def model_forward(x, net, down_samples, class_preds, box_preds, cap_transforms,
sizes, ratios):
# extract feature with the body network
x = net(x)
# for each scale, add anchors, box and class predictions,
# then compute the input to next scale
default_anchors = []
predicted_boxes = []
predicted_classes = []
for i in range(5):
default_anchors.append(
MultiBoxPrior(x, sizes=sizes[i], ratios=ratios[i]))
prime_out = cap_transforms[i](x)
class_capout = class_preds[i * 2](prime_out)
class_pred = class_preds[i * 2 + 1](class_capout)
class_pred = nd.flatten(nd.transpose(class_pred, (0, 2, 3, 1)))
'''
class_pred = class_preds[i](x)
class_pred = nd.flatten(nd.transpose(class_pred, (0,2,3,1)))
'''
box_pred = nd.flatten(nd.transpose(box_preds[i](x), (0, 2, 3, 1)))
# print class_pred.shape, box_pred.shape
print class_pred.shape
# print class_pred.shape
predicted_boxes.append(box_pred)
predicted_classes.append(class_pred)
if i < 3:
x = down_samples[i](x)
elif i == 3:
# simply use the pooling layer
x = nd.Pooling(x, global_pool=True, pool_type='max', kernel=(4, 4))
return default_anchors, predicted_classes, predicted_boxes
def net(X, verbose=False):
X = X.reshape((batch_size, 1, 28, 28))
out1 = nd.Convolution(data=X,
weight=w1,
bias=b1,
kernel=w1.shape[2:],
num_filter=w1.shape[0])
out2 = nd.relu(out1)
out3 = nd.Pooling(data=out2, pool_type="max", kernel=(2, 2), stride=(2, 2))
out4 = nd.Convolution(data=out3,
weight=w2,
bias=b2,
kernel=w2.shape[2:],
num_filter=w2.shape[0])
out5 = nd.relu(out4)
out6 = nd.Pooling(data=out5, pool_type="max", kernel=(2, 2), stride=(2, 2))
out7 = nd.dot(nd.flatten(out6), w3) + b3
out8 = nd.relu(out7)
out9 = nd.dot(out8, w4) + b4
if verbose:
print('1st conv block:', out3.shape)
print('2nd conv block:', out5.shape)
print('2nd conv block:', out6.shape)
print('2nd conv block:', out7.shape)
print('1st dense:', out8.shape)
print('2nd dense:', out9.shape)
# print('output:', out9)
return out9
def LetNet_Direct(X, verbose=False):
# 初始化参数
params = initialize_params()
W1, b1, W2, b2, W3, b3, W4, b4 = params
# 1. 将数据集copy到GPU
X = X.as_in_context(ctx)
# 2. 定义卷积层
# 2.1 卷积层一
h1_conv = nd.Convolution(data=X, weight=W1, bias=b1, kernel=W1.shape[2:],num_filter=W1.shape[0])
h1_activation = nd.relu(h1_conv)
h1 = nd.Pooling(data=h1_activation, pool_type='max', kernel=(2,2),stride=(2,2))
# 2.2 卷积层二
h2_conv = nd.Convolution(data=h1, weight=W2, bias=b2,kernel=W2.shape[2:], num_filter=W2.shape[0])
h2_activation = nd.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation, pool_type='max', kernel=(2,2), stride=(2,2)) #Plooing的Kernel值，决定输出图像大小
# Flatten成2-dimens矩阵，以作为Dense层输入
h2 = nd.flatten(h2)
# 3. 定义全连接层
# 3.1 第一层全连接：激活函数非线性
h3_linear = nd.dot(h2, W3) + b3
h3 = nd.relu(h3_linear)
# 3.2 第二层全连接
h4 = nd.dot(h3_linear, W4) + b4
# 是否显示详细信息：各层代码块的输出shape
if verbose:
print('1st conv block: ', h1.shape)
print('2st conv block: ', h2.shape)
print('3st dense: ', h3.shape)
print('4st dense: ', h4.shape)
return h4
def net(self, X, debug=False):
########################
# Define the computation of the first convolutional layer
########################
h1_conv = nd.Convolution(data=X, weight=self.W1, bias=self.b1, kernel=(2, 3, 3), num_filter=20)
h1_activation = self.relu(h1_conv)
h1 = nd.Pooling(data=h1_activation, pool_type="avg", kernel=(2, 2, 2), stride=(2, 2, 2))
########################
# Define the computation of the second convolutional layer
########################
h2_conv = nd.Convolution(data=h1, weight=self.W2, bias=self.b2, kernel=(1, 5, 5), num_filter=50)
h2_activation = self.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation, pool_type="avg", kernel=(1, 2, 2), stride=(1, 2, 2))
########################
# Flattening h2 so that we can feed it into a fully-connected layer
########################
h2 = nd.flatten(h2)
########################
# Define the computation of the third (fully-connected) layer
########################
h3_linear = nd.dot(h2, self.W3) + self.b3
h3 = self.relu(h3_linear)
########################
# Define the computation of the output layer
########################
yhat_linear = nd.dot(h3, self.W4) + self.b4
#yhat = self.softmax(yhat_linear)
return yhat_linear
def lenet(X, params):
h1_conv = nd.Convolution(data=X,
weight=params[0],
bias=params[1],
kernel=(3, 3),
num_filter=20)
h1_activation = nd.reshape(h1_conv)
h1 = nd.Pooling(data=h1_activation,
pool_type='avg',
kernel=(2, 2),
stride=(2, 2))
h2_conv = nd.Convolution(data=h1,
weight=params[2],
bias=params[3],
kernel=(5, 5),
num_filter=50)
h2_activation = nd.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation,
pool_type='avg',
kernel=(2, 2),
stride=(2, 2))
h2 = nd.flatten(h2)
h3_linear = nd.dot(h2, params[4]) + params[5]
h3 = nd.relu(h3_linear)
y_hat = nd.dot(h3, params[6]) + params[7]
return y_hat
def net(X, verbose=False):
h1_conv = nd.Convolution(data=X,
weight=w1,
bias=b1,
kernel=w1.shape[2:],
num_filter=w1.shape[0])
h1_activation = nd.relu(h1_conv)
h1 = nd.Pooling(data=h1_activation,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
h2_conv = nd.Convolution(data=h1,
weight=w2,
bias=b2,
kernel=w2.shape[2:],
num_filter=w2.shape[0])
h2_activation = nd.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
h2 = nd.flatten(h2)
h3_linear = nd.dot(h2, w3) + b3
h3 = nd.relu(h3_linear)
h4_linear = nd.dot(h3, w4) + b4
return h4_linear
def net(X, params, is_training=True, debug=False):
W1, b1, gamma1, beta1, W2, b2, gamma2, beta2, W3, b3, gamma3, beta3, W4, b4 = params
h1_conv = nd.Convolution(data=X, weight=W1, bias=b1, kernel=(3, 3), num_filter=20)
h1_normed = batch_norm(h1_conv, gamma1, beta1, scope_name='bn1', is_training=is_training)
h1_activation = relu(h1_normed)
h1 = nd.Pooling(data=h1_activation, pool_type="avg", kernel=(2, 2), stride=(2, 2))
if debug:
print("h1 shape: %s" % (np.array(h1.shape)))
h2_conv = nd.Convolution(data=h1, weight=W2, bias=b2, kernel=(5, 5), num_filter=50)
h2_normed = batch_norm(h2_conv, gamma2, beta2, scope_name='bn2', is_training=is_training)
h2_activation = relu(h2_normed)
h2 = nd.Pooling(data=h2_activation, pool_type="avg", kernel=(2, 2), stride=(2, 2))
if debug:
print("h2 shape: %s" % (np.array(h2.shape)))
h2 = nd.flatten(h2)
if debug:
print("Flat h2 shape: %s" % (np.array(h2.shape)))
h3_linear = nd.dot(h2, W3) + b3
h3_normed = batch_norm(h3_linear, gamma3, beta3, scope_name="bn3", is_training=is_training)
h3 = relu(h3_normed)
yhat_linear = nd.dot(h3, W4) + b4
if debug:
print("yhat_linear shape: %s" % (np.array(yhat_linear.shape)))
return yhat_linear
def forward(self, x):
# Initial hidden state
# sentence representation할 때 hidden의 context가 cpu여서 오류 발생. context를 gpu로 전환
x, att = self.sen_rep(x)
x = nd.flatten(x)
res = self.classifier(x)
return res, att
def test_flatten():
a = create_2d_tensor(rows=LARGE_X, columns=SMALL_Y).reshape(
(LARGE_X // 2, 2, SMALL_Y))
b = nd.flatten(a)
assert b[-1][-1] == (LARGE_X - 1)
assert b[-1][0] == (LARGE_X - 2)
assert b.shape == (LARGE_X // 2, SMALL_Y * 2)
def net(X, debug=False):
########################
# Define the computation of the first convolutional layer
########################
h1_conv = nd.Convolution(data=X,
weight=W1,
bias=b1,
kernel=(3, 3),
num_filter=20)
h1_activation = relu(h1_conv)
h1 = nd.Pooling(data=h1_activation,
pool_type="avg",
kernel=(2, 2),
stride=(2, 2))
if debug:
print("h1 shape: %s" % (np.array(h1.shape)))
########################
# Define the computation of the second convolutional layer
########################
h2_conv = nd.Convolution(data=h1,
weight=W2,
bias=b2,
kernel=(5, 5),
num_filter=50)
h2_activation = relu(h2_conv)
h2 = nd.Pooling(data=h2_activation,
pool_type="avg",
kernel=(2, 2),
stride=(2, 2))
if debug:
print("h2 shape: %s" % (np.array(h2.shape)))
########################
# Flattening h2 so that we can feed it into a fully-connected layer
########################
h2 = nd.flatten(h2)
if debug:
print("Flat h2 shape: %s" % (np.array(h2.shape)))
########################
# Define the computation of the third (fully-connected) layer
########################
h3_linear = nd.dot(h2, W3) + b3
h3 = relu(h3_linear)
if debug:
print("h3 shape: %s" % (np.array(h3.shape)))
########################
# Define the computation of the output layer
########################
yhat_linear = nd.dot(h3, W4) + b4
if debug:
print("yhat_linear shape: %s" % (np.array(yhat_linear.shape)))
return yhat_linear
def forward(self, x):
x = nd.relu(self.bn1(self.conv1(x)))
x = nd.relu(self.bn2(self.conv2(x)))
x = nd.relu(self.bn3(self.conv3(x)))
x = nd.flatten(x).expand_dims(0)
#x, self.states = self.lstm(x, self.states)
x = self.dense1(x)
x = self.dense2(x)
probs = self.action_pred(x)
values = self.value_pred(x)
return mx.ndarray.softmax(probs), values
def forward(self, inputs):
# 将两个形状是(批量大小, 词数, 词向量维度)的嵌入层的输出按词向量连结
embeddings = nd.concat(
self.embedding(inputs), self.constant_embedding(inputs), dim=2)
# 根据Conv1D要求的输入格式，将词向量维，即一维卷积层的通道维，变换到前一维
embeddings = embeddings.transpose((0, 2, 1))
# 对于每个一维卷积层，在时序最大池化后会得到一个形状为(批量大小, 通道大小, 1)的
# NDArray。使用flatten函数去掉最后一维，然后在通道维上连结
encoding = nd.concat(*[nd.flatten(
self.pool(conv(embeddings))) for conv in self.convs], dim=1)
# 应用丢弃法后使用全连接层得到输出
outputs = self.decoder(self.dropout(encoding))
return outputs
def net(X, verbose=False):
X = X.as_in_context(W1.context)
# 第一层卷积
h1_conv = nd.Convolution(data=X,
weight=W1,
bias=b1,
kernel=W1.shape[2:],
num_filter=W1.shape[0])
h1_activation = nd.relu(h1_conv)
h1 = nd.Pooling(data=h1_activation,
pool_type='max',
kernel=(2, 2),
stride=(2, 2))
# 第二层卷积
h2_conv = nd.Convolution(data=h1,
weight=W2,
bias=b2,
kernel=W2.shape[2:],
num_filter=W2.shape[0])
h2_activation = nd.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation,
pool_type='max',
kernel=(2, 2),
stride=(2, 2))
h2 = nd.flatten(h2)
# 第一层全连接
h3_linear = nd.dot(h2, W3) + b3
h3 = nd.relu(h3_linear)
# 第二层全连接
h4_linear = nd.dot(h3, W4) + b4
if verbose:
print('h1_conv:', h1_conv.shape)
print('h1_activation:', h1_activation.shape)
print('1st conv block:', h1.shape)
print('-------------------\n')
print('h2_conv:', h2_conv.shape)
print('h2_activation:', h2_activation.shape)
print('2nd conv block:', h2.shape)
print('-------------------\n')
print('h3_linear:', h3_linear.shape)
print('1st dense:', h3.shape)
print('2nd dense:', h4_linear.shape)
print('output:', h4_linear)
return h4_linear
def forward(self, inputs):
# (batchsize, len(vocab), emb_size) contaced by emb_size
embeddings = nd.concat(self.embedding(inputs),
self.constant_embedding(inputs),
dim=2)
# Output:(batchsize, emb_size, len(vocab))
embeddings = embeddings.transpose((0, 2, 1))
# Output:(batchsize, outputchannels,1) and then connect in the second dimension
encoding = nd.concat(
*[nd.flatten(self.pool(conv(embeddings))) for conv in self.convs],
dim=1)
# Apply dropout and then fully connected layer
outputs = self.decoder(self.dropout(encoding))
return outputs
def forward(self, inputs):
# 将两个形状是（批量⼤⼩，词数，词向量维度）的嵌⼊层的输出按词向量连结。
embeddings = nd.concat(self.embedding(inputs),
self.constant_embedding(inputs),
dim=2)
# 根据 Conv1D 要求的输⼊格式，将词向量维，即⼀维卷积层的通道维，变换到前⼀维。
embeddings = embeddings.transpose((0, 2, 1))
# 对于每个⼀维卷积层，在时序最⼤池化后会得到⼀个形状为（批量⼤⼩，通道⼤⼩，1）的
# NDArray。使⽤ flatten 函数去掉最后⼀维，然后在通道维上连结。
encoding = nd.concat(
*[nd.flatten(self.pool(conv(embeddings))) for conv in self.convs],
dim=1)
# 应⽤丢弃法后使⽤全连接层得到输出。
outputs = self.decoder(self.dropout(encoding))
return outputs
def forward(self, inputs):
# 将两个形状是(批量大小, 词数, 词向量维度)的嵌入层的输出按词向量连结
# NxTx2D
embeddings = nd.concat(self.embedding(inputs), self.constant_embedding(inputs), dim=2)
# 根据Conv1D要求的输入格式，将词向量维，即一维卷积层的通道维，变换到前一维
# Nx2DxT T相当于图像中的W 2D相当于图像中的C NxCxW
embeddings = embeddings.transpose((0, 2, 1))
# 对于每个一维卷积层，在时序最大池化后会得到一个形状为(批量大小, 通道大小, 1)的
# NDArray。使用flatten函数去掉最后一维，然后在通道维上连结
# Nx2DxT Cxk -> NxCx(T-k+1) -> NxCx1 -> NxC -> [NxC, NxC, NxC] -> Nx3C
encoding = nd.concat(*[nd.flatten(self.pool(conv(embeddings))) for conv in self.convs], dim=1) # dim=1 是相对于concat而说的
# 应用丢弃法后使用全连接层得到输出
# Nx3C -> Nx2
outputs = self.decoder(self.dropout(encoding))
return outputs
def net(X, is_training=False, verbose=False):
X = X.as_in_context(W1.context)
# X = X.flatten().reshape((X.shape[0], X.shape[3], X.shape[1], X.shape[2]))
# 第一层卷积
h1_conv = nd.Convolution(data=X,
weight=W1,
bias=b1,
kernel=W1.shape[2:],
num_filter=W1.shape[0])
# 第一层归一化
h1_bn = batch_norm(h1_conv, gamma1, beta1, is_training, moving_mean1,
moving_variance1)
# 第一层激活函数
h1_activation = nd.relu(h1_bn)
# 第一层pooling
h1 = nd.Pooling(data=h1_activation,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
# 第二层卷积
h2_conv = nd.Convolution(data=h1,
weight=W2,
bias=b2,
kernel=W2.shape[2:],
num_filter=W2.shape[0])
h2_bn = batch_norm(h2_conv, gamma2, beta2, is_training, moving_mean2,
moving_variance2)
h2_activation = nd.relu(h2_bn)
h2 = nd.Pooling(data=h2_activation,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
h2 = nd.flatten(h2)
# 第一层全连接
h3_linear = nd.dot(h2, W3) + b3
h3 = nd.relu(h3_linear)
# 第二层全连接
h4_linear = nd.dot(h3, W4) + b4
if verbose:
print('1st conv block:', h1.shape)
print('2nd conv block:', h2.shape)
print('1st dense:', h3.shape)
print('2nd dense:', h4_linear.shape)
print('output:', h4_linear)
return h4_linear
def forward(self, inputs):
# 将两个形状 (批量大小，词数，词向量维度)的嵌入层的输出按词向量连结
embeddings = nd.concat(self.embedding(inputs),
self.constant_embedding(inputs),
dim=2)
# 根据 Conv1D 要求的输入格式，将词向量维，即：一维卷积层的通道维，变换到前一维
embeddings = embeddings.transpose((0, 2, 1))
# 对每一维卷积层，在时序最大池化后会得到一个形状为 (批量大小，通道大小，1) 的
# NDArray；使用 flatten 函数去掉最后一维，然后再通道维上进行连结
encoding = nd.concat(
*[nd.flatten(self.pool(conv(embeddings))) for conv in self.convs],
dim=1)
outputs = self.decoder(self.dropout(encoding))
return outputs
def net(self, X, debug=False):
########################
# Define the computation of the first convolutional layer
########################
h1_conv = nd.Convolution(data=X, weight=self.W1, bias=self.b1, kernel=(2, 8, 8), num_filter=16)
if debug: print("h1 shape: %s" % (np.array(h1_conv.shape)))
h1_activation = self.relu(h1_conv)
if debug: print("h1 shape: %s" % (np.array(h1_activation.shape)))
#h1 = nd.Pooling(data=h1_activation, pool_type="avg", kernel=(2, 2, 2), stride=(2, 2, 2))
h1 = h1_activation
if debug: print("h1 shape: %s" % (np.array(h1.shape)))
########################
# Define the computation of the second convolutional layer
########################
h2_conv = nd.Convolution(data=h1, weight=self.W2, bias=self.b2, kernel=(1, 4, 4), num_filter=32)
if debug: print("h2 shape: %s" % (np.array(h2_conv.shape)))
h2_activation = self.relu(h2_conv)
if debug: print("h2 shape: %s" % (np.array(h2_activation.shape)))
#h2 = nd.Pooling(data=h2_activation, pool_type="avg", kernel=(1, 2, 2), stride=(1, 2, 2))
h2 = h2_activation
if debug: print("h2 shape: %s" % (np.array(h2.shape)))
########################
# Flattening h2 so that we can feed it into a fully-connected layer
########################
h2 = nd.flatten(h2)
if debug: print("Flat h2 shape: %s" % (np.array(h2.shape)))
########################
# Define the computation of the third (fully-connected) layer
########################
h3_linear = nd.dot(h2, self.W3) + self.b3
h3 = self.relu(h3_linear)
if debug: print("h3 shape: %s" % (np.array(h3.shape)))
########################
# Define the computation of the output layer
########################
yhat_linear = nd.dot(h3, self.W4) + self.b4
#yhat = self.softmax(yhat_linear)
if debug: print("yhat_linear shape: %s" % (np.array(yhat_linear.shape)))
return yhat_linear
def net(X, verbose=False):
X = X.as_in_context(W1.context)
# first convolution layer
h1_conv = nd.Convolution(data=X,
weight=W1,
bias=b1,
kernel=W1.shape[2:],
num_filter=W1.shape[0])
h1_activation = nd.relu(h1_conv)
h1 = nd.Pooling(data=h1_activation,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
# print(h1.shape)
# second convolution layer
h2_conv = nd.Convolution(data=h1,
weight=W2,
bias=b2,
kernel=W2.shape[2:],
num_filter=W2.shape[0])
h2_activation = nd.relu(h2_conv)
h2 = nd.Pooling(data=h2_activation,
pool_type="max",
kernel=(2, 2),
stride=(2, 2))
# print(h2.shape)
h2 = nd.flatten(h2)
# first layer fully connected
h3_linear = nd.dot(h2, W3) + b3
h3 = nd.relu(h3_linear)
# second layer fully connected
h4_linear = nd.dot(h3, W4) + b4
if verbose:
print('1st conv block:', h1.shape)
print('2nd conv block:', h2.shape)
print('1st dense:', h3.shape)
print('2nd dense:', h4_linear.shape)
print('output:', h4_linear)
return h4_linear
def forward(self, inputs):
# Concatenate the output of two embedding layers with shape of
# (batch size, number of words, word vector dimension) by word vector
embeddings = nd.concat(
self.embedding(inputs), self.constant_embedding(inputs), dim=2)
# According to the input format required by Conv1D, the word vector
# dimension, that is, the channel dimension of the one-dimensional
# convolutional layer, is transformed into the previous dimension
embeddings = embeddings.transpose((0, 2, 1))
# For each one-dimensional convolutional layer, after max-over-time
# pooling, an NDArray with the shape of (batch size, channel size, 1)
# can be obtained. Use the flatten function to remove the last
# dimension and then concatenate on the channel dimension
encoding = nd.concat(*[nd.flatten(
self.pool(conv(embeddings))) for conv in self.convs], dim=1)
# After applying the dropout method, use a fully connected layer to
# obtain the output
outputs = self.decoder(self.dropout(encoding))
return outputs
def Net(X, is_training=False, verbose=False):
# 第一层卷积
h1_conv = nd.Convolution(X,
weight=W1,
bias=b1,
kernel=W1.shape[2:],
num_filter=c1)
h1_bn = batch_norm(h1_conv, gamma1, beta1, is_training, moving_mean1,
moving_variance1)
h1_activation = nd.Activation(h1_bn, act_type='relu')
h1 = nd.Pooling(h1_activation,
pool_type='max',
kernel=(2, 2),
stride=(2, 2))
# second Convolution
h2_conv = nd.Convolution(h1,
weight=W2,
bias=b2,
kernel=W2.shape[2:],
num_filter=c2)
h2_bn = batch_norm(h2_conv, gamma2, beta2, is_training, moving_mean2,
moving_variance2)
h2_activation = nd.relu(h2_bn)
h2 = nd.Pooling(h2_activation,
pool_type='max',
kernel=(2, 2),
stride=(2, 2))
h2 = nd.flatten(h2)
h3_linear = nd.dot(h2, W3) + b3
h3 = nd.relu(h3_linear)
h4_linear = nd.dot(h3, W4) + b4
if verbose:
print('1st conv block:', h1.shape)
print('2nd conv block:', h2.shape)
print('1st dense block:', h3.shape)
print('2nd dense block:', h4_linear.shape)
print('output:', h4_linear)
return h4_linear
def forward(self, inputs):
"""
inputs: (batch_size, max_length)
"""
embeddings = nd.concat(
self.embedding(inputs), self.constant_embedding(inputs),
dim=2) # return (batch_size, max_length, embed_size * 2)
embeddings = embeddings.transpose((0, 2, 1)) # Conv1D required
"""
for each convs, the output of conv(embeddings) should be: (batch_size, num_channels of that convs, max_length - kerner_size + 1)
for each convs, the output of maxpool(conv(embeddings)) should be: (batch_size, num_channels of that convs, 1)
for each convs, the output of nd.flatten(maxpool(conv(embeddings))) should be (batch_size, num_channels of that convs)
encoding: concate all the convs and return (batch_size, sum(all kernel_sizes))
"""
encoding = nd.concat(*[
nd.flatten(self.maxpool(conv(embeddings))) for conv in self.convs
],
dim=1)
# use the dropout to prevent overfit
outputs = self.decoder(self.dropout(encoding))
return outputs
def forward(self, inputs):
# 将 inputs 的形状由（批量大小，词数）变换为（词数，批量大小）。
inputs = inputs.T
# 根据 Conv1D 要求的输入形状，embeddings_static 和 embeddings_non_static
# 的形状由（词数，批量大小，词向量维度）变换为（批量大小，词向量维度，词数）。
embeddings_static = self.embedding_static(inputs).transpose((1, 2, 0))
embeddings_non_static = self.embedding_non_static(inputs).transpose(
(1, 2, 0))
# 将 embeddings_static 和 embeddings_non_static 按词向量维度连结。
embeddings = nd.concat(embeddings_static, embeddings_non_static, dim=1)
# 对于第 i 个卷积核，在时序最大池化后会得到一个形状为
# （批量大小，nums_channels[i]，1）的矩阵。使用 flatten 函数将它形状压成
# （批量大小，nums_channels[i]）。
encoding = [
nd.flatten(
self.get_pool(i)(self.get_bn(i)(self.get_conv(i)(embeddings))))
for i in range(len(self.ngram_kernel_sizes))
]
# 将批量按各通道的输出连结。encoding 的形状：
# （批量大小，nums_channels 各元素之和）。
encoding = nd.concat(*encoding, dim=1)
outputs = self.decoder(self.dropout(encoding))
return outputs
def net(X, verbose=False):
X = X.as_in_context(W1.context) #将X转到和W1一样的ctx上
#第一层卷积
h1_cov = nd.Convolution(data=X, weight=W1, bias=b1, kernel=W1.shape[2:], num_filter=W1.shape[0])
h1_activation = nd.relu(h1_cov)
h1 = h1.Pooling(data=h1_activation, pool_type="max", kernel=(2,2), stride=(2,2))
#第二层卷积
h2_cov = nd.Convolution(data=h1, weight=W2, bias=b2, kernel=W2.shape[2:], num_filter=W2.shape[0])
h2_activation = nd.relu(h2_cov)
h2 = nd.Pooling(data=h2_activation, pool_type="max", kernel=(2,2), stride=(2,2))
#拉平，变为全连接
h2 = nd.flatten(h2)
#第一层全连接
h3_linear = nd.dot(h2, W3) + b3
h3 = nd.relu(h3_linear)
#第四层全连接
h4_linear = nd.dot(h3, W4) + b4
#用于打印网络规模
if verbose:
print('1st conv block:', h1.shape)
print('2nd conv block:', h2.shape)
print('1st dense:', h3.shape)
print('2nd dense:', h4_linear.shape)
print('output:', h4_linear)
def flatten_prediction(pred):
return nd.flatten(nd.transpose(pred, axes=(0, 2, 3, 1)))
def _flatten_prediction(self, pred):
'''
Helper function to flatten the predicted bounding boxes and categories
'''
return nd.flatten(nd.transpose(pred, axes=(0, 2, 3, 1)))
# this make mxnet faster then numpy. mxnet uses a optimized library: Intel MKL for CPU operation.
# with gpu its 30x faster then numpy
# convert mxnet ndarray to numpy ndarray
# this operation is synchronous (= further computation is blocked until function is finished.)
matrix = matrix.asnumpy()
print(matrix)
# convert numpy to mxnet ndarray
matrix = nd.array(matrix)
print(matrix)
# calculate the average intensity for every pixel
matrix = nd.random.randint(low=0, high=255, shape=(1, 3, 3))
print(matrix)
flat = nd.flatten(matrix)
flat = flat[0]
print(flat.shape)
a = np.array([[0, 1, 2, 3], [4, 5, 6, 7]])
print(a.shape)
for row in a:
for col in row:
print(col)
a = np.arange(8).reshape(2, 4)
print(a.shape)
# https://mxnet.apache.org/api/python/docs/tutorials/packages/gluon/training/normalization/index.html#Data-Normalization
print('''
