• Hardware Platform (Jetson)
• DeepStream Version 7.1
• JetPack Version (6.2)
• Issue Type(Question)
Hi everyone,
I’m building a DeepStream Python pipeline based on the multi-input multi-output python sample with an integrated tracker. My goal is to run a ResNet34 gaze-estimation ONNX model model repo as a secondary inference on each face detected by a PeopleNet primary GIE that should output pitch and yaw. The engine builds at runtime, but my pad probe never sees any tensor buffers (always buffer address = 0).
This is the structure of the pipeline
[ uridecodebin ]
↓
[ nvstreammux ] ──► queue1 ──► pgie (PeopleNet, unique-id=1) ──► nvtracker ──► sgie (ResNet34 gaze, unique-id=2) ──► nvstreamdemux ──► [ display bins… ]
SGIE File
[property]
gpu-id=0
batch-size=1
network-type=1
network-mode=0
onnx-file=/opt/nvidia/deepstream/deepstream-7.1/models/gaze-estimation/resnet34_gaze.onnx
process-mode=2
gie-unique-id=2
operate-on-gie-id=1
operate-on-class-ids=2
output-tensor-meta=1
Here is my probe
def sgie_src_pad_buffer_probe(pad, info, u_data):
#print("[DBG] → ENTER SGIE probe")
gst_buffer = info.get_buffer()
if not gst_buffer:
#print("[DBG] no GstBuffer!")
return Gst.PadProbeReturn.OK
# 1) get batch metadata
batch_meta = pyds.gst_buffer_get_nvds_batch_meta(hash(gst_buffer))
"""if not batch_meta:
print("[DBG] no batch_meta!") """
l_frame = batch_meta.frame_meta_list
"""if not l_frame:
print("[DBG] no frames in this batch!") # <-- (D)"""
# 2) for each frame
while l_frame:
frame_meta = pyds.NvDsFrameMeta.cast(l_frame.data)
#print(f"[DBG] frame #{frame_meta.frame_num} has {frame_meta.num_obj_meta} objs")
# 3) walk all object metas (e.g. faces) in this frame
l_obj = frame_meta.obj_meta_list
while l_obj:
obj_meta = pyds.NvDsObjectMeta.cast(l_obj.data)
count = 0
u = obj_meta.obj_user_meta_list
while u:
count += 1
u = u.next
#print(f"[DBG] obj_id={obj_meta.object_id} class={obj_meta.class_id} has {count} user_meta entries") # <-- (F)
# only process faces
if obj_meta.class_id == PGIE_CLASS_ID_FACE:
# 4) each object has its own user_meta_list
l_user = obj_meta.obj_user_meta_list
# index 0 will be pitch, 1 will be yaw
tensor_idx = 0
while l_user:
user_meta = pyds.NvDsUserMeta.cast(l_user.data)
#print(f"[DBG] user_meta_type={user_meta.base_meta.meta_type}")
# here i changed to nvdsinfer
if user_meta.base_meta.meta_type == pyds.NvDsMetaType.NVDSINFER_TENSOR_OUTPUT_META:
infer_meta = pyds.NvDsInferTensorMeta.cast(user_meta.user_meta_data)
#print(f"[DBG] num_output_layers={infer_meta.num_output_layers}")
#print(range(infer_meta.num_output_layers))
for layer_idx in range(infer_meta.num_output_layers):
#print(f"[DBG] layer_idx={layer_idx}")
layer_info = infer_meta.output_layers_info(layer_idx)
#print(layer_info)
# pick up dims (older binding uses layer_info.dims)
dims_obj = getattr(layer_info, "inferDims", None) or layer_info.dims
#print(f"[DBG] dims_obj.numDims={dims_obj.numDims}, dims={tuple(dims_obj.d[i] for i in range(dims_obj.numDims))}")
raw_buf = pyds.get_ptr(layer_info.buffer)
#print(f"[DBG] layer_info.buffer address = {raw_buf}")
if raw_buf == 0:
#print("[DBG] buffer was NULL, skipping decode")
continue
#print(dims_obj.d[0], dims_obj.d[1], dims_obj.d[2], dims_obj.d[3])
nd = dims_obj.numDims
shape = tuple(int(dims_obj.d[i]) for i in range(nd))
# total floats in this tensor
nelems = 1
for s in shape:
nelems *= s
# pull out the integer pointer
addr = pyds.get_ptr(layer_info.buffer)
if addr == 0:
# engine didn’t attach this tensor
continue
# build a ctypes array from that integer address
buffer_type = c_float * nelems
float_array = buffer_type.from_address(addr)
# wrap it in numpy
arr = np.ctypeslib.as_array(float_array).reshape(shape)
#print(f"[DBG] successfully mapped array, arr.shape={arr.shape}")
# now decode your bins → angle
probs = arr[0] # 1×90 → [90]
idxs = np.arange(len(probs), dtype=np.float32)
angle_degs = float(np.sum(probs * idxs) * _BINWIDTH - _ANGLE_OFFSET)
angle_rad = math.radians(angle_degs)
label = "pitch" if layer_idx == 0 else "yaw"
print(f"[SGIE] obj_id={obj_meta.object_id} → {label} = {angle_degs:.2f}° / {angle_rad:.3f} rad")
l_user = l_user.next
l_obj = l_obj.next
l_frame = l_frame.next
return Gst.PadProbeReturn.OK
SGIE starts but my debug prints show:
num_output_layers=2dims=(90,)layer_info.buffer address = 0for every face object
Here is some of the output
<pyds.NvDsInferLayerInfo object at 0xffff8ddc1130>
[DBG] dims_obj.numDims=1, dims=(90,)
[DBG] layer_info.buffer address = 0
[DBG] obj_id=6 class=2 has 0 user_meta entries
[DBG] obj_id=0 class=2 has 0 user_meta entries
[DBG] obj_id=1 class=2 has 0 user_meta entries
I am running this through the deepstream:7.1-triton-multiarch docker container and ran two sh files that are required for python dev.
I don’t know if I am probing erroneously or if I have to modify the configuration file to include something else.
