dilating drawings to better match training data

app.py

@@ -1,24 +1,25 @@
 import gradio as gr
 import tensorflow as tf
 import numpy as np
+import cv2
 
 model = tf.keras.models.load_model("mymodel/mymodel")
 
 def predict(img):
     z = tf.keras.preprocessing.image.img_to_array(img)
+    kernel = np.ones((5,5),np.uint8)
+    z = cv2.dilate(z ,kernel,iterations = 1)
     z = np.expand_dims(z, axis=0)
-    print(z.shape, z)
     y = model.predict(z)
     ysoft = tf.nn.softmax(y)
     ymax = np.argmax(ysoft)
-    return ymax
+    return int(ymax)
 
-sp = gr.Sketchpad(shape=(140,100), image_mode="L", label='arabic numeral', invert_colors=False).style(height=200, width=280)
+sp = gr.Sketchpad(tool="sketch", shape=(140,100), image_mode="L", label='arabic numeral', invert_colors=False).style(height=200, width=280)
 gr.Label()
 gr.Interface(fn=predict, 
              inputs=sp,
              outputs="label",
-             share=True,
              live=True,
              examples=[
              ["writer001_pass01_digit2.png"],

requirements.txt

@@ -1,2 +1,3 @@
 numpy == 1.21.3
-tensorflow == 2.10.1
+tensorflow == 2.10.1
+cv2