v1.3.0 release bump.

- pad tolerance for lstm test
- deprecate deterministic argument
- Fix py test to use 2x2 winograd

ChangeLog

@@ -1,6 +1,17 @@
 # ChangeLog
 
-## v1.2.2 (2106-02-24):
+## v1.3.0 (2016-02-29):
+
+### Modifications
+
+* benchmarking scripts
+* winograd kernels and associated autotuning routines
+* deprecation of deterministic argument for backend constructor
+* improve batch norm stability with fp16 backend
+* allow strided support for dimshuffle kernel
+* speed up zero momentum gradient descent
+
+## v1.2.2 (2016-02-24):
 
 ### Modifications
 

conftest.py

@@ -50,6 +50,8 @@ def backend_default(request):
                      datatype=np.float32,
                      batch_size=128,
                      rng_seed=0)
+    if request.param == 'gpu':
+        be.enable_winograd = 2 if be.enable_winograd else be.enable_winograd
 
     # add a cleanup call - will run after all test in module are done
     def cleanup():

neon/backends/__init__.py

@@ -31,7 +31,7 @@
 def gen_backend(backend='cpu', rng_seed=None, datatype=np.float32,
                 batch_size=0, stochastic_round=False, device_id=0,
                 max_devices=get_device_count(), compat_mode=None,
-                deterministic_update=False, deterministic=True,
+                deterministic_update=None, deterministic=None,
                 cache_dir=os.path.join(os.path.expanduser('~'), 'nervana/cache')):
     """
     Construct and return a backend instance of the appropriate type based on
@@ -81,9 +81,10 @@ def gen_backend(backend='cpu', rng_seed=None, datatype=np.float32,
         # NervanaObject.be instead of a global
         atexit.register(cleanup_backend)
 
-    if deterministic_update:
-       deterministic = True
-       logger.warning("--deterministic_update is deprecated in favor of --deterministic")
+    if deterministic_update is not None or deterministic is not None:
+       logger.warning('deterministic_update and deterministic args are deprecated in favor of '
+                      'specifying random seed')
+       deterministic = None
 
     if backend == 'cpu' or backend is None:
         from neon.backends.nervanacpu import NervanaCPU
@@ -115,7 +116,6 @@ def gen_backend(backend='cpu', rng_seed=None, datatype=np.float32,
                                  num_devices=max_devices,
                                  compat_mode=compat_mode,
                                  deterministic=deterministic,
-                                 # TODO add cache_dir to mgpu
                                  cache_dir=cache_dir)
             except ImportError:
                 logger.error("Multi-GPU support is a premium feature "

neon/backends/backend.py

@@ -431,7 +431,7 @@ class Backend(object):
                                        backends.
     """
     def __init__(self, rng_seed=None, default_dtype=np.float32,
-                 compat_mode=None, deterministic=False):
+                 compat_mode=None, deterministic=None):
         # dtype
         self.default_dtype = default_dtype
 
@@ -451,7 +451,10 @@ def __init__(self, rng_seed=None, default_dtype=np.float32,
         else:
             self.compat_mode = None
 
-        self.deterministic = deterministic
+        if deterministic is not None:
+            logger.warning('deterministic arg is deprecated in favor of specifying random seed')
+
+        self.deterministic = self.rng_seed is not None
 
     def output_dim(self, X, S, padding, strides, pooling=False):
         """

neon/backends/nervanagpu.py

@@ -657,7 +657,7 @@ def __init__(self,
                  rng_seed=None,
                  default_dtype=np.float32,
                  stochastic_round=False,
-                 deterministic=True,
+                 deterministic=None,
                  device_id=0,
                  bench=False,
                  scratch_size=0,
@@ -707,7 +707,6 @@ def __init__(self,
             stochastic_round = 0
 
         # attributes
-        self.deterministic = deterministic
         self.scratch_size = scratch_size
         self.scratch_offset = 0
         self.round_mode = stochastic_round

neon/util/argparser.py

@@ -312,6 +312,9 @@ def parse_args(self, gen_be=True):
         else:
             args.compat_mode = None
 
+        if args.deterministic:
+            logger.warn('--deterministic flag is deprecated.  Specify random seed for '
+                        'deterministic behavior.')
         # extended parsers may need to generate backend after argparsing
         if gen_be:
             # generate the backend
@@ -321,8 +324,7 @@ def parse_args(self, gen_be=True):
                         batch_size=args.batch_size,
                         datatype=args.datatype,
                         max_devices=args.max_devices,
-                        compat_mode=args.compat_mode,
-                        deterministic=args.deterministic)
+                        compat_mode=args.compat_mode)
 
         # display what command line / config options were set (and from where)
         logger.info(self.format_values())

setup.py

@@ -19,7 +19,7 @@
 import subprocess
 
 # Define version information
-VERSION = '1.2.2'
+VERSION = '1.3'
 FULLVERSION = VERSION
 write_version = True
 

tests/test_lstm.py

@@ -141,19 +141,19 @@ def check_lstm(seq_len, input_size, hidden_size,
     allclose_with_out(lstm.ifog_buffer.get(),
                       IFOGf_ref,
                       rtol=0.0,
-                      atol=1.0e-5)
+                      atol=1.5e-5)
 
     print '====Verifying cell states===='
     allclose_with_out(lstm.c_act_buffer.get(),
                       Ct_ref,
                       rtol=0.0,
-                      atol=1.0e-5)
+                      atol=1.5e-5)
 
     print '====Verifying hidden states===='
     allclose_with_out(lstm.outputs.get(),
                       Hout_ref,
                       rtol=0.0,
-                      atol=1.0e-5)
+                      atol=1.5e-5)
 
     print 'fprop is verified'
 
@@ -182,25 +182,25 @@ def check_lstm(seq_len, input_size, hidden_size,
     assert allclose_with_out(dWrecur_neon,
                              dWrecur_ref.T,
                              rtol=0.0,
-                             atol=1.0e-5)
+                             atol=1.5e-5)
 
     print '====Verifying update on W_input===='
     assert allclose_with_out(dWinput_neon,
                              dWinput_ref.T,
                              rtol=0.0,
-                             atol=1.0e-5)
+                             atol=1.5e-5)
 
     print '====Verifying update on bias===='
     assert allclose_with_out(db_neon.flatten(),
                              db_ref,
                              rtol=0.0,
-                             atol=1.0e-5)
+                             atol=1.5e-5)
 
     print '====Verifying output delta===='
     assert allclose_with_out(lstm.out_deltas_buffer.get(),
                              dX_ref,
                              rtol=0.0,
-                             atol=1.0e-5)
+                             atol=1.5e-5)
 
     print 'bprop is verified'
 

tests/test_mergesum_layer.py

@@ -80,7 +80,7 @@ def test_proj_upsample():
 
 
 def mergesum_test_config(modfunc, use_stride=1):
-    NervanaObject.be = gen_backend("gpu", batch_size=64)
+    NervanaObject.be = gen_backend("gpu", batch_size=64, rng_seed=0)
     be = NervanaObject.be
     l1 = Conv(**conv_params(3, 16))
     neon_layer = modfunc(16, use_stride)