实施卷积序列到序列网络的需求指南

时间:2019-06-18 11:49:12

标签: computer-vision conv-neural-network recurrent-neural-network seq2seq encoder-decoder

我正在尝试构建卷积序列到序列网络,该网络将输入作为卫星图像并预测下一个图像序列!这样我们可以预测天气。我已经实现了模型,但是我需要更多关于模型的指南。我对模型有疑问!

Research Paper Link

Model Image

我尝试在pytorch中建立模型。这是我的代码-

from convlstm2d import ConvLSTMCell
from torch.autograd import Variable
import torch
from torch.nn import Conv2d, ConvTranspose2d
from torch.nn.modules import UpsamplingNearest2d, Conv2d, Upsample, 
import torch.nn.functional as F

input = Variable(torch.rand( 6, 3, 256, 256))
decoder_input = Variable(torch.rand( 6, 80, 8, 8))


conv = Conv2d(3, 16, kernel_size=(3, 3), stride=2, padding=1)(input)

conv1 = Conv2d(16, 32, kernel_size=(3, 3), stride=2, padding=1)(F.relu(conv))
conv2 = Conv2d(32, 48, kernel_size=(3, 3), stride=2, padding=1)(F.relu(conv1))

conv3 = Conv2d(48, 64, kernel_size=(3, 3), stride=2, padding=1)(F.relu(conv2))
conv4 = Conv2d(64, 80, kernel_size=(3, 3), stride=2, padding=1)(F.relu(conv3))

Encoder, prev_c,= ConvLSTMCell(80, 80)(F.relu(conv4), None)
states = (Encoder, prev_c)

decoder_cell = ConvLSTMCell(80, 80)

decoder, _ = decoder_cell(decoder_input, states)

layer = ConvTranspose2d(80, 80, kernel_size=(1, 1))(F.relu(decoder))
m = torch.cat([F.relu(layer), conv4], dim=1)

layer1 = ConvTranspose2d(160, 64, kernel_size=(2, 2), stride=2)(m)
m1 = torch.cat([F.relu(layer1), conv3], dim=1)

layer2 = ConvTranspose2d(128, 48, kernel_size=(2, 2), stride=2)(m1)
m2 = torch.cat([F.relu(layer2), conv2], dim=1)

layer3 = ConvTranspose2d(96, 32, kernel_size=(2, 2), stride=2)(m2)
m3 = torch.cat([F.relu(layer3), conv1], dim=1)

layer4 = ConvTranspose2d(64, 16, kernel_size=(2, 2), stride=2)(m3)
m4 = torch.cat([F.relu(layer4), conv], dim=1)

output = ConvTranspose2d(32, 3, kernel_size=(2, 2), stride=2)(m4)



import torch
from torch import nn
import torch.nn.functional as f
from torch.autograd import Variable


# Define some constants
KERNEL_SIZE = 3
PADDING = KERNEL_SIZE // 2


class ConvLSTMCell(nn.Module):
    """
    Generate a convolutional LSTM cell
    """
    def __init__(self, input_size, hidden_size):
        super().__init__()
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.Gates = nn.Conv2d(input_size + hidden_size, 4 * hidden_size, KERNEL_SIZE, padding=PADDING)

    def forward(self, input_, prev_state):

        # get batch and spatial sizes
        batch_size = input_.data.size()[0]
        spatial_size = input_.data.size()[2:]

        # generate empty prev_state, if None is provided
        if prev_state is None:
            state_size = [batch_size, self.hidden_size] + list(spatial_size)
            print(state_size)
            prev_state = (
                Variable(torch.zeros(state_size)),
                Variable(torch.zeros(state_size))
            )


        prev_hidden, prev_cell = prev_state

        # data size is [batch, channel, height, width]
        stacked_inputs = torch.cat((input_, prev_hidden), 1)
        gates = self.Gates(stacked_inputs)

        # chunk across channel dimension
        in_gate, remember_gate, out_gate, cell_gate = gates.chunk(4, 1)

        # apply sigmoid non linearity
        in_gate = f.sigmoid(in_gate)
        remember_gate = f.sigmoid(remember_gate)
        out_gate = f.sigmoid(out_gate)

        # apply tanh non linearity
        cell_gate = f.tanh(cell_gate)

        # compute current cell and hidden state
        cell = (remember_gate * prev_cell) + (in_gate * cell_gate)
        hidden = out_gate * f.tanh(cell)

        return hidden, cell
  • 什么是编码器模型卷积或ConvLSTMCell的第一层?
  • 解码器模型卷积或ConvLSTMCell的第一层是什么?
  • 解码器输入层的暗淡是什么? 如果我做错了,请指导我!

0 个答案:

没有答案