testdata/dnn/tensorflow/generate_tf2_models.py

@@ -76,9 +76,9 @@ model = tf.keras.models.Sequential([
 save(model, 'tf2_dense', flatten_input=tf.TensorSpec(shape=[None, 1, 2, 3], dtype=tf.float32))
 ################################################################################
 model = tf.keras.models.Sequential([
-  tf.keras.layers.PReLU(input_shape=(1, 2, 3)),
+  tf.keras.layers.PReLU(input_shape=(1, 4, 6), alpha_initializer='random_normal'),
 ])
-save(model, 'tf2_prelu', p_re_lu_input=tf.TensorSpec(shape=[None, 1, 2, 3], dtype=tf.float32))
+save(model, 'tf2_prelu', p_re_lu_input=tf.TensorSpec(shape=[None, 1, 4, 6], dtype=tf.float32))
 ################################################################################
 model = tf.keras.models.Sequential([
   tf.keras.layers.AveragePooling2D(input_shape=(4, 6, 3), pool_size=(2, 2)),

testdata/dnn/tflite/generate.py

@@ -45,3 +45,33 @@ run_tflite_model("face_landmark", (192, 192))
 run_tflite_model("face_detection_short_range", (128, 128))
 
 run_mediapipe_solution(mp.solutions.selfie_segmentation.SelfieSegmentation(model_selection=0), (256, 256))
+
+# Save TensorFlow model as TFLite
+def save_tflite_model(model, inp, name):
+    func = model.get_concrete_function()
+    converter = tf.lite.TFLiteConverter.from_concrete_functions([func])
+    tflite_model = converter.convert()
+
+    interpreter = tf.lite.Interpreter(model_content=tflite_model)
+
+    with open(f'{name}.tflite', 'wb') as f:
+        f.write(tflite_model)
+
+    out = model(inp)
+
+    np.save(f'{name}_inp.npy', inp.transpose(0, 3, 1, 2))
+    np.save(f'{name}_out_Identity.npy', np.array(out).transpose(0, 3, 1, 2))
+
+
+@tf.function(input_signature=[tf.TensorSpec(shape=[1, 3, 3, 1], dtype=tf.float32)])
+def replicate_by_pack(x):
+    pack_1 = tf.stack([x, x], axis=3)
+    reshape_1 = tf.reshape(pack_1, [1, 3, 6, 1])
+    pack_2 = tf.stack([reshape_1, reshape_1], axis=2)
+    reshape_2 = tf.reshape(pack_2, [1, 6, 6, 1])
+    scaled = tf.image.resize(reshape_2, size=(3, 3), method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
+    return scaled + x
+
+inp = np.random.standard_normal((1, 3, 3, 1)).astype(np.float32)
+save_tflite_model(replicate_by_pack, inp, 'replicate_by_pack')
+