Aten::Index converter #2277
Merged
Aten::Index converter #2277
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
github-actions
bot
added
component: api [Python]
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
There are some changes that do not conform to Python style guidelines:
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-08-29 21:00:52.894681+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-08-29 21:03:31.334736+00:00
@@ -69,28 +69,28 @@
network: TRTNetwork,
target: Target,
source_ir: Optional[SourceIR],
name: str,
input: TRTTensor,
- index: Union[TRTTensor, Sequence[TRTTensor]]
+ index: Union[TRTTensor, Sequence[TRTTensor]],
) -> TRTTensor:
adv_indx_indices = []
tensor_indices = []
for i in len(index):
ind = index[i]
- #FIXME: check if the datatype for the indices needs to be casted to INT32
- #TRTInterpretor should take care
+ # FIXME: check if the datatype for the indices needs to be casted to INT32
+ # TRTInterpretor should take care
adv_indx_indices.append(i)
tensor_indices.append(ind)
if not tensor_indices:
identity_layer = network.add_identity(input)
identity_layer.set_output_type(0, trt.int32)
set_layer_name(identity_layer, target, name + "_index_identity", source_ir)
return identity_layer.get_output(0)
- elif (len(tensor_indices) == 1):
+ elif len(tensor_indices) == 1:
indices_tensor = tensor_indices[0]
gather_layer = network.add_gather(input, indices_tensor, adv_indx_indices[0])
set_layer_name(gather_layer, target, name + "_index_gather", source_ir)
return gather_layer.get_output(0)
else:
@@ -99,7 +99,5 @@
adv_indx_count = len(adv_indx_indices)
input_shape_layer = network.add_shape(input)
set_layer_name(input_shape_layer, target, name + "_index_shape", source_ir)
input_shape_tensor = input_shape_layer.get_output(0)
return input_shape_tensor.get_output(0)
-
-
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py 2023-08-29 21:00:52.894681+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py 2023-08-29 21:03:31.387359+00:00
@@ -169,11 +169,11 @@
)
@dynamo_tensorrt_converter(torch.ops.aten.index.Tensor)
def aten_ops_index(
-network: TRTNetwork,
+ network: TRTNetwork,
target: Target,
args: Tuple[Argument, ...],
kwargs: Dict[str, Argument],
name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
apbose
force-pushed
the
dynamo_converter_index
branch
from
71ca151 to
302b962
Compare
There was a problem hiding this comment.
There are some changes that do not conform to Python style guidelines:
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py 2023-09-01 17:57:45.620889+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/aten_ops_converters.py 2023-09-01 18:00:04.480860+00:00
@@ -169,11 +169,11 @@
)
@dynamo_tensorrt_converter(torch.ops.aten.index.Tensor)
def aten_ops_index(
-network: TRTNetwork,
+ network: TRTNetwork,
target: Target,
args: Tuple[Argument, ...],
kwargs: Dict[str, Argument],
name: str,
) -> Union[TRTTensor, Sequence[TRTTensor]]:
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-09-01 17:57:45.620889+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-09-01 18:00:04.530477+00:00
@@ -71,28 +71,28 @@
network: TRTNetwork,
target: Target,
source_ir: Optional[SourceIR],
name: str,
input: TRTTensor,
- index: Union[TRTTensor, Sequence[TRTTensor]]
+ index: Union[TRTTensor, Sequence[TRTTensor]],
) -> TRTTensor:
adv_indx_indices = []
tensor_indices = []
for i in len(index):
ind = index[i]
- #FIXME: check if the datatype for the indices needs to be casted to INT32
- #TRTInterpretor should take care
+ # FIXME: check if the datatype for the indices needs to be casted to INT32
+ # TRTInterpretor should take care
adv_indx_indices.append(i)
tensor_indices.append(ind)
if not tensor_indices:
identity_layer = network.add_identity(input)
identity_layer.set_output_type(0, trt.int32)
set_layer_name(identity_layer, target, name + "_index_identity", source_ir)
return identity_layer.get_output(0)
- elif (len(tensor_indices) == 1):
+ elif len(tensor_indices) == 1:
indices_tensor = tensor_indices[0]
gather_layer = network.add_gather(input, indices_tensor, adv_indx_indices[0])
set_layer_name(gather_layer, target, name + "_index_gather", source_ir)
return gather_layer.get_output(0)
else:
@@ -102,24 +102,26 @@
input_shape_layer = network.add_shape(input)
set_layer_name(input_shape_layer, target, name + "_index_shape", source_ir)
input_shape_tensor = input_shape_layer.get_output(0)
dim_tensor_list = []
for i in range(rank):
- #check this
- dim_tensor_layer = network.add_gather(input_shape_tensor, i ,0)
- set_layer_name(input_shape_layer, target, name + "_index_gather_rank", source_ir)
+ # check this
+ dim_tensor_layer = network.add_gather(input_shape_tensor, i, 0)
+ set_layer_name(
+ input_shape_layer, target, name + "_index_gather_rank", source_ir
+ )
dim_tensor = dim_tensor_layer.get_output(0)
dim_tensor_list.append(dim_tensor)
- #for cases like
- #t: [x_1, y_1, y_2, ..., x_m, ..., y_n] -> t: [x_1, x_2, ..., x_m, y_1, y_2, ..., y_n],
- #where t is a tensor of rank m+n, {x_i} are axes where tensor index is provided, and {y_i} are axes
- #for ":"
- #Examples: x.shape = (10,20,30,40,50)
- #ind_1, ind_2 broadcasted to (2,3,4)
- #x[:, ind_1, ind_2] = 10, 2, 3, 4, 40, 50
- #x[:,ind_1, :, ind_2] = 2, 3, 4, 10, 30, 50
+ # for cases like
+ # t: [x_1, y_1, y_2, ..., x_m, ..., y_n] -> t: [x_1, x_2, ..., x_m, y_1, y_2, ..., y_n],
+ # where t is a tensor of rank m+n, {x_i} are axes where tensor index is provided, and {y_i} are axes
+ # for ":"
+ # Examples: x.shape = (10,20,30,40,50)
+ # ind_1, ind_2 broadcasted to (2,3,4)
+ # x[:, ind_1, ind_2] = 10, 2, 3, 4, 40, 50
+ # x[:,ind_1, :, ind_2] = 2, 3, 4, 10, 30, 50
transpose_layer = network.add_shuffle(input)
new_order = []
for i in range(adv_indx_count):
new_order.append(adv_indx_indices[i])
for i in range(rank):
@@ -130,166 +132,194 @@
permute_order(new_order)
transpose_layer.set_second_transpose(permute_order)
set_layer_name(transpose_layer, target, name + "_index_transpose", source_ir)
transpose_tensor = transpose_layer.get_output(0)
- #Flatten [x_1, x_2,.......x_m, y_1, y_2,.....y_m]
+ # Flatten [x_1, x_2,.......x_m, y_1, y_2,.....y_m]
transpose_tensor_shape = network.add_shape(transpose_tensor)
d0 = 1
d0 = get_trt_tensor(network, d0, "d0_initial")
for i in range(adv_indx_count):
dim_tensor_layer = network.add_gather(transpose_tensor_shape, i, 0)
- set_layer_name(dim_tensor_layer, target, name + "_index_gather_concatOne", source_ir)
+ set_layer_name(
+ dim_tensor_layer, target, name + "_index_gather_concatOne", source_ir
+ )
d0_gather = gather_layer.get_output(0)
mult_d0 = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_concatOne_shape",
- trt.ElementWisePROD,
- mult_d0,
- d0_gather,
- )
-
+ network,
+ target,
+ source_ir,
+ name + "index_concatOne_shape",
+ trt.ElementWisePROD,
+ mult_d0,
+ d0_gather,
+ )
+
d1 = 1
d1 = get_trt_tensor(network, d0, "d0_initial")
for i in range(adv_indx_count, rank):
dim_tensor_layer = network.add_gather(transpose_tensor_shape, i, 0)
- set_layer_name(dim_tensor_layer, target, name + "_index_gather_concatTwo", source_ir)
+ set_layer_name(
+ dim_tensor_layer, target, name + "_index_gather_concatTwo", source_ir
+ )
d1_gather = gather_layer.get_output(0)
mult_d1 = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_concatTwo_shape",
- trt.ElementWisePROD,
+ network,
+ target,
+ source_ir,
+ name + "index_concatTwo_shape",
+ trt.ElementWisePROD,
mult_d1,
d1_gather,
)
concat_tensor_layer = network.add_concatenation([mult_d0, mult_d1])
set_layer_name(concat_tensor_layer, target, name + "_index_Concat", source_ir)
concat_tensor = concat_tensor_layer.get_output(0)
reshape_layer = network.add_shuffle(transpose_tensor)
- #check this
+ # check this
reshape_layer.set_input(1, concat_tensor)
flatten_tensor = reshape_layer.get_output(0)
- #tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j)), ind_i is input indices[i], x_j is the
- #// j dimension of input x.
- multiplier = get_trt_tensor(network, dim_tensor_list[adv_indx_indices[adv_indx_count - 1]], "dim_last")
+ # tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j)), ind_i is input indices[i], x_j is the
+ # // j dimension of input x.
+ multiplier = get_trt_tensor(
+ network, dim_tensor_list[adv_indx_indices[adv_indx_count - 1]], "dim_last"
+ )
cum_adv_index = tensor_indices[adv_indx_count - 1]
- for i in range(adv_indx_count-2, 0):
+ for i in range(adv_indx_count - 2, 0):
adv_index = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_intermediate",
- trt.ElementWisePROD,
+ network,
+ target,
+ source_ir,
+ name + "index_intermediate",
+ trt.ElementWisePROD,
multiplier,
tensor_indices[i],
)
cum_adv_index = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_sum_intermediate",
- trt.ElementWiseSUM,
+ network,
+ target,
+ source_ir,
+ name + "index_sum_intermediate",
+ trt.ElementWiseSUM,
cum_adv_index,
adv_index,
)
multiplier = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_intermediate",
- trt.ElementWisePROD,
+ network,
+ target,
+ source_ir,
+ name + "index_intermediate",
+ trt.ElementWisePROD,
multiplier,
dim_tensor_list[adv_indx_count[i]],
)
gather_layer_element = network.add_gather(flatten_tensor, cum_adv_index, 0)
- set_layer_name(gather_layer_element, target, name + "_index_gather_element", source_ir)
+ set_layer_name(
+ gather_layer_element, target, name + "_index_gather_element", source_ir
+ )
gather_out = gather_layer.get_output(0)
cum_adv_index_shape_tensor = cum_adv_index.add_shape(cum_adv_index_shape_tensor)
- #check if all advanced indices are consecutive
+ # check if all advanced indices are consecutive
concat_tensor_reshape = []
- if(adv_indx_count == adv_indx_indices[adv_indx_count - 1] - adv_indx_indices[0] + 1):
- #concat_tensor_reshape_initial = -1
- #concat_tensor_reshape_initial_tensor = get_trt_tensor(network, concat_tensor_reshape_initial, "concat_tensor_reshape_initial")
+ if (
+ adv_indx_count
+ == adv_indx_indices[adv_indx_count - 1] - adv_indx_indices[0] + 1
+ ):
+ # concat_tensor_reshape_initial = -1
+ # concat_tensor_reshape_initial_tensor = get_trt_tensor(network, concat_tensor_reshape_initial, "concat_tensor_reshape_initial")
concat_tensor_reshape.append(-1)
for i in range(0, rank):
if i not in adv_indx_indices:
curr_dim = dim_tensor_list[i]
concat_tensor_reshape.append(curr_dim)
-
+
concat_tensor_layer = network.add_concatenation(concat_tensor_reshape)
- set_layer_name(concat_tensor_layer, target, name + "_index_Concat_reshape", source_ir)
+ set_layer_name(
+ concat_tensor_layer, target, name + "_index_Concat_reshape", source_ir
+ )
concat_tensor = concat_tensor_layer.get_output(0)
regular_index_shuffle_layer = network.add_shuffle(gather_out)
- set_layer_name(regular_index_shuffle_layer, target, name + "_index_regular_index", source_ir)
+ set_layer_name(
+ regular_index_shuffle_layer,
+ target,
+ name + "_index_regular_index",
+ source_ir,
+ )
unfold_tensor = regular_index_shuffle_layer.get_output(0)
transpose_advanced_shuffle_layer = network.add_shuffle(unfold_tensor)
new_order = []
- for i in range(1, adv_indx_count[0]+1):
+ for i in range(1, adv_indx_count[0] + 1):
new_order.append(i)
new_order.append(0)
- for i in range(adv_indx_indices[0]+1, rank - adv_indx_count):
+ for i in range(adv_indx_indices[0] + 1, rank - adv_indx_count):
new_order.append(i)
permute_order = trt.Permutation()
permute_order(new_order)
transpose_advanced_shuffle_layer.set_second_transpose(permute_order)
- set_layer_name(transpose_advanced_shuffle_layer, target, name + "_index_advanced_shuffle_transpose", source_ir)
+ set_layer_name(
+ transpose_advanced_shuffle_layer,
+ target,
+ name + "_index_advanced_shuffle_transpose",
+ source_ir,
+ )
transpose_tensor = transpose_advanced_shuffle_layer.get_output(0)
- #unfold advanced layer
+ # unfold advanced layer
concat_final_tensor = []
for i in range(0, adv_indx_indices[0]):
current_dim = dim_tensor_list[i]
concat_final_tensor.push_back(curr_dim)
concat_final_tensor.push_back(cum_adv_index_shape_tensor)
for i in range(adv_indx_indices[0], rank):
- if(i not in (adv_indx_indices)):
+ if i not in (adv_indx_indices):
current_dim = dim_tensor_list[i]
concat_final_tensor.append(current_dim)
-
+
concat_final_shape_layer = network.add_concatenation(concat_final_tensor)
- set_layer_name(concat_final_shape_layer, target, name + "_index_concat_final_shape_layer", source_ir)
+ set_layer_name(
+ concat_final_shape_layer,
+ target,
+ name + "_index_concat_final_shape_layer",
+ source_ir,
+ )
concat_final_tensor = concat_final_shape_layer.get_output(0)
unfold_advanced_shuffle_layer = network.add_shuffle(transpose_tensor)
- #check this
+ # check this
reshape_layer.set_input(1, concat_final_tensor)
reshape_output = reshape_layer.get_output(0)
-
+
else:
- concat_tensor= []
+ concat_tensor = []
for i in range(0, rank):
if i not in adv_indx_indices:
curr_dim = dim_tensor_list[i]
concat_tensor.append(curr_dim)
-
+
concat_layer = network.add_concatenation(concat_tensor)
- set_layer_name(concat_layer, target, name + "_index_concat_final_shape_layer", source_ir)
+ set_layer_name(
+ concat_layer,
+ target,
+ name + "_index_concat_final_shape_layer",
+ source_ir,
+ )
concat_final_tensor = concat_final_shape_layer.get_output(0)
reshape_layer = network.add_shuffle(gather_out)
reshape_layer.setInput(1, concat_final_tensor)
- set_layer_name(reshape_layer, target, name + "_index_shuffle_final_shape_layer", source_ir)
+ set_layer_name(
+ reshape_layer,
+ target,
+ name + "_index_shuffle_final_shape_layer",
+ source_ir,
+ )
reshape_output = reshape_layer.get_output(0)
return reshape_output
-
-
-
-
-
-
-
-
-
-
apbose
force-pushed
the
dynamo_converter_index
branch
from
4f2a738 to
42798cc
Compare
There was a problem hiding this comment.
There are some changes that do not conform to Python style guidelines:
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-09-07 03:07:48.113307+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-09-07 03:10:32.769440+00:00
@@ -71,28 +71,28 @@
network: TRTNetwork,
target: Target,
source_ir: Optional[SourceIR],
name: str,
input: TRTTensor,
- index: Union[TRTTensor, Sequence[TRTTensor]]
+ index: Union[TRTTensor, Sequence[TRTTensor]],
) -> TRTTensor:
adv_indx_indices = []
tensor_indices = []
for i in len(index):
ind = index[i]
- #FIXME: check if the datatype for the indices needs to be casted to INT32
- #TRTInterpretor should take care
+ # FIXME: check if the datatype for the indices needs to be casted to INT32
+ # TRTInterpretor should take care
adv_indx_indices.append(i)
tensor_indices.append(ind)
if not tensor_indices:
identity_layer = network.add_identity(input)
identity_layer.set_output_type(0, trt.int32)
set_layer_name(identity_layer, target, name + "_index_identity", source_ir)
return identity_layer.get_output(0)
- elif (len(tensor_indices) == 1):
+ elif len(tensor_indices) == 1:
indices_tensor = tensor_indices[0]
gather_layer = network.add_gather(input, indices_tensor, adv_indx_indices[0])
set_layer_name(gather_layer, target, name + "_index_gather", source_ir)
return gather_layer.get_output(0)
else:
@@ -102,24 +102,26 @@
input_shape_layer = network.add_shape(input)
set_layer_name(input_shape_layer, target, name + "_index_shape", source_ir)
input_shape_tensor = input_shape_layer.get_output(0)
dim_tensor_list = []
for i in range(rank):
- #check this
- dim_tensor_layer = network.add_gather(input_shape_tensor, i ,0)
- set_layer_name(input_shape_layer, target, name + "_index_gather_rank", source_ir)
+ # check this
+ dim_tensor_layer = network.add_gather(input_shape_tensor, i, 0)
+ set_layer_name(
+ input_shape_layer, target, name + "_index_gather_rank", source_ir
+ )
dim_tensor = dim_tensor_layer.get_output(0)
dim_tensor_list.append(dim_tensor)
- #for cases like
- #t: [x_1, y_1, y_2, ..., x_m, ..., y_n] -> t: [x_1, x_2, ..., x_m, y_1, y_2, ..., y_n],
- #where t is a tensor of rank m+n, {x_i} are axes where tensor index is provided, and {y_i} are axes
- #for ":"
- #Examples: x.shape = (10,20,30,40,50)
- #ind_1, ind_2 broadcasted to (2,3,4)
- #x[:, ind_1, ind_2] = 10, 2, 3, 4, 40, 50
- #x[:,ind_1, :, ind_2] = 2, 3, 4, 10, 30, 50
+ # for cases like
+ # t: [x_1, y_1, y_2, ..., x_m, ..., y_n] -> t: [x_1, x_2, ..., x_m, y_1, y_2, ..., y_n],
+ # where t is a tensor of rank m+n, {x_i} are axes where tensor index is provided, and {y_i} are axes
+ # for ":"
+ # Examples: x.shape = (10,20,30,40,50)
+ # ind_1, ind_2 broadcasted to (2,3,4)
+ # x[:, ind_1, ind_2] = 10, 2, 3, 4, 40, 50
+ # x[:,ind_1, :, ind_2] = 2, 3, 4, 10, 30, 50
transpose_layer = network.add_shuffle(input)
new_order = []
for i in range(adv_indx_count):
new_order.append(adv_indx_indices[i])
for i in range(rank):
@@ -130,166 +132,194 @@
permute_order(new_order)
transpose_layer.set_second_transpose(permute_order)
set_layer_name(transpose_layer, target, name + "_index_transpose", source_ir)
transpose_tensor = transpose_layer.get_output(0)
- #Flatten [x_1, x_2,.......x_m, y_1, y_2,.....y_m]
+ # Flatten [x_1, x_2,.......x_m, y_1, y_2,.....y_m]
transpose_tensor_shape = network.add_shape(transpose_tensor)
d0 = 1
d0 = get_trt_tensor(network, d0, "d0_initial")
for i in range(adv_indx_count):
dim_tensor_layer = network.add_gather(transpose_tensor_shape, i, 0)
- set_layer_name(dim_tensor_layer, target, name + "_index_gather_concatOne", source_ir)
+ set_layer_name(
+ dim_tensor_layer, target, name + "_index_gather_concatOne", source_ir
+ )
d0_gather = gather_layer.get_output(0)
mult_d0 = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_concatOne_shape",
- trt.ElementWisePROD,
- mult_d0,
- d0_gather,
- )
-
+ network,
+ target,
+ source_ir,
+ name + "index_concatOne_shape",
+ trt.ElementWisePROD,
+ mult_d0,
+ d0_gather,
+ )
+
d1 = 1
d1 = get_trt_tensor(network, d0, "d0_initial")
for i in range(adv_indx_count, rank):
dim_tensor_layer = network.add_gather(transpose_tensor_shape, i, 0)
- set_layer_name(dim_tensor_layer, target, name + "_index_gather_concatTwo", source_ir)
+ set_layer_name(
+ dim_tensor_layer, target, name + "_index_gather_concatTwo", source_ir
+ )
d1_gather = gather_layer.get_output(0)
mult_d1 = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_concatTwo_shape",
- trt.ElementWisePROD,
+ network,
+ target,
+ source_ir,
+ name + "index_concatTwo_shape",
+ trt.ElementWisePROD,
mult_d1,
d1_gather,
)
concat_tensor_layer = network.add_concatenation([mult_d0, mult_d1])
set_layer_name(concat_tensor_layer, target, name + "_index_Concat", source_ir)
concat_tensor = concat_tensor_layer.get_output(0)
reshape_layer = network.add_shuffle(transpose_tensor)
- #check this
+ # check this
reshape_layer.set_input(1, concat_tensor)
flatten_tensor = reshape_layer.get_output(0)
- #tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j)), ind_i is input indices[i], x_j is the
- #// j dimension of input x.
- multiplier = get_trt_tensor(network, dim_tensor_list[adv_indx_indices[adv_indx_count - 1]], "dim_last")
+ # tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j)), ind_i is input indices[i], x_j is the
+ # // j dimension of input x.
+ multiplier = get_trt_tensor(
+ network, dim_tensor_list[adv_indx_indices[adv_indx_count - 1]], "dim_last"
+ )
cum_adv_index = tensor_indices[adv_indx_count - 1]
- for i in range(adv_indx_count-2, 0):
+ for i in range(adv_indx_count - 2, 0):
adv_index = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_intermediate",
- trt.ElementWisePROD,
+ network,
+ target,
+ source_ir,
+ name + "index_intermediate",
+ trt.ElementWisePROD,
multiplier,
tensor_indices[i],
)
cum_adv_index = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_sum_intermediate",
- trt.ElementWiseSUM,
+ network,
+ target,
+ source_ir,
+ name + "index_sum_intermediate",
+ trt.ElementWiseSUM,
cum_adv_index,
adv_index,
)
multiplier = convert_binary_elementwise(
- network,
- target,
- source_ir,
- name + "index_intermediate",
- trt.ElementWisePROD,
+ network,
+ target,
+ source_ir,
+ name + "index_intermediate",
+ trt.ElementWisePROD,
multiplier,
dim_tensor_list[adv_indx_count[i]],
)
gather_layer_element = network.add_gather(flatten_tensor, cum_adv_index, 0)
- set_layer_name(gather_layer_element, target, name + "_index_gather_element", source_ir)
+ set_layer_name(
+ gather_layer_element, target, name + "_index_gather_element", source_ir
+ )
gather_out = gather_layer.get_output(0)
cum_adv_index_shape_tensor = cum_adv_index.add_shape(cum_adv_index_shape_tensor)
- #check if all advanced indices are consecutive
+ # check if all advanced indices are consecutive
concat_tensor_reshape = []
- if(adv_indx_count == adv_indx_indices[adv_indx_count - 1] - adv_indx_indices[0] + 1):
- #concat_tensor_reshape_initial = -1
- #concat_tensor_reshape_initial_tensor = get_trt_tensor(network, concat_tensor_reshape_initial, "concat_tensor_reshape_initial")
+ if (
+ adv_indx_count
+ == adv_indx_indices[adv_indx_count - 1] - adv_indx_indices[0] + 1
+ ):
+ # concat_tensor_reshape_initial = -1
+ # concat_tensor_reshape_initial_tensor = get_trt_tensor(network, concat_tensor_reshape_initial, "concat_tensor_reshape_initial")
concat_tensor_reshape.append(-1)
for i in range(0, rank):
if i not in adv_indx_indices:
curr_dim = dim_tensor_list[i]
concat_tensor_reshape.append(curr_dim)
-
+
concat_tensor_layer = network.add_concatenation(concat_tensor_reshape)
- set_layer_name(concat_tensor_layer, target, name + "_index_Concat_reshape", source_ir)
+ set_layer_name(
+ concat_tensor_layer, target, name + "_index_Concat_reshape", source_ir
+ )
concat_tensor = concat_tensor_layer.get_output(0)
regular_index_shuffle_layer = network.add_shuffle(gather_out)
- set_layer_name(regular_index_shuffle_layer, target, name + "_index_regular_index", source_ir)
+ set_layer_name(
+ regular_index_shuffle_layer,
+ target,
+ name + "_index_regular_index",
+ source_ir,
+ )
unfold_tensor = regular_index_shuffle_layer.get_output(0)
transpose_advanced_shuffle_layer = network.add_shuffle(unfold_tensor)
new_order = []
- for i in range(1, adv_indx_count[0]+1):
+ for i in range(1, adv_indx_count[0] + 1):
new_order.append(i)
new_order.append(0)
- for i in range(adv_indx_indices[0]+1, rank - adv_indx_count):
+ for i in range(adv_indx_indices[0] + 1, rank - adv_indx_count):
new_order.append(i)
permute_order = trt.Permutation()
permute_order(new_order)
transpose_advanced_shuffle_layer.set_second_transpose(permute_order)
- set_layer_name(transpose_advanced_shuffle_layer, target, name + "_index_advanced_shuffle_transpose", source_ir)
+ set_layer_name(
+ transpose_advanced_shuffle_layer,
+ target,
+ name + "_index_advanced_shuffle_transpose",
+ source_ir,
+ )
transpose_tensor = transpose_advanced_shuffle_layer.get_output(0)
- #unfold advanced layer
+ # unfold advanced layer
concat_final_tensor = []
for i in range(0, adv_indx_indices[0]):
current_dim = dim_tensor_list[i]
concat_final_tensor.push_back(curr_dim)
concat_final_tensor.push_back(cum_adv_index_shape_tensor)
for i in range(adv_indx_indices[0], rank):
- if(i not in (adv_indx_indices)):
+ if i not in (adv_indx_indices):
current_dim = dim_tensor_list[i]
concat_final_tensor.append(current_dim)
-
+
concat_final_shape_layer = network.add_concatenation(concat_final_tensor)
- set_layer_name(concat_final_shape_layer, target, name + "_index_concat_final_shape_layer", source_ir)
+ set_layer_name(
+ concat_final_shape_layer,
+ target,
+ name + "_index_concat_final_shape_layer",
+ source_ir,
+ )
concat_final_tensor = concat_final_shape_layer.get_output(0)
unfold_advanced_shuffle_layer = network.add_shuffle(transpose_tensor)
- #check this
+ # check this
reshape_layer.set_input(1, concat_final_tensor)
reshape_output = reshape_layer.get_output(0)
-
+
else:
- concat_tensor= []
+ concat_tensor = []
for i in range(0, rank):
if i not in adv_indx_indices:
curr_dim = dim_tensor_list[i]
concat_tensor.append(curr_dim)
-
+
concat_layer = network.add_concatenation(concat_tensor)
- set_layer_name(concat_layer, target, name + "_index_concat_final_shape_layer", source_ir)
+ set_layer_name(
+ concat_layer,
+ target,
+ name + "_index_concat_final_shape_layer",
+ source_ir,
+ )
concat_final_tensor = concat_final_shape_layer.get_output(0)
reshape_layer = network.add_shuffle(gather_out)
reshape_layer.setInput(1, concat_final_tensor)
- set_layer_name(reshape_layer, target, name + "_index_shuffle_final_shape_layer", source_ir)
+ set_layer_name(
+ reshape_layer,
+ target,
+ name + "_index_shuffle_final_shape_layer",
+ source_ir,
+ )
reshape_output = reshape_layer.get_output(0)
return reshape_output
-
-
-
-
-
-
-
-
-
-
--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_index_aten.py 2023-09-07 03:07:48.137309+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/conversion/test_index_aten.py 2023-09-07 03:10:36.917639+00:00
@@ -2,10 +2,11 @@
import torch.nn as nn
from torch.testing._internal.common_utils import run_tests
from torch_tensorrt import Input
from .harness import DispatchTestCase
+
class TestIndexConverter(DispatchTestCase):
def test_index(self):
class TestModule(nn.Module):
def forward(self, x):
@@ -13,6 +14,6 @@
index0 = torch.randint(0, 16, (1, 16))
index1 = torch.randint(0, 16, (1, 16))
out = torch.ops.aten.index(None, None, index0, index1)
inputs = [torch.randn(1, 10)]
- self.run_test(TestModule(), inputs, expected_ops={torch.ops.aten.index.Tensor})
\ No newline at end of file
+ self.run_test(TestModule(), inputs, expected_ops={torch.ops.aten.index.Tensor})
apbose
force-pushed
the
dynamo_converter_index
branch
from
42798cc to
6b186b6
Compare
gs-olive
reviewed
Sep 8, 2023
Contributor
gs-olive
left a comment
gs-olive
left a comment
There was a problem hiding this comment.
Overall looks good! Added some style/naming comments, and will run tests on a model which uses this layer, to verify
Comment on lines
87
to
88
| for i in range(0, len(index)): | ||
| ind = index[i] |
Contributor
There was a problem hiding this comment.
Consider rewriting as: for i, ind in enumerate(index)
Comment on lines
141
to
142
| permute_order = trt.Permutation(new_order) | ||
| transpose_layer.second_transpose = permute_order |
Contributor
There was a problem hiding this comment.
Can shorten to transpose_layer.second_transpose = tuple(new_order)
| target, | ||
| source_ir, | ||
| name + "index_intermediate", | ||
| trt.ElementWisePROD, |
Contributor
There was a problem hiding this comment.
Should be trt.ElementWiseOperation.PROD
| target, | ||
| source_ir, | ||
| name + "index_sum_intermediate", | ||
| trt.ElementWiseSUM, |
Contributor
There was a problem hiding this comment.
trt.ElementWiseOperation.SUM
| target, | ||
| source_ir, | ||
| name + "index_intermediate", | ||
| trt.ElementWisePROD, |
Contributor
There was a problem hiding this comment.
trt.ElementWiseOperation.PROD
Comment on lines
265
to
266
| permute_order = trt.Permutation(new_order) | ||
| transpose_advanced_shuffle_layer.second_transpose = permute_order |
Contributor
There was a problem hiding this comment.
permute_order could be replaced with tuple(new_order)
gs-olive
reviewed
Sep 8, 2023
| ) -> TRTTensor: | ||
| adv_indx_indices = [] | ||
| tensor_indices = [] | ||
| _LOGGER.debug(f"The index shape is", index.shape) |
Contributor
There was a problem hiding this comment.
index.shape is not valid, since index could be a python list
There was a problem hiding this comment.
There are some changes that do not conform to Python style guidelines:
--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-09-25 20:21:10.951934+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/conversion/impl/select.py 2023-09-25 20:24:07.893237+00:00
@@ -94,11 +94,13 @@
_LOGGER.debug(f"Shape of {i} index is {ind.shape}")
adv_indx_indices.append(i)
# torch.nn.parameter.Parameter=> torch.Tensor
ind = get_trt_tensor(network, ind, name + f"_parameter_to_fp32_tensor_{i}")
if last_index is not None:
- assert broadcastable(ind, last_index), "The indices should be broadcastable!"
+ assert broadcastable(
+ ind, last_index
+ ), "The indices should be broadcastable!"
last_index = ind
tensor_indices.append(ind)
if not tensor_indices:
identity_layer = network.add_identity(input)
@@ -177,11 +179,13 @@
_LOGGER.debug(f"The flatten tensor shape is {flatten_tensor.shape}")
# tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j)), ind_i is input indices[i], x_j is the
# // j dimension of input x.
multiplier = get_trt_tensor(
- network, dim_tensor_list[adv_indx_indices[adv_indx_count - 1]], name + "dim_last"
+ network,
+ dim_tensor_list[adv_indx_indices[adv_indx_count - 1]],
+ name + "dim_last",
)
cum_adv_index = tensor_indices[adv_indx_count - 1]
for i in range(adv_indx_count - 2, -1, -1):
adv_index = convert_binary_elementwise(
network,
@@ -231,11 +235,13 @@
if (
adv_indx_count
== adv_indx_indices[adv_indx_count - 1] - adv_indx_indices[0] + 1
):
_LOGGER.debug(f"The indices are continuous in this case")
- concat_tensor_reshape.append(get_trt_tensor(network, -1, name + "dynamic_concat"))
+ concat_tensor_reshape.append(
+ get_trt_tensor(network, -1, name + "dynamic_concat")
+ )
for i in range(0, rank):
if i not in adv_indx_indices:
curr_dim = dim_tensor_list[i]
concat_tensor_reshape.append(curr_dim)
apbose
force-pushed
the
dynamo_converter_index
branch
from
7d215af to
fcbd767
Compare
gs-olive
suggested changes
Sep 28, 2023
|
|
||
| import torch | ||
| import torch.nn as nn | ||
| from harness import DispatchTestCase |
gs-olive
reviewed
Sep 29, 2023
…adcast and broadcasting cases
… for non continuous indices
apbose
force-pushed
the
dynamo_converter_index
branch
from
742c7c2 to
709d626
Compare
gs-olive
suggested changes
Oct 3, 2023
Contributor
gs-olive
left a comment
gs-olive
left a comment
There was a problem hiding this comment.
Added a few suggestions for input data types and imports
Comment on lines
15
to
20
| from torch_tensorrt.fx.converters.converter_utils import ( | ||
| get_positive_dim, | ||
| has_dynamic_shape, | ||
| set_layer_name, | ||
| to_numpy, | ||
| ) |
Contributor
There was a problem hiding this comment.
Switch get_positive_dim and to_numpy to the torch_tensorrt.dynamo.conversion.converter_utils version
| ) | ||
|
|
||
|
|
||
| @dynamo_tensorrt_converter(torch.ops.aten.index.Tensor) |
Contributor
There was a problem hiding this comment.
Consider adding @enforce_tensor_types( {0: (TRTTensor,)} ), to ensure the input is a TRTTensor
… to_numpy to dynamo converter_utils
gs-olive
approved these changes
Oct 4, 2023
Contributor
gs-olive
left a comment
gs-olive
left a comment
There was a problem hiding this comment.
Looks good to me! Approved, pending CI
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Aten::index converter
#2231