Update app.py

app.py

@@ -10,6 +10,33 @@ vgg.trainable = False
 STYLE_LAYERS = [...]  # same layers as in your notebook
 CONTENT_LAYER = [...] # e.g. [('block5_conv4', 1)]
 
+
+def gram_matrix(A):
+    """Compute Gram matrix for style representation."""
+    A = tf.transpose(A, (0, 3, 1, 2))   # (batch, channels, height, width)
+    features = tf.reshape(A, (A.shape[0], A.shape[1], -1))
+    gram = tf.matmul(features, features, transpose_b=True)
+    return gram / tf.cast(tf.shape(features)[-1], tf.float32)
+
+def compute_content_cost(a_C, a_G):
+    """Content cost between content and generated image features."""
+    return tf.reduce_mean(tf.square(a_C - a_G))
+
+def compute_style_cost(a_S, a_G):
+    """Style cost using Gram matrices."""
+    J_style = 0
+    for s, g in zip(a_S, a_G):
+        J_style += tf.reduce_mean(tf.square(gram_matrix(s) - gram_matrix(g)))
+    return J_style / len(a_S)
+
+def total_cost(J_content, J_style, alpha=10, beta=40):
+    """Weighted sum of content + style."""
+    return alpha * J_content + beta * J_style
+
+def clip_0_1(img):
+    """Keep pixel values in [0,1]."""
+    return tf.clip_by_value(img, 0.0, 1.0)
+
 def preprocess(img):
     img = Image.fromarray(img).resize((256, 256))
     arr = np.expand_dims(np.array(img) / 255.0, axis=0).astype(np.float32)