def check_fn(weights):
return layers.softmax_loss(f(x=x, fc_weight=weights), fake_y)
def loss(self, predict, y):
return layers.softmax_loss(predict, y)
def check_gamma(g):
y, _, _ = layers.batchnorm(x, g, beta)
return layers.softmax_loss(y, fake_y)
def check_beta(b):
y, _, _ = layers.batchnorm(x, gamma, b)
return layers.softmax_loss(y, fake_y)
def check_fn(w):
return layers.softmax_loss(layers.affine(x, w, b), fake_y)
def check_fn(x):
return layers.softmax_loss(layers.relu(x), fake_y)
def check_fn(weights):
return layers.softmax_loss(f(x=x, fc_weight=weights), fake_y)
def loss(self, predict, y):
return layers.softmax_loss(predict, y)
def check_beta(b):
y, _, _ = layers.batchnorm(x, gamma, b)
return layers.softmax_loss(y, fake_y)
def check_fn(x):
return layers.softmax_loss(x, lbl)
def check_gamma(g):
y, _, _ = layers.batchnorm(x, g, beta)
return layers.softmax_loss(y, fake_y)
def check_fn(x):
return layers.softmax_loss(layers.relu(x), fake_y)
def check_fn(w):
return layers.softmax_loss(layers.affine(x, w, b), fake_y)
def check_fn(x):
return layers.softmax_loss(x, lbl)
def loss(self, prediction, Y):
return layers.softmax_loss(prediction, Y)
def loss(self, predict, y):
# Compute softmax loss between the output and the label.
return layers.softmax_loss(predict, y)
network.extend(convolution_module(num_filter=64))
network.extend([
bind(lambda data : Pooling(data=data, pool_type='avg', kernel=(8, 8), stride=(1, 1), pad=(0, 0))),
bind(lambda data : Flatten(data=data)),
bind(lambda data : FullyConnected(data=data, num_hidden=10)),
])
data = Variable('data', shape=(128, 10))
labels = Variable('labels', shape=(128,))
loss_symbol = SoftmaxOutput(data=data, label=labels)
loss = loss_symbol.simple_bind(context)
import numpy as np
X = mx.nd.random_normal(0, 1, (50000, 3, 32, 32), context)
Y = mx.nd.array(np.random.choice(np.arange(10), (50000,)), context)
iterator = mx.io.NDArrayIter(data=X, label=Y, batch_size=128)
while True:
iterator.reset()
for i, batch in enumerate(iterator):
data, labels = batch.data[0], batch.label[0]
scores = forward(network, data)
results = loss.forward(data=scores, labels=labels, is_train=True)
softmax_loss(results[0].asnumpy(), labels.asnumpy())
loss.backward()
backward(network, loss.grad_dict['data'])
if (i + 1) % 100 == 0:
print i + 1