Skip to content
Projects
Groups
Snippets
Help
Loading...
Help
Support
Keyboard shortcuts
?
Submit feedback
Contribute to GitLab
Sign in
Toggle navigation
S
seminar-breakout
Project overview
Project overview
Details
Activity
Releases
Repository
Repository
Files
Commits
Branches
Tags
Contributors
Graph
Compare
Issues
0
Issues
0
List
Boards
Labels
Milestones
Merge Requests
0
Merge Requests
0
CI / CD
CI / CD
Pipelines
Jobs
Schedules
Analytics
Analytics
CI / CD
Repository
Value Stream
Wiki
Wiki
Members
Members
Collapse sidebar
Close sidebar
Activity
Graph
Create a new issue
Jobs
Commits
Issue Boards
Open sidebar
Shashank Suhas
seminar-breakout
Commits
ba4fcf2a
Commit
ba4fcf2a
authored
Jul 17, 2016
by
Yuxin Wu
Browse files
Options
Browse Files
Download
Email Patches
Plain Diff
update docs
parent
f8969459
Changes
3
Show whitespace changes
Inline
Side-by-side
Showing
3 changed files
with
10 additions
and
8 deletions
+10
-8
examples/Atari2600/README.md
examples/Atari2600/README.md
+6
-5
tensorpack/callbacks/param.py
tensorpack/callbacks/param.py
+3
-3
tensorpack/dataflow/remote.py
tensorpack/dataflow/remote.py
+1
-0
No files found.
examples/Atari2600/README.md
View file @
ba4fcf2a
...
@@ -13,11 +13,14 @@ Claimed performance in the paper can be reproduced, on several games I've tested
...
@@ -13,11 +13,14 @@ Claimed performance in the paper can be reproduced, on several games I've tested


DQN was trained on 1 GPU and it typically took 2~3 days of training to reach a score of 400 on breakout game.
DQN typically took 2 days of training to reach a score of 400 on breakout game.
My Batch-A3C implementation only took <2 hours with 2 GPUs (one for training and one for simulation).
My Batch-A3C implementation only took <2 hours (one for training and one for simulation).
Both were trained on one GPU with an extra GPU for simulation.
This is probably the fastest RL trainer you'd find.
This is probably the fastest RL trainer you'd find.
The x-axis is the number of iterations not wall time, but iteration speed is about 7.8it/s for both models.
The x-axis is the number of iterations, not wall time.
Iteration speed on Tesla M40 is about 10.7it/s for B-A3C.
D-DQN is faster at the beginning but will converge to 12it/s due of exploration annealing.
A demo trained with Double-DQN on breakout is available at
[
youtube
](
https://youtu.be/o21mddZtE5Y
)
.
A demo trained with Double-DQN on breakout is available at
[
youtube
](
https://youtu.be/o21mddZtE5Y
)
.
...
@@ -30,8 +33,6 @@ To train:
...
@@ -30,8 +33,6 @@ To train:
```
```
./DQN.py --rom breakout.bin --gpu 0
./DQN.py --rom breakout.bin --gpu 0
```
```
Training speed is about 7.3 iteration/s on 1 Tesla M40
(faster than this at the beginning, but will slow down due to exploration annealing).
To visualize the agent:
To visualize the agent:
```
```
...
...
tensorpack/callbacks/param.py
View file @
ba4fcf2a
...
@@ -142,9 +142,9 @@ class HumanHyperParamSetter(HyperParamSetter):
...
@@ -142,9 +142,9 @@ class HumanHyperParamSetter(HyperParamSetter):
ret
=
dic
[
self
.
param
.
readable_name
]
ret
=
dic
[
self
.
param
.
readable_name
]
return
ret
return
ret
except
:
except
:
logger
.
warn
(
#
logger.warn(
"Failed to
find {} in {}"
.
format
(
#"Cannot
find {} in {}".format(
self
.
param
.
readable_name
,
self
.
file_name
))
#
self.param.readable_name, self.file_name))
return
None
return
None
class
ScheduledHyperParamSetter
(
HyperParamSetter
):
class
ScheduledHyperParamSetter
(
HyperParamSetter
):
...
...
tensorpack/dataflow/remote.py
View file @
ba4fcf2a
...
@@ -3,6 +3,7 @@
...
@@ -3,6 +3,7 @@
# File: remote.py
# File: remote.py
# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
# Author: Yuxin Wu <ppwwyyxxc@gmail.com>
from
..utils
import
logger
try
:
try
:
import
zmq
import
zmq
except
ImportError
:
except
ImportError
:
...
...
Write
Preview
Markdown
is supported
0%
Try again
or
attach a new file
Attach a file
Cancel
You are about to add
0
people
to the discussion. Proceed with caution.
Finish editing this message first!
Cancel
Please
register
or
sign in
to comment