def _setup(self, data):
image = data['image']
label = data['label']
t4.Conv.default.filter_initializer = t4.GlorotNormalInitializer
conv = t4.DeepConv(1, 64, 1024)
lin = t4.Linear(7 * 7 * 1024, 10)
dense = t4.image.to_dense(image)
flat = t4.image.to_flat(conv(dense, t4.relu))
pred = tf.nn.softmax(lin(flat))
loss = t4.categorical_cross_entropy_loss(
pred, tf.one_hot(label, 10, dtype=t4.Widget.default.float_dtype))
acc = t4.correct_prediction(tf.argmax(pred, axis=1), label)
optimizer = tf.train.AdamOptimizer().minimize(loss)
self.train = self._add_slot('train',
outputs=(loss, acc),
updates=optimizer,
summaries=lin.summaries)
self.valid = self._add_slot(
'valid',
outputs=(loss, acc),
# givens={t4.BatchNorm.default.shift: False}
)
def _setup(self, data):
image = data['image']
label = data['label']
dense = t4.image.to_dense(image)
flat = t4.image.to_flat(dense)
lin = t4.Linear(28 * 28, 10)
pred = tf.nn.softmax(lin(flat))
loss = t4.categorical_cross_entropy_loss(pred, tf.one_hot(label, 10, dtype=t4.Widget.default.float_dtype))
acc = t4.correct_prediction(tf.argmax(pred, axis=1), label)
optimizer = tf.train.AdamOptimizer().minimize(loss)
self.train = self._add_slot(
'train',
outputs=(loss, acc),
updates=optimizer,
summaries=lin.summaries
)
self.valid = self._add_slot(
'valid',
outputs=(loss, acc)
)
def _setup(self, data):
image = data['image']
label = data['label'] - 1
k = 2
n = 4
t4.L2Regularizer.default.rate = 0.0005
t4.SimpleResidualConv.default.block = t4.ConvNoBias
conv = t4.Conv(3, 16, 3, 3, 1, 1)
residual = t4.DeepResidualConv((16 * k, n), (32 * k, n), (64 * k, n),
(128 * k, n),
input_channel=16,
block=t4.SimpleResidualConv)
lin = t4.Linear(128 * k, 1000)
shift = tf.placeholder_with_default(True, ())
image = t4.image.randomize_flip(image, 0.5, switch=shift)
image = t4.image.randomize_crop(image, switch=shift)
dense = t4.image.to_dense(image, std=64)
dense = conv(dense)
dense = t4.batch_normalization(dense)
dense = t4.relu(dense)
dense = residual(dense, t4.batch_normalization, t4.relu)
flat = t4.flat_pool(dense)
pred = tf.nn.softmax(t4.batch_normalization(lin(flat)))
loss = t4.categorical_cross_entropy_loss(pred,
t4.OneHot(1000)(label),
with_false=False)
acc = t4.correct_prediction(t4.argmax(pred), label)
loss_ma = t4.MovingAverage()(loss)
acc_ma = t4.MovingAverage()(acc)
self.lr = lr = tf.placeholder(t4.Widget.default.float_dtype)
regularizers = tf.reduce_sum(self.losses)
optimizer = tf.train.MomentumOptimizer(lr, 0.9).minimize(
loss + regularizers, colocate_gradients_with_ops=True)
self.train = self._add_slot('train',
outputs=(loss_ma, acc_ma, regularizers),
updates=optimizer)
self.valid = self._add_slot('valid',
outputs=(loss, acc),
givens={
t4.BatchNorm.default.shift: False,
shift: False
})
def _setup(self, data):
image = data['image']
label = data['label']
t4.ResidualConv.default.block = t4.ConvNoBias
conv = t4.ConvNoBias(3, 64, 7, 7, 2, 2)
residual = t4.DeepResidualConv(
(256, 3), (512, 4), (1024, 6), (2048, 3),
input_channel=64, block=t4.BottleNeckResidualConvOS
)
lin = t4.Linear(2048, 90)
# image = tf.image.resize_images(image, [224, 224])
mean = tf.Variable(tf.constant([128, 128, 128], dtype=self.default_float_dtype), trainable=False)
dense = t4.image.to_dense(image, mean=mean, std=1)
dense = conv(dense)
dense = t4.batch_normalization(dense)
dense = t4.relu(dense)
dense = t4.max_pool(dense, 3, 3, 2, 2)
dense = residual(dense, t4.batch_normalization, t4.relu)
flat = t4.flat_pool(dense)
pred = tf.nn.softmax(lin(flat))
loss = t4.categorical_cross_entropy_loss(pred, t4.OneHot(90)(label))
acc = t4.correct_prediction(t4.argmax(pred), label)
with tf.device('/gpu:1'):
optimizer = tf.train.GradientDescentOptimizer(0.01).minimize(loss)
self.train = self._add_slot(
'train',
outputs=(loss, acc),
updates=optimizer
)
self.valid = self._add_slot(
'valid',
outputs=(loss, acc),
givens={
t4.BatchNorm.default.shift: False
}
)
#!/usr/bin/env python3
# -- coding: utf8 --
# :author: nvagus
# :time: 11/8/17-5:19 PM
# :package: tforce.test
import code
import tensorflow as tf
import tforce as t4
if __name__ == '__main__':
with tf.Session() as sess:
x = tf.constant([1., 2., 3.], dtype=t4.Widget.default.float_dtype)
lin = t4.Linear(3, 5)
y = lin(x)
sess.run(tf.global_variables_initializer())
print(sess.run(y))
print(
lin.global_variables,
lin.trainable_variables,
lin.weight_variables,
lin.bias_variables,
lin.losses,
lin.summaries,
sep='\n'
)
code.interact(local=locals())