Skip to content

tensorrt7 to tensorrt #37

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mosout wants to merge 3 commits into
base: main
Choose a base branch
from
Open

tensorrt7 to tensorrt #37

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
c29fa10
tensorrt7 to tensorrt
mosout Jun 23, 2021
ef45f6e
refine quantization conversion
mosout Jun 27, 2021
bdac569
fix fake quantization
mosout Jul 8, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 5 additions & 4 deletions examples/tensorrt_qat/models.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -47,7 +47,7 @@ def Lenet(data):
pool2 = flow.nn.max_pool2d(
conv2, ksize=2, strides=2, padding="VALID", name="pool2", data_format="NCHW"
)
reshape = flow.reshape(pool2, [pool2.shape[0], -1])
reshape = flow.reshape(pool2, [pool2.shape[0], -1], name="reshape")
hidden = flow.layers.dense(
reshape,
512,
Expand All @@ -57,6 +57,7 @@ def Lenet(data):
)
return flow.layers.dense(hidden, 10, kernel_initializer=initializer, name="dense2")


def _get_regularizer(model_name):
# all decay
return flow.regularizers.l2(0.00004)
Expand Down Expand Up @@ -337,7 +338,7 @@ def build_network(
name="pool5",
)
fc = flow.layers.dense(
flow.reshape(pool, (pool.shape[0], -1)),
flow.reshape(pool, (pool.shape[0], -1), name="reshape"),
units=class_num,
use_bias=False,
kernel_initializer=_get_initializer("dense_weight"),
Expand Down Expand Up @@ -392,7 +393,7 @@ def get_lenet_job_function(
func_config.qat.symmetric(True)
func_config.qat.per_channel_weight_quantization(False)
func_config.qat.moving_min_max_stop_update_after_iters(1000)
func_config.qat.target_backend("tensorrt7")
func_config.qat.target_backend("tensorrt")
if func_type == "train":

@flow.global_function(type="train", function_config=func_config)
Expand Down Expand Up @@ -438,7 +439,7 @@ def get_mobilenet_job_function(
func_config.qat.symmetric(True)
func_config.qat.per_channel_weight_quantization(False)
func_config.qat.moving_min_max_stop_update_after_iters(1000)
func_config.qat.target_backend("tensorrt7")
func_config.qat.target_backend("tensorrt")
if func_type == "train":

@flow.global_function(type="train", function_config=func_config)
Expand Down
16 changes: 11 additions & 5 deletions oneflow_onnx/oneflow2onnx/handlers/quantize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -90,8 +90,14 @@ def get_min_or_max_value(get_min: bool, pre_func: Optional[Callable] = None):
raise ValueError("invalid quantization formula: " + formula)

ctx.RemoveNode(node.name)
ctx.MakeConst(node.output_tensor_names[0], scale)
ctx.MakeConst(node.output_tensor_names[1], zero_point)
ctx.MakeConst(
node.output_tensor_names[0],
scale.squeeze() if formula == "cambricon" or per_layer else scale,
)
ctx.MakeConst(
node.output_tensor_names[1],
zero_point.squeeze() if formula == "cambricon" or per_layer else zero_point,
)

@classmethod
def Version_10(cls, ctx: Graph, node: Node, **kwargs):
Expand Down Expand Up @@ -144,8 +150,8 @@ def Version_10(cls, ctx: Graph, node: Node, **kwargs):
raise ValueError("invalid quantization formula: " + formula)

ctx.RemoveNode(node.name)
ctx.MakeConst(node.output_tensor_names[0], scale.flatten())
ctx.MakeConst(node.output_tensor_names[1], zero_point)
ctx.MakeConst(node.output_tensor_names[0], scale.squeeze())
ctx.MakeConst(node.output_tensor_names[1], zero_point.squeeze())


@flow_op(
Expand All @@ -167,7 +173,7 @@ def _Convert(cls, ctx: Graph, node: Node, opset: int, **kwargs):
)
if opset < 13:
scale_shape = ctx.get_shape(node.input_tensor_names[1])
if not (len(scale_shape) == 1 and scale_shape[0] == 1):
if not len(scale_shape) == 0:
raise RuntimeError("per-channel mode is not supported in version 10")

else:
Expand Down