def template(cond_shape, x0_shape, x1_shape, description: str = ""):
cond = tf.placeholder(np.bool, cond_shape)
x0 = tf.placeholder(np.float32, x0_shape, "x0")
x1 = tf.placeholder(np.float32, x1_shape, "x1")
y = tf.where(cond, x0, x1)
vcond = np.random.rand(*cond_shape) > 0.5
vx0 = np.random.rand(*x0_shape).astype(np.float32)
vx1 = np.random.rand(*x1_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {cond: vcond, x0: vx0, x1: vx1})
graph = TensorFlowConverter(sess, batch_size=2).convert([cond, x0, x1],
[y])
generate_kernel_test_case(
description=f"[TensorFlow] Select {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={
graph.inputs[0]: vcond,
graph.inputs[1]: vx0,
graph.inputs[2]: vx1
},
expected={graph.outputs[0]: vy},
)
def template(x_shape=[2, 3, 4, 5], data_format="NHWC", description: str = ""):
from tensorflow.python.ops import nn
x = tf.placeholder(np.float32, x_shape)
scale = tf.placeholder(
np.float32, x_shape[-1] if data_format == "NHWC" else x_shape[1])
bias = tf.placeholder(np.float32,
x_shape[-1] if data_format == "NHWC" else x_shape[1])
mean = tf.placeholder(np.float32,
x_shape[-1] if data_format == "NHWC" else x_shape[1])
variance = tf.placeholder(
np.float32, x_shape[-1] if data_format == "NHWC" else x_shape[1])
y, _, _ = nn.fused_batch_norm(x,
scale,
bias,
mean,
variance,
data_format=data_format,
is_training=False)
vx = np.random.rand(*x_shape).astype(np.float32)
vs = np.random.rand(
*[x_shape[-1] if data_format == "NHWC" else x_shape[1]]).astype(
np.float32)
vb = np.random.rand(
*[x_shape[-1] if data_format == "NHWC" else x_shape[1]]).astype(
np.float32)
vm = np.random.rand(
*[x_shape[-1] if data_format == "NHWC" else x_shape[1]]).astype(
np.float32)
vv = np.random.rand(
*[x_shape[-1] if data_format == "NHWC" else x_shape[1]]).astype(
np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {
x: vx,
scale: vs,
bias: vb,
mean: vm,
variance: vv
})
graph = TensorFlowConverter(sess, batch_size=2).convert(
[x, scale, bias, mean, variance], [y])
generate_kernel_test_case(
description=f"[TensorFlow] FusedBatchNorm {description}",
graph=graph,
inputs={
graph.inputs[0]: vx,
graph.inputs[1]: vs,
graph.inputs[2]: vb,
graph.inputs[3]: vm,
graph.inputs[4]: vv
},
expected={graph.outputs[0]: vy},
)
def template(x_shape=[2, 3, 4, 5], description: str = ""):
from tensorflow.python.ops import nn
x = tf.placeholder(np.float32, x_shape, "x")
y = nn.relu6(x)
vx = np.random.rand(*x_shape).astype(np.float32) - 0.5
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(description=f"[TensorFlow] Relu6 {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy})
def template(x_shape=[2, 3, 4, 5], description: str = ""):
x = tf.placeholder(np.float32, x_shape, "x")
y = tf.log(x)
vx = np.random.rand(*x_shape).astype(np.float32) + 1
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(description=f"[TensorFlow] Log {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
EPS=1e-2)
def template(x_shape=[2, 3, 4, 5], ksize=[1, 3, 3, 1], strides=[1, 1, 1, 1], padding="VALID", description: str = ""):
from tensorflow.python.ops import nn
x = tf.placeholder(np.float32, x_shape, "x")
y = nn.avg_pool(x, ksize=ksize, strides=strides, padding=padding)
vx = np.random.rand(*x_shape).astype(np.float32) - 0.5
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(
description=f"[TensorFlow] AvgPool {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy}
)
def template(x_shape, description: str = ""):
x = tf.placeholder(np.float32, x_shape, "x")
y = tf.exp(x)
vx = np.random.rand(*x_shape) - 0.5
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(
description=f"[TensorFlow] Exp {description}",
graph=graph,
inputs={graph.inputs[0]: vx, },
expected={graph.outputs[0]: vy},
)
def template(x_shape, description: str = ""):
x = tf.placeholder(np.float32, x_shape, "x")
y = tf.depth_to_space(x, 2)
vx = np.random.rand(*x_shape).astype(np.float32) - 0.5
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(
description=f"[TensorFlow] Depth2Space {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(x_shape=[2, 4, 6, 8], y_shape=[8, 48], description: str = ""):
x = tf.placeholder(np.float32, x_shape)
y = tf.reshape(x, y_shape)
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
assert list(vy.shape) == list(y.shape)
generate_kernel_test_case(
description=f"[TensorFlow] Reshape {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(N=2,
Hin=5,
Hout=3,
Win=5,
Wout=3,
Cin=1,
Cout=1,
KH=3,
KW=3,
strides=[1, 1, 1, 1],
padding="VALID",
data_format="NHWC",
description: str = ""):
x_shape = [N, Hin, Win, Cin
] if data_format == "NHWC" else [N, Cin, Hin, Win]
y_shape = [N, Hout, Wout, Cout
] if data_format == "NHWC" else [N, Cout, Hout, Wout]
w_shape = [KH, KW, Cin, Cout]
w = tf.placeholder(np.float32, w_shape)
gy = tf.placeholder(np.float32, y_shape)
x = tf.nn.conv2d_backprop_input(x_shape,
w,
gy,
strides=strides,
padding=padding,
data_format=data_format)
vw = np.random.rand(*w_shape).astype(np.float32)
vgy = np.random.rand(*y_shape).astype(np.float32)
with tf.Session() as sess:
vx, = sess.run([x], {w: vw, gy: vgy})
graph = TensorFlowConverter(sess, batch_size=2).convert([w, gy], [x])
assert list(vx.shape) == list(
graph.outputs[0].shape
), f"(vx.shape)={vx.shape}, (graph.outputs[0].shape)={graph.outputs[0].shape}"
generate_kernel_test_case(
description=f"[TensorFlow] Conv2DBackPropInput {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={
graph.inputs[0]: vw,
graph.inputs[1]: vgy
},
expected={graph.outputs[0]: vx})
def template(start, limit, delta, x1_shape, description: str = ""):
x0 = tf.range(start, limit, delta, dtype=np.float32)
x1 = tf.placeholder(np.float32, x1_shape)
y = x0 + x1
vx1 = np.random.rand(*x1_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x1: vx1})
graph = TensorFlowConverter(sess, batch_size=2).convert([x1], [y])
generate_kernel_test_case(
description=f"[TensorFlow] Range {description}",
graph=graph,
inputs={graph.inputs[0]: vx1},
expected={graph.outputs[0]: vy},
)
def template(x_shape, paddings, constant_values, description: str = ""):
x = tf.placeholder(np.float32, x_shape)
y = tf.pad(x, paddings, constant_values=constant_values)
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
assert list(vy.shape) == list(y.shape)
generate_kernel_test_case(
description=f"[TensorFlow] PadV2 {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(x_shape=[2, 4, 6, 8], multiplier=[1, 3, 2, 1], description: str = ""):
x = tf.placeholder(np.float32, x_shape)
y = tf.tile(x, multiples=multiplier)
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(
description=f"[TensorFlow] Tile {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(x_shape, paddings, mode, description: str = ""):
x = tf.placeholder(np.float32, x_shape)
y = tf.pad(x, paddings, mode=mode)
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
assert list(vy.shape) == list(y.shape)
generate_kernel_test_case(
description=f"[TensorFlow] MirrorPad {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(x_shape=[2, 3, 4, 5], pred=True, description: str = ""):
from tensorflow.python.ops import control_flow_ops
x = tf.placeholder(np.float32, x_shape)
y_f, y_t = control_flow_ops.switch(x, pred)
y = y_t if pred else y_f
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
generate_kernel_test_case(
description=f"[TensorFlow] Switch {description}",
graph=graph,
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(x: tf.Tensor, y, description: str = ""):
vx = np.random.rand(*[dim.value
for dim in x.shape.dims]).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
assert list(vy.shape) == list(
graph.outputs[0].shape
), f"vy_shape: {vy.shape}, y.shape: {graph.outputs[0].shape}"
generate_kernel_test_case(
description=f"[TensorFlow] StridedSlice {description}",
graph=graph,
backend=["webgpu", "webgl", "webassembly"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(axis, keep_dims, description: str = ""):
x = tf.placeholder(np.float32, [2, 3, 4, 5])
y = tf.reduce_sum(x, axis=axis, keep_dims=keep_dims)
vx = np.random.rand(2, 3, 4, 5).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
x = graph.inputs[0]
y = graph.outputs[0]
assert list(vy.shape) == list(y.shape), f"{vy.shape}, {y.shape}"
generate_kernel_test_case(
description=f"[TensorFlow] Sum {description}",
graph=graph,
backend=["webgpu", "webgl", "webassembly"],
inputs={x: vx},
expected={y: vy},
)
def template(x0_shape, x1_shape, description: str = ""):
x0 = tf.placeholder(np.float32, x0_shape, "x0")
x1 = tf.placeholder(np.float32, x1_shape, "x1")
y = x0 * x1
vx0 = np.random.rand(*x0_shape).astype(np.float32)
vx1 = np.random.rand(*x1_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x0: vx0, x1: vx1})
graph = TensorFlowConverter(sess, batch_size=2).convert([x0, x1], [y])
generate_kernel_test_case(
description=f"[TensorFlow] Mul {description}",
graph=graph,
inputs={
graph.inputs[0]: vx0,
graph.inputs[1]: vx1
},
expected={graph.outputs[0]: vy},
)
def template(x_shape=[2, 3, 4, 5], x_order=OrderNHWC, axis=Axis.C, keep_dims=False, description: str = ""):
x = tf.placeholder(np.float32, x_shape)
y = tf.reduce_min(x, axis=x_order.axes_dict[axis], keep_dims=keep_dims)
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
assert list(vy.shape) == list(graph.outputs[0].shape)
generate_kernel_test_case(
description=f"[TensorFlow] Min {description}",
graph=graph,
backend=["webgpu", "webassembly", "webgl"],
inputs={
graph.inputs[0]: vx
},
expected={graph.outputs[0]: vy},
)
def template(x_shape, y_shape, description: str = ""):
x = tf.placeholder(np.float32, x_shape, "x")
y = tf.placeholder(np.float32, y_shape, "y")
z = tf.squared_difference(x, y)
vx = np.random.rand(*x_shape).astype(np.float32)
vy = np.random.rand(*y_shape).astype(np.float32)
with tf.Session() as sess:
vz, = sess.run([z], {x: vx, y: vy})
graph = TensorFlowConverter(sess, batch_size=2).convert([x, y], [z])
generate_kernel_test_case(
description=f"[TensorFlow] SquaredDifference {description}",
graph=graph,
inputs={
graph.inputs[0]: vx,
graph.inputs[1]: vy
},
expected={graph.outputs[0]: vz},
)
def template(x_shape, begin, size, description: str = ""):
x = tf.placeholder(np.float32, x_shape)
y = tf.slice(x, begin, size)
vx = np.random.rand(*x_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx})
graph = TensorFlowConverter(sess, batch_size=2).convert([x], [y])
assert list(vy.shape) == list(
graph.outputs[0].shape
), f"vy_shape: {vy.shape}, y.shape: {graph.outputs[0].shape}"
generate_kernel_test_case(
description=f"[TensorFlow] Slice {description}",
graph=graph,
backend=["webgpu", "webgl", "webassembly"],
inputs={graph.inputs[0]: vx},
expected={graph.outputs[0]: vy},
)
def template(N=2,
H=5,
W=5,
Cin=7,
Cout=9,
KH=3,
KW=3,
strides=[1, 1, 1, 1],
padding="VALID",
data_format="NHWC",
description: str = ""):
x_shape = [N, H, W, Cin] if data_format == "NHWC" else [N, Cin, H, W]
w_shape = [KH, KW, Cin, Cout]
x = tf.placeholder(np.float32, x_shape)
w = tf.placeholder(np.float32, w_shape)
y = tf.nn.conv2d(x,
w,
strides=strides,
padding=padding,
data_format=data_format)
vx = np.random.rand(*x_shape).astype(np.float32)
vw = np.random.rand(*w_shape).astype(np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx, w: vw})
graph = TensorFlowConverter(sess, batch_size=2).convert([x, w], [y])
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"[TensorFlow] Conv2D {description}",
graph=graph,
inputs={
graph.inputs[0]: vx,
graph.inputs[1]: vw
},
expected={graph.outputs[0]: vy})
def template(x_shape=[2, 3, 4, 5], data_format="NHWC", description: str = ""):
from tensorflow.python.ops import nn
x = tf.placeholder(np.float32, x_shape)
b = tf.placeholder(np.float32,
x_shape[-1] if data_format == "NHWC" else x_shape[1])
y = nn.bias_add(x, b, data_format)
vx = np.random.rand(*x_shape).astype(np.float32)
vb = np.random.rand(
*[x_shape[-1] if data_format == "NHWC" else x_shape[1]]).astype(
np.float32)
with tf.Session() as sess:
vy, = sess.run([y], {x: vx, b: vb})
graph = TensorFlowConverter(sess, batch_size=2).convert([x, b], [y])
generate_kernel_test_case(
description=f"[TensorFlow] BiasAdd {description}",
graph=graph,
inputs={
graph.inputs[0]: vx,
graph.inputs[1]: vb
},
expected={graph.outputs[0]: vy},
)
