Fixed ICUDAUnit @Implementer and initialize(snapshot)

Change-Id: I7a639f14cf713390b6bb13ea79e2ef9c28db60af

activation.py

@@ -10,7 +10,7 @@
 import numpy
 from zope.interface import implementer
 
-from veles.accelerated_units import AcceleratedUnit, IOpenCLUnit
+from veles.accelerated_units import AcceleratedUnit, IOpenCLUnit, ICUDAUnit
 import veles.error as error
 from veles.memory import eq_addr, ravel
 from veles.znicz.nn_units import Forward, GradientDescentBase
@@ -22,7 +22,7 @@ def init_unpickled(self):
         self.sources_["activation"] = {}
 
 
-@implementer(IOpenCLUnit)
+@implementer(IOpenCLUnit, ICUDAUnit)
 class ActivationForward(Forward, Activation):
     MAPPING = set()
 
@@ -87,7 +87,7 @@ def cuda_run(self):
         self._gpu_run()
 
 
-@implementer(IOpenCLUnit)
+@implementer(IOpenCLUnit, ICUDAUnit)
 class ActivationBackward(GradientDescentBase, Activation):
     """Backward activation pass: err_input = err_output * F'(output).
 
@@ -178,7 +178,6 @@ def cuda_run(self):
         self._gpu_run()
 
 
-@implementer(IOpenCLUnit)
 class ForwardTanh(ActivationForward):
     """Forward pass for y = 1.7159 * tanh(0.6666 * x).
     """
@@ -192,7 +191,6 @@ def cpu_run(self):
         out *= 1.7159
 
 
-@implementer(IOpenCLUnit)
 class BackwardTanh(ActivationBackward):
     """Backward pass for :class:`ForwardTanh`.
     """
@@ -208,7 +206,6 @@ def cpu_run(self):
             err_input)
 
 
-@implementer(IOpenCLUnit)
 class ForwardSigmoid(ActivationForward):
     """Forward pass for y = 1.0 / (1.0 + exp(-x)).
     """
@@ -221,7 +218,6 @@ def cpu_run(self):
         numpy.reciprocal(1.0 + numpy.exp(-out), out)
 
 
-@implementer(IOpenCLUnit)
 class BackwardSigmoid(ActivationBackward):
     """Backward pass for :class:`ForwardSigmoid`.
     """
@@ -235,7 +231,6 @@ def cpu_run(self):
         numpy.multiply(err_output, output * (1.0 - output), err_input)
 
 
-@implementer(IOpenCLUnit)
 class ForwardMul(ActivationForward):
     """Forward pass for :math:`y = k x`.
     """
@@ -302,7 +297,6 @@ def cpu_run(self):
         out *= self.factor
 
 
-@implementer(IOpenCLUnit)
 class BackwardMul(ActivationBackward):
     """Backward pass for :class:`ForwardMul`.
     """
@@ -342,7 +336,6 @@ def cpu_run(self):
         err_input[:] = err_output[:] * self.factor
 
 
-@implementer(IOpenCLUnit)
 class ForwardRELU(ActivationForward):
     """
     This activation is taken from article
@@ -361,7 +354,6 @@ def cpu_run(self):
         out[:] = numpy.where(inp > 15, inp, numpy.log(numpy.exp(inp) + 1.0))
 
 
-@implementer(IOpenCLUnit)
 class BackwardRELU(ActivationBackward):
     """Backward pass for :class:`ForwardRELU`
     """
@@ -375,7 +367,6 @@ def cpu_run(self):
         numpy.multiply(err_output, 1.0 - numpy.exp(-output), err_input)
 
 
-@implementer(IOpenCLUnit)
 class ForwardStrictRELU(ActivationForward):
     """
     Forward pass for :math:`y = \\max(0, x)`.
@@ -405,7 +396,6 @@ def drop_slave(self, slave):
         pass
 
 
-@implementer(IOpenCLUnit)
 class BackwardStrictRELU(ActivationBackward):
     """
     Backward pass for :class:`ForwardStrictRELU`.
@@ -438,7 +428,6 @@ def drop_slave(self, slave):
         pass
 
 
-@implementer(IOpenCLUnit)
 class ForwardLog(ActivationForward):
     """Forward pass for :math:`y = \\log(x + \\sqrt{x^2 + 1})`.
     """
@@ -461,7 +450,6 @@ def cpu_run(self):
         numpy.log(inp + numpy.sqrt(numpy.square(inp) + 1), out)
 
 
-@implementer(IOpenCLUnit)
 class BackwardLog(ActivationBackward):
     """Backward pass for :class:`ForwardLog`.
     """
@@ -488,7 +476,6 @@ def cpu_run(self):
             err_input)
 
 
-@implementer(IOpenCLUnit)
 class ForwardTanhLog(ActivationForward):
     """Forward pass for hybrid tanh-log function.
     """
@@ -517,7 +504,6 @@ def cpu_run(self):
             out[i] = y
 
 
-@implementer(IOpenCLUnit)
 class BackwardTanhLog(ActivationBackward):
     """Backward pass for hybrid tanh-log function.
     """
@@ -554,7 +540,6 @@ def _set_activation_args(self):
         self.set_args(self.input, self.output, self.err_output, self.err_input)
 
 
-@implementer(IOpenCLUnit)
 class ForwardSinCos(ActivationForward):
     """Forward pass for y = sin(x) if idx(x) is odd else cos(x).
     """
@@ -575,7 +560,6 @@ def cpu_run(self):
         out[0::2] = numpy.cos(inp[0::2])
 
 
-@implementer(IOpenCLUnit)
 class BackwardSinCos(ActivationBackward):
     """Backward pass for :class:`ForwardSinCos`.
     """

all2all.py

@@ -14,13 +14,13 @@
 import numpy
 from zope.interface import implementer
 
-from veles.accelerated_units import IOpenCLUnit
+from veles.accelerated_units import IOpenCLUnit, ICUDAUnit
 import veles.error as error
 from veles.memory import reshape, roundup, Vector
 import veles.znicz.nn_units as nn_units
 
 
-@implementer(IOpenCLUnit)
+@implementer(IOpenCLUnit, ICUDAUnit)
 class All2All(nn_units.NNLayerBase):
     """All2All with linear activation f(x) = x.
 

evaluator.py

@@ -15,7 +15,7 @@
 from veles.distributable import TriviallyDistributable
 import veles.error as error
 from veles.memory import assert_addr, ravel, Vector
-from veles.accelerated_units import AcceleratedUnit, IOpenCLUnit
+from veles.accelerated_units import AcceleratedUnit, IOpenCLUnit, ICUDAUnit
 from veles.opencl_types import numpy_dtype_to_opencl
 
 
@@ -44,7 +44,7 @@ def initialize(self, device, **kwargs):
             vec.initialize(self.device)
 
 
-@implementer(IOpenCLUnit)
+@implementer(IOpenCLUnit, ICUDAUnit)
 class EvaluatorSoftmax(EvaluatorBase, TriviallyDistributable):
     """Evaluator for nn softmax output from the batch labels.
 

gd.py

@@ -23,7 +23,7 @@
 from zope.interface import implementer
 
 from veles.memory import reshape, roundup, Vector
-from veles.accelerated_units import IOpenCLUnit
+from veles.accelerated_units import IOpenCLUnit, ICUDAUnit
 import veles.znicz.nn_units as nn_units
 from collections import namedtuple
 
@@ -41,7 +41,7 @@
                                                    "bias"))
 
 
-@implementer(IOpenCLUnit)
+@implementer(IOpenCLUnit, ICUDAUnit)
 class GradientDescent(nn_units.GradientDescentBase):
     """Gradient Descent unit for :class:`veles.znicz.all2all.All2All`.
 

samples/MnistSimple/mnist.py

@@ -194,7 +194,7 @@ def __init__(self, workflow, **kwargs):
     def initialize(self, learning_rate, weights_decay, device, **kwargs):
         return super(MnistWorkflow, self).initialize(
             learning_rate=learning_rate, weights_decay=weights_decay,
-            device=device)
+            device=device, **kwargs)
 
 
 def run(load, main):

tests/functional/test_approximator.py

@@ -66,7 +66,8 @@ def test_approximator(self):
             device=self.device,
             learning_rate=root.approximator.learning_rate,
             weights_decay=root.approximator.weights_decay,
-            minibatch_size=root.approximator.loader.minibatch_size)
+            minibatch_size=root.approximator.loader.minibatch_size,
+            snapshot=False)
         self.w.run()
         file_name = self.w.snapshotter.file_name
 
@@ -83,7 +84,8 @@ def test_approximator(self):
             device=self.device,
             learning_rate=root.approximator.learning_rate,
             weights_decay=root.approximator.weights_decay,
-            minibatch_size=root.approximator.loader.minibatch_size)
+            minibatch_size=root.approximator.loader.minibatch_size,
+            snapshot=True)
         self.wf.run()
 
         avg_mse = self.wf.decision.epoch_metrics[2][0]

tests/functional/test_channels.py

@@ -88,7 +88,8 @@ def test_channels_all2all(self):
         self.assertEqual(self.w.evaluator.labels,
                          self.w.loader.minibatch_labels)
         self.w.initialize(device=self.device,
-                          minibatch_size=root.channels.loader.minibatch_size)
+                          minibatch_size=root.channels.loader.minibatch_size,
+                          snapshot=False)
         self.assertEqual(self.w.evaluator.labels,
                          self.w.loader.minibatch_labels)
         self.w.run()
@@ -107,7 +108,8 @@ def test_channels_all2all(self):
         self.assertEqual(self.wf.evaluator.labels,
                          self.wf.loader.minibatch_labels)
         self.wf.initialize(device=self.device,
-                           minibatch_size=root.channels.loader.minibatch_size)
+                           minibatch_size=root.channels.loader.minibatch_size,
+                           snapshot=True)
         self.assertEqual(self.wf.evaluator.labels,
                          self.wf.loader.minibatch_labels)
         self.wf.run()

tests/functional/test_cifar_all2all.py

@@ -84,7 +84,8 @@ def test_cifar_all2all(self):
         self.assertEqual(self.w.evaluator.labels,
                          self.w.loader.minibatch_labels)
         self.w.initialize(device=self.device,
-                          minibatch_size=root.cifar.loader.minibatch_size)
+                          minibatch_size=root.cifar.loader.minibatch_size,
+                          snapshot=False)
         self.assertEqual(self.w.evaluator.labels,
                          self.w.loader.minibatch_labels)
         self.w.run()
@@ -102,7 +103,8 @@ def test_cifar_all2all(self):
         self.assertEqual(self.wf.evaluator.labels,
                          self.wf.loader.minibatch_labels)
         self.wf.initialize(device=self.device,
-                           minibatch_size=root.cifar.loader.minibatch_size)
+                           minibatch_size=root.cifar.loader.minibatch_size,
+                           snapshot=True)
         self.assertEqual(self.wf.evaluator.labels,
                          self.wf.loader.minibatch_labels)
         self.wf.run()

tests/research/Approximator/approximator.py

@@ -304,7 +304,7 @@ def initialize(self, learning_rate, weights_decay, minibatch_size,
                    device, **kwargs):
         super(ApproximatorWorkflow, self).initialize(
             learning_rate=learning_rate, weights_decay=weights_decay,
-            minibatch_size=minibatch_size, device=device)
+            minibatch_size=minibatch_size, device=device, **kwargs)
 
 
 def run(load, main):

tests/research/VideoAE/video_ae.py

@@ -219,7 +219,7 @@ def initialize(self, learning_rate, weights_decay, device, **kwargs):
             forward.device = device
         return super(VideoAEWorkflow, self).initialize(
             learning_rate=learning_rate, weights_decay=weights_decay,
-            device=device)
+            device=device, **kwargs)
 
 
 def run(load, main):

tests/research/WineRelu/wine_relu.py

@@ -132,7 +132,7 @@ def __init__(self, workflow, **kwargs):
     def initialize(self, learning_rate, weights_decay, device, **kwargs):
         super(WineReluWorkflow, self).initialize(
             learning_rate=learning_rate, weights_decay=weights_decay,
-            device=device)
+            device=device, **kwargs)
 
 
 def run(load, main):

tests/unit/test_caffe_complex.py

@@ -471,7 +471,7 @@ def test_all(self):
         self._create_gd_units(cur_iter)
         self.workflow.end_point.link_from(self.workflow["gd_conv1"])
 
-        self.workflow.initialize(device=self.device)
+        self.workflow.initialize(device=self.device, snapshot=False)
         self.info("self.workflow.initialize() completed")
 
         self._load_labels_and_data(cur_iter)

tests/unit/test_matrix_multiplication.py

@@ -21,7 +21,7 @@
 import veles.prng as prng
 from veles import opencl_types
 from veles.dummy import DummyWorkflow
-from veles.accelerated_units import TrivialOpenCLUnit
+from veles.accelerated_units import TrivialAcceleratedUnit
 import veles.znicz as znicz
 znicz.nothing()
 
@@ -92,7 +92,7 @@ def _cleanup_after_tsts(self):
     def _do_tst(self, device, BLOCK_SIZE):
         """Do test for specific context
         """
-        obj = TrivialOpenCLUnit(DummyWorkflow())
+        obj = TrivialAcceleratedUnit(DummyWorkflow())
         obj.initialize(device=device)
 
         self.a.initialize(device)

tests/unit/test_matrix_reduce.py

@@ -19,7 +19,7 @@
 import veles.opencl_types as opencl_types
 import veles.prng as prng
 from veles.dummy import DummyWorkflow
-from veles.accelerated_units import TrivialOpenCLUnit
+from veles.accelerated_units import TrivialAcceleratedUnit
 
 
 class TestMatrixReduce(unittest.TestCase):
@@ -42,7 +42,7 @@ def _build_program(self, a, b, A_WIDTH, A_HEIGHT, A_COL, REDUCE_SIZE):
         if A_COL:
             defines["A_COL"] = 1
 
-        tmp = TrivialOpenCLUnit(DummyWorkflow())
+        tmp = TrivialAcceleratedUnit(DummyWorkflow())
         tmp.initialize(device=self.device)
         tmp.sources_["test_matrix_reduce"] = {}
 

tests/unit/test_matrix_transpose.py

@@ -16,7 +16,7 @@
 from veles.backends import Device
 import veles.prng as prng
 from veles.dummy import DummyWorkflow
-from veles.accelerated_units import TrivialOpenCLUnit
+from veles.accelerated_units import TrivialAcceleratedUnit
 
 
 class TestMatrixTranspose(unittest.TestCase):
@@ -33,7 +33,7 @@ def test_transpose(self):
         prng.get().fill(a.mem)
         b = Vector(numpy.zeros([WIDTH * 2, HEIGHT], dtype=dtype))
 
-        obj = TrivialOpenCLUnit(DummyWorkflow())
+        obj = TrivialAcceleratedUnit(DummyWorkflow())
         obj.initialize(device=device)
 
         a.initialize(device)