ekalda commented on a change in pull request #9576:
URL: https://github.com/apache/tvm/pull/9576#discussion_r759508529
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -435,6 +435,110 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
[email protected](
+ "accel_type",
+ ACCEL_TYPES,
+)
[email protected](
+ "ifm_shape, axis, keep_dims, use_same_quantization",
+ [
+ # mean to depthwise + multiply
+ [(1, 8, 16, 16), (1, 2), True, False],
+ [(1, 3, 4), (0, 1), True, False],
+ [(1, 65, 2, 1), (1, 2), True, False], # special case when h > 64
+ # mean to average pool
+ [(1, 8, 16, 16), (2,), False, True],
+ [(3, 3, 4), (0,), True, True],
+ [(8, 5), (0,), False, True],
+ # mean to depthwise
+ [(1, 8, 16, 16), (2,), True, False],
+ [(1, 8, 16, 16), (2, 1), False, False],
+ [(8, 4), (0,), False, False],
+ ],
+)
+def test_mean(accel_type, ifm_shape, axis, keep_dims, use_same_quantization):
+ dtype = "int8"
+
+ def create_mod_from_tflite():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x):
+ op = tf.math.reduce_mean(x, axis=axis, keepdims=keep_dims)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ yield [data.astype(np.float32)]
+
+ converter =
tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops =
[tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_graph = converter.convert()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"ifm": ifm_shape},
+ dtype_dict={"ifm": dtype},
+ )
+ input_data, output_data = infra.generate_ref_data_tflite(tflite_graph)
+ return mod, input_data, output_data
+
+ def create_mod_from_relay():
+ ifm = relay.var("input", shape=ifm_shape, dtype=dtype)
+ cast = relay.cast(ifm, dtype="int32")
+ mean = relay.mean(cast, axis=axis, keepdims=keep_dims)
+ requantize = relay.qnn.op.requantize(
+ mean,
+ input_scale=relay.const(1.0, dtype="float32"),
+ input_zero_point=relay.const(0, dtype="int32"),
+ output_scale=relay.const(1.0, dtype="float32"),
+ output_zero_point=relay.const(0, dtype="int32"),
+ )
+
+ func = relay.Function(relay.analysis.free_vars(requantize), requantize)
+ mod = tvm.IRModule.from_expr(func)
+
+ input_data = {"input": np.random.randint(low=-127, high=128,
size=ifm_shape, dtype=dtype)}
+ output_data = generate_ref_data(mod, input_data)
+ return mod, input_data, output_data
+
+ mod, input_data, output_data = (
+ create_mod_from_relay() if use_same_quantization else
create_mod_from_tflite()
+ )
Review comment:
Cool, thanks for clarifying this! Yeah I think creating the graph
directly in Relay is fine, especially since TVM is not very adventurous in
lowering TFLite's mean :D
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
