Littlehongman
commited on
Commit
β’
4cea813
1
Parent(s):
20dcad7
First version success
Browse files- .gitignore +2 -0
- app.py +10 -0
- model.py +137 -0
- predict.py +42 -0
- requirements.txt +7 -0
.gitignore
ADDED
@@ -0,0 +1,2 @@
|
|
|
|
|
|
|
1 |
+
artifacts/
|
2 |
+
wandb/
|
app.py
CHANGED
@@ -2,6 +2,9 @@ import streamlit as st
|
|
2 |
import streamlit.components.v1 as components
|
3 |
from PIL import Image
|
4 |
|
|
|
|
|
|
|
5 |
|
6 |
# Configure Streamlit page
|
7 |
st.set_page_config(page_title="Caption Machine", page_icon="π₯")
|
@@ -40,4 +43,11 @@ if upload_file is not None:
|
|
40 |
st.image(img)
|
41 |
st.write("Image Uploaded Successfully")
|
42 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
43 |
|
|
|
2 |
import streamlit.components.v1 as components
|
3 |
from PIL import Image
|
4 |
|
5 |
+
from predict import generate_text
|
6 |
+
from model import load_clip_model, load_gpt_model, load_model
|
7 |
+
|
8 |
|
9 |
# Configure Streamlit page
|
10 |
st.set_page_config(page_title="Caption Machine", page_icon="π₯")
|
|
|
43 |
st.image(img)
|
44 |
st.write("Image Uploaded Successfully")
|
45 |
|
46 |
+
# gpt_model, tokenizer = load_gpt_model()
|
47 |
+
|
48 |
+
model, image_transform, tokenizer = load_model()
|
49 |
+
caption = generate_text(model, img, tokenizer, image_transform)
|
50 |
+
|
51 |
+
st.write(caption)
|
52 |
+
|
53 |
|
model.py
CHANGED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import wandb
|
4 |
+
import streamlit as st
|
5 |
+
|
6 |
+
import clip
|
7 |
+
from transformers import GPT2Tokenizer, GPT2LMHeadModel
|
8 |
+
|
9 |
+
|
10 |
+
class ImageEncoder(nn.Module):
|
11 |
+
|
12 |
+
def __init__(self, base_network):
|
13 |
+
super(ImageEncoder, self).__init__()
|
14 |
+
self.base_network = base_network
|
15 |
+
self.embedding_size = self.base_network.token_embedding.weight.shape[1]
|
16 |
+
|
17 |
+
def forward(self, images):
|
18 |
+
with torch.no_grad():
|
19 |
+
x = self.base_network.encode_image(images)
|
20 |
+
x = x / x.norm(dim=1, keepdim=True)
|
21 |
+
x = x.float()
|
22 |
+
|
23 |
+
return x
|
24 |
+
|
25 |
+
class Mapping(nn.Module):
|
26 |
+
# Map the featureMap from CLIP model to GPT2
|
27 |
+
def __init__(self, clip_embedding_size, gpt_embedding_size, length=30): # length: sentence length
|
28 |
+
super(Mapping, self).__init__()
|
29 |
+
|
30 |
+
self.clip_embedding_size = clip_embedding_size
|
31 |
+
self.gpt_embedding_size = gpt_embedding_size
|
32 |
+
self.length = length
|
33 |
+
|
34 |
+
self.fc1 = nn.Linear(clip_embedding_size, gpt_embedding_size * length)
|
35 |
+
|
36 |
+
def forward(self, x):
|
37 |
+
x = self.fc1(x)
|
38 |
+
|
39 |
+
return x.view(-1, self.length, self.gpt_embedding_size)
|
40 |
+
|
41 |
+
|
42 |
+
class TextDecoder(nn.Module):
|
43 |
+
def __init__(self, base_network):
|
44 |
+
super(TextDecoder, self).__init__()
|
45 |
+
self.base_network = base_network
|
46 |
+
self.embedding_size = self.base_network.transformer.wte.weight.shape[1]
|
47 |
+
self.vocab_size = self.base_network.transformer.wte.weight.shape[0]
|
48 |
+
|
49 |
+
def forward(self, concat_embedding, mask=None):
|
50 |
+
return self.base_network(inputs_embeds=concat_embedding, attention_mask=mask)
|
51 |
+
|
52 |
+
|
53 |
+
def get_embedding(self, texts):
|
54 |
+
return self.base_network.transformer.wte(texts)
|
55 |
+
|
56 |
+
|
57 |
+
import pytorch_lightning as pl
|
58 |
+
|
59 |
+
|
60 |
+
class ImageCaptioner(pl.LightningModule):
|
61 |
+
def __init__(self, clip_model, gpt_model, tokenizer, total_steps, max_length=20):
|
62 |
+
super(ImageCaptioner, self).__init__()
|
63 |
+
|
64 |
+
self.padding_token_id = tokenizer.pad_token_id
|
65 |
+
#self.stop_token_id = tokenizer.encode('.')[0]
|
66 |
+
|
67 |
+
# Define networks
|
68 |
+
self.clip = ImageEncoder(clip_model)
|
69 |
+
self.gpt = TextDecoder(gpt_model)
|
70 |
+
self.mapping_network = Mapping(self.clip.embedding_size, self.gpt.embedding_size, max_length)
|
71 |
+
|
72 |
+
# Define variables
|
73 |
+
self.total_steps = total_steps
|
74 |
+
self.max_length = max_length
|
75 |
+
self.clip_embedding_size = self.clip.embedding_size
|
76 |
+
self.gpt_embedding_size = self.gpt.embedding_size
|
77 |
+
self.gpt_vocab_size = self.gpt.vocab_size
|
78 |
+
|
79 |
+
|
80 |
+
def forward(self, images, texts, masks):
|
81 |
+
texts_embedding = self.gpt.get_embedding(texts)
|
82 |
+
images_embedding = self.clip(images)
|
83 |
+
|
84 |
+
images_projection = self.mapping_network(images_embedding).view(-1, self.max_length, self.gpt_embedding_size)
|
85 |
+
embedding_concat = torch.cat((images_projection, texts_embedding), dim=1)
|
86 |
+
|
87 |
+
out = self.gpt(embedding_concat, masks)
|
88 |
+
|
89 |
+
return out
|
90 |
+
|
91 |
+
@st.cache_resource
|
92 |
+
def download_trained_model():
|
93 |
+
wandb.init(anonymous="must")
|
94 |
+
|
95 |
+
api = wandb.Api()
|
96 |
+
artifact = api.artifact('hungchiehwu/CLIP-L14_GPT/model-ql03493w:v3')
|
97 |
+
artifact_dir = artifact.download()
|
98 |
+
|
99 |
+
wandb.finish()
|
100 |
+
|
101 |
+
return artifact_dir
|
102 |
+
|
103 |
+
@st.cache_resource
|
104 |
+
def load_clip_model():
|
105 |
+
|
106 |
+
clip_model, image_transform = clip.load("ViT-L/14", device="cpu")
|
107 |
+
|
108 |
+
return clip_model, image_transform
|
109 |
+
|
110 |
+
@st.cache_resource
|
111 |
+
def load_gpt_model():
|
112 |
+
tokenizer = GPT2Tokenizer.from_pretrained('gpt2')
|
113 |
+
gpt_model = GPT2LMHeadModel.from_pretrained('gpt2')
|
114 |
+
|
115 |
+
tokenizer.pad_token = tokenizer.eos_token
|
116 |
+
|
117 |
+
return gpt_model, tokenizer
|
118 |
+
|
119 |
+
@st.cache_resource
|
120 |
+
def load_model():
|
121 |
+
|
122 |
+
# # Load fine-tuned model from wandb
|
123 |
+
artifact_dir = download_trained_model()
|
124 |
+
PATH = f"{artifact_dir[2:]}/model.ckpt"
|
125 |
+
|
126 |
+
# Load pretrained GPT, CLIP model from OpenAI
|
127 |
+
clip_model, image_transfrom = load_clip_model()
|
128 |
+
gpt_model, tokenizer = load_gpt_model()
|
129 |
+
|
130 |
+
|
131 |
+
|
132 |
+
# Load weights
|
133 |
+
model = ImageCaptioner(clip_model, gpt_model, tokenizer, 0)
|
134 |
+
checkpoint = torch.load(PATH, map_location=torch.device('cpu'))
|
135 |
+
model.load_state_dict(checkpoint["state_dict"])
|
136 |
+
|
137 |
+
return model, image_transfrom, tokenizer
|
predict.py
ADDED
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
def generate_text(model, image, tokenizer, image_transfrom, max_length=30):
|
4 |
+
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
5 |
+
|
6 |
+
model.eval()
|
7 |
+
# model = model.to(device)
|
8 |
+
|
9 |
+
temperature = 0.9
|
10 |
+
stop_token_id = tokenizer.pad_token_id
|
11 |
+
output_ids = []
|
12 |
+
|
13 |
+
|
14 |
+
image = image_transfrom(image)
|
15 |
+
img_tensor = image.unsqueeze(0)#.to(device)
|
16 |
+
images_embedding = model.clip(img_tensor)
|
17 |
+
|
18 |
+
images_projection = model.mapping_network(images_embedding).view(-1, model.max_length, model.gpt_embedding_size)
|
19 |
+
|
20 |
+
input_state = images_projection
|
21 |
+
|
22 |
+
with torch.no_grad():
|
23 |
+
for i in range(max_length):
|
24 |
+
outputs = model.gpt(input_state, None).logits
|
25 |
+
|
26 |
+
next_token_scores = outputs[0, -1, :].detach().div(temperature).softmax(dim=0)
|
27 |
+
|
28 |
+
#next_token_id = np.random.choice(len(next_token_scores), p = next_token_scores.cpu().numpy())
|
29 |
+
next_token_id = next_token_scores.max(dim=0).indices.item()
|
30 |
+
|
31 |
+
if next_token_id == stop_token_id:
|
32 |
+
break
|
33 |
+
|
34 |
+
output_ids.append(next_token_id)
|
35 |
+
|
36 |
+
|
37 |
+
# Update state
|
38 |
+
next_token_id = torch.tensor([next_token_id]).unsqueeze(0)#.to(device)
|
39 |
+
next_token_embed = model.gpt.base_network.transformer.wte(next_token_id)
|
40 |
+
input_state = torch.cat((input_state, next_token_embed), dim=1)
|
41 |
+
|
42 |
+
return tokenizer.decode(output_ids)
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
ftfy
|
2 |
+
regex
|
3 |
+
tqdm
|
4 |
+
git+https://github.com/openai/CLIP.git
|
5 |
+
transformers
|
6 |
+
pytorch-lightning==1.9.0
|
7 |
+
wandb
|