import numpy as np
import matplotlib.pyplot as plt
from otter import Variable
from otter.ops.loss import mean_squared_error
from otter.dam.graph import Graph
from otter.ops.activation import sigmoid
from otter.layers.common import Dense
from otter.optimizer import *
np.random.seed(2019)
with Graph() as g:
n = 10000
p = 10
m = 5
x = Variable(np.random.normal(0, 1, (n, p)))
y = Variable(np.random.normal(0, 1, (n, m)))
layer1 = Dense(output_shape=10, activation=sigmoid)
layer2 = Dense(output_shape=m, activation=sigmoid)
optimizer = GradientDescent(0.8)
loss = mean_squared_error
loss_array = []
for i in range(300):
if i % 50 == 0:
print(i)
a = layer1.forward(x)
b = layer2.forward(a)
conv2 = Conv2D(out_channel=16,
kernel_size=(3, 3),
stride=(2, 2))
flatten = Flatten()
dense1 = Dense(output_shape=128)
dense2 = Dense(output_shape=10)
optimizer = GradientDescent(1e-4)
loss_list = []
acc_list = []
norm1_list = []
norm2_list = []
g = Graph()
iteration = 1000
batch_size = 1024
total_epoch = int(n / batch_size)
for it_idx in range(iteration):
print(f"The {it_idx}th iteration.")
for epoch in tqdm(range(total_epoch)):
x = x_train[epoch*batch_size: (epoch+1) * batch_size]
y = y_train[epoch*batch_size: (epoch+1) * batch_size]
x = Variable(x)
y = Variable(y)
# inverse_ones = a.reshape(n, 1, m) * np.ones((m, m)) # # ds = eyed - np.multiply(ones, inverse_ones) # avg_ds = np.average(ds, axis=0) # # print(ones) # print(inverse_ones) # print(eyed) # print(ds) from otter.dam.structure import Variable a = Variable(np.random.normal(0, 1, (20, 10))) # b = softmax(a) g = Graph() # # c = a.sum(axis=1) # # g.set_and_update_gradient(c, np.arange(20).reshape(20,1)) # g.update_gradient(b) # # print(a.gradient) c = Variable(np.random.randint(0, 9, (20, 1))) sliced = b.slice(c.value.reshape((len(c.value), )), axis=1) mid1 = sliced.safe_log() mid2 = mid1.average() mid3 = mid2.neg() g.update_gradient(mid3)
x_train = np.frombuffer(imgpath.read(), np.uint8,
offset=16).reshape(len(y_train), 28, 28)
with gzip.open(paths[2], 'rb') as lbpath:
y_test = np.frombuffer(lbpath.read(), np.uint8, offset=8)
with gzip.open(paths[3], 'rb') as imgpath:
x_test = np.frombuffer(imgpath.read(), np.uint8,
offset=16).reshape(len(y_test), 28, 28)
return (x_train, y_train), (x_test, y_test)
(x_train, y_train), (x_test, y_test) = read_data()
conv = Conv2D(1, (5, 5), relu, (2, 2), bias=False)
conv.w = w
t = Variable(x_train[0].reshape(1, 1, 28, 28))
ans = conv.forward(t)
print(conv.w)
print(ans)
print(ans.shape)
plt.imshow(ans.value[0][0])
plt.show()
g = Graph()
g.update_gradient(ans)
print(t.gradient)