Create app.py

app.py

@@ -0,0 +1,55 @@
+import json
+
+import gradio as gr
+import tensorflow as tf
+import tensorflow.keras as keras
+from gradio import inputs, outputs
+
+SIZE = 256
+DEVICE = "/CPU:0"
+
+
+with open("./tags.json", "rt", encoding="utf-8") as f:
+    tags = json.load(f)
+
+
+with tf.device(DEVICE):
+    base_model = keras.applications.resnet.ResNet50(
+        include_top=False, weights=None, input_shape=(SIZE, SIZE, 3)
+    )
+    model = keras.Sequential(
+        [
+            base_model,
+            keras.layers.Conv2D(filters=len(tags), kernel_size=(1, 1), padding="same"),
+            keras.layers.BatchNormalization(epsilon=1.001e-5),
+            keras.layers.GlobalAveragePooling2D(name="avg_pool"),
+            keras.layers.Activation("sigmoid"),
+        ]
+    )
+    model.load_weights("tf_model.h5")
+
+
+@tf.function
+def process_data(content):
+    img = tf.io.decode_jpeg(content, channels=3)
+    img = tf.image.resize_with_pad(img, SIZE, SIZE)
+    img = tf.image.per_image_standardization(img)
+    return img
+
+
+def predict(img, size):
+    with tf.device(DEVICE):
+        img = tf.image.resize_with_pad(img, size, size)
+        img = tf.image.per_image_standardization(img)
+        data = process_data(image)
+        data = tf.expand_dims(data, 0)
+        out = model(data)[0]
+    return dict((tags[i], out[i].numpy()) for i in range(len(tags)))
+
+
+image = inputs.Image(label="Upload your image here!")
+size = inputs.Number(label="Image resize", default=SIZE)
+
+labels = outputs.Label(label="Tags")
+
+gr.Interface(predict, inputs=[image, size], outputs=[labels])