Beispiel #1
0
def test_record_mode_bad():
    """
    Like test_record_bad, but some events are recorded by the
    theano RecordMode, as is the event that triggers the mismatch
    error.
    """

    # Record a sequence of events
    output = StringIO()

    recorder = Record(file_object=output, replay=False)

    record_mode = RecordMode(recorder)

    i = iscalar()
    f = function([i], i, mode=record_mode, name='f')

    num_lines = 10

    for i in range(num_lines):
        recorder.handle_line(str(i) + '\n')
        f(i)

    # Make sure that the playback functionality doesn't raise any errors
    # when we repeat them
    output_value = output.getvalue()
    output = StringIO(output_value)

    playback_checker = Record(file_object=output, replay=True)

    playback_mode = RecordMode(playback_checker)

    i = iscalar()
    f = function([i], i, mode=playback_mode, name='f')

    for i in range(num_lines // 2):
        playback_checker.handle_line(str(i) + '\n')
        f(i)

    # Make sure a wrong event causes a MismatchError
    try:
        f(0)
    except MismatchError:
        return
    raise AssertionError("Failed to detect a mismatch.")
    def run(replay, log=None):

        if not replay:
            log = StringIO()
        else:
            log = StringIO(log)
        record = Record(replay=replay, file_object=log)

        disturb_mem.disturb_mem()

        mode = RecordMode(record=record)

        b = sharedX(np.zeros((2, )), name='b')
        channels = OrderedDict()

        disturb_mem.disturb_mem()

        v_max = b.max(axis=0)
        v_min = b.min(axis=0)
        v_range = v_max - v_min

        updates = []
        for i, val in enumerate([
                v_max.max(),
                v_max.min(),
                v_range.max(),
        ]):
            disturb_mem.disturb_mem()
            s = sharedX(0., name='s_' + str(i))
            updates.append((s, val))

        for var in theano.gof.graph.ancestors(update for _, update in updates):
            if var.name is not None and var.name is not 'b':
                if var.name[0] != 's' or len(var.name) != 2:
                    var.name = None

        for key in channels:
            updates.append((s, channels[key]))
        f = theano.function([],
                            mode=mode,
                            updates=updates,
                            on_unused_input='ignore',
                            name='f')
        for output in f.maker.fgraph.outputs:
            mode.record.handle_line(var_descriptor(output) + '\n')
        disturb_mem.disturb_mem()
        f()

        mode.record.f.flush()

        if not replay:
            return log.getvalue()
Beispiel #3
0
def test_record_mode_good():
    """
    Like test_record_good, but some events are recorded by the
    theano RecordMode. We don't attempt to check the
    exact string value of the record in this case.
    """

    # Record a sequence of events
    output = StringIO()

    recorder = Record(file_object=output, replay=False)

    record_mode = RecordMode(recorder)

    i = iscalar()
    f = function([i], i, mode=record_mode, name='f')

    num_lines = 10

    for i in range(num_lines):
        recorder.handle_line(str(i) + '\n')
        f(i)

    # Make sure that the playback functionality doesn't raise any errors
    # when we repeat them
    output_value = output.getvalue()
    output = StringIO(output_value)

    playback_checker = Record(file_object=output, replay=True)

    playback_mode = RecordMode(playback_checker)

    i = iscalar()
    f = function([i], i, mode=playback_mode, name='f')

    for i in range(num_lines):
        playback_checker.handle_line(str(i) + '\n')
        f(i)
Beispiel #4
0
def test_determinism_2():
    """
    A more aggressive determinism test. Tests that apply nodes are all passed
    inputs with the same md5sums, apply nodes are run in same order, etc.  Uses
    disturb_mem to try to cause dictionaries to iterate in different orders,
    etc.
    """
    def run_sgd(mode):
        # Must be seeded the same both times run_sgd is called
        disturb_mem.disturb_mem()
        rng = np.random.RandomState([2012, 11, 27])

        batch_size = 5
        train_batches = 3
        valid_batches = 4
        num_features = 2

        # Synthesize dataset with a linear decision boundary
        w = rng.randn(num_features)

        def make_dataset(num_batches):
            disturb_mem.disturb_mem()
            m = num_batches * batch_size
            X = rng.randn(m, num_features)
            y = np.zeros((m, 1))
            y[:, 0] = np.dot(X, w) > 0.

            rval = DenseDesignMatrix(X=X, y=y)

            rval.yaml_src = ""  # suppress no yaml_src warning

            X = rval.get_batch_design(batch_size)
            assert X.shape == (batch_size, num_features)

            return rval

        train = make_dataset(train_batches)
        valid = make_dataset(valid_batches)

        num_chunks = 10
        chunk_width = 2

        class ManyParamsModel(Model):
            """
            Make a model with lots of parameters, so that there are many
            opportunities for their updates to get accidentally re-ordered
            non-deterministically. This makes non-determinism bugs manifest
            more frequently.
            """
            def __init__(self):
                super(ManyParamsModel, self).__init__()
                self.W1 = [
                    sharedX(rng.randn(num_features, chunk_width))
                    for i in xrange(num_chunks)
                ]
                disturb_mem.disturb_mem()
                self.W2 = [
                    sharedX(rng.randn(chunk_width)) for i in xrange(num_chunks)
                ]
                self._params = safe_union(self.W1, self.W2)
                self.input_space = VectorSpace(num_features)
                self.output_space = VectorSpace(1)

        disturb_mem.disturb_mem()
        model = ManyParamsModel()
        disturb_mem.disturb_mem()

        class LotsOfSummingCost(Cost):
            """
            Make a cost whose gradient on the parameters involves summing many
            terms together, so that T.grad is more likely to sum things in a
            random order.
            """

            supervised = True

            def expr(self, model, data, **kwargs):
                self.get_data_specs(model)[0].validate(data)
                X, Y = data
                disturb_mem.disturb_mem()

                def mlp_pred(non_linearity):
                    Z = [T.dot(X, W) for W in model.W1]
                    H = map(non_linearity, Z)
                    Z = [T.dot(h, W) for h, W in safe_izip(H, model.W2)]
                    pred = sum(Z)
                    return pred

                nonlinearity_predictions = map(
                    mlp_pred, [T.nnet.sigmoid, T.nnet.softplus, T.sqr, T.sin])
                pred = sum(nonlinearity_predictions)
                disturb_mem.disturb_mem()

                return abs(pred - Y[:, 0]).sum()

            def get_data_specs(self, model):
                data = CompositeSpace(
                    (model.get_input_space(), model.get_output_space()))
                source = (model.get_input_source(), model.get_target_source())
                return (data, source)

        cost = LotsOfSummingCost()

        disturb_mem.disturb_mem()

        algorithm = SGD(
            cost=cost,
            batch_size=batch_size,
            learning_rule=Momentum(.5),
            learning_rate=1e-3,
            monitoring_dataset={
                'train': train,
                'valid': valid
            },
            update_callbacks=[ExponentialDecay(decay_factor=2., min_lr=.0001)],
            termination_criterion=EpochCounter(max_epochs=5))

        disturb_mem.disturb_mem()

        train_object = Train(dataset=train,
                             model=model,
                             algorithm=algorithm,
                             extensions=[
                                 PolyakAveraging(start=0),
                                 MomentumAdjustor(final_momentum=.9,
                                                  start=1,
                                                  saturate=5),
                             ],
                             save_freq=0)

        disturb_mem.disturb_mem()

        train_object.main_loop()

    output = cStringIO.StringIO()
    record = Record(file_object=output, replay=False)
    record_mode = RecordMode(record)

    run_sgd(record_mode)

    output = cStringIO.StringIO(output.getvalue())
    playback = Record(file_object=output, replay=True)
    playback_mode = RecordMode(playback)

    run_sgd(playback_mode)