Skip to content

Commit

Permalink
fix pool sign error
Browse files Browse the repository at this point in the history 
Change-Id: I7f63193085c4564dfc7484eabe4a09ec7fe93e83
  • Loading branch information
@HarmonyHu
HarmonyHu committed Mar 8, 2023
1 parent 2de83ca commit 2a68827
Show file tree
Hide file tree
Showing 2 changed files with 128 additions and 66 deletions.
39 changes: 39 additions & 0 deletions lib/Conversion/TopToTpu/TopToTpuPass.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -107,6 +107,40 @@ struct ForwardCalibartion : public OpRewritePattern<TyOp> {
}
};

template <typename TyOp>
struct KeepSignPattern : public OpRewritePattern<TyOp> {
using OpRewritePattern<TyOp>::OpRewritePattern;

LogicalResult matchAndRewrite(TyOp op,
PatternRewriter &rewriter) const override {
Value in = op.getInput();
Value out = op.getOutput();
if (!module::isCalibratedType(in, out)) {
return failure();
}
auto in_qtype = module::getCalibratedType(in);
auto out_qtype = module::getCalibratedType(out);
float min;
if (in_qtype.getMin() < 0) {
if (out_qtype.getMin() < 0) {
return failure();
}
min = -out_qtype.getMax();
} else {
if (out_qtype.getMin() >= 0) {
return failure();
}
min = 0;
}
auto etype = module::getStorageType(out);
auto new_qtype =
quant::CalibratedQuantizedType::get(etype, min, out_qtype.getMax());
auto new_type = RankedTensorType::get(module::getShape(out), new_qtype);
out.setType(new_type);
return success();
}
};

template <typename TyOp, bool KeepMin = false>
struct BackwardCalibartion : public OpRewritePattern<TyOp> {
using OpRewritePattern<TyOp>::OpRewritePattern;
Expand Down Expand Up @@ -376,6 +410,11 @@ struct ConvertTopToTpu : public ::impl::ConvertTopToTpuBase<ConvertTopToTpu> {
patterns.add<ForwardCalibartion<top::AvgPoolOp>>(ctx_);
}
applyPatternsAndFoldGreedily(module_, std::move(patterns));
// keep sign for some ops
// backend not support in out not the same sign
patterns.clear();
patterns.add<KeepSignPattern<top::AvgPoolOp>>(ctx_);
applyPatternsAndFoldGreedily(module_, std::move(patterns));
}

void all_int8_process() {
Expand Down
155 changes: 89 additions & 66 deletions python/test/test_onnx.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,9 @@
import onnxruntime
import multiprocessing

BM1684X_Failed_Cases = ["QDQ", "QDQConv", "TorchArgmax", "TorchActivation", "TorchChannelShuffle", "TorchNonZero"]
BM1684X_Failed_Cases = [
"QDQ", "QDQConv", "TorchArgmax", "TorchActivation", "TorchChannelShuffle", "TorchNonZero"
]
CV18XX_Failed_Cases = [
"Conv3d", "Compare", "CompareConst", "Erf", "GRU3", "LeakyRelu", "LogSoftmax", "Reshape",
"ReshapeFuse", "PadEdge", "ScatterND", "Sqrt", "Sub2", "Where", "TopK", "TorchGelu", "TorchGRU",
Expand Down Expand Up @@ -152,13 +154,13 @@ def __init__(self, chip: str = "bm1684x", mode: str = "all", dynamic: bool = Tru
"TorchConv3dTranspose": self.test_TorchConv3dTranspose,
"TorchHardSwish": self.test_TorchHardSwish,
"TorchHardSigmoid": self.test_TorchHardSigmoid,
"TorchIdentity": self.test_TorchIdentity,
"TorchInstanceNorm": self.test_TorchInstanceNorm,
"TorchInstanceNorm2": self.test_TorchInstanceNorm2,
"TorchGelu": self.test_TorchGelu,
"TorchGroupNorm": self.test_TorchGroupNorm,
"TorchGroupNorm2": self.test_TorchGroupNorm2,
"TorchGRU": self.test_TorchGRU,
"TorchIdentity": self.test_TorchIdentity,
"TorchInstanceNorm": self.test_TorchInstanceNorm,
"TorchInstanceNorm2": self.test_TorchInstanceNorm2,
"TorchLayerGroup": self.test_TorchLayerGroup,
"TorchLayerNorm": self.test_TorchLayerNorm,
"TorchLayerNorm2": self.test_TorchLayerNorm2,
Expand All @@ -175,13 +177,11 @@ def __init__(self, chip: str = "bm1684x", mode: str = "all", dynamic: bool = Tru
#############################
# Special Pass test case, Alphabetically
#############################
"ArgReducefull": self.test_ArgReducefull,
"ConcatFuse": self.test_ConcatFuse,
"ConcatToSpace": self.test_ConcatToSpace,
"Conv3dTo2d": self.test_Conv3dTo2d,
"Div2Mul": self.test_Div2Mul,
"PermuteFuse": self.test_PermuteFuse,
"PixelNorm": self.test_PixelNorm,
"PixelNorm2": self.test_PixelNorm2,
"GatherToSlice": self.test_GatherToSlice,
"Mul2Scale": self.test_Mul2Scale,
"MatMulTranspose": self.test_MatMulTranspose,
Expand All @@ -191,18 +191,21 @@ def __init__(self, chip: str = "bm1684x", mode: str = "all", dynamic: bool = Tru
# "PadPool1d": self.test_PadPool1d,
# "PadPool2d": self.test_PadPool2d,
# "PadPool3d": self.test_PadPool3d,
"PixelNorm": self.test_PixelNorm,
"PixelNorm2": self.test_PixelNorm2,
"PermuteFuse": self.test_PermuteFuse,
"PermuteToReorg": self.test_PermuteToReorg,
"PermuteToReorg2": self.test_PermuteToReorg2,
"PermuteToReshape": self.test_PermuteToReshape,
"Permute5dSplit": self.test_Permute5dSplit,
"PoolSignError": self.test_PoolSignError,
"ReshapeFuse": self.test_ReshapeFuse,
"SwapDimInner": self.test_SwapDimInner,
"ReduceTranspose": self.test_ReduceTranspose,
"ReduceFusePattern": self.test_ReduceFusePattern,
"SwapDimInner": self.test_SwapDimInner,
"SliceToReverse": self.test_SliceToReverse,
"StaticDynMixed": self.test_StaticDynMixed,
"ReduceFusePattern": self.test_ReduceFusePattern,
"ArgReducefull": self.test_ArgReducefull,
"TransposeArg": self.test_TransposeArg,
"PermuteToReorg": self.test_PermuteToReorg,
"PermuteToReorg2": self.test_PermuteToReorg2,
"PermuteToReshape": self.test_PermuteToReshape,
"Permute5dSplit": self.test_Permute5dSplit,
}

# no quantization when quant_mode == "f32"
Expand Down Expand Up @@ -1800,6 +1803,7 @@ def test_Slice(self, case_name):
self.onnx_and_test(graph_def)

def test_Slice2(self, case_name):

class Model(nn.Module):

def __init__(self):
Expand Down Expand Up @@ -2031,9 +2035,8 @@ def forward(self, x, boxes):
y = torchvision.ops.roi_align(x, boxes, [8, 8])
return y

def gen_rand_rois(N, H, W, roi_num) -> torch.Tensor :
batch_indice = torch.randint(0, N, (roi_num,),
dtype=torch.int32).float()
def gen_rand_rois(N, H, W, roi_num) -> torch.Tensor:
batch_indice = torch.randint(0, N, (roi_num, ), dtype=torch.int32).float()
roi_xl = torch.rand(roi_num, dtype=torch.float32) * (W - 1)
roi_xh = torch.rand(roi_num, dtype=torch.float32) * (W - 1)
roi_yl = torch.rand(roi_num, dtype=torch.float32) * (H - 1)
Expand Down Expand Up @@ -2253,6 +2256,7 @@ def forward(self, x):
self.torch_and_test(input_data, Net(), case_name)

def test_TorchNonZero(self, case_name):

class Net(nn.Module):

def __init__(self):
Expand Down Expand Up @@ -2520,8 +2524,8 @@ def forward(self, a, b):
x = torch.where(a >= b, a, b)
return x

a = torch.randint(-128,127,(4, 3, 100, 100)).float()
b = torch.randint(-128,127,(4, 3, 100, 100)).float()
a = torch.randint(-128, 127, (4, 3, 100, 100)).float()
b = torch.randint(-128, 127, (4, 3, 100, 100)).float()
self.torch_and_test((a, b), Net(), case_name)

def test_TorchSize(self, case_name):
Expand Down Expand Up @@ -3562,9 +3566,8 @@ def test_TransposeArg(self, case_name):
perm=transpose_order)
arg_keepdims = False
arg_axis = 1
reduce_output_shape = [8,1,16,64]
arg_output = helper.make_tensor_value_info('output', TensorProto.INT64,
reduce_output_shape)
reduce_output_shape = [8, 1, 16, 64]
arg_output = helper.make_tensor_value_info('output', TensorProto.INT64, reduce_output_shape)
arg_max_def = helper.make_node(
'ArgMax',
['transpose_output'],
Expand All @@ -3577,33 +3580,32 @@ def test_TransposeArg(self, case_name):
self.onnx_and_test(graph_def)

def test_ArgReducefull(self, case_name):
input_shape = [2,3,4]
input_shape = [2, 3, 4]
arg_axis = 0
reduce_axes = [0]
reduce_axes_num = 1
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, input_shape)
output_shape = [1,3,4]
arg_output = helper.make_tensor_value_info('arg_output', TensorProto.INT64,
output_shape)
output_shape = [1, 3, 4]
arg_output = helper.make_tensor_value_info('arg_output', TensorProto.INT64, output_shape)
arg_def = helper.make_node("ArgMax",
inputs=['input'],
outputs=['arg_output'],
axis=arg_axis)
inputs=['input'],
outputs=['arg_output'],
axis=arg_axis)
reduce_output_1 = helper.make_tensor_value_info('reduce_output_1', TensorProto.FLOAT,
output_shape)
output_shape)
reduce_def_1 = helper.make_node("ReduceMax",
inputs=['input'],
outputs=['reduce_output_1'],
axes=reduce_axes)
inputs=['input'],
outputs=['reduce_output_1'],
axes=reduce_axes)
reduce_output_2 = helper.make_tensor_value_info('reduce_output_2', TensorProto.FLOAT,
output_shape)
output_shape)
reduce_def_2 = helper.make_node("ReduceMax",
inputs=['input'],
outputs=['reduce_output_2'],
axes=reduce_axes)
inputs=['input'],
outputs=['reduce_output_2'],
axes=reduce_axes)

graph_def = helper.make_graph([arg_def, reduce_def_1, reduce_def_2], case_name, [input],
[arg_output,reduce_output_1,reduce_output_2])
[arg_output, reduce_output_1, reduce_output_2])
self.onnx_and_test(graph_def)

# def test_LayerNorm(self, case_name):
Expand Down Expand Up @@ -3820,8 +3822,8 @@ def __init__(self):
super(Net, self).__init__()

def forward(self, x):
x = torch.sum(x,1)
x = torch.sum(x,1)
x = torch.sum(x, 1)
x = torch.sum(x, 1)
return x

x = torch.randn(2, 2, 3, 4).float()
Expand All @@ -3843,42 +3845,37 @@ def test_PermuteToReshape(self, case_name):
[output])
self.onnx_and_test(graph_def)


def test_PermuteToReorg(self, case_name):
input_shape = [1, 4, 6, 6]
output_shape = [1, 16, 3, 3]
input = helper.make_tensor_value_info(
'input', TensorProto.FLOAT, input_shape)
output = helper.make_tensor_value_info(
'output', TensorProto.FLOAT, output_shape)
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, input_shape)
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, output_shape)
rshape1 = [6]
rshape1_data = np.array([1, 4, 3, 2, 3, 2], dtype=np.int64)
r1 = helper.make_tensor('shape1', TensorProto.INT64, rshape1, rshape1_data)
reshape1_def = helper.make_node("Reshape", inputs=['input', 'shape1'], outputs=['out1'])
order = [0, 1, 3, 5, 2, 4]
permute_def = helper.make_node("Transpose", inputs=['out1'], outputs=['out2'], perm = order)
permute_def = helper.make_node("Transpose", inputs=['out1'], outputs=['out2'], perm=order)
rshape2 = [4]
rshape2_data = np.array(output_shape, dtype=np.int64)
r2 = helper.make_tensor('shape2', TensorProto.INT64, rshape2, rshape2_data)
reshape2_def = helper.make_node("Reshape", inputs=['out2', 'shape2'], outputs=['output'])
graph_def = helper.make_graph([reshape1_def, permute_def, reshape2_def],
case_name, [input], [output],
initializer=[r1, r2])
case_name, [input], [output],
initializer=[r1, r2])
self.onnx_and_test(graph_def)

def test_PermuteToReorg2(self, case_name):
input_shape = [1, 16, 200, 200]
output_shape = [1, 64, 100, 100]
input = helper.make_tensor_value_info(
'input', TensorProto.FLOAT, input_shape)
output = helper.make_tensor_value_info(
'output', TensorProto.FLOAT, output_shape)
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, input_shape)
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, output_shape)
rshape1 = [6]
rshape1_data = np.array([1, 16, 100, 2, 100, 2], dtype=np.int64)
r1 = helper.make_tensor('shape1', TensorProto.INT64, rshape1, rshape1_data)
reshape1_def = helper.make_node("Reshape", inputs=['input', 'shape1'], outputs=['out1'])
order = [0, 1, 3, 5, 2, 4]
permute_def = helper.make_node("Transpose", inputs=['out1'], outputs=['out2'], perm = order)
permute_def = helper.make_node("Transpose", inputs=['out1'], outputs=['out2'], perm=order)
rshape2 = [4]
rshape2_data = np.array(output_shape, dtype=np.int64)
r2 = helper.make_tensor('shape2', TensorProto.INT64, rshape2, rshape2_data)
Expand All @@ -3888,25 +3885,28 @@ def test_PermuteToReorg2(self, case_name):
kernel = [3, 3]
padding = [1, 1, 1, 1]
stride = [1, 1]
dilation = [1, 1]
dilation = [1, 1]
weight_data = np.random.randn(*filter_shape).astype(np.float32)
bias_data = np.random.randn(output_shape[1]).astype(np.float32)
weight = helper.make_tensor("weight", TensorProto.FLOAT, filter_shape, weight_data)
bias = helper.make_tensor("bias", TensorProto.FLOAT, list(bias_data.shape), bias_data)
conv_def = helper.make_node("Conv", inputs = ['out3', 'weight', 'bias'], outputs = ['output'],
kernel_shape = kernel,
pads = padding,
strides = stride,
dilations = dilation,
group = 1)
conv_def = helper.make_node("Conv",
inputs=['out3', 'weight', 'bias'],
outputs=['output'],
kernel_shape=kernel,
pads=padding,
strides=stride,
dilations=dilation,
group=1)
graph_def = helper.make_graph([reshape1_def, permute_def, reshape2_def, conv_def],
case_name, [input], [output],
initializer = [r1, r2, weight, bias])
case_name, [input], [output],
initializer=[r1, r2, weight, bias])
self.onnx_and_test(graph_def)

def test_Permute5dSplit(self, case_name):
input_shape = [2, 4, 16, 20, 32]
orders = [[0, 4, 2, 1, 3], [3, 1, 0, 4, 2], [2, 1, 3, 4, 0], [1, 0, 2, 4, 3], [4, 0, 3, 2, 1], [4, 3, 2, 1, 0]]
orders = [[0, 4, 2, 1, 3], [3, 1, 0, 4, 2], [2, 1, 3, 4, 0], [1, 0, 2, 4, 3],
[4, 0, 3, 2, 1], [4, 3, 2, 1, 0]]
for i in range(0, len(orders)):
order = orders[i]
if i >= 3 and self.chip.startswith("cv18"):
Expand All @@ -3916,11 +3916,34 @@ def test_Permute5dSplit(self, case_name):
output_shape = [input_shape[order[i]] for i in range(0, len(order))]
input = helper.make_tensor_value_info('input', TensorProto.FLOAT, input_shape)
output = helper.make_tensor_value_info('output', TensorProto.FLOAT, output_shape)
permute_def = helper.make_node("Transpose", inputs = ['input'],
outputs = ['output'], perm = order)
graph_def = helper.make_graph([permute_def], "{}_{}".format(case_name, i), [input], [output])
permute_def = helper.make_node("Transpose",
inputs=['input'],
outputs=['output'],
perm=order)
graph_def = helper.make_graph([permute_def], "{}_{}".format(case_name, i), [input],
[output])
self.onnx_and_test(graph_def)

def test_PoolSignError(self, case_name):

class Net(torch.nn.Module):

def __init__(self):
super(Net, self).__init__()
self.relu = nn.ReLU()
self.conv = nn.Conv2d(8, 8, 3, 2, 1)
self.pool = nn.AvgPool2d(2, 2)

def forward(self, x):
a = self.relu(x)
b = self.pool(a)
c = self.conv(x)
d = b + c
return d

x = torch.randn(4, 8, 32, 32).float()
self.torch_and_test(x, Net(), case_name)


def test_one_case_in_all(tester: ONNX_IR_TESTER, case, error_cases, success_cases):
try:
Expand Down

0 comments on commit 2a68827

Please sign in to comment.