forked from mindspore-Ecosystem/mindspore
support pad mode for conv2d transpose
This commit is contained in:
parent
982033b6a1
commit
4061c11e18
|
@ -892,9 +892,10 @@ ReplaceGraphPtr Conv2DInfo::replace_graph(const CNodePtr &cnode) {
|
|||
int64_t all_send_lens = std::accumulate(send_lens_.begin(), send_lens_.end(), 0);
|
||||
int64_t all_recv_lens = std::accumulate(recv_lens_.begin(), recv_lens_.end(), 0);
|
||||
if (all_send_lens + all_recv_lens == 0) {
|
||||
int64_t pad_mode = 0; // 0 is "pad" mode
|
||||
auto prim = GetValueNode<PrimitivePtr>(cnode->input(0));
|
||||
prim->set_attr(OUT_CHANNEL, MakeValue(new_out_channel_));
|
||||
prim->set_attr(PAD_MODE, MakeValue(PAD));
|
||||
prim->set_attr(PAD_MODE, MakeValue(pad_mode)); // need to use int64_t to define pad_mode
|
||||
prim->set_attr(PAD, MakeValue(new_pad_list_));
|
||||
MS_LOG(INFO) << name_ << ": the send lens and recv lens is 0, no need exchange data";
|
||||
return nullptr;
|
||||
|
@ -986,12 +987,12 @@ Status Conv2DBackpropInputInfo::CheckStrategy(const StrategyPtr &strategy) {
|
|||
}
|
||||
}
|
||||
|
||||
// kernel size larger than stride and the h/w dimension is split, need to exchange overlap
|
||||
if ((kernel_size_use_dilation_[0] > stride_[2]) && (input_strategy[2] > 1)) {
|
||||
// if the h/w dimension is split, need to exchange overlap
|
||||
if (input_strategy[2] > 1) {
|
||||
h_dim_need_exchange_overlap_ = true;
|
||||
}
|
||||
|
||||
if ((kernel_size_use_dilation_[1] > stride_[3]) && (input_strategy[3] > 1)) {
|
||||
if (input_strategy[3] > 1) {
|
||||
w_dim_need_exchange_overlap_ = true;
|
||||
}
|
||||
return SUCCESS;
|
||||
|
@ -1002,21 +1003,11 @@ Status Conv2DBackpropInputInfo::CheckHWStrategy(int64_t h_strategy, int64_t w_st
|
|||
return FAILED;
|
||||
}
|
||||
|
||||
if (pad_mode_ != 1) { // only support same mode
|
||||
if (pad_mode_ != 0 && pad_mode_ != 1) { // only support pad mode and same mode
|
||||
MS_LOG(ERROR) << name_ << ": Do not support the pad mode " << pad_mode_ << " when split H or W dimension";
|
||||
return FAILED;
|
||||
}
|
||||
|
||||
if (h_strategy > 1 && inputs_shape_[0][2] * stride_[2] != outputs_shape_[0][2]) {
|
||||
MS_LOG(ERROR) << name_ << ": Do not support to split h dimension when in_shape * stride != out_shape";
|
||||
return FAILED;
|
||||
}
|
||||
|
||||
if (w_strategy > 1 && inputs_shape_[0][3] * stride_[3] != outputs_shape_[0][3]) {
|
||||
MS_LOG(ERROR) << name_ << ": Do not support to split w dimension when in_shape * stride != out_shape";
|
||||
return FAILED;
|
||||
}
|
||||
|
||||
return SUCCESS;
|
||||
}
|
||||
|
||||
|
@ -1265,8 +1256,9 @@ void Conv2DBackpropInputInfo::InferNewPadListByDimension(const std::string &dime
|
|||
// if (o/n + k - o + ws - s - x) is divisible by s, real_left_pad = s - 1.
|
||||
// otherwise, real_left_pad = (o/n + k - o + ws - s - x) % s - 1
|
||||
// 3) the middle rank:
|
||||
// if (r*on - k + x + 1) is divisible by s, real_left_pad = 0.
|
||||
// otherwise, real_left_pad = s - (r*on - k + x + 1) % s
|
||||
// if (r*o/n - k + x + 1) < 0, real_left_pad = -(r*o/n - k + x + 1);
|
||||
// otherwise, if (r*o/n - k + x + 1) is divisible by s, real_left_pad = 0.
|
||||
// otherwise, real_left_pad = s - (r*o/n - k + x + 1) % s
|
||||
int64_t current_rank_required_size = 0;
|
||||
int64_t real_top_or_left_pad = 0;
|
||||
int64_t h_or_w_output_shape = -1;
|
||||
|
@ -1324,7 +1316,9 @@ void Conv2DBackpropInputInfo::InferNewPadListByDimension(const std::string &dime
|
|||
|
||||
int64_t tmp =
|
||||
h_or_w_rank_bias * h_or_w_output_shape / h_or_w_dim_shard_num - h_or_w_kernel_size + top_or_left_pad + 1;
|
||||
if (tmp % h_or_w_stride == 0) {
|
||||
if (tmp < 0) {
|
||||
real_top_or_left_pad = -tmp;
|
||||
} else if (tmp % h_or_w_stride == 0) {
|
||||
real_top_or_left_pad = 0;
|
||||
} else {
|
||||
real_top_or_left_pad = h_or_w_stride - tmp % h_or_w_stride;
|
||||
|
|
|
@ -41,11 +41,12 @@ class Net(Cell):
|
|||
|
||||
|
||||
class Net2(Cell):
|
||||
def __init__(self, conv2d_weight, out_channel, kernel_size, pad_mode, stride, group=1, dilation=1,
|
||||
def __init__(self, conv2d_weight, out_channel, kernel_size, pad_mode, stride, pad=0, group=1, dilation=1,
|
||||
strategy1=None, strategy2=None):
|
||||
super().__init__()
|
||||
self.conv2d_transpose = P.Conv2DTranspose(out_channel=out_channel, kernel_size=kernel_size, pad_mode=pad_mode,
|
||||
stride=stride, group=group, dilation=dilation).shard(strategy1)
|
||||
stride=stride, pad=pad, group=group,
|
||||
dilation=dilation).shard(strategy1)
|
||||
self.neg = P.Neg().shard(strategy2)
|
||||
self.weight = Parameter(conv2d_weight, "w1")
|
||||
|
||||
|
@ -61,6 +62,9 @@ _w2 = Tensor(np.ones([8, 16, 4, 4]), dtype=ms.float32)
|
|||
_w3 = Tensor(np.ones([8, 16, 10, 10]), dtype=ms.float32)
|
||||
_w4 = Tensor(np.ones([8, 16, 3, 3]), dtype=ms.float32)
|
||||
_w5 = Tensor(np.ones([8, 8, 4, 4]), dtype=ms.float32)
|
||||
_w6 = Tensor(np.ones([8, 16, 5, 5]), dtype=ms.float32)
|
||||
_w7 = Tensor(np.ones([8, 16, 1, 1]), dtype=ms.float32)
|
||||
_w8 = Tensor(np.ones([8, 16, 4, 4]), dtype=ms.float32)
|
||||
_b = Tensor(np.ones([32, 16, 8, 8]), dtype=ms.float32)
|
||||
|
||||
|
||||
|
@ -158,7 +162,7 @@ def test_conv2d_transpose_model_parallel3():
|
|||
def test_conv2d_transpose_model_parallel4():
|
||||
"""
|
||||
Feature: test model parallel strategy
|
||||
Description: shard batch dimension, channel dimension and w dimension
|
||||
Description: shard h dimension and w dimension
|
||||
Expectation: compile success
|
||||
"""
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
|
||||
|
@ -187,15 +191,14 @@ def test_conv2d_transpose_split_h_or_w_in_pad_mode():
|
|||
"""
|
||||
Feature: test pad mode
|
||||
Description: shard batch dimension, channel dimension and w dimension in pad mode
|
||||
Expectation: compile failed
|
||||
Expectation: compile success
|
||||
"""
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=16, global_rank=0)
|
||||
strategy1 = ((2, 2, 1, 4), (2, 1, 1, 1))
|
||||
strategy2 = ((2, 2, 1, 4),)
|
||||
net = Net2(_w1, out_channel=8, kernel_size=(2, 2), pad_mode="pad", stride=2,
|
||||
strategy1=strategy1, strategy2=strategy2)
|
||||
with pytest.raises(RuntimeError):
|
||||
compile_net(net)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_conv2d_transpose_split_h_in_same_mode():
|
||||
|
@ -225,3 +228,59 @@ def test_conv2d_transpose_overlap_size_too_large():
|
|||
strategy1=strategy1, strategy2=strategy2)
|
||||
with pytest.raises(RuntimeError):
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_conv2d_transpose_pad_mode_no_need_exchange():
|
||||
"""
|
||||
Feature: pad mode, and two direction send, w = 8, o = 16, s = 2, k = 1, n = 8, pad = (0, 0, 0, 0)
|
||||
Description: shard h and w dimension
|
||||
Expectation: compile success
|
||||
"""
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=64, global_rank=13)
|
||||
strategy1 = ((1, 1, 8, 8), (1, 1, 1, 1))
|
||||
strategy2 = ((8, 1, 1, 1),)
|
||||
net = Net2(_w7, out_channel=8, kernel_size=1, pad_mode="pad", pad=(0, 0, 0, 0), stride=2, strategy1=strategy1,
|
||||
strategy2=strategy2)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_conv2d_transpose_pad_mode_two_direction_send_all_slice_pad_different():
|
||||
"""
|
||||
Feature: pad mode, and two direction send, w = 8, o = 16, s = 2, k = 5, n = 8, pad = (1, 2, 1, 2)
|
||||
Description: shard h and w dimension
|
||||
Expectation: compile success
|
||||
"""
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=64, global_rank=13)
|
||||
strategy1 = ((1, 1, 8, 8), (1, 1, 1, 1))
|
||||
strategy2 = ((8, 1, 1, 1),)
|
||||
net = Net2(_w6, out_channel=8, kernel_size=5, pad_mode="pad", pad=(1, 2, 1, 2), stride=2, strategy1=strategy1,
|
||||
strategy2=strategy2)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_conv2d_transpose_pad_mode_two_direction_send_all_slice():
|
||||
"""
|
||||
Feature: pad mode, and two direction send, w = 8, o = 16, s = 2, k = 4, n = 8, pad = (1, 1, 1, 1)
|
||||
Description: shard h and w dimension
|
||||
Expectation: compile success
|
||||
"""
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=64, global_rank=13)
|
||||
strategy1 = ((1, 1, 8, 8), (1, 1, 1, 1))
|
||||
strategy2 = ((8, 1, 1, 1),)
|
||||
net = Net2(_w8, out_channel=8, kernel_size=4, pad_mode="pad", pad=(1, 1, 1, 1), stride=2, strategy1=strategy1,
|
||||
strategy2=strategy2)
|
||||
compile_net(net)
|
||||
|
||||
|
||||
def test_conv2d_transpose_pad_mode_single_direction_send():
|
||||
"""
|
||||
Feature: pad mode, and single direction send, w = 8, o = 16, s = 2, k = 3, n = 8, pad = (0, 1, 0, 1)
|
||||
Description: shard h and w dimension
|
||||
Expectation: compile success
|
||||
"""
|
||||
context.set_auto_parallel_context(parallel_mode="semi_auto_parallel", device_num=64, global_rank=13)
|
||||
strategy1 = ((1, 1, 8, 8), (1, 1, 1, 1))
|
||||
strategy2 = ((8, 1, 1, 1),)
|
||||
net = Net2(_w4, out_channel=8, kernel_size=3, pad_mode="pad", pad=(0, 1, 0, 1), stride=2, strategy1=strategy1,
|
||||
strategy2=strategy2)
|
||||
compile_net(net)
|
||||
|
|
Loading…
Reference in New Issue