This is an automated email from the ASF dual-hosted git repository.
jroesch pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 3ebd353 [VTA] Make vta graph_pack compatible with latest TVM, and
bring back object detection tutorials. (#8731)
3ebd353 is described below
commit 3ebd353a7f526cdf21293055a00eeeabe6efae1f
Author: Hua Jiang <[email protected]>
AuthorDate: Sat Aug 14 17:48:50 2021 -0700
[VTA] Make vta graph_pack compatible with latest TVM, and bring back object
detection tutorials. (#8731)
* [VTA] Make vta graph_pack compatible with latest TVM, and bring back
object detection tutorials.
* remove deploy_detection.py.
* move out deploy_detection.py from legacy folder.
* fix build error.
---
vta/python/vta/top/graphpack.py | 33 ++++++++++++++++++----
.../frontend/{legacy => }/deploy_detection.py | 7 ++---
2 files changed, 30 insertions(+), 10 deletions(-)
diff --git a/vta/python/vta/top/graphpack.py b/vta/python/vta/top/graphpack.py
index a982b88..f15e492 100644
--- a/vta/python/vta/top/graphpack.py
+++ b/vta/python/vta/top/graphpack.py
@@ -56,13 +56,24 @@ def _pack_batch_channel(data, dshape, bfactor, cfactor):
return data
-def _unpack_batch_channel(data, old_shape):
+def _unpack_batch_channel(data, old_shape, unpack_transpose=False):
"""Unpack the data channel dimension."""
- data = op.transpose(data, axes=(0, 4, 1, 5, 2, 3))
+ if unpack_transpose:
+ data = op.transpose(data, axes=(0, 4, 1, 5, 2, 3))
data = op.reshape(data, newshape=old_shape)
return data
+def _channel_const_match(channel_length, cfactor_out):
+ """Round the chanel const variant if the value not divisible by
cfactor_out"""
+ diff = int(channel_length) % cfactor_out
+ if diff != 0:
+ diff = cfactor_out - diff
+ channel_length = channel_length + diff
+
+ return diff, channel_length
+
+
def _const_shape_match(data, dshape, cfactor_out):
"""Pad the constant if the shape[0] not divisible by cfactor_out."""
assert len(dshape) == 3
@@ -299,6 +310,7 @@ class ExprPack(ExprMutator):
self.upsampling = op.op.get("nn.upsampling")
self.reshape = op.op.get("reshape")
self.number_of_conv2d = 0
+ self.unpack_transpose = True
super().__init__()
def visit_call(self, call):
@@ -319,7 +331,7 @@ class ExprPack(ExprMutator):
self.start_pack = False
data = args[0]
data_shape = _get_tensor_shape(call.args[0])
- return _unpack_batch_channel(data, data_shape)
+ return _unpack_batch_channel(data, data_shape,
self.unpack_transpose)
if self.start_pack:
# Operator cases
if call.op == self.conv2d and odtype == "int32":
@@ -429,12 +441,12 @@ class ExprPack(ExprMutator):
if len(pad_width) == 6:
pass
elif len(pad_width) == 4:
- (data,) = args
+ (data, pad_value) = args
new_pad_width = []
new_pad_width.extend(pad_width)
for _ in range(2):
new_pad_width.append([0, 0])
- return op.nn.pad(data, pad_value=call.attrs.pad_value,
pad_width=new_pad_width)
+ return op.nn.pad(data, pad_value=pad_value,
pad_width=new_pad_width)
elif call.op == self.upsampling:
(data,) = args
scale_h = call.attrs.scale_h
@@ -445,8 +457,17 @@ class ExprPack(ExprMutator):
return op.nn.upsampling(data, scale_h, scale_w, data_layout,
method, align_corners)
elif call.op == self.reshape and len(input_types[0].shape) == 4:
(data,) = args
+ self.unpack_transpose = False
data = op.transpose(data, axes=(0, 4, 1, 5, 2, 3))
- return op.reshape(data, [int(x) for x in input_types[0].shape])
+ new_shape = [int(x) for x in input_types[0].shape]
+ # Check if the reshape match with such shape after pad
+ pad, new_shape[1] = _channel_const_match(new_shape[1],
self.cfactor)
+ data = op.reshape(data, new_shape)
+ # remove pad data
+ if pad != 0:
+ new_pad_width = [[0, 0], [0, -pad], [0, 0], [0, 0]]
+ data = op.nn.pad(data, pad_width=new_pad_width)
+ return data
return relay.Call(self.visit(call.op), args, call.attrs)
diff --git a/vta/tutorials/frontend/legacy/deploy_detection.py
b/vta/tutorials/frontend/deploy_detection.py
similarity index 99%
rename from vta/tutorials/frontend/legacy/deploy_detection.py
rename to vta/tutorials/frontend/deploy_detection.py
index 1d78786..7718018 100644
--- a/vta/tutorials/frontend/legacy/deploy_detection.py
+++ b/vta/tutorials/frontend/deploy_detection.py
@@ -177,9 +177,9 @@ else:
# Get execution context from remote
ctx = remote.ext_dev(0) if device == "vta" else remote.cpu(0)
-####################################
+#####################################
# Build the inference graph executor.
-# ----------------------------------
+# -----------------------------------
# Using Darknet library load downloaded vision model and compile with Relay.
# The compilation steps are:
#
@@ -191,7 +191,6 @@ ctx = remote.ext_dev(0) if device == "vta" else
remote.cpu(0)
# 5. Perform relay build to object file.
# 6. Load the object file onto remote (FPGA device).
# 7. Generate graph executor, `m`.
-#
# Load pre-configured AutoTVM schedules
with autotvm.tophub.context(target):
@@ -212,7 +211,7 @@ with autotvm.tophub.context(target):
# Note: We set opt_level to 3 in order to fold batch norm
with tvm.transform.PassContext(opt_level=3):
with relay.quantize.qconfig(
- global_scale=33.0,
+ global_scale=23.0,
skip_conv_layers=[0],
store_lowbit_output=True,
round_for_shift=True,
