def main():
# define lenet
layers = get_lenet()
# load data
# change the following value to true to load the entire dataset
fullset = True
print("Loading MNIST Dataset...")
xtrain, ytrain, xval, yval, xtest, ytest = cnn_lenet.load_mnist(fullset)
print("MNIST Dataset Loading Complete!\n")
xtrain = np.hstack([xtrain, xval])
ytrain = np.hstack([ytrain, yval])
m_train = xtrain.shape[1]
# cnn parameters
batch_size = 64
mu = 0.9
epsilon = 0.01
gamma = 0.0001
power = 0.75
weight_decay = 0.0005
w_lr = 1
b_lr = 2
test_interval = 15
display_interval = 15
snapshot = 5000
max_iter = 30
# initialize parameters
print("Initializing Parameters...")
params = cnn_lenet.init_convnet(layers)
param_winc = copy.deepcopy(params)
print("Initilization Complete!\n")
for l_idx in range(1, len(layers)):
param_winc[l_idx]['w'] = np.zeros(param_winc[l_idx]['w'].shape)
param_winc[l_idx]['b'] = np.zeros(param_winc[l_idx]['b'].shape)
# learning iterations
random.seed(100000)
indices = range(m_train)
random.shuffle(indices)
print("Training Started. Printing report on training data every " +
str(display_interval) + " steps.")
print("Printing report on test data every " + str(test_interval) +
" steps.\n")
train_acc_100 = []
test_acc_100 = []
train_cost_100 = []
test_cost_100 = []
program_starts = time.time()
for step in range(max_iter):
# get mini-batch and setup the cnn with the mini-batch
start_idx = step * batch_size % m_train
end_idx = (step + 1) * batch_size % m_train
if start_idx > end_idx:
random.shuffle(indices)
continue
idx = indices[start_idx:end_idx]
layers[2]['iteration'] = step
[cp, param_grad] = cnn_lenet.conv_net(params, layers, xtrain[:, idx],
ytrain[idx], True)
# we have different epsilons for w and b
w_rate = cnn_lenet.get_lr(step, epsilon * w_lr, gamma, power)
b_rate = cnn_lenet.get_lr(step, epsilon * b_lr, gamma, power)
params, param_winc = cnn_lenet.sgd_momentum(w_rate, b_rate, mu,
weight_decay, params,
param_winc, param_grad)
# display training loss
if (step + 1) % display_interval == 0:
train_cost_100.append(cp['cost'])
train_acc_100.append(cp['percent'])
print 'training_cost = %f training_accuracy = %f' % (
cp['cost'], cp['percent']) + ' current_step = ' + str(step + 1)
now = time.time()
print "It has been {0} seconds since the training started".format(
now - program_starts)
# display test accuracy
if (step + 1) % test_interval == 0:
layers[1]['batch_size'] = xtest.shape[1]
cptest, _ = cnn_lenet.conv_net(params, layers, xtest, ytest, False)
layers[1]['batch_size'] = 64
test_cost_100.append(cptest['cost'])
test_acc_100.append(cptest['percent'])
print 'test_cost = %f test_accuracy = %f' % (
cptest['cost'],
cptest['percent']) + ' current_step = ' + str(step + 1) + '\n'
# save params peridocally to recover from any crashes
if (step + 1) % snapshot == 0:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
def main():
# define lenet
layers = get_lenet()
# load data
# change the following value to true to load the entire dataset
fullset = False
xtrain, ytrain, xval, yval, xtest, ytest = cnn_lenet.load_mnist(fullset)
xtrain = np.hstack([xtrain, xval])
ytrain = np.hstack([ytrain, yval])
m_train = xtrain.shape[1]
# cnn parameters
batch_size = 64
mu = 0.9
epsilon = 0.01
gamma = 0.0001
power = 0.75
weight_decay = 0.0005
w_lr = 1
b_lr = 2
test_interval = 200
display_interval = 50
snapshot = 500
max_iter = 10000
# initialize parameters
params = cnn_lenet.init_convnet(layers)
param_winc = copy.deepcopy(params)
for l_idx in range(1, len(layers)):
param_winc[l_idx]['w'] = np.zeros(param_winc[l_idx]['w'].shape)
param_winc[l_idx]['b'] = np.zeros(param_winc[l_idx]['b'].shape)
# learning iterations
indices = range(m_train)
random.shuffle(indices)
for step in range(max_iter):
# get mini-batch and setup the cnn with the mini-batch
start_idx = step * batch_size % m_train
end_idx = (step+1) * batch_size % m_train
if start_idx > end_idx:
random.shuffle(indices)
continue
idx = indices[start_idx: end_idx]
[cp, param_grad] = cnn_lenet.conv_net(params,
layers,
xtrain[:, idx],
ytrain[idx])
# we have different epsilons for w and b
w_rate = cnn_lenet.get_lr(step, epsilon*w_lr, gamma, power)
b_rate = cnn_lenet.get_lr(step, epsilon*b_lr, gamma, power)
params, param_winc = cnn_lenet.sgd_momentum(w_rate,
b_rate,
mu,
weight_decay,
params,
param_winc,
param_grad)
# display training loss
if (step+1) % display_interval == 0:
print 'cost = %f training_percent = %f' % (cp['cost'], cp['percent'])
# display test accuracy
if (step+1) % test_interval == 0:
layers[1]['batch_size'] = xtest.shape[1]
cptest, _ = cnn_lenet.conv_net(params, layers, xtest, ytest)
layers[1]['batch_size'] = 64
print '\ntest accuracy: %f\n' % (cptest['percent'])
# save params peridocally to recover from any crashes
if (step+1) % snapshot == 0:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
def trainNet():
# define lenet
layers = get_lenet()
# load data
# change the following value to true to load the entire dataset
fullset = True
print("Loading MNIST Dataset...")
xtrain, ytrain, xval, yval, xtest, ytest = cnn_lenet.load_mnist(fullset)
print("MNIST Dataset Loading Complete!\n")
xtrain = np.hstack([xtrain, xval])
ytrain = np.hstack([ytrain, yval])
m_train = xtrain.shape[1]
# cnn parameters
batch_size = 64
mu = 0.9
epsilon = 0.01
gamma = 0.0001
power = 0.75
weight_decay = 0.0005
w_lr = 1
b_lr = 2
test_interval = 100
display_interval = 100
snapshot = 5000
max_iter = 10000
# Lets it run the entire way
# initialize parameters
print("Initializing Parameters...")
# You can make the params your params, and not the initialized ones, in order to visualize the results
params = cnn_lenet.init_convnet(layers)
param_winc = copy.deepcopy(params)
print("Initilization Complete!\n")
for l_idx in range(1, len(layers)):
param_winc[l_idx]['w'] = np.zeros(param_winc[l_idx]['w'].shape)
param_winc[l_idx]['b'] = np.zeros(param_winc[l_idx]['b'].shape)
# learning iterations
random.seed(100000)
indices = range(m_train)
random.shuffle(indices)
train_cost = np.array([])
train_accuracy = np.array([])
test_cost = np.array([])
test_accuracy = np.array([])
print("Training Started. Printing report on training data every " +
str(display_interval) + " steps.")
print("Printing report on test data every " + str(test_interval) +
" steps.\n")
for step in range(max_iter):
# get mini-batch and setup the cnn with the mini-batch
start_idx = step * batch_size % m_train
end_idx = (step + 1) * batch_size % m_train
if start_idx > end_idx:
random.shuffle(indices)
continue
idx = indices[start_idx:end_idx]
[cp, param_grad] = cnn_lenet.conv_net(params, layers, xtrain[:, idx],
ytrain[idx], True)
# True there is to get backtracking, but you can just use it for forward, to visualize
# You have to make the function return output for you, so that you can reshape it into an image matrix, to show the image
# we have different epsilons for w and b
w_rate = cnn_lenet.get_lr(step, epsilon * w_lr, gamma, power)
b_rate = cnn_lenet.get_lr(step, epsilon * b_lr, gamma, power)
params, param_winc = cnn_lenet.sgd_momentum(w_rate, b_rate, mu,
weight_decay, params,
param_winc, param_grad)
# display training loss
if (step + 1) % display_interval == 0:
print 'training_cost = %f training_accuracy = %f' % (
cp['cost'], cp['percent']) + ' current_step = ' + str(step + 1)
train_cost = np.append(train_cost, cp['cost'])
train_accuracy = np.append(train_accuracy, cp['percent'])
# display test accuracy
if (step + 1) % test_interval == 0:
layers[1]['batch_size'] = xtest.shape[1]
cptest, _ = cnn_lenet.conv_net(params, layers, xtest, ytest, False)
layers[1]['batch_size'] = 64
print 'test_cost = %f test_accuracy = %f' % (
cptest['cost'],
cptest['percent']) + ' current_step = ' + str(step + 1) + '\n'
test_cost = np.append(test_cost, cptest['cost'])
test_accuracy = np.append(test_accuracy, cptest['percent'])
# save params peridocally to recover from any crashes
if (step + 1) % snapshot == 0:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
# Saves params at 30 for Question 4
if (step + 1) == 30:
pickle_path = 'lenetAt30Iterations.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
if (step + 1) == max_iter:
np.savetxt('trainCost.txt', train_cost)
np.savetxt('trainAccuracy.txt', train_accuracy)
np.savetxt('testCost.txt', test_cost)
np.savetxt('testAccuracy.txt', test_accuracy)
# np.savetxt('costsStacked.txt', np.column_stack(train_cost, test_cost))
# np.savetxt('accuracyStacked.txt', np.column_stack(train_accuracy, test_accuracy))
pickle_path = 'lenetAt10000Iterations.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
if (step) == max_iter:
np.savetxt('trainCost1.txt', train_cost)
np.savetxt('trainAccuracy1.txt', train_accuracy)
np.savetxt('testCost1.txt', test_cost)
np.savetxt('testAccuracy1.txt', test_accuracy)
# np.savetxt('costsStacked1.txt', np.column_stack(train_cost, test_cost))
# np.savetxt('accuracyStacked1.txt', np.column_stack(train_accuracy, test_accuracy))
pickle_path = 'lenetAtMAXPLUSONEIterations.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
def main():
# define lenet
layers = get_lenet()
# load data
# change the following value to true to load the entire dataset
fullset = True
print("Loading MNIST Dataset...")
xtrain, ytrain, xval, yval, xtest, ytest = cnn_lenet.load_mnist(fullset)
print("MNIST Dataset Loading Complete!\n")
xtrain = np.hstack([xtrain, xval])
ytrain = np.hstack([ytrain, yval])
m_train = xtrain.shape[1]
# cnn parameters
batch_size = 64
mu = 0.9
epsilon = 0.01
gamma = 0.0001
power = 0.75
weight_decay = 0.0005
w_lr = 1
b_lr = 2
test_interval = 100
display_interval = 100
snapshot = 5000
max_iter = 10000
# initialize parameters
print("Initializing Parameters...")
params = cnn_lenet.init_convnet(layers)
param_winc = copy.deepcopy(params)
print("Initilization Complete!\n")
for l_idx in range(1, len(layers)):
param_winc[l_idx]['w'] = np.zeros(param_winc[l_idx]['w'].shape)
param_winc[l_idx]['b'] = np.zeros(param_winc[l_idx]['b'].shape)
# learning iterations
random.seed(100000)
indices = range(m_train)
random.shuffle(indices)
print("Training Started. Printing report on training data every " +
str(display_interval) + " steps.")
print("Printing report on test data every " + str(test_interval) +
" steps.\n")
startTime = time.time()
numSamples = int(max_iter / display_interval)
costResults = np.zeros((2, numSamples))
accuracyResults = np.zeros((2, numSamples))
displayIndex = 0
for step in range(max_iter):
print("current step is: {}".format(step))
# get mini-batch and setup the cnn with the mini-batch
start_idx = step * batch_size % m_train
end_idx = (step + 1) * batch_size % m_train
if start_idx > end_idx:
random.shuffle(indices)
continue
idx = indices[start_idx:end_idx]
[output, cp,
param_grad] = cnn_lenet.conv_net(params, layers, xtrain[:, idx],
ytrain[idx], True)
# we have different epsilons for w and b
w_rate = cnn_lenet.get_lr(step, epsilon * w_lr, gamma, power)
b_rate = cnn_lenet.get_lr(step, epsilon * b_lr, gamma, power)
params, param_winc = cnn_lenet.sgd_momentum(w_rate, b_rate, mu,
weight_decay, params,
param_winc, param_grad)
# display training loss
if (step + 1) % display_interval == 0:
print 'training_cost = %f training_accuracy = %f' % (
cp['cost'], cp['percent']) + ' current_step = ' + str(step + 1)
costResults[0, displayIndex] = cp['cost']
accuracyResults[0, displayIndex] = cp['percent']
# display test accuracy
if (step + 1) % test_interval == 0:
layers[1]['batch_size'] = xtest.shape[1]
_, cptest, _ = cnn_lenet.conv_net(params, layers, xtest, ytest,
False)
layers[1]['batch_size'] = 64
print 'test_cost = %f test_accuracy = %f' % (
cptest['cost'],
cptest['percent']) + ' current_step = ' + str(step + 1) + '\n'
costResults[1, displayIndex] = cptest['cost']
accuracyResults[1, displayIndex] = cptest['percent']
displayIndex += 1
# save params peridocally to recover from any crashes
if (step + 1) % snapshot == 0:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
# print(output[1]['data'].shape, output[2]['data'].shape, output[3]['data'].shape)
if step == 29:
layers[1]['batch_size'] = 1
output, _, _ = cnn_lenet.conv_net(params, layers, xtest, ytest,
False)
layers[1]['batch_size'] = 64
printableData = output[2]['data'][:, 0]
printableData.shape = [24, 24, 20]
plt.figure()
for num in range(20):
plt.subplot(4, 5, num + 1)
plt.imshow(printableData[:, :, num])
plt.show()
endTime = time.time()
print(endTime - startTime)
pickle.dump(accuracyResults, open('accuracyResults.p', 'wb'))
pickle.dump(costResults, open('costResults.p', 'wb'))
layers[1]['batch_size'] = 1
output, _, _ = cnn_lenet.conv_net(params, layers, xtest, ytest, False)
layers[1]['batch_size'] = 64
pickle.dump(output, open('output.p', 'wb'))
# print(output[1]['data'].shape, output[2]['data'].shape,)
print("printing original image")
printableData = output[1]['data'][:, 0]
printableData.shape = [28, 28]
plt.figure()
plt.imshow(printableData)
plt.show()
print("printing output 2")
printableData = output[2]['data'][:, 0]
printableData.shape = [24, 24, 20]
plt.figure()
for num in range(20):
plt.subplot(4, 5, num + 1)
plt.imshow(printableData[:, :, num])
plt.show()
print("printing output 3")
printableData = output[3]['data'][:, 0]
printableData.shape = [12, 12, 20]
plt.figure()
for num in range(20):
plt.subplot(4, 5, num + 1)
plt.imshow(printableData[:, :, num])
plt.show()
plt.figure()
plt.plot(np.linspace(1, numSamples, numSamples),
costResults[0, :],
color='k',
linewidth=3,
label='train cost')
plt.plot(np.linspace(1, numSamples, numSamples),
costResults[1, :],
color='r',
linewidth=3,
linestyle=':',
label='test cost')
plt.title('cost comparison')
plt.legend()
plt.xlabel('iteration/100')
plt.ylabel('cost')
plt.grid()
plt.show()
plt.figure()
plt.plot(np.linspace(1, numSamples, numSamples),
accuracyResults[0, :],
color='k',
linewidth=3,
label='train accuracy')
plt.plot(np.linspace(1, numSamples, numSamples),
accuracyResults[1, :],
color='r',
linewidth=3,
linestyle=':',
label='test accuracy')
plt.title('accuracy comparison')
plt.legend()
plt.xlabel('iteration/100')
plt.ylabel('accuracy')
plt.grid()
plt.show()
def main():
# define lenet
layers = get_lenet()
mat1 = np.zeros((1,1))
mat2 = np.zeros((1,1))
mat1traic = np.zeros((1,1))
mat1testc = np.zeros((1,1))
mat1traiac = np.zeros((1,1))
mat1testac = np.zeros((1,1))
# load data
# change the following value to true to load the entire dataset
fullset = True
print("Loading MNIST Dataset...")
xtrain, ytrain, xval, yval, xtest, ytest = cnn_lenet.load_mnist(fullset)
print("MNIST Dataset Loading Complete!\n")
xtrain = np.hstack([xtrain, xval])
ytrain = np.hstack([ytrain, yval])
m_train = xtrain.shape[1]
# cnn parameters
batch_size = 64
mu = 0.9
epsilon = 0.01
gamma = 0.0001
power = 0.75
weight_decay = 0.0005
w_lr = 1
b_lr = 2
test_interval = 100
display_interval = 100
snapshot = 5000
max_iter = 10000
# initialize parameters
print("Initializing Parameters...")
params = cnn_lenet.init_convnet(layers)
param_winc = copy.deepcopy(params)
print("Initilization Complete!\n")
for l_idx in range(1, len(layers)):
param_winc[l_idx]['w'] = np.zeros(param_winc[l_idx]['w'].shape)
param_winc[l_idx]['b'] = np.zeros(param_winc[l_idx]['b'].shape)
# learning iterations
random.seed(100000)
indices = range(m_train)
random.shuffle(indices)
print("Training Started. Printing report on training data every " + str(display_interval) + " steps.")
print("Printing report on test data every " + str(test_interval) + " steps.\n")
for step in range(max_iter):
# get mini-batch and setup the cnn with the mini-batch
start_idx = step * batch_size % m_train
end_idx = (step+1) * batch_size % m_train
if start_idx > end_idx:
random.shuffle(indices)
continue
idx = indices[start_idx: end_idx]
[cp, param_grad] = cnn_lenet.conv_net(params,
layers,
xtrain[:, idx],
ytrain[idx], True)
# we have different epsilons for w and b
w_rate = cnn_lenet.get_lr(step, epsilon*w_lr, gamma, power)
b_rate = cnn_lenet.get_lr(step, epsilon*b_lr, gamma, power)
params, param_winc = cnn_lenet.sgd_momentum(w_rate,
b_rate,
mu,
weight_decay,
params,
param_winc,
param_grad)
# display training loss
if (step+1) % display_interval == 0:
print 'training_cost = %f training_accuracy = %f' % (cp['cost'], cp['percent']) + ' current_step = ' + str(step + 1)
#show input to maxpooling
# display test accuracy
if (step+1) % test_interval == 0:
layers[1]['batch_size'] = xtest.shape[1]
cptest, _ = cnn_lenet.conv_net(params, layers, xtest, ytest, False)
layers[1]['batch_size'] = 64
print 'test_cost = %f test_accuracy = %f' % (cptest['cost'], cptest['percent']) + ' current_step = ' + str(step + 1) + '\n'
# save params peridocally to recover from any crashes
if (step+1) % snapshot == 0:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()
if (step+1) % test_interval == 0:
mat1traic = np.insert(mat1traic, 0, cp['cost'])
mat1testc = np.insert(mat1testc, 0, cptest['cost'])
mat1traiac = np.insert(mat1traiac, 0, cp['percent'])
mat1testac = np.insert(mat1testac, 0, cptest['percent'])
print mat1traic
if (step+1) == 30:
pickle_path = 'lay2it30w.mat'
pickle_file = open(pickle_path, 'wb')
mat1 = params[1]['w']
pickle.dump(mat1, pickle_file)
pickle_file.close()
pickle_path = 'lay2it30b.mat'
pickle_file = open(pickle_path, 'wb')
mat1 = params[1]['b']
pickle.dump(mat1, pickle_file)
pickle_file.close()
pickle_path = 'inputdata30it.mat'
pickle_file = open(pickle_path, 'wb')
mat1 = xtrain[:, idx]
pickle.dump(mat1, pickle_file)
pickle_file.close()
if (step+1) == 1000:
pickle_path = 'lay2it10000w.mat'
pickle_file = open(pickle_path, 'wb')
mat2 = params[1]['w']
pickle.dump(mat2, pickle_file)
pickle_file.close()
pickle_path = 'lay2it10000b.mat'
pickle_file = open(pickle_path, 'wb')
mat2 = params[1]['b']
pickle.dump(mat2, pickle_file)
pickle_file.close()
pickle_path = 'inputdata10000it.mat'
pickle_file = open(pickle_path, 'wb')
mat2 = xtrain[:, idx]
pickle.dump(mat2, pickle_file)
pickle_file.close()
pickle_path = 'trainingcost.mat'
pickle_file = open(pickle_path, 'wb')
#mat1 = np.zeros((1,1))
pickle.dump(mat1traic, pickle_file)
pickle_file.close()
pickle_path = 'testcost.mat'
pickle_file = open(pickle_path, 'wb')
#mat1 = np.zeros((1,1))
pickle.dump(mat1testc, pickle_file)
pickle_file.close()
pickle_path = 'trainingaccu.mat'
pickle_file = open(pickle_path, 'wb')
#mat1 = np.zeros((1,1))
pickle.dump(mat1traiac, pickle_file)
pickle_file.close()
pickle_path = 'testaccu.mat'
pickle_file = open(pickle_path, 'wb')
#mat1 = np.zeros((1,1))
pickle.dump(mat1testac, pickle_file)
pickle_file.close()
def main():
# define lenet
layers = get_lenet()
# load data
# change the following value to true to load the entire dataset
fullset = False
"""xtrain, ytrain, xval, yval, xtest, ytest = cnn_lenet.load_mnist(fullset)"""
negPath = "/home/lui/CMU/Clinical-feature-learning/dataset/negative.txt"
posPath = "/home/lui/CMU/Clinical-feature-learning/dataset/positive.txt"
XTrain, yTrain = seq2mat.genTrain(posPath, negPath)
print XTrain.shape, yTrain.shape
threshold = XTrain.shape[1] / 10 * 9
print "threshold: ", threshold
XTrainTrue = XTrain[:, :threshold]
Xtest = XTrain[:, threshold:]
yTrainTrue = yTrain[:threshold]
ytest = yTrain[threshold:]
print "finish loading data"
print "shape: xTrain, yTrain, xTest, yTest"
print XTrainTrue.shape, yTrainTrue.shape, Xtest.shape, ytest.shape
"""xtrain = np.hstack([xtrain, xval])
ytrain = np.hstack([ytrain, yval])"""
m_train = XTrainTrue.shape[1]
# cnn parameters
batch_size = 64
mu = 0.9
epsilon = 0.01
gamma = 0.0001
power = 0.75
weight_decay = 0.0005
w_lr = 1
b_lr = 2
test_interval = 500
display_interval = 10
snapshot = 100
max_iter = 10000
loadPara = bool(raw_input("load parameters, True or False: "))
if loadPara:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'rb')
params = pickle.load(pickle_file)
param_winc = copy.deepcopy(params)
else:
# initialize parameters
params = cnn_lenet.init_convnet(layers)
param_winc = copy.deepcopy(params)
for l_idx in range(1, len(layers)):
param_winc[l_idx]['w'] = np.zeros(param_winc[l_idx]['w'].shape)
param_winc[l_idx]['b'] = np.zeros(param_winc[l_idx]['b'].shape)
# learning iterations
indices = range(m_train)
random.shuffle(indices)
for step in range(1500, max_iter):
# get mini-batch and setup the cnn with the mini-batch
start_idx = step * batch_size % m_train
end_idx = (step + 1) * batch_size % m_train
if start_idx > end_idx:
random.shuffle(indices)
continue
idx = indices[start_idx:end_idx]
[cp, param_grad] = cnn_lenet.conv_net(params, layers, XTrainTrue[:,
idx],
yTrainTrue[idx])
# we have different epsilons for w and b
w_rate = cnn_lenet.get_lr(step, epsilon * w_lr, gamma, power)
b_rate = cnn_lenet.get_lr(step, epsilon * b_lr, gamma, power)
params, param_winc = cnn_lenet.sgd_momentum(w_rate, b_rate, mu,
weight_decay, params,
param_winc, param_grad)
# display training loss
if (step + 1) % display_interval == 0:
print 'cost = %f training_percent = %f' % (cp['cost'],
cp['percent'])
# display test accuracy
if (step + 1) % test_interval == 0:
layers[1]['batch_size'] = Xtest.shape[1]
cptest, _ = cnn_lenet.conv_net(params, layers, Xtest, ytest)
layers[1]['batch_size'] = 64
print '\ntest accuracy: %f\n' % (cptest['percent'])
# save params peridocally to recover from any crashes
if (step + 1) % snapshot == 0:
pickle_path = 'lenet.mat'
pickle_file = open(pickle_path, 'wb')
pickle.dump(params, pickle_file)
pickle_file.close()