, 8))
+        y = conv_t(x)
+
+        self.assertEqual(len(y.shape), 5)
+        self.assertEqual(y.shape[1], self.out_channels)  # channels first
+        self.assertTrue(conv_t.channel_first)
+
+    def test_stride_upsampling(self):
+        """Test 3D upsampling with stride=2."""
+        conv_t = brainstate.nn.ConvTranspose3d(
+            in_size=(8, 8, 8, 16),
+            out_channels=8,
+            kernel_size=4,
+            stride=2,
+            padding='SAME'
+        )
+        x = jnp.ones((2, 8, 8, 8, 16))
+        y = conv_t(x)
+
+        # With stride=2, output should be approximately 2x larger
+        self.assertGreater(y.shape[1], x.shape[1])
+        self.assertGreater(y.shape[2], x.shape[2])
+        self.assertGreater(y.shape[3], x.shape[3])
+
+
+class TestErrorHandlingConvTranspose(unittest.TestCase):
+    """Test error handling for transposed convolutions."""
+
+    def test_invalid_groups(self):
+        """Test that invalid groups raises assertion error."""
+        with self.assertRaises(AssertionError):
+            brainstate.nn.ConvTranspose2d(
+                in_size=(16, 16, 32),
+                out_channels=15,  # Not divisible by groups
                 kernel_size=3,
-                stride=1,
-                padding="SAME",
-                w_initializer=brainstate.init.Constant(),
-                b_initializer=brainstate.init.Constant() if use_bias else None,
+                groups=4
             )
-            out = net(data)
-            self.assertEqual(out.shape, (1, 3, 1))
-            out = jnp.squeeze(out, axis=(0, 2))
-            expected_out = jnp.asarray([2, 3, 2])
-            if use_bias:
-                expected_out += 1
-            self.assertTrue(jnp.allclose(out, expected_out, rtol=1e-5))
-
-
-@pytest.mark.skip(reason="not implemented yet")
-class TestConvTranspose2d(parameterized.TestCase):
-    def test_conv_transpose(self):
-
-        x = jnp.ones((1, 8, 8, 3))
-        for use_bias in [True, False]:
-            conv_transpose_module = brainstate.nn.ConvTranspose2d(
-                in_channels=3,
-                out_channels=4,
-                kernel_size=(3, 3),
-                padding='VALID',
-                w_initializer=brainstate.init.Constant(),
-                b_initializer=brainstate.init.Constant() if use_bias else None,
-            )
-        self.assertEqual(conv_transpose_module.w.shape, (3, 3, 3, 4))
-        y = conv_transpose_module(x)
-        print(y.shape)
-
-    def test_single_input_masked_conv_transpose(self):
-
-        x = jnp.ones((1, 8, 8, 3))
-        m = jnp.tril(jnp.ones((3, 3, 3, 4)))
-        conv_transpose_module = brainstate.nn.ConvTranspose2d(
-            in_channels=3,
-            out_channels=4,
-            kernel_size=(3, 3),
-            padding='VALID',
-            mask=m,
-            w_initializer=brainstate.init.Constant(),
-        )
-        y = conv_transpose_module(x)
-        print(y.shape)
-
-    def test_computation_padding_same(self):
-
-        x = jnp.ones((1, 8, 8, 3))
-        for use_bias in [True, False]:
-            conv_transpose_module = brainstate.nn.ConvTranspose2d(
-                in_channels=3,
-                out_channels=4,
-                kernel_size=(3, 3),
-                stride=1,
-                padding='SAME',
-                w_initializer=brainstate.init.Constant(),
-                b_initializer=brainstate.init.Constant() if use_bias else None,
-            )
-        y = conv_transpose_module(x)
-        print(y.shape)
-
-
-@pytest.mark.skip(reason="not implemented yet")
-class TestConvTranspose3d(parameterized.TestCase):
-    def test_conv_transpose(self):
-
-        x = jnp.ones((1, 8, 8, 8, 3))
-        for use_bias in [True, False]:
-            conv_transpose_module = brainstate.nn.ConvTranspose3d(
-                in_channels=3,
-                out_channels=4,
-                kernel_size=(3, 3, 3),
-                padding='VALID',
-                w_initializer=brainstate.init.Constant(),
-                b_initializer=brainstate.init.Constant() if use_bias else None,
-            )
-        y = conv_transpose_module(x)
-        print(y.shape)
-
-    def test_single_input_masked_conv_transpose(self):
-
-        x = jnp.ones((1, 8, 8, 8, 3))
-        m = jnp.tril(jnp.ones((3, 3, 3, 3, 4)))
-        conv_transpose_module = brainstate.nn.ConvTranspose3d(
-            in_channels=3,
-            out_channels=4,
-            kernel_size=(3, 3, 3),
-            padding='VALID',
-            mask=m,
-            w_initializer=brainstate.init.Constant(),
-        )
-        y = conv_transpose_module(x)
-        print(y.shape)
-
-    def test_computation_padding_same(self):
-
-        x = jnp.ones((1, 8, 8, 8, 3))
-        for use_bias in [True, False]:
-            conv_transpose_module = brainstate.nn.ConvTranspose3d(
-                in_channels=3,
-                out_channels=4,
-                kernel_size=(3, 3, 3),
-                stride=1,
-                padding='SAME',
-                w_initializer=brainstate.init.Constant(),
-                b_initializer=brainstate.init.Constant() if use_bias else None,
-            )
-        y = conv_transpose_module(x)
-        print(y.shape)
 
+    def test_dimension_mismatch(self):
+        """Test that wrong input dimensions raise error."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=3
+        )
+        x = jnp.ones((2, 16, 16, 16))  # Wrong number of channels
+
+        with self.assertRaises(ValueError):
+            conv_t(x)
 
-if __name__ == '__main__':
-    absltest.main()
+    def test_invalid_input_shape(self):
+        """Test that invalid input shape raises error."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=8,
+            kernel_size=3
+        )
+        x = jnp.ones((2, 2, 28, 16))  # Too many dimensions
+
+        with self.assertRaises(ValueError):
+            conv_t(x)
+
+
+class TestOutputShapesConvTranspose(unittest.TestCase):
+    """Test output shape computation for transposed convolutions."""
+
+    def test_out_size_attribute_1d(self):
+        """Test that out_size attribute is correctly computed for 1D."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=8,
+            kernel_size=4,
+            stride=2
+        )
+
+        self.assertIsNotNone(conv_t.out_size)
+        self.assertEqual(len(conv_t.out_size), 2)
+
+    def test_out_size_attribute_2d(self):
+        """Test that out_size attribute is correctly computed for 2D."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=4,
+            stride=2
+        )
+
+        self.assertIsNotNone(conv_t.out_size)
+        self.assertEqual(len(conv_t.out_size), 3)
+
+    def test_upsampling_factor(self):
+        """Test that stride=2 approximately doubles spatial dimensions."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=4,
+            stride=2,
+            padding='SAME'
+        )
+        x = jnp.ones((2, 16, 16, 32))
+        y = conv_t(x)
+
+        # For SAME padding and stride=2, output should be approximately 2x input
+        self.assertGreaterEqual(y.shape[1], 28)
+        self.assertGreaterEqual(y.shape[2], 28)
+
+
+class TestChannelFormatConsistencyConvTranspose(unittest.TestCase):
+    """Test consistency between different channel formats."""
+
+    def test_conv_transpose_1d_output_channels(self):
+        """Test that output channels are in correct position for both formats."""
+        # Channels-last
+        conv_t_last = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=8,
+            kernel_size=3
+        )
+        x_last = jnp.ones((2, 28, 16))
+        y_last = conv_t_last(x_last)
+        self.assertEqual(y_last.shape[-1], 8)
+
+        # Channels-first
+        conv_t_first = brainstate.nn.ConvTranspose1d(
+            in_size=(16, 28),
+            out_channels=8,
+            kernel_size=3,
+            channel_first=True
+        )
+        x_first = jnp.ones((2, 16, 28))
+        y_first = conv_t_first(x_first)
+        self.assertEqual(y_first.shape[1], 8)
+
+    def test_conv_transpose_2d_output_channels(self):
+        """Test that output channels are in correct position for both formats."""
+        # Channels-last
+        conv_t_last = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=3
+        )
+        x_last = jnp.ones((2, 16, 16, 32))
+        y_last = conv_t_last(x_last)
+        self.assertEqual(y_last.shape[-1], 16)
+
+        # Channels-first
+        conv_t_first = brainstate.nn.ConvTranspose2d(
+            in_size=(32, 16, 16),
+            out_channels=16,
+            kernel_size=3,
+            channel_first=True
+        )
+        x_first = jnp.ones((2, 32, 16, 16))
+        y_first = conv_t_first(x_first)
+        self.assertEqual(y_first.shape[1], 16)
+
+
+class TestReproducibilityConvTranspose(unittest.TestCase):
+    """Test deterministic behavior of transposed convolutions."""
+
+    def test_deterministic_output(self):
+        """Test that same input produces same output."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=3
+        )
+        x = jnp.ones((2, 16, 16, 32))
+
+        y1 = conv_t(x)
+        y2 = conv_t(x)
+
+        self.assertTrue(jnp.allclose(y1, y2))
+
+
+class TestKernelShapeConvTranspose(unittest.TestCase):
+    """Test kernel shape computation for transposed convolutions."""
+
+    def test_kernel_shape_1d(self):
+        """Test that kernel shape is correct for transposed conv 1D."""
+        conv_t = brainstate.nn.ConvTranspose1d(
+            in_size=(28, 16),
+            out_channels=8,
+            kernel_size=4,
+            groups=2
+        )
+        # For transpose conv: (kernel_size, out_channels, in_channels // groups)
+        expected_shape = (4, 8, 16 // 2)
+        self.assertEqual(conv_t.kernel_shape, expected_shape)
+
+    def test_kernel_shape_2d(self):
+        """Test that kernel shape is correct for transposed conv 2D."""
+        conv_t = brainstate.nn.ConvTranspose2d(
+            in_size=(16, 16, 32),
+            out_channels=16,
+            kernel_size=4,
+            groups=4
+        )
+        # For transpose conv: (kernel_h, kernel_w, out_channels, in_channels // groups)
+        expected_shape = (4, 4, 16, 32 // 4)
+        self.assertEqual(conv_t.kernel_shape, expected_shape)