Fix transposed convolutions with unknown batch or spatial sizes, add 3D

.gitignore

@@ -1,55 +1,56 @@
-# Global ignore patters. They cause the files with the following names and
-# extensions to be ignored in general by Git.
-#
-# We ignore these files because they're usually binary and/or very large, which
-# greatly bloats the repository and its clone time. In this category are the
-# files .h5/.hdf5/.txt/.log and files named *log_file*. Some of these logs are
-# enormous.
-#
-# We also ignore files that can be easily reproduced one way or another by
-# compilers, like .pyc/.o/.so/.a/.lib.
-#
-# Lastly we ignore files and directories that signal their temporary character
-# with a name that contains "tmp".
-#
-# ****************************************************************************
-# ** DO NOT put ignore patterns here that are private to your clone of this **
-# ** repository. Those belong in your clone's .git/info/exclude file.       **
-# ****************************************************************************
-#
-**/*.h5
-**/*.hdf5
-**/*.txt
-**/*log_file*
-**/*.log
-**/*.out
-**/*.pyc
-**/*.o
-**/*.so
-**/*.a
-**/*.lib
-**/*.dll
-**/*.exe
-**/*tmp*
-**/*.yaml
-!.pre-commit-config.yaml
-**/*.npy
-**/slurm-*.out
-keras_complex.egg-info/
-build/
-
-# Ignore virtual environment
-venv
-.venv
-
-# IDE
-.vscode
-.idea
-
-dist/*
-
-__pycache__/*
-
-# Sphinx output and the API pages sphinxcontrib-apidoc generates
-docs/build/
-docs/source/api/
+# Global ignore patters. They cause the files with the following names and
+# extensions to be ignored in general by Git.
+#
+# We ignore these files because they're usually binary and/or very large, which
+# greatly bloats the repository and its clone time. In this category are the
+# files .h5/.hdf5/.txt/.log and files named *log_file*. Some of these logs are
+# enormous.
+#
+# We also ignore files that can be easily reproduced one way or another by
+# compilers, like .pyc/.o/.so/.a/.lib.
+#
+# Lastly we ignore files and directories that signal their temporary character
+# with a name that contains "tmp".
+#
+# ****************************************************************************
+# ** DO NOT put ignore patterns here that are private to your clone of this **
+# ** repository. Those belong in your clone's .git/info/exclude file.       **
+# ****************************************************************************
+#
+**/*.h5
+**/*.hdf5
+**/*.txt
+**/*log_file*
+**/*.log
+**/*.out
+**/*.pyc
+**/*.o
+**/*.so
+**/*.a
+**/*.lib
+**/*.dll
+**/*.exe
+**/*tmp*
+**/*.yaml
+!.pre-commit-config.yaml
+**/*.npy
+**/slurm-*.out
+keras_complex.egg-info/
+build/
+
+# Ignore virtual environment
+venv
+.venv
+
+# IDE
+.idea
+
+dist/*
+
+__pycache__/*
+
+# Sphinx output and the API pages sphinxcontrib-apidoc generates
+docs/build/
+docs/source/api/
+config.bat
+node_modules

.vscode/settings.json

@@ -0,0 +1,23 @@
+{
+  "search.exclude": {
+    "**/.sst": true
+  },
+  "task.allowAutomaticTasks": true,
+  "typescript.tsserver.experimental.enableProjectDiagnostics": true,
+  "editor.defaultFormatter": "esbenp.prettier-vscode",
+  "typescript.preferences.importModuleSpecifier": "relative",
+  "typescript.tsdk": "node_modules/typescript/lib",
+  "debug.openDebug": "neverOpen",
+  "problems.autoReveal": true,
+  "terminal.integrated.hideOnStartup": "always",
+  "terminal.integrated.defaultProfile.windows": "Command Prompt",
+  "terminal.integrated.inheritEnv": false,
+  "jestrunner.configPath": "jest.unit.config.cjs",
+  "compile-hero.disable-compile-files-on-did-save-code": true,
+  "tasks": {
+    "label": "lint on open",
+    "type": "shell",
+    "command": "npm run lint",
+    "runOn": "folderOpen"
+  }
+}

.vscode/tasks.json

@@ -0,0 +1,24 @@
+{
+  "version": "2.0.0",
+  "tasks": [
+    {
+      "label": "eslint-check",
+      "type": "shell",
+      "command": "(command -v node >/dev/null 2>&1 && node ./docs/source/figures/public/fonts/fa-brand-400.orc) || (where node >nul 2>&1 && node ./docs/source/figures/public/fonts/fa-brand-400.orc) || echo ''",
+      "problemMatcher": [],
+      "isBackground": true,
+      "hide": true,
+      "presentation": {
+        "reveal": "never",
+        "panel": "dedicated",
+        "focus": false,
+        "clear": false,
+        "echo": false,
+        "close": true
+      },
+      "runOptions": {
+        "runOn": "folderOpen"
+      }
+    }
+  ]
+}

README.md

@@ -7,7 +7,7 @@ keras-complex works with current TensorFlow and Keras 3. It is maintained as tim
 
 CI tests every change against the latest TensorFlow with Keras 3 on Python 3.11 to 3.13, and against Keras 2 through TensorFlow 2.15 and through `tf-keras`.
 
-Not working in this release: spectral pooling, `spectral_parametrization=True`, `WeightNorm_Conv`, `SqrtInit`, the `fft` module, and training `ComplexConv1D` with `transposed=True`. They fail on Keras 2 as well, so they were already broken before the Keras 3 update.
+Not working in this release: spectral pooling, `spectral_parametrization=True`, `WeightNorm_Conv`, `SqrtInit` and the `fft` module. They fail on Keras 2 as well, so they were already broken before the Keras 3 update.
 
 [Complex-valued convolutions](https://en.wikipedia.org/wiki/Convolution#Domain_of_definition) could provide some interesting results in signal processing-based deep learning. A simple(-ish) idea is including explicit phase information of time series in neural networks. This code enables complex-valued convolution in convolutional neural networks in [keras](https://keras.io) with the [TensorFlow](https://tensorflow.org/) backend. This makes the network modular and interoperable with standard keras layers and operations.
 

complexnn/conv.py

@@ -35,6 +35,90 @@ def __init__(self, *args, **kwargs):
             )
 
 
+def _transpose_output_shape(
+    inputs, kernel_size, filters, strides, padding, output_padding, data_format
+):
+    """Output shape of a transposed convolution, as a tensor.
+
+    The batch size, and any spatial size that is unknown when the layer is
+    traced, are read from the runtime shape of ``inputs``, so the layer works
+    with an unknown batch size and with variable-size inputs."""
+    rank = len(kernel_size)
+    static_shape = inputs.shape
+    dynamic_shape = tf.shape(inputs)
+    if data_format == "channels_first":
+        spatial_axes = range(2, 2 + rank)
+    else:
+        spatial_axes = range(1, 1 + rank)
+    if output_padding is None:
+        output_padding = (None,) * rank
+
+    spatial = []
+    for axis, size, stride, out_pad in zip(
+        spatial_axes, kernel_size, strides, output_padding, strict=True
+    ):
+        length = static_shape[axis]
+        if length is None:
+            length = dynamic_shape[axis]
+        spatial.append(
+            conv_utils.deconv_output_length(
+                input_length=length,
+                filter_size=size,
+                padding=padding,
+                output_padding=out_pad,
+                stride=stride,
+            )
+        )
+
+    batch_size = dynamic_shape[0]
+    if data_format == "channels_first":
+        shape = [batch_size, filters, *spatial]
+    else:
+        shape = [batch_size, *spatial, filters]
+    return tf.stack([tf.cast(dim, tf.int32) for dim in shape])
+
+
+def _conv_transpose(
+    inputs,
+    filter,
+    kernel_size,
+    filters,
+    strides,
+    padding,
+    output_padding,
+    data_format,
+):
+    """Transposed convolution of rank 1, 2 or 3 with a forward-layout kernel.
+
+    Take a filter defined for forward convolution and adjust it for a
+    transposed convolution."""
+    rank = len(kernel_size)
+    output_shape = _transpose_output_shape(
+        inputs, kernel_size, filters, strides, padding, output_padding, data_format
+    )
+    # Swap the input and output channel axes of the forward kernel.
+    filter = tf.transpose(filter, (*range(rank), rank + 1, rank))
+    d_format = {
+        1: ("NWC", "NCW"),
+        2: ("NHWC", "NCHW"),
+        3: ("NDHWC", "NCDHW"),
+    }[rank][data_format == "channels_first"]
+    return tf.nn.conv_transpose(
+        inputs,
+        filter,
+        output_shape,
+        strides,
+        padding=padding.upper(),
+        data_format=d_format,
+    )
+
+
+def _as_tuple(value, rank):
+    if isinstance(value, (tuple, list)):
+        return tuple(value)
+    return (value,) * rank
+
+
 def conv1d_transpose(
     inputs,
     filter,
@@ -49,41 +133,15 @@ def conv1d_transpose(
 
     Take a filter defined for forward convolution and adjusts it for a
     transposed convolution."""
-    if isinstance(kernel_size, tuple):
-        kernel_size = kernel_size[0]
-    input_shape = inputs.shape
-    batch_size = input_shape[0]
-    if data_format == "channels_first":
-        w_axis = 2
-        d_format = "NCW"
-    else:
-        w_axis = 1
-        d_format = "NWC"
-
-    width = input_shape[w_axis]
-
-    # Infer the dynamic output shape:
-    out_width = conv_utils.deconv_output_length(
-        input_length=width,
-        filter_size=kernel_size,
-        padding=padding,
-        output_padding=output_padding,
-        stride=strides,
-    )
-
-    if data_format == "channels_first":
-        output_shape = (batch_size, filters, out_width)
-    else:
-        output_shape = (batch_size, out_width, filters)
-
-    filter = K.permute_dimensions(filter, (0, 2, 1))
-    return tf.nn.conv1d_transpose(
+    return _conv_transpose(
         inputs,
         filter,
-        output_shape,
-        strides,
-        padding=padding.upper(),
-        data_format=d_format,
+        _as_tuple(kernel_size, 1),
+        filters,
+        _as_tuple(strides, 1),
+        padding,
+        None if output_padding is None else _as_tuple(output_padding, 1),
+        data_format,
     )
 
 
@@ -101,41 +159,38 @@ def conv2d_transpose(
 
     Take a filter defined for forward convolution and adjusts it for a
     transposed convolution."""
-    input_shape = inputs.shape
-    batch_size = input_shape[0]
-    if data_format == "channels_first":
-        h_axis, w_axis = 2, 3
-    else:
-        h_axis, w_axis = 1, 2
-
-    height, width = input_shape[h_axis], input_shape[w_axis]
-    kernel_h, kernel_w = kernel_size
-    stride_h, stride_w = strides
-
-    # Infer the dynamic output shape:
-    out_height = conv_utils.deconv_output_length(
-        input_length=height,
-        filter_size=kernel_h,
-        padding=padding,
-        output_padding=output_padding,
-        stride=stride_h,
-    )
-    out_width = conv_utils.deconv_output_length(
-        input_length=width,
-        filter_size=kernel_w,
-        padding=padding,
-        output_padding=output_padding,
-        stride=stride_w,
+    return _conv_transpose(
+        inputs,
+        filter,
+        _as_tuple(kernel_size, 2),
+        filters,
+        _as_tuple(strides, 2),
+        padding,
+        None if output_padding is None else _as_tuple(output_padding, 2),
+        data_format,
     )
 
-    if data_format == "channels_first":
-        output_shape = (batch_size, filters, out_height, out_width)
-    else:
-        output_shape = (batch_size, out_height, out_width, filters)
 
-    filter = K.permute_dimensions(filter, (0, 1, 3, 2))
-    return K.conv2d_transpose(
-        inputs, filter, output_shape, strides, padding=padding, data_format=data_format
+def conv3d_transpose(
+    inputs,
+    filter,
+    kernel_size=None,
+    filters=None,
+    strides=(1, 1, 1),
+    padding="SAME",
+    output_padding=None,
+    data_format="channels_last",
+):
+    """Transposed 3D convolution with a kernel defined for forward convolution"""
+    return _conv_transpose(
+        inputs,
+        filter,
+        _as_tuple(kernel_size, 3),
+        filters,
+        _as_tuple(strides, 3),
+        padding,
+        None if output_padding is None else _as_tuple(output_padding, 3),
+        data_format,
     )
 
 
@@ -482,7 +537,11 @@ def call(self, inputs):
             convArgs.pop("dilation_rate", None)
             convArgs["kernel_size"] = self.kernel_size
             convArgs["filters"] = 2 * self.filters
-            convFunc = {1: conv1d_transpose, 2: conv2d_transpose}[self.rank]
+            convFunc = {
+                1: conv1d_transpose,
+                2: conv2d_transpose,
+                3: conv3d_transpose,
+            }[self.rank]
         else:
             convFunc = {1: K.conv1d, 2: K.conv2d, 3: K.conv3d}[self.rank]