Cannot couple RandomBBoxCrop and Slice #855
Description
I have a detection pipeline similar to the COCO example pipeline:
import nvidia.dali.ops as ops
import nvidia.dali.types as types
import nvidia.dali.tfrecord as tfrecord
from nvidia.dali.pipeline import Pipeline
from nvidia.dali import plugin_manager
plugin_manager.load_library('./ConcatOp/build/libconcat.so')
class DetectionPipeline(Pipeline):
def __init__(self,
tfrecords,
tfrecords_idx,
batch_size,
num_workers,
device_id,
shard_id,
is_training=True):
super(DetectionPipeline, self).__init__(batch_size, num_workers, device_id)
self.is_training = is_training
features = {
'image/encoded': tfrecord.FixedLenFeature((), tfrecord.string, ''),
# Object boxes and classes.
'image/object/bbox/xmin': tfrecord.VarLenFeature(tfrecord.float32, 0.0),
'image/object/bbox/ymin': tfrecord.VarLenFeature(tfrecord.float32, 0.0),
'image/object/bbox/xmax': tfrecord.VarLenFeature(tfrecord.float32, 0.0),
'image/object/bbox/ymax': tfrecord.VarLenFeature(tfrecord.float32, 0.0),
'image/object/class/label': tfrecord.VarLenFeature(tfrecord.int64, -1),
}
self.input = ops.TFRecordReader(path=tfrecords,
index_path=tfrecords_idx,
features=features,
shard_id=shard_id,
num_shards=num_workers,
random_shuffle=False)
self.decode = ops.nvJPEGDecoder(device="mixed", output_type=types.RGB)
self.concat = ops.Concat(device="cpu")
self.cast_to_int32 = ops.Cast(device="cpu", dtype=types.INT32)
self.paste = ops.Paste(device='gpu', fill_value=0.)
self.bbox_paste = ops.BBoxPaste(device='cpu', ltrb=True)
self.paste_ratio = ops.Uniform(range=[1.0, 4.0])
self.paste_pos = ops.Uniform(range=[0.0, 1.0])
# Augumentation techniques
self.crop = ops.RandomBBoxCrop(
device="cpu",
aspect_ratio=[0.5, 2.0],
thresholds=[0, 0.1, 0.3, 0.5, 0.7, 0.9],
scaling=[0.3, 1.0],
ltrb=True,
allow_no_crop=True,
num_attempts=10)
self.slice = ops.Slice(device="gpu")
self.resize = ops.Resize(
device="gpu",
resize_x=512,
resize_y=512,
min_filter=types.DALIInterpType.INTERP_TRIANGULAR)
def define_graph(self):
# Read images and labels
inputs = self.input(name="Reader")
images = inputs["image/encoded"]
xmin = inputs["image/object/bbox/xmin"]
ymin = inputs["image/object/bbox/ymin"]
xmax = inputs["image/object/bbox/xmax"]
ymax = inputs["image/object/bbox/ymax"]
labels = inputs["image/object/class/label"]
images = self.decode(images)
boxes = self.concat(xmin, ymin, xmax, ymax)
labels = self.cast_to_int32(labels)
paste_ratio = self.paste_ratio()
paste_pos_x = self.paste_pos()
paste_pos_y = self.paste_pos()
images = self.paste(images, ratio=paste_ratio, paste_x=paste_pos_x, paste_y=paste_pos_y)
boxes = self.bbox_paste(boxes, ratio=paste_ratio, paste_x=paste_pos_x, paste_y=paste_pos_y)
crop_begin, crop_size, boxes, labels = self.crop(boxes, labels)
images = self.slice(images, crop_begin, crop_size)
images = self.resize(images)
return images.gpu(), boxes.gpu(), labels.gpu()
Trying to run it, I am getting an error:
DALI daliCreatePipeline(&pipe_handle_, serialized_pipeline.c_str(), serialized_pipeline.length(), batch_size, num_threads, device_id, prefetch_queue_depth_) failed: [/opt/dali/dali/pipeline/pipeline.cc:160] Assert on "it != edge_names_.end()" failed: Input 'contiguous_RandomBBoxCrop_id_9_output_0' to op 'Slice' is not known to the pipeline.
However, if I comment out the cropping/slicing code:
crop_begin, crop_size, boxes, labels = self.crop(boxes, labels)
images = self.slice(images, crop_begin, crop_size)
it works just fine.
I think this is due to the fact that RandomBBoxCrop is CPU op, and Slice is the GPU one.
If I move Slice to CPU, it understands the output from RandomBBoxCrop pretty well.
But how to couple the CPU RandomBBoxCrop and GPU Slice?
Thanks in advance.