Use np-conv2d from PyPI

.nengobones.yml

@@ -35,6 +35,7 @@ setup_py:
       - nengo = pytest_nengo
   install_req:
     - numpy>=1.19
+    - np-conv2d>=1.0.0
   docs_req:
     - sphinx>=1.8
     - jupyter

nengo/builder/tests/test_transforms_conv.py

@@ -3,7 +3,7 @@
 
 from nengo.builder.signal import Signal
 from nengo.builder.tests.test_operator import _test_operator_arg_attributes
-from nengo.builder.transforms import ConvInc, ConvTransposeInc, conv2d_gradx
+from nengo.builder.transforms import ConvInc, ConvTransposeInc
 from nengo.transforms import ChannelShape, Convolution, ConvolutionTranspose
 
 
@@ -281,58 +281,6 @@ def debug_convtransposeinc_2d(w, wt, x, xt, y, yt, conv, conv_transpose, plt):
                     assert False, f"Output {i} channel {k} differs"
 
 
-@pytest.mark.parametrize("x_size", (1, 2, 3, 4))
-@pytest.mark.parametrize("k_size", (1, 2, 3, 4))
-@pytest.mark.parametrize("stride", (1, 2, 3, 4))
-@pytest.mark.parametrize("padding", ("same", "valid"))
-def test_conv2d_gradx(padding, stride, k_size, x_size, rng, plt, allclose):
-    tf = pytest.importorskip("tensorflow")
-
-    in_channels = 1
-    out_channels = 1
-
-    if padding == "same":
-        y_min = (x_size - 1) * stride + 1
-        y_max = x_size * stride
-    else:
-        y_min = (x_size - 1) * stride + k_size
-        y_max = x_size * stride + k_size - 1
-
-    bad_sizes = []
-    for y_size in range(y_min, y_max + 1):
-        x_shape = (1, x_size, x_size, in_channels)
-        y_shape = (1, y_size, y_size, out_channels)
-        k_shape = (k_size, k_size, out_channels, in_channels)
-
-        x = rng.uniform(-1, 1, size=x_shape)
-        kernel = rng.uniform(-1, 1, size=k_shape)
-        y_tf = tf.nn.conv2d_transpose(
-            x, kernel, y_shape, stride, padding=padding.upper()
-        ).numpy()[0]
-
-        kernel_t = np.transpose(kernel, (0, 1, 3, 2))
-        y_np = conv2d_gradx(
-            kernel_t,
-            x,
-            xsize=y_shape[1:3],
-            pad=padding.upper(),
-            stride=(stride, stride),
-        )[0]
-
-        assert y_np.shape == y_tf.shape
-
-        if not allclose(y_np, y_tf):
-            plt.subplot(1, 2, 1)
-            plt.imshow(y_tf[..., 0], vmin=y_tf.min(), vmax=y_tf.max())
-            plt.subplot(1, 2, 2)
-            plt.imshow(y_np[..., 0], vmin=y_tf.min(), vmax=y_tf.max())
-            bad_sizes.append(y_size)
-
-    if len(bad_sizes) == 0:
-        plt.saveas = None
-    assert len(bad_sizes) == 0, (y_min, y_max, bad_sizes)
-
-
 def test_convinc_attrs_decstr():
     argnames = ["W", "X", "Y", "conv"]
     non_signals = ["conv"]