PyTorch 中使用深度学习（CNN和LSTM）的自动图像描述

git clone https://github.com/pdollar/coco.gitcd coco/PythonAPI/
make
python setup.py build
python setup.py installcd ../../

git clone https://github.com/yunjey/pytorch-tutorial.gitcd pytorch-tutorial/tutorials/03-advanced/image_captioning/

pip install -r requirements.txt

chmod +x download.sh
./download.sh

python train.py --num_epochs 10 --learning_rate 0.01

import torch
import torch.nn as nn
import torchvision.models as models
from torch.nn.utils.rnn import pack_padded_sequence
from torch.autograd import Variable

class EncoderCNN(nn.Module):
    def __init__(self, embed_size):
        """Load the pretrained ResNet-152 and replace top fc layer."""
        super(EncoderCNN, self).__init__()
        resnet = models.resnet152(pretrained=True)
        modules = list(resnet.children())[:-1]      # delete the last fc layer.
        self.resnet = nn.Sequential(*modules)
        self.linear = nn.Linear(resnet.fc.in_features, embed_size)
        self.bn = nn.BatchNorm1d(embed_size, momentum=0.01)
        self.init_weights()
        
    def init_weights(self):
        """Initialize the weights."""
        self.linear.weight.data.normal_(0.0, 0.02)
        self.linear.bias.data.fill_(0)
        
    def forward(self, images):
        """Extract the image feature vectors."""
        features = self.resnet(images)
        features = Variable(features.data)
        features = features.view(features.size(0), -1)
        features = self.bn(self.linear(features))
        return features
    
    
class DecoderRNN(nn.Module):
    def __init__(self, embed_size, hidden_size, vocab_size, num_layers):
        """Set the hyper-parameters and build the layers."""
        super(DecoderRNN, self).__init__()
        self.embed = nn.Embedding(vocab_size, embed_size)
        self.lstm = nn.LSTM(embed_size, hidden_size, num_layers, batch_first=True)
        self.linear = nn.Linear(hidden_size, vocab_size)
        self.init_weights()
    
    def init_weights(self):
        """Initialize weights."""
        self.embed.weight.data.uniform_(-0.1, 0.1)
        self.linear.weight.data.uniform_(-0.1, 0.1)
        self.linear.bias.data.fill_(0)
        
    def forward(self, features, captions, lengths):
        """Decode image feature vectors and generates captions."""
        embeddings = self.embed(captions)
        embeddings = torch.cat((features.unsqueeze(1), embeddings), 1)
        packed = pack_padded_sequence(embeddings, lengths, batch_first=True) 
        hiddens, _ = self.lstm(packed)
        outputs = self.linear(hiddens[0])
        return outputs
    
    def sample(self, features, states=None):
        """Samples captions for given image features (Greedy search)."""
        sampled_ids = []
        inputs = features.unsqueeze(1)
        for i in range(20):                                    # maximum sampling length
            hiddens, states = self.lstm(inputs, states)        # (batch_size, 1, hidden_size), 
            outputs = self.linear(hiddens.squeeze(1))          # (batch_size, vocab_size)
            predicted = outputs.max(1)[1]
            sampled_ids.append(predicted)
            inputs = self.embed(predicted)
            inputs = inputs.unsqueeze(1)                       # (batch_size, 1, embed_size)
        sampled_ids = torch.cat(sampled_ids, 1)                # (batch_size, 20)
        return sampled_ids.squeeze()

python sample.py --image='png/example.png'