Hugging Face's logo

Spaces:
akhaliq
/
deeplab2
Runtime error

App Files Community
deeplab2 / tensorflow_ops /python /kernel_tests /merge_semantic_and_instance_maps_op_test.py
akhaliq3
spaces demo
506da10
raw history blame contribute delete
No virus
10.3 kB
# coding=utf-8
# Copyright 2021 The Deeplab2 Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
r"""Tests for merge_semantic_and_instance_maps_op."""
import numpy as np
import tensorflow as tf
from deeplab2.tensorflow_ops.python.ops import merge_semantic_and_instance_maps_op
class MergeSemanticAndInstanceMapsOpTest(tf.test.TestCase):
def testMergeSemanticAndInstanceMaps(self):
"""Test the op with 2 images."""
batch = 2
height = 4
width = 4
# Create the instance labels.
instance_maps = np.zeros((batch, height, width), dtype=np.int32)
instance_maps[0, :, :] = np.array([[0, 2, 1, 0], [0, 1, 1, 0], [2, 0, 1, 2],
[0, 0, 1, 1]])
instance_maps[1, :, :] = np.array([[1, 2, 3, 1], [0, 2, 1, 3], [0, 2, 2, 0],
[3, 3, 2, 0]])
# Create the semantic labels.
# The instances with the instance label equal to 0 and 2 have the same
# semantic label. The other instances all have different semantic labels.
semantic_maps = np.zeros((batch, height, width), dtype=np.int32)
# Instance 0 has 4 pixels predicted as 0 and 3 pixels predicted as 3.
# Instance 1 has 6 pixels predicted as 1.
# Instance 2 has 2 pixels predicted as 0 and 1 pixel predicted as 3.
semantic_maps[0, :, :] = np.array([[0, 0, 1, 0], [0, 1, 1, 0], [3, 3, 1, 0],
[3, 3, 1, 1]])
# Instance 0 has 3 pixels predicted as 0 and 1 pixel predicted as 3.
# Instance 1 has 3 pixels predicted as 1.
# Instance 2 has 3 pixels predicted as 0 and 2 pixels predicted as 2.
# Instance 3 has 1 pixel predicted as 0 and 3 pixels predicted as 2.
semantic_maps[1, :, :] = np.array([[1, 0, 2, 1], [0, 0, 1, 2], [0, 2, 2, 3],
[0, 2, 0, 0]])
# Create the ID list for things.
thing_ids = [0, 2]
# Groundtruth semantic segmentation maps after majority voting.
gt_semantic_maps = np.zeros((batch, height, width), dtype=np.int32)
gt_semantic_maps[0, :, :] = np.array([[0, 0, 1, 0], [0, 1, 1, 0],
[3, 3, 1, 0], [3, 3, 1, 1]])
# Instance 2 takes semantic label 0 after majority voting.
# Instance 3 takes semantic label 2 after majority voting.
gt_semantic_maps[1, :, :] = np.array([[1, 0, 2, 1], [0, 0, 1, 2],
[0, 0, 0, 3], [2, 2, 0, 0]])
# Groundtruth instance segmentation maps.
gt_instance_maps = np.zeros((batch, 2, height, width), dtype=np.int32)
# There are two cases for gt_instance_maps in batch 1.
# Case 1:
# Instance 0 is re-assigned instance label 1.
# Instance 2 is re-assigned instance label 2.
gt_instance_maps[0, 0, :, :] = np.array([[1, 2, 0, 1], [1, 0, 0, 1],
[0, 0, 0, 2], [0, 0, 0, 0]])
# Case 2:
# Instance 0 is re-assigned instance label 2.
# Instance 2 is re-assigned instance label 1.
gt_instance_maps[0, 1, :, :] = np.array([[2, 1, 0, 2], [2, 0, 0, 2],
[0, 0, 0, 1], [0, 0, 0, 0]])
# There are two cases for gt_instance_maps in batch 2.
# Case 1:
# Instance 0 is re-assigned instance label 1.
# Instance 2 is re-assigned instance label 2.
# Instance 3 is re-assigned instance label 1.
gt_instance_maps[1, 0, :, :] = np.array([[0, 2, 1, 0], [1, 2, 0, 1],
[1, 2, 2, 0], [1, 1, 2, 1]])
# Case 2:
# Instance 0 is re-assigned instance label 2.
# Instance 2 is re-assigned instance label 1.
# Instance 3 is re-assigned instance label 1.
gt_instance_maps[1, 1, :, :] = np.array([[0, 1, 1, 0], [2, 1, 0, 1],
[2, 1, 1, 0], [1, 1, 1, 2]])
# Groundtruth parsing maps.
label_divisor = 256
# Run the op.
parsing_maps = (
merge_semantic_and_instance_maps_op.merge_semantic_and_instance_maps(
semantic_maps,
instance_maps,
thing_ids,
label_divisor=label_divisor))
pass_test = False
for i in range(2):
for j in range(2):
current_gt_instance_maps = np.stack(
[gt_instance_maps[0, i, :, :], gt_instance_maps[1, j, :, :]],
axis=0)
gt_parsing_maps = (
gt_semantic_maps * label_divisor + current_gt_instance_maps)
if np.array_equal(parsing_maps, gt_parsing_maps):
pass_test = True
self.assertTrue(pass_test)
def testMergeSemanticAndInstanceMapsWithStuffAreaLimit(self):
batch = 1
height = 4
width = 4
# Create the instance labels.
instance_maps = np.zeros((batch, height, width), dtype=np.int32)
instance_maps[0, :, :] = np.array([[0, 0, 0, 0],
[0, 0, 1, 1],
[0, 0, 0, 0],
[0, 0, 0, 0]])
# Create the semantic labels.
semantic_maps = np.zeros((batch, height, width), dtype=np.int32)
semantic_maps[0, :, :] = np.array([[0, 0, 0, 0],
[0, 0, 1, 1],
[0, 0, 2, 2],
[0, 0, 2, 2]])
thing_ids = [0, 2]
stuff_area_limit = 3
void_label = 3
# Groundtruth semantic segmentation maps after majority voting.
# Instance 0 takes semantic label 0.
# Instance 1 is re-assigned with void label.
gt_semantic_maps = np.zeros((batch, height, width), dtype=np.int32)
gt_semantic_maps[0, :, :] = np.array([[0, 0, 0, 0],
[0, 0, void_label, void_label],
[0, 0, 0, 0],
[0, 0, 0, 0]])
# Groundtruth instance segmentation maps.
gt_instance_maps = np.zeros((batch, height, width), dtype=np.int32)
gt_instance_maps[0, :, :] = np.array([[1, 1, 1, 1],
[1, 1, 0, 0],
[1, 1, 1, 1],
[1, 1, 1, 1]])
label_divisor = 256
gt_parsing_maps = gt_semantic_maps * label_divisor + gt_instance_maps
# Run the op.
parsing_maps = (
merge_semantic_and_instance_maps_op.merge_semantic_and_instance_maps(
semantic_maps,
instance_maps,
thing_ids,
label_divisor=label_divisor,
stuff_area_limit=stuff_area_limit,
void_label=void_label))
self.assertTrue(np.array_equal(parsing_maps, gt_parsing_maps))
class MergeSemanticAndInstanceMapsOpGpuTest(MergeSemanticAndInstanceMapsOpTest):
def session(self, use_gpu=True):
return super(MergeSemanticAndInstanceMapsOpGpuTest,
self).session(use_gpu=use_gpu)
def testMergeSemanticAndInstanceMapsWithRandomInputs(self):
batch = 1
height = 1441
width = 1441
rng = np.random.RandomState(0)
instance_maps = rng.randint(0, 255, (batch, height, width), dtype=np.int32)
semantic_maps = rng.randint(0, 3, (batch, height, width), dtype=np.int32)
thing_ids = [0, 2]
stuff_area_limit = 400
void_label = 3
label_divisor = 256
with self.session(use_gpu=False):
parsing_maps_cpu = (
merge_semantic_and_instance_maps_op.merge_semantic_and_instance_maps(
semantic_maps,
instance_maps,
thing_ids,
label_divisor=label_divisor,
stuff_area_limit=stuff_area_limit,
void_label=void_label))
parsing_maps_cpu = parsing_maps_cpu.numpy()
with self.session():
parsing_maps_gpu = (
merge_semantic_and_instance_maps_op.merge_semantic_and_instance_maps(
semantic_maps,
instance_maps,
thing_ids,
label_divisor=label_divisor,
stuff_area_limit=stuff_area_limit,
void_label=void_label))
parsing_maps_gpu = parsing_maps_gpu.numpy()
# Checks semantic maps are the same.
semantic_maps_cpu = parsing_maps_cpu // label_divisor
semantic_maps_gpu = parsing_maps_gpu // label_divisor
np.testing.assert_array_equal(semantic_maps_cpu, semantic_maps_gpu)
# Checks instance maps are the same, despite of label order.
instance_maps_cpu = parsing_maps_cpu % label_divisor
instance_maps_gpu = parsing_maps_gpu % label_divisor
thing_labels_cpu = np.unique(semantic_maps_cpu[instance_maps_cpu > 0])
for semantic_label in thing_labels_cpu:
semantic_mask = semantic_maps_cpu == semantic_label
instance_labels_cpu = np.unique(instance_maps_cpu[semantic_mask])
instance_labels_gpu = np.unique(instance_maps_gpu[semantic_mask])
self.assertEqual(len(instance_labels_cpu), len(instance_labels_gpu))
# For each instance (cpu reference) of this semantic label, we check:
# 1. Within this instance mask, GPU produces one and only one instance
# label.
# 2. GPU results with the current semantic and instance label matches
# CPU instance mask.
for instance_label in instance_labels_cpu:
instance_mask_cpu = np.logical_and(
instance_maps_cpu == instance_label, semantic_mask)
instance_labels_gpu = set(instance_maps_gpu[instance_mask_cpu])
self.assertLen(instance_labels_gpu, 1)
instance_label_gpu = instance_labels_gpu.pop()
# Here GPU must use the same semantic mask (given we have checked
# semantic maps are the same).
instance_mask_gpu = np.logical_and(
instance_maps_gpu == instance_label_gpu, semantic_mask)
np.testing.assert_array_equal(instance_mask_cpu, instance_mask_gpu)
if __name__ == '__main__':
tf.test.main()