cherry.nn.init¶
cherry.nn.init.robotics_init_(module, gain = None)
¶
Description¶
Default initialization for robotic control.
Credit¶
Adapted from Ilya Kostrikov's implementation, itself inspired from OpenAI Baslines.
Arguments¶
module(nn.Module) - Module to initialize.gain(float, optional, default=sqrt(2.0)) - Gain of orthogonal initialization.
Returns¶
- Module, whose weight and bias have been modified in-place.
Example¶
linear = nn.Linear(23, 5)
kostrikov_robotics_(linear)
cherry.nn.init.atari_init_(module, gain = None)
¶
Description¶
Default initialization for Atari environments.
Credit¶
Adapted from Ilya Kostrikov's implementation, itself inspired from OpenAI Baslines.
Arguments¶
module(nn.Module) - Module to initialize.gain(float, optional, default=None) - Gain of orthogonal initialization. Default is computed for ReLU activation withtorch.nn.init.calculate_gain('relu').
Returns¶
- Module, whose weight and bias have been modified in-place.
Example¶
linear = nn.Linear(23, 5)
atari_init_(linear)