manupa-arm commented on a change in pull request #9508:
URL: https://github.com/apache/tvm/pull/9508#discussion_r749232944
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -48,122 +46,147 @@ def get_shape_expr(in_expr, out_expr):
return shape
[email protected](
- "accel_type",
- ACCEL_TYPES,
-)
-def test_ethosu_conv2d(accel_type):
- def create_graph_single(input_tensor_name, input_tensor_shape,
input_tensor_dtype):
- c1_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c1_params.ifm.shape = input_tensor_shape
- c1_params.kernel.shape = (3, 3, c1_params.ifm.shape[3], 32)
- c1_params.kernel.sc = relay.const(np.random.rand(32) * 2, "float32")
- c1_params.strides = (1, 1)
- c1_params.pad = "VALID"
- c1_params.update_output_qnn_params(
- input_tensor_dtype, input_tensor_dtype, input_tensor_dtype
- )
- input0 = relay.var(input_tensor_name, shape=c1_params.ifm.shape,
dtype=c1_params.ifm.dtype)
- c1, new_params = relay_ir_builder.create_qnn_conv2d(c1_params, input0)
- c1_params.ofm.shape = get_shape_expr(input0, c1)
[email protected]("ifm_shape", [(1, 299, 299, 3), (1, 55, 55, 3)])
[email protected]("kernel_shape", [(3, 2, 3, 3), (1, 3, 3, 3)])
[email protected]("padding", ["SAME", "VALID"])
[email protected]("accel_type", ACCEL_TYPES)
+def test_ethosu_conv2d(ifm_shape, kernel_shape, padding, accel_type):
+ dtype = "int8"
- f = relay.Function([input0], c1)
- mod = tvm.IRModule()
- mod["main"] = f
- return mod, [c1_params]
-
- def create_graph_double(input_tensor_name, input_tensor_shape,
input_tensor_dtype):
- c1_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c1_params.ifm.shape = input_tensor_shape
- c1_params.kernel.shape = (7, 7, c1_params.ifm.shape[3], 8)
- c1_params.strides = (2, 2)
- c1_params.pad = "VALID"
- c1_params.update_output_qnn_params(
- input_tensor_dtype, input_tensor_dtype, input_tensor_dtype
- )
- input0 = relay.var(input_tensor_name, shape=c1_params.ifm.shape,
dtype=c1_params.ifm.dtype)
- c1, new_params = relay_ir_builder.create_qnn_conv2d(c1_params, input0)
- c1_params.ofm.shape = get_shape_expr(input0, c1)
-
- c2_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c2_params.ifm.shape = c1_params.ofm.shape
- c2_params.kernel.shape = (5, 5, c2_params.ifm.shape[3], 16)
- c2_params.strides = (1, 1)
- c2_params.pad = "SAME"
- c2_params.update_output_qnn_params()
- c2, new_params = relay_ir_builder.create_qnn_conv2d(c2_params, c1)
- c2_params.ofm.shape = get_shape_expr(input0, c2)
-
- f = relay.Function([input0], c2)
- mod = tvm.IRModule()
- mod["main"] = f
- return mod, [c2_params, c1_params]
-
- def create_graph_activation(input_tensor_name, input_tensor_shape,
input_tensor_dtype):
- c1_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c1_params.ifm.shape = input_tensor_shape
- c1_params.kernel.shape = (7, 7, c1_params.ifm.shape[3], 8)
- c1_params.strides = (2, 2)
- c1_params.pad = "VALID"
- c1_params.activation = "CLIP"
- c1_params.clip_min = 90
- c1_params.clip_max = 110
- c1_params.update_output_qnn_params(
- input_tensor_dtype, input_tensor_dtype, input_tensor_dtype
+ def create_tflite_graph_single():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x):
+ # Use tf.nn API to create the model
+ op = tf.nn.conv2d(
+ x,
+ filters=tf.constant(np.random.uniform(size=kernel_shape),
dtype=tf.float32),
+ strides=(1, 1),
+ padding=padding,
+ data_format="NHWC",
+ dilations=1,
+ )
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32)
)
- input0 = relay.var(input_tensor_name, shape=c1_params.ifm.shape,
dtype=c1_params.ifm.dtype)
- c1, new_params = relay_ir_builder.create_qnn_conv2d(c1_params, input0)
- c1_params.ofm.shape = get_shape_expr(input0, c1)
-
- c2_params = relay_ir_builder.QnnConv2DParams(input_tensor_dtype)
- c2_params.ifm.shape = c1_params.ofm.shape
- c2_params.kernel.shape = (5, 5, c2_params.ifm.shape[3], 16)
- c2_params.strides = (1, 1)
- c2_params.pad = "SAME"
- c2_params.update_output_qnn_params()
- c2, new_params = relay_ir_builder.create_qnn_conv2d(c2_params, c1)
- c2_params.ofm.shape = get_shape_expr(input0, c2)
-
- f = relay.Function([input0], c2)
- mod = tvm.IRModule()
- mod["main"] = f
- return mod, [c2_params, c1_params]
-
- test_cases = [
- (create_graph_single, ["input", (1, 300, 300, 3), "int8"]),
- (create_graph_double, ["input", (1, 128, 256, 4), "int8"]),
- (create_graph_activation, ["input", (1, 64, 100, 4), "int8"]),
- ]
- np.random.seed(42)
- for test_case in test_cases:
- relay_module, conv_params = test_case[0](*test_case[1])
- input_tensor, input_shape, input_dtype = test_case[1]
- mod = partition_for_ethosu(relay_module)
-
- # Generate reference data
- in_min, in_max = util.get_range_for_dtype_str(input_dtype)
- input_data = {
- input_tensor: np.random.randint(
- in_min, high=in_max, size=input_shape, dtype=input_dtype
- )
- }
- output_data = generate_ref_data(relay_module, input_data)
-
- compiled_models = infra.build_source(
- mod, input_data, output_data, accel_type, output_tolerance=1
+
+ # 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_model = converter.convert()
+ return tflite_model
+
+ def create_tflite_graph_double():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function_double(self, x):
+ # Use tf.nn API to create the model with two convolutions
+ op = tf.nn.conv2d(
+ x,
+ filters=tf.constant(np.random.uniform(size=kernel_shape),
dtype=tf.float32),
+ strides=(1, 1),
+ padding=padding,
+ data_format="NHWC",
+ dilations=1,
+ )
+ # Second convolution
+ op2 = tf.nn.conv2d(
+ op,
+ filters=tf.constant(np.random.uniform(size=kernel_shape),
dtype=tf.float32),
+ strides=(1, 1),
+ padding=padding,
+ data_format="NHWC",
+ dilations=2,
+ )
+ return op2
+
+ model = Model()
+ concrete_func = model.tf_function_double.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32)
)
- # Assumes only two runtime.Modules are created -- i.e. single offload
module
- imported_modules =
compiled_models[0].executor_factory.lib.imported_modules
- assert len(imported_modules) == 2
- ethosu_module = imported_modules[0]
-
- # Verify generated C source
- get_cs = tvm._ffi.get_global_func("runtime.module.ethosu.getcs")
- cmms = get_cs(ethosu_module)
- cmms = bytes.fromhex(cmms)
- infra.print_payload(cmms)
- infra.verify_source(compiled_models, accel_type)
+ # 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_model = converter.convert()
+ return tflite_model
+
+ tflite_graph_single = create_tflite_graph_single()
+ tflite_model_single =
tflite.Model.Model.GetRootAsModel(tflite_graph_single, 0)
+
+ tflite_graph_double = create_tflite_graph_double()
+ tflite_model_double =
tflite.Model.Model.GetRootAsModel(tflite_graph_double, 0)
Review comment:
I think tflite_model_double and tflite_model_single beyond this point
has the same code, if I am not mistaken. Could we refactor that bit to re-use
the same code ?
--
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]
