spacr 0.5.0__py3-none-any.whl → 0.9.1__py3-none-any.whl

spacr/resources/MEDIAR/core/MEDIAR/EnsemblePredictor.py

@@ -1,105 +0,0 @@
-import torch
-import os, sys, copy
-import numpy as np
-
-sys.path.insert(0, os.path.abspath(os.path.join(os.getcwd(), "../../")))
-
-from core.MEDIAR.Predictor import Predictor
-
-__all__ = ["EnsemblePredictor"]
-
-
-class EnsemblePredictor(Predictor):
-    def __init__(
-        self,
-        model,
-        model_aux,
-        device,
-        input_path,
-        output_path,
-        make_submission=False,
-        exp_name=None,
-        algo_params=None,
-    ):
-        super(EnsemblePredictor, self).__init__(
-            model,
-            device,
-            input_path,
-            output_path,
-            make_submission,
-            exp_name,
-            algo_params,
-        )
-        self.model_aux = model_aux
-
-    @torch.no_grad()
-    def _inference(self, img_data):
-
-        self.model_aux.to(self.device)
-        self.model_aux.eval()
-
-        img_data = img_data.to(self.device)
-        img_base = img_data
-
-        outputs_base = self._window_inference(img_base)
-        outputs_base = outputs_base.cpu().squeeze()
-
-        outputs_aux = self._window_inference(img_base, aux=True)
-        outputs_aux = outputs_aux.cpu().squeeze()
-        img_base.cpu()
-
-        if not self.use_tta:
-            pred_mask = (outputs_base + outputs_aux) / 2
-            return pred_mask
-
-        else:
-            # HorizontalFlip TTA
-            img_hflip = self.hflip_tta.apply_aug_image(img_data, apply=True)
-
-            outputs_hflip = self._window_inference(img_hflip)
-            outputs_hflip_aux = self._window_inference(img_hflip, aux=True)
-
-            outputs_hflip = self.hflip_tta.apply_deaug_mask(outputs_hflip, apply=True)
-            outputs_hflip_aux = self.hflip_tta.apply_deaug_mask(
-                outputs_hflip_aux, apply=True
-            )
-
-            outputs_hflip = outputs_hflip.cpu().squeeze()
-            outputs_hflip_aux = outputs_hflip_aux.cpu().squeeze()
-            img_hflip = img_hflip.cpu()
-
-            # VertricalFlip TTA
-            img_vflip = self.vflip_tta.apply_aug_image(img_data, apply=True)
-
-            outputs_vflip = self._window_inference(img_vflip)
-            outputs_vflip_aux = self._window_inference(img_vflip, aux=True)
-
-            outputs_vflip = self.vflip_tta.apply_deaug_mask(outputs_vflip, apply=True)
-            outputs_vflip_aux = self.vflip_tta.apply_deaug_mask(
-                outputs_vflip_aux, apply=True
-            )
-
-            outputs_vflip = outputs_vflip.cpu().squeeze()
-            outputs_vflip_aux = outputs_vflip_aux.cpu().squeeze()
-            img_vflip = img_vflip.cpu()
-
-            # Merge Results
-            pred_mask = torch.zeros_like(outputs_base)
-            pred_mask[0] = (outputs_base[0] + outputs_hflip[0] - outputs_vflip[0]) / 3
-            pred_mask[1] = (outputs_base[1] - outputs_hflip[1] + outputs_vflip[1]) / 3
-            pred_mask[2] = (outputs_base[2] + outputs_hflip[2] + outputs_vflip[2]) / 3
-
-            pred_mask_aux = torch.zeros_like(outputs_aux)
-            pred_mask_aux[0] = (
-                outputs_aux[0] + outputs_hflip_aux[0] - outputs_vflip_aux[0]
-            ) / 3
-            pred_mask_aux[1] = (
-                outputs_aux[1] - outputs_hflip_aux[1] + outputs_vflip_aux[1]
-            ) / 3
-            pred_mask_aux[2] = (
-                outputs_aux[2] + outputs_hflip_aux[2] + outputs_vflip_aux[2]
-            ) / 3
-
-            pred_mask = (pred_mask + pred_mask_aux) / 2
-
-        return pred_mask

spacr/resources/MEDIAR/core/MEDIAR/Predictor.py

@@ -1,234 +0,0 @@
-import torch
-import numpy as np
-import os, sys
-from monai.inferers import sliding_window_inference
-
-sys.path.insert(0, os.path.abspath(os.path.join(os.getcwd(), "../../")))
-
-from core.BasePredictor import BasePredictor
-from core.MEDIAR.utils import compute_masks
-
-__all__ = ["Predictor"]
-
-
-class Predictor(BasePredictor):
-    def __init__(
-        self,
-        model,
-        device,
-        input_path,
-        output_path,
-        make_submission=False,
-        exp_name=None,
-        algo_params=None,
-    ):
-        super(Predictor, self).__init__(
-            model,
-            device,
-            input_path,
-            output_path,
-            make_submission,
-            exp_name,
-            algo_params,
-        )
-        self.hflip_tta = HorizontalFlip()
-        self.vflip_tta = VerticalFlip()
-
-    @torch.no_grad()
-    def _inference(self, img_data):
-        """Conduct model prediction"""
-
-        img_data = img_data.to(self.device)
-        img_base = img_data
-        outputs_base = self._window_inference(img_base)
-        outputs_base = outputs_base.cpu().squeeze()
-        img_base.cpu()
-
-        if not self.use_tta:
-            pred_mask = outputs_base
-            return pred_mask
-
-        else:
-            # HorizontalFlip TTA
-            img_hflip = self.hflip_tta.apply_aug_image(img_data, apply=True)
-            outputs_hflip = self._window_inference(img_hflip)
-            outputs_hflip = self.hflip_tta.apply_deaug_mask(outputs_hflip, apply=True)
-            outputs_hflip = outputs_hflip.cpu().squeeze()
-            img_hflip = img_hflip.cpu()
-
-            # VertricalFlip TTA
-            img_vflip = self.vflip_tta.apply_aug_image(img_data, apply=True)
-            outputs_vflip = self._window_inference(img_vflip)
-            outputs_vflip = self.vflip_tta.apply_deaug_mask(outputs_vflip, apply=True)
-            outputs_vflip = outputs_vflip.cpu().squeeze()
-            img_vflip = img_vflip.cpu()
-
-            # Merge Results
-            pred_mask = torch.zeros_like(outputs_base)
-            pred_mask[0] = (outputs_base[0] + outputs_hflip[0] - outputs_vflip[0]) / 3
-            pred_mask[1] = (outputs_base[1] - outputs_hflip[1] + outputs_vflip[1]) / 3
-            pred_mask[2] = (outputs_base[2] + outputs_hflip[2] + outputs_vflip[2]) / 3
-
-        return pred_mask
-
-    def _window_inference(self, img_data, aux=False):
-        """Inference on RoI-sized window"""
-        outputs = sliding_window_inference(
-            img_data,
-            roi_size=512,
-            sw_batch_size=4,
-            predictor=self.model if not aux else self.model_aux,
-            padding_mode="constant",
-            mode="gaussian",
-            overlap=0.6,
-        )
-
-        return outputs
-
-    def _post_process(self, pred_mask):
-        """Generate cell instance masks."""
-        dP, cellprob = pred_mask[:2], self._sigmoid(pred_mask[-1])
-        H, W = pred_mask.shape[-2], pred_mask.shape[-1]
-
-        if np.prod(H * W) < (5000 * 5000):
-            pred_mask = compute_masks(
-                dP,
-                cellprob,
-                use_gpu=True,
-                flow_threshold=0.4,
-                device=self.device,
-                cellprob_threshold=0.5,
-            )[0]
-
-        else:
-            print("\n[Whole Slide] Grid Prediction starting...")
-            roi_size = 2000
-
-            # Get patch grid by roi_size
-            if H % roi_size != 0:
-                n_H = H // roi_size + 1
-                new_H = roi_size * n_H
-            else:
-                n_H = H // roi_size
-                new_H = H
-
-            if W % roi_size != 0:
-                n_W = W // roi_size + 1
-                new_W = roi_size * n_W
-            else:
-                n_W = W // roi_size
-                new_W = W
-
-            # Allocate values on the grid
-            pred_pad = np.zeros((new_H, new_W), dtype=np.uint32)
-            dP_pad = np.zeros((2, new_H, new_W), dtype=np.float32)
-            cellprob_pad = np.zeros((new_H, new_W), dtype=np.float32)
-
-            dP_pad[:, :H, :W], cellprob_pad[:H, :W] = dP, cellprob
-
-            for i in range(n_H):
-                for j in range(n_W):
-                    print("Pred on Grid (%d, %d) processing..." % (i, j))
-                    dP_roi = dP_pad[
-                        :,
-                        roi_size * i : roi_size * (i + 1),
-                        roi_size * j : roi_size * (j + 1),
-                    ]
-                    cellprob_roi = cellprob_pad[
-                        roi_size * i : roi_size * (i + 1),
-                        roi_size * j : roi_size * (j + 1),
-                    ]
-
-                    pred_mask = compute_masks(
-                        dP_roi,
-                        cellprob_roi,
-                        use_gpu=True,
-                        flow_threshold=0.4,
-                        device=self.device,
-                        cellprob_threshold=0.5,
-                    )[0]
-
-                    pred_pad[
-                        roi_size * i : roi_size * (i + 1),
-                        roi_size * j : roi_size * (j + 1),
-                    ] = pred_mask
-
-            pred_mask = pred_pad[:H, :W]
-
-        return pred_mask
-
-    def _sigmoid(self, z):
-        return 1 / (1 + np.exp(-z))
-
-
-"""
-Adapted from the following references:
-[1] https://github.com/qubvel/ttach/blob/master/ttach/transforms.py
-
-"""
-
-
-def hflip(x):
-    """flip batch of images horizontally"""
-    return x.flip(3)
-
-
-def vflip(x):
-    """flip batch of images vertically"""
-    return x.flip(2)
-
-
-class DualTransform:
-    identity_param = None
-
-    def __init__(
-        self, name: str, params,
-    ):
-        self.params = params
-        self.pname = name
-
-    def apply_aug_image(self, image, *args, **params):
-        raise NotImplementedError
-
-    def apply_deaug_mask(self, mask, *args, **params):
-        raise NotImplementedError
-
-
-class HorizontalFlip(DualTransform):
-    """Flip images horizontally (left -> right)"""
-
-    identity_param = False
-
-    def __init__(self):
-        super().__init__("apply", [False, True])
-
-    def apply_aug_image(self, image, apply=False, **kwargs):
-        if apply:
-            image = hflip(image)
-        return image
-
-    def apply_deaug_mask(self, mask, apply=False, **kwargs):
-        if apply:
-            mask = hflip(mask)
-        return mask
-
-
-class VerticalFlip(DualTransform):
-    """Flip images vertically (up -> down)"""
-
-    identity_param = False
-
-    def __init__(self):
-        super().__init__("apply", [False, True])
-
-    def apply_aug_image(self, image, apply=False, **kwargs):
-        if apply:
-            image = vflip(image)
-
-        return image
-
-    def apply_deaug_mask(self, mask, apply=False, **kwargs):
-        if apply:
-            mask = vflip(mask)
-
-        return mask

spacr/resources/MEDIAR/core/MEDIAR/Trainer.py

@@ -1,172 +0,0 @@
-import torch
-import torch.nn as nn
-import numpy as np
-import os, sys
-from tqdm import tqdm
-from monai.inferers import sliding_window_inference
-
-sys.path.insert(0, os.path.abspath(os.path.join(os.getcwd(), "../../")))
-
-from core.BaseTrainer import BaseTrainer
-from core.MEDIAR.utils import *
-
-__all__ = ["Trainer"]
-
-
-class Trainer(BaseTrainer):
-    def __init__(
-        self,
-        model,
-        dataloaders,
-        optimizer,
-        scheduler=None,
-        criterion=None,
-        num_epochs=100,
-        device="cuda:0",
-        no_valid=False,
-        valid_frequency=1,
-        amp=False,
-        algo_params=None,
-    ):
-        super(Trainer, self).__init__(
-            model,
-            dataloaders,
-            optimizer,
-            scheduler,
-            criterion,
-            num_epochs,
-            device,
-            no_valid,
-            valid_frequency,
-            amp,
-            algo_params,
-        )
-
-        self.mse_loss = nn.MSELoss(reduction="mean")
-        self.bce_loss = nn.BCEWithLogitsLoss(reduction="mean")
-
-    def mediar_criterion(self, outputs, labels_onehot_flows):
-        """loss function between true labels and prediction outputs"""
-
-        # Cell Recognition Loss
-        cellprob_loss = self.bce_loss(
-            outputs[:, -1],
-            torch.from_numpy(labels_onehot_flows[:, 1] > 0.5).to(self.device).float(),
-        )
-
-        # Cell Distinction Loss
-        gradient_flows = torch.from_numpy(labels_onehot_flows[:, 2:]).to(self.device)
-        gradflow_loss = 0.5 * self.mse_loss(outputs[:, :2], 5.0 * gradient_flows)
-
-        loss = cellprob_loss + gradflow_loss
-
-        return loss
-
-    def _epoch_phase(self, phase):
-        phase_results = {}
-
-        # Set model mode
-        self.model.train() if phase == "train" else self.model.eval()
-
-        # Epoch process
-        for batch_data in tqdm(self.dataloaders[phase]):
-            images, labels = batch_data["img"], batch_data["label"]
-
-            if self.with_public:
-                # Load batches sequentially from the unlabeled dataloader
-                try:
-                    batch_data = next(self.public_iterator)
-                    images_pub, labels_pub = batch_data["img"], batch_data["label"]
-
-                except:
-                    # Assign memory loader if the cycle ends
-                    self.public_iterator = iter(self.public_loader)
-                    batch_data = next(self.public_iterator)
-                    images_pub, labels_pub = batch_data["img"], batch_data["label"]
-
-                # Concat memory data to the batch
-                images = torch.cat([images, images_pub], dim=0)
-                labels = torch.cat([labels, labels_pub], dim=0)
-
-            images = images.to(self.device)
-            labels = labels.to(self.device)
-
-            self.optimizer.zero_grad()
-
-            # Forward pass
-            with torch.cuda.amp.autocast(enabled=self.amp):
-                with torch.set_grad_enabled(phase == "train"):
-                    # Output shape is B x [grad y, grad x, cellprob] x H x W
-                    outputs = self._inference(images, phase)
-
-                    # Map label masks to graidnet and onehot
-                    labels_onehot_flows = labels_to_flows(
-                        labels, use_gpu=True, device=self.device
-                    )
-                    # Calculate loss
-                    loss = self.mediar_criterion(outputs, labels_onehot_flows)
-                    self.loss_metric.append(loss)
-
-                    # Calculate valid statistics
-                    if phase != "train":
-                        outputs, labels = self._post_process(outputs, labels)
-                        f1_score = self._get_f1_metric(outputs, labels)
-                        self.f1_metric.append(f1_score)
-
-                # Backward pass
-                if phase == "train":
-                    # For the mixed precision training
-                    if self.amp:
-                        self.scaler.scale(loss).backward()
-                        self.scaler.unscale_(self.optimizer)
-                        self.scaler.step(self.optimizer)
-                        self.scaler.update()
-
-                    else:
-                        loss.backward()
-                        self.optimizer.step()
-
-        # Update metrics
-        phase_results = self._update_results(
-            phase_results, self.loss_metric, "dice_loss", phase
-        )
-        if phase != "train":
-            phase_results = self._update_results(
-                phase_results, self.f1_metric, "f1_score", phase
-            )
-
-        return phase_results
-
-    def _inference(self, images, phase="train"):
-        """inference methods for different phase"""
-
-        if phase != "train":
-            outputs = sliding_window_inference(
-                images,
-                roi_size=512,
-                sw_batch_size=4,
-                predictor=self.model,
-                padding_mode="constant",
-                mode="gaussian",
-                overlap=0.5,
-            )
-        else:
-            outputs = self.model(images)
-
-        return outputs
-
-    def _post_process(self, outputs, labels=None):
-        """Predict cell instances using the gradient tracking"""
-        outputs = outputs.squeeze(0).cpu().numpy()
-        gradflows, cellprob = outputs[:2], self._sigmoid(outputs[-1])
-        outputs = compute_masks(gradflows, cellprob, use_gpu=True, device=self.device)
-        outputs = outputs[0]  # (1, C, H, W) -> (C, H, W)
-
-        if labels is not None:
-            labels = labels.squeeze(0).squeeze(0).cpu().numpy()
-
-        return outputs, labels
-
-    def _sigmoid(self, z):
-        """Sigmoid function for numpy arrays"""
-        return 1 / (1 + np.exp(-z))

spacr/resources/MEDIAR/core/MEDIAR/__init__.py

@@ -1,3 +0,0 @@
-from .Trainer import *
-from .Predictor import *
-from .EnsemblePredictor import *