1、输入通过Inception模块的4个分支分别计算，得到的输出宽和高相同(因为使用了padding)，而通道不同。 2、将4个分支的通道进行简单的合并，即得到Inception模块的输出。 3、每次卷积之后都使用批正则化`BatchNorm2d`，并使用relu函数进行激

VGG model

VGG model

Computes and stores the average and current value

Keeps track of most recent, average, sum, and count of a metric.

Implement label smoothing.

Implement label smoothing.

Construct a layernorm module (See citation for details).

Construct a layernorm module (See citation for details).

Construct a layernorm module (See citation for details).

(conv => BN => ReLU) * 2

Generic N layer decoder with masking.

Generic N layer decoder with masking.

Decoder is made of self-attn, src-attn, and feed forward (defined below)

Decoder is made of self-attn, src-attn, and feed forward (defined below)

Encoder The encoder is composed of a stack of N=6 identical layers.

Encoder The encoder is composed of a stack of N=6 identical layers.

Encoder is made up of self-attn and feed forward (defined below)

Encoder is made up of self-attn and feed forward (defined below)

Define standard linear + softmax generation step.

Define standard linear + softmax generation step.

Optim wrapper that implements rate.

A PyTorch Dataset class to be used in a PyTorch DataLoader to create batches.

Implement the PE function.

Implement the PE function.

Implements FFN equation.

Implements FFN equation.

A simple loss compute and train function.

A simple loss compute and train function.

A simple loss compute and train function.

A residual connection followed by a layer norm. Note for code simplicity the norm is first as opposed to last.

A residual connection followed by a layer norm. Note for code simplicity the norm is first as opposed to last.

实现子层连接结构的类

Object for holding a batch of data with mask during training.

Generic N layer decoder with masking.

Decoder is made of self-attn, src-attn, and feed forward (defined below)

Encoder The encoder is composed of a stack of N=6 identical layers.

A standard Encoder-Decoder architecture. Base for this and many other models.

A standard Encoder-Decoder architecture. Base for this and many other models.

A standard Encoder-Decoder architecture. Base for this and many other models.

EncoderLayer is made up of two sublayer: self-attn and feed forward

Define standard linear + softmax generation step.

Implement label smoothing.

The loss function for object detection. 对于Loss的计算，完全遵循SSD的定义，即 MultiBox Loss This is a combination of: (1) a localization loss for t

A standard Encoder-Decoder architecture. Base for this and many other models.

A standard Encoder-Decoder architecture. Base for this and many other models.

Implement the PE function.

Implements FFN equation.

Convolutions to predict class scores and bounding boxes using feature maps. The bounding boxes (locations) are predicted as encoded offsets

VGG base convolutions to produce feature maps. 完全采用vgg16的结构作为特征提取模块，丢掉fc6和fc7两个全连接层。 因为vgg16的ImageNet预训练模型是使用224×224尺寸训练的，因此我们的网络输入也固定为22

The tiny_detector network 包含一个VGG作为特征提取模块，并在最后一个特征图上添加一个输出头来预测目标框信息