def single_gpu_test(model, dataset, show=False):
# create a loader for this runner
tf_dataset, num_examples = build_dataloader(dataset,
1,
1,
num_gpus=1,
dist=False)
results = []
start = time.time()
for i, data_batch in enumerate(tf_dataset):
if i >= num_examples:
break
_, img_meta = data_batch
print(dataset.img_ids[i])
outputs = model(data_batch, training=False)
bboxes = outputs['bboxes']
# # map boxes back to original scale
bboxes = transforms.bbox_mapping_back(bboxes, img_meta)
# # print('>>>>', bboxes)
labels = outputs['labels']
scores = outputs['scores']
result = transforms.bbox2result(bboxes, labels, scores, num_classes=81)
#for b, l, s in zip(bboxes, labels, scores):
# print(b, l, s)
#print(result)
results.append(result)
print("Forward pass through test set took {}s".format(time.time() - start))
evaluate(dataset, results)
return results
def _unmold_single_detection(self, detections, img_meta):
zero_ix = tf.where(tf.not_equal(detections[:, 4], 0))
detections = tf.gather_nd(detections, zero_ix)
# Extract boxes, class_ids, scores, and class-specific masks
boxes = detections[:, :4]
class_ids = tf.cast(detections[:, 4], tf.int32)
scores = detections[:, 5]
boxes = transforms.bbox_mapping_back(boxes, img_meta)
return {'rois': boxes.numpy(),
'class_ids': class_ids.numpy(),
'scores': scores.numpy()}
def after_train_epoch(self, runner):
if not self.every_n_epochs(runner, self.interval):
return
# create a loader for this runner
tf_dataset, num_examples = build_dataloader(self.dataset,
1,
1,
num_gpus=runner.local_size,
dist=True)
# num_examples=8
results = [None for _ in range(num_examples * runner.local_size)
] # REVISIT - may require a lot of memory
#if runner.model.mask:
if self.dataset.mask:
masks = [None for _ in range(num_examples * runner.local_size)]
if runner.rank == 0:
prog_bar = ProgressBar(num_examples)
for i, data_batch in enumerate(tf_dataset):
if i >= num_examples:
break
_, img_meta = data_batch
outputs = runner.model(data_batch, training=False)
assert isinstance(outputs, dict)
bboxes = outputs['bboxes']
# map boxes back to original scale
bboxes = transforms.bbox_mapping_back(bboxes, img_meta)
labels = outputs['labels']
scores = outputs['scores']
result = transforms.bbox2result(bboxes,
labels,
scores,
num_classes=self.dataset.CLASSES +
1) # add background class
#if runner.model.mask:
if self.dataset.mask:
mask = mask2result(outputs['masks'], labels, img_meta[0])
results[i * runner.local_size + runner.local_rank] = (result,
mask)
else:
results[i * runner.local_size + runner.local_rank] = result
if runner.rank == 0:
prog_bar.update()
# write to a file
tmp_file = osp.join(runner.work_dir, 'temp_{}.pkl'.format(runner.rank))
if runner.rank != 0:
dump(results, tmp_file)
# open(tmp_file+'.done', 'w').close()
# MPI barrier through horovod
_ = get_barrier()
self._accumulate_results(runner, results, num_examples)