def __call__(self, engine: Union[SupervisedTrainer, SupervisedEvaluator],
batchdata: Dict[str, torch.Tensor]):
if batchdata is None:
raise ValueError("Must provide batch data for current iteration.")
if np.random.choice(
[True, False],
p=[self.deepgrow_probability, 1 - self.deepgrow_probability]):
# Run the inner loop only once
inputs, _ = engine.prepare_batch(batchdata)
inputs = inputs.to(engine.state.device)
engine.fire_event(IterationEvents.INNER_ITERATION_STARTED)
engine.network.eval()
with torch.no_grad():
if engine.amp:
with torch.cuda.amp.autocast():
predictions = engine.inferer(inputs, engine.network)
else:
predictions = engine.inferer(inputs, engine.network)
batchdata.update({CommonKeys.PRED: predictions})
# decollate/collate batchdata to execute click transforms
batchdata_list = decollate_batch(batchdata, detach=True)
for i in range(len(batchdata_list)):
batchdata_list[i][self.click_probability_key] = 1.0
batchdata_list[i] = self.transforms(batchdata_list[i])
batchdata = list_data_collate(batchdata_list)
engine.fire_event(IterationEvents.INNER_ITERATION_COMPLETED)
#
else:
# zero out input guidance channels
batchdata_list = decollate_batch(batchdata, detach=True)
for i in range(1, len(batchdata_list[0][CommonKeys.IMAGE])):
batchdata_list[0][CommonKeys.IMAGE][i] *= 0
batchdata = list_data_collate(batchdata_list)
# first item in batch only
engine.state.batch = batchdata
return engine._iteration(engine, batchdata)
def __call__(self, engine: Union[SupervisedTrainer, SupervisedEvaluator],
batchdata: Dict[str, torch.Tensor]):
if batchdata is None:
raise ValueError("Must provide batch data for current iteration.")
for j in range(self.max_interactions):
inputs, _ = engine.prepare_batch(batchdata)
inputs = inputs.to(engine.state.device)
engine.fire_event(IterationEvents.INNER_ITERATION_STARTED)
engine.network.eval()
with torch.no_grad():
if engine.amp:
with torch.cuda.amp.autocast():
predictions = engine.inferer(inputs, engine.network)
else:
predictions = engine.inferer(inputs, engine.network)
engine.fire_event(IterationEvents.INNER_ITERATION_COMPLETED)
batchdata.update({CommonKeys.PRED: predictions})
# decollate batch data to execute click transforms
batchdata_list = decollate_batch(batchdata, detach=True)
for i in range(len(batchdata_list)):
batchdata_list[i][self.key_probability] = ((1.0 - (
(1.0 / self.max_interactions) * j)) if self.train else 1.0)
batchdata_list[i] = self.transforms(batchdata_list[i])
# collate list into a batch for next round interaction
batchdata = list_data_collate(batchdata_list)
return engine._iteration(engine, batchdata)
def test_type_shape(self, input_data, expected_type, expected_shape):
result = list_data_collate(input_data)
self.assertIsInstance(result, expected_type)
if isinstance(result, dict):
data = result['image']
else:
data = result[0]
self.assertEqual(data.shape, expected_shape)
def test_type_shape(self, input_data, expected_type, expected_shape):
result = list_data_collate(input_data)
self.assertIsInstance(result, expected_type)
if isinstance(result, dict):
image = result["image"]
label = result["label"]
else:
image = result[0]
label = result[1]
self.assertEqual(image.shape, expected_shape)
self.assertEqual(label.shape, expected_shape)
self.assertTrue(isinstance(label, MetaTensor))
self.assertTrue(label.is_batch, True)
def __call__(self, engine: Union[SupervisedTrainer, SupervisedEvaluator],
batchdata: Dict[str, torch.Tensor]):
if batchdata is None:
raise ValueError("Must provide batch data for current iteration.")
pos_click_sum = 0
neg_click_sum = 0
if np.random.choice(
[True, False],
p=[self.deepgrow_probability, 1 - self.deepgrow_probability]):
pos_click_sum += 1 # increase pos_click_sum by 1-click for AddInitialSeedPointd pre_transform
for j in range(self.max_interactions):
# print("Inner iteration (click simulations running): ", str(j))
inputs, _ = engine.prepare_batch(batchdata)
inputs = inputs.to(engine.state.device)
engine.fire_event(IterationEvents.INNER_ITERATION_STARTED)
engine.network.eval()
with torch.no_grad():
if engine.amp:
with torch.cuda.amp.autocast():
predictions = engine.inferer(
inputs, engine.network)
else:
predictions = engine.inferer(inputs, engine.network)
batchdata.update({CommonKeys.PRED: predictions})
# decollate/collate batchdata to execute click transforms
batchdata_list = decollate_batch(batchdata, detach=True)
for i in range(len(batchdata_list)):
batchdata_list[i][self.click_probability_key] = (
(1.0 - ((1.0 / self.max_interactions) * j))
if self.train else 1.0)
batchdata_list[i] = self.transforms(batchdata_list[i])
batchdata = list_data_collate(batchdata_list)
# first item in batch only
pos_click_sum += (batchdata_list[0].get("is_pos", 0)) * 1
neg_click_sum += (batchdata_list[0].get("is_neg", 0)) * 1
engine.fire_event(IterationEvents.INNER_ITERATION_COMPLETED)
else:
# zero out input guidance channels
batchdata_list = decollate_batch(batchdata, detach=True)
for i in range(len(batchdata_list)):
batchdata_list[i][CommonKeys.IMAGE][-1] *= 0
batchdata_list[i][CommonKeys.IMAGE][-2] *= 0
batchdata = list_data_collate(batchdata_list)
# first item in batch only
engine.state.batch = batchdata
engine.state.batch.update(
{"pos_click_sum": torch.tensor(pos_click_sum)})
engine.state.batch.update(
{"neg_click_sum": torch.tensor(neg_click_sum)})
return engine._iteration(engine, batchdata)