Spaces:
Sleeping
Sleeping
nishantguvvada
commited on
Commit
•
19c08c2
1
Parent(s):
dead93e
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import streamlit as st
|
2 |
+
import tensorflow as tf
|
3 |
+
import cv2
|
4 |
+
import numpy as np
|
5 |
+
from PIL import Image, ImageOps
|
6 |
+
import imageio.v3 as iio
|
7 |
+
import time
|
8 |
+
from textwrap import wrap
|
9 |
+
|
10 |
+
import matplotlib.pylab as plt
|
11 |
+
import numpy as np
|
12 |
+
import tensorflow as tf
|
13 |
+
import tensorflow_datasets as tfds
|
14 |
+
import tensorflow_hub as hub
|
15 |
+
from tensorflow.keras import Input
|
16 |
+
from tensorflow.keras.layers import (
|
17 |
+
GRU,
|
18 |
+
Add,
|
19 |
+
AdditiveAttention,
|
20 |
+
Attention,
|
21 |
+
Concatenate,
|
22 |
+
Dense,
|
23 |
+
Embedding,
|
24 |
+
LayerNormalization,
|
25 |
+
Reshape,
|
26 |
+
StringLookup,
|
27 |
+
TextVectorization,
|
28 |
+
)
|
29 |
+
|
30 |
+
@st.cache_resource()
|
31 |
+
def load_image_model():
|
32 |
+
image_model=tf.keras.models.load_model('./image_caption_model.h5')
|
33 |
+
return image_model
|
34 |
+
|
35 |
+
@st.cache_resource()
|
36 |
+
def load_decoder_model():
|
37 |
+
decoder_model=tf.keras.models.load_model('./decoder_pred_model.h5')
|
38 |
+
return decoder_model
|
39 |
+
|
40 |
+
@st.cache_resource()
|
41 |
+
def load_encoder_model():
|
42 |
+
encoder=tf.keras.models.load_model('./encoder_model.h5')
|
43 |
+
return encoder
|
44 |
+
|
45 |
+
st.title(":blue[Nishant Guvvada's] :red[AI Journey] Image Caption Generation")
|
46 |
+
image = Image.open('./title.jpg')
|
47 |
+
st.image(image)
|
48 |
+
st.write("""
|
49 |
+
# Multi-Modal Machine Learning
|
50 |
+
"""
|
51 |
+
)
|
52 |
+
|
53 |
+
file = st.file_uploader("Upload any image and the model will try to provide a caption to it!", type= ['png', 'jpg'])
|
54 |
+
|
55 |
+
MAX_CAPTION_LEN = 64
|
56 |
+
MINIMUM_SENTENCE_LENGTH = 5
|
57 |
+
IMG_HEIGHT = 299
|
58 |
+
IMG_WIDTH = 299
|
59 |
+
IMG_CHANNELS = 3
|
60 |
+
ATTENTION_DIM = 512 # size of dense layer in Attention
|
61 |
+
VOCAB_SIZE = 20000
|
62 |
+
|
63 |
+
|
64 |
+
# We will override the default standardization of TextVectorization to preserve
|
65 |
+
# "<>" characters, so we preserve the tokens for the <start> and <end>.
|
66 |
+
def standardize(inputs):
|
67 |
+
inputs = tf.strings.lower(inputs)
|
68 |
+
return tf.strings.regex_replace(
|
69 |
+
inputs, r"[!\"#$%&\(\)\*\+.,-/:;=?@\[\\\]^_`{|}~]?", ""
|
70 |
+
)
|
71 |
+
|
72 |
+
|
73 |
+
# Choose the most frequent words from the vocabulary & remove punctuation etc.
|
74 |
+
tokenizer = TextVectorization(
|
75 |
+
max_tokens=VOCAB_SIZE,
|
76 |
+
standardize=standardize,
|
77 |
+
output_sequence_length=MAX_CAPTION_LEN,
|
78 |
+
)
|
79 |
+
|
80 |
+
# Lookup table: Word -> Index
|
81 |
+
word_to_index = StringLookup(
|
82 |
+
mask_token="", vocabulary=tokenizer.get_vocabulary()
|
83 |
+
)
|
84 |
+
|
85 |
+
# Lookup table: Index -> Word
|
86 |
+
index_to_word = StringLookup(
|
87 |
+
mask_token="", vocabulary=tokenizer.get_vocabulary(), invert=True
|
88 |
+
)
|
89 |
+
|
90 |
+
|
91 |
+
## Probabilistic prediction using the trained model
|
92 |
+
def predict_caption(file):
|
93 |
+
gru_state = tf.zeros((1, ATTENTION_DIM))
|
94 |
+
|
95 |
+
img = tf.image.decode_jpeg(tf.io.read_file(filename), channels=IMG_CHANNELS)
|
96 |
+
img = tf.image.resize(img, (IMG_HEIGHT, IMG_WIDTH))
|
97 |
+
img = img / 255
|
98 |
+
|
99 |
+
encoder = load_encoder_model()
|
100 |
+
features = encoder(tf.expand_dims(img, axis=0))
|
101 |
+
dec_input = tf.expand_dims([word_to_index("<start>")], 1)
|
102 |
+
result = []
|
103 |
+
decoder_pred_model = load_decoder_model()
|
104 |
+
for i in range(MAX_CAPTION_LEN):
|
105 |
+
predictions, gru_state = decoder_pred_model(
|
106 |
+
[dec_input, gru_state, features]
|
107 |
+
)
|
108 |
+
|
109 |
+
# draws from log distribution given by predictions
|
110 |
+
top_probs, top_idxs = tf.math.top_k(
|
111 |
+
input=predictions[0][0], k=10, sorted=False
|
112 |
+
)
|
113 |
+
chosen_id = tf.random.categorical([top_probs], 1)[0].numpy()
|
114 |
+
predicted_id = top_idxs.numpy()[chosen_id][0]
|
115 |
+
|
116 |
+
result.append(tokenizer.get_vocabulary()[predicted_id])
|
117 |
+
|
118 |
+
if predicted_id == word_to_index("<end>"):
|
119 |
+
return img, result
|
120 |
+
|
121 |
+
dec_input = tf.expand_dims([predicted_id], 1)
|
122 |
+
|
123 |
+
return img, result
|
124 |
+
|
125 |
+
|
126 |
+
filename = "../sample_images/surf.jpeg" # you can also try surf.jpeg
|
127 |
+
|
128 |
+
for i in range(5):
|
129 |
+
image, caption = predict_caption(filename)
|
130 |
+
print(" ".join(caption[:-1]) + ".")
|
131 |
+
|
132 |
+
img = tf.image.decode_jpeg(tf.io.read_file(filename), channels=IMG_CHANNELS)
|
133 |
+
plt.imshow(img)
|
134 |
+
plt.axis("off")
|
135 |
+
|
136 |
+
|
137 |
+
filename = np.array(Image.open(file).convert('RGB'))
|
138 |
+
|
139 |
+
def model_prediction(path):
|
140 |
+
resize = tf.image.resize(path, (256,256))
|
141 |
+
with st.spinner('Model is being loaded..'):
|
142 |
+
model=load_image_model()
|
143 |
+
yhat = model.predict(np.expand_dims(resize/255, 0))
|
144 |
+
return yhat
|
145 |
+
|
146 |
+
def on_click():
|
147 |
+
if file is None:
|
148 |
+
st.text("Please upload an image file")
|
149 |
+
else:
|
150 |
+
image = Image.open(file)
|
151 |
+
st.image(image, use_column_width=True)
|
152 |
+
image = image.convert('RGB')
|
153 |
+
predictions = model_prediction(np.array(image))
|
154 |
+
if (predictions>0.5):
|
155 |
+
st.write("""# Prediction : Implant is loose""")
|
156 |
+
else:
|
157 |
+
st.write("""# Prediction : Implant is in control""")
|
158 |
+
|
159 |
+
st.button('Predict', on_click=on_click)
|