def inference(x, output_num):
conv1_out = op.Conv2D(input_variable=x, kernel_shape=(5, 5, 1, 12), name='conv1', padding=0).output_variables
relu1_out = activation.Relu(input_variable=conv1_out, name='relu1').output_variables
#dropout1_out = op.DropOut(input_variable=relu1_out, name='dropout1', phase='train', prob=0.7).output_variables
pool1_out = op.MaxPooling(input_variable=relu1_out, ksize=2, name='pool1').output_variables
conv2_out = op.Conv2D(input_variable=pool1_out, kernel_shape=(3, 3, 12, 24), name='conv2', padding=1).output_variables
relu2_out = activation.Relu(input_variable=conv2_out, name='relu2').output_variables
#dropout2_out = op.DropOut(input_variable=relu2_out, name='dropout2', phase='train', prob=0.7).output_variables
pool2_out = op.MaxPooling(input_variable=relu1_out, ksize=2, name='pool2').output_variables
fc_out = op.FullyConnect(input_variable=pool2_out, output_num=output_num, name='fc').output_variables
return fc_out
def inference(x, output_num):
conv1_out = op.Conv2D((5, 5, 1, 12),
input_variable=x,
name='conv1',
padding='VALID').output_variables
relu1_out = op.Relu(input_variable=conv1_out,
name='relu1').output_variables
pool1_out = op.MaxPooling(ksize=2, input_variable=relu1_out,
name='pool1').output_variables
conv2_out = op.Conv2D((3, 3, 12, 24),
input_variable=pool1_out,
name='conv2').output_variables
relu2_out = op.Relu(input_variable=conv2_out,
name='relu2').output_variables
pool2_out = op.MaxPooling(ksize=2, input_variable=relu2_out,
name='pool2').output_variables
fc_out = op.FullyConnect(output_num=output_num,
input_variable=pool2_out,
name='fc').output_variables
return fc_out
from layers.softmax import Softmax
import cv2
import numpy as np
img = cv2.imread('layers/test.jpg')
img = img[np.newaxis, :]
a = var.Variable((1, 128, 128, 3), 'a')
label = var.Variable([1, 1], 'label')
import random
label.data = np.array([random.randint(1, 9)])
label.data = label.data.astype(int)
conv1_out = op.Conv2D((3, 3, 3, 3),
input_variable=a,
name='conv1',
padding='VALID').output_variables
relu1_out = op.Relu(input_variable=conv1_out, name='relu1').output_variables
pool1_out = op.MaxPooling(ksize=2, input_variable=relu1_out,
name='pool1').output_variables
fc1_out = op.FullyConnect(output_num=10, input_variable=pool1_out,
name='fc1').output_variables
sf_out = op.SoftmaxLoss(predict=fc1_out, label=label, name='sf').loss
new_conv1 = op.GLOBAL_VARIABLE_SCOPE['conv1']
new_fc1 = op.GLOBAL_VARIABLE_SCOPE['fc1']
conv1 = Conv2D([1, 128, 128, 3], 3, 3, 1, method='VALID')
relu1 = Relu(conv1.output_shape)
pool1 = MaxPooling(conv1.output_shape)
fc1 = FullyConnect(pool1.output_shape, 10)
batch_size = 64
global_step = 0
# set method
for k in var.GLOBAL_VARIABLE_SCOPE:
s = var.GLOBAL_VARIABLE_SCOPE[k]
if isinstance(s, var.Variable) and s.learnable:
s.set_method_adam()
img_placeholder = var.Variable((batch_size, 28, 28, 1), 'input')
label_placeholder = var.Variable([batch_size, 1], 'label')
# set train_op
prediction = inference(img_placeholder, 10)
sf = op.SoftmaxLoss(prediction, label_placeholder, 'sf')
images, labels = load_mnist('./data/mnist')
test_images, test_labels = load_mnist('./data/mnist', 't10k')
# save train curve config
loss_collect = []
acc_collect = []
print ('new')
with open('logs/%s_log.txt'%VERSION, 'wb') as logf:
for epoch in range(20):
# random shuffle
order = np.arange(images.shape[0])
np.random.shuffle(order)
_images = images[order]
e=1e-3 a = var.Variable((1, 128, 128, 3), 'a') b = var.Variable((1, 128, 128, 3), 'b') b.data = a.data.copy() a.data[0,0,0,1] += e b.data[0,0,0,1] -= e # label = var.Variable([1, 1], 'label') # import random # label.data = np.array([random.randint(1,9)]) # label.data = label.data.astype(int) import numpy as np conv1_out = op.Conv2D((3, 3, 3, 3), input_variable=a, name='conv1',padding='VALID').output_variables conv2_out = op.Conv2D((3, 3, 3, 3), input_variable=b, name='conv2',padding='VALID').output_variables conv1 = var.GLOBAL_VARIABLE_SCOPE['conv1'] conv2 = var.GLOBAL_VARIABLE_SCOPE['conv2'] var.GLOBAL_VARIABLE_SCOPE['conv1'].weights.data = var.GLOBAL_VARIABLE_SCOPE['conv2'].weights.data var.GLOBAL_VARIABLE_SCOPE['conv1'].bias.data = var.GLOBAL_VARIABLE_SCOPE['conv2'].bias.data # print conv1.weights.data - conv2.weights.data # print conv1_out.eval()-conv2_out.eval() conv1_out.eval() conv1_out.diff.data = (np.ones(conv1_out.diff.shape)) print a.wait_bp, conv1.wait_forward
import numpy as np
### grad_check
e = 1e-3
a = var.Variable((1, 28, 28, 3), 'a')
b = var.Variable((1, 28, 28, 3), 'b')
c = var.Variable((1, 28, 28, 3), 'c')
b.data = a.data.copy()
c.data = a.data.copy()
a.data += e
b.data -= e
## conv2d
conv1_out = op.Conv2D(a, (3, 3, 3, 3), name='conv1', stride=1,
padding=1).output_variables
conv2_out = op.Conv2D(b, (3, 3, 3, 3), name='conv2', stride=1,
padding=1).output_variables
conv3_out = op.Conv2D(c, (3, 3, 3, 3), name='conv3', stride=1,
padding=1).output_variables
conv1 = var.GLOBAL_VARIABLE_SCOPE['conv1']
conv2 = var.GLOBAL_VARIABLE_SCOPE['conv2']
conv3 = var.GLOBAL_VARIABLE_SCOPE['conv3']
var.GLOBAL_VARIABLE_SCOPE['conv1'].weights.data = var.GLOBAL_VARIABLE_SCOPE[
'conv2'].weights.data
var.GLOBAL_VARIABLE_SCOPE['conv1'].bias.data = var.GLOBAL_VARIABLE_SCOPE[
'conv2'].bias.data
var.GLOBAL_VARIABLE_SCOPE['conv3'].weights.data = var.GLOBAL_VARIABLE_SCOPE[
'conv2'].weights.data
var.GLOBAL_VARIABLE_SCOPE['conv3'].bias.data = var.GLOBAL_VARIABLE_SCOPE[