def export_coreml(model, im, file, prefix=colorstr('CoreML:')):
# YOLOv5 CoreML export
ct_model = None
try:
check_requirements(('coremltools',))
import coremltools as ct
print(f'\n{prefix} starting export with coremltools {ct.__version__}...')
f = file.with_suffix('.mlmodel')
model.train() # CoreML exports should be placed in model.train() mode
ts = torch.jit.trace(model, im, strict=False) # TorchScript model
ct_model = ct.convert(ts, inputs=[ct.ImageType('image', shape=im.shape, scale=1 / 255.0, bias=[0, 0, 0])])
ct_model.save(f)
print(f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)')
except Exception as e:
print(f'\n{prefix} export failure: {e}')
return ct_model
def test_mil_default_value_runtime(self):
program_input_spec = [
ct.TensorType(name="x",
shape=[1],
default_value=np.array([1.0]).astype(np.float32)),
ct.TensorType(name="y", shape=[1])
]
mlmodel = ct.convert(self.basic_network,
convert_to="mlprogram",
inputs=program_input_spec)
if _macos_version() < (12, 0):
# Can only get predictions for ml program on macOS 12+
return
res = mlmodel.predict({"x": np.array([3.]), "y": np.array([2.])})
assert res["output"][0] == 5.0
res = mlmodel.predict({"y": np.array([2.])})
assert res["output"][0] == 3.0
def test():
torch_model = torchvision.models.mobilenet_v2(pretrained=True)
# Set the model in evaluation mode
# torch_model.eval()
# example_input = torch.rand(1, 3, 224, 224) # after test, will get 'size mismatch' error message with size 256x256
# traced_model = torch.jit.trace(torch_model, example_input)
fastpunct = FastPunct()
traced_model = fastpunct.model
traced_model.save('aaa.h5')
# Convert to Core ML using the Unified Conversion API
model = ct.convert(
traced_model,
'pytorch',
inputs=[ct.ImageType(name="input_1", shape=example_input.shape)
] # provide only if step 2 was performed
)
# Save model
model.save("MobileNetV2.mlmodel")
def test_tf2keras_enumerated_shapes():
# Test examples in https://coremltools.readme.io/docs/flexible-inputs
import tensorflow as tf
input_shape = (28, 28, 3)
# None denotes seq_len dimension
x = tf.keras.Input(shape=input_shape, name="input")
C_out = 2
kHkW = 3
y = tf.keras.layers.Conv2D(C_out,
kHkW,
activation='relu',
input_shape=input_shape)(x)
keras_model = tf.keras.Model(inputs=[x], outputs=[y])
# One RangeDim shared by two inputs
shapes = [(1, 28, 28, 3), (1, 56, 56, 3)]
enumerated_shapes = ct.EnumeratedShapes(shapes=shapes)
tensor_input = ct.TensorType(name="input", shape=enumerated_shapes)
mlmodel = ct.convert(keras_model, inputs=[tensor_input])
# Test (1, 28, 28, 3) shape
test_input_x = np.random.rand(*shapes[0]).astype(np.float32)
expected_val = keras_model([test_input_x])
if ct.utils._is_macos():
results = mlmodel.predict({"input": test_input_x})
np.testing.assert_allclose(results["Identity"],
expected_val,
rtol=1e-2)
# Test (1, 56, 56, 3) shape (can't verify numerical parity with Keras
# which doesn't support enumerated shape)
test_input_x = np.random.rand(*shapes[1]).astype(np.float32)
results = mlmodel.predict({"input": test_input_x})
# Test with a wrong shape
with pytest.raises(
RuntimeError,
match=r"not compatible with the model\'s feature"):
test_input_x = np.random.rand(1, 29, 29, 3).astype(np.float32)
results = mlmodel.predict({"input": test_input_x})
def test_tf2_image_enumerated_shapes():
import tensorflow as tf
keras_model = tf.keras.applications.MobileNetV2(
input_shape=(
None,
None,
3,
),
classes=1000,
include_top=False,
)
input_shapes = ct.EnumeratedShapes(shapes=[(1, 192, 192,
3), (1, 224, 224, 3)])
image_input = ct.ImageType(shape=input_shapes,
bias=[-1, -1, -1],
scale=1 / 127)
model = ct.convert(keras_model, inputs=[image_input])
assert model is not None
spec = model.get_spec()
assert len(spec.description.input[0].type.imageType.enumeratedSizes.
sizes) == 2
def test_classifier(self):
torch_model = torch.nn.ReLU().eval()
traced_model = torch.jit.trace(torch_model, torch.rand(3, ))
model = ct.convert(
traced_model,
inputs=[ct.TensorType(shape=(3, ), dtype=np.float16)],
outputs=[ct.TensorType(dtype=np.float16)],
classifier_config=ct.ClassifierConfig(['a', 'b', 'c']),
convert_to='mlprogram',
minimum_deployment_target=ct.target.macOS13,
)
assert_input_dtype(model, expected_type_str="fp16")
assert_ops_in_mil_program(model, ["relu", "cast", "classify"])
spec = model.get_spec()
input_name = spec.description.input[0].name
out_dict = model.predict({input_name: np.array([1.0, 2.0, 3.0])})
assert 'classLabel' in out_dict
assert out_dict['classLabel'] == 'c'
assert len(spec.description.output) == 2
assert "classLabel_probs" in out_dict
assert isinstance(out_dict["classLabel_probs"], dict)
def test_convert_tf_keras_applications_model():
import tensorflow as tf
tf_keras_model = tf.keras.applications.MobileNet(
weights="imagenet", input_shape=(224, 224, 3)
)
# inputs / outputs are optional, we can get from tf.keras model
# this can be extremely helpful when we want to extract sub-graphs
input_name = tf_keras_model.inputs[0].name.split(":")[0]
# note that the `convert()` requires tf.Graph's outputs instead of
# tf.keras.Model's outputs, to access that, we can do the following
output_name = tf_keras_model.outputs[0].name.split(":")[0]
tf_graph_output_name = output_name.split("/")[-1]
mlmodel = ct.convert(
tf_keras_model,
inputs=[ct.TensorType(shape=(1, 224, 224, 3))],
outputs=[tf_graph_output_name],
)
mlmodel.save("./mobilenet.mlmodel")
def test_fusion_with_image_full(self):
# Avoid circular import
from coremltools import convert
@mb.program(input_specs=[mb.TensorSpec(shape=(10, 20, 30, 3))])
def prog(x):
x1 = mb.transpose(x=x, perm=[0, 3, 1, 2])
x2 = mb.relu(x=x)
x3 = mb.transpose(x=x2, perm=[0, 3, 1, 2])
x4 = mb.add(x=x1, y=x3)
return mb.relu(x=x4)
mlmodel = convert(prog,
inputs=[
ImageType(name="x",
shape=(10, 20, 30, 3),
channel_first=False)
],
source="milinternal",
convert_to="neuralnetwork")
assert mlmodel is not None
assert len(mlmodel.get_spec().neuralNetwork.layers) == 3
def test_mil_enumerated_multiarray_with_default(self):
enumerated_shapes = tuple([(1, 3, 10, 10), (1, 3, 10, 20),
(1, 3, 10, 30)])
input_shape = [
ct.TensorType(name="x",
shape=ct.EnumeratedShapes(shapes=enumerated_shapes,
default=(1, 3, 10, 30)))
]
mlmodel = ct.convert(self.basic_network,
source="milinternal",
convert_to="mlprogram",
inputs=input_shape)
input_spec = mlmodel.get_spec().description.input
assert len(input_spec) == 1, "1 input expected, got {} instead".format(
len(input_spec))
assert input_spec[
0].name == "x", "input name in MLModel is {}, 'x' is expected".format(
input_spec[0].name)
assert input_spec[0].type.WhichOneof(
"Type"
) == "multiArrayType", "Expected multiArrayType, got {}".format(
input_spec[0].type.WhichOneof("Type"))
assert input_spec[0].type.multiArrayType.WhichOneof(
"ShapeFlexibility"
) == "enumeratedShapes", "Expected enumeratedShapes in ShapeFlexibility"
spec_default_shape = [
s for s in input_spec[0].type.multiArrayType.shape
]
spec_enumerated_shapes = set()
for enumerated in input_spec[
0].type.multiArrayType.enumeratedShapes.shapes:
spec_enumerated_shapes.add(tuple([s for s in enumerated.shape]))
assert spec_default_shape == [
1, 3, 10, 30
], "Expected default shape to be [1, 3, 10, 10], got {} instead".format(
str(spec_default_shape))
assert spec_enumerated_shapes == set(
enumerated_shapes), "Enumerated shape mismatch"
def test_image_input_rangedim(self, convert_to):
example_input = torch.rand(1, 3, 50, 50) * 255
traced_model = torch.jit.trace(TestConvModule().eval(), example_input)
input_shape = ct.Shape(shape=(1, 3, ct.RangeDim(25, 100, default=45),
ct.RangeDim(25, 100, default=45)))
model = ct.convert(traced_model,
inputs=[ct.ImageType(shape=input_shape)],
convert_to=convert_to)
spec = model.get_spec()
assert spec.description.input[0].type.imageType.width == 45
assert spec.description.input[0].type.imageType.height == 45
assert spec.description.input[
0].type.imageType.imageSizeRange.widthRange.lowerBound == 25
assert spec.description.input[
0].type.imageType.imageSizeRange.widthRange.upperBound == 100
_assert_torch_coreml_output_shapes(model,
spec,
traced_model,
example_input,
is_image_input=True)
def test_grayscale_input_image(self, rank4_grayscale_input_model):
mlmodel = ct.convert(
rank4_grayscale_input_model,
inputs=[
ct.ImageType(name="input_image",
shape=(1, 1, 10, 20),
color_layout=ct.colorlayout.GRAYSCALE)
],
outputs=[ct.TensorType(name="output")],
minimum_deployment_target=ct.target.macOS13,
)
sample_input = np.random.randint(low=0, high=246, size=(1, 1, 10, 20))
img_input = Image.fromarray(sample_input[0, 0, :, :].astype(np.uint8),
'L')
model_output = mlmodel.predict({"input_image": img_input})['output']
reference_output = rank4_grayscale_input_model(
torch.from_numpy(sample_input.astype(
np.float32))).detach().numpy()
np.testing.assert_allclose(reference_output,
model_output,
rtol=1e-2,
atol=1e-2)
def convert(input_model_path, output_model_path, input_config_json,
compiler_options_json):
input_config = json.loads(input_config_json)
compiler_options = None
if compiler_options_json is not None:
compiler_options = json.loads(compiler_options_json)
try:
inputs = get_input_list(input_config)
classifier_config = get_classifier_config(compiler_options)
mlmodel = ct.convert(input_model_path,
inputs=inputs,
classifier_config=classifier_config)
mlmodel.save(output_model_path)
print("MLMODEL was saved to ", output_model_path)
except ValueError as e:
print('ValueError:', e)
err_str = str(e)
print("Error (str) {}".format(e))
for x in DISPLAY_TO_USER_ERRORS:
if x in err_str:
sys.exit(4)
sys.exit(1)
def test_tf2keras_shared_range_dim(use_symbol):
# Test examples in https://coremltools.readme.io/docs/flexible-inputs
import tensorflow as tf
input_dim = 3
# None denotes seq_len dimension
x1 = tf.keras.Input(shape=(None, input_dim), name="seq1")
x2 = tf.keras.Input(shape=(None, input_dim), name="seq2")
y = x1 + x2
keras_model = tf.keras.Model(inputs=[x1, x2], outputs=[y])
# One RangeDim shared by two inputs
if use_symbol:
seq_len_dim = ct.RangeDim(symbol='seq_len')
else:
# symbol is optional
seq_len_dim = ct.RangeDim()
seq1_input = ct.TensorType(name="seq1",
shape=(1, seq_len_dim, input_dim))
seq2_input = ct.TensorType(name="seq2",
shape=(1, seq_len_dim, input_dim))
mlmodel = ct.convert(keras_model, inputs=[seq1_input, seq2_input])
batch = 1
seq_len = 5
test_input_x1 = np.random.rand(batch, seq_len,
input_dim).astype(np.float32)
test_input_x2 = np.random.rand(batch, seq_len,
input_dim).astype(np.float32)
expected_val = keras_model([test_input_x1, test_input_x2])
if ct.utils._is_macos():
results = mlmodel.predict({
"seq1": test_input_x1,
"seq2": test_input_x2
})
np.testing.assert_allclose(results["Identity"],
expected_val,
rtol=1e-4,
atol=1e-3)
def test_prediction_with_fp16_io(self):
torch_model = torch.nn.Linear(30, 5).eval()
traced_model = torch.jit.trace(torch_model, torch.rand(1, 30))
mlmodel = ct.convert(
traced_model,
inputs=[
ct.TensorType(name="input", shape=(1, 30), dtype=np.float32)
],
outputs=[ct.TensorType(dtype=np.float32)],
minimum_deployment_target=ct.target.macOS13,
compute_units=ct.ComputeUnit.CPU_ONLY,
)
# test prediction
sample_input = np.random.rand(1, 30).astype(np.float32) * 10
model_output = mlmodel.predict(
{"input": sample_input})[mlmodel._spec.description.output[0].name]
reference_output = traced_model(
torch.from_numpy(sample_input)).detach().numpy()
np.testing.assert_allclose(reference_output,
model_output,
rtol=1e-2,
atol=1e-2)
def test_grayscale_input(self, rank4_input_model, rank3_input_model, rank4_grayscale_input_model):
with pytest.raises(ValueError, match="must have rank 4"):
mlmodel = ct.convert(rank3_input_model,
inputs=[ct.ImageType(color_layout=ct.colorlayout.GRAYSCALE)],
minimum_deployment_target=ct.target.macOS13,
)
# invalid shape
with pytest.raises(ValueError):
mlmodel = ct.convert(rank4_input_model,
inputs=[ct.ImageType(color_layout=ct.colorlayout.GRAYSCALE)],
minimum_deployment_target=ct.target.macOS13,
)
mlmodel = ct.convert(rank4_grayscale_input_model,
inputs=[ct.ImageType(color_layout=ct.colorlayout.GRAYSCALE)],
minimum_deployment_target=ct.target.macOS13,
)
assert_ops_in_mil_program(mlmodel, expected_op_list=["cast", "transpose", "add", "cast"])
assert_spec_input_image_type(mlmodel._spec, expected_feature_type=ft.ImageFeatureType.GRAYSCALE)
assert_prog_input_type(mlmodel._mil_program, expected_dtype_str="fp32")
assert_prog_output_type(mlmodel._mil_program, expected_dtype_str="fp32")
verify_prediction(mlmodel)
with pytest.raises(TypeError, match="float16 dtype for inputs is only supported for deployment target >= iOS16/macOS13"):
mlmodel = ct.convert(rank4_grayscale_input_model,
inputs=[ct.ImageType(color_layout=ct.colorlayout.GRAYSCALE_FLOAT16)],
minimum_deployment_target=ct.target.macOS12,
)
# test that grayscale_16 raises error when used with neural network
with pytest.raises(TypeError, match="float16 dtype for inputs is only supported for deployment target >= iOS16/macOS13"):
mlmodel = ct.convert(rank4_grayscale_input_model,
inputs=[ct.ImageType(color_layout=ct.colorlayout.GRAYSCALE_FLOAT16)],
)
mlmodel = ct.convert(rank4_grayscale_input_model,
inputs=[ct.ImageType(color_layout=ct.colorlayout.GRAYSCALE_FLOAT16)],
outputs=[ct.TensorType(dtype=np.float16)],
minimum_deployment_target=ct.target.macOS13,
)
assert_ops_in_mil_program(mlmodel, expected_op_list=["transpose", "add"])
assert_spec_input_image_type(mlmodel._spec, expected_feature_type=ft.ImageFeatureType.GRAYSCALE_FLOAT16)
assert_prog_input_type(mlmodel._mil_program, expected_dtype_str="fp16")
assert_output_dtype(mlmodel, expected_type_str="fp16")
verify_prediction(mlmodel)
def convertToCoremlSpec(torchScript, sampleInput):
"""
Converts a torchscript to a coreml model
"""
try:
print(f"Starting CoreML conversion with coremltools {ct.__version__}")
nnSpec = ct.convert(
torchScript,
inputs=[
ct.ImageType(
name="image",
shape=sampleInput.shape,
scale=1 / 255.0,
bias=[0, 0, 0],
)
],
).get_spec()
print(f"CoreML conversion success")
except Exception as e:
print(f"CoreML conversion failure: {e}")
return
return nnSpec
def test_multiarray_input_rangedim(self, convert_to):
if convert_to == "mlprogram" and ct.utils._macos_version() < (12, 0):
return
example_input = torch.rand(1, 3, 50, 50) * 100
traced_model = torch.jit.trace(TestConvModule().eval(), example_input)
input_shape = ct.Shape(shape=(1, 3, ct.RangeDim(25, 100, default=45),
ct.RangeDim(25, 100, default=45)))
model = ct.convert(traced_model,
inputs=[ct.TensorType(shape=input_shape)],
convert_to=convert_to)
spec = model.get_spec()
assert list(spec.description.input[0].type.multiArrayType.shape) == [
1, 3, 45, 45
]
assert spec.description.input[
0].type.multiArrayType.shapeRange.sizeRanges[2].lowerBound == 25
assert spec.description.input[
0].type.multiArrayType.shapeRange.sizeRanges[2].upperBound == 100
_assert_torch_coreml_output_shapes(model, spec, traced_model,
example_input)
def test_mil_ranged_image_with_default(self):
input_shape = [
ct.ImageType(name="x",
shape=(1, 3, 10, ct.RangeDim(10, 30, default=20)))
]
mlmodel = ct.convert(self.basic_network,
source="milinternal",
convert_to="mlprogram",
inputs=input_shape)
input_spec = mlmodel.get_spec().description.input
assert len(input_spec) == 1, "1 input expected, got {} instead".format(
len(input_spec))
assert input_spec[
0].name == "x", "input name in MLModel is {}, 'x' is expected".format(
input_spec[0].name)
assert input_spec[0].type.WhichOneof(
"Type") == "imageType", "Expected imageType, got {}".format(
input_spec[0].type.WhichOneof("Type"))
assert input_spec[0].type.imageType.WhichOneof(
"SizeFlexibility"
) == "imageSizeRange", "Expected imageSizeRange in ShapeFlexibility"
spec_H = input_spec[0].type.imageType.height
spec_W = input_spec[0].type.imageType.width
assert spec_H == 10 and spec_W == 20, "expected [H, W] == [10, 20], got [{}, {}] instead".format(
spec_H, spec_W)
spec_H_range = [
input_spec[0].type.imageType.imageSizeRange.heightRange.lowerBound,
input_spec[0].type.imageType.imageSizeRange.heightRange.upperBound
]
spec_W_range = [
input_spec[0].type.imageType.imageSizeRange.widthRange.lowerBound,
input_spec[0].type.imageType.imageSizeRange.widthRange.upperBound
]
assert spec_H_range == [10, 10], "Ranged height mismatch"
assert spec_W_range == [10, 30], "Ranged width mismatch"
def test_mil_ranged_multiarray_with_default(self):
input_shape = [
ct.TensorType(name="x",
shape=(1, 3, 10, ct.RangeDim(10, 30, default=20)))
]
mlmodel = ct.convert(self.basic_network,
source="milinternal",
convert_to="mlprogram",
inputs=input_shape)
input_spec = mlmodel.get_spec().description.input
assert len(input_spec) == 1, "1 input expected, got {} instead".format(
len(input_spec))
assert input_spec[
0].name == "x", "input name in MLModel is {}, 'x' is expected".format(
input_spec[0].name)
assert input_spec[0].type.WhichOneof(
"Type"
) == "multiArrayType", "Expected multiArrayType, got {}".format(
input_spec[0].type.WhichOneof("Type"))
assert input_spec[0].type.multiArrayType.WhichOneof(
"ShapeFlexibility"
) == "shapeRange", "Expected shapeRange in ShapeFlexibility"
spec_default_shape = [
s for s in input_spec[0].type.multiArrayType.shape
]
ranged_shapes = [(1, 1), (3, 3), (10, 10), (10, 30)]
spec_ranged_shapes = []
for range_dim in input_spec[
0].type.multiArrayType.shapeRange.sizeRanges:
spec_ranged_shapes.append(
tuple([range_dim.lowerBound, range_dim.upperBound]))
assert spec_default_shape == [
1, 3, 10, 20
], "Expected default shape to be [1, 3, 10, 20], got {} instead".format(
str(spec_default_shape))
assert spec_ranged_shapes == ranged_shapes, "Enumerated shape mismatch"
def convertCoreML(keras_model, labelbinarizer):
# modelName for coreML model
modelName = "rooms_coreml"
cleaned_path = "."
class_labels = labelbinarizer.classes_.tolist()
mlconfig = coremltools.ClassifierConfig(class_labels)
# load the trained convolutional neural network
model = coremltools.convert(
keras_model,
input_names=["dense_input"],
classifier_config=mlconfig,
)
spec = model.get_spec()
model.author = "Rooms - https://github.com/st0nedB/rooms"
model.license = "MIT"
model.short_description = "This model can be used to predict in which room a device resides based on BLE beacon measurements."
model.versionString = "Version 0.1"
model.input_description["dense_input"] = "Vector of input RSSI"
model.output_description["Identity"] = "Predicted Room"
model.save(cleaned_path + "/" + modelName + ".mlmodel")
def test_can_build_keras_to_coreml_to_relay():
"""Test multiple conversion paths and importing from
a saved file."""
model = keras.models.Sequential()
model.add(
keras.layers.Conv2D(
filters=6,
kernel_size=(1, 1),
activation="relu",
padding="same",
input_shape=(3, 3, 1),
data_format="channels_first",
)
)
with tempfile.TemporaryDirectory() as tmpdir:
kmodel_fn = path.join(tmpdir, "c1mdl.h5")
model.save(kmodel_fn)
mdl = cm.convert(kmodel_fn)
model_file = path.join(tmpdir, "c1.mlmodel")
mdl.save(model_file)
mdl = cm.models.MLModel(model_file)
desc = mdl.get_spec().description
iname = desc.input[0].name
ishape = desc.input[0].type.multiArrayType.shape
shape_dict = {}
for i in mdl.get_spec().description.input:
iname = i.name
ishape = i.type.multiArrayType.shape
shape_dict[iname] = ishape
mod, params = relay.frontend.from_coreml(mdl, shape_dict)
with tvm.transform.PassContext(opt_level=3):
relay.build(mod, "llvm", params=params)
def test_mil_enumerated_image_with_default(self):
enumerated_shapes = tuple([(1, 3, 10, 10), (1, 3, 10, 20),
(1, 3, 10, 30)])
input_shape = [
ct.ImageType(name="x",
shape=ct.EnumeratedShapes(shapes=enumerated_shapes,
default=(1, 3, 10, 30)))
]
mlmodel = ct.convert(self.basic_network,
source="milinternal",
convert_to="mlprogram",
inputs=input_shape)
input_spec = mlmodel.get_spec().description.input
assert len(input_spec) == 1, "1 input expected, got {} instead".format(
len(input_spec))
assert input_spec[
0].name == "x", "input name in MLModel is {}, 'x' is expected".format(
input_spec[0].name)
assert input_spec[0].type.WhichOneof(
"Type") == "imageType", "Expected imageType, got {}".format(
input_spec[0].type.WhichOneof("Type"))
assert input_spec[0].type.imageType.WhichOneof(
"SizeFlexibility"
) == "enumeratedSizes", "Expected enumeratedShapes in ShapeFlexibility"
spec_H = input_spec[0].type.imageType.height
spec_W = input_spec[0].type.imageType.width
assert spec_H == 10 and spec_W == 30, "expected [H, W] == [10, 30], got [{}, {}] instead".format(
spec_H, spec_W)
spec_enumerated_shapes = set()
for enumerated in input_spec[0].type.imageType.enumeratedSizes.sizes:
spec_enumerated_shapes.add(
tuple([1, 3, enumerated.height, enumerated.width]))
assert spec_enumerated_shapes == set(
enumerated_shapes), "Enumerated shape mismatch"
def test_grayscale_fp16_input_image(self, rank4_grayscale_input_model):
mlmodel = ct.convert(
rank4_grayscale_input_model,
inputs=[
ct.ImageType(name="input_image",
shape=(1, 1, 10, 20),
color_layout=ct.colorlayout.GRAYSCALE_FLOAT16)
],
outputs=[ct.TensorType(name="output")],
minimum_deployment_target=ct.target.macOS13,
)
# incorrect way to do prediction
with pytest.raises(
TypeError,
match="must be of type PIL.Image.Image with mode=='F'",
):
sample_input = np.random.randint(low=0,
high=246,
size=(1, 1, 10, 20))
img_input = Image.fromarray(
sample_input[0, 0, :, :].astype(np.uint8), 'L')
mlmodel.predict({"input_image": img_input})
# correct way to do prediction
sample_input = np.random.rand(1, 1, 10, 20) # in between [0, 1]
img_input = Image.fromarray(
sample_input[0, 0, :, :].astype(np.float32), 'F')
model_output = mlmodel.predict({"input_image": img_input})['output']
reference_output = rank4_grayscale_input_model(
torch.from_numpy(sample_input.astype(
np.float32))).detach().numpy()
np.testing.assert_allclose(reference_output,
model_output,
rtol=1e-2,
atol=1e-2)
def test_rename_feature_mlprogram(self):
torch_model = _torch.nn.ReLU().eval()
model = coremltools.convert(
_torch.jit.trace(torch_model, _torch.rand(3, )),
inputs=[coremltools.TensorType(shape=(3,))],
convert_to='mlprogram'
)
spec = model.get_spec()
input_name = spec.description.input[0].name
output_name = spec.description.output[0].name
# rename input
rename_feature(spec, input_name, "new_input_name")
self.assertEqual(spec.description.input[0].name, "new_input_name")
model = coremltools.models.MLModel(spec)
out = model.predict({"new_input_name": np.array([1.0, 2.0, 3.0])})[output_name]
self.assertEqual(out[0], 1.0)
# rename output
rename_feature(spec, output_name, "new_output_name")
self.assertEqual(spec.description.output[0].name, "new_output_name")
model = coremltools.models.MLModel(spec)
out = model.predict({"new_input_name": np.array([1.0, 2.0, 3.0])})["new_output_name"]
self.assertEqual(out[1], 2.0)
print(onnx.helper.printable_graph(
onnx_model.graph)) # print a human readable model
print('ONNX export success, saved as %s' % f)
except Exception as e:
print('ONNX export failure: %s' % e)
# CoreML export
try:
import coremltools as ct
print('\nStarting CoreML export with coremltools %s...' %
ct.__version__)
# convert model from torchscript and apply pixel scaling as per detect.py
model = ct.convert(ts,
inputs=[
ct.ImageType(name='image',
shape=img.shape,
scale=1 / 255.0,
bias=[0, 0, 0])
])
f = opt.weights.replace('.pt', '.mlmodel') # filename
model.save(f)
print('CoreML export success, saved as %s' % f)
except Exception as e:
print('CoreML export failure: %s' % e)
# Finish
print(
'\nExport complete (%.2fs). Visualize with https://github.com/lutzroeder/netron.'
% (time.time() - t))
def run(
weights='./yolov5s.pt', # weights path
img_size=(640, 640), # image (height, width)
batch_size=1, # batch size
device='cpu', # cuda device, i.e. 0 or 0,1,2,3 or cpu
include=('torchscript', 'onnx', 'coreml'), # include formats
half=False, # FP16 half-precision export
inplace=False, # set YOLOv5 Detect() inplace=True
train=False, # model.train() mode
optimize=False, # TorchScript: optimize for mobile
dynamic=False, # ONNX: dynamic axes
simplify=False, # ONNX: simplify model
opset_version=12, # ONNX: opset version
):
t = time.time()
include = [x.lower() for x in include]
img_size *= 2 if len(img_size) == 1 else 1 # expand
# Load PyTorch model
device = select_device(device)
assert not (
device.type == 'cpu' and opt.half
), '--half only compatible with GPU export, i.e. use --device 0'
model = attempt_load(weights, map_location=device) # load FP32 model
labels = model.names
# Input
gs = int(max(model.stride)) # grid size (max stride)
img_size = [check_img_size(x, gs)
for x in img_size] # verify img_size are gs-multiples
img = torch.zeros(batch_size, 3, *img_size).to(
device) # image size(1,3,320,192) iDetection
# Update model
if half:
img, model = img.half(), model.half() # to FP16
model.train() if train else model.eval(
) # training mode = no Detect() layer grid construction
for k, m in model.named_modules():
m._non_persistent_buffers_set = set() # pytorch 1.6.0 compatibility
if isinstance(m, Conv): # assign export-friendly activations
if isinstance(m.act, nn.Hardswish):
m.act = Hardswish()
elif isinstance(m.act, nn.SiLU):
m.act = SiLU()
elif isinstance(m, Detect):
m.inplace = inplace
m.onnx_dynamic = dynamic
# m.forward = m.forward_export # assign forward (optional)
for _ in range(2):
y = model(img) # dry runs
print(
f"\n{colorstr('PyTorch:')} starting from {weights} ({file_size(weights):.1f} MB)"
)
# TorchScript export -----------------------------------------------------------------------------------------------
if 'torchscript' in include or 'coreml' in include:
prefix = colorstr('TorchScript:')
try:
print(
f'\n{prefix} starting export with torch {torch.__version__}...'
)
f = weights.replace('.pt', '.torchscript.pt') # filename
ts = torch.jit.trace(model, img, strict=False)
(optimize_for_mobile(ts) if optimize else ts).save(f)
print(
f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)'
)
except Exception as e:
print(f'{prefix} export failure: {e}')
# ONNX export ------------------------------------------------------------------------------------------------------
if 'onnx' in include:
prefix = colorstr('ONNX:')
try:
import onnx
print(f'{prefix} starting export with onnx {onnx.__version__}...')
f = weights.replace('.pt', '.onnx') # filename
torch.onnx.export(
model,
img,
f,
verbose=False,
opset_version=opset_version,
training=torch.onnx.TrainingMode.TRAINING
if train else torch.onnx.TrainingMode.EVAL,
do_constant_folding=not train,
input_names=['images'],
output_names=['output'],
dynamic_axes={
'images': {
0: 'batch',
2: 'height',
3: 'width'
}, # shape(1,3,640,640)
'output': {
0: 'batch',
1: 'anchors'
} # shape(1,25200,85)
} if dynamic else None)
# Checks
model_onnx = onnx.load(f) # load onnx model
onnx.checker.check_model(model_onnx) # check onnx model
# print(onnx.helper.printable_graph(model_onnx.graph)) # print
# Simplify
if simplify:
try:
check_requirements(['onnx-simplifier'])
import onnxsim
print(
f'{prefix} simplifying with onnx-simplifier {onnxsim.__version__}...'
)
model_onnx, check = onnxsim.simplify(
model_onnx,
dynamic_input_shape=dynamic,
input_shapes={'images': list(img.shape)}
if dynamic else None)
assert check, 'assert check failed'
onnx.save(model_onnx, f)
except Exception as e:
print(f'{prefix} simplifier failure: {e}')
print(
f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)'
)
except Exception as e:
print(f'{prefix} export failure: {e}')
# CoreML export ----------------------------------------------------------------------------------------------------
if 'coreml' in include:
prefix = colorstr('CoreML:')
try:
import coremltools as ct
print(
f'{prefix} starting export with coremltools {ct.__version__}...'
)
assert train, 'CoreML exports should be placed in model.train() mode with `python export.py --train`'
model = ct.convert(ts,
inputs=[
ct.ImageType('image',
shape=img.shape,
scale=1 / 255.0,
bias=[0, 0, 0])
])
f = weights.replace('.pt', '.mlmodel') # filename
model.save(f)
print(
f'{prefix} export success, saved as {f} ({file_size(f):.1f} MB)'
)
except Exception as e:
print(f'{prefix} export failure: {e}')
# Finish
print(
f'\nExport complete ({time.time() - t:.2f}s). Visualize with https://github.com/lutzroeder/netron.'
)
import onnx
print('\nStarting ONNX export with onnx %s...' % onnx.__version__)
f = opt.weights.replace('.pt', '.onnx') # filename
model.fuse() # only for ONNX
torch.onnx.export(model, img, f, verbose=False, opset_version=12, input_names=['images'],
output_names=['classes', 'boxes'] if y is None else ['output'])
# Checks
onnx_model = onnx.load(f) # load onnx model
onnx.checker.check_model(onnx_model) # check onnx model
print(onnx.helper.printable_graph(onnx_model.graph)) # print a human readable model
print('ONNX export success, saved as %s' % f)
except Exception as e:
print('ONNX export failure: %s' % e)
# CoreML export
try:
import coremltools as ct
print('\nStarting CoreML export with coremltools %s...' % ct.__version__)
model = ct.convert(ts, inputs=[ct.ImageType(name='images', shape=img.shape)]) # convert
f = opt.weights.replace('.pt', '.mlmodel') # filename
model.save(f)
print('CoreML export success, saved as %s' % f)
except Exception as e:
print('CoreML export failure: %s' % e)
# Finish
print('\nExport complete. Visualize with https://github.com/lutzroeder/netron.')
def test_mil_as_package(self):
import torch
num_tokens = 3
embedding_size = 5
class TestModule(torch.nn.Module):
def __init__(self):
super(TestModule, self).__init__()
self.embedding = torch.nn.Embedding(num_tokens, embedding_size)
def forward(self, x):
return self.embedding(x)
model = TestModule()
model.eval()
example_input = torch.randint(high=num_tokens,
size=(2, ),
dtype=torch.int64)
traced_model = torch.jit.trace(model, example_input)
temp_package_dir = tempfile.TemporaryDirectory(suffix=".mlpackage")
for converted_package_path in [None, temp_package_dir.name]:
mlmodel = coremltools.convert(
traced_model,
package_dir=converted_package_path,
source='pytorch',
convert_to='mlprogram',
compute_precision=coremltools.precision.FLOAT32,
inputs=[
coremltools.TensorType(
name="input",
shape=example_input.shape,
dtype=example_input.numpy().dtype,
)
],
)
assert isinstance(mlmodel, MLModel)
package_path = tempfile.mkdtemp(suffix=".mlpackage")
mlmodel.save(package_path)
assert ModelPackage.isValid(package_path)
assert os.path.exists(
ModelPackage(package_path).getRootModel().path())
# Read back the saved bundle and compile
mlmodel2 = MLModel(package_path,
compute_units=ComputeUnit.CPU_ONLY)
if utils._macos_version() >= (12, 0):
result = mlmodel2.predict({
"input":
example_input.cpu().detach().numpy().astype(np.float32)
})
# Verify outputs
expected = model(example_input)
name = list(result.keys())[0]
np.testing.assert_allclose(result[name],
expected.cpu().detach().numpy())
# Cleanup package
shutil.rmtree(package_path)
tmp_package_path = mlmodel.package_path
assert os.path.exists(tmp_package_path)
del mlmodel
if converted_package_path is not None:
# Verify we leave the provided package dir alone
assert os.path.exists(tmp_package_path)
temp_package_dir.cleanup()
help='Net file to be converted to a model checkpoint.')
argparser.add_argument('--start',
type=int,
default=0,
help='Offset to set global_step to.')
argparser.add_argument('--cfg',
type=argparse.FileType('r'),
help='yaml configuration with training parameters')
argparser.add_argument('-e',
'--ignore-errors',
action='store_true',
help='Ignore missing and wrong sized values.')
args = argparser.parse_args()
cfg = yaml.safe_load(args.cfg.read())
print(yaml.dump(cfg, default_flow_style=False))
START_FROM = args.start
tfp = tfprocess.TFProcess(cfg)
tfp.init_net_v2()
tfp.replace_weights_v2(args.net, args.ignore_errors)
tfp.global_step.assign(START_FROM)
root_dir = os.path.join(cfg['training']['path'], cfg['name'])
if not os.path.exists(root_dir):
os.makedirs(root_dir)
tfp.manager.save(checkpoint_number=START_FROM)
print("Wrote model to {}".format(tfp.manager.latest_checkpoint))
coreml_model = ct.convert(tfp.model, source='tensorflow')
coreml_model.save(args.net + '.mlmodel')
### tensorflow==2.3.1
import tensorflow as tf
import coremltools as ct
mlmodel = ct.convert('saved_model_age', source='tensorflow')
mlmodel.save("age_gender_recognition_62x62_float32.mlmodel")
