def instancenorm_3_constraints(pattern):
epsilon_var = _get_var(pattern.add_epsilon, pattern.mean1)
gamma_var = mb.const(
val=np.ones(shape=(1, pattern.root_var.shape[1], 1, 1)), name="gamma_var"
)
beta_var = mb.const(
val=np.zeros(shape=(1, pattern.root_var.shape[1], 1, 1)),
name="_fuse_layernorm_or_instancenorm_beta",
)
passed = True
passed = passed and _check_reduce_op(pattern.main_reduce)
passed = passed and pattern.sub.x == pattern.root_var and pattern.sub.y == pattern.main_reduce.outputs[0]
passed = passed and _check_reduce_op(pattern.mean1)
passed = passed and pattern.mul_sub.y.val is not None and pattern.mul_sub.y.val == -1
passed = passed and _general_constraints(pattern, epsilon_var, gamma_var, beta_var)
passed = passed and _instancenorm_constraints(pattern)
return passed
def linear_prog(input):
W = mb.const(val=np.random.rand(100, 5000), name="const_W")
out = mb.linear(x=input, weight=W, name="output")
return out
def transform_pattern(pattern):
bias_value = _get_bias_var(pattern).val
is_conv_op = (pattern.conv.op_type == "conv")
is_bias_scalar = False
if not isinstance(bias_value, np.ndarray):
is_bias_scalar = True
bias_value = np.array([bias_value
]) if is_bias_scalar else np.squeeze(bias_value)
if pattern.add_or_sub.op_type == "sub":
bias_value *= -1
# everything looks good, now find the new updated bias
old_bias = pattern.conv.inputs.get("bias", None)
old_bias_value = None
if old_bias is not None and old_bias.val is not None:
old_bias_value = old_bias.val
if old_bias is None:
# need to create a fresh numpy array for bias
if np.prod(bias_value.shape) == 1:
# its a scalar bias
# need to find the value of Cout to form a new bias
# conv_transpose has weight format [K, C_out, spatial dims]
# conv has weight format [C_out, K, spatial dims]
Cout = pattern.conv.weight.val.shape[0 if is_conv_op else 1]
new_bias_value = np.broadcast_to(bias_value, (Cout, ))
else:
new_bias_value = bias_value
else:
# just need to update the existing bias array
new_bias_value = old_bias_value + bias_value
# create a new conv op with the new bias value, copying rest of the attributes
out_name = pattern.add_or_sub.outputs[0].name
if new_bias_value.dtype != np.float32 and new_bias_value.dtype != np.float16:
# cast the bias to match the weight type
weight_np_type = types.nptype_from_builtin(
pattern.conv.inputs["weight"].sym_type.get_primitive())
logging.warning(
"conv_bias_fusion pass: casting bias "
"from {} to {} to match the dtype of the weight of the conv layer".
format(new_bias_value.dtype, weight_np_type))
new_bias_value = new_bias_value.astype(weight_np_type)
new_bias_var = mb.const(val=new_bias_value, before_op=pattern.conv)
conv_kargs = {
"bias": new_bias_var,
"name": out_name,
"before_op": pattern.conv
}
for k, v in pattern.conv.inputs.items():
if k == "bias":
continue
conv_kargs[k] = v
if is_conv_op:
x = mb.conv(**conv_kargs)
else:
x = mb.conv_transpose(**conv_kargs)
pattern.add_or_sub.enclosing_block.replace_uses_of_var_after_op(
anchor_op=pattern.add_or_sub,
old_var=pattern.add_or_sub.outputs[0],
new_var=x)
# Remove all the ops at once
pattern.block.remove_ops(pattern.op_list())
