def argsort(input, dim, descending):
idx = nd.argsort(input, dim, is_ascend=not descending)
idx = nd.cast(idx, dtype='int64')
return idx
def sort_1d(input):
# TODO: this isn't an ideal implementation.
val = nd.sort(input, axis=None, is_ascend=True)
idx = nd.argsort(input, is_ascend=True)
idx = nd.cast(idx, dtype='int64')
return val, idx
def astype(input, ty):
return nd.cast(input, ty)
def merge(conv_w, gamma, beta, running_mean, running_var):
gamma_over_var = gamma / nd.sqrt(running_var + 1e-5)
gamma_over_var_expanded = nd.reshape(gamma_over_var, (gamma_over_var.shape[0], 1, 1, 1))
new_w = gamma_over_var_expanded * nd.cast(conv_w, 'float32')
new_b = beta - running_mean * gamma_over_var
return new_w, new_b
def acc(output, label):
# output: (batch, num_output) float32 ndarray
# label: (batch, ) int32 ndarray
return (cast(output.argmax(axis=1),
dtype="int64") == label).mean().asscalar()
def mxnet_cast(in1, in2):
return [mxnd.cast(t, dtype='float32')
for t in in1], [mxnd.cast(t, dtype='int64') for t in in2]
def __normalize_rgb_image(arr):
return nd.cast(arr, "float32") / 255.0