gryf/coach
1
0
Fork 0
mirror of https://github.com/gryf/coach.git synced 2025-12-17 19:20:19 +01:00
Code Issues Projects Releases Wiki Activity

parameter noise exploration - using Noisy Nets

Browse Source
This commit is contained in:
Gal Leibovich
2018-08-27 18:19:01 +03:00
parent 658b437079
commit 1aa2ab0590
49 changed files with 536 additions and 433 deletions
Download Patch File Download Diff File Expand all files Collapse all files

70
rl_coach/architectures/tensorflow_components/embedders/image_embedder.py
View File

@@ -18,7 +18,7 @@ from typing import List
import tensorflow as tf
from rl_coach.architectures.tensorflow_components.architecture import Conv2d
from rl_coach.architectures.tensorflow_components.architecture import Conv2d, Dense
from rl_coach.architectures.tensorflow_components.embedders.embedder import InputEmbedder
from rl_coach.base_parameters import EmbedderScheme
from rl_coach.core_types import InputImageEmbedding
@@ -30,45 +30,49 @@ class ImageEmbedder(InputEmbedder):
The embedder is intended for image like inputs, where the channels are expected to be the last axis.
The embedder also allows custom rescaling of the input prior to the neural network.
"""
schemes = {
EmbedderScheme.Empty:
[],
EmbedderScheme.Shallow:
[
Conv2d([32, 3, 1])
],
# atari dqn
EmbedderScheme.Medium:
[
Conv2d([32, 8, 4]),
Conv2d([64, 4, 2]),
Conv2d([64, 3, 1])
],
# carla
EmbedderScheme.Deep: \
[
Conv2d([32, 5, 2]),
Conv2d([32, 3, 1]),
Conv2d([64, 3, 2]),
Conv2d([64, 3, 1]),
Conv2d([128, 3, 2]),
Conv2d([128, 3, 1]),
Conv2d([256, 3, 2]),
Conv2d([256, 3, 1])
]
}
def __init__(self, input_size: List[int], activation_function=tf.nn.relu,
scheme: EmbedderScheme=EmbedderScheme.Medium, batchnorm: bool=False, dropout: bool=False,
name: str= "embedder", input_rescaling: float=255.0, input_offset: float=0.0, input_clipping=None):
name: str= "embedder", input_rescaling: float=255.0, input_offset: float=0.0, input_clipping=None,
dense_layer=Dense):
super().__init__(input_size, activation_function, scheme, batchnorm, dropout, name, input_rescaling,
input_offset, input_clipping)
input_offset, input_clipping, dense_layer=dense_layer)
self.return_type = InputImageEmbedding
if len(input_size) != 3 and scheme != EmbedderScheme.Empty:
raise ValueError("Image embedders expect the input size to have 3 dimensions. The given size is: {}"
.format(input_size))
@property
def schemes(self):
return {
EmbedderScheme.Empty:
[],
EmbedderScheme.Shallow:
[
Conv2d([32, 3, 1])
],
# atari dqn
EmbedderScheme.Medium:
[
Conv2d([32, 8, 4]),
Conv2d([64, 4, 2]),
Conv2d([64, 3, 1])
],
# carla
EmbedderScheme.Deep: \
[
Conv2d([32, 5, 2]),
Conv2d([32, 3, 1]),
Conv2d([64, 3, 2]),
Conv2d([64, 3, 1]),
Conv2d([128, 3, 2]),
Conv2d([128, 3, 1]),
Conv2d([256, 3, 2]),
Conv2d([256, 3, 1])
]
}