Merge branch 'master' into fix-sarsa-gpu

.pfnci/run.sh

@@ -54,10 +54,19 @@ main() {
       --hint="/chainerrl/.pfnci/hint.pbtxt"
   )
 
+  UBUNTU_VERSION_ID=$(grep DISTRIB_RELEASE /etc/lsb-release | cut -d "=" -f2)
+  if [ "$UBUNTU_VERSION_ID" = "16.04" ]; then
+    # Because ffmpeg of ubuntu 16.04 causes segmentation fault,
+    # we use jonathonf/ffmpeg-3
+    apt-get update -q
+    apt-get install -qy --no-install-recommends software-properties-common
+    add-apt-repository ppa:jonathonf/ffmpeg-3
+  fi
+
   apt-get update -q
   apt-get install -qy --no-install-recommends \
       "${PYTHON}-dev" "${PYTHON}-pip" "${PYTHON}-setuptools" \
-      zlib1g-dev make cmake g++ git
+      zlib1g-dev make cmake g++ git ffmpeg freeglut3-dev xvfb
 
   if [ "${CHAINER}" != '' ]; then
     "${PYTHON}" -m pip install "chainer==${CHAINER}"
@@ -77,7 +86,10 @@ main() {
   git config --global user.email "you@example.com"
   git config --global user.name "Your Name"
 
+  # Xvfb's default screen is 1280x1024x8, which seems to cause a problem.
+  # https://bugzilla.redhat.com/show_bug.cgi?id=904851
   OMP_NUM_THREADS=1 PYTHONHASHSEED=0 \
+      xvfb-run --server-args="-screen 0 1280x800x24" \
       xpytest "${xpytest_args[@]}" '/chainerrl/tests/**/test_*.py'
 }
 

chainerrl/agents/a2c.py

@@ -110,7 +110,7 @@ def __init__(self, model, optimizer, gamma, num_processes,
         self.model = model
         self.gpu = gpu
         if gpu is not None and gpu >= 0:
-            chainer.cuda.get_device(gpu).use()
+            chainer.cuda.get_device_from_id(gpu).use()
             self.model.to_gpu(device=gpu)
 
         self.optimizer = optimizer

chainerrl/agents/ddpg.py

@@ -105,7 +105,7 @@ def __init__(self, model, actor_optimizer, critic_optimizer, replay_buffer,
         self.model = model
 
         if gpu is not None and gpu >= 0:
-            cuda.get_device(gpu).use()
+            cuda.get_device_from_id(gpu).use()
             self.model.to_gpu(device=gpu)
 
         self.xp = self.model.xp

chainerrl/agents/dqn.py

@@ -159,7 +159,7 @@ def __init__(self, q_function, optimizer, replay_buffer, gamma,
         self.q_function = q_function  # For backward compatibility
 
         if gpu is not None and gpu >= 0:
-            cuda.get_device(gpu).use()
+            cuda.get_device_from_id(gpu).use()
             self.model.to_gpu(device=gpu)
 
         self.xp = self.model.xp

chainerrl/agents/pgt.py

@@ -81,7 +81,7 @@ def __init__(self, model, actor_optimizer, critic_optimizer, replay_buffer,
         self.model = model
 
         if gpu is not None and gpu >= 0:
-            cuda.get_device(gpu).use()
+            cuda.get_device_from_id(gpu).use()
             self.model.to_gpu(device=gpu)
 
         self.xp = self.model.xp

examples/gym/train_pcl_gym.py

@@ -157,7 +157,7 @@ def make_env(process_idx, test):
         )
 
     if not args.train_async and args.gpu >= 0:
-        chainer.cuda.get_device(args.gpu).use()
+        chainer.cuda.get_device_from_id(args.gpu).use()
         model.to_gpu(args.gpu)
 
     if args.train_async:

examples/gym/train_reinforce_gym.py

@@ -109,7 +109,7 @@ def make_env(test):
         os.path.join(args.outdir, 'model'))
 
     if args.gpu >= 0:
-        chainer.cuda.get_device(args.gpu).use()
+        chainer.cuda.get_device_from_id(args.gpu).use()
         model.to_gpu(args.gpu)
 
     opt = chainer.optimizers.Adam(alpha=args.lr)

examples/mujoco/train_trpo_gym.py

@@ -156,7 +156,7 @@ def make_env(test):
 
     # Draw the computational graph and save it in the output directory.
     fake_obs = chainer.Variable(
-        policy.xp.zeros_like(obs_space.low, dtype=np.float32)[None],
+        policy.xp.zeros(obs_space.low.shape, dtype=np.float32)[None],
         name='observation')
     chainerrl.misc.draw_computational_graph(
         [policy(fake_obs)], os.path.join(args.outdir, 'policy'))

tests/agents_tests/test_reinforce.py

@@ -116,7 +116,7 @@ def phi(x):
                 )
 
         if gpu >= 0:
-            chainer.cuda.get_device(gpu).use()
+            chainer.cuda.get_device_from_id(gpu).use()
             model.to_gpu()
 
         opt = optimizers.Adam()

tests/links_tests/test_stateless_recurrent_sequential.py

@@ -10,7 +10,12 @@
 
 import chainer
 from chainer import functions as F
-from chainer.functions.connection.n_step_lstm import _lstm
+try:
+    # chainer<=7.0.0b2
+    from chainer.functions.connection.n_step_lstm import _lstm
+except ImportError:
+    # chainer>=7.0.0b3 (https://github.com/chainer/chainer/pull/7725)
+    from chainer.functions.rnn.n_step_lstm import _lstm
 from chainer import links as L
 from chainer import testing
 import numpy as np

tests/test_distribution.py

@@ -21,7 +21,7 @@ class TestSampleDiscreteActions(unittest.TestCase):
 
     def _test(self, gpu):
         if gpu >= 0:
-            chainer.cuda.get_device(gpu).use()
+            chainer.cuda.get_device_from_id(gpu).use()
             xp = chainer.cuda.cupy
         else:
             xp = np