def on_batch_end(self, runner: "IRunner"):
"""Batch end hook.
Args:
runner: current runner
"""
if self.step % self.summary_step == 0:
pred_mask = runner.output[self.output_key][0]
pred_mask = self.activation(pred_mask)
pred_mask = utils.detach(pred_mask)
pred_mask = self._prob2mask(pred_mask)
if self.input_mask_key is not None:
gt_mask = runner.input[self.input_mask_key][0]
gt_mask = utils.detach(gt_mask)
gt_mask = self._prob2mask(gt_mask)
else:
gt_mask = None
if self.input_image_key is not None:
image = runner.input[self.input_image_key][0].cpu()
image = tensor_to_ndimage(image)
else:
# white background
image = np.ones_like(pred_mask, dtype=np.uint8) * 255
pred_colors = self._get_colors(pred_mask)
image_over_predicted_mask = label2rgb(pred_mask,
image,
bg_label=0,
colors=pred_colors)
if gt_mask is not None:
gt_colors = self._get_colors(gt_mask)
image_over_gt_mask = label2rgb(gt_mask,
image,
bg_label=0,
colors=gt_colors)
else:
image_over_gt_mask = None
self._draw_masks(
self.loggers[runner.loader_key],
runner.global_sample_step,
image_over_predicted_mask,
image_over_gt_mask,
)
self.step += 1
def next_data(self):
# Code copy pasted for performance reasons
try:
batch = next(self.iter)['image']
logits, tables = self.model(batch)
mask = np.array([
utils.detach(logit[0].sigmoid() > 0.07).astype(bool)
for logit in logits
])
table = tables.detach().numpy()
return mask, table
except StopIteration:
return None, None
def on_batch_end(self, state: RunnerState):
if state.loader_name.startswith("valid"):
detections = decode_centernet_predictions(state.output["hm"],
state.output["wh"],
state.output["reg"],
K=self.max_objs)
detections = detach(detections).reshape(
(detections.shape[0], -1, detections.shape[2]))
detections[:, :, :4] *= self.down_ratio
bboxes = detections[:, :, :4].astype(int)
scores = detections[:, :, 4]
labels = detections[:, :, 5].astype(int)
result = dict(
bboxes=bboxes,
labels=labels,
scores=scores,
)
state.output.update(result)
def predict_table_mask(self, batch: torch.Tensor) -> np.ndarray:
return np.array([
utils.detach(logit[0].sigmoid() > 0.07).astype(bool)
for logit in self.model(batch)[0]
])
),
)
)
)
print(type(predictions))
print(predictions.shape)
threshold = 0.5
max_count = 5
for i, (features, logits) in enumerate(zip(test_dataset, predictions)):
image = utils.tensor_to_ndimage(features["image"])
mask_ = torch.from_numpy(logits[0]).sigmoid()
mask = utils.detach(mask_ > threshold).astype("float")
show_examples(name="", image=image, mask=mask)
if i >= max_count:
break
batch = next(iter(loaders["valid"]))
# saves to `logdir` and returns a `ScriptModule` class
runner.trace(model=model, batch=batch, logdir=logdir, fp16=is_fp16_used)
!ls {logdir}/trace/
from catalyst.utils import tracing
if is_fp16_used:
threshold = 0.5
max_count = 5
for i, (features, logits) in enumerate(zip(train_dataset, predictions)):
image = utils.tensor_to_ndimage(features["image"], denormalize=False)
# filename_mask = os.path.splitext(features["filename"])[0]
filename_mask = os.path.splitext(features["filename_mask"])[0]
gt = imread(mask_path / f"{filename_mask}.png")
gt_res = gt.copy()
gt_res.resize((224, 224))
gt_im = Image.fromarray(gt_res * 255)
mask_ = torch.from_numpy(logits[0]).sigmoid()
mask = utils.detach(mask_ > threshold).astype("uint8")
# # Replace mask with real image
# # filename_mask = os.path.splitext(features["filename_img"])[0]
# # mask = imread(test_image_path / f"{filename_mask}.jpg")
show_examples(
name="",
image=image,
mask=mask,
gt=gt_res,
save=True,
fig_path=f"predictions/{model.name}/",
fig_name=f"prediction_{i}.png",
overlay=True,
)