activation_func="tanh",
input_noise=0.3)
cnn.compileOutputFunctions(
target="regression") #compiles the cnn.get_error function as well
cnn.setOptimizerParams(SGD={'LR': 5e-1, 'momentum': 0.9}, weight_decay=0)
print "training..."
for i in range(40000):
d, l = data.getbatch(batch_size)
loss, loss_instance, time_per_step = cnn.trainingStep(d, d, mode="SGD")
if i % 100 == 0:
print "update:", i, "; Training error:", loss
cnn.layers[3].input_noise.set_value(0.0)
loss, _, test_predictions = cnn.get_error(data.valid_d.reshape((10000, 784)),
data.valid_d.reshape((10000, 784)))
print "Final error:", loss
print "Done."
plt.figure(figsize=(14, 6))
plt.subplot(121)
images = embedMatricesInGray(data.valid_d[:200].reshape((200, 28, 28)), 1)
plt.imshow(images, interpolation='none', cmap='gray')
plt.title('Data')
plt.subplot(122)
recon = embedMatricesInGray(test_predictions[:200].reshape((200, 28, 28)), 1)
plt.imshow(recon, interpolation='none', cmap='gray')
plt.title('Reconstruction')
d_v = data.valid_d
force_no_dropout=True)
cnn.addPerceptronLayer(200, activation_func="tanh")
cnn.addPerceptronLayer(150, activation_func="tanh")
cnn.addPerceptronLayer(
10, activation_func="tanh", force_no_dropout=True
) # need 10 outputs as there are 10 classes in the data set
cnn.compileOutputFunctions()
cnn.setOptimizerParams(SGD={'LR': 1e-2, 'momentum': 0.9}, weight_decay=0)
print "training..."
for i in range(10000):
d, l = data.getbatch(batch_size)
loss, loss_instance, time_per_step = cnn.trainingStep(d, l, mode="SGD")
if i % 100 == 0:
r = cnn.getDropoutRates()
cnn.setDropoutRates([
0.0,
] * len(r))
valid_loss, valid_error, valid_predictions = cnn.get_error(
data.valid_d, data.valid_l)
print "update:", i, "; Validation loss:", valid_loss, "Validation error:", valid_error * 100., "%"
cnn.setDropoutRates(r)
cnn.setDropoutRates([
0.0,
] * len(r))
loss, error, test_predictions = cnn.get_error(data.test_d, data.test_l)
print "Test loss:", loss, "Test error:", error * 100., "%"
print "Done."
# Create CNN #
batch_size = 100
cnn = MixedConvNN((28, 28), input_depth=1)
cnn.addConvLayer(10,
5,
pool_shape=2,
activation_func="abs") # (nof, filtersize)
cnn.addConvLayer(8, 5, pool_shape=2, activation_func="abs")
cnn.addPerceptronLayer(100, activation_func="abs")
cnn.addPerceptronLayer(80, activation_func="abs")
cnn.addPerceptronLayer(
10,
activation_func="abs"
) # need 10 outputs as there are 10 classes in the data set
cnn.compileOutputFunctions()
cnn.setOptimizerParams(SGD={'LR': 1e-2, 'momentum': 0.9}, weight_decay=0)
print "training..."
for i in range(5000):
d, l = data.getbatch(batch_size)
loss, loss_instance, time_per_step = cnn.trainingStep(d, l, mode="SGD")
if i % 100 == 0:
valid_loss, valid_error, valid_predictions = cnn.get_error(data.valid_d, data.valid_l)
print "update:", i, "; Validation loss:", valid_loss, "Validation error:", valid_error * 100., "%"
loss, error, test_predictions = cnn.get_error(data.test_d, data.test_l)
print "Test loss:", loss, "Test error:", error * 100., "%"
print "Done."
cnn.addTiedAutoencoderChain(n_layers=None,
activation_func="tanh",
input_noise=0.3)
cnn.compileOutputFunctions(target="regression") #compiles the cnn.get_error function as well
cnn.setOptimizerParams(SGD={'LR': 5e-1, 'momentum': 0.9}, weight_decay=0)
print "training..."
for i in range(40000):
d, l = data.getbatch(batch_size)
loss, loss_instance, time_per_step = cnn.trainingStep(d, d, mode="SGD")
if i % 100 == 0:
print "update:", i, "; Training error:", loss
cnn.layers[3].input_noise.set_value(0.0)
loss, _, test_predictions = cnn.get_error(data.valid_d.reshape((10000, 784)), data.valid_d.reshape((10000, 784)))
print "Final error:", loss
print "Done."
plt.figure(figsize=(14, 6))
plt.subplot(121)
images = embedMatricesInGray(data.valid_d[:200].reshape((200, 28, 28)), 1)
plt.imshow(images, interpolation='none', cmap='gray')
plt.title('Data')
plt.subplot(122)
recon = embedMatricesInGray(test_predictions[:200].reshape((200, 28, 28)), 1)
plt.imshow(recon, interpolation='none', cmap='gray')
plt.title('Reconstruction')
d_v = data.valid_d
