Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -7,50 +7,113 @@ from tensorflow.keras.preprocessing.image import img_to_array
|
|
7 |
from tensorflow.keras.preprocessing.sequence import pad_sequences
|
8 |
import pickle
|
9 |
|
10 |
-
# Custom
|
11 |
-
|
12 |
-
|
13 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
|
15 |
def compute_output_shape(self, input_shape):
|
16 |
-
if self.
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
return input_shape
|
21 |
-
return
|
22 |
|
23 |
-
# Define custom objects
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
|
|
|
|
|
|
|
|
28 |
|
29 |
-
# Multiple loading strategies
|
30 |
def load_model_safely():
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
lambda: tf.keras.models.load_model("caption_model.h5", compile=False),
|
36 |
-
# Strategy 3: Load with different custom objects
|
37 |
-
lambda: tf.keras.models.load_model("caption_model.h5",
|
38 |
-
custom_objects={'Lambda': tf.keras.layers.Lambda}),
|
39 |
]
|
40 |
|
41 |
-
for i,
|
42 |
try:
|
43 |
-
|
44 |
-
|
|
|
|
|
45 |
return model
|
46 |
except Exception as e:
|
47 |
-
print(f"
|
48 |
continue
|
49 |
|
50 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
51 |
|
52 |
# Load your pre-trained model and tokenizer
|
53 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
|
55 |
with open("tokenizer.pkl", "rb") as handle:
|
56 |
tokenizer = pickle.load(handle)
|
@@ -62,6 +125,9 @@ feature_extractor = tf.keras.Model(feature_extractor.input, feature_extractor.la
|
|
62 |
# Description generation function
|
63 |
def generate_caption(image):
|
64 |
try:
|
|
|
|
|
|
|
65 |
# Preprocess the image
|
66 |
image = image.resize((224, 224))
|
67 |
image = img_to_array(image)
|
@@ -84,7 +150,7 @@ def generate_caption(image):
|
|
84 |
yhat = np.argmax(yhat)
|
85 |
except Exception as e:
|
86 |
print(f"Prediction error: {e}")
|
87 |
-
return "Error
|
88 |
|
89 |
word = ''
|
90 |
for w, i in tokenizer.word_index.items():
|
@@ -97,7 +163,7 @@ def generate_caption(image):
|
|
97 |
input_text += ' ' + word
|
98 |
|
99 |
caption = input_text.replace('startseq', '').strip()
|
100 |
-
return caption
|
101 |
|
102 |
except Exception as e:
|
103 |
return f"Error processing image: {str(e)}"
|
|
|
7 |
from tensorflow.keras.preprocessing.sequence import pad_sequences
|
8 |
import pickle
|
9 |
|
10 |
+
# Custom function to handle attention mechanism
|
11 |
+
def attention_function(inputs):
|
12 |
+
"""
|
13 |
+
Custom attention function that likely combines two inputs
|
14 |
+
Input 1: (None, 34, 34) - attention weights
|
15 |
+
Input 2: (None, 34, 512) - feature vectors
|
16 |
+
Output: (None, 34, 512) - attended features
|
17 |
+
"""
|
18 |
+
attention_weights, features = inputs
|
19 |
+
# Expand attention weights to match feature dimensions
|
20 |
+
attention_weights = tf.expand_dims(attention_weights, axis=-1)
|
21 |
+
# Apply attention weights to features
|
22 |
+
attended_features = attention_weights * features
|
23 |
+
return attended_features
|
24 |
+
|
25 |
+
def attention_output_shape(input_shapes):
|
26 |
+
"""Define the output shape for attention mechanism"""
|
27 |
+
# Return the shape of the feature input (second input)
|
28 |
+
return input_shapes[1] # (None, 34, 512)
|
29 |
+
|
30 |
+
# Alternative attention functions to try
|
31 |
+
def attention_function_v2(inputs):
|
32 |
+
"""Alternative attention mechanism - weighted sum"""
|
33 |
+
attention_weights, features = inputs
|
34 |
+
# Normalize attention weights
|
35 |
+
attention_weights = tf.nn.softmax(attention_weights, axis=-1)
|
36 |
+
attention_weights = tf.expand_dims(attention_weights, axis=-1)
|
37 |
+
return attention_weights * features
|
38 |
+
|
39 |
+
def attention_function_v3(inputs):
|
40 |
+
"""Another alternative - dot product attention"""
|
41 |
+
attention_weights, features = inputs
|
42 |
+
# Sum along the second dimension of attention weights
|
43 |
+
attention_weights = tf.reduce_sum(attention_weights, axis=-1, keepdims=True)
|
44 |
+
attention_weights = tf.expand_dims(attention_weights, axis=-1)
|
45 |
+
return attention_weights * features
|
46 |
+
|
47 |
+
# Custom Lambda layer class
|
48 |
+
class AttentionLambda(tf.keras.layers.Lambda):
|
49 |
+
def __init__(self, function, output_shape_func=None, **kwargs):
|
50 |
+
super().__init__(function, **kwargs)
|
51 |
+
self.output_shape_func = output_shape_func
|
52 |
|
53 |
def compute_output_shape(self, input_shape):
|
54 |
+
if self.output_shape_func:
|
55 |
+
return self.output_shape_func(input_shape)
|
56 |
+
# Default: return the shape of the second input (features)
|
57 |
+
if isinstance(input_shape, list) and len(input_shape) >= 2:
|
58 |
+
return input_shape[1]
|
59 |
+
return input_shape
|
60 |
|
61 |
+
# Define multiple custom objects to try different attention mechanisms
|
62 |
+
def get_custom_objects(attention_func, output_shape_func):
|
63 |
+
return {
|
64 |
+
'Lambda': lambda function=None, **kwargs: AttentionLambda(
|
65 |
+
attention_func if function is None else function,
|
66 |
+
output_shape_func,
|
67 |
+
**kwargs
|
68 |
+
)
|
69 |
+
}
|
70 |
|
71 |
+
# Multiple loading strategies with different attention mechanisms
|
72 |
def load_model_safely():
|
73 |
+
attention_strategies = [
|
74 |
+
(attention_function, attention_output_shape),
|
75 |
+
(attention_function_v2, attention_output_shape),
|
76 |
+
(attention_function_v3, attention_output_shape),
|
|
|
|
|
|
|
|
|
77 |
]
|
78 |
|
79 |
+
for i, (att_func, shape_func) in enumerate(attention_strategies, 1):
|
80 |
try:
|
81 |
+
print(f"Trying attention strategy {i}...")
|
82 |
+
custom_objects = get_custom_objects(att_func, shape_func)
|
83 |
+
model = tf.keras.models.load_model("caption_model.h5", custom_objects=custom_objects)
|
84 |
+
print(f"Model loaded successfully using attention strategy {i}!")
|
85 |
return model
|
86 |
except Exception as e:
|
87 |
+
print(f"Attention strategy {i} failed: {e}")
|
88 |
continue
|
89 |
|
90 |
+
# If all attention strategies fail, try loading without compilation
|
91 |
+
try:
|
92 |
+
print("Trying to load without compilation...")
|
93 |
+
model = tf.keras.models.load_model("caption_model.h5", compile=False)
|
94 |
+
print("Model loaded without compilation!")
|
95 |
+
return model
|
96 |
+
except Exception as e:
|
97 |
+
print(f"Loading without compilation failed: {e}")
|
98 |
+
|
99 |
+
# Last resort: try to load and rebuild the model
|
100 |
+
try:
|
101 |
+
print("Attempting to load model weights only...")
|
102 |
+
# This is a more complex approach that would require knowing the model architecture
|
103 |
+
raise Exception("Model architecture reconstruction needed")
|
104 |
+
except:
|
105 |
+
pass
|
106 |
+
|
107 |
+
raise Exception("All loading strategies failed. The model may need to be retrained or converted.")
|
108 |
|
109 |
# Load your pre-trained model and tokenizer
|
110 |
+
try:
|
111 |
+
model = load_model_safely()
|
112 |
+
except Exception as e:
|
113 |
+
print(f"Failed to load model: {e}")
|
114 |
+
print("Creating a dummy model for testing...")
|
115 |
+
# Create a simple dummy model for testing the interface
|
116 |
+
model = None
|
117 |
|
118 |
with open("tokenizer.pkl", "rb") as handle:
|
119 |
tokenizer = pickle.load(handle)
|
|
|
125 |
# Description generation function
|
126 |
def generate_caption(image):
|
127 |
try:
|
128 |
+
if model is None:
|
129 |
+
return "Model failed to load. Please check the model file."
|
130 |
+
|
131 |
# Preprocess the image
|
132 |
image = image.resize((224, 224))
|
133 |
image = img_to_array(image)
|
|
|
150 |
yhat = np.argmax(yhat)
|
151 |
except Exception as e:
|
152 |
print(f"Prediction error: {e}")
|
153 |
+
return f"Error during prediction: {str(e)}"
|
154 |
|
155 |
word = ''
|
156 |
for w, i in tokenizer.word_index.items():
|
|
|
163 |
input_text += ' ' + word
|
164 |
|
165 |
caption = input_text.replace('startseq', '').strip()
|
166 |
+
return caption if caption else "Unable to generate caption"
|
167 |
|
168 |
except Exception as e:
|
169 |
return f"Error processing image: {str(e)}"
|