Spaces:
Build error
Build error
Update app.py
Browse files
app.py
CHANGED
@@ -1,280 +1,53 @@
|
|
1 |
-
import pickle
|
2 |
-
import tensorflow as tf
|
3 |
-
import pandas as pd
|
4 |
-
import numpy as np
|
5 |
-
import os
|
6 |
import io
|
|
|
7 |
import streamlit as st
|
8 |
import requests
|
9 |
from PIL import Image
|
|
|
10 |
|
11 |
-
# Set environment variable
|
12 |
-
os.environ['PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION'] = 'python'
|
13 |
-
|
14 |
-
# Constants
|
15 |
-
MAX_LENGTH = 40
|
16 |
-
BATCH_SIZE = 32
|
17 |
-
BUFFER_SIZE = 1000
|
18 |
-
EMBEDDING_DIM = 512
|
19 |
-
UNITS = 512
|
20 |
-
|
21 |
-
# Load vocabulary
|
22 |
-
vocab = pickle.load(open('saved_vocabulary/vocab_coco.file', 'rb'))
|
23 |
-
|
24 |
-
tokenizer = tf.keras.layers.TextVectorization(
|
25 |
-
standardize=None,
|
26 |
-
output_sequence_length=MAX_LENGTH,
|
27 |
-
vocabulary=vocab
|
28 |
-
)
|
29 |
-
|
30 |
-
idx2word = tf.keras.layers.StringLookup(
|
31 |
-
mask_token="",
|
32 |
-
vocabulary=tokenizer.get_vocabulary(),
|
33 |
-
invert=True
|
34 |
-
)
|
35 |
-
|
36 |
-
# Model Definitions
|
37 |
-
def CNN_Encoder():
|
38 |
-
inception_v3 = tf.keras.applications.InceptionV3(
|
39 |
-
include_top=False,
|
40 |
-
weights='imagenet'
|
41 |
-
)
|
42 |
-
output = inception_v3.output
|
43 |
-
output = tf.keras.layers.Reshape(
|
44 |
-
(-1, output.shape[-1]))(output)
|
45 |
-
cnn_model = tf.keras.models.Model(inception_v3.input, output)
|
46 |
-
return cnn_model
|
47 |
-
|
48 |
-
class TransformerEncoderLayer(tf.keras.layers.Layer):
|
49 |
-
def __init__(self, embed_dim, num_heads):
|
50 |
-
super().__init__()
|
51 |
-
self.layer_norm_1 = tf.keras.layers.LayerNormalization()
|
52 |
-
self.layer_norm_2 = tf.keras.layers.LayerNormalization()
|
53 |
-
self.attention = tf.keras.layers.MultiHeadAttention(
|
54 |
-
num_heads=num_heads, key_dim=embed_dim)
|
55 |
-
self.dense = tf.keras.layers.Dense(embed_dim, activation="relu")
|
56 |
-
|
57 |
-
def call(self, x, training):
|
58 |
-
x = self.layer_norm_1(x)
|
59 |
-
x = self.dense(x)
|
60 |
-
attn_output = self.attention(
|
61 |
-
query=x,
|
62 |
-
value=x,
|
63 |
-
key=x,
|
64 |
-
attention_mask=None,
|
65 |
-
training=training
|
66 |
-
)
|
67 |
-
x = self.layer_norm_2(x + attn_output)
|
68 |
-
return x
|
69 |
-
|
70 |
-
class Embeddings(tf.keras.layers.Layer):
|
71 |
-
def __init__(self, vocab_size, embed_dim, max_len):
|
72 |
-
super().__init__()
|
73 |
-
self.token_embeddings = tf.keras.layers.Embedding(
|
74 |
-
vocab_size, embed_dim)
|
75 |
-
self.position_embeddings = tf.keras.layers.Embedding(
|
76 |
-
max_len, embed_dim, input_shape=(None, max_len))
|
77 |
-
|
78 |
-
def call(self, input_ids):
|
79 |
-
length = tf.shape(input_ids)[-1]
|
80 |
-
position_ids = tf.range(start=0, limit=length, delta=1)
|
81 |
-
position_ids = tf.expand_dims(position_ids, axis=0)
|
82 |
-
token_embeddings = self.token_embeddings(input_ids)
|
83 |
-
position_embeddings = self.position_embeddings(position_ids)
|
84 |
-
return token_embeddings + position_embeddings
|
85 |
-
|
86 |
-
class TransformerDecoderLayer(tf.keras.layers.Layer):
|
87 |
-
def __init__(self, embed_dim, units, num_heads):
|
88 |
-
super().__init__()
|
89 |
-
self.embedding = Embeddings(
|
90 |
-
tokenizer.vocabulary_size(), embed_dim, MAX_LENGTH)
|
91 |
-
self.attention_1 = tf.keras.layers.MultiHeadAttention(
|
92 |
-
num_heads=num_heads, key_dim=embed_dim, dropout=0.1
|
93 |
-
)
|
94 |
-
self.attention_2 = tf.keras.layers.MultiHeadAttention(
|
95 |
-
num_heads=num_heads, key_dim=embed_dim, dropout=0.1
|
96 |
-
)
|
97 |
-
self.layernorm_1 = tf.keras.layers.LayerNormalization()
|
98 |
-
self.layernorm_2 = tf.keras.layers.LayerNormalization()
|
99 |
-
self.layernorm_3 = tf.keras.layers.LayerNormalization()
|
100 |
-
self.ffn_layer_1 = tf.keras.layers.Dense(units, activation="relu")
|
101 |
-
self.ffn_layer_2 = tf.keras.layers.Dense(embed_dim)
|
102 |
-
self.out = tf.keras.layers.Dense(tokenizer.vocabulary_size(), activation="softmax")
|
103 |
-
self.dropout_1 = tf.keras.layers.Dropout(0.3)
|
104 |
-
self.dropout_2 = tf.keras.layers.Dropout(0.5)
|
105 |
-
|
106 |
-
def call(self, input_ids, encoder_output, training, mask=None):
|
107 |
-
embeddings = self.embedding(input_ids)
|
108 |
-
combined_mask = None
|
109 |
-
padding_mask = None
|
110 |
-
|
111 |
-
if mask is not None:
|
112 |
-
causal_mask = self.get_causal_attention_mask(embeddings)
|
113 |
-
padding_mask = tf.cast(mask[:, :, tf.newaxis], dtype=tf.int32)
|
114 |
-
combined_mask = tf.cast(mask[:, tf.newaxis, :], dtype=tf.int32)
|
115 |
-
combined_mask = tf.minimum(combined_mask, causal_mask)
|
116 |
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
key=embeddings,
|
121 |
-
attention_mask=combined_mask,
|
122 |
-
training=training
|
123 |
-
)
|
124 |
-
out_1 = self.layernorm_1(embeddings + attn_output_1)
|
125 |
|
126 |
-
|
127 |
-
query=out_1,
|
128 |
-
value=encoder_output,
|
129 |
-
key=encoder_output,
|
130 |
-
attention_mask=padding_mask,
|
131 |
-
training=training
|
132 |
-
)
|
133 |
-
out_2 = self.layernorm_2(out_1 + attn_output_2)
|
134 |
|
135 |
-
ffn_out = self.ffn_layer_1(out_2)
|
136 |
-
ffn_out = self.dropout_1(ffn_out, training=training)
|
137 |
-
ffn_out = self.ffn_layer_2(ffn_out)
|
138 |
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
return preds
|
143 |
|
144 |
-
|
145 |
-
|
146 |
-
batch_size, sequence_length = input_shape[0], input_shape[1]
|
147 |
-
i = tf.range(sequence_length)[:, tf.newaxis]
|
148 |
-
j = tf.range(sequence_length)
|
149 |
-
mask = tf.cast(i >= j, dtype="int32")
|
150 |
-
mask = tf.reshape(mask, (1, input_shape[1], input_shape[1]))
|
151 |
-
mult = tf.concat(
|
152 |
-
[tf.expand_dims(batch_size, -1), tf.constant([1, 1], dtype=tf.int32)],
|
153 |
-
axis=0
|
154 |
-
)
|
155 |
-
return tf.tile(mask, mult)
|
156 |
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
self.encoder = encoder
|
162 |
-
self.decoder = decoder
|
163 |
-
self.image_aug = image_aug
|
164 |
-
self.loss_tracker = tf.keras.metrics.Mean(name="loss")
|
165 |
-
self.acc_tracker = tf.keras.metrics.Mean(name="accuracy")
|
166 |
-
|
167 |
-
def calculate_loss(self, y_true, y_pred, mask):
|
168 |
-
loss = self.loss(y_true, y_pred)
|
169 |
-
mask = tf.cast(mask, dtype=loss.dtype)
|
170 |
-
loss *= mask
|
171 |
-
return tf.reduce_sum(loss) / tf.reduce_sum(mask)
|
172 |
-
|
173 |
-
def calculate_accuracy(self, y_true, y_pred, mask):
|
174 |
-
accuracy = tf.equal(y_true, tf.argmax(y_pred, axis=2))
|
175 |
-
accuracy = tf.math.logical_and(mask, accuracy)
|
176 |
-
accuracy = tf.cast(accuracy, dtype=tf.float32)
|
177 |
-
mask = tf.cast(mask, dtype=tf.float32)
|
178 |
-
return tf.reduce_sum(accuracy) / tf.reduce_sum(mask)
|
179 |
-
|
180 |
-
def compute_loss_and_acc(self, img_embed, captions, training=True):
|
181 |
-
encoder_output = self.encoder(img_embed, training=True)
|
182 |
-
y_input = captions[:, :-1]
|
183 |
-
y_true = captions[:, 1:]
|
184 |
-
mask = (y_true != 0)
|
185 |
-
y_pred = self.decoder(
|
186 |
-
y_input, encoder_output, training=True, mask=mask
|
187 |
-
)
|
188 |
-
loss = self.calculate_loss(y_true, y_pred, mask)
|
189 |
-
acc = self.calculate_accuracy(y_true, y_pred, mask)
|
190 |
-
return loss, acc
|
191 |
-
|
192 |
-
def train_step(self, batch):
|
193 |
-
imgs, captions = batch
|
194 |
-
if self.image_aug:
|
195 |
-
imgs = self.image_aug(imgs)
|
196 |
-
img_embed = self.cnn_model(imgs)
|
197 |
-
with tf.GradientTape() as tape:
|
198 |
-
loss, acc = self.compute_loss_and_acc(
|
199 |
-
img_embed, captions
|
200 |
-
)
|
201 |
-
train_vars = (
|
202 |
-
self.encoder.trainable_variables + self.decoder.trainable_variables
|
203 |
-
)
|
204 |
-
grads = tape.gradient(loss, train_vars)
|
205 |
-
self.optimizer.apply_gradients(zip(grads, train_vars))
|
206 |
-
self.loss_tracker.update_state(loss)
|
207 |
-
self.acc_tracker.update_state(acc)
|
208 |
-
return {"loss": self.loss_tracker.result(), "acc": self.acc_tracker.result()}
|
209 |
-
|
210 |
-
def test_step(self, batch):
|
211 |
-
imgs, captions = batch
|
212 |
-
img_embed = self.cnn_model(imgs)
|
213 |
-
loss, acc = self.compute_loss_and_acc(
|
214 |
-
img_embed, captions, training=False
|
215 |
-
)
|
216 |
-
self.loss_tracker.update_state(loss)
|
217 |
-
self.acc_tracker.update_state(acc)
|
218 |
-
return {"loss": self.loss_tracker.result(), "acc": self.acc_tracker.result()}
|
219 |
-
|
220 |
-
@property
|
221 |
-
def metrics(self):
|
222 |
-
return [self.loss_tracker, self.acc_tracker]
|
223 |
-
|
224 |
-
def load_image_from_path(img_path):
|
225 |
-
img = tf.io.read_file(img_path)
|
226 |
-
img = tf.io.decode_jpeg(img, channels=3)
|
227 |
-
img = tf.keras.layers.Resizing(299, 299)(img)
|
228 |
-
img = tf.keras.applications.inception_v3.preprocess_input(img)
|
229 |
-
return img
|
230 |
-
|
231 |
-
def generate_caption(img, caption_model, add_noise=False):
|
232 |
-
if isinstance(img, str):
|
233 |
-
img = load_image_from_path(img)
|
234 |
-
if add_noise:
|
235 |
-
noise = tf.random.normal(img.shape) * 0.1
|
236 |
-
img = (img + noise)
|
237 |
-
img = (img - tf.reduce_min(img)) / (tf.reduce_max(img) - tf.reduce_min(img))
|
238 |
-
img = tf.expand_dims(img, 0) # Add batch dimension
|
239 |
-
img_embed = caption_model.cnn_model(img, training=False)
|
240 |
-
encoder_output = caption_model.encoder(img_embed, training=False)
|
241 |
-
caption = [tokenizer.token_to_id("[START]")]
|
242 |
-
for _ in range(MAX_LENGTH):
|
243 |
-
input_caption = tf.convert_to_tensor([caption], dtype=tf.int32)
|
244 |
-
pred = caption_model.decoder(input_caption, encoder_output, training=False)
|
245 |
-
pred = tf.argmax(pred[0, -1, :]).numpy()
|
246 |
-
caption.append(pred)
|
247 |
-
if pred == tokenizer.token_to_id("[END]"):
|
248 |
-
break
|
249 |
-
return ' '.join([idx2word(word).numpy().decode('utf-8') for word in caption[1:-1]])
|
250 |
-
|
251 |
-
# Load saved model weights
|
252 |
-
cnn_model = CNN_Encoder()
|
253 |
-
encoder = TransformerEncoderLayer(embed_dim=EMBEDDING_DIM, num_heads=8)
|
254 |
-
decoder = TransformerDecoderLayer(embed_dim=EMBEDDING_DIM, units=UNITS, num_heads=8)
|
255 |
-
caption_model = ImageCaptioningModel(cnn_model=cnn_model, encoder=encoder, decoder=decoder)
|
256 |
-
caption_model.load_weights('saved_model_weights/caption_model')
|
257 |
-
|
258 |
-
# Streamlit App
|
259 |
-
st.title('Image Captioning with Transformer')
|
260 |
-
st.write('Upload an image to generate a caption.')
|
261 |
-
|
262 |
-
uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])
|
263 |
-
|
264 |
-
if uploaded_file is not None:
|
265 |
-
image = Image.open(uploaded_file)
|
266 |
-
st.image(image, caption='Uploaded Image', use_column_width=True)
|
267 |
-
st.write("")
|
268 |
-
st.write("Generating caption...")
|
269 |
|
270 |
-
|
271 |
-
|
272 |
-
|
273 |
-
|
274 |
-
|
275 |
-
|
276 |
-
|
277 |
-
|
278 |
-
|
279 |
-
|
280 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import io
|
2 |
+
import os
|
3 |
import streamlit as st
|
4 |
import requests
|
5 |
from PIL import Image
|
6 |
+
from model import get_caption_model, generate_caption
|
7 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
|
9 |
+
@st.cache(allow_output_mutation=True)
|
10 |
+
def get_model():
|
11 |
+
return get_caption_model()
|
|
|
|
|
|
|
|
|
|
|
12 |
|
13 |
+
caption_model = get_model()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
|
|
|
|
|
|
15 |
|
16 |
+
def predict():
|
17 |
+
captions = []
|
18 |
+
pred_caption = generate_caption('tmp.jpg', caption_model)
|
|
|
19 |
|
20 |
+
st.markdown('#### Predicted Captions:')
|
21 |
+
captions.append(pred_caption)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
+
for _ in range(4):
|
24 |
+
pred_caption = generate_caption('tmp.jpg', caption_model, add_noise=True)
|
25 |
+
if pred_caption not in captions:
|
26 |
+
captions.append(pred_caption)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
|
28 |
+
for c in captions:
|
29 |
+
st.write(c)
|
30 |
+
|
31 |
+
st.title('Image Captioner')
|
32 |
+
img_url = st.text_input(label='Enter Image URL')
|
33 |
+
|
34 |
+
if (img_url != "") and (img_url != None):
|
35 |
+
img = Image.open(requests.get(img_url, stream=True).raw)
|
36 |
+
img = img.convert('RGB')
|
37 |
+
st.image(img)
|
38 |
+
img.save('tmp.jpg')
|
39 |
+
predict()
|
40 |
+
os.remove('tmp.jpg')
|
41 |
+
|
42 |
+
|
43 |
+
st.markdown('<center style="opacity: 70%">OR</center>', unsafe_allow_html=True)
|
44 |
+
img_upload = st.file_uploader(label='Upload Image', type=['jpg', 'png', 'jpeg'])
|
45 |
+
|
46 |
+
if img_upload != None:
|
47 |
+
img = img_upload.read()
|
48 |
+
img = Image.open(io.BytesIO(img))
|
49 |
+
img = img.convert('RGB')
|
50 |
+
img.save('tmp.jpg')
|
51 |
+
st.image(img)
|
52 |
+
predict()
|
53 |
+
os.remove('tmp.jpg')
|