ConvNet¶
- class torchrl.modules.ConvNet(in_features: ~typing.Optional[int] = None, depth: ~typing.Optional[int] = None, num_cells: ~typing.Optional[~typing.Union[~typing.Sequence[int], int]] = None, kernel_sizes: ~typing.Union[~typing.Sequence[int], int] = 3, strides: ~typing.Union[~typing.Sequence[int], int] = 1, paddings: ~typing.Union[~typing.Sequence[int], int] = 0, activation_class: ~typing.Union[~typing.Type[~torch.nn.modules.module.Module], ~typing.Callable] = <class 'torch.nn.modules.activation.ELU'>, activation_kwargs: ~typing.Optional[~typing.Union[dict, ~typing.List[dict]]] = None, norm_class: ~typing.Optional[~typing.Union[~typing.Type[~torch.nn.modules.module.Module], ~typing.Callable]] = None, norm_kwargs: ~typing.Optional[~typing.Union[dict, ~typing.List[dict]]] = None, bias_last_layer: bool = True, aggregator_class: ~typing.Optional[~typing.Union[~typing.Type[~torch.nn.modules.module.Module], ~typing.Callable]] = <class 'torchrl.modules.models.utils.SquashDims'>, aggregator_kwargs: ~typing.Optional[dict] = None, squeeze_output: bool = False, device: ~typing.Optional[~typing.Union[~torch.device, str, int]] = None)[source]¶
A convolutional neural network.
- Parameters:
in_features (int, optional) – number of input features. If
None
, aLazyConv2d
module is used for the first layer.;depth (int, optional) – depth of the network. A depth of 1 will produce a single linear layer network with the desired input size, and with an output size equal to the last element of the num_cells argument. If no depth is indicated, the depth information should be contained in the
num_cells
argument (see below). Ifnum_cells
is an iterable anddepth
is indicated, both should match:len(num_cells)
must be equal to thedepth
.num_cells (int or Sequence of int, optional) – number of cells of every layer in between the input and output. If an integer is provided, every layer will have the same number of cells. If an iterable is provided, the linear layers
out_features
will match the content of num_cells. Defaults to[32, 32, 32]
.kernel_sizes (int, sequence of int, optional) – Kernel size(s) of the conv network. If iterable, the length must match the depth, defined by the
num_cells
or depth arguments. Defaults to3
.strides (int or sequence of int, optional) – Stride(s) of the conv network. If iterable, the length must match the depth, defined by the
num_cells
or depth arguments. Defaults to1
.activation_class (Type[nn.Module] or callable, optional) – activation class or constructor to be used. Defaults to
Tanh
.activation_kwargs (dict or list of dicts, optional) – kwargs to be used with the activation class. A list of kwargs of length
depth
can also be passed, with one element per layer.norm_class (Type or callable, optional) – normalization class or constructor, if any.
norm_kwargs (dict or list of dicts, optional) – kwargs to be used with the normalization layers. A list of kwargs of length
depth
can also be passed, with one element per layer.bias_last_layer (bool) – if
True
, the last Linear layer will have a bias parameter. Defaults toTrue
.aggregator_class (Type[nn.Module] or callable) – aggregator class or constructor to use at the end of the chain. Defaults to
torchrl.modules.utils.models.SquashDims
;aggregator_kwargs (dict, optional) – kwargs for the
aggregator_class
.squeeze_output (bool) – whether the output should be squeezed of its singleton dimensions. Defaults to
False
.device (torch.device, optional) – device to create the module on.
Examples
>>> # All of the following examples provide valid, working MLPs >>> cnet = ConvNet(in_features=3, depth=1, num_cells=[32,]) # MLP consisting of a single 3 x 6 linear layer >>> print(cnet) ConvNet( (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1)) (1): ELU(alpha=1.0) (2): SquashDims() ) >>> cnet = ConvNet(in_features=3, depth=4, num_cells=32) >>> print(cnet) ConvNet( (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1)) (1): ELU(alpha=1.0) (2): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1)) (3): ELU(alpha=1.0) (4): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1)) (5): ELU(alpha=1.0) (6): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1)) (7): ELU(alpha=1.0) (8): SquashDims() ) >>> cnet = ConvNet(in_features=3, num_cells=[32, 33, 34, 35]) # defines the depth by the num_cells arg >>> print(cnet) ConvNet( (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1)) (1): ELU(alpha=1.0) (2): Conv2d(32, 33, kernel_size=(3, 3), stride=(1, 1)) (3): ELU(alpha=1.0) (4): Conv2d(33, 34, kernel_size=(3, 3), stride=(1, 1)) (5): ELU(alpha=1.0) (6): Conv2d(34, 35, kernel_size=(3, 3), stride=(1, 1)) (7): ELU(alpha=1.0) (8): SquashDims() ) >>> cnet = ConvNet(in_features=3, num_cells=[32, 33, 34, 35], kernel_sizes=[3, 4, 5, (2, 3)]) # defines kernels, possibly rectangular >>> print(cnet) ConvNet( (0): Conv2d(3, 32, kernel_size=(3, 3), stride=(1, 1)) (1): ELU(alpha=1.0) (2): Conv2d(32, 33, kernel_size=(4, 4), stride=(1, 1)) (3): ELU(alpha=1.0) (4): Conv2d(33, 34, kernel_size=(5, 5), stride=(1, 1)) (5): ELU(alpha=1.0) (6): Conv2d(34, 35, kernel_size=(2, 3), stride=(1, 1)) (7): ELU(alpha=1.0) (8): SquashDims() )
- forward(inputs: Tensor) Tensor [source]¶
Define the computation performed at every call.
Should be overridden by all subclasses.
Note
Although the recipe for forward pass needs to be defined within this function, one should call the
Module
instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.