300 * B: 0.0001
})
for i, l in enumerate(losses):
optimizers.append(
Adam(l, learning_rate, params=dnn[i * 2 + 1:(i + 1) * 2 + 1].params))
minimizer = tf.group(
sum([opt.updates for opt in optimizers], []) + dnn.updates)
# Workers
#---------
min1 = sknet.Worker(op_name='minimizer',
context='train_set',
op=minimizer,
instruction='execute every batch',
deterministic=False)
loss1 = sknet.Worker(op_name='loss',
context='train_set',
op=tf.stack(losses, -1),
instruction='save & print every 100 batch',
deterministic=False)
op_accu = tf.reshape(tf.stack(accus), (1, 6))
accus1 = sknet.Worker(
op_name='prediction',
context='test_set',
op=op_accu,
lr = sknet.schedules.PiecewiseConstant(0.002, {
100 * B: 0.002,
200 * B: 0.001,
250 * B: 0.0005
})
optimizer = sknet.optimizers.Adam(loss,
lr,
params=dnn.variables(trainable=True))
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
min1 = sknet.Worker(name='minimizer',
context='train_set',
op=[minimizer, loss],
deterministic=False,
period=[1, 1],
verbose=[0, 3])
accu1 = sknet.Worker(name='accu',
context='test_set',
op=accu,
deterministic=True,
transform_function=np.mean,
verbose=1)
queue = sknet.Queue((min1, accu1))
# Pipeline
#---------
workplace = sknet.utils.Workplace(dnn, dataset=dataset)
loss = sknet.losses.crossentropy_logits(p=dataset.labels, q=dnn[-1])
accuracy = sknet.losses.StreamingAccuracy(dataset.labels, dnn[-1])
auc = sknet.losses.StreamingAUC(dataset.labels, tf.nn.softmax(dnn[-1])[:, 1])
# we aim at minimizing the loss, so we create the optimizer (Adam in this case)
# with a stepwise learning rate, the minimizer is the operation that applies
# the changes to the model parameters, we also specify that this process
# should also include some possible network dependencies present in UPDATE_OPS
optimizer = sknet.optimizers.Adam(loss, dnn.variables(trainable=True), 0.001)
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
work1 = sknet.Worker(name='minimizer',
context='train_set',
op=[minimizer, loss, accuracy, auc],
deterministic=False,
period=[1, 100, 1, 1],
verbose=[0, 2, 1, 1])
work2 = sknet.Worker(name='AUC',
context='test_set',
op=[accuracy, auc],
deterministic=True,
verbose=[1, 1],
period=[1, 1])
queue = sknet.Queue((work1, work2))
# Pipeline
#---------
B = dataset.N('train_set') // 64
lr = sknet.schedules.PiecewiseConstant(0.001, {
100 * B: 0.0005,
200 * B: 0.0001,
250 * B: 0.00005
})
optimizer = sknet.optimizers.Adam(loss, dnn.variables(trainable=True), lr)
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
min_worker = sknet.Worker(name='minimizer',
context='train_set',
op=[minimizer, loss] + mus + [dataset.images],
deterministic=False,
period=[1, 100] + [5000] * (len(mus) + 1))
accu_worker = sknet.Worker(name='accu',
context='test_set',
op=[accu] + mus + [dataset.images],
deterministic=True,
period=[1] + [5] * (len(mus) + 1))
PATH = '/mnt/drive1/rbalSpace/centroids/'
queue = sknet.Queue((min_worker, accu_worker),
filename=PATH + 'saved_mus_' + MODEL + '.h5')
# Pipeline
#---------
reconstruction = dnn[-1]
# Loss and Optimizer
#-------------------
# Compute some quantities that we want to keep track and/or act upon
loss = MSE(dataset.signals, dnn[-1])
optimizer = sknet.optimize.Adam(loss, 0.00001, params=dnn.params)
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
work1 = sknet.Worker(op_name='minimizer',
context='train_set',
op=minimizer,
instruction='execute every batch',
deterministic=False)
work2 = sknet.Worker(op_name='loss',
context='train_set',
op=loss,
instruction='save & print every 300 batch',
deterministic=False)
work3 = work2.alter(context='test_set',
deterministic=True,
instruction='save every batch & print average')
work4 = sknet.Worker(op_name='reconstruction',
context='test_set',
B = dataset.N('train_set') // 32
lr = sknet.optimize.PiecewiseConstant(0.005, {
100 * B: 0.003,
200 * B: 0.001,
250 * B: 0.0005
})
optimizer = Adam(loss, lr, params=dnn.variables(trainable=True))
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
minimize = sknet.Worker(name='loss',
context='train_set',
op=[minimizer, loss, hessian, dataset.labels],
deterministic=False,
period=[1, 100, 100, 100])
accu = sknet.Worker(name='accu',
context='test_set',
op=[accu, hessian, dataset.labels],
deterministic=True,
transform_function=[np.mean, None, None])
queue = sknet.Queue((minimize, accu),filename='cifar10_'\
+str(data_augmentation)+'_'+str(D)+'_'+str(W)+'.h5')
# Pipeline
#---------
workplace = sknet.utils.Workplace(dnn, dataset=dataset)
test_accu = sknet.losses.StreamingAccuracy(dataset.labels, prediction[:16])
B = dataset.N('train_set') // 16
lr = sknet.schedules.PiecewiseConstant(0.01, {75 * B: 0.005, 125 * B: 0.001})
optimizer = sknet.optimizers.Adam(train_classif + train_recons * PARAMETER,
dnn.variables(trainable=True), lr)
minimizer = tf.group(optimizer.updates + dnn.updates)
reset_op = tf.group(dnn.reset_variables_op, optimizer.reset_variables_op)
# Workers
# ---------
train_w = sknet.Worker(
name='minimizer',
context='train_set',
op=[minimizer, train_recons, train_classif, train_uaccu],
deterministic=False,
period=[1, 100, 100, 1],
verbose=[0, 2, 2, 1])
valid_w = sknet.Worker(name='accu',
context='valid_set',
op=[test_accu, test_recons],
deterministic=True,
verbose=1)
test_w = sknet.Worker(name='accu',
context='test_set',
op=[test_accu, test_recons],
deterministic=True,
verbose=1)
W1[:, 1] -= W1[:, 0] * (W1[:, 1] * W1[:, 0]).sum() / (W1[:, 0]**2).sum()
W1[:, 1] *= 2
W1 = np.repeat(W1, 3, 1)
dnn.append(ops.Dense(dataset.input, 6, W=W1, b=b1[[0, 5, 1, 2, 4, 3]]))
output = dnn[0]
else:
# RANK 2 ARBITRARY
dnn = sknet.network.Network(name='simple_model')
np.random.seed(111)
W1 = (np.random.randn(2, 6) / 4).astype('float32')
dnn.append(ops.Dense(dataset.input, 6, W=W1, b=b1))
output = dnn[0]
output = sknet.Worker(op_name='poly',
context='train_set',
op=output,
instruction='save every batch',
deterministic=False)
# Pipeline
#---------
workplace = sknet.utils.Workplace(dnn, dataset=dataset)
workplace.execute_worker(output)
fig = plt.figure(figsize=(6, 9))
#
mask = output.data[0] > 0
boundary = sknet.utils.geometry.get_input_space_partition(mask, N, N,
2).astype('bool')
boundary = boundary.astype('float32')
# we aim at minimizing the loss, so we create the optimizer (Adam in this case)
# with a stepwise learning rate, the minimizer is the operation that applies
# the changes to the model parameters, we also specify that this process
# should also include some possible network dependencies present in UPDATE_OPS
optimizer = sknet.optimize.Adam(loss, 0.01, params=dnn.params)
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
work1 = sknet.Worker(op_name='minimizer',
context='train_set',
op=minimizer,
instruction='execute every batch',
deterministic=False,
sampling='random',
description='minimize')
work2 = sknet.Worker(op_name='loss',
context='train_set',
op=loss,
instruction='save & print every 30 batch',
deterministic=False,
description='saving the loss every 30 batches',
sampling='random')
work3 = sknet.Worker(
op_name='accuracy',
context='train_set',
# last layer going to 1d for binary classification
dnn.append(
ops.Dense(dnn[-1],
1,
W=1 * np.random.randn(4, 1).astype('float32'),
b=np.array([0.5]).astype('float32')))
# extracts the outputs prior nonlinearity
outputs = dnn[0::2].as_list()
# Workers
#---------
outputs = tf.concat(outputs, -1)
output = sknet.Worker(outputs=outputs,
context='train_set',
feed_dict=dnn.deter_dict(False))
# Pipeline
#---------
# now get the feature maps prior activation for all the points of the 2D grid
workplace = sknet.Workplace(dataset=dataset)
workplace.execute_worker(output)
# format those data for plotting the partitioning
features = output.epoch_data['outputs'][0]
features = features.reshape((-1, features.shape[-1]))
masks = [features[:, :4] > 0, features[:, 4:8] > 0, features[:, [8]] > 0]
boundarys_l = [
100 * B: 0.005,
200 * B: 0.001,
250 * B: 0.0005
})
optimizer = sknet.optimizers.NesterovMomentum(
loss, lr, params=dnn.variables(trainable=True))
minimizer = tf.group(optimizer.updates + dnn.updates)
# Workers
#---------
def period_function(batch, epoch):
if batch % 100 == 0:
return True
return False
train = sknet.Worker(context='train_set',
deterministic=False,
minimizer=minimizer,
loss=(loss, period_function))
test = sknet.Worker(context='test_set', deterministic=True, accuracy=accu)
queue = sknet.Queue((train, test))
workplace = sknet.utils.Workplace(dataset)
workplace.execute_queue(queue, repeat=350, deter_func=dnn.deter_dict)
# should also include some possible network dependencies present in UPDATE_OPS
learning_rate_schedule = schedule.stepwise({
0: 0.01,
5000: 0.001,
100000: 0.0001
})
optimizer = Adam(schedule=learning_rate_schedule,
dependencies=tf.get_collection(tf.GraphKeys.UPDATE_OPS))
minimizer = optimizer.minimize(loss)
# Workers
#---------
work1 = sknet.Worker(op_name='minimizer',
context='train_set',
op=minimizer,
instruction='execute every batch',
deterministic=False,
description=optimizer.description)
work2 = sknet.Worker(op_name='loss',
context='train_set',
op=loss,
instruction='save & print every 30 batch',
deterministic=False,
description='saving the loss every 30 batches')
work3 = sknet.Worker(
op_name='accuracy',
context='test_set',
op=accuracy,
instruction='execute every batch and save & print & average',
lr = sknet.schedules.PiecewiseConstant(0.001, {
90 * B: 0.0005,
160 * B: 0.0001
})
optimizer = sknet.optimizers.Adam(loss_classif + PARAMETER * loss_recons,
dnn.variables(trainable=True), lr)
minimizer = tf.group(*optimizer.updates, *dnn.updates)
reset_op = tf.group(optimizer.reset_variables_op, dnn.reset_variables_op)
# Workers
# ---------
period = sknet.dataset.BatchPeriod(1000)
train = sknet.Worker(minimizer,
loss_recons=loss_recons,
images=(dataset.images, period),
reconstructions=(reconstruction, period),
loss_classif=loss_classif,
context='train_set',
deterministic=False)
valid = sknet.Worker(loss_recons=loss_recons,
loss_classif=loss_classif,
context='valid_set',
deterministic=True)
test = sknet.Worker(loss_recons=loss_recons,
loss_classif=loss_classif,
images=(dataset.images, period),
reconstructions=(reconstruction, period),
context='test_set',
deterministic=True)