Upload 3 files
Browse files- app.py +170 -0
- im2text_model_full.pt +3 -0
- vocab_full.json +0 -0
app.py
ADDED
@@ -0,0 +1,170 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from torchvision import transforms, models
|
6 |
+
from flask import Flask, request, jsonify , render_template
|
7 |
+
import imageio.v3 as imageio
|
8 |
+
import numpy as np
|
9 |
+
from io import BytesIO
|
10 |
+
from PIL import Image
|
11 |
+
# 建立詞彙表
|
12 |
+
class tokenizer():
|
13 |
+
def __init__(self, threshold=5):
|
14 |
+
self.word2idx = {}
|
15 |
+
self.idx2word = {}
|
16 |
+
self.threshold = threshold
|
17 |
+
self.word2count = {}
|
18 |
+
|
19 |
+
def build_vocab(self, corpus):
|
20 |
+
print('buiding vocab......')
|
21 |
+
tokens = corpus.lower().split()
|
22 |
+
for token in tokens:
|
23 |
+
self.word2count[token] = self.word2count.get(token, 0) + 1
|
24 |
+
idx = 0
|
25 |
+
for word, count in self.word2count.items():
|
26 |
+
if count >= self.threshold:
|
27 |
+
self.word2idx[word] = idx
|
28 |
+
self.idx2word[idx] = word
|
29 |
+
idx += 1
|
30 |
+
print(f'Vocab size: {len(self.idx2word)}')
|
31 |
+
|
32 |
+
def encode(self, sentence):
|
33 |
+
tokens = sentence.lower().split()
|
34 |
+
return [self.word2idx.get(token, self.word2idx['<unk>']) for token in tokens]
|
35 |
+
|
36 |
+
def decode(self, indices):
|
37 |
+
return ' '.join([self.idx2word.get(idx, '<unk>') for idx in indices])
|
38 |
+
|
39 |
+
def save_vocab(self, filepath):
|
40 |
+
with open(filepath, 'w') as f:
|
41 |
+
json.dump({'word2idx': self.word2idx, 'idx2word': self.idx2word}, f)
|
42 |
+
|
43 |
+
def load_vocab(self, filepath):
|
44 |
+
with open(filepath, 'r') as f:
|
45 |
+
data = json.load(f)
|
46 |
+
self.word2idx = data['word2idx']
|
47 |
+
self.idx2word = {int(k): v for k, v in data['idx2word'].items()}
|
48 |
+
|
49 |
+
# 定義CNN編碼器
|
50 |
+
class CNNEncoder(nn.Module):
|
51 |
+
def __init__(self, embed_size, num_groups=32):
|
52 |
+
super(CNNEncoder, self).__init__()
|
53 |
+
resnet = models.resnet50(weights=models.ResNet50_Weights.DEFAULT)
|
54 |
+
for param in resnet.parameters():
|
55 |
+
param.requires_grad = False
|
56 |
+
self.resnet = nn.Sequential(*list(resnet.children())[:-1])
|
57 |
+
self.linear = nn.Linear(resnet.fc.in_features, embed_size)
|
58 |
+
self.gn = nn.GroupNorm(num_groups, embed_size)
|
59 |
+
|
60 |
+
def forward(self, images):
|
61 |
+
with torch.no_grad():
|
62 |
+
features = self.resnet(images)
|
63 |
+
features = features.view(features.size(0), -1)
|
64 |
+
features = self.gn(self.linear(features))
|
65 |
+
return features
|
66 |
+
|
67 |
+
# 定義RNN解碼器
|
68 |
+
class RNNDecoder(nn.Module):
|
69 |
+
def __init__(self, embed_size, hidden_size, vocab_size, num_layers):
|
70 |
+
super(RNNDecoder, self).__init__()
|
71 |
+
self.embed = nn.Embedding(vocab_size, embed_size)
|
72 |
+
self.lstm = nn.LSTM(embed_size, hidden_size, num_layers, batch_first=True)
|
73 |
+
self.linear = nn.Linear(hidden_size, vocab_size)
|
74 |
+
self.embed_size = embed_size
|
75 |
+
self.hidden_size = hidden_size
|
76 |
+
self.num_layers = num_layers
|
77 |
+
|
78 |
+
def forward(self, features, captions):
|
79 |
+
embeddings = self.embed(captions)
|
80 |
+
embeddings = torch.cat((features.unsqueeze(1), embeddings), 1)
|
81 |
+
hiddens, _ = self.lstm(embeddings)
|
82 |
+
outputs = self.linear(hiddens[:, 1:, :])
|
83 |
+
return outputs
|
84 |
+
|
85 |
+
def sample(self, features, states=None, max_len=20):
|
86 |
+
sampled_ids = [vocab.word2idx['<start>']]
|
87 |
+
inputs = features.unsqueeze(1)
|
88 |
+
start_token = torch.tensor([vocab.word2idx['<start>']]).to(device).unsqueeze(0)
|
89 |
+
inputs = torch.cat((features.unsqueeze(1), self.embed(start_token)), 1)
|
90 |
+
for i in range(max_len):
|
91 |
+
hiddens, states = self.lstm(inputs, states)
|
92 |
+
outputs = self.linear(hiddens[:, -1, :]) # take the output of the last time step
|
93 |
+
_, predicted = outputs.max(1)
|
94 |
+
sampled_ids.append(predicted.item())
|
95 |
+
if predicted.item() == vocab.word2idx['<end>']:
|
96 |
+
break
|
97 |
+
inputs = self.embed(predicted).unsqueeze(1)
|
98 |
+
return sampled_ids
|
99 |
+
|
100 |
+
# 定義ImageToText模型
|
101 |
+
class im2text_model(nn.Module):
|
102 |
+
def __init__(self, cnn_encoder, rnn_decoder):
|
103 |
+
super(im2text_model, self).__init__()
|
104 |
+
self.encoder = cnn_encoder
|
105 |
+
self.decoder = rnn_decoder
|
106 |
+
|
107 |
+
def forward(self, images, captions):
|
108 |
+
features = self.encoder(images)
|
109 |
+
outputs = self.decoder(features, captions)
|
110 |
+
return outputs
|
111 |
+
|
112 |
+
def sample(self, images, states=None):
|
113 |
+
features = self.encoder(images)
|
114 |
+
sampled_ids = self.decoder.sample(features, states)
|
115 |
+
return sampled_ids
|
116 |
+
|
117 |
+
# 初始化應用
|
118 |
+
app = Flask(__name__)
|
119 |
+
|
120 |
+
# 加載詞彙表
|
121 |
+
vocab = tokenizer()
|
122 |
+
vocab.load_vocab('vocab_full.json')
|
123 |
+
|
124 |
+
# 加載模型
|
125 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
126 |
+
model = torch.load('im2text_model_full.pt', map_location=torch.device('cpu'))
|
127 |
+
model.to(device)
|
128 |
+
model.eval()
|
129 |
+
|
130 |
+
transform = transforms.Compose([
|
131 |
+
transforms.Resize((224, 224), antialias=True),
|
132 |
+
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
|
133 |
+
])
|
134 |
+
@app.route('/')
|
135 |
+
def index():
|
136 |
+
return render_template('index.html')
|
137 |
+
|
138 |
+
@app.route('/upload', methods=['POST'])
|
139 |
+
def upload_image():
|
140 |
+
if 'file' not in request.files:
|
141 |
+
return jsonify({'error': 'No file part'})
|
142 |
+
file = request.files['file']
|
143 |
+
if file.filename == '':
|
144 |
+
return jsonify({'error': 'No selected file'})
|
145 |
+
if file:
|
146 |
+
# Convert image to RGB format if necessary and process in memory
|
147 |
+
image = Image.open(file.stream)
|
148 |
+
if image.format in ['GIF', 'WebP', 'PNG']:
|
149 |
+
image = image.convert('RGB')
|
150 |
+
|
151 |
+
# Save image to a BytesIO object
|
152 |
+
byte_io = BytesIO()
|
153 |
+
image.save(byte_io, 'JPEG')
|
154 |
+
byte_io.seek(0)
|
155 |
+
|
156 |
+
image = imageio.imread(byte_io)
|
157 |
+
if len(image.shape) == 2:
|
158 |
+
image = np.stack([image] * 3, axis=0)
|
159 |
+
else:
|
160 |
+
image = np.transpose(image, (2, 0, 1))
|
161 |
+
image = torch.tensor(image / 255.0).float()
|
162 |
+
image = transform(image).unsqueeze(0).to(device)
|
163 |
+
|
164 |
+
with torch.no_grad():
|
165 |
+
generated_caption = model.sample(image)
|
166 |
+
generated_caption_text = vocab.decode(generated_caption)
|
167 |
+
|
168 |
+
return jsonify({'caption': generated_caption_text})
|
169 |
+
if __name__ == '__main__':
|
170 |
+
app.run(debug=True)
|
im2text_model_full.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c6f3a66c5ee31f06ee6ab67860d01c6c3230b318bfa8e441db91e7f52e5768d2
|
3 |
+
size 145253389
|
vocab_full.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|