Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,210 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import pandas as pd
|
2 |
+
from torch.utils.data import DataLoader
|
3 |
+
from torchvision import transforms
|
4 |
+
from tqdm.notebook import tqdm
|
5 |
+
import torch
|
6 |
+
from torch.autograd import Variable
|
7 |
+
import torchvision
|
8 |
+
import pickle
|
9 |
+
from PIL import Image
|
10 |
+
import torch.nn as nn
|
11 |
+
import math
|
12 |
+
import random
|
13 |
+
import gradio as gr
|
14 |
+
device = "cpu"
|
15 |
+
max_seq_len=67
|
16 |
+
with open('index_to_word.pkl', 'rb') as handle:
|
17 |
+
index_to_word = pickle.load(handle)
|
18 |
+
with open('word_to_index.pkl', 'rb') as handle:
|
19 |
+
word_to_index = pickle.load(handle)
|
20 |
+
|
21 |
+
resnet18 = torchvision.models.resnet18(pretrained=True).to(device)
|
22 |
+
resnet18.eval()
|
23 |
+
resNet18Layer4 = resnet18._modules.get('layer4').to(device)
|
24 |
+
|
25 |
+
def create_df(img):
|
26 |
+
df = pd.DataFrame({"image": [img]})
|
27 |
+
return df
|
28 |
+
|
29 |
+
def get_vector(t_img):
|
30 |
+
|
31 |
+
t_img = Variable(t_img)
|
32 |
+
my_embedding = torch.zeros(1, 512, 7, 7)
|
33 |
+
def copy_data(m, i, o):
|
34 |
+
my_embedding.copy_(o.data)
|
35 |
+
|
36 |
+
h = resNet18Layer4.register_forward_hook(copy_data)
|
37 |
+
resnet18(t_img)
|
38 |
+
|
39 |
+
h.remove()
|
40 |
+
return my_embedding
|
41 |
+
|
42 |
+
class extractImageFeatureResNetDataSet():
|
43 |
+
from PIL import Image
|
44 |
+
def __init__(self, data):
|
45 |
+
self.data = data
|
46 |
+
self.scaler = transforms.Resize([224, 224])
|
47 |
+
self.normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
|
48 |
+
std=[0.229, 0.224, 0.225])
|
49 |
+
self.to_tensor = transforms.ToTensor()
|
50 |
+
def __len__(self):
|
51 |
+
return len(self.data)
|
52 |
+
|
53 |
+
def __getitem__(self, idx):
|
54 |
+
|
55 |
+
image_name = self.data.iloc[idx]['image']
|
56 |
+
img_loc = str(image_name) #os.getcwd()+'/imput_img/'+str(image_name)
|
57 |
+
img = Image.open(img_loc)
|
58 |
+
t_img = self.normalize(self.to_tensor(self.scaler(img)))
|
59 |
+
|
60 |
+
return image_name, t_img
|
61 |
+
|
62 |
+
def feature_exctractor(df):
|
63 |
+
extract_imgFtr_ResNet_input = {}
|
64 |
+
input_ImageDataset_ResNet = extractImageFeatureResNetDataSet(df[['image']])
|
65 |
+
input_ImageDataloader_ResNet = DataLoader(input_ImageDataset_ResNet, batch_size = 1, shuffle=False)
|
66 |
+
for image_name, t_img in tqdm(input_ImageDataloader_ResNet):
|
67 |
+
t_img = t_img.to("cpu")
|
68 |
+
embdg = get_vector(t_img)
|
69 |
+
extract_imgFtr_ResNet_input[image_name[0]] = embdg
|
70 |
+
return extract_imgFtr_ResNet_input
|
71 |
+
|
72 |
+
class PositionalEncoding(nn.Module):
|
73 |
+
|
74 |
+
def __init__(self, d_model, dropout=0.1, max_len=max_seq_len):
|
75 |
+
super(PositionalEncoding, self).__init__()
|
76 |
+
self.dropout = nn.Dropout(p=dropout)
|
77 |
+
|
78 |
+
pe = torch.zeros(max_len, d_model)
|
79 |
+
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
|
80 |
+
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
|
81 |
+
pe[:, 0::2] = torch.sin(position * div_term)
|
82 |
+
pe[:, 1::2] = torch.cos(position * div_term)
|
83 |
+
pe = pe.unsqueeze(0)
|
84 |
+
self.register_buffer('pe', pe)
|
85 |
+
|
86 |
+
|
87 |
+
def forward(self, x):
|
88 |
+
if self.pe.size(0) < x.size(0):
|
89 |
+
self.pe = self.pe.repeat(x.size(0), 1, 1).to(device)
|
90 |
+
self.pe = self.pe[:x.size(0), : , : ]
|
91 |
+
|
92 |
+
x = x + self.pe
|
93 |
+
return self.dropout(x)
|
94 |
+
|
95 |
+
class ImageCaptionModel(nn.Module):
|
96 |
+
def __init__(self, n_head, n_decoder_layer, vocab_size, embedding_size):
|
97 |
+
super(ImageCaptionModel, self).__init__()
|
98 |
+
self.pos_encoder = PositionalEncoding(embedding_size, 0.1)
|
99 |
+
self.TransformerDecoderLayer = nn.TransformerDecoderLayer(d_model = embedding_size, nhead = n_head)
|
100 |
+
self.TransformerDecoder = nn.TransformerDecoder(decoder_layer = self.TransformerDecoderLayer, num_layers = n_decoder_layer)
|
101 |
+
self.embedding_size = embedding_size
|
102 |
+
self.embedding = nn.Embedding(vocab_size , embedding_size)
|
103 |
+
self.last_linear_layer = nn.Linear(embedding_size, vocab_size)
|
104 |
+
self.init_weights()
|
105 |
+
|
106 |
+
def init_weights(self):
|
107 |
+
initrange = 0.1
|
108 |
+
self.embedding.weight.data.uniform_(-initrange, initrange)
|
109 |
+
self.last_linear_layer.bias.data.zero_()
|
110 |
+
self.last_linear_layer.weight.data.uniform_(-initrange, initrange)
|
111 |
+
|
112 |
+
def generate_Mask(self, size, decoder_inp):
|
113 |
+
decoder_input_mask = (torch.triu(torch.ones(size, size)) == 1).transpose(0, 1)
|
114 |
+
decoder_input_mask = decoder_input_mask.float().masked_fill(decoder_input_mask == 0, float('-inf')).masked_fill(decoder_input_mask == 1, float(0.0))
|
115 |
+
|
116 |
+
decoder_input_pad_mask = decoder_inp.float().masked_fill(decoder_inp == 0, float(0.0)).masked_fill(decoder_inp > 0, float(1.0))
|
117 |
+
decoder_input_pad_mask_bool = decoder_inp == 0
|
118 |
+
|
119 |
+
return decoder_input_mask, decoder_input_pad_mask, decoder_input_pad_mask_bool
|
120 |
+
|
121 |
+
def forward(self, encoded_image, decoder_inp):
|
122 |
+
encoded_image = encoded_image.permute(1,0,2)
|
123 |
+
|
124 |
+
|
125 |
+
decoder_inp_embed = self.embedding(decoder_inp)* math.sqrt(self.embedding_size)
|
126 |
+
|
127 |
+
decoder_inp_embed = self.pos_encoder(decoder_inp_embed)
|
128 |
+
decoder_inp_embed = decoder_inp_embed.permute(1,0,2)
|
129 |
+
|
130 |
+
|
131 |
+
decoder_input_mask, decoder_input_pad_mask, decoder_input_pad_mask_bool = self.generate_Mask(decoder_inp.size(1), decoder_inp)
|
132 |
+
decoder_input_mask = decoder_input_mask.to(device)
|
133 |
+
decoder_input_pad_mask = decoder_input_pad_mask.to(device)
|
134 |
+
decoder_input_pad_mask_bool = decoder_input_pad_mask_bool.to(device)
|
135 |
+
|
136 |
+
|
137 |
+
decoder_output = self.TransformerDecoder(tgt = decoder_inp_embed, memory = encoded_image, tgt_mask = decoder_input_mask, tgt_key_padding_mask = decoder_input_pad_mask_bool)
|
138 |
+
|
139 |
+
final_output = self.last_linear_layer(decoder_output)
|
140 |
+
|
141 |
+
return final_output, decoder_input_pad_mask
|
142 |
+
|
143 |
+
|
144 |
+
def generate_caption(K, img_nm, extract_imgFtr_ResNet_input):
|
145 |
+
from PIL import Image
|
146 |
+
img_loc = str(img_nm)#os.getcwd()+'/imput_img/'+
|
147 |
+
image = Image.open(img_loc).convert("RGB")
|
148 |
+
#plt.imshow(image)
|
149 |
+
|
150 |
+
model.eval()
|
151 |
+
img_embed = extract_imgFtr_ResNet_input[img_nm].to(device)
|
152 |
+
|
153 |
+
|
154 |
+
img_embed = img_embed.permute(0,2,3,1)
|
155 |
+
img_embed = img_embed.view(img_embed.size(0), -1, img_embed.size(3))
|
156 |
+
|
157 |
+
|
158 |
+
input_seq = [pad_token]*max_seq_len
|
159 |
+
input_seq[0] = start_token
|
160 |
+
|
161 |
+
input_seq = torch.tensor(input_seq).unsqueeze(0).to(device)
|
162 |
+
predicted_sentence = []
|
163 |
+
with torch.no_grad():
|
164 |
+
for eval_iter in range(0, max_seq_len):
|
165 |
+
output, padding_mask = model.forward(img_embed, input_seq)
|
166 |
+
|
167 |
+
output = output[eval_iter, 0, :]
|
168 |
+
|
169 |
+
values = torch.topk(output, K).values.tolist()
|
170 |
+
indices = torch.topk(output, K).indices.tolist()
|
171 |
+
|
172 |
+
next_word_index = random.choices(indices, values, k = 1)[0]
|
173 |
+
|
174 |
+
next_word = index_to_word[next_word_index]
|
175 |
+
|
176 |
+
input_seq[:, eval_iter+1] = next_word_index
|
177 |
+
|
178 |
+
|
179 |
+
if next_word == '<end>' :
|
180 |
+
break
|
181 |
+
|
182 |
+
predicted_sentence.append(next_word)
|
183 |
+
return " ".join(predicted_sentence + ["."])
|
184 |
+
|
185 |
+
device = torch.device('cpu')
|
186 |
+
model = torch.load('./BestModel_20000_Datos', map_location=device)
|
187 |
+
start_token = word_to_index['<start>']
|
188 |
+
end_token = word_to_index['<end>']
|
189 |
+
pad_token = word_to_index['<pad>']
|
190 |
+
max_seq_len = 67
|
191 |
+
|
192 |
+
def predict(inp):
|
193 |
+
device = "cpu"
|
194 |
+
max_seq_len=67
|
195 |
+
with open('index_to_word.pkl', 'rb') as handle:
|
196 |
+
index_to_word = pickle.load(handle)
|
197 |
+
with open('word_to_index.pkl', 'rb') as handle:
|
198 |
+
word_to_index = pickle.load(handle)
|
199 |
+
|
200 |
+
resnet18 = torchvision.models.resnet18(pretrained=True).to(device)
|
201 |
+
resnet18.eval()
|
202 |
+
resNet18Layer4 = resnet18._modules.get('layer4').to(device)
|
203 |
+
df = create_df(inp)
|
204 |
+
extract_imgFtr_ResNet_input = feature_exctractor(df)
|
205 |
+
prediction = generate_caption(1, inp, extract_imgFtr_ResNet_input)
|
206 |
+
return prediction
|
207 |
+
|
208 |
+
gr.Interface(fn=predict,
|
209 |
+
inputs=gr.Image(type="filepath"),
|
210 |
+
outputs=gr.Text()).launch(debug=True)
|