import numpy as np
from xuance.common import Sequence, Optional, Union, Callable
from xuance.tensorflow import tf, tk, Module
from xuance.tensorflow.utils.layers import cnn_block, mlp_block
from xuance.tensorflow.utils import ModuleType
[docs]
class Basic_CNN(Module):
def __init__(self,
input_shape: Sequence[int],
kernels: Sequence[int],
strides: Sequence[int],
filters: Sequence[int],
normalize: Optional[tk.layers.Layer] = None,
initialize: Optional[tk.initializers.Initializer] = None,
activation: Optional[tk.layers.Layer] = None,
device: Optional[str] = None,
**kwargs):
super(Basic_CNN, self).__init__()
self.input_shapes = (input_shape[2], input_shape[0], input_shape[1])
self.kernels = kernels
self.strides = strides
self.filters = filters
self.normalize = normalize
self.initialize = initialize
self.activation = activation
self.device = device
self.output_shapes = {'state': (filters[-1],)}
self.model = self._create_network()
def _create_network(self):
layers = []
input_shape = self.input_shape
for k, s, f in zip(self.kernels, self.strides, self.filters):
cnn, input_shape = cnn_block(input_shape, f, k, s, self.normalize, self.activation, self.initialize,
self.device)
layers.extend(cnn)
layers.append(tfa.layers.AdaptiveMaxPooling2D((1, 1)))
layers.append(tk.layers.Flatten())
return tk.Sequential(*layers)
@tf.function
def call(self, observations: np.ndarray, **kwargs):
with tf.device(self.device):
tensor_observation = tf.convert_to_tensor(np.transpose(observations, (0, 3, 1, 2)), dtype=tf.float32)
return {'state': self.model(tensor_observation)}
[docs]
class AC_CNN_Atari(Module):
def __init__(self,
input_shape: Sequence[int],
kernels: Sequence[int],
strides: Sequence[int],
filters: Sequence[int],
normalize: Optional[tk.layers.Layer] = None,
initialize: Optional[tk.initializers.Initializer] = None,
activation: Optional[tk.layers.Layer] = None,
device: Optional[str] = None,
fc_hidden_sizes: Sequence[int] = (),
**kwargs):
super(AC_CNN_Atari, self).__init__()
self.input_shapes = (input_shape[2], input_shape[0], input_shape[1]) # Channels x Height x Width
self.kernels = kernels
self.strides = strides
self.filters = filters
self.normalize = normalize
self.initialize = initialize
self.activation = activation
self.device = device
self.fc_hidden_sizes = fc_hidden_sizes
self.output_shapes = {'state': (fc_hidden_sizes[-1],)}
self.model = self._create_network()
def _init_layer(self, layer, gain=np.sqrt(2), bias=0.0):
initializer = tf.keras.initializers.Orthogonal(gain=gain)
nn.init.orthogonal_(layer.weight, gain=gain)
nn.init.constant_(layer.bias, bias)
return layer
def _create_network(self):
layers = []
input_shape = self.input_shape
for k, s, f in zip(self.kernels, self.strides, self.filters):
cnn, input_shape = cnn_block(input_shape, f, k, s, None, self.activation, None, self.device)
cnn[0] = self._init_layer(cnn[0])
layers.extend(cnn)
layers.append(nn.Flatten())
input_shape = (np.prod(input_shape, dtype=np.int32), )
for h in self.fc_hidden_sizes:
mlp, input_shape = mlp_block(input_shape[0], h, None, self.activation, None, self.device)
mlp[0] = self._init_layer(mlp[0])
layers.extend(mlp)
return nn.Sequential(*layers)
[docs]
def forward(self, observations: np.ndarray):
observations = observations / 255.0
tensor_observation = torch.as_tensor(np.transpose(observations, (0, 3, 1, 2)), dtype=torch.float32,
device=self.device)
return {'state': self.model(tensor_observation)}