Conv1D and Conv2D building blocks for convolutional neural networks.
Provides composable convolutional blocks and full model builders for both 1D (sequence) and 2D (image) processing. Each block follows the pattern: convolution -> batch normalization -> activation -> dropout.
Architecture (Conv1D)
Input [batch, seq_len, channels]
|
v
+---------------------------+
| Conv1D -> BN -> Act -> Drop| Block 1
+---------------------------+
|
v
+---------------------------+
| Conv1D -> BN -> Act -> Drop| Block 2
+---------------------------+
| (optional pooling)
v
Output [batch, seq_len', channels']Usage
# Build a 1D convolutional model for sequence processing
model = Conv.build_conv1d(
input_size: 64,
channels: [128, 256, 512],
kernel_sizes: [3, 3, 3],
activation: :relu,
dropout: 0.1,
pooling: :max
)
# Build a 2D convolutional model for image processing
model = Conv.build_conv2d(
input_shape: {nil, 32, 32, 3},
channels: [32, 64, 128],
kernel_sizes: [3, 3, 3],
activation: :relu,
dropout: 0.1,
pooling: :max
)
# Use individual blocks as building blocks
input = Axon.input("input", shape: {nil, 100, 64})
block = Conv.conv1d_block(input, 128, kernel_size: 3, activation: :gelu)Summary
Functions
Alias for build_conv1d/1.
Build a 1D convolutional model for sequence processing.
Build a 2D convolutional model for image processing.
Build a single conv1d -> batch_norm -> activation -> dropout block.
Build a single conv2d -> batch_norm -> activation -> dropout block.
Get the output channel count for a convolutional model with the given options.
Types
@type build_opt() :: {:input_size, pos_integer()} | {:channels, [pos_integer()]} | {:kernel_sizes, pos_integer() | [pos_integer()]} | {:activation, atom()} | {:dropout, float()} | {:pooling, atom() | nil} | {:seq_len, pos_integer() | nil}
Options for build/1.
Functions
Alias for build_conv1d/1.
Build a 1D convolutional model for sequence processing.
Options
:input_size- Number of input channels/features (required):channels- List of output channel counts per layer (default: [64, 128, 256]):kernel_sizes- List of kernel sizes per layer, or single integer for all layers (default: 3):activation- Activation function atom (default: :relu):dropout- Dropout rate (default: 0.1):pooling- Pooling type::max,:avg, ornilfor no pooling (default: nil):seq_len- Expected sequence length, or nil for dynamic (default: nil)
Returns
An Axon model taking [batch, seq_len, input_size] and outputting
[batch, seq_len', last_channel_count].
Build a 2D convolutional model for image processing.
Options
:input_shape- Input shape as{nil, height, width, channels}(required):channels- List of output channel counts per layer (default: [32, 64, 128]):kernel_sizes- List of kernel sizes per layer, or single integer for all layers (default: 3):activation- Activation function atom (default: :relu):dropout- Dropout rate (default: 0.1):pooling- Pooling type::max,:avg, ornilfor no pooling (default: nil)
Returns
An Axon model taking [batch, height, width, channels] (NHWC) and outputting
[batch, height', width', last_channel_count].
@spec conv1d_block(Axon.t(), pos_integer(), keyword()) :: Axon.t()
Build a single conv1d -> batch_norm -> activation -> dropout block.
Parameters
input- Input Axon node with shape[batch, seq_len, channels]out_channels- Number of output channels
Options
:kernel_size- Convolution kernel size (default: 3):activation- Activation function atom (default: :relu):dropout- Dropout rate, 0.0 to skip (default: 0.1):name- Layer name prefix (default: "conv1d_block"):padding- Padding mode::sameor:valid(default: :same):strides- Convolution stride (default: 1)
Returns
An Axon node with shape [batch, seq_len', out_channels].
@spec conv2d_block(Axon.t(), pos_integer(), keyword()) :: Axon.t()
Build a single conv2d -> batch_norm -> activation -> dropout block.
Parameters
input- Input Axon node with shape[batch, height, width, channels]out_channels- Number of output channels
Options
:kernel_size- Convolution kernel size (default: 3):activation- Activation function atom (default: :relu):dropout- Dropout rate, 0.0 to skip (default: 0.1):name- Layer name prefix (default: "conv2d_block"):padding- Padding mode::sameor:valid(default: :same):strides- Convolution stride (default: 1)
Returns
An Axon node with shape [batch, height', width', out_channels].
@spec output_size(keyword()) :: pos_integer()
Get the output channel count for a convolutional model with the given options.