def test_scalar_value():
npr.seed(1)
for ii in xrange(NUM_TRIALS):
np_X = npr.randn()
X = kayak.Parameter(np_X)
out = kayak.L2Norm(X)
assert close_float(out.value, np_X**2)
def test_matrix_value():
npr.seed(7)
for ii in xrange(NUM_TRIALS):
np_X = npr.randn(10, 20)
wt = np.exp(npr.randn())
X = kayak.Parameter(np_X)
out = kayak.L2Norm(X, weight=wt)
assert close_float(out.value, wt * np.sum(np_X**2))
def test_scalar_value_2():
npr.seed(3)
for ii in xrange(NUM_TRIALS):
np_X = npr.randn()
wt = np.exp(npr.randn())
X = kayak.Parameter(np_X)
out = kayak.L2Norm(X, weight=wt)
assert close_float(out.value, wt * np_X**2)
def test_scalar_grad():
npr.seed(2)
for ii in xrange(NUM_TRIALS):
np_X = npr.randn()
X = kayak.Parameter(np_X)
out = kayak.L2Norm(X)
assert close_float(out.grad(X), 2 * np_X)
assert kayak.util.checkgrad(X, out) < MAX_GRAD_DIFF
def test_vector_grad():
npr.seed(6)
for ii in xrange(NUM_TRIALS):
np_X = npr.randn(10, 1)
wt = np.exp(npr.randn())
X = kayak.Parameter(np_X)
out = kayak.L2Norm(X, weight=wt)
assert np.all(close_float(out.grad(X), 2 * wt * np_X))
assert kayak.util.checkgrad(X, out) < MAX_GRAD_DIFF
def test_scalar_grad_2():
npr.seed(4)
for ii in xrange(NUM_TRIALS):
np_X = npr.randn()
wt = np.exp(npr.randn())
X = kayak.Parameter(np_X)
out = kayak.L2Norm(X, weight=wt)
assert close_float(out.grad(X), 2 * wt * np_X)
assert kayak.util.checkgrad(X, out) < MAX_GRAD_DIFF
def train(inputs, targets, batch_size, learn_rate, momentum, l1_weight,
l2_weight, dropout, improvement_thresh):
# Create a batcher object.
batcher = kayak.Batcher(batch_size, inputs.shape[0])
# Inputs and targets need access to the batcher.
X = kayak.Inputs(inputs, batcher)
T = kayak.Targets(targets, batcher)
# Put some dropout regularization on the inputs
H = kayak.Dropout(X, dropout)
# Weights and biases, with random initializations.
W = kayak.Parameter(0.1 * npr.randn(inputs.shape[1], 10))
B = kayak.Parameter(0.1 * npr.randn(1, 10))
# Nothing fancy here: inputs times weights, plus bias, then softmax.
Y = kayak.LogSoftMax(kayak.ElemAdd(kayak.MatMult(H, W), B))
# The training loss is negative multinomial log likelihood.
loss = kayak.MatAdd(kayak.MatSum(kayak.LogMultinomialLoss(Y, T)),
kayak.L2Norm(W, l2_weight), kayak.L1Norm(W, l1_weight))
# Use momentum for the gradient-based optimization.
mom_grad_W = np.zeros(W.shape)
best_loss = np.inf
best_epoch = -1
# Loop over epochs.
for epoch in range(100):
# Track the total loss.
total_loss = 0.0
# Loop over batches -- using batcher as iterator.
for batch in batcher:
# Draw new random dropouts
H.draw_new_mask()
# Compute the loss of this minibatch by asking the Kayak
# object for its value and giving it reset=True.
total_loss += loss.value
# Now ask the loss for its gradient in terms of the
# weights and the biases -- the two things we're trying to
# learn here.
grad_W = loss.grad(W)
grad_B = loss.grad(B)
# Use momentum on the weight gradient.
mom_grad_W *= momentum
mom_grad_W += (1.0 - momentum) * grad_W
# Now make the actual parameter updates.
W.value -= learn_rate * mom_grad_W
B.value -= learn_rate * grad_B
print("Epoch: %d, total loss: %f" % (epoch, total_loss))
if not np.isfinite(total_loss):
print("Training diverged. Returning constraint violation.")
break
if total_loss < best_loss:
best_epoch = epoch
else:
if (epoch - best_epoch) > improvement_thresh:
print("Has been %d epochs without improvement. Aborting." %
(epoch - best_epoch))
break
# After we've trained, we return a sugary little function handle
# that makes things easy. Basically, what we're doing here is
# simply replacing the inputs in the above defined graph and then
# running through it to produce the outputs.
# The point here is that we wind up with a function
# handle the can be called with a numpy object and it produces the
# target values for novel data, using the parameters we just learned.
def predict(x):
X.value = x
H.reinstate_units()
return Y.value
return predict
def kayak_mlp(X, y):
"""
Kayak implementation of a mlp with relu hidden layers and dropout
"""
# Create a batcher object.
batcher = kayak.Batcher(batch_size, X.shape[0])
# count number of rows and columns
num_examples, num_features = np.shape(X)
X = kayak.Inputs(X, batcher)
T = kayak.Targets(y, batcher)
# ----------------------------- first hidden layer -------------------------------
# set up weights for our input layer
# use the same scheme as our numpy mlp
input_range = 1.0 / num_features**(1 / 2)
weights_1 = kayak.Parameter(0.1 * np.random.randn(X.shape[1], layer1_size))
bias_1 = kayak.Parameter(0.1 * np.random.randn(1, layer1_size))
# linear combination of weights and inputs
hidden_1_input = kayak.ElemAdd(kayak.MatMult(X, weights_1), bias_1)
# apply activation function to hidden layer
hidden_1_activation = kayak.HardReLU(hidden_1_input)
# apply a dropout for regularization
hidden_1_out = kayak.Dropout(hidden_1_activation,
layer1_dropout,
batcher=batcher)
# ----------------------------- output layer -----------------------------------
weights_out = kayak.Parameter(0.1 * np.random.randn(layer1_size, 9))
bias_out = kayak.Parameter(0.1 * np.random.randn(1, 9))
# linear combination of layer2 output and output weights
out = kayak.ElemAdd(kayak.MatMult(hidden_1_out, weights_out), bias_out)
# apply activation function to output
yhat = kayak.SoftMax(out)
# ----------------------------- loss function -----------------------------------
loss = kayak.MatAdd(kayak.MatSum(kayak.L2Loss(yhat, T)),
kayak.L2Norm(weights_1, layer1_l2))
# Use momentum for the gradient-based optimization.
mom_grad_W1 = np.zeros(weights_1.shape)
mom_grad_W2 = np.zeros(weights_out.shape)
# Loop over epochs.
plot_loss = np.ones((iterations, 2))
for epoch in xrange(iterations):
# Track the total loss.
total_loss = 0.0
for batch in batcher:
# Compute the loss of this minibatch by asking the Kayak
# object for its value and giving it reset=True.
total_loss += loss.value
# Now ask the loss for its gradient in terms of the
# weights and the biases -- the two things we're trying to
# learn here.
grad_W1 = loss.grad(weights_1)
grad_B1 = loss.grad(bias_1)
grad_W2 = loss.grad(weights_out)
grad_B2 = loss.grad(bias_out)
# Use momentum on the weight gradients.
mom_grad_W1 = momentum * mom_grad_W1 + (1.0 - momentum) * grad_W1
mom_grad_W2 = momentum * mom_grad_W2 + (1.0 - momentum) * grad_W2
# Now make the actual parameter updates.
weights_1.value -= learn_rate * mom_grad_W1
bias_1.value -= learn_rate * grad_B1
weights_out.value -= learn_rate * mom_grad_W2
bias_out.value -= learn_rate * grad_B2
# save values into table to print learning curve at the end of trianing
plot_loss[epoch, 0] = epoch
plot_loss[epoch, 1] = total_loss
print epoch, total_loss
#pyplot.plot(plot_loss[:,0], plot_loss[:,1], linewidth=2.0)
#pyplot.show()
def compute_predictions(x):
X.data = x
batcher.test_mode()
return yhat.value
return compute_predictions
def train(inputs, targets, batch_size, learn_rate, momentum, l1_weight, l2_weight, dropout):
# Create a batcher object.
batcher = kayak.Batcher(batch_size, inputs.shape[0])
# Inputs and targets need access to the batcher.
X = kayak.Inputs(inputs, batcher)
T = kayak.Targets(targets, batcher)
# Weights and biases, with random initializations.
W = kayak.Parameter( 0.1*npr.randn( inputs.shape[1], 10 ))
B = kayak.Parameter( 0.1*npr.randn(1,10) )
# Nothing fancy here: inputs times weights, plus bias, then softmax.
dropout_layer = kayak.Dropout(X, dropout, batcher=batcher)
Y = kayak.LogSoftMax( kayak.ElemAdd( kayak.MatMult(dropout_layer, W), B ) )
# The training loss is negative multinomial log likelihood.
loss = kayak.MatAdd(kayak.MatSum(kayak.LogMultinomialLoss(Y, T)),
kayak.L2Norm(W, l2_weight),
kayak.L1Norm(W, l1_weight))
# Use momentum for the gradient-based optimization.
mom_grad_W = np.zeros(W.shape)
# Loop over epochs.
for epoch in xrange(10):
# Track the total loss and the overall gradient.
total_loss = 0.0
total_grad_W = np.zeros(W.shape)
# Loop over batches -- using batcher as iterator.
for batch in batcher:
# Compute the loss of this minibatch by asking the Kayak
# object for its value and giving it reset=True.
total_loss += loss.value
# Now ask the loss for its gradient in terms of the
# weights and the biases -- the two things we're trying to
# learn here.
grad_W = loss.grad(W)
grad_B = loss.grad(B)
# Use momentum on the weight gradient.
mom_grad_W = momentum*mom_grad_W + (1.0-momentum)*grad_W
# Now make the actual parameter updates.
W.value -= learn_rate * mom_grad_W
B.value -= learn_rate * grad_B
# Keep track of the gradient to see if we're converging.
total_grad_W += grad_W
#print epoch, total_loss, np.sum(total_grad_W**2)
# After we've trained, we return a sugary little function handle
# that makes things easy. Basically, what we're doing here is
# handing the output object (not the loss!) a dictionary where the
# key is the Kayak input object 'X' (that is the features being
# used here for logistic regression) and the value in that
# dictionary is being determined by the argument to the lambda
# expression. The point here is that we wind up with a function
# handle the can be called with a numpy object and it produces the
# target values for novel data, using the parameters we just learned.
def compute_predictions(x):
X.data = x
batcher.test_mode()
return Y.value
return compute_predictions