fix minor problems in the docs (#138)

* fix minor probmels in the docs

* typo

* remove pip source

* fix monitor

* add performance of A2C

* Update README.md

* modify logger for GPU detection

.teamcity/build.sh

@@ -155,8 +155,8 @@ function main() {
               echo Running tests in $env ..
               echo `which pip`
               echo ========================================
-              pip install -i https://pypi.tuna.tsinghua.edu.cn/simple .
-              pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r .teamcity/requirements.txt
+              pip install .
+              pip install -r .teamcity/requirements.txt
               run_test_with_cpu $env
               run_test_with_cpu $env "DIS_TESTING_SERIALLY"
           done

docs/parallel_training/overview.rst

-Summary
+Overview
 =======
 
 Easy-to-use
@@ -22,10 +22,11 @@ Web UI for computation resources
  to the cluster. Users can view the cluster status at a WEB UI. It shows the
  detailed information for each worker(e.g, memory used) and each task submitted.
 
-Board compatibility
+
+Supporting vairous frameworks
 ###################
-| Our framework for distributed training is compatible with any other
- frameworks, like tensorflow, pytorch or mxnet. By adding `@parl.remote_class`
+| PARL for distributed training is compatible with any other
+ frameworks, like tensorflow, pytorch and mxnet. By adding `@parl.remote_class`
  decorator to their codes, users can easily convert their codes to distributed
  computation.
 
@@ -33,7 +34,7 @@ Why PARL
 ########
 
 High throughput
-###############
+-------------------------
 | PARL uses a point-to-point connection for network communication in the
  cluster. Unlike other framework like RLlib which replies on redis for
  communication, PARL is able to achieve much higher throughput. The results
@@ -41,7 +42,7 @@ High throughput
  achieved an increase of 160% on data throughout over Ray(RLlib).
 
 Automatic deployment
-####################
+-------------------------
 | Unlike other parallel frameworks which fail to import modules from
  external file, PARL will automatically package all related files and send
  them to remote machines.

docs/parallel_training/setup.rst

@@ -61,4 +61,4 @@ Further Reading
 | Now we know how to set up a cluster and use this cluster by simply adding ``@parl.remote_class``. 
 | In `next_tutorial`_, we will show how this decorator help us implement the **real** multi-thread computation in Python, breaking the limitation of Python Global Interpreter Lock(GIL).
 
-.. _`next_tutorial`: https://parl.readthedocs.io/parallel_training/recommended_practice.html
+.. _`next_tutorial`: https://parl.readthedocs.io/en/latest/parallel_training/recommended_practice.html

examples/A2C/README.md

@@ -8,11 +8,14 @@ Please see [here](https://gym.openai.com/envs/#atari) to know more about Atari g
 
 ### Benchmark result
 Mean episode reward in training process after 10 million sample steps.
+![learninng_curve](learning_curve.png)
+
+Performance of A2C on various envrionments
+
+<p align="center">
+<img src="result.png" alt="result" width="700"/>
+</p>
 
-|              |                |                  |               |                     |
-|--------------|----------------|------------------|---------------|---------------------|
-| Alien (1278) | Amidar (380)   | Assault (4659)   | Aterix (3883) | Atlantis (3040000)  |
-| Pong (20)    | Breakout (405) | Beamrider (3394) | Qbert (14528) | SpaceInvaders (819) |
 
 ## How to use
 ### Dependencies

parl/remote/tests/cluster_monitor_3_test.py

@@ -81,12 +81,12 @@ class TestClusterMonitor(unittest.TestCase):
 
         for i in range(10):
             workers[i].exit()
-        time.sleep(40)
+        time.sleep(60)
         self.assertEqual(10, len(cluster_monitor.data['workers']))
 
         for i in range(10, 20):
             workers[i].exit()
-        time.sleep(40)
+        time.sleep(60)
         self.assertEqual(0, len(cluster_monitor.data['workers']))
 
         master.exit()

parl/utils/machine_info.py

@@ -77,7 +77,7 @@ def get_gpu_count():
             logger.info(
                 'CUDA_VISIBLE_DEVICES found gpu count: {}'.format(gpu_count))
         except:
-            logger.warning('Cannot find available GPU devices, using CPU now.')
+            logger.info('Cannot find available GPU devices, using CPU now.')
             gpu_count = 0
     else:
         try:
@@ -85,7 +85,7 @@ def get_gpu_count():
                                                      "-L"])).count('UUID')
             logger.info('nvidia-smi -L found gpu count: {}'.format(gpu_count))
         except:
-            logger.warning('Cannot find available GPU devices, using CPU now.')
+            logger.info('Cannot find available GPU devices, using CPU now.')
             gpu_count = 0
     return gpu_count