def test():
for kwargs in [
{},
# {"use_bias": False}, # FIXME: Not Supported Yet
# {"stateful": True}, # FIXME: Not Supported Yet
# {"go_backwards": True}, # FIXME: Not Supported Yet
{"recurrent_activation": "hard_sigmoid"},
{"recurrent_activation": "sigmoid"},
{"return_state": True},
{"return_sequences": True},
{"return_state": True, "return_sequences": True},
]:
return_state = ("return_state" in kwargs) and (kwargs["return_state"])
x = keras.layers.Input((14, 15))
vx = np.random.rand(2, 14, 15)
outputs = keras.layers.LSTM(units=16, **kwargs)(x)
if return_state:
y, _, c = outputs
model = keras.models.Model([x], [y, c])
graph = KerasConverter(batch_size=2).convert(model)
vy, vc = model.predict(vx, batch_size=2)
expected = {
graph.outputs[0]: vy,
graph.outputs[1]: vc,
}
else:
y = outputs
model = keras.models.Model([x], [y])
graph = KerasConverter(batch_size=2).convert(model)
vy = model.predict(vx, batch_size=2)
expected = {
graph.outputs[0]: vy,
}
generate_kernel_test_case(
description="[keras] LSTM " + (", ".join([f"{k}={v}" for k, v in kwargs.items()])),
graph=graph,
backend=["webgpu", "webassembly"], # FIXME: fallback backend doesn't support LSTM
inputs={graph.inputs[0]: vx},
expected=expected,
raise_skip=False
)
raise SkipTest
def template(units=16,
return_sequences=False,
return_state=False,
go_backwards=False,
stateful=False,
activation="tanh",
recurrent_activation="hard_sigmoid",
use_bias=True,
description: str = ""):
x = keras.layers.Input((14, 15))
vx = np.random.rand(2, 14, 15).astype(np.float32)
outputs = keras.layers.LSTM(units=units,
return_sequences=return_sequences,
return_state=return_state,
go_backwards=go_backwards,
stateful=stateful,
activation=activation,
recurrent_activation=recurrent_activation,
use_bias=use_bias)(x)
if return_state:
y, _, c = outputs
model = keras.models.Model([x], [y, c])
graph = KerasConverter(batch_size=2,
use_tensorflow_converter=False).convert(model)
vy, vc = model.predict(vx, batch_size=2)
expected = {
graph.outputs[0]: vy,
graph.outputs[1]: vc,
}
else:
y = outputs
model = keras.models.Model([x], [y])
graph = KerasConverter(batch_size=2,
use_tensorflow_converter=False).convert(model)
vy = model.predict(vx, batch_size=2)
expected = {
graph.outputs[0]: vy,
}
generate_kernel_test_case(description=f"[keras] LSTM {description}",
graph=graph,
backend=["webgpu", "webassembly"],
inputs={graph.inputs[0]: vx},
expected=expected,
EPS=1e-2)
def template(pool_size=(3, 3),
strides=2,
padding="valid",
data_format=None,
description: str = ""):
x = keras.layers.Input(
(15, 17, 16)
) # (height + pad * 2 - pool_size) % stride == 0 to avoid edge difference
y = keras.layers.AveragePooling2D(pool_size=pool_size,
strides=strides,
padding=padding,
data_format=data_format)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 15, 17, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2,
use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] AveragePooling2D {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def test():
for kwargs in [
# {"data_format": "channels_first"}, # FIXME: Not Supported Yet
{
"data_format": "channels_last"
},
]:
channels_first = ("data_format" in kwargs) and (kwargs["data_format"]
== "channels_first")
x = keras.layers.Input((14, 15, 16))
y = keras.layers.GlobalAveragePooling2D(**kwargs)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(
model, input_orders=[OrderNCHW if channels_first else OrderNHWC])
generate_kernel_test_case(
description="[keras] GlobalAveragePooling2D " +
(", ".join([f"{k}={v}" for k, v in kwargs.items()])),
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
raise_skip=False)
raise SkipTest
def template(x1_shape=[5, 10, 15],
x2_shape=[5, 9, 15],
axis=2,
description: str = ""):
x1 = keras.layers.Input(x1_shape)
x2 = keras.layers.Input(x2_shape)
y = keras.layers.Concatenate(axis=axis)([x1, x2])
model = keras.models.Model([x1, x2], [y])
vx1 = np.random.rand(2, *x1_shape)
vx2 = np.random.rand(2, *x2_shape)
vy = model.predict([vx1, vx2], batch_size=2)
graph = KerasConverter(batch_size=2,
use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] Concat {description}",
graph=graph,
inputs={
graph.inputs[0]: vx1,
graph.inputs[1]: vx2
},
expected={graph.outputs[0]: vy},
)
def test():
for kwargs in [
{"output_dim": 1},
{"output_dim": 10},
]:
N_VOCABULARY = 30
x = keras.layers.Input((14,))
y = keras.layers.Embedding(input_dim=N_VOCABULARY, **kwargs)(x)
model = keras.models.Model([x], [y])
vx = np.random.randint(low=0, high=N_VOCABULARY, size=(2, 14))
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description="[keras] Embedding " + (", ".join([f"{k}={v}" for k, v in kwargs.items()])),
graph=graph,
backend=["webgpu", "webassembly"], # FIXME: fallback backend doesn't support ZeroPadding1D
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
raise_skip=False
)
raise SkipTest
def test():
for kwargs in [
{},
{"use_bias": False},
{"activation": None},
{"use_bias": False, "activation": None}
]:
x = keras.layers.Input((4,))
y = keras.layers.Dense(8, **kwargs)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description="[keras] Dense " + (", ".join([f"{k}={v}" for k, v in kwargs.items()])),
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
raise_skip=False
)
raise SkipTest
def template(filters=17,
kernel_size=3,
strides=(1, 1),
padding='valid',
data_format=None,
dilation_rate=(1, 1),
activation=None,
use_bias=True,
description: str = ""):
x = keras.layers.Input((14, 15, 16))
y = keras.layers.Conv2D(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
activation=activation,
use_bias=use_bias)(x)
model = keras.models.Model([x], [y])
vx = np.random.randint(low=0, high=100,
size=(2, 14, 15, 16)).astype(np.float32)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(
model,
input_orders=[
OrderNCHW if data_format == "channels_first" else OrderNHWC
])
generate_kernel_test_case(description=f"[keras] Conv2D {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
EPS=1e-2)
def template(pool_size=(3, 3),
shape=(15, 17, 16),
strides=2,
padding="valid",
data_format=None,
description: str = ""):
x = keras.layers.Input(shape)
y = keras.layers.MaxPooling2D(pool_size=pool_size,
strides=strides,
padding=padding,
data_format=data_format)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, *shape) - 0.5
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2,
use_tensorflow_converter=False).convert(model)
assert list(vy.shape) == list(
graph.outputs[0].shape
), f"(vy.shape)={vy.shape}, (graph.outputs[0].shape)={graph.outputs[0].shape}"
generate_kernel_test_case(
description=f"[keras] MaxPooling2D {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def test():
for kwargs in [
{},
{
"epsilon": 0.8
},
{
"scale": False
},
{
"center": False
},
]:
x = keras.layers.Input((14, 15, 16))
y = keras.layers.BatchNormalization(axis=3, **kwargs)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description="[keras] BatchNormalization " +
(", ".join([f"{k}={v}" for k, v in kwargs.items()])),
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
raise_skip=False)
raise SkipTest
def template(shape=(14, 15, 4), filters=5, kernel_size=3, strides=(1, 1), padding='valid', data_format="channels_last",
dilation_rate=(1, 1), activation=None, depth_multiplier=1, use_bias=True, description: str = ""):
x = keras.layers.Input(shape)
y = keras.layers.SeparableConv2D(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding=padding,
data_format=data_format,
dilation_rate=dilation_rate,
depth_multiplier=depth_multiplier,
activation=activation,
use_bias=use_bias)(x)
model = keras.models.Model([x], [y])
vx = np.random.randint(low=0, high=100, size=(2, *shape)).astype(np.float32)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] SeparableConv2D {description}",
graph=graph,
backend=["webgpu", "webgl", "webassembly"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
# TODO: replace computation algorithm with more accurate one
EPS=1e-2
)
def test():
x = keras.layers.Input((14, ))
y = keras.layers.RepeatVector(n=5)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(description="[keras] Reshape",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy})
def test():
x = keras.layers.Input((3, 4, 5))
y = keras.layers.Flatten()(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 3, 4, 5)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(description="[keras] Flatten",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy})
def test():
x = keras.layers.Input((14, 15))
y = keras.layers.GlobalAveragePooling1D()(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(description="[keras] GlobalAveragePooling1D",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy})
def template(activation="", EPS=1e-5, description: str = ""):
x = keras.layers.Input((4, ))
y = keras.layers.Activation(activation)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4) - 0.5
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(description=f"[keras] Activation {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
EPS=EPS)
def template(alpha=1.0, description: str = ""):
x = keras.layers.Input((4, ))
y = keras.layers.ELU(alpha=alpha)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4).astype(np.float32) - 0.5
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] Elu {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(description: str = ""):
x = keras.layers.Input((3, 4, 5))
y = keras.layers.Flatten()(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 3, 4, 5)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] Flatten {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(activation=None, use_bias=True, description: str = ""):
x = keras.layers.Input((4,))
y = keras.layers.Dense(8, activation=activation, use_bias=use_bias)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4) - 0.5
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] Dense {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(axis=-1, epsilon=1e-3, center=True, scale=True, description: str = ""):
x = keras.layers.Input((14, 15, 16))
y = keras.layers.BatchNormalization(axis=axis, epsilon=epsilon, center=center, scale=scale)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] BatchNormalization {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(pool_size=3, strides=1, padding="valid", description: str = ""):
x = keras.layers.Input((14, 15))
y = keras.layers.AveragePooling1D(pool_size=pool_size, strides=strides, padding=padding)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] AveragePooling1D {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(theta=1.0, description: str = ""):
x = keras.layers.Input((4,))
y = keras.layers.ThresholdedReLU(theta=theta)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4) - 0.5
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] ThresholdedReLU {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(description: str = ""):
x = keras.layers.Input((4, 5, 6))
y = keras.layers.Reshape(target_shape=(12, 10))(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4, 5, 6)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] Reshape {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(description: str = ""):
x = keras.layers.Input((14, 15, 16))
y = keras.layers.GlobalMaxPooling2D()(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] GlobalMaxPooling2D {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(n=5, description: str = ""):
x = keras.layers.Input((3,))
y = keras.layers.RepeatVector(n=n)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 3)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] RepeatVector {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def test():
x = keras.layers.Input((4, ))
y = keras.layers.LeakyReLU()(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 4) - 0.5
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] LeakyReLU",
graph=graph,
backend=["webgpu", "webassembly"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def test_itself():
x = keras.layers.Input([5, 10, 15])
y = keras.layers.Concatenate(axis=2)([x, x])
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 5, 10, 15)
vy = model.predict([vx], batch_size=2)
graph = KerasConverter(batch_size=2,
use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] Concat itself",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(padding=1, description: str = ""):
x = keras.layers.Input((14, 15))
y = keras.layers.ZeroPadding1D(padding=padding)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] ZeroPadding1D {description}",
graph=graph,
backend=["webgpu", "webassembly"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
EPS=1e-2)
def template(description: str = ""):
x = keras.layers.Input((14, 15, 16))
y = keras.layers.GlobalAveragePooling2D()(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2, use_tensorflow_converter=False).convert(model)
generate_kernel_test_case(
description=f"[keras] GlobalAveragePooling2D {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(input_dim=30, output_dim=10, description: str = ""):
x = keras.layers.Input((14, ))
y = keras.layers.Embedding(input_dim=input_dim, output_dim=output_dim)(x)
model = keras.models.Model([x], [y])
vx = np.random.randint(low=0, high=input_dim, size=(2, 14))
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(model)
generate_kernel_test_case(
description=f"[keras] Embedding {description}",
graph=graph,
backend=["webgpu", "webassembly"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def test():
for kwargs in [
{},
# {"data_format": "channels_first"}, # FIXME: Not Supported Yet
{
"data_format": "channels_last"
},
{
"use_bias": False
},
{
"activation": None
},
{
"padding": "valid"
},
{
"padding": "same"
},
{
"dilation_rate": (2, 2)
},
]:
channels_first = ("data_format" in kwargs) and (kwargs["data_format"]
== "channels_first")
x = keras.layers.Input((14, 15, 16))
y = keras.layers.Conv2D(kernel_size=3, filters=8, **kwargs)(x)
model = keras.models.Model([x], [y])
vx = np.random.rand(2, 14, 15, 16)
vy = model.predict(vx, batch_size=2)
graph = KerasConverter(batch_size=2).convert(
model, input_orders=[OrderNCHW if channels_first else OrderNHWC])
generate_kernel_test_case(
description="[keras] Conv2D " +
(", ".join([f"{k}={v}" for k, v in kwargs.items()])),
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
raise_skip=False)
raise SkipTest
