from typing import TYPE_CHECKING, Optional, Tuple
import warnings
import logging
import torch
from torch.nn import functional as F
import torch.distributed as dist
from rastervision.pytorch_learner.learner import Learner
from rastervision.pytorch_learner.utils import (
compute_conf_mat_metrics, compute_conf_mat, aggregate_metrics)
from rastervision.pytorch_learner.dataset.visualizer import (
SemanticSegmentationVisualizer)
if TYPE_CHECKING:
from torch import nn
warnings.filterwarnings('ignore')
log = logging.getLogger(__name__)
[docs]class SemanticSegmentationLearner(Learner):
[docs] def get_visualizer_class(self):
return SemanticSegmentationVisualizer
[docs] def train_step(self, batch, batch_ind):
x, y = batch
out = self.post_forward(self.model(x))
return {'train_loss': self.loss(out, y)}
[docs] def validate_step(self, batch, batch_ind):
x, y = batch
out = self.post_forward(self.model(x))
val_loss = self.loss(out, y)
num_labels = len(self.cfg.data.class_names)
y = y.view(-1)
out = self.prob_to_pred(out).view(-1)
conf_mat = compute_conf_mat(out, y, num_labels)
return {'val_loss': val_loss, 'conf_mat': conf_mat}
[docs] def validate_end(self, outputs):
metrics = aggregate_metrics(outputs, exclude_keys={'conf_mat'})
conf_mat = sum([o['conf_mat'] for o in outputs])
if self.is_ddp_process:
metrics = self.reduce_distributed_metrics(metrics)
dist.reduce(conf_mat, dst=0, op=dist.ReduceOp.SUM)
if not self.is_ddp_master:
return metrics
ignored_idx = self.cfg.solver.ignore_class_index
if ignored_idx is not None and ignored_idx < 0:
ignored_idx += self.cfg.data.num_classes
class_names = self.cfg.data.class_names
conf_mat_metrics = compute_conf_mat_metrics(
conf_mat, class_names, ignore_idx=ignored_idx)
metrics.update(conf_mat_metrics)
return metrics
[docs] def post_forward(self, x):
if isinstance(x, dict):
return x['out']
return x
[docs] def predict(self,
x: torch.Tensor,
raw_out: bool = False,
out_shape: Optional[Tuple[int, int]] = None) -> torch.Tensor:
if out_shape is None:
out_shape = x.shape[-2:]
x = self.to_batch(x).float()
x = self.to_device(x, self.device)
with torch.inference_mode():
out = self.model(x)
out = self.post_forward(out)
out = self.postprocess_model_output(
out, raw_out=raw_out, out_shape=out_shape)
return out
[docs] def predict_onnx(
self,
x: torch.Tensor,
raw_out: bool = False,
out_shape: Optional[Tuple[int, int]] = None) -> torch.Tensor:
if out_shape is None:
out_shape = x.shape[-2:]
x = self.to_batch(x).float()
out = self.model(x)
out = self.post_forward(out)
out = self.postprocess_model_output(
out, raw_out=raw_out, out_shape=out_shape)
return out
[docs] def postprocess_model_output(self, out: torch.Tensor, raw_out: bool,
out_shape: Tuple[int, int]):
out = out.softmax(dim=1)
# ensure correct output shape
if out.shape[-2:] != out_shape:
out = F.interpolate(
out, size=out_shape, mode='bilinear', align_corners=False)
if not raw_out:
out = self.prob_to_pred(out)
out = self.to_device(out, 'cpu')
return out
[docs] def prob_to_pred(self, x):
return x.argmax(1)
[docs] def export_to_onnx(self,
path: str,
model: Optional['nn.Module'] = None,
sample_input: Optional[torch.Tensor] = None,
**kwargs) -> None:
args = dict(
input_names=['x'],
output_names=['out'],
dynamic_axes={
'x': {
0: 'batch_size',
2: 'height',
3: 'width',
},
'out': {
0: 'batch_size',
2: 'height',
3: 'width',
},
},
)
args.update(kwargs)
return super().export_to_onnx(path, model, sample_input, **args)