如何主动结束python程序
目前我有两个文件分别为a.py和b.py然后a.py在运行时需要import b执行一些功能。然后a.py是始终保持运行的，而b.py在运行执行完功能后并没有结束，也是保持运行。
也就导致b.py只能运行一次，没法运行第二次。除非结束a.py。我试过在b.py里加上sys.exit(0)。但b.py结束后会把a.py也结束。
所以有没有办法可以使a.py始终保持运行的情况下，b.py运行完一次后就结束程序，然后可以再运行第二次，第三次...即多次运行 能把b.py的具体代码贴一下么 子函数应该会有return的，你的b里到底做了啥
把代码发全。 我觉得你该重新设计你的代码逻辑了，看你发了好几个帖子感觉你的程序为了实现一个简单的逻辑绕了一大圈路 你发一下代码吧，我给你改一改，我觉得你的代码逻辑好像有问题 你想在一个程序中执行另一个程序的话最好用 exec，而不应该用 import
eg:
with open('b.py') as fd:
code = fd.read()
exec(code)
当然，你想用 import 也不是不行，你可以成功 import b 就已经说明 b 已经执行结束了，
你再次 import b 没有效果不是因为 b 没有结束，而是因为 python 的导入机制，
python 在导入一个模块之后，会将模块名和模块对象的对应关系放入 sys.modules 中，
当你导入这个模块的时候会先看 sys.modules 中有没有这个模块，如果有的话就不会再导入，
所以你想能多次导入同一个模块的话得先从 sys.modules 中删掉这个模块，
eg:
import sys
import b
del sys.modules['b']
import b 我就好奇b里到底干了啥，这种运行方式也是不按套路出牌 tommyyu 发表于 2022-12-23 09:23
能把b.py的具体代码贴一下么
import torch
#import cv2
import torch.nn.functional as F
import numpy as np
import json
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import matplotlib.cm as cm
import skimage.transform
import argparse
from PIL import Image
torch.cuda.set_device(-1)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
def caption_image_beam_search(encoder, decoder, image_path, word_map, beam_size=3):
"""
Reads an image and captions it with beam search.
:param encoder: encoder model
:param decoder: decoder model
:param image_path: path to image
:param word_map: word map
:param beam_size: number of sequences to consider at each decode-step
:return: caption, weights for visualization
"""
k = beam_size
vocab_size = len(word_map)
# Read image and process
img = Image.open(image_path)
#result = np.zeros(img.shape, dtype=np.float32)
#result = img / 255.0
#img = np.uint8(result * 255.0)
img = np.array(img)
width, high, channel = img.shape
width_new, high_new = (256, 256)
img = img
#img = cv2.imread(image_path)
#img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
# 当为单通道图像时，转化为三通道
if len(img.shape) == 2:
img = img[:, :, np.newaxis]# 增加纬度
img = np.concatenate(, axis=2)# 拼接为三通道
#img = img.resize(img, (256, 256), Image.ANTIALIAS)
img = np.array(img)
img = img.transpose(2, 0, 1)# 矩阵转置 通道数放在前面
img = img / 255.
img = torch.FloatTensor(img).to(device)
normalize = transforms.Normalize(mean=,
std=)
transform = transforms.Compose()
image = transform(img)# (3, 256, 256)
# Encode
image = image.unsqueeze(0)# (1, 3, 256, 256)
encoder_out = encoder(image)# (1, enc_image_size, enc_image_size, encoder_dim) 1,14,14,2048
enc_image_size = encoder_out.size(1)
print('enc_image_size:', enc_image_size)
encoder_dim = encoder_out.size(3)
print('encoder_dim:', encoder_dim)
# Flatten encoding
encoder_out = encoder_out.view(1, -1, encoder_dim)# (1, num_pixels, encoder_dim) 1,196,2048
# 表示了图像的196个区域各自的特征
# print('encoder_out:',encoder_out)
num_pixels = encoder_out.size(1)# 第二位 196
# print('num_pixels:',num_pixels)
# We'll treat the problem as having a batch size of k
# print(encoder_out.size())
encoder_out = encoder_out.expand(k, num_pixels, encoder_dim)# (k, num_pixels, encoder_dim)1->k纬度扩展，五份特征
# print(encoder_out.size())
# Tensor to store top k previous words at each step; now they're just <start>
k_prev_words = torch.LongTensor([]] * k).to(device)# (k, 1)
# print('k_prev_words:',k_prev_words)
# Tensor to store top k sequences; now they're just <start>
seqs = k_prev_words# (k, 1)
# Tensor to store top k sequences' scores; now they're just 0
top_k_scores = torch.zeros(k, 1).to(device)# (k, 1)
# Tensor to store top k sequences' alphas; now they're just 1s 这里其实就是存储每个字对应图像上的关注区域，映射在14*14的张量上面
seqs_alpha = torch.ones(k, 1, enc_image_size, enc_image_size).to(device)# (k, 1, enc_image_size, enc_image_size)
# Lists to store completed sequences, their alphas and scores
complete_seqs = list()
complete_seqs_alpha = list()
complete_seqs_scores = list()
# Start decoding
step = 1
h, c = decoder.init_hidden_state(encoder_out)# h0
print('h, c', h.size(), c.size())
# s is a number less than or equal to k, because sequences are removed from this process once they hit <end>
while True:
embeddings = decoder.embedding(k_prev_words).squeeze(1)# (s, embed_dim) （5,隐层512）
print('embeddings', embeddings.size())
# encode的图片表示 和隐状态
awe, alpha = decoder.attention(encoder_out,
h)# (s, encoder_dim), (s, num_pixels)（5，2048（）,5，196（attention 存储字对应图像各部分的权重））
print(' awe, alpha', awe.size(), alpha.size())
# 0/0
alpha = alpha.view(-1, enc_image_size, enc_image_size)# (s, enc_image_size, enc_image_size)（5,14,14）
gate = decoder.sigmoid(decoder.f_beta(h))# gating scalar, (s, encoder_dim)
awe = gate * awe# 给特征赋予权重
h, c = decoder.decode_step(torch.cat(, dim=1),
(h, c))# (s, decoder_dim)输入(512,2048),(512,512)带权重的特征和上一次的lstm输出和细胞状态值
scores = decoder.fc(h)# (s, vocab_size)
scores = F.log_softmax(scores, dim=1)
print('scores', scores.size())
# Add 每一句 含有多少词 更新
scores = top_k_scores.expand_as(scores) + scores# (s, vocab_size)
print('top_k_scores，scores', top_k_scores.size(), scores.size())
# For the first step, all k points will have the same scores (since same k previous words, h, c)
if step == 1:
top_k_scores, top_k_words = scores.topk(k, 0, True, True)# (s)
else:
# Unroll and find top scores, and their unrolled indices
top_k_scores, top_k_words = scores.view(-1).topk(k, 0, True, True)# (s) 取词，top
print('top_k_scores，top_k_words', top_k_scores.size(), top_k_words.size())
# Convert unrolled indices to actual indices of scores
prev_word_inds = torch.floor_divide(top_k_words, vocab_size)
# prev_word_inds = top_k_words / vocab_size# (s)
next_word_inds = top_k_words % vocab_size# (s)
print('top_k_scores,top_k_words，prev_word_inds，next_word_inds', top_k_words, top_k_scores, prev_word_inds,
next_word_inds)
# Add new words to sequences, alphas
seqs = torch.cat(, next_word_inds.unsqueeze(1)], dim=1)# (s, step+1)#词加一
seqs_alpha = torch.cat(, alpha.unsqueeze(1)],# 词对应图像区域加一
dim=1)# (s, step+1, enc_image_size, enc_image_size)
# Which sequences are incomplete (didn't reach <end>)? 挑出这次循环完结的 句子
incomplete_inds = [ind for ind, next_word in enumerate(next_word_inds) if
next_word != word_map['<end>']]
complete_inds = list(set(range(len(next_word_inds))) - set(incomplete_inds))
# Set aside complete sequences 挑出完整序列
if len(complete_inds) > 0:
complete_seqs.extend(seqs.tolist())# 追加全部序列
complete_seqs_alpha.extend(seqs_alpha.tolist())
complete_seqs_scores.extend(top_k_scores)
k -= len(complete_inds)# reduce beam length accordingly
# Proceed with incomplete sequences
if k == 0:
break
# 更新参数 只保留未完全序列参数
seqs = seqs
seqs_alpha = seqs_alpha
h = h]
c = c]
encoder_out = encoder_out]
top_k_scores = top_k_scores.unsqueeze(1)
k_prev_words = next_word_inds.unsqueeze(1)
# Break if things have been going on too long
if step > 50:
break
step += 1
# 标记 scores分数最高序列作为返回值。
i = complete_seqs_scores.index(max(complete_seqs_scores))
seq = complete_seqs
alphas = complete_seqs_alpha
return seq, alphas
def visualize_att(image_path, seq, alphas, rev_word_map, smooth=True):
"""
Visualizes caption with weights at every word.
Adapted from paper authors' repo: https://github.com/kelvinxu/arctic-captions/blob/master/alpha_visualization.ipynb
:param image_path: path to image that has been captioned
:param seq: caption
:param alphas: weights
:param rev_word_map: reverse word mapping, i.e. ix2word
:param smooth: smooth weights?
"""
image = Image.open(image_path)
image = image.resize(, Image.LANCZOS)
words = for ind in seq]
#print(words)
for t in range(1,len(words)-1):
if t > 50:
break
plt.subplot(int(np.ceil(len(words)) / 5.), 6, t)
plt.text(0, 1, '%s' % (words), color='black', backgroundcolor='white', fontsize=12)
plt.imshow(image)
current_alpha = alphas
if smooth:
alpha = skimage.transform.pyramid_expand(current_alpha.numpy(), upscale=12, sigma=8)
else:
alpha = skimage.transform.resize(current_alpha.numpy(), )
if t == 0:
plt.imshow(alpha, alpha=0)
else:
plt.imshow(alpha, alpha=0.8)
plt.set_cmap(cm.Greys_r)
plt.axis('off')
#plt.show()
import scipy
print(scipy.__version__)
#checkpoint = torch.load('./BEST_checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar', map_location=str(device))
checkpoint = torch.load(r'C:/Users/Ternence/PycharmProjects/pythonProject/tuxiang/BEST_checkpoint_flickr8k_5_cap_per_img_5_min_word_freq.pth.tar', map_location=str(device))
decoder = checkpoint['decoder']
decoder = decoder.to(device)
decoder.eval()
encoder = checkpoint['encoder']
encoder = encoder.to(device)
encoder.eval()
# Load word map (word2ix)
with open(r'C:/Users/Ternence/kind2/Flickr8k/data/WORDMAP_flickr8k_5_cap_per_img_5_min_word_freq.json', 'r') as j:
word_map = json.load(j,)
rev_word_map = {v: k for k, v in word_map.items()}# ix2word
# Encode, decode with attention and beam search
seq, alphas = caption_image_beam_search(encoder, decoder, r'C:/Users/Ternence/kind2/Flickr8k/tupian/ren2.jpg', word_map, 5)
alphas = torch.FloatTensor(alphas)
# Visualize caption and attention of best sequence
visualize_att(r'C:/Users/Ternence/kind2/Flickr8k/tupian/ren2.jpg', seq, alphas, rev_word_map, True)
words = for ind in seq]
print(words)
这个是图像字幕的可视化功能 编程追风梦 发表于 2022-12-23 10:02
你发一下代码吧，我给你改一改，我觉得你的代码逻辑好像有问题
import torch
#import cv2
import torch.nn.functional as F
import numpy as np
import json
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import matplotlib.cm as cm
import skimage.transform
import argparse
from PIL import Image
torch.cuda.set_device(-1)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
def caption_image_beam_search(encoder, decoder, image_path, word_map, beam_size=3):
"""
Reads an image and captions it with beam search.
:param encoder: encoder model
:param decoder: decoder model
:param image_path: path to image
:param word_map: word map
:param beam_size: number of sequences to consider at each decode-step
:return: caption, weights for visualization
"""
k = beam_size
vocab_size = len(word_map)
# Read image and process
img = Image.open(image_path)
#result = np.zeros(img.shape, dtype=np.float32)
#result = img / 255.0
#img = np.uint8(result * 255.0)
img = np.array(img)
width, high, channel = img.shape
width_new, high_new = (256, 256)
img = img
#img = cv2.imread(image_path)
#img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
# 当为单通道图像时，转化为三通道
if len(img.shape) == 2:
img = img[:, :, np.newaxis]# 增加纬度
img = np.concatenate(, axis=2)# 拼接为三通道
#img = img.resize(img, (256, 256), Image.ANTIALIAS)
img = np.array(img)
img = img.transpose(2, 0, 1)# 矩阵转置 通道数放在前面
img = img / 255.
img = torch.FloatTensor(img).to(device)
normalize = transforms.Normalize(mean=,
std=)
transform = transforms.Compose()
image = transform(img)# (3, 256, 256)
# Encode
image = image.unsqueeze(0)# (1, 3, 256, 256)
encoder_out = encoder(image)# (1, enc_image_size, enc_image_size, encoder_dim) 1,14,14,2048
enc_image_size = encoder_out.size(1)
print('enc_image_size:', enc_image_size)
encoder_dim = encoder_out.size(3)
print('encoder_dim:', encoder_dim)
# Flatten encoding
encoder_out = encoder_out.view(1, -1, encoder_dim)# (1, num_pixels, encoder_dim) 1,196,2048
# 表示了图像的196个区域各自的特征
# print('encoder_out:',encoder_out)
num_pixels = encoder_out.size(1)# 第二位 196
# print('num_pixels:',num_pixels)
# We'll treat the problem as having a batch size of k
# print(encoder_out.size())
encoder_out = encoder_out.expand(k, num_pixels, encoder_dim)# (k, num_pixels, encoder_dim)1->k纬度扩展，五份特征
# print(encoder_out.size())
# Tensor to store top k previous words at each step; now they're just <start>
k_prev_words = torch.LongTensor([]] * k).to(device)# (k, 1)
# print('k_prev_words:',k_prev_words)
# Tensor to store top k sequences; now they're just <start>
seqs = k_prev_words# (k, 1)
# Tensor to store top k sequences' scores; now they're just 0
top_k_scores = torch.zeros(k, 1).to(device)# (k, 1)
# Tensor to store top k sequences' alphas; now they're just 1s 这里其实就是存储每个字对应图像上的关注区域，映射在14*14的张量上面
seqs_alpha = torch.ones(k, 1, enc_image_size, enc_image_size).to(device)# (k, 1, enc_image_size, enc_image_size)
# Lists to store completed sequences, their alphas and scores
complete_seqs = list()
complete_seqs_alpha = list()
complete_seqs_scores = list()
# Start decoding
step = 1
h, c = decoder.init_hidden_state(encoder_out)# h0
print('h, c', h.size(), c.size())
# s is a number less than or equal to k, because sequences are removed from this process once they hit <end>
while True:
embeddings = decoder.embedding(k_prev_words).squeeze(1)# (s, embed_dim) （5,隐层512）
print('embeddings', embeddings.size())
# encode的图片表示 和隐状态
awe, alpha = decoder.attention(encoder_out,
h)# (s, encoder_dim), (s, num_pixels)（5，2048（）,5，196（attention 存储字对应图像各部分的权重））
print(' awe, alpha', awe.size(), alpha.size())
# 0/0
alpha = alpha.view(-1, enc_image_size, enc_image_size)# (s, enc_image_size, enc_image_size)（5,14,14）
gate = decoder.sigmoid(decoder.f_beta(h))# gating scalar, (s, encoder_dim)
awe = gate * awe# 给特征赋予权重
h, c = decoder.decode_step(torch.cat(, dim=1),
(h, c))# (s, decoder_dim)输入(512,2048),(512,512)带权重的特征和上一次的lstm输出和细胞状态值
scores = decoder.fc(h)# (s, vocab_size)
scores = F.log_softmax(scores, dim=1)
print('scores', scores.size())
# Add 每一句 含有多少词 更新
scores = top_k_scores.expand_as(scores) + scores# (s, vocab_size)
print('top_k_scores，scores', top_k_scores.size(), scores.size())
# For the first step, all k points will have the same scores (since same k previous words, h, c)
if step == 1:
top_k_scores, top_k_words = scores.topk(k, 0, True, True)# (s)
else:
# Unroll and find top scores, and their unrolled indices
top_k_scores, top_k_words = scores.view(-1).topk(k, 0, True, True)# (s) 取词，top
print('top_k_scores，top_k_words', top_k_scores.size(), top_k_words.size())
# Convert unrolled indices to actual indices of scores
prev_word_inds = torch.floor_divide(top_k_words, vocab_size)
# prev_word_inds = top_k_words / vocab_size# (s)
next_word_inds = top_k_words % vocab_size# (s)
print('top_k_scores,top_k_words，prev_word_inds，next_word_inds', top_k_words, top_k_scores, prev_word_inds,
next_word_inds)
# Add new words to sequences, alphas
seqs = torch.cat(, next_word_inds.unsqueeze(1)], dim=1)# (s, step+1)#词加一
seqs_alpha = torch.cat(, alpha.unsqueeze(1)],# 词对应图像区域加一
dim=1)# (s, step+1, enc_image_size, enc_image_size)
# Which sequences are incomplete (didn't reach <end>)? 挑出这次循环完结的 句子
incomplete_inds = [ind for ind, next_word in enumerate(next_word_inds) if
next_word != word_map['<end>']]
complete_inds = list(set(range(len(next_word_inds))) - set(incomplete_inds))
# Set aside complete sequences 挑出完整序列
if len(complete_inds) > 0:
complete_seqs.extend(seqs.tolist())# 追加全部序列
complete_seqs_alpha.extend(seqs_alpha.tolist())
complete_seqs_scores.extend(top_k_scores)
k -= len(complete_inds)# reduce beam length accordingly
# Proceed with incomplete sequences
if k == 0:
break
# 更新参数 只保留未完全序列参数
seqs = seqs
seqs_alpha = seqs_alpha
h = h]
c = c]
encoder_out = encoder_out]
top_k_scores = top_k_scores.unsqueeze(1)
k_prev_words = next_word_inds.unsqueeze(1)
# Break if things have been going on too long
if step > 50:
break
step += 1
# 标记 scores分数最高序列作为返回值。
i = complete_seqs_scores.index(max(complete_seqs_scores))
seq = complete_seqs
alphas = complete_seqs_alpha
return seq, alphas
def visualize_att(image_path, seq, alphas, rev_word_map, smooth=True):
"""
Visualizes caption with weights at every word.
Adapted from paper authors' repo: https://github.com/kelvinxu/arctic-captions/blob/master/alpha_visualization.ipynb
:param image_path: path to image that has been captioned
:param seq: caption
:param alphas: weights
:param rev_word_map: reverse word mapping, i.e. ix2word
:param smooth: smooth weights?
"""
image = Image.open(image_path)
image = image.resize(, Image.LANCZOS)
words = for ind in seq]
#print(words)
for t in range(1,len(words)-1):
if t > 50:
break
plt.subplot(int(np.ceil(len(words)) / 5.), 6, t)
plt.text(0, 1, '%s' % (words), color='black', backgroundcolor='white', fontsize=12)
plt.imshow(image)
current_alpha = alphas
if smooth:
alpha = skimage.transform.pyramid_expand(current_alpha.numpy(), upscale=12, sigma=8)
else:
alpha = skimage.transform.resize(current_alpha.numpy(), )
if t == 0:
plt.imshow(alpha, alpha=0)
else:
plt.imshow(alpha, alpha=0.8)
plt.set_cmap(cm.Greys_r)
plt.axis('off')
#plt.show()
import scipy
print(scipy.__version__)
#checkpoint = torch.load('./BEST_checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar', map_location=str(device))
checkpoint = torch.load(r'C:/Users/Ternence/PycharmProjects/pythonProject/tuxiang/BEST_checkpoint_flickr8k_5_cap_per_img_5_min_word_freq.pth.tar', map_location=str(device))
decoder = checkpoint['decoder']
decoder = decoder.to(device)
decoder.eval()
encoder = checkpoint['encoder']
encoder = encoder.to(device)
encoder.eval()
# Load word map (word2ix)
with open(r'C:/Users/Ternence/kind2/Flickr8k/data/WORDMAP_flickr8k_5_cap_per_img_5_min_word_freq.json', 'r') as j:
word_map = json.load(j,)
rev_word_map = {v: k for k, v in word_map.items()}# ix2word
# Encode, decode with attention and beam search
seq, alphas = caption_image_beam_search(encoder, decoder, r'C:/Users/Ternence/kind2/Flickr8k/tupian/ren2.jpg', word_map, 5)
alphas = torch.FloatTensor(alphas)
# Visualize caption and attention of best sequence
visualize_att(r'C:/Users/Ternence/kind2/Flickr8k/tupian/ren2.jpg', seq, alphas, rev_word_map, True)
words = for ind in seq]
print(words)
这个是图像字幕的可视化功能，就是b.py 莫凡辰 发表于 2022-12-23 10:49
import torch
#import cv2
import torch.nn.functional as F
我感觉这两个函数都可以定义到 a.py 里面，把下面的那一部分也可以封装成一个函数，全都放到 a.py 里面 suchocolate 发表于 2022-12-23 09:26
子函数应该会有return的，你的b里到底做了啥
把代码发全。
import torch
#import cv2
import torch.nn.functional as F
import numpy as np
import json
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import matplotlib.cm as cm
import skimage.transform
import argparse
from PIL import Image
torch.cuda.set_device(-1)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)
def caption_image_beam_search(encoder, decoder, image_path, word_map, beam_size=3):
"""
Reads an image and captions it with beam search.
:param encoder: encoder model
:param decoder: decoder model
:param image_path: path to image
:param word_map: word map
:param beam_size: number of sequences to consider at each decode-step
:return: caption, weights for visualization
"""
k = beam_size
vocab_size = len(word_map)
# Read image and process
img = Image.open(image_path)
#result = np.zeros(img.shape, dtype=np.float32)
#result = img / 255.0
#img = np.uint8(result * 255.0)
img = np.array(img)
width, high, channel = img.shape
width_new, high_new = (256, 256)
img = img
#img = cv2.imread(image_path)
#img = Image.fromarray(cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
# 当为单通道图像时，转化为三通道
if len(img.shape) == 2:
img = img[:, :, np.newaxis]# 增加纬度
img = np.concatenate(, axis=2)# 拼接为三通道
#img = img.resize(img, (256, 256), Image.ANTIALIAS)
img = np.array(img)
img = img.transpose(2, 0, 1)# 矩阵转置 通道数放在前面
img = img / 255.
img = torch.FloatTensor(img).to(device)
normalize = transforms.Normalize(mean=,
std=)
transform = transforms.Compose()
image = transform(img)# (3, 256, 256)
# Encode
image = image.unsqueeze(0)# (1, 3, 256, 256)
encoder_out = encoder(image)# (1, enc_image_size, enc_image_size, encoder_dim) 1,14,14,2048
enc_image_size = encoder_out.size(1)
print('enc_image_size:', enc_image_size)
encoder_dim = encoder_out.size(3)
print('encoder_dim:', encoder_dim)
# Flatten encoding
encoder_out = encoder_out.view(1, -1, encoder_dim)# (1, num_pixels, encoder_dim) 1,196,2048
# 表示了图像的196个区域各自的特征
# print('encoder_out:',encoder_out)
num_pixels = encoder_out.size(1)# 第二位 196
# print('num_pixels:',num_pixels)
# We'll treat the problem as having a batch size of k
# print(encoder_out.size())
encoder_out = encoder_out.expand(k, num_pixels, encoder_dim)# (k, num_pixels, encoder_dim)1->k纬度扩展，五份特征
# print(encoder_out.size())
# Tensor to store top k previous words at each step; now they're just <start>
k_prev_words = torch.LongTensor([]] * k).to(device)# (k, 1)
# print('k_prev_words:',k_prev_words)
# Tensor to store top k sequences; now they're just <start>
seqs = k_prev_words# (k, 1)
# Tensor to store top k sequences' scores; now they're just 0
top_k_scores = torch.zeros(k, 1).to(device)# (k, 1)
# Tensor to store top k sequences' alphas; now they're just 1s 这里其实就是存储每个字对应图像上的关注区域，映射在14*14的张量上面
seqs_alpha = torch.ones(k, 1, enc_image_size, enc_image_size).to(device)# (k, 1, enc_image_size, enc_image_size)
# Lists to store completed sequences, their alphas and scores
complete_seqs = list()
complete_seqs_alpha = list()
complete_seqs_scores = list()
# Start decoding
step = 1
h, c = decoder.init_hidden_state(encoder_out)# h0
print('h, c', h.size(), c.size())
# s is a number less than or equal to k, because sequences are removed from this process once they hit <end>
while True:
embeddings = decoder.embedding(k_prev_words).squeeze(1)# (s, embed_dim) （5,隐层512）
print('embeddings', embeddings.size())
# encode的图片表示 和隐状态
awe, alpha = decoder.attention(encoder_out,
h)# (s, encoder_dim), (s, num_pixels)（5，2048（）,5，196（attention 存储字对应图像各部分的权重））
print(' awe, alpha', awe.size(), alpha.size())
# 0/0
alpha = alpha.view(-1, enc_image_size, enc_image_size)# (s, enc_image_size, enc_image_size)（5,14,14）
gate = decoder.sigmoid(decoder.f_beta(h))# gating scalar, (s, encoder_dim)
awe = gate * awe# 给特征赋予权重
h, c = decoder.decode_step(torch.cat(, dim=1),
(h, c))# (s, decoder_dim)输入(512,2048),(512,512)带权重的特征和上一次的lstm输出和细胞状态值
scores = decoder.fc(h)# (s, vocab_size)
scores = F.log_softmax(scores, dim=1)
print('scores', scores.size())
# Add 每一句 含有多少词 更新
scores = top_k_scores.expand_as(scores) + scores# (s, vocab_size)
print('top_k_scores，scores', top_k_scores.size(), scores.size())
# For the first step, all k points will have the same scores (since same k previous words, h, c)
if step == 1:
top_k_scores, top_k_words = scores.topk(k, 0, True, True)# (s)
else:
# Unroll and find top scores, and their unrolled indices
top_k_scores, top_k_words = scores.view(-1).topk(k, 0, True, True)# (s) 取词，top
print('top_k_scores，top_k_words', top_k_scores.size(), top_k_words.size())
# Convert unrolled indices to actual indices of scores
prev_word_inds = torch.floor_divide(top_k_words, vocab_size)
# prev_word_inds = top_k_words / vocab_size# (s)
next_word_inds = top_k_words % vocab_size# (s)
print('top_k_scores,top_k_words，prev_word_inds，next_word_inds', top_k_words, top_k_scores, prev_word_inds,
next_word_inds)
# Add new words to sequences, alphas
seqs = torch.cat(, next_word_inds.unsqueeze(1)], dim=1)# (s, step+1)#词加一
seqs_alpha = torch.cat(, alpha.unsqueeze(1)],# 词对应图像区域加一
dim=1)# (s, step+1, enc_image_size, enc_image_size)
# Which sequences are incomplete (didn't reach <end>)? 挑出这次循环完结的 句子
incomplete_inds = [ind for ind, next_word in enumerate(next_word_inds) if
next_word != word_map['<end>']]
complete_inds = list(set(range(len(next_word_inds))) - set(incomplete_inds))
# Set aside complete sequences 挑出完整序列
if len(complete_inds) > 0:
complete_seqs.extend(seqs.tolist())# 追加全部序列
complete_seqs_alpha.extend(seqs_alpha.tolist())
complete_seqs_scores.extend(top_k_scores)
k -= len(complete_inds)# reduce beam length accordingly
# Proceed with incomplete sequences
if k == 0:
break
# 更新参数 只保留未完全序列参数
seqs = seqs
seqs_alpha = seqs_alpha
h = h]
c = c]
encoder_out = encoder_out]
top_k_scores = top_k_scores.unsqueeze(1)
k_prev_words = next_word_inds.unsqueeze(1)
# Break if things have been going on too long
if step > 50:
break
step += 1
# 标记 scores分数最高序列作为返回值。
i = complete_seqs_scores.index(max(complete_seqs_scores))
seq = complete_seqs
alphas = complete_seqs_alpha
return seq, alphas
def visualize_att(image_path, seq, alphas, rev_word_map, smooth=True):
"""
Visualizes caption with weights at every word.
Adapted from paper authors' repo: https://github.com/kelvinxu/arctic-captions/blob/master/alpha_visualization.ipynb
:param image_path: path to image that has been captioned
:param seq: caption
:param alphas: weights
:param rev_word_map: reverse word mapping, i.e. ix2word
:param smooth: smooth weights?
"""
image = Image.open(image_path)
image = image.resize(, Image.LANCZOS)
words = for ind in seq]
#print(words)
for t in range(1,len(words)-1):
if t > 50:
break
plt.subplot(int(np.ceil(len(words)) / 5.), 6, t)
plt.text(0, 1, '%s' % (words), color='black', backgroundcolor='white', fontsize=12)
plt.imshow(image)
current_alpha = alphas
if smooth:
alpha = skimage.transform.pyramid_expand(current_alpha.numpy(), upscale=12, sigma=8)
else:
alpha = skimage.transform.resize(current_alpha.numpy(), )
if t == 0:
plt.imshow(alpha, alpha=0)
else:
plt.imshow(alpha, alpha=0.8)
plt.set_cmap(cm.Greys_r)
plt.axis('off')
#plt.show()
import scipy
print(scipy.__version__)
#checkpoint = torch.load('./BEST_checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar', map_location=str(device))
checkpoint = torch.load(r'C:/Users/Ternence/PycharmProjects/pythonProject/tuxiang/BEST_checkpoint_flickr8k_5_cap_per_img_5_min_word_freq.pth.tar', map_location=str(device))
decoder = checkpoint['decoder']
decoder = decoder.to(device)
decoder.eval()
encoder = checkpoint['encoder']
encoder = encoder.to(device)
encoder.eval()
# Load word map (word2ix)
with open(r'C:/Users/Ternence/kind2/Flickr8k/data/WORDMAP_flickr8k_5_cap_per_img_5_min_word_freq.json', 'r') as j:
word_map = json.load(j,)
rev_word_map = {v: k for k, v in word_map.items()}# ix2word
# Encode, decode with attention and beam search
seq, alphas = caption_image_beam_search(encoder, decoder, r'C:/Users/Ternence/kind2/Flickr8k/tupian/ren2.jpg', word_map, 5)
alphas = torch.FloatTensor(alphas)
# Visualize caption and attention of best sequence
visualize_att(r'C:/Users/Ternence/kind2/Flickr8k/tupian/ren2.jpg', seq, alphas, rev_word_map, True)
words = for ind in seq]
print(words)
这个是b.py。是图像字幕的可视化功能 isdkz 发表于 2022-12-23 10:22
你想在一个程序中执行另一个程序的话最好用 exec，而不应该用 import
eg:
问题解决了，多谢大佬
页:
[1]