AndrewZhaoLuo commented on a change in pull request #9126:
URL: https://github.com/apache/tvm/pull/9126#discussion_r719108112
##########
File path: python/tvm/relay/frontend/paddlepaddle.py
##########
@@ -17,73 +17,178 @@
# pylint: disable=invalid-name, import-self, len-as-condition,
unused-argument, too-many-lines
# pylint: disable=import-outside-toplevel
"""Paddle: PArallel Distributed Deep LEarning."""
-import warnings
import numpy as np
import tvm
from tvm.ir import IRModule
from .. import analysis
+from .. import ty as _ty
from .. import expr as _expr
from .. import function as _function
from .. import ty as _ty
from .. import op as _op
from .common import (
fold_constant,
+ get_relay_op,
infer_shape,
infer_type,
infer_value,
+ try_infer_value,
new_var,
)
__all__ = ["from_paddle"]
+def _get_pad_size(in_size, dilated_kernel_size, stride_size):
+ """Calculate the paddings size."""
Review comment:
Should describe padding size for what
##########
File path: python/tvm/relay/frontend/paddlepaddle.py
##########
@@ -191,57 +307,80 @@ def convert_dropout(g, op, block):
"""Operator converter for dropout."""
x = g.get_node(op.input("X")[0])
- out = _op.copy(x)
+ g.add_node(op.output("Out")[0], x)
+
+
+def convert_dot(g, op, block):
+ """Operator converter for dot."""
+
+ # x, y should be 1D or 2D tensor
+ # when it's 2D tensor, the first dimension means batch dimension
+ x = g.get_node(op.input("X")[0])
+ y = g.get_node(op.input("Y")[0])
+
+ out = _op.sum(_op.multiply(x, y), axis=[-1], keepdims=True)
g.add_node(op.output("Out")[0], out)
def convert_elementwise_op(g, op, block):
"""Operator converter for all the elementwise operators."""
op_map = {
- "elementwise_div": lambda x, y: x / y,
- "elementwise_add": lambda x, y: x + y,
- "elementwise_mul": lambda x, y: x * y,
- "elementwise_sub": lambda x, y: x - y,
- "elementwise_mod": lambda x, y: x % y,
+ "elementwise_div": "divide",
+ "elementwise_add": "add",
+ "elementwise_mul": "multiply",
+ "elementwise_sub": "subtract",
+ "elementwise_mod": "mod",
+ "elementwise_max": "maximum",
+ "elementwise_min": "minimum",
+ "elementwise_pow": "power",
+ "elementwise_floordiv": "floor_divide",
+ "equal": "equal",
+ "greater_equal": "greater_equal",
+ "greater_than": "greater",
+ "less_equal": "less_equal",
+ "less_than": "less",
+ "not_equal": "not_equal",
}
op_func = op_map[op.type]
ipt0 = g.get_node(op.input("X")[0])
ipt1 = g.get_node(op.input("Y")[0])
- ipt0_shape = block.var(op.input("X")[0]).shape
- ipt1_shape = block.var(op.input("Y")[0]).shape
+ ipt0_shape = infer_shape(ipt0)
+ ipt1_shape = infer_shape(ipt1)
axis = op.attr("axis")
if len(ipt0_shape) != len(ipt1_shape):
if axis < 0:
axis = axis + len(ipt0_shape)
if axis != len(ipt0_shape) - 1:
ipt1 = _op.expand_dims(ipt1, axis=axis,
num_newaxis=(len(ipt0_shape) - axis - 1))
+ op_func = get_relay_op(op_func)
out = op_func(ipt0, ipt1)
g.add_node(op.output("Out")[0], out)
-def convert_equal(g, op, block):
- """Operator converter for equal."""
+def convert_expand(g, op, block):
+ """Operator converter for expand."""
x = g.get_node(op.input("X")[0])
- y = g.get_node(op.input("Y")[0])
- out = _op.equal(x, y)
+ if op.input("Shape"):
+ sizes = g.get_node(op.input("Shape")[0])
+ sizes = try_infer_value(sizes, g.get_params())[0]
+ else:
+ sizes = op.attr("shape")
+
+ if isinstance(sizes, np.ndarray):
+ sizes = sizes.tolist()
+
+ out = _op.broadcast_to(x, sizes)
g.add_node(op.output("Out")[0], out)
-def convert_activation(g, op, block):
- """Operator converter for all the activation."""
+def convert_expand_as(g, op, block):
+ """Operator converter for expand_as."""
- op_map = {
- "exp": _op.exp,
- "relu": _op.nn.relu,
- "tanh": _op.tanh,
- "sqrt": _op.sqrt,
- "erf": _op.erf,
- "abs": _op.abs,
- }
- act_func = op_map[op.type]
- out = act_func(g.get_node(op.input("X")[0]))
+ x = g.get_node(op.input("X")[0])
+ target_shape = op.attr("target_shape")
+ out = _op.broadcast_to(x, target_shape)
Review comment:
I believe you might run into a similar issue as
https://github.com/apache/tvm/blob/main/python/tvm/relay/frontend/onnx.py#L2257
--
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]
