MNIST Solver cannot check accuracy when y is onehot encoded

[kkweon]

Setup

• Ubuntu 16.04
• MXNet (ver 0.10.0)
• Minpy (ver 0.3.4)

What I was trying to do

• Run MNIST but with y_onehot
  • X.shape = (n, 784)
  • y.shape = (n, 10)
• Then I get the following message
  • It seems accuracy step cannot be performed due to wrong shape because it expects y.shape to be (n,) instead

(skipped)
(Iteration 91 / 1074) loss: 0.710763
(Iteration 101 / 1074) loss: 0.691108
[15:44:48] /home/travis/build/dmlc/mxnet-distro/mxnet-build/dmlc-core/include/dmlc/logging.h:304: [15:44:48] src/operator/tensor/./elemwise_binary_broadcast_op.h:48: Check failed: l == 1 || r == 1 operands could not be broadcast together with shapes (512,) (512,10)

Stack trace returned 10 entries:
[bt] (0) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x18b0dc) [0x7f1d2c2f30dc]
[bt] (1) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x4c2ce2) [0x7f1d2c62ace2]
[bt] (2) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0xb3dadd) [0x7f1d2cca5add]
[bt] (3) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(MXImperativeInvoke+0x4b8) [0x7f1d2cca8c78]
[bt] (4) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(ffi_call_unix64+0x4c) [0x7f1d50390550]
[bt] (5) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(ffi_call+0x1f5) [0x7f1d5038fcf5]
[bt] (6) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(_ctypes_callproc+0x3dc) [0x7f1d5038783c]
[bt] (7) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(+0x9da3) [0x7f1d5037fda3]
[bt] (8) /home/kkweon/anaconda3/bin/../lib/libpython3.6m.so.1.0(_PyObject_FastCallDict+0x9e) [0x7f1d51adfade]
[bt] (9) /home/kkweon/anaconda3/bin/../lib/libpython3.6m.so.1.0(+0x1482bb) [0x7f1d51bbc2bb]

Traceback (most recent call last):
  File "lab-10-1-mnist-softmax.py", line 64, in <module>
    solver.train()
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/minpy-0.3.4-py3.6.egg/minpy/nn/solver.py", line 275, in train
    self.train_dataiter, num_samples=self.train_acc_num_samples)
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/minpy-0.3.4-py3.6.egg/minpy/nn/solver.py", line 231, in check_accuracy
    acc_count += np.sum(np.argmax(predict, axis=1) == each_batch.label[0])
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/minpy-0.3.4-py3.6.egg/minpy/array.py", line 74, in __eq__
    return Value._ns.equal(self, other)
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/minpy-0.3.4-py3.6.egg/minpy/primitive.py", line 141, in __call__
    return self.call(args, kwargs)
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/minpy-0.3.4-py3.6.egg/minpy/primitive.py", line 238, in call
    result_value = self._func(*arg_values, **kwarg_values)
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/ndarray.py", line 1723, in equal
    None)
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/ndarray.py", line 1247, in _ufunc_helper
    return fn_array(lhs, rhs)
  File "<string>", line 14, in broadcast_equal
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/_ctypes/ndarray.py", line 72, in _imperative_invoke
    c_array(ctypes.c_char_p, [c_str(str(val)) for val in vals])))
  File "/home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/base.py", line 84, in check_call
    raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [15:44:48] src/operator/tensor/./elemwise_binary_broadcast_op.h:48: Check failed: l == 1 || r == 1 operands could not be broadcast together with shapes (512,) (512,10)

Stack trace returned 10 entries:
[bt] (0) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x18b0dc) [0x7f1d2c2f30dc]
[bt] (1) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x4c2ce2) [0x7f1d2c62ace2]
[bt] (2) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(+0xb3dadd) [0x7f1d2cca5add]
[bt] (3) /home/kkweon/anaconda3/lib/python3.6/site-packages/mxnet/libmxnet.so(MXImperativeInvoke+0x4b8) [0x7f1d2cca8c78]
[bt] (4) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(ffi_call_unix64+0x4c) [0x7f1d50390550]
[bt] (5) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(ffi_call+0x1f5) [0x7f1d5038fcf5]
[bt] (6) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(_ctypes_callproc+0x3dc) [0x7f1d5038783c]
[bt] (7) /home/kkweon/anaconda3/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(+0x9da3) [0x7f1d5037fda3]
[bt] (8) /home/kkweon/anaconda3/bin/../lib/libpython3.6m.so.1.0(_PyObject_FastCallDict+0x9e) [0x7f1d51adfade]
[bt] (9) /home/kkweon/anaconda3/bin/../lib/libpython3.6m.so.1.0(+0x1482bb) [0x7f1d51bbc2bb]

If I run without onehot_encoded

• It works fine but I get a bunch of deprecation warning messages regarding onehot_encode

[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
[15:56:36] src/ndarray/./ndarray_function-inl.h:68: The operator onehot_encode is deprecated; use one_hot instead.
(Epoch 10 / 10) train acc: 0.9033203125, val acc: 0.921875, time: 0.3320729732513428.

Full code

from minpy.nn import layers
from minpy.nn import model
from minpy.nn import solver
from minpy.nn import io
from tensorflow.examples.tutorials.mnist import input_data

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)

X_train = mnist.train.images
y_train = mnist.train.labels
print(X_train.shape, y_train.shape)  # (55000, 784) (55000, 10)
X_val = mnist.validation.images
y_val = mnist.validation.labels
print(X_val.shape, y_val.shape)  # (5000, 784) (5000, 10)
X_test = mnist.test.images
y_test = mnist.test.labels
print(X_test.shape, y_test.shape)  # (10000, 784) (10000, 10)

class SingleLayerNetwork(model.ModelBase):

    def __init__(self):
        super(SingleLayerNetwork, self).__init__()
        self.add_param("W1", shape=(784, 10))
        self.add_param("b1", shape=(10,))

    def forward(self, X, mode):
        net = X
        net = layers.affine(net, self.params['W1'], self.params['b1'])
        return net

    def loss(self, predict, y):
        return layers.softmax_loss(predict, y)


BATCH_SIZE = 512

train_iter = io.NDArrayIter(data=X_train,
                            label=y_train,
                            batch_size=BATCH_SIZE,
                            shuffle=True)
val_iter = io.NDArrayIter(data=X_val,
                          label=y_val,
                          batch_size=BATCH_SIZE)
test_iter = io.NDArrayIter(data=X_test,
                           label=y_test,
                           batch_size=BATCH_SIZE)


model = SingleLayerNetwork()

solver_config = {
    'model': model,
    'train_dataiter': train_iter,
    'test_dataiter': val_iter,
    'update_rule': 'adam',
    'num_epochs': 10,
    'init_rule': 'gaussian',
}
solver = solver.Solver(**solver_config)
solver.init()
solver.train()

[kkweon]

In code: solver.check_accuracy

it doesn't check each_batch.label is onehot encoded

acc_count += np.sum(np.argmax(predict, axis=1) == each_batch.label[0])

but minpy.nn.layers.softmax_loss can take either onehot(N, C) or labels (N,)