test failure on windows

[Bartzi]

The test test_train_main fails on Windows:

    def test_train():
        np.seterr(divide='raise')

        for optimizer in [SGD(0.001), Adam(0.001)]:
            for data_set in [MNIST(64), FashionMNIST(64)]:
                for i in range(10):
                    model = MLP()

                    for iteration, batch in enumerate(data_set.train):
                        model.forward(batch)
>                       model.backward(optimizer)

length\tests\test_train_main.py:18:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
length\models\mlp.py:32: in backward
    self.loss.backward(optimizer)
length\graph.py:45: in backward
    gradients = candidate_layer.creator.backward(candidate_layer.grad)
length\layer.py:28: in backward
    parameter_deltas = self.optimizer.run_update_rule(parameter_gradients, self)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <length.optimizers.adam.Adam object at 0x07C85370>
gradients = (array([[0., 0., 0., ..., 0., 0., 0.],
       [0., 0., 0., ..., 0., 0., 0.],
       [0., 0., 0., ..., 0., 0., 0.],
   ...-02, -4.15754272e-03,
       -3.68550362e-04, -4.95784450e-03, -7.36804074e-03, -5.72426897e-03],
      dtype=float32))
layer = <length.layers.fully_connected.FullyConnected object at 0x07C855B0>

    def run_update_rule(self, gradients, layer):
        self.current_id = id(layer)

        if not self.initialized:
            self.initialize(gradients)

        self.t_values[self.current_id] += 1
        t = self.t_values[self.current_id]

        param_deltas = []

        for i, gradient in enumerate(gradients):
            m = self.m_values[self.current_id][i]
            v = self.v_values[self.current_id][i]

>           m += (1 - self.beta1) * (gradient - m)
E           ValueError: operands could not be broadcast together with shapes (512,784) (10,512)

length\optimizers\adam.py:47: ValueError

It does not fail on linux 😕