repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
AgML
AgML-main/experiments/benchmarking/classification_distributed.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import argparse import agml import albumentations as A import torch import torch.nn as nn import pytorch_lightning as pl from torchvision.models import efficientnet_b4 class EfficientNetB4Transfer(nn.Module): """Represents a transfer learning EfficientNetB4 model. This is the base benchmarking model for image classification, using the EfficientNetB4 model with two added linear fully-connected layers. """ def __init__(self, num_classes, pretrained = True): super(EfficientNetB4Transfer, self).__init__() self.base = efficientnet_b4(pretrained = pretrained) self.l1 = nn.Linear(1000, 256) self.dropout = nn.Dropout(0.1) self.relu = nn.ReLU() self.l2 = nn.Linear(256, num_classes) def forward(self, x, **kwargs): # noqa x = self.base(x) x = x.view(x.size(0), -1) x = self.dropout(self.relu(self.l1(x))) x = self.l2(x) return x def get_num_gpus(args): """Resolves the number of GPUs to use.""" arg_gpus = getattr(args, 'gpus', None) if isinstance(arg_gpus, int): return arg_gpus return torch.cuda.device_count() class ClassificationBenchmark(pl.LightningModule): """Represents an image classification benchmark model.""" def __init__(self, dataset, pretrained = False): # Initialize the module. super(ClassificationBenchmark, self).__init__() # Construct the network. self._source = agml.data.source(dataset) self._pretrained = pretrained self.net = EfficientNetB4Transfer( self._source.num_classes, self._pretrained ) # Construct the loss for training. self.loss = nn.CrossEntropyLoss() def forward(self, x): return self.net.forward(x) def training_step(self, batch, *args, **kwargs): # noqa x, y = batch y_pred = self(x) loss = self.loss(y_pred, y) acc = accuracy(y_pred, torch.argmax(y, 1)).item() self.log('accuracy', acc, prog_bar = True, sync_dist = True) return { 'loss': loss, 'accuracy': acc } def validation_step(self, batch, *args, **kwargs): # noqa x, y = batch y_pred = self(x) val_loss = self.loss(y_pred, y) val_acc = accuracy(y_pred, torch.argmax(y, 1)) self.log('val_loss', val_loss.item(), prog_bar = True, sync_dist = True) self.log('val_accuracy', val_acc.item(), prog_bar = True, sync_dist = True) return { 'val_loss': val_loss, 'val_accuracy': val_acc } def configure_optimizers(self): return torch.optim.Adam(self.parameters()) def get_progress_bar_dict(self): tqdm_dict = super(ClassificationBenchmark, self) \ .get_progress_bar_dict() tqdm_dict.pop('v_num', None) return tqdm_dict def accuracy(output, target): """Computes the accuracy between `output` and `target`.""" with torch.no_grad(): batch_size = target.size(0) _, pred = torch.topk(output, 1, 1) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) correct_k = correct[:1].reshape(-1).float().sum(0, keepdim = True) return correct_k.mul_(100.0 / batch_size) # Build the data loaders. def build_loaders(name): loader = agml.data.AgMLDataLoader(name) loader.split(train = 0.8, val = 0.1, test = 0.1) loader.batch(batch_size = 16) loader.resize_images('imagenet') loader.normalize_images('imagenet') loader.labels_to_one_hot() train_data = loader.train_data train_data.transform(transform = A.RandomRotate90()) train_ds = train_data.copy().as_torch_dataset() val_ds = loader.val_data.as_torch_dataset() val_ds.shuffle_data = False test_ds = loader.test_data.as_torch_dataset() return train_ds, val_ds, test_ds # Main training loop method. def train(dataset, pretrained, epochs, gpus, save_dir = None): """Constructs the training loop and trains a model.""" if save_dir is None: save_dir = os.path.join(f"/data2/amnjoshi/checkpoints/{dataset}") os.makedirs(save_dir, exist_ok = True) # Set up the checkpoint saving callback. callbacks = [ pl.callbacks.ModelCheckpoint( dirpath = save_dir, mode = 'min', filename = f"{dataset}" + "-epoch{epoch:02d}-val_loss_{val_loss:.2f}", monitor = 'val_loss', save_top_k = 3, auto_insert_metric_name = False ), pl.callbacks.EarlyStopping( monitor = 'val_loss', min_delta = 0.001, patience = 3, ) ] # Construct the model. model = ClassificationBenchmark( dataset = dataset, pretrained = pretrained) # Construct the data loaders. train_ds, val_ds, test_ds = build_loaders(dataset) # Create the trainer and train the model. trainer = pl.Trainer( max_epochs = epochs, gpus = get_num_gpus(gpus), callbacks = callbacks) trainer.fit( model = model, train_dataloaders = train_ds, val_dataloaders = val_ds ) if __name__ == '__main__': # Parse input arguments. ap = argparse.ArgumentParser() ap.add_argument( '--dataset', type = str, help = "The name of the dataset.") ap.add_argument( '--not-pretrained', action = 'store_false', default = True, help = "Whether to load a pretrained model.") ap.add_argument( '--checkpoint_dir', type = str, default = None, help = "The checkpoint directory to save to.") ap.add_argument( '--epochs', type = int, default = 50, help = "How many epochs to train for. Default is 50.") ap.add_argument( '--gpus', type = int, default = None, help = "How many GPUs to use when training.") args = ap.parse_args() # Train the model. train(args.dataset, args.not_pretrained, epochs = args.epochs, gpus = args.gpus, save_dir = args.checkpoint_dir)
6,726
31.814634
83
py
AgML
AgML-main/experiments/benchmarking/detection_lightning_local.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Some of the training code in this file is adapted from the following sources: 1. https://github.com/rwightman/efficientdet-pytorch 2. https://gist.github.com/Chris-hughes10/73628b1d8d6fc7d359b3dcbbbb8869d7 """ import os import argparse from typing import List, Union import numpy as np from PIL import Image import torch from torch.utils.data import Dataset, DataLoader import pytorch_lightning as pl from pytorch_lightning.loggers import CSVLogger, TensorBoardLogger from mean_average_precision import MeanAveragePrecision import agml import albumentations as A from albumentations.pytorch import ToTensorV2 from agml.models.training_resources.model import get_efficientdet_config, HeadNet, create_model_from_config from agml.models.training_resources.bench import DetBenchTrain, DetBenchPredict from ensemble_boxes import ensemble_boxes_wbf from tools import gpus, checkpoint_dir, MetricLogger, auto_move_data # Constants IMAGE_SIZE = 512 def create_model(num_classes = 1, architecture = "tf_efficientdet_d4", pretrained = (False, False)): if isinstance(pretrained, bool): if pretrained is False: pretrained = True else: if not pretrained[0]: pretrained = True if isinstance(pretrained, str): return create_model_from_pretrained(num_classes, architecture, pretrained) config = get_efficientdet_config(architecture) config.update({'image_size': (IMAGE_SIZE, IMAGE_SIZE)}) net = create_model_from_config(config, pretrained = pretrained, num_classes = num_classes) net.class_net = HeadNet( config, num_outputs = num_classes, ) return DetBenchTrain(net, config) # Modification of the above to load pretrained weights from a path. def create_model_from_pretrained( num_classes = 1, architecture = "tf_efficientdet_d4", pretrained_path = None): config = get_efficientdet_config(architecture) config.update({'image_size': (IMAGE_SIZE, IMAGE_SIZE)}) net = create_model_from_config( config, num_classes = pretrained_path[1], pretrained = False) if net.num_classes != num_classes: net.reset_head(num_classes = num_classes) net.load_state_dict( torch.load(pretrained_path[0], map_location = 'cpu')) return DetBenchTrain(net, config) class AgMLDatasetAdaptor(object): """Adapts an AgML dataset for use in a `LightningDataModule`.""" def __init__(self, loader, adapt_class = False): self.loader = loader self.adapt_class = adapt_class def __len__(self) -> int: return len(self.loader) def get_image_and_labels_by_idx(self, index): image, annotation = self.loader[index] image = Image.fromarray(image) bboxes = np.array(annotation['bbox']).astype(np.int32) x_min = bboxes[:, 0] y_min = bboxes[:, 1] x_max = bboxes[:, 2] + x_min y_max = bboxes[:, 3] + y_min x_min, y_min = np.clip(x_min, 0, image.width), np.clip(y_min, 0, image.height) x_max, y_max = np.clip(x_max, 0, image.width), np.clip(y_max, 0, image.height) bboxes = np.dstack((x_min, y_min, x_max, y_max)).squeeze(axis = 0) class_labels = np.array(annotation['category_id']).squeeze() if self.adapt_class: class_labels = np.ones_like(class_labels) return image, bboxes, class_labels, index def get_transforms(mode = 'inference'): """Returns a set of transforms corresponding to the mode.""" if mode == 'train': return A.Compose( [A.HorizontalFlip(p = 0.5), A.Resize(height = IMAGE_SIZE, width = IMAGE_SIZE, p = 1), ToTensorV2(p = 1)], p = 1.0, bbox_params = A.BboxParams( format = "pascal_voc", min_area = 0, min_visibility = 0, label_fields = ["labels"])) elif mode in ['val', 'validation']: return A.Compose( [A.Resize(height = IMAGE_SIZE, width = IMAGE_SIZE, p = 1), ToTensorV2(p = 1)], p = 1.0, bbox_params = A.BboxParams( format = "pascal_voc", min_area = 0, min_visibility = 0, label_fields = ["labels"])) elif mode == 'inference': return A.Compose( [A.Resize(height = IMAGE_SIZE, width = IMAGE_SIZE, p = 1), ToTensorV2(p = 1)], p = 1.0) class EfficientDetDataset(Dataset): def __init__(self, adaptor, transforms = None): self.ds = adaptor if transforms is None: transforms = get_transforms('val') self.transforms = transforms def __len__(self): return len(self.ds) def __getitem__(self, index): image, pascal_bboxes, class_labels, image_id = \ self.ds.get_image_and_labels_by_idx(index) # Add a label dimension for consistency. if class_labels.ndim == 0: class_labels = np.expand_dims(class_labels, axis = 0) # Construct the sample. sample = { "image": np.array(image, dtype = np.float32), "bboxes": pascal_bboxes, "labels": class_labels} try: sample = self.transforms(**sample) except: # debugging raise Exception(f"Failed sample: {sample}") sample["bboxes"] = np.array(sample["bboxes"]) image = sample["image"] labels = sample["labels"] # Convert 1-channel and 4-channel to 3-channel. if image.shape[0] == 1: image = torch.tile(image, (3, 1, 1)) if image.shape[0] == 4: image = image[:3] # Convert to yxyx from xyxy. _, new_h, new_w = image.shape sample["bboxes"][:, [0, 1, 2, 3]] = \ sample["bboxes"][:, [1, 0, 3, 2]] # Create the target from the annotations. target = { "bboxes": torch.as_tensor(sample["bboxes"], dtype = torch.float32), "labels": torch.as_tensor(labels), "image_id": torch.tensor([image_id]), "img_size": (new_h, new_w), "img_scale": torch.tensor([1.0])} return image, target, image_id class EfficientDetDataModule(pl.LightningDataModule): """A `LightningDataModule` for the `LightningModule`.""" def __init__(self, train_dataset_adaptor, validation_dataset_adaptor, train_transforms = None, val_transforms = None, num_workers = 4, batch_size = 8): self.train_ds = train_dataset_adaptor self.valid_ds = validation_dataset_adaptor if train_transforms is None: train_transforms = get_transforms('train') self.train_tfms = train_transforms if val_transforms is None: val_transforms = get_transforms('val') self.val_tfms = val_transforms self.num_workers = num_workers self.batch_size = batch_size super().__init__() def train_dataset(self) -> EfficientDetDataset: return EfficientDetDataset( adaptor = self.train_ds, transforms = self.train_tfms) def train_dataloader(self) -> DataLoader: return DataLoader( self.train_dataset(), batch_size = self.batch_size, shuffle = True, pin_memory = True, drop_last = True, num_workers = self.num_workers, collate_fn = self.collate_fn, ) def val_dataset(self) -> EfficientDetDataset: return EfficientDetDataset( adaptor = self.valid_ds, transforms = self.val_tfms) def val_dataloader(self) -> DataLoader: return DataLoader( self.val_dataset(), batch_size = self.batch_size, shuffle = False, pin_memory = True, drop_last = True, num_workers = self.num_workers, collate_fn = self.collate_fn, ) @staticmethod def collate_fn(batch): images, targets, image_ids = tuple(zip(*batch)) images = torch.stack(images) images = images.float() boxes = [target["bboxes"].float() for target in targets] labels = [target["labels"].float() for target in targets] img_size = torch.tensor([target["img_size"] for target in targets]).float() img_scale = torch.tensor([target["img_scale"] for target in targets]).float() annotations = { "bbox": boxes, "cls": labels, "img_size": img_size, "img_scale": img_scale} return images, annotations, targets, image_ids class EfficientDetModel(pl.LightningModule): def __init__(self, num_classes = 1, confidence_threshold = 0.3, learning_rate = 0.0002, wbf_iou_threshold = 0.44, inference_transforms = None, architecture = 'efficientdet_d4', save_dir = None, pretrained = False, validation_dataset_adaptor = None): super().__init__() self.model = create_model( num_classes, architecture = architecture, pretrained = pretrained) self.confidence_threshold = confidence_threshold self.lr = learning_rate self.wbf_iou_threshold = wbf_iou_threshold if inference_transforms is None: inference_transforms = get_transforms('inference') self.inference_tfms = inference_transforms # Construct the metric. self.val_dataset_adaptor = None if validation_dataset_adaptor is not None: # Add a metric calculator. self.val_dataset_adaptor = AgMLDatasetAdaptor( validation_dataset_adaptor) self.map = MeanAveragePrecision() self.metric_logger = DetectionMetricLogger({ 'map': MeanAveragePrecision()}, os.path.join(save_dir, f'logs-{self._version}.csv')) self._sanity_check_passed = False @auto_move_data def forward(self, images, targets): return self.model(images, targets) def configure_optimizers(self): return torch.optim.AdamW(self.model.parameters(), lr = self.lr) def training_step(self, batch, batch_idx): # Run a forward pass through the model. images, annotations, _, _ = batch losses = self.model(images, annotations) # Calculate and log losses. self.log("train_loss", losses["loss"], on_step = True, on_epoch = True, prog_bar = True, logger = True) self.log("train_class_loss", losses["class_loss"], on_step = True, on_epoch = True, logger = True) self.log("train_box_loss", losses["box_loss"], on_step = True, on_epoch = True, logger = True) return losses['loss'] @torch.no_grad() def validation_step(self, batch, batch_idx): images, annotations, targets, image_ids = batch outputs = self.model(images, annotations) detections = outputs["detections"] # Update the metric. if self.val_dataset_adaptor is not None and self._sanity_check_passed: predicted_bboxes, predicted_class_confidences, predicted_class_labels = \ self.post_process_detections(detections) for idx, pred_box, pred_conf, pred_labels in zip( image_ids, predicted_bboxes, predicted_class_confidences, predicted_class_labels): image, truth_boxes, truth_cls, _ = \ self.val_dataset_adaptor.get_image_and_labels_by_idx(idx) metric_update_values = \ [[pred_box, pred_labels, pred_conf], [truth_boxes, truth_cls]] self.metric_logger.update_metrics(*metric_update_values) self.map.update(*metric_update_values) # Compute the metric result. self.log("map", self.map.compute(), on_step = True, on_epoch = True, prog_bar = True, logger = True, sync_dist = True) batch_predictions = { "predictions": detections, "targets": targets, "image_ids": image_ids, } logging_losses = { "class_loss": outputs["class_loss"].detach(), "box_loss": outputs["box_loss"].detach(), } self.log("valid_loss", outputs["loss"], on_step = True, on_epoch = True, prog_bar = True, logger = True, sync_dist = True) self.log("valid_class_loss", logging_losses["class_loss"], on_step = True, on_epoch = True, logger = True, sync_dist = True) self.log("valid_box_loss", logging_losses["box_loss"], on_step = True, on_epoch = True, logger = True, sync_dist = True) return {'loss': outputs["loss"], 'batch_predictions': batch_predictions} def predict(self, images: Union[torch.Tensor, List]): """Runs inference on a set of images. Parameters ---------- images : {torch.Tensor, list} Either a list of images (which can be numpy arrays, tensors, or another type), or a torch.Tensor returned from a DataLoader. Returns ------- A tuple containing bounding boxes, confidence scores, and class labels. """ if isinstance(images, list): image_sizes = [(image.size[1], image.size[0]) for image in images] images_tensor = torch.stack([ self.inference_tfms( image = np.array(image, dtype = np.float32), )["image"] for image in images]) return self._run_inference(images_tensor, image_sizes) elif isinstance(images, torch.Tensor): image_tensor = images if image_tensor.ndim == 3: image_tensor = image_tensor.unsqueeze(0) if image_tensor.shape[-1] != IMAGE_SIZE \ or image_tensor.shape[-2] != IMAGE_SIZE: raise ValueError( f"Input tensors must be of shape " f"(N, 3, {IMAGE_SIZE}, {IMAGE_SIZE})") num_images = image_tensor.shape[0] image_sizes = [(IMAGE_SIZE, IMAGE_SIZE)] * num_images return self._run_inference(image_tensor, image_sizes) else: raise TypeError( "Expected either a list of images or a " "torch.Tensor of images for `predict()`.") def _run_inference(self, images_tensor, image_sizes): dummy_targets = self._create_dummy_inference_targets( images_tensor.shape[0], self.device, IMAGE_SIZE) detections = self.model( images_tensor.to(self.device), dummy_targets)["detections"] predicted_bboxes, predicted_class_confidences, predicted_class_labels = \ self.post_process_detections(detections) scaled_bboxes = self._rescale_bboxes( predicted_bboxes = predicted_bboxes, image_sizes = image_sizes) return scaled_bboxes, predicted_class_labels, predicted_class_confidences @staticmethod def _create_dummy_inference_targets(num_images, device, size): return { "bbox": [ torch.tensor([[0.0, 0.0, 0.0, 0.0]], device = device) for _ in range(num_images) ], "cls": [torch.tensor([1.0], device = device) for _ in range(num_images)], "img_size": torch.tensor( [(size, size)] * num_images, device = device).float(), "img_scale": torch.ones(num_images, device = device).float(), } def post_process_detections(self, detections): predictions = [self._postprocess_single_prediction_detections(d) for d in detections] predicted_bboxes, predicted_class_confidences, predicted_class_labels = self.run_wbf( predictions, image_size = IMAGE_SIZE, iou_thr = self.wbf_iou_threshold) return predicted_bboxes, predicted_class_confidences, predicted_class_labels def _postprocess_single_prediction_detections(self, detections): # Extract the bounding boxes, confidence scores, # and class labels from the output detections. boxes = detections.detach().cpu().numpy()[:, :4] scores = detections.detach().cpu().numpy()[:, 4] classes = detections.detach().cpu().numpy()[:, 5] # Only return boxes which are above the confidence threshold. valid_indexes = np.where(scores > self.confidence_threshold)[0] boxes = boxes[valid_indexes] scores = scores[valid_indexes] classes = classes[valid_indexes] return {"boxes": boxes, "scores": scores, "classes": classes} @staticmethod def _rescale_bboxes(predicted_bboxes, image_sizes): scaled_bboxes = [] for bboxes, img_dims in zip(predicted_bboxes, image_sizes): im_h, im_w = img_dims if len(bboxes) > 0: scaled_bboxes.append( (np.array(bboxes) * [ im_w / IMAGE_SIZE, im_h / IMAGE_SIZE, im_w / IMAGE_SIZE, im_h / IMAGE_SIZE ]).tolist()) else: scaled_bboxes.append(bboxes) return scaled_bboxes @staticmethod def run_wbf(predictions, image_size = 512, iou_thr = 0.44, skip_box_thr = 0.43, weights = None): bboxes, confidences, class_labels = [], [], [] for prediction in predictions: boxes = [(prediction["boxes"] / image_size).tolist()] scores = [prediction["scores"].tolist()] labels = [prediction["classes"].tolist()] boxes, scores, labels = ensemble_boxes_wbf.weighted_boxes_fusion( boxes, scores, labels, weights = weights, iou_thr = iou_thr, skip_box_thr = skip_box_thr) boxes = boxes * (image_size - 1) bboxes.append(boxes.tolist()) confidences.append(scores.tolist()) class_labels.append(labels.tolist()) return bboxes, confidences, class_labels def on_validation_epoch_end(self) -> None: if not self._sanity_check_passed: self._sanity_check_passed = True return if hasattr(self, 'metric_logger'): self.metric_logger.compile_epoch() def on_fit_end(self) -> None: if hasattr(self, 'metric_logger'): self.metric_logger.save() def get_progress_bar_dict(self): p_bar = super(EfficientDetModel, self).get_progress_bar_dict() p_bar.pop('v_num', None) return p_bar # Calculate and log the metrics. class DetectionMetricLogger(MetricLogger): def update_metrics(self, y_pred, y_true) -> None: self.metrics['map'].update(y_pred, y_true) def train(dataset, epochs, save_dir = None, overwrite = None, pretrained_path = None): """Constructs the training loop and trains a model.""" save_dir = checkpoint_dir(save_dir, dataset) log_dir = save_dir.replace('checkpoints', 'logs') # Check if the dataset already has benchmarks. if os.path.exists(save_dir) and os.path.isdir(save_dir): if not overwrite and len(os.listdir(save_dir)) >= 4: print(f"Checkpoints already exist for {dataset} " f"at {save_dir}, skipping generation.") return # Set up the checkpoint saving callback. callbacks = [ pl.callbacks.ModelCheckpoint( dirpath = save_dir, mode = 'min', filename = f"{dataset}" + "-epoch{epoch:02d}-valid_loss_{valid_loss:.2f}", monitor = 'valid_loss', save_top_k = 3, auto_insert_metric_name = False ) ] # Create the loggers. loggers = [ CSVLogger(log_dir), TensorBoardLogger(log_dir) ] # Construct the data. pl.seed_everything(2499751) loader = agml.data.AgMLDataLoader(dataset) loader.shuffle() loader.split(train = 0.8, val = 0.1, test = 0.1) dm = EfficientDetDataModule( train_dataset_adaptor = AgMLDatasetAdaptor(loader.train_data), validation_dataset_adaptor = AgMLDatasetAdaptor(loader.val_data), num_workers = 12, batch_size = 4) # Construct the model. model = EfficientDetModel( num_classes = loader.info.num_classes, architecture = 'tf_efficientdet_d4', pretrained = True) # Create the trainer and train the model. msg = f"Training dataset {dataset}!" print("\n" + "=" * len(msg) + "\n" + msg + "\n" + "=" * len(msg) + "\n") trainer = pl.Trainer( max_epochs = epochs, gpus = gpus(None), callbacks = callbacks, logger = loggers) trainer.fit(model, dm) # Save the final state. torch.save(model.state_dict(), os.path.join(save_dir, 'final_model.pth')) def train_per_class(dataset, epochs, save_dir = None, overwrite = None): """Constructs the training loop and trains a model.""" save_dir = checkpoint_dir(save_dir, dataset) log_dir = save_dir.replace('checkpoints', 'logs') # Check if the dataset already has benchmarks. if os.path.exists(save_dir) and os.path.isdir(save_dir): if not overwrite and len(os.listdir(save_dir)) >= 4: print(f"Checkpoints already exist for {dataset} " f"at {save_dir}, skipping generation.") return # Construct the loader. pl.seed_everything(2499751) loader = agml.data.AgMLDataLoader(dataset) loader.shuffle() # Create the loop for each class. for cl in range(agml.data.source(dataset).num_classes): # Create the data module with the new, reduced class. cls = cl + 1 new_loader = loader.take_class(cls) new_loader.split(train = 0.8, val = 0.1, test = 0.1) dm = EfficientDetDataModule( train_dataset_adaptor = AgMLDatasetAdaptor( new_loader.train_data, adapt_class = True), validation_dataset_adaptor = AgMLDatasetAdaptor( new_loader.val_data, adapt_class = True), num_workers = 12, batch_size = 4) this_save_dir = os.path.join( save_dir, f'{new_loader.num_to_class[cls]}-{cls}') this_log_dir = this_save_dir.replace('checkpoints', 'logs') # Set up the checkpoint saving callback. callbacks = [ pl.callbacks.ModelCheckpoint( dirpath = this_save_dir, mode = 'min', filename = f"{dataset}" + "-epoch{epoch:02d}-valid_loss_{valid_loss:.2f}", monitor = 'valid_loss', save_top_k = 3, auto_insert_metric_name = False ) ] # Create the loggers. loggers = [ CSVLogger(this_log_dir), TensorBoardLogger(this_log_dir) ] # Construct the model. model = EfficientDetModel( num_classes = 1, architecture = 'tf_efficientdet_d4', pretrained = True) model.load_state_dict( torch.load('/data2/amnjoshi/full_grape/checkpoints/final_model.pth', map_location = 'cpu')) # Create the trainer and train the model. msg = f"Training dataset {dataset} for class {cls}: {loader.num_to_class[cls]}!" print("\n" + "=" * len(msg) + "\n" + msg + "\n" + "=" * len(msg) + "\n") trainer = pl.Trainer( max_epochs = epochs, gpus = gpus(None), callbacks = callbacks, logger = loggers) trainer.fit(model, dm) # Save the final state. torch.save(model.state_dict(), os.path.join(this_save_dir, 'final_model.pth')) if __name__ == '__main__': # Parse input arguments. ap = argparse.ArgumentParser() ap.add_argument( '--dataset', type = str, nargs = '+', help = "The name of the dataset.") ap.add_argument( '--regenerate-existing', action = 'store_true', default = False, help = "Whether to re-generate existing benchmarks.") ap.add_argument( '--checkpoint_dir', type = str, default = None, help = "The checkpoint directory to save to.") ap.add_argument( '--epochs', type = int, default = 50, help = "How many epochs to train for. Default is 50.") ap.add_argument( '--per-class-for-dataset', action = 'store_true', default = False, help = "Whether to generate benchmarks per class.") ap.add_argument( '--pretrained-model-path', type = str, default = None, help = "The path to a set of pretrained weights for the model.") ap.add_argument( '--pretrained-num-classes', type = str, default = None, help = "The number of classes in the pretrained model..") args = ap.parse_args() # Train the model. if args.per_class_for_dataset: train_per_class(args.dataset[0], epochs = args.epochs, save_dir = args.checkpoint_dir) elif args.dataset[0] in agml.data.public_data_sources(ml_task = 'object_detection') \ and len(args.dataset) > 1: train(args.dataset, epochs = args.epochs, save_dir = args.checkpoint_dir, pretrained_path = (args.pretrained_model_path, args.pretrained_num_classes)) else: if args.dataset[0] == 'all': datasets = [ds for ds in agml.data.public_data_sources( ml_task = 'object_detection')] else: datasets = args.dataset for ds in datasets: train(ds, epochs = args.epochs, save_dir = args.checkpoint_dir, overwrite = args.regenerate_existing, pretrained_path = (args.pretrained_model_path, args.pretrained_num_classes))
26,638
37.607246
107
py
AgML
AgML-main/experiments/benchmarking/miou_evaluation.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import argparse import pandas as pd from tqdm import tqdm import torch import pytorch_lightning as pl import agml from torchmetrics import IoU from segmentation_lightning import SegmentationBenchmark # Define device. device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') def run_evaluation(model, name): """Runs evaluation for mean intersection over union..""" # Load the test dataset. pl.seed_everything(2499751) loader = agml.data.AgMLDataLoader(name) loader.split(train = 0.8, val = 0.1, test = 0.1) loader.batch(batch_size = 2) loader.resize_images((512, 512)) loader.normalize_images('imagenet') loader.mask_to_channel_basis() ds = loader.test_data.as_torch_dataset() # Create the metric. iou = IoU(num_classes = ds.num_classes + 1) # Run inference for all of the images in the test dataset. for i in tqdm(range(len(ds)), leave = False): image, annotation = ds[i] y_pred = model(image.to(device)).float().squeeze() iou(y_pred.detach().cpu(), annotation.int().cpu()) # Compute the mAP for all of the thresholds. print(iou.compute().detach().cpu().numpy()) return iou.compute().detach().cpu().numpy() def make_checkpoint(name): """Gets a checkpoint for the model name.""" ckpt_path = os.path.join( "/data2/amnjoshi/final/segmentation_checkpoints", name, "final_model.pth") state = torch.load(ckpt_path, map_location = 'cpu') model = SegmentationBenchmark(dataset = name) model.load_state_dict(state) model.eval().to(device) return model def evaluate(names, log_file = None): """Runs the evaluation and saves results to a file.""" print(f"Running mIoU evaluation for {names}.") # Create the log file. if log_file is None: log_file = os.path.join(os.getcwd(), 'miou_evaluation.csv') # Run the evaluation. log_contents = {} bar = tqdm(names) for name in bar: ckpt = make_checkpoint(name) bar.set_description(f"Evaluating {name}") if hasattr(name, 'name'): name = name.name log_contents[name] = run_evaluation(ckpt, name) # Save the results. df = pd.DataFrame(columns = ('name', 'miou')) for name, value in log_contents.items(): df.loc[len(df.index)] = [name, value] df.to_csv(log_file) if __name__ == '__main__': ap = argparse.ArgumentParser() ap.add_argument( '--dataset', type = str, nargs = '+', help = "The name of the dataset.") ap.add_argument( '--log_file', type = str, default = None, help = "The name of the output log file.") args = ap.parse_args() # Train the model. if args.dataset[0] in agml.data.public_data_sources(ml_task = 'semantic_segmentation'): datasets = args.dataset[0] else: if args.dataset[0] == 'all': datasets = [ds for ds in agml.data.public_data_sources( ml_task = 'semantic_segmentation')] elif args.dataset[0] == 'except': exclude_datasets = args.dataset[1:] datasets = [ dataset for dataset in agml.data.public_data_sources( ml_task = 'semantic_segmentation') if dataset.name not in exclude_datasets] else: datasets = args.dataset evaluate(datasets, args.log_file)
3,988
31.169355
91
py
AgML
AgML-main/experiments/benchmarking/detection_modeling.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os from typing import Union import torch import numpy as np from effdet import get_efficientdet_config, DetBenchTrain, create_model_from_config from agml.models.detection import DetectionModel from experiments.benchmarking.mean_average_precision_torch import MeanAveragePrecision class DetectionTrainingModel(DetectionModel): """Wraps an `EfficientDet` model for a training experiment.""" def __init__(self, num_classes: int = 1, pretrained_weights: str = None, confidence_threshold: float = 0.3, learning_rate: float = 0.0002, wbf_iou_threshold: float = 0.44, **kwargs): # Initialize the super module. super(DetectionTrainingModel, self).__init__(model_initialized = True) # Initialize the model using the provided arguments for customizability. self.model = self.make_model( num_classes = num_classes, pretrained_weights = pretrained_weights, architecture = kwargs.get('architecture', 'tf_efficientdet_d4')) # Set the training parameters. self._confidence_threshold = confidence_threshold self.lr = learning_rate self.wbf_iou_threshold = wbf_iou_threshold # Construct the metric. self.val_dataset_adaptor = None self.map = MeanAveragePrecision(self.model.num_classes) self._sanity_check_passed = False @staticmethod def make_model(num_classes: int, pretrained_weights: str, image_size: Union[int, tuple] = 512, architecture: str = 'tf_efficientdet_d4'): """Constructs the `EfficientDet` model from the provided parameters.""" # Parse the input arguments. if isinstance(image_size, int): image_size = (image_size, image_size) if pretrained_weights is None or pretrained_weights is False: pretrained_weights = '' # Construct the configuration. cfg = get_efficientdet_config(architecture) cfg.update({'image_size': image_size}) # Build the model. model_default_pretrained = False if pretrained_weights == 'coco': model_default_pretrained = True net = create_model_from_config( cfg, pretrained = model_default_pretrained, num_classes = num_classes) # Load the pretrained weights if they are provided. if os.path.exists(pretrained_weights): # Auto-inference the number of classes from the state dict. state = torch.load(pretrained_weights, map_location = 'cpu') weight_key = [i for i in state.keys() if 'class_net.predict.conv_pw.weight' in i][0] pretrained_num_classes = int(state[weight_key].shape[0] / 9) # Load the pretrained weights. net.reset_head(num_classes = pretrained_num_classes) net.load_state_dict(state) # Restore the number of classes. if num_classes != pretrained_num_classes: net.reset_head(num_classes = num_classes) # Return the network. return DetBenchTrain(net) @staticmethod def _rescale_bboxes_xyxy(predicted_bboxes, image_sizes): """Re-scales output bounding boxes to the original image sizes. This is re-written in this training subclass since annotations here are also mapped to the YXYX format, and that needs to be accounted for. """ scaled_boxes = [] for bboxes, img_dims in zip(predicted_bboxes, image_sizes): h, w = img_dims if len(bboxes) > 0: # Re-scale the bounding box to the appropriate format. scale_ratio = [h / 512, w / 512, h / 512, w / 512] scaled = (np.array(bboxes.detach().cpu()) * scale_ratio).astype(np.int32) # Convert the Pascal-VOC (yxyx) format to COCO (xywh). y, x = scaled[:, 0], scaled[:, 1] h, w = scaled[:, 2] - y, scaled[:, 3] - x scaled_boxes.append(np.dstack((x, y, w, h))) continue # Otherwise, there is no prediction for this image. scaled_boxes.append(np.array([])) return scaled_boxes def configure_optimizers(self): return torch.optim.Adam(self.model.parameters(), lr = self.lr) def training_step(self, batch, batch_idx): # Run a forward pass through the model. images, annotations, _ = batch losses = self.model(images, annotations) # Calculate and log losses. self.log("train_loss", losses["loss"], on_step = True, on_epoch = True, prog_bar = True, logger = True) self.log("train_class_loss", losses["class_loss"], on_step = True, on_epoch = True, logger = True) self.log("train_box_loss", losses["box_loss"], on_step = True, on_epoch = True, logger = True) return losses['loss'] @torch.no_grad() def validation_step(self, batch, batch_idx): images, annotations, targets = batch outputs = self.model(images, annotations) detections = outputs["detections"] # Calculate the mean average precision. if self._sanity_check_passed: boxes, confidences, labels = self._process_detections(self._to_out(detections)) boxes = self._rescale_bboxes(boxes, [[512, 512]] * len(images)) annotations['bbox'] = self._rescale_bboxes_xyxy( annotations['bbox'], [[512, 512, ]] * len(images)) for pred_box, pred_label, pred_conf, true_box, true_label in zip( boxes, labels, confidences, annotations['bbox'], annotations['cls']): metric_update_values = \ dict(boxes = self._to_out(torch.tensor(pred_box, dtype = torch.float32)), labels = self._to_out(torch.tensor(pred_label, dtype = torch.int32)), scores = self._to_out(torch.tensor(pred_conf))), \ dict(boxes = self._to_out(torch.tensor(true_box, dtype = torch.float32)), labels = self._to_out(torch.tensor(true_label, dtype = torch.int32))) self.map.update(*metric_update_values) # Log the MAP values. map_ = self.map.compute().detach().cpu().numpy().item() self.log("map", map_, prog_bar = True, on_epoch = True, logger = True, sync_dist = True) batch_predictions = { "predictions": detections, "targets": targets, } logging_losses = { "class_loss": outputs["class_loss"].detach(), "box_loss": outputs["box_loss"].detach(), } self.log("valid_loss", outputs["loss"], on_step = True, on_epoch = True, prog_bar = True, logger = True, sync_dist = True) self.log("valid_class_loss", logging_losses["class_loss"], on_step = True, on_epoch = True, logger = True, sync_dist = True) self.log("valid_box_loss", logging_losses["box_loss"], on_step = True, on_epoch = True, logger = True, sync_dist = True) return {'loss': outputs["loss"], 'batch_predictions': batch_predictions} def on_validation_epoch_end(self) -> None: """Log mean average precision at the end of each epoch.""" # No validation should be run during the sanity check step. if not self._sanity_check_passed: self._sanity_check_passed = True return # Compute the mean average precision and reset it. if hasattr(self, 'map'): map_ = self.map.compute().detach().cpu().numpy().item() self.log("map_epoch", map_, prog_bar = True, on_epoch = True, logger = True, sync_dist = True) self.map.reset() def get_progress_bar_dict(self) -> dict: """Remove the `v_num` from the bar; it takes away valuable space.""" p_bar = super(DetectionTrainingModel, self).get_progress_bar_dict() p_bar.pop('v_num', None) return p_bar
8,868
41.845411
96
py
AgML
AgML-main/experiments/benchmarking/classification_lightning_resnet50.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import argparse import torch import torch.nn as nn import pytorch_lightning as pl from pytorch_lightning.loggers import CSVLogger, TensorBoardLogger from torchmetrics.classification import Precision, Recall, Accuracy from torchvision.models import resnet50 import agml import albumentations as A from tools import gpus, checkpoint_dir, MetricLogger class ResNet50Transfer(nn.Module): """Represents a transfer learning ResNet50 model. This is the base benchmarking model for image classification, using the ResNet50 model with two added linear fully-connected layers. """ def __init__(self, num_classes, pretrained = True): super(ResNet50Transfer, self).__init__() self.base = resnet50(pretrained = pretrained) self.l1 = nn.Linear(1000, 256) self.dropout = nn.Dropout(0.1) self.relu = nn.ReLU() self.l2 = nn.Linear(256, num_classes) def forward(self, x, **kwargs): # noqa x = self.base(x) x = x.view(x.size(0), -1) x = self.dropout(self.relu(self.l1(x))) x = self.l2(x) return x class ClassificationBenchmark(pl.LightningModule): """Represents an image classification benchmark model.""" def __init__(self, dataset, pretrained = False, save_dir = None): # Initialize the module. super(ClassificationBenchmark, self).__init__() # Construct the network. self._source = agml.data.source(dataset) self._pretrained = pretrained self.net = ResNet50Transfer( self._source.num_classes, self._pretrained ) # Construct the loss for training. self.loss = nn.CrossEntropyLoss() # Add a metric calculator. self.metric_logger = ClassificationMetricLogger({ 'accuracy': Accuracy(num_classes = self._source.num_classes), 'precision': Precision(num_classes = self._source.num_classes), 'recall': Recall(num_classes = self._source.num_classes)}, os.path.join(save_dir, f'logs-{self._version}.csv')) self._sanity_check_passed = False def forward(self, x): return self.net.forward(x) def training_step(self, batch, *args, **kwargs): # noqa x, y = batch y_pred = self(x) loss = self.loss(y_pred, y) acc = accuracy(y_pred, torch.argmax(y, 1)).item() self.log('accuracy', acc, prog_bar = True, logger = True) self.log('loss', loss, logger = True) return { 'loss': loss, 'accuracy': acc } def validation_step(self, batch, *args, **kwargs): # noqa x, y = batch y_pred = self(x) val_loss = self.loss(y_pred, y) val_acc = accuracy(y_pred, torch.argmax(y, 1)) if self._sanity_check_passed: self.metric_logger.update(y_pred, torch.argmax(y, 1)) self.log('val_loss', val_loss.item(), prog_bar = True, logger = True) self.log('val_accuracy', val_acc.item(), prog_bar = True, logger = True) return { 'val_loss': val_loss, 'val_accuracy': val_acc } def configure_optimizers(self): return torch.optim.Adam(self.parameters()) def get_progress_bar_dict(self): tqdm_dict = super(ClassificationBenchmark, self)\ .get_progress_bar_dict() tqdm_dict.pop('v_num', None) return tqdm_dict def on_validation_epoch_end(self) -> None: if not self._sanity_check_passed: self._sanity_check_passed = True return self.metric_logger.compile_epoch() def on_fit_end(self) -> None: self.metric_logger.save() # Calculate and log the metrics. class ClassificationMetricLogger(MetricLogger): def update_metrics(self, y_pred, y_true) -> None: for metric in self.metrics.values(): metric.update(y_pred.cpu(), y_true.cpu()) def accuracy(output, target): """Computes the accuracy between `output` and `target`.""" with torch.no_grad(): batch_size = target.size(0) _, pred = torch.topk(output, 1, 1) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) correct_k = correct[:1].reshape(-1).float().sum(0, keepdim = True) return correct_k.mul_(100.0 / batch_size) # Build the data loaders. def build_loaders(name): pl.seed_everything(2499751) loader = agml.data.AgMLDataLoader(name) loader.split(train = 0.8, val = 0.1, test = 0.1) loader.batch(batch_size = 16) loader.resize_images('imagenet') loader.normalize_images('imagenet') loader.labels_to_one_hot() train_data = loader.train_data train_data.transform(transform = A.RandomRotate90()) train_ds = train_data.copy().as_torch_dataset() val_ds = loader.val_data.as_torch_dataset() val_ds.shuffle_data = False test_ds = loader.test_data.as_torch_dataset() return train_ds, val_ds, test_ds def train(dataset, pretrained, epochs, save_dir = None, overwrite = None): """Constructs the training loop and trains a model.""" save_dir = "/data2/amnjoshi/resnet50_pretrained/checkpoints" os.makedirs(save_dir, exist_ok = True) log_dir = save_dir.replace('checkpoints', 'logs') os.makedirs(log_dir, exist_ok = True) # Check if the dataset already has benchmarks. if os.path.exists(save_dir) and os.path.isdir(save_dir): if not overwrite and len(os.listdir(save_dir)) >= 4: print(f"Checkpoints already exist for {dataset} " f"at {save_dir}, skipping generation.") return # Set up the checkpoint saving callback. callbacks = [ pl.callbacks.ModelCheckpoint( dirpath = save_dir, mode = 'min', filename = f"{dataset}" + "-epoch{epoch:02d}-val_loss_{val_loss:.2f}", monitor = 'val_loss', save_top_k = 3, auto_insert_metric_name = False ), pl.callbacks.EarlyStopping( monitor = 'val_loss', min_delta = 0.0001, patience = 10, ) ] # Construct the model. model = ClassificationBenchmark( dataset = dataset, pretrained = pretrained, save_dir = save_dir) # Construct the data loaders. train_ds, val_ds, test_ds = build_loaders(dataset) # Create the loggers. loggers = [ CSVLogger(log_dir), TensorBoardLogger(log_dir) ] # Create the trainer and train the model. msg = f"Training dataset {dataset}!" print("\n" + "=" * len(msg) + "\n" + msg + "\n" + "=" * len(msg) + "\n") trainer = pl.Trainer( max_epochs = epochs, gpus = gpus(None), callbacks = callbacks, logger = loggers, log_every_n_steps = 5) trainer.fit( model = model, train_dataloaders = train_ds, val_dataloaders = val_ds, ) if __name__ == '__main__': # Parse input arguments. ap = argparse.ArgumentParser() ap.add_argument( '--dataset', type = str, nargs = '+', help = "The name of the dataset.") ap.add_argument( '--regenerate-existing', action = 'store_true', default = False, help = "Whether to re-generate existing benchmarks.") ap.add_argument( '--not-pretrained', action = 'store_false', default = True, help = "Whether to load a pretrained model.") ap.add_argument( '--checkpoint_dir', type = str, default = None, help = "The checkpoint directory to save to.") ap.add_argument( '--epochs', type = int, default = 20, help = "How many epochs to train for. Default is 20.") args = ap.parse_args() # Train the model. if args.dataset[0] in agml.data.public_data_sources(ml_task = 'image_classification'): train(args.dataset, args.not_pretrained, epochs = args.epochs, save_dir = args.checkpoint_dir) else: if args.dataset[0] == 'all': datasets = [ds for ds in agml.data.public_data_sources( ml_task = 'image_classification')] else: datasets = args.dataset for dataset in datasets: train(dataset, args.not_pretrained, epochs = args.epochs, save_dir = args.checkpoint_dir, overwrite = args.regenerate_existing)
9,011
33.136364
90
py
AgML
AgML-main/agml/models/base.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import abc from typing import List, Union, overload import cv2 import numpy as np import torch from pytorch_lightning import LightningModule from agml.framework import AgMLSerializable from agml.backend.config import model_save_path from agml.backend.tftorch import is_array_like from agml.utils.image import imread_context from agml.utils.downloads import download_model from agml.models.benchmarks import BenchmarkMetadata class AgMLModelBase(AgMLSerializable, LightningModule): """Base class for all AgML pretrained models. All pretrained model variants in AgML inherit from this base class, which provides common methods which each use, such as weight loading and image input preprocessing, as well as other stubs for common methods. """ def __init__(self): self._benchmark = BenchmarkMetadata(None) super(AgMLModelBase, self).__init__() @property def original(self): """Returns the original model architecture (without weights).""" return self.net.base @overload def preprocess_input(self, images: str) -> "torch.Tensor": ... @overload def preprocess_input(self, images: List[str]) -> "torch.Tensor": ... @overload def preprocess_input(self, images: Union[np.ndarray, torch.Tensor]) -> "torch.Tensor": ... @overload def preprocess_input(self, images: List[Union[np.ndarray, torch.Tensor]]) -> "torch.Tensor": ... @abc.abstractmethod def preprocess_input(self, *args, **kwargs): """Preprocesses input images to model specifications.""" raise NotImplementedError @overload def predict(self, images: str) -> "torch.Tensor": ... @overload def predict(self, images: List[str]) -> "torch.Tensor": ... @overload def predict(self, images: Union[np.ndarray, torch.Tensor]) -> "torch.Tensor": ... @overload def predict(self, images: List[Union[np.ndarray, torch.Tensor]]) -> "torch.Tensor": ... @abc.abstractmethod def predict(self, *args, **kwargs): """Runs model inference on input image(s).""" raise NotImplementedError @staticmethod def _expand_input_images(images): """Expands the input list of images to a specification. This is particularly useful because the model accepts numerous inputs, ranging from a list of image paths all the way to an already pre-processed image batch. This method standardizes all inputs to a common format, in particular, a list of all of the images that are going to be then passed to the input preprocessing method, before being passed through the model for inference. """ # First check for a path or a list of paths, for speed. if isinstance(images, str): with imread_context(images) as image: return [cv2.cvtColor(image, cv2.COLOR_BGR2RGB), ] elif isinstance(images, list) and isinstance(images[0], str): parsed_images = [] for path in images: with imread_context(path) as image: parsed_images.append(cv2.cvtColor(image, cv2.COLOR_BGR2RGB)) return parsed_images # Then check if we already have read-in images, either one # single image or a batch of images. if is_array_like(images): # Check if there is one single input image, or a batch of input # images. If there is a single input image, just return this. if images.ndim == 3: return [images, ] # Check if we have a batch of images first. This check is # done by seeing if the input is 4-dimensional. if images.ndim == 4: # If so, unstack the images along the first dimension. return [i for i in images] # Finally, the only remaining viable input type is a list of images. if isinstance(images, list) and is_array_like(images[0]): return images # Otherwise, we need to raise an error. raise TypeError( "Expected an input of a list of paths or images, a " "single path or image, or a batched image tensor for " f"preprocessing inputs, instead got {type(images)}.") @staticmethod def _to_out(tensor: "torch.Tensor") -> "torch.Tensor": return tensor.detach().cpu().numpy() @staticmethod def _get_shapes(images: list) -> list: """Gets the height and width of each of the input images.""" shapes = [] for image in images: if image.ndim == 2: shapes.append(image.shape) continue if image.shape[0] <= 3: # channels first shapes.append(image.shape[1:]) else: # channels last shapes.append(image.shape[:2]) return shapes @property def benchmark(self): """Information about the loaded benchmark.""" return self._benchmark @benchmark.setter def benchmark(self, value): self._benchmark = value @staticmethod def _get_benchmark(name): """Returns the `state_dict` for a pretrained model benchmark.""" # Check if the benchmark exists; if not, download it. benchmark_path = os.path.join(model_save_path(), name + '.pth') if not os.path.exists(benchmark_path): download_model(name, os.path.dirname(benchmark_path)) # Load the benchmark. return torch.load(benchmark_path, map_location = 'cpu') @abc.abstractmethod def load_benchmark(self, dataset: str): """Loads a benchmark for the given AgML dataset.""" raise NotImplementedError @abc.abstractmethod def evaluate(self, loader, **kwargs): """Evaluates the model on the given loader.""" raise NotImplementedError
6,551
33.851064
96
py
AgML
AgML-main/agml/models/classification.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from tqdm import tqdm import torch import torch.nn as nn try: from torchvision.models import efficientnet_b4 except ImportError: raise ImportError("To use image classification models in `agml.models`, you " "need to install Torchvision first. You can do this by " "running `pip install torchvision`.") from agml.models.base import AgMLModelBase from agml.models.tools import auto_move_data, imagenet_style_process from agml.models.metrics.accuracy import Accuracy class EfficientNetB4Transfer(nn.Module): """Wraps an EfficientDetB4 model with a classification head.""" def __init__(self, num_classes): super(EfficientNetB4Transfer, self).__init__() self.base = efficientnet_b4(pretrained = False) self.l1 = nn.Linear(1000, 256) self.dropout = nn.Dropout(0.1) self.relu = nn.ReLU() self.l2 = nn.Linear(256, num_classes) def forward(self, x, **kwargs): # noqa x = self.base(x) x = x.view(x.size(0), -1) x = self.dropout(self.relu(self.l1(x))) return self.l2(x) class ClassificationModel(AgMLModelBase): """Wraps an `EfficientNetB4` model for agricultural image classification. When using the model for inference, you should use the `predict()` method on any set of input images. This method wraps the `forward()` call with additional steps which perform the necessary preprocessing on the inputs, including resizing, normalization, and batching, as well as additional post-processing on the outputs (such as converting one-hot labels to integer labels), which will allow for a streamlined inference pipeline. If you want to use the `forward()` method directly, so that you can just call `model(inputs)`, then make sure your inputs are properly processed. This can be done by calling `model.preprocess_input(inputs)` on the input list of images/single image and then passing that result to the model. This will also return a one-hot label feature vector instead of integer labels, in the case that you want further customization of the outputs. This model can be subclassed in order to run a full training job; the actual transfer `EfficientNetB4` model can be accessed through the parameter `net`, and you'll need to implement methods like `training_step`, `configure_optimizers`, etc. See PyTorch Lightning for more information. """ serializable = frozenset(("model", "regression")) state_override = frozenset(("model",)) def __init__(self, num_classes = None, regression = False, **kwargs): # Construct the network and load in pretrained weights. super(ClassificationModel, self).__init__() self._regression = regression if not kwargs.get('model_initialized', False): self._num_classes = num_classes self.net = self._construct_sub_net(num_classes) @auto_move_data def forward(self, batch): return self.net(batch) @staticmethod def _construct_sub_net(num_classes): return EfficientNetB4Transfer(num_classes) @staticmethod def _preprocess_image(image, **kwargs): """Preprocesses a single input image to EfficientNet standards. The preprocessing steps are applied logically; if the images are passed with preprocessing already having been applied, for instance, the images are already resized or they are already been normalized, the operation is not applied again, for efficiency. Preprocessing includes the following steps: 1. Resizing the image to size (224, 224). 2. Performing normalization with ImageNet parameters. 3. Converting the image into a PyTorch tensor format. as well as other intermediate steps such as adding a channel dimension for two-channel inputs, for example. """ return imagenet_style_process(image, **kwargs) @staticmethod def preprocess_input(images = None, **kwargs) -> "torch.Tensor": """Preprocesses the input image to the specification of the model. This method takes in a set of inputs and preprocesses them into the expected format for the `EfficientNetB4` image classification model. There are a variety of inputs which are accepted, including images, image paths, as well as fully-processed image tensors. The inputs are expanded and standardized, then run through a preprocessing pipeline which formats them into a single tensor ready for the model. Preprocessing steps include normalization, resizing, and converting to the channels-first format used by PyTorch models. The output of this method will be a single `torch.Tensor`, which has shape [N, C, H, W], where `N` is the batch dimension. If only a single image is passed, this will have a value of 1. This method is largely beneficial when you just want to preprocess images into the specification of the model, without getting the output. Namely, `predict()` is essentially just a wrapper around this method and `forward()`, so you can run this externally and then run `forward()` to get the original model outputs, without any extra post-processing. Parameters ---------- images : Any One of the following formats (and types): 1. A single image path (str) 2. A list of image paths (List[str]) 3. A single image (np.ndarray, torch.Tensor) 4. A list of images (List[np.ndarray, torch.Tensor]) 5. A batched tensor of images (np.ndarray, torch.Tensor) Returns ------- A 4-dimensional, preprocessed `torch.Tensor`. """ images = ClassificationModel._expand_input_images(images) return torch.stack( [ClassificationModel._preprocess_image( image, **kwargs) for image in images], dim = 0) @torch.no_grad() def predict(self, images, **kwargs): """Runs `EfficientNetB4` inference on the input image(s). This method is the primary inference method for the model; it accepts a set of input images (see `preprocess_input()` for a detailed specification on the allowed input parameters), then preprocesses them to the model specifications, forward passes them through the model, and finally returns the predictions. In essence, this method is a wrapper for `forward()` that allows for passing a variety of inputs. If, on the other hand, you have pre-processed inputs and simply want to forward pass through the model without having to spend computational time on what is now unnecessary preprocessing, you can simply call `forward()` and then run `torch.argmax()` on the outputs to get predictions. Parameters ---------- images : Any See `preprocess_input()` for the allowed input images. Returns ------- A `np.ndarray` with integer labels for each image. """ images = self.preprocess_input(images, **kwargs) out = self.forward(images) if not self._regression: # standard classification out = torch.argmax(out, 1) return self._to_out(torch.squeeze(out)) def evaluate(self, loader, **kwargs): """Runs an accuracy evaluation on the given loader. This method will loop over the `AgMLDataLoader` and compute accuracy. Parameters ---------- loader : AgMLDataLoader A semantic segmentation loader with the dataset you want to evaluate. Returns ------- The final calculated accuracy. """ # Construct the metric and run the calculations. acc = Accuracy() bar = tqdm(loader, desc = "Calculating Accuracy") for sample in bar: image, truth = sample pred_label = self.predict(image, **kwargs) acc.update([pred_label], [truth]) bar.set_postfix({'accuracy': acc.compute().numpy().item()}) # Compute the final accuracy. return acc.compute().numpy().item()
8,921
40.887324
81
py
AgML
AgML-main/agml/models/losses.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ A collection of useful loss functions for agricultural ML tasks. Some of these have been put into use for benchmarking (see the `training` directory for examples of usage in training scripts). """ import torch import torch.nn as nn def dice_loss(y_pred, y): """Calculates `dice loss` for semantic segmentation. See https://arxiv.org/abs/1707.03237 for an in-depth explanation. """ # Convert ground truth to float for compatibility in operations. y = y.float() # Determine whether this is a multi-class or binary task. try: # Multi-class segmentation c, h, w = y.shape[1:] except: # Binary segmentation h, w = y.shape[1:]; c = 1 # noqa # Sigmoid for the outputs (since this is automatically done by binary # cross-entropy loss with logits, the actual base model for semantic # segmentation does not include any sigmoid activation in it, only the # `predict()` function wraps it. So, we do it here as well. y_pred = torch.sigmoid(y_pred) # Run the dice loss calculations. pred_flat = torch.reshape(y_pred, [-1, c * h * w]) y_flat = torch.reshape(y, [-1, c * h * w]) intersection = 2.0 * torch.sum(pred_flat * y_flat, dim = 1) + 1e-6 denominator = torch.sum(pred_flat, dim = 1) \ + torch.sum(y_flat, dim = 1) + 1e-6 return 1. - torch.mean(intersection / denominator) class DiceLoss(nn.Module): def forward(self, x, target, **kwargs): # noqa return dice_loss(x, target)
2,104
34.677966
74
py
AgML
AgML-main/agml/models/preprocessing.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ A set of data preprocessing functions for `AgMLDataLoaders`. """ import inspect import torch import numpy as np from PIL import Image import albumentations as A from albumentations.pytorch import ToTensorV2 from agml.models.tools import imagenet_style_process as _isp def imagenet_preprocess(image, size = None): """Preprocesses a single input image to ImageNet standards. The preprocessing steps are applied logically; if the images are passed with preprocessing already having been applied, for instance, the images are already resized or they are already been normalized, the operation is not applied again, for efficiency. Preprocessing includes the following steps: 1. Resizing the image to size (224, 224). 2. Performing normalization with ImageNet parameters. 3. Converting the image into a PyTorch tensor format. as well as other intermediate steps such as adding a channel dimension for two-channel inputs, for example. """ return _isp(image, size = size) class EfficientDetPreprocessor(object): """A preprocessor which prepares a data sample for `EfficientDet`. This class can be used to construct a preprocessing pipeline which auto-formats the data in an `AgMLDataLoader` for object detection into the format necessary for training an `EfficientDet` model. By default, this includes resizing images, converting bounding boxes to `yxyx`, and finally preparing the image and annotation for PyTorch. Using this class supersedes the need for any other transformations or even image resizing. It can be used as follows: > loader = agml.data.AgMLDataLoader('grape_detection_californiaday') > processor = agml.models.EfficientDetPreprocessor( > image_size = 512, augmentation = [A.HorizontalFlip(0.5)]) > loader.transform(dual_transform = processor) Parameters ---------- image_size : int The size to which images will be resized (default of 512). augmentation : Any Either a list of albumentations transforms (without being wrapped into a Compose object), or a custom method which accepts three arguments: `image`, `bboxes`, and `labels`, and returns the same. The `bboxes` will be in the XYXY format. Notes ----- - Passing `augmentation = None`, the default, is equivalent to preparing a validation or test loader: preprocessing is applied, but no transforms. - Note that if you pass a custom augmentation method, the resulting output is expected to be in PyTorch's format (image should be a tensor with its first dimension being the image's channels, for example). """ def __init__(self, image_size = 512, augmentation = None): # Parse the image size. if isinstance(image_size, int): image_size = (image_size, image_size) elif isinstance(image_size, (list, tuple)): if not len(image_size) == 2: raise ValueError( "Expected either an integer or sequence of 2 values " "for `image_size`, instead got ({image_size}).") # Construct the applied input augmentation. self._albumentations_params = dict( bbox_params = A.BboxParams( format = "pascal_voc", min_area = 0, min_visibility = 0, label_fields = ["labels"]), standard_augmentations = [ A.Resize(height = image_size[0], width = image_size[1], p = 1), ToTensorV2(p = 1)]) self._check_and_make_augmentation(augmentation) def _check_and_make_augmentation(self, augmentation): """Constructs the applied augmentation from the inputs.""" # If no augmentation is provided, then use the defaults. if augmentation is None: self._augmentation = A.Compose([ *self._albumentations_params['standard_augmentations'] ], p = 1.0, bbox_params = self._albumentations_params['bbox_params']) # If a list of augmentations are provided, then use those augmentations # wrapped alongside the default ones in a `Compose` object. elif isinstance(augmentation, list): if not all(isinstance(a, A.BasicTransform) for a in augmentation): raise ValueError( "If providing a list of transforms, all of them must be " f"albumentations augmentations, instead got {augmentation} " f"of types {[type(i) for i in augmentation]}.") self._augmentation = A.Compose([ *augmentation, *self._albumentations_params['standard_augmentations'] ], p = 1.0, bbox_params = self._albumentations_params['bbox_params']) # Otherwise, the augmentation should be a method with three input # arguments, so check it and then wrap it into an application method. else: if not len(inspect.signature(augmentation).parameters): raise ValueError( f"The input augmentation should have three input arguments, " f"instead got {inspect.signature(augmentation).parameters}.") self._method_augmentation = augmentation self._augmentation = self._apply_method_augmentation def _apply_method_augmentation(self, image, bboxes, labels): # Wrapper method to apply a user-provided method augmentation. image, bboxes, labels = self._method_augmentation(image, bboxes, labels) return {'image': image, 'bboxes': bboxes, 'labels': labels} def __call__(self, image, annotation): # Convert the image type. image = Image.fromarray(image) # Clip the bounding boxes to the image shape to prevent errors. bboxes = np.array(annotation['bbox']).astype(np.int32) x_min = bboxes[:, 0] y_min = bboxes[:, 1] x_max = bboxes[:, 2] + x_min y_max = bboxes[:, 3] + y_min x_min, y_min = np.clip(x_min, 0, image.width), \ np.clip(y_min, 0, image.height) x_max, y_max = np.clip(x_max, 0, image.width), \ np.clip(y_max, 0, image.height) # Reconstruct the boxes and get the class labels. bboxes = np.dstack((x_min, y_min, x_max, y_max)).squeeze(axis = 0) class_labels = np.array(annotation['category_id']).squeeze() # Add an extra dimension to labels for consistency. if class_labels.ndim == 0: class_labels = np.expand_dims(class_labels, axis = 0) # Add an albumentations augmentation. sample = {'image': np.array(image, dtype = np.float32), 'bboxes': bboxes, 'labels': class_labels} sample = self._augmentation(**sample) image = sample['image'] bboxes = np.array(sample['bboxes']) labels = sample['labels'] # Convert 1-channel and 4-channel to 3-channel. if image.shape[0] == 1: image = torch.tile(image, (3, 1, 1)) if image.shape[0] == 4: image = image[:3] # Convert to yxyx from xyxy. _, new_h, new_w = image.shape if bboxes.ndim == 1: bboxes = np.expand_dims(bboxes, axis = 0) bboxes[:, [0, 1, 2, 3]] = bboxes[:, [1, 0, 3, 2]] # Create the target from the annotations. target = { "bboxes": torch.as_tensor( bboxes, dtype = torch.float32), "labels": torch.as_tensor(labels), "img_size": torch.tensor([new_h, new_w]), "img_scale": torch.tensor([1.0])} return image, target
8,321
41.676923
85
py
AgML
AgML-main/agml/models/segmentation.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np from tqdm import tqdm import torch import torch.nn as nn import torch.nn.functional as F try: from torchvision.models.segmentation import deeplabv3_resnet50 except ImportError: raise ImportError("To use image classification models in `agml.models`, you " "need to install Torchvision first. You can do this by " "running `pip install torchvision`.") from agml.models.base import AgMLModelBase from agml.models.benchmarks import BenchmarkMetadata from agml.models.tools import auto_move_data, imagenet_style_process from agml.data.public import source from agml.utils.general import resolve_list_value from agml.utils.image import resolve_image_size from agml.viz.masks import show_image_and_overlaid_mask, show_image_and_mask # This is last since `agml.models.base` will check for PyTorch Lightning, # and PyTorch Lightning automatically installed torchmetrics with it. from torchmetrics import JaccardIndex as IoU class DeepLabV3Transfer(nn.Module): """Wraps a DeepLabV3 model with the right number of classes.""" def __init__(self, num_classes): super(DeepLabV3Transfer, self).__init__() self.base = deeplabv3_resnet50( pretrained = False, num_classes = num_classes) def forward(self, x, **kwargs): # noqa return self.base(x)['out'] class SegmentationModel(AgMLModelBase): """Wraps a `DeepLabV3` model for agricultural semantic segmentation. When using the model for inference, you should use the `predict()` method on any set of input images. This method wraps the `forward()` call with additional steps which perform the necessary preprocessing on the inputs, including resizing, normalization, and batching, as well as additional post-processing on the outputs (such as converting one-hot labels to integer labels), which will allow for a streamlined inference pipeline. If you want to use the `forward()` method directly, so that you can just call `model(inputs)`, then make sure your inputs are properly processed. This can be done by calling `model.preprocess_input(inputs)` on the input list of images/single image and then passing that result to the model. This will also return a one-hot label feature vector instead of integer labels, in the case that you want further customization of the outputs. This model can be subclassed in order to run a full training job; the actual transfer `EfficientNetB4` model can be accessed through the parameter `net`, and you'll need to implement methods like `training_step`, `configure_optimizers`, etc. See PyTorch Lightning for more information. """ serializable = frozenset(( "net", "num_classes", "conf_thresh", "image_size")) state_override = frozenset(("net",)) def __init__(self, num_classes = 1, image_size = 512, **kwargs): # Construct the network and load in pretrained weights. super(SegmentationModel, self).__init__() # If being initialized by a subclass, then don't do any of # model construction logic (since that's already been done). if not kwargs.get('model_initialized', False): self._num_classes = num_classes self._image_size = resolve_image_size(image_size) self.net = self._construct_sub_net(num_classes) if self._num_classes == 1: conf_threshold = kwargs.get('conf_threshold', 0.2) if not 0 < conf_threshold < 1: raise ValueError( "The given confidence threshold " "must be between 0 and 1.") self._conf_thresh = conf_threshold # By default, the model starts in inference mode. self.eval() @auto_move_data def forward(self, batch): return self.net(batch) @staticmethod def _construct_sub_net(num_classes): return DeepLabV3Transfer(num_classes) @staticmethod def _preprocess_image(image, image_size, **kwargs): """Preprocesses a single input image to EfficientNet standards. The preprocessing steps are applied logically; if the images are passed with preprocessing already having been applied, for instance, the images are already resized or they are already been normalized, the operation is not applied again, for efficiency. Preprocessing includes the following steps: 1. Resizing the image to size (224, 224). 2. Performing normalization with ImageNet parameters. 3. Converting the image into a PyTorch tensor format. as well as other intermediate steps such as adding a channel dimension for two-channel inputs, for example. """ return imagenet_style_process(image, size = image_size, **kwargs) def preprocess_input(self, images, return_shapes = False, **kwargs): """Preprocesses the input image to the specification of the model. This method takes in a set of inputs and preprocesses them into the expected format for the `DeepLabV3` semantic segmentation model. There are a variety of inputs which are accepted, including images, image paths, as well as fully-processed image tensors. The inputs are expanded and standardized, then run through a preprocessing pipeline which formats them into a single tensor ready for the model. Preprocessing steps include normalization, resizing, and converting to the channels-first format used by PyTorch models. The output of this method will be a single `torch.Tensor`, which has shape [N, C, H, W], where `N` is the batch dimension. If only a single image is passed, this will have a value of 1. This method is largely beneficial when you just want to preprocess images into the specification of the model, without getting the output. Namely, `predict()` is essentially just a wrapper around this method and `forward()`, so you can run this externally and then run `forward()` to get the original model outputs, without any extra post-processing. Parameters ---------- images : Any One of the following formats (and types): 1. A single image path (str) 2. A list of image paths (List[str]) 3. A single image (np.ndarray, torch.Tensor) 4. A list of images (List[np.ndarray, torch.Tensor]) 5. A batched tensor of images (np.ndarray, torch.Tensor) return_shapes : bool Whether to return the original shapes of the input images. Returns ------- A 4-dimensional, preprocessed `torch.Tensor`. If `return_shapes` is set to True, it also returns the original shapes of the images. """ images = self._expand_input_images(images) shapes = self._get_shapes(images) images = torch.stack( [self._preprocess_image( image, self._image_size, **kwargs) for image in images], dim = 0) if return_shapes: return images, shapes return images @torch.no_grad() def predict(self, images, **kwargs): """Runs `DeepLabV3` inference on the input image(s). This method is the primary inference method for the model; it accepts a set of input images (see `preprocess_input()` for a detailed specification on the allowed input parameters), then preprocesses them to the model specifications, forward passes them through the model, and finally returns the predictions. In essence, this method is a wrapper for `forward()` that allows for passing a variety of inputs. If, on the other hand, you have pre-processed inputs and simply want to forward pass through the model without having to spend computational time on what is now unnecessary preprocessing, you can simply call `forward()` and then run `torch.argmax()` on the outputs to get predictions. Parameters ---------- images : Any See `preprocess_input()` for the allowed input images. Returns ------- A list of `np.ndarray`s with resized output masks. """ # Process the images and run inference. images, shapes = self.preprocess_input(images, return_shapes = True, **kwargs) out = torch.sigmoid(self.forward(images)) # Post-process the output masks to a valid format. if out.shape[1] == 1: # binary class predictions out[out >= self._conf_thresh] = 1 out[out != 1] = 0 out = torch.squeeze(out, dim = 1) else: # multi-class predictions to integer labels out = torch.argmax(out, 1) # Resize the masks to their original shapes. masks = [] for mask, shape in zip( torch.index_select(out, 0, torch.arange(len(out))), shapes): masks.append(self._to_out(torch.squeeze(F.interpolate( torch.unsqueeze(torch.unsqueeze(mask, 0), 0).float(), size = shape).int()))) return resolve_list_value(masks) def show_prediction(self, image, overlay = False, **kwargs): """Shows the output predictions for one input image. This method is useful for instantly visualizing the predictions for a single input image. It accepts a single input image (or any type of valid 'image' input, as described in the method `preprocess_input()`), and then runs inference on that input image and displays its predictions in a matplotlib window. Parameters ---------- image : Any See `preprocess_input()` for allowed types of inputs. method : str Either `True` for overlaid masks, or `False` for visualizing the mask separately from the image. kwargs Visualization keyword arguments. Returns ------- The matplotlib figure containing the image. """ image = self._expand_input_images(image)[0] mask = self.predict(image, **kwargs) if overlay: return show_image_and_overlaid_mask(image, mask, **kwargs) return show_image_and_mask(image, mask, **kwargs) def load_benchmark(self, dataset): """Loads a benchmark for the given semantic segmentation dataset. This method is used to load pretrained weights for a specific AgML dataset. In essence, it serves as a wrapper for `load_state_dict`, directly getting the model from its save path in the AWS storage bucket. You can then use the `benchmark` property to access the metric value of the benchmark, as well as additional training parameters which you can use to train your own models. Parameters ---------- dataset : str The name of the semantic segmentation benchmark to load. Notes ----- If the benchmark has a different number of classes than this input model, then this method will raise an error. This issue may be adapted in the future. """ if source(dataset).tasks.ml != 'semantic_segmentation': raise ValueError( f"You are trying to load a benchmark for a " f"{source(dataset).tasks.ml} task ({dataset}) " f"in a semantic segmentation model.") # Number of classes must be the same for semantic segmentation. if source(dataset).num_classes != self._num_classes: raise ValueError( f"You cannot load a benchmark for a dataset '{dataset}' " f"with {source(dataset).num_classes} classes, while your " f"model has {self._num_classes} classes.") # Load the benchmark. state = self._get_benchmark(dataset) self.load_state_dict(state) self._benchmark = BenchmarkMetadata(dataset) def evaluate(self, loader, **kwargs): """Runs a mean intersection over union evaluation on the given loader. This method will loop over the provided `AgMLDataLoader` and compute the mean intersection over union (mIOU). Parameters ---------- loader : AgMLDataLoader A semantic segmentation loader with the dataset you want to evaluate. Returns ------- The final calculated mIoU. """ # Construct the metric and run the calculations. iou = IoU(num_classes = self._num_classes + 1) bar = tqdm(loader, desc = "Calculating Mean Intersection Over Union") for sample in bar: image, truth = sample pred_mask = self.predict(image, **kwargs) if pred_mask.ndim == 3: pred_mask = np.transpose(pred_mask, (2, 0, 1)) truth = np.transpose(truth, (2, 0, 1)) iou(torch.from_numpy(pred_mask).int().unsqueeze(0), torch.from_numpy(truth).unsqueeze(0)) bar.set_postfix({'miou': iou.compute().numpy().item()}) # Compute the final mIoU. return iou.compute().numpy().item()
14,006
42.231481
86
py
AgML
AgML-main/agml/models/tools.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from typing import Callable from functools import wraps import torch import numpy as np import albumentations as A from agml.backend.tftorch import is_array_like def imagenet_style_process(image, size = None, **kwargs): """Preprocesses a single input image to ImageNet standards. The preprocessing steps are applied logically; if the images are passed with preprocessing already having been applied, for instance, the images are already resized or they are already been normalized, the operation is not applied again, for efficiency. Preprocessing includes the following steps: 1. Resizing the image to size (224, 224). 2. Performing normalization with ImageNet parameters. 3. Converting the image into a PyTorch tensor format. as well as other intermediate steps such as adding a channel dimension for two-channel inputs, for example. """ # Convert the image to a NumPy array. if is_array_like(image) and hasattr(image, 'numpy'): image = image.numpy() # Add a channel dimension for grayscale imagery. if image.ndim == 2: image = np.expand_dims(image, axis = -1) # If the image is already in channels-first format, convert # it back temporarily until preprocessing has concluded. if image.shape[0] <= 3: image = np.transpose(image, (1, 2, 0)) # Resize the image to ImageNet standards. h = w = 224 if size is not None: h, w = size rz = A.Resize(height = h, width = w) if image.shape[0] != h or image.shape[1] != w: image = rz(image = image)['image'] # Normalize the image to ImageNet standards. if kwargs.get('normalize', True): if 1 <= image.max() <= 255: mean = [0.485, 0.456, 0.406] std = [0.229, 0.224, 0.225] image = image.astype(np.float32) / 255. mean = np.array(mean, dtype = np.float32) std = np.array(std, dtype = np.float32) denominator = np.reciprocal(std, dtype = np.float32) image = (image - mean) * denominator else: # Otherwise, just scale the image. if 1 <= image.max() <= 255: image = image.astype(np.float32) / 255.0 # Convert the image into a PyTorch tensor. image = torch.from_numpy(image).permute(2, 0, 1) # Return the processed image. return image # Ported from PyTorch Lightning v1.3.0. def auto_move_data(fn: Callable) -> Callable: """ Decorator for :class:`~pytorch_lightning.core.lightning.LightningModule` methods for which input arguments should be moved automatically to the correct device. It has no effect if applied to a method of an object that is not an instance of :class:`~pytorch_lightning.core.lightning.LightningModule` and is typically applied to ``__call__`` or ``forward``. Args: fn: A LightningModule method for which the arguments should be moved to the device the parameters are on. Example:: # directly in the source code class LitModel(LightningModule): @auto_move_data def forward(self, x): return x # or outside LitModel.forward = auto_move_data(LitModel.forward) model = LitModel() model = model.to('cuda') model(torch.zeros(1, 3)) # input gets moved to device # tensor([[0., 0., 0.]], device='cuda:0') """ @wraps(fn) def auto_transfer_args(self, *args, **kwargs): from pytorch_lightning import LightningModule if not isinstance(self, LightningModule): return fn(self, *args, **kwargs) args, kwargs = self.transfer_batch_to_device( (args, kwargs), device = self.device, dataloader_idx = None) # noqa return fn(self, *args, **kwargs) return auto_transfer_args
4,452
32.734848
103
py
AgML
AgML-main/agml/models/__init__.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ This module contains PyTorch pre-trained weights and benchmarks for commonly used deep learning models on agricultural datasets within AgML. """ # Before anything can be imported, we need to run checks for PyTorch and # PyTorch Lightning, as these are not imported on their own. try: import torch except ImportError: raise ImportError('Could not find a PyTorch installation. If you want to use ' 'the models in `agml.models`, you will need to install PyTorch ' 'first. Try `pip install torch` to do so.') try: import pytorch_lightning except ImportError: raise ImportError('Could not find a PyTorch Lightning installation. If you want to ' 'use the models in `agml.models`, you will need to install PyTorch ' 'Lightning first. Try `pip install pytorch-lightning` to do so.') from .classification import ClassificationModel from .segmentation import SegmentationModel from .detection import DetectionModel from . import metrics from . import losses from . import preprocessing
1,690
37.431818
90
py
AgML
AgML-main/agml/models/detection.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import warnings import torch import numpy as np import albumentations as A from tqdm import tqdm try: from ensemble_boxes.ensemble_boxes_wbf import weighted_boxes_fusion except ImportError: raise ImportError( "Could not find an installation of the `ensemble_boxes` package. " "Try `pip install ensemble-boxes` to install it." ) try: from effdet import ( create_model_from_config, get_efficientdet_config, DetBenchPredict, DetBenchTrain ) except ImportError: raise ImportError( "Could not find an installation of the `effdet` package. " "Try `pip install effdet==0.2.4` to install it (note that " "the version is important for proper functionality)." ) from agml.models.base import AgMLModelBase from agml.models.benchmarks import BenchmarkMetadata from agml.models.tools import auto_move_data from agml.models.metrics.map import MeanAveragePrecision from agml.data.public import source from agml.backend.tftorch import is_array_like from agml.utils.image import resolve_image_size from agml.utils.logging import log from agml.viz.boxes import show_image_and_boxes class DetectionModel(AgMLModelBase): """Wraps an `EfficientDetD4` model for agricultural object detection. When using the model for inference, you should use the `predict()` method on any set of input images. This method wraps the `forward()` call with additional steps which perform the necessary preprocessing on the inputs, including resizing, normalization, and batching, as well as additional post-processing on the outputs (such as converting one-hot labels to integer labels), which will allow for a streamlined inference pipeline. If you want to use the `forward()` method directly, so that you can just call `model(inputs)`, then make sure your inputs are properly processed. This can be done by calling `model.preprocess_input(inputs)` on the input list of images/single image and then passing that result to the model. This will also return a one-hot label feature vector instead of integer labels, in the case that you want further customization of the outputs. By default, when instantiating a `DetectionModel`, it is prepared in inference mode. In order to use this model for training, you need to convert it to training mode by using `DetectionModel.switch_train()`. To convert back to inference mode, use `DetectionModel.switch_predict()`. If you want to use your own custom model and/or training pipeline, without the existing input restrictions, then you can subclass this model. In the `super` call in the `__init__` method, pass the parameter `model_initialized = True`, which will enable you to initialize the model in your own format. Parameters ---------- num_classes : int The number of classes for the `EfficientDet` model. image_size : int, tuple The shape of image inputs to the model. conf_threshold : float Filters bounding boxes by their level of confidence based on this threshold. """ serializable = frozenset(( "model", "num_classes", "conf_thresh", "image_size")) state_override = frozenset(("model",)) def __init__(self, num_classes = 1, image_size = 512, conf_threshold = 0.3, **kwargs): # Initialize the base modules. super(DetectionModel, self).__init__() # If being initialized by a subclass, then don't do any of # model construction logic (since that's already been done). if not kwargs.get('model_initialized', False): # Construct the network and load in pretrained weights. self._image_size = resolve_image_size(image_size) self._confidence_threshold = conf_threshold self._num_classes = num_classes self.model = self._construct_sub_net( self._num_classes, self._image_size) # Filter out unnecessary warnings. warnings.filterwarnings( 'ignore', category = UserWarning, module = 'ensemble_boxes') warnings.filterwarnings( 'ignore', category = UserWarning, module = 'effdet.bench') # noqa @auto_move_data def forward(self, batch): """Ensures that the input is valid for the model.""" return self.model(batch) @staticmethod def _construct_sub_net(num_classes, image_size): cfg = get_efficientdet_config('tf_efficientdet_d4') cfg.update({"image_size": image_size}) model = create_model_from_config( cfg, pretrained = False, num_classes = num_classes) return DetBenchPredict(model) def switch_predict(self): """Prepares the model for evaluation mode.""" state = self.model.state_dict() if not isinstance(self.model, DetBenchPredict): self.model = DetBenchPredict(self.model.model) self.model.load_state_dict(state) def switch_train(self): """Prepares the model for training mode.""" state = self.model.state_dict() if not isinstance(self.model, DetBenchTrain): self.model = DetBenchTrain(self.model.model) self.model.load_state_dict(state) @property def original(self): # override for detection models. return self.model.model @torch.jit.ignore() def reset_class_net(self, num_classes = 1): """Reconfigures the output class net for a new number of classes. Parameters ---------- num_classes : int The number of classes to reconfigure the output net to use. """ if num_classes != self._num_classes: self.model.model.reset_head(num_classes = num_classes) @staticmethod def _preprocess_image(image, image_size): """Preprocesses a single input image to EfficientNet standards. The preprocessing steps are applied logically; if the images are passed with preprocessing already having been applied, for instance, the images are already resized or they are already been normalized, the operation is not applied again, for efficiency. Preprocessing includes the following steps: 1. Resizing the image to size (224, 224). 2. Performing normalization with ImageNet parameters. 3. Converting the image into a PyTorch tensor format. as well as other intermediate steps such as adding a channel dimension for two-channel inputs, for example. """ # Convert the image to a NumPy array. if is_array_like(image) and hasattr(image, 'numpy'): image = image.numpy() # Add a channel dimension for grayscale imagery. if image.ndim == 2: image = np.expand_dims(image, axis = -1) # If the image is already in channels-first format, convert # it back temporarily until preprocessing has concluded. if image.shape[0] <= 3: image = np.transpose(image, (1, 2, 0)) # Resize the image to ImageNet standards. (w, h) = image_size rz = A.Resize(height = h, width = w) if image.shape[0] != h or image.shape[1] != w: image = rz(image = image)['image'] # Normalize the image to ImageNet standards. if 1 <= image.max() <= 255: image = image.astype(np.float32) / 255. # Convert the image into a PyTorch tensor. image = torch.from_numpy(image).permute(2, 0, 1) # Return the processed image. return image @torch.jit.ignore() def preprocess_input(self, images, return_shapes = False): """Preprocesses the input image to the specification of the model. This method takes in a set of inputs and preprocesses them into the expected format for the `EfficientDetD4` object detection model. There are a variety of inputs which are accepted, including images, image paths, as well as fully-processed image tensors. The inputs are expanded and standardized, then run through a preprocessing pipeline which formats them into a single tensor ready for the model. Preprocessing steps include normalization, resizing, and converting to the channels-first format used by PyTorch models. The output of this method will be a single `torch.Tensor`, which has shape [N, C, H, W], where `N` is the batch dimension. If only a single image is passed, this will have a value of 1. This method is largely beneficial when you just want to preprocess images into the specification of the model, without getting the output. Namely, `predict()` is essentially just a wrapper around this method and `forward()`, so you can run this externally and then run `forward()` to get the original model outputs, without any extra post-processing. Parameters ---------- images : Any One of the following formats (and types): 1. A single image path (str) 2. A list of image paths (List[str]) 3. A single image (np.ndarray, torch.Tensor) 4. A list of images (List[np.ndarray, torch.Tensor]) 5. A batched tensor of images (np.ndarray, torch.Tensor) return_shapes : bool Whether to return the original shapes of the input images. Returns ------- A 4-dimensional, preprocessed `torch.Tensor`. If `return_shapes` is set to True, it also returns the original shapes of the images. """ images = self._expand_input_images(images) shapes = self._get_shapes(images) images = torch.stack( [self._preprocess_image( image, self._image_size) for image in images], dim = 0) if return_shapes: return images, shapes return images def _process_detections(self, detections): """Post-processes the output detections (boxes, labels) from the model.""" predictions = [] # Convert all of the output detections into predictions. This involves # selecting the bounding boxes, confidence scores, and classes, and then # dropping any of them which do not have a confidence score about the # threshold as determined when the class is initialized. for d in detections: # Extract the bounding boxes, confidence scores, # and class labels from the output detections. boxes, scores, classes = d[:, :4], d[:, 4], d[:, 5] # Only return boxes which are above the confidence threshold. valid_indexes = np.where(scores > self._confidence_threshold)[0] boxes = boxes[valid_indexes] scores = scores[valid_indexes] classes = classes[valid_indexes] predictions.append({"boxes": boxes, "scores": scores, "classes": classes}) # Run weighted boxes fusion. For an exact description of how this # works, see the paper: https://arxiv.org/pdf/1910.13302.pdf. (predicted_bboxes, predicted_class_confidences, predicted_class_labels) = self._wbf(predictions) return predicted_bboxes, predicted_class_confidences, predicted_class_labels @staticmethod def _rescale_bboxes(predicted_bboxes, image_sizes): """Re-scales output bounding boxes to the original image sizes.""" scaled_boxes = [] for bboxes, img_dims in zip(predicted_bboxes, image_sizes): h, w = img_dims if len(bboxes) > 0: # Re-scale the bounding box to the appropriate format. scale_ratio = [w / 512, h / 512, w / 512, h / 512] scaled = (np.array(bboxes) * scale_ratio).astype(np.int32) # Convert the Pascal-VOC (xyxy) format to COCO (xywh). x, y = scaled[:, 0], scaled[:, 1] w, h = scaled[:, 2] - x, scaled[:, 3] - y scaled_boxes.append(np.dstack((x, y, w, h))) continue # Otherwise, there is no prediction for this image. scaled_boxes.append(np.array([])) return scaled_boxes @staticmethod def _wbf(predictions): """Runs weighted boxes fusion on the output predictions.""" bboxes, confidences, class_labels = [], [], [] # Fuse the predictions for each in the batch. for prediction in predictions: boxes = [(prediction["boxes"] / 512).tolist()] scores = [prediction["scores"].tolist()] labels = [prediction["classes"].tolist()] # Run the actual fusion and update the containers. boxes, scores, labels = weighted_boxes_fusion( boxes, scores, labels, iou_thr = 0.44, skip_box_thr = 0.43) boxes = boxes * (512 - 1) bboxes.append(boxes) confidences.append(scores) class_labels.append(labels.astype(np.int32)) return bboxes, confidences, class_labels @staticmethod def _remap_outputs(boxes, labels, confidences): """Remaps the outputs to the format described in `predict()`.""" squeeze = lambda *args: tuple(list( np.squeeze(a).tolist() for a in args)) return [squeeze(b, l, c) for b, l, c in zip(boxes, labels, confidences)] def _to_out(self, tensor: "torch.Tensor") -> "torch.Tensor": if isinstance(tensor, dict): tensor = tensor['detections'] return super()._to_out(tensor) @torch.no_grad() def predict(self, images): """Runs `EfficientNetD4` inference on the input image(s). This method is the primary inference method for the model; it accepts a set of input images (see `preprocess_input()` for a detailed specification on the allowed input parameters), then preprocesses them to the model specifications, forward passes them through the model, and finally returns the predictions. In essence, this method is a wrapper for `forward()` that allows for passing a variety of inputs. If, on the other hand, you have pre-processed inputs and simply want to forward pass through the model without having to spend computational time on what is now unnecessary preprocessing, you can simply call `forward()` and then run the post-processing as described in this method. Parameters ---------- images : Any See `preprocess_input()` for the allowed input images. Returns ------- A tuple of `n` lists, where `n` is the number of input images. Each of the `n` lists will contain three values consisting of the bounding boxes, class labels, and prediction confidences for the corresponding input image. """ # Process the images and run inference. images, shapes = self.preprocess_input(images, return_shapes = True) out = self._to_out(self.forward(images)) # Post-process the output detections. boxes, confidences, labels = self._process_detections(out) boxes = self._rescale_bboxes(boxes, shapes) ret = self._remap_outputs(boxes, labels, confidences) return ret[0] if len(ret) == 1 else ret def show_prediction(self, image): """Shows the output predictions for one input image. This method is useful for instantly visualizing the predictions for a single input image. It accepts a single input image (or any type of valid 'image' input, as described in the method `preprocess_input()`), and then runs inference on that input image and displays its predictions in a matplotlib window. Parameters ---------- image : Any See `preprocess_input()` for allowed types of inputs. Returns ------- The matplotlib figure containing the image. """ image = self._expand_input_images(image)[0] bboxes, labels, _ = self.predict(image) if isinstance(labels, int): bboxes, labels = [bboxes], [labels] return show_image_and_boxes(image, bboxes, labels) def load_benchmark(self, dataset, strict = False): """Loads a benchmark for the given semantic segmentation dataset. This method is used to load pretrained weights for a specific AgML dataset. In essence, it serves as a wrapper for `load_state_dict`, directly getting the model from its save path in the AWS storage bucket. You can then use the `benchmark` property to access the metric value of the benchmark, as well as additional training parameters which you can use to train your own models. Parameters ---------- dataset : str The name of the object detection benchmark to load. strict : bool Whether to require the same number of classes. Notes ----- If the given benchmark has a different number of classes than this input model, then the class network will be loaded with random weights, while the remaining network weights (backbone, box network, etc.) will use the pretrained weights for the benchmark. This can be disabled by setting `strict = True`. """ if source(dataset).tasks.ml != 'object_detection': raise ValueError( f"You are trying to load a benchmark for a " f"{source(dataset).tasks.ml} task ({dataset}) " f"in an object detection model.") # Check loading strictness. cs = source(dataset).num_classes == self._num_classes if strict: if not cs: raise ValueError( f"You cannot load a benchmark for a dataset '{dataset}' " f"with {source(dataset).num_classes} classes, while your " f"model has {self._num_classes} classes. If you want to, " f"then you need to set `strict = False`.") # Load the benchmark. state = self._get_benchmark(dataset) if not strict and not cs: self.reset_class_net(source(dataset).num_classes) log(f"Loading a state dict for {dataset} with " f"{source(dataset).num_classes}, while your " f"model has {self._num_classes} classes. The " f"class network will use random weights.") self.load_state_dict(state) if not strict and not cs: self.reset_class_net(self._num_classes) self._benchmark = BenchmarkMetadata(dataset) def evaluate(self, loader, iou_threshold = 0.5, method = 'accumulate'): """Runs a mean average precision evaluation on the given loader. This method will loop over the provided `AgMLDataLoader` and compute the mean average precision at the provided `iou_threshold`. This can be done using two methods. Using the method `average` will compute the mean average precision for each individual sample and then average over all of the samples, while using the method `accumulate` will compute the mean average precision over the entire dataset. Parameters ---------- loader : AgMLDataLoader An object detection loader with the dataset you want to evaluate. iou_threshold : float The IoU threshold between a ground truth and predicted bounding box at which point they are considered the same. method : str The method to use, as described above. Returns ------- The final calculated mean average precision. """ if not 0 < iou_threshold < 1: raise ValueError( f"The `iou_threshold` must be between 0 and 1, got {iou_threshold}.") if method not in ['accumulate', 'average']: raise ValueError( f"Method must be either `accumulate` or `average`, got {method}.") # Construct the mean average precision accumulator and run the calculations. mean_ap = MeanAveragePrecision( num_classes = self._num_classes, iou_threshold = iou_threshold) bar = tqdm(loader, desc = "Calculating Mean Average Precision") if method == 'average': cumulative_maps = [] for sample in bar: image, truth = sample true_box, true_label = truth['bbox'], truth['category_id'] bboxes, labels, conf = self.predict(image) mean_ap.update(*( dict(boxes = bboxes, labels = labels, scores = conf), dict(boxes = true_box, labels = true_label))) # If averaging, then calculate and reset the mAP, otherwise continue. if method == 'average': res = mean_ap.compute() cumulative_maps.append(res) # noqa bar.set_postfix({'map': float(res)}) mean_ap.reset() # Compute the final mAP. if method == 'average': result = sum(cumulative_maps) / len(cumulative_maps) else: result = mean_ap.compute() return result
22,088
41.397313
87
py
AgML
AgML-main/agml/models/metrics/accuracy.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import torch import torch.nn as nn @torch.jit.script def accuracy(output, target): """Computes the accuracy between `output` and `target`.""" batch_size = target.size(0) _, pred = torch.topk(output, 1, 1) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) correct_k = correct[:1].reshape(-1).float().sum(0, keepdim = True) return correct_k.mul_(100.0 / batch_size) class Accuracy(nn.Module): """A metric to compute accuracy for image classification tasks. This class is used as a wrapper around accuracy calculations, which allows for accumulation of predictions over time. The `update` method can be used to update data, then `compute` to get the calculated accuracy, and finally `reset` can be used to reset the accumulators to an empty state, allowing new calculations. """ def __init__(self): # Construct the data accumulators. super(Accuracy, self).__init__() self._prediction_data, self._truth_data = [], [] def update(self, pred_data, gt_data): """Updates the state of the accuracy metric. The `pred_data` and `gt_data` arguments should both be sequences of integer labels (e.g., an array of one-dimension), *not* one-hot labels. So, any activation operations or softmax operations must be applied before inputting data into the accuracy metric. """ if not len(pred_data) == len(gt_data): raise ValueError("Predictions and truths should be the same length.") self._prediction_data.extend(pred_data) self._truth_data.extend(gt_data) def compute(self): """Computes the accuracy between the predictions and ground truths.""" return accuracy(torch.tensor(self._prediction_data), torch.tensor(self._truth_data)) def reset(self): """Resets the accumulator states.""" del self._prediction_data, self._truth_data self._prediction_data, self._truth_data = [], []
2,632
38.298507
84
py
AgML
AgML-main/agml/models/metrics/map.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import torch import torch.nn as nn from agml.utils.logging import log @torch.jit.script def bbox_iou(predicted_box, truth_box): """Calculates the IOU of predicted and truth bounding boxes.""" # Calculate the coordinates of the bounding box overlap. box1_x1 = predicted_box[..., 0:1] box1_y1 = predicted_box[..., 1:2] box1_x2 = predicted_box[..., 0:1] + predicted_box[..., 2:3] box1_y2 = predicted_box[..., 1:2] + predicted_box[..., 3:4] box2_x1 = truth_box[..., 0:1] box2_y1 = truth_box[..., 1:2] box2_x2 = truth_box[..., 0:1] + truth_box[..., 2:3] box2_y2 = truth_box[..., 1:2] + truth_box[..., 3:4] x1 = torch.max(box1_x1, box2_x1) y1 = torch.max(box1_y1, box2_y1) x2 = torch.min(box1_x2, box2_x2) y2 = torch.min(box1_y2, box2_y2) # Get the area of the union. intersection = (x2 - x1).clamp(0) * (y2 - y1).clamp(0) box1_area = abs((box1_x2 - box1_x1) * (box1_y2 - box1_y1)) box2_area = abs((box2_x2 - box2_x1) * (box2_y2 - box2_y1)) union_area = (box1_area + box2_area - intersection) # Return the intersection over union. return intersection / (union_area + 1e-6) @torch.no_grad() def mean_average_precision( predicted_boxes, truth_boxes, num_classes = 1, iou_thresh = 0.5): """Calculates the mean average precision for predicted and true boxes.""" average_precisions = [] # Check whether to add confidence scores to the ground truth boxes. if len(truth_boxes[0]) == 6: out_boxes = [] for a in truth_boxes: out_boxes.append([*a[:2], 1.0, *a[2:]]) truth_boxes = out_boxes.copy() # Calculate average precision for each class. pred_boxes, true_boxes = torch.tensor(predicted_boxes), torch.tensor(truth_boxes) for c in range(num_classes): # If there are no predictions, then the per-class AP is 0. if len(pred_boxes) == 0: average_precisions.append(torch.tensor(0.0)) continue # Get the predictions and targets corresponding to this class. detections = pred_boxes[torch.where(pred_boxes[:, 1] == c)[0]].tolist() ground_truths = true_boxes[torch.where(true_boxes[:, 1] == c)[0]].tolist() torch_gt = torch.tensor(ground_truths) # If there are no ground truths, then the per-class AP is 0. if len(ground_truths) == 0: average_precisions.append(torch.tensor(0.0)) continue # Determine the number of boxes for each of the training samples. numpy_gt = torch.tensor(ground_truths) amount_bboxes = {int(k.numpy().item()): torch.zeros(v) for k, v in zip( *torch.unique(numpy_gt[:, 0], return_counts = True))} # Sort the boxes by probabilities. detections.sort(key = lambda x: x[2], reverse = True) tp = torch.zeros((len(detections))) fp = torch.zeros((len(detections))) total_true_bboxes = len(ground_truths) # If there are no boxes for this class, then there are # no calculations to do, so skip it. if total_true_bboxes == 0: continue for detection_idx, detection in enumerate(detections): # Get the update number for this detection. update_num = int(detection[0]) # Only take out the ground_truths that have the same # training idx as the detection. ground_truth_img = torch_gt[ torch.where(torch_gt[:, 0] == update_num)[0]] # Get the bounding box with the highest IoU. ious = torch.tensor([bbox_iou( torch.tensor(detection[3:]), gt[3:].clone()) for gt in ground_truth_img]) best_iou = torch.max(ious) best_gt_idx = torch.argmax(ious) # If the IoU is above the threshold, then it may be a true positive.' if best_iou > iou_thresh: # This should be the first time the box is detected. Otherwise, # that would mean that there are multiple predicted bounding # boxes for the same object, which is a false positive. try: if amount_bboxes[update_num][best_gt_idx] == 0: tp[detection_idx] = 1 amount_bboxes[update_num][best_gt_idx] = 1 else: fp[detection_idx] = 1 except KeyError: # A false detection. fp[detection_idx] = 1 # If the IoU is below the threshold, then it is a false positive. else: fp[detection_idx] = 1 # Calculate the prediction/recalls and update the array. tp_cumsum = torch.cumsum(tp, dim = 0) recalls = tp_cumsum / (total_true_bboxes + 1e-6) precisions = tp_cumsum / (tp_cumsum + torch.cumsum(fp, dim = 0) + 1e-6) precisions = torch.cat((torch.tensor([1]), precisions)) recalls = torch.cat((torch.tensor([0]), recalls)) average_precisions.append(torch.trapz(precisions, recalls)) # Calculate the mean of all of the average precisions. return sum(average_precisions) / len(average_precisions) class MeanAveragePrecision(nn.Module): """A metric to calculate mean average precision for object detection tasks. This class is used as a wrapper around mean average precision calculations, which allows for accumulation of predictions over time. The `update` (and `batch_update`) methods can be used to update data in the class, then `compute` to get the calculated mean average precision, and finally `reset` to restore the accumulators to an empty state, allowing from-scratch calculations. """ def __init__(self, num_classes = 1, iou_threshold = 0.5): # Set base parameters. super(MeanAveragePrecision, self).__init__() self._num_classes = num_classes self._iou_threshold = iou_threshold # Store the truth and prediction data in containers. self._prediction_data, self._truth_data = [], [] self._num_updates = 0 @staticmethod def _scalar_to_array(*args): """Converts 0-dimensional scalar arrays to 1-d arrays.""" cvt = lambda x: np.expand_dims(x, 0) if x.ndim == 0 else x outs = [cvt(arg) for arg in args] return outs[0] if len(args) == 1 else outs def update(self, pred_data, gt_data): """Update the tracker with prediction and ground truth data. The arguments `pred_data` and `gt_data` should be either dictionaries (with the following keys), or lists of values which correspond in order to the same keys listed below: - `pred_data`: `boxes`, `labels`, and `scores`. - `gt_data`: `boxes` and `labels`. Note: To update a batch of data, use `batch_update()`. """ # Get the relevant data from the input arguments. if isinstance(pred_data, dict): pred_boxes, pred_labels, pred_scores = \ pred_data['boxes'], pred_data['labels'], pred_data['scores'] else: pred_boxes, pred_labels, pred_scores = pred_data if isinstance(gt_data, dict): gt_boxes, gt_labels = \ gt_data['boxes'], gt_data['labels'] else: gt_boxes, gt_labels = gt_data # Format the data. pred_boxes = np.squeeze(pred_boxes) pred_labels = np.squeeze(pred_labels) pred_scores = np.squeeze(pred_scores) pred_labels, gt_labels, pred_scores = \ self._scalar_to_array(pred_labels, gt_labels, pred_scores) gt_boxes = np.squeeze(gt_boxes) if pred_boxes.ndim == 1: pred_boxes = np.expand_dims(pred_boxes, axis = 0) if gt_boxes.ndim == 1: gt_boxes = np.expand_dims(gt_boxes, axis = 0) # Create the data in the proper format. for bbox, label in zip(gt_boxes, gt_labels): self._truth_data.append( [self._num_updates, int(label - 1), 1.0, *bbox]) if pred_boxes.ndim == 1: pred_boxes = np.expand_dims(pred_boxes, axis = 0) for bbox, label, score in zip(pred_boxes, pred_labels, pred_scores): self._prediction_data.append( [self._num_updates, int(label - 1), score, *bbox]) # Increment the number of updates. self._num_updates += 1 def compute(self): """Computes the mean average precision with the given data.""" if self._num_updates == 0: log("Tried to compute mean average precision " "without any data updates; returning 0.0.") return 0.0 return mean_average_precision( predicted_boxes = self._prediction_data, truth_boxes = self._truth_data, num_classes = self._num_classes, iou_thresh = self._iou_threshold, ) def reset(self): """Resets the mean average precision.""" del self._prediction_data, self._truth_data self._prediction_data, self._truth_data = [], [] self._num_updates = 0
9,795
39.147541
85
py
AgML
AgML-main/agml/backend/random.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from agml.backend.tftorch import torch, tf, get_backend def set_seed(seed = None): """Sets a new random seed. If None, uses a random seed.""" import random random.seed(seed) import numpy as np np.random.seed(seed) if get_backend() == 'torch': torch.random.manual_seed(seed) elif get_backend() == 'tensorflow': tf.random.set_seed(seed)
983
32.931034
74
py
AgML
AgML-main/agml/backend/tftorch.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ This part of the backend controls the AgML methods where either TensorFlow or PyTorch methods can be used, and prevents unnecessary importing of either library (which takes a significant amount of time). """ import types import inspect import logging import importlib import functools import numpy as np from agml.utils.logging import log from agml.utils.image import consistent_shapes # Suppress any irrelevant warnings which will pop up from either backend. import warnings warnings.filterwarnings( 'ignore', category = UserWarning, message = '.*Named tensors.*Triggered internally.*') class StrictBackendError(ValueError): def __init__(self, message = None, change = None, obj = None): if message is None: message = f"Backend was manually set to " \ f"'{get_backend()}', but got an object " \ f"from backend '{change}': {obj}." super(StrictBackendError, self).__init__(message) # Check if TensorFlow and PyTorch exist in the environment. _HAS_TENSORFLOW: bool _HAS_TORCH: bool @functools.lru_cache(maxsize = None) def _check_tf_torch(): global _HAS_TENSORFLOW, _HAS_TORCH try: import tensorflow except ImportError: _HAS_TENSORFlOW = False else: _HAS_TENSORFLOW = True try: import torch except ImportError: _HAS_TORCH = False else: _HAS_TORCH = True # Default backend is PyTorch. _BACKEND = None _USER_SET_BACKEND = False def get_backend(): """Returns the current AgML backend.""" return _BACKEND def set_backend(backend): """Change the AgML backend for the current session. By default, AgML uses PyTorch as a backend, but it is compatible with both TensorFlow and PyTorch. AgML can also automatically inference the backend from the different parameters passed into `AgMLDataLoader` and other internal library methods. This method allows a user to automatically set the backend. """ global _USER_SET_BACKEND, _BACKEND # Check whether the user has modified the backend. mod = inspect.getmodule(inspect.stack()[1][0]) if mod is None: # IPython shell _USER_SET_BACKEND = True elif 'agml.' not in mod.__name__: _USER_SET_BACKEND = True # If the backend is the same, don't do anything. if backend == _BACKEND: return _check_tf_torch() if backend not in ['tensorflow', 'tf', 'torch', 'pytorch']: raise ValueError(f"Invalid backend: {backend}.") if backend in ['tensorflow', 'tf'] and _BACKEND != 'tensorflow': if not _HAS_TENSORFLOW: raise ImportError( "TensorFlow not found on system, cannot be used as " "backend. Try running `pip install tensorflow`.") _BACKEND = 'tf' log("Switched backend to TensorFlow.", level = logging.INFO) elif backend in ['torch', 'pytorch'] and _BACKEND != 'torch': if not _HAS_TORCH: raise ImportError( "PyTorch not found on system, cannot be used as " "backend. Try running `pip install torch`.") _BACKEND = 'torch' log("Switched backend to PyTorch.", level = logging.INFO) def user_changed_backend(): """Returns whether the backend has been manually changed.""" return _USER_SET_BACKEND # Ported from https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/util/lazy_loader.py class LazyLoader(types.ModuleType): """Lazily import a module, mainly to avoid pulling in large dependencies. """ def __init__(self, local_name, parent_module_globals, name): self._local_name = local_name self._parent_module_globals = parent_module_globals super(LazyLoader, self).__init__(name) def _load(self): """Load the module and insert it into the parent's globals.""" # Import the target module and insert it into the parent's namespace. module = importlib.import_module(self.__name__) self._parent_module_globals[self._local_name] = module # Update this object's dict so that if someone keeps a reference to the # LazyLoader, lookups are efficient (__getattr__ is only called on lookups # that fail). self.__dict__.update(module.__dict__) return module def __getattr__(self, item): module = self._load() return getattr(module, item) def __dir__(self): module = self._load() return dir(module) # Load TensorFlow and PyTorch lazily to prevent pulling them in when unnecessary. torch = LazyLoader('torch', globals(), 'torch') torch_data = LazyLoader('torch_data', globals(), 'torch.utils.data') torchvision = LazyLoader('torchvision', globals(), 'torchvision') tf = LazyLoader('tensorflow', globals(), 'tensorflow') ######### GENERAL METHODS ######### def _convert_image_to_torch(image): """Converts an image (np.ndarray) to a torch Tensor.""" if isinstance(image, (list, tuple)): return torch.tensor(image) if isinstance(image, torch.Tensor) or image.ndim == 4: if image.shape[0] == 1 and image.shape[-1] <= 3 and image.ndim == 4: return torch.from_numpy(image).permute(0, 3, 1, 2).float() return image if image.shape[0] > image.shape[-1]: return torch.from_numpy(image).permute(2, 0, 1).float() return torch.from_numpy(image) def _postprocess_torch_annotation(image): """Post-processes a spatially augmented torch annotation.""" try: if image.dtype.is_floating_point: image = (image * 255).int() except AttributeError: pass return image def as_scalar(inp): """Converts an input value to a scalar.""" if isinstance(inp, (int, float)): return inp if np.isscalar(inp): return inp.item() if isinstance(inp, np.ndarray): return inp.item() if isinstance(inp, torch.Tensor): return inp.item() if isinstance(inp, tf.Tensor): return inp.numpy() raise TypeError(f"Unsupported variable type {type(inp)}.") def scalar_unpack(inp): """Unpacks a 1-d array into a list of scalars.""" return [as_scalar(item) for item in inp] def is_array_like(inp, no_list = False): """Determines if an input is a np.ndarray, torch.Tensor, or tf.Tensor.""" if isinstance(inp, (list, tuple)): # no need to import tensorflow for this if no_list: return False return True if isinstance(inp, np.ndarray): return True if isinstance(inp, torch.Tensor): return True if isinstance(inp, tf.Tensor): return True return False def convert_to_batch(images): """Converts a set of images to a batch.""" # If `images` is already an array type, nothing to do. if is_array_like(images): return images # NumPy Arrays. if isinstance(images[0], np.ndarray): if not consistent_shapes(images): images = np.array(images, dtype = object) log("Created a batch of images with different " "shapes. If you want the shapes to be consistent, " "run `loader.resize_images('auto')`.") else: images = np.array(images) return images # Torch Tensors. if isinstance(images[0], torch.Tensor): if not consistent_shapes(images): images = [image.numpy() for image in images] images = np.array(images, dtype = object) log("Created a batch of images with different " "shapes. If you want the shapes to be consistent, " "run `loader.resize_images('auto')`.") else: images = torch.stack(images) return images # TensorFlow Tensors. if isinstance(images[0], tf.Tensor): if not consistent_shapes(images): images = tf.ragged.stack(images) log("Created a batch of images with different " "shapes. If you want the shapes to be consistent, " "run `loader.resize_images('auto')`.") else: images = tf.stack(images) return images ######### AGMLDATALOADER METHODS ######### class AgMLObject(object): """Base class for the `AgMLDataLoader` to enable inheritance. This class solves a bug which arises when trying to dynamically inherit from `tf.keras.utils.Sequence` and/or `torch.utils.data.Dataset`. The fact that the `AgMLDataLoader` has this `AgMLObject` as a subclass enables it to be able to handle dynamic inheritance. This is the sole purpose of this subclass, it does not have any features. """ def _add_dataset_to_mro(inst, mode): """Adds the relevant backend class to the `AgMLDataLoader` MRO. This allows for the loader to dynamically inherent from the `tf.keras.utils.Sequence` and `torch.utils.data.Dataset`. """ if mode == 'tf': if not get_backend() == 'tf': if user_changed_backend(): raise StrictBackendError(change = 'tf', obj = inst) set_backend('tf') if tf.keras.utils.Sequence not in inst.__class__.__bases__: inst.__class__.__bases__ += (tf.keras.utils.Sequence, ) if mode == 'torch': if not get_backend() == 'torch': if user_changed_backend(): raise StrictBackendError(change = 'torch', obj = inst) if torch_data.Dataset not in inst.__class__.__bases__: inst.__class__.__bases__ += (torch_data.Dataset,)
10,087
33.081081
104
py
AgML
AgML-main/agml/backend/__init__.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Controls the AgML backend system. The first part of the backend is the backend deep learning library. The backend, one of {TensorFlow, PyTorch}, primarily exists for internal purposes, e.g. figuring out which methods to use in the data module or the actual model configuring/training modules. Secondly, the backend controls the loading/saving procedure for files within AgML, specifically data loaded or generated from the data module. """ from .config import ( data_save_path, set_data_save_path, synthetic_data_save_path, set_synthetic_save_path, model_save_path, set_model_save_path, clear_all_datasets, downloaded_datasets ) from .tftorch import get_backend, set_backend from .random import set_seed from . import experimental
1,371
33.3
76
py
AgML
AgML-main/agml/viz/labels.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np import matplotlib.pyplot as plt from agml.backend.tftorch import as_scalar, is_array_like from agml.viz.tools import format_image, _inference_best_shape, convert_figure_to_image from agml.viz.display import display_image def show_images_and_labels(images, labels = None, info = None, shape = None, **kwargs): """Visualizes a set of images with their classification labels. Given a set of images and their corresponding labels, this method will generate a grid for the images and display them with their image classification labels displayed underneath them. The shape of the grid will by default be inferenced to be the two closest factors of the number of images (to be as close to square as possible). If you provide an `info` parameter, which will consist of the `info` property of an AgMLDataLoader (literally pass `loader.info`), then the method will convert the classification numbers to their label names. If you don't want to display the image (and just get the output), pass `no_show` as true in order to bypass this. Parameters ---------- images : Any Either a list of images, a tuple of images and labels, or a list of image/label pairs (like you would get as the output of a dataset). labels : Any A list or array of classification labels. info : DatasetMetadata The `loader.info` attribute of a dataloader. shape : Any The shape of the display grid. Returns ------- The matplotlib figure with the plotted info. """ if images is not None and labels is None: if is_array_like(images[0], no_list = True): if images[0].ndim >= 3: images, labels = images[0], images[1] else: raise ValueError( "If passing a numpy array for `images`, expected at " "least three dimensions: (batch, height, width).") elif isinstance(images[0], (tuple, list)): if isinstance(images[0][0], np.ndarray): if len(images[0]) == 2: _images, _labels = [], [] for content in images: _images.append(content[0]) _labels.append(content[1]) images, labels = _images, _labels else: images, labels = images[0], images[1] if labels is None: raise TypeError( "Invalid format for `images` and `labels`, see documentation.") if isinstance(images, np.ndarray) and images.shape[0] > 100: images, labels = [images], [labels] # Check if the labels are converted to one-hot, and re-convert them back. if is_array_like(labels): if not isinstance(labels, (list, tuple)): if labels.ndim == 2: # noqa labels = np.argmax(labels, axis = -1) # If a prime number is passed, e.g. 23, then the `_inference_best_shape` # method will return the shape of (23, 1). Likely, the user is expecting # a non-rectangular shape such as (6, 4), where the bottom right axis is # empty. This method does not support such computations (yet). if shape is None: shape = _inference_best_shape(len(images)) if max(shape) > 20: raise NotImplementedError( "Length of maximum shape length is greater than 20. " "This method does not support non-rectangular shapes.") fig, axes = plt.subplots( shape[0], shape[1], figsize = (shape[1] * 2, shape[0] * 2)) try: iter_ax = axes.flat except AttributeError: # If showing only a single image. iter_ax = [axes] for image, label, ax in zip(images, labels, iter_ax): ax.imshow(format_image(image)) ax.set_aspect(1) label = as_scalar(label) if info is not None: label = info.num_to_class[label] ax.set_xticklabels([]) ax.set_yticklabels([]) ax.tick_params(axis = 'both', which = 'both', bottom = False, top = False, left = False, right = False) plt.setp(ax.spines.values(), visible = False) ax.set_xlabel(label) # Display and return the image. image = convert_figure_to_image() if not kwargs.get('no_show', False): _ = display_image(image) return image
5,095
38.8125
87
py
AgML
AgML-main/agml/viz/tools.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ A tools module for `agml.viz`, which also serves as almost a mini-backend to control ops such as the colormap being used. """ import os import io import json import functools import cv2 import numpy as np from PIL import Image import matplotlib.pyplot as plt from agml.backend.tftorch import tf, torch from agml.backend.config import _get_config, _update_config from agml.utils.logging import log # Sets the colormaps used in the other `agml.viz` methods. @functools.lru_cache(maxsize = None) def _load_colormaps(): with open(os.path.join(os.path.dirname(os.path.dirname(__file__)), '_assets', 'viz_colormaps.json'), 'r') as f: cmaps = json.load(f) ret_dict = {} for map_ in cmaps.items(): ret_dict[map_[0]] = map_[1] * 5 return ret_dict _COLORMAPS = _load_colormaps() _COLORMAP_CHOICE: str = 'default' # Sets the visualization backend: either `matplotlib` or `cv2`. @functools.lru_cache(maxsize = None) def _load_backend(): if _get_config('viz_backend') is None: _update_config('viz_backend', 'matplotlib') return _get_config('viz_backend') _BACKEND: str = _load_backend() def get_colormap(): """Returns the current AgML colormap.""" global _COLORMAPS, _COLORMAP_CHOICE return _COLORMAPS[_COLORMAP_CHOICE] def set_colormap(colormap): """Sets the current AgML colormap used in color displays. This method accepts one argument, `colormap`, which can be any of the colormaps listed in `_assets/viz_colormaps.json`, namely one of the following: 1. "default": Traditional matplotlib RGB colors. 2. "agriculture": Various shades of green (for agriculture). If you want to set a custom colormap, then pass a list of RGB values which will be used as the colormap. Parameters ---------- colormap : str The colormap to set. """ global _COLORMAP_CHOICE, _COLORMAPS if isinstance(colormap, list): if not all(len(i) == 3 for i in colormap): raise ValueError( "If you want a custom colormap, then pass a list of RGB values.") elif isinstance(colormap, str): colormap = colormap.lower() if colormap not in _COLORMAPS.keys(): raise ValueError(f"Invalid colormap {colormap} received.") else: raise TypeError(f"Invalid colormap of type {type(colormap)}.") _COLORMAPS['custom'] = colormap _COLORMAP_CHOICE = 'custom' def get_viz_backend(): """Returns the current AgML visualization backend.""" global _BACKEND return _BACKEND def set_viz_backend(backend): """Sets the global AgML visualization backend.""" if backend not in ['matplotlib', 'cv2']: raise ValueError(f"Invalid backend {backend} received.") global _BACKEND _update_config('viz_backend', backend) _BACKEND = backend def format_image(img, mask = False): """Formats an image to be used in a Matplotlib visualization. This method takes in one of a number of common image/array types and returns a formatted NumPy array with formatted image data as expected by matplotlib. This method is primarily necessary to serve as convenience in a few situations: converting images from PyTorch's channels first format to channels last, or removing the extra grayscale dimension in the case of grayscale images. Parameters ---------- img : Any An np.ndarray, torch.Tensor, tf.Tensor, or PIL.Image. mask : Any Whether the image is a segmentation mask. Returns ------- An np.ndarray formatted correctly for a Matplotlib visualization. """ # Get the numpy array from the image type. if isinstance(img, np.ndarray): img = img elif isinstance(img, Image.Image): img = np.array(img).reshape((img.height, img.width, len(img.getbands()))) elif isinstance(img, torch.Tensor): if img.is_cuda: img = img.cpu().detach().numpy() img = img.numpy() elif isinstance(img, tf.Tensor): img = img.numpy() else: raise TypeError( f"Expected either an np.ndarray, torch.Tensor, " f"tf.Tensor, or PIL.Image, got {type(img)}.") # Convert channels_first to channels_last. if img.ndim == 4: if img.shape[0] > 1: raise ValueError( f"Got a batch of images with shape {img.shape}, " f"expected at most a batch of one image.") img = np.squeeze(img) if img.shape[0] <= 3: img = np.transpose(img, (1, 2, 0)) # Remove the grayscale axis. if img.shape[-1] == 1: img = np.squeeze(img) # If the image is in range 0-255 but a float image, then # we need to convert it to an integer type. if mask: img = img.astype(np.uint8) else: if np.issubdtype(img.dtype, np.inexact): if not img.max() <= 1: # noqa img = img.astype(np.uint8) else: img = (img * 255).astype(np.uint8) # Convert 64-bit integer to unsigned 8-bit. if img.dtype == np.int64: log("Converting image of dtype `np.int64` to `np.uint8` for display. " "This may cause a loss in precision/invalid result.") img = img.astype(np.uint8) # Return the formatted image. return img def convert_figure_to_image(fig = None): """This method is used to convert a Matplotlib figure to an image array.""" # Use PIL to get the image, then convert to an array. buf = io.BytesIO() fig = fig if fig is not None else plt.gcf() fig.savefig(buf, format = 'png') buf.seek(0) arr = np.fromstring(buf.read(), dtype = np.uint8) return cv2.imdecode(arr, cv2.IMREAD_COLOR) def _inference_best_shape(n_images): """Inferences the best matplotlib row/column layout. This method searches for the two closest factors of the number `n_images`, and returns this tuple as the best shape, since this is the closest to a square grid as possible. """ a, b, i = 1, n_images, 0 while a < b: i += 1 if n_images % i == 0: a = i b = n_images // a return [b, a]
6,828
31.061033
82
py
AgML
AgML-main/agml/viz/general.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import matplotlib.pyplot as plt from agml.backend.tftorch import is_array_like from agml.viz.boxes import show_image_and_boxes from agml.viz.masks import show_image_and_overlaid_mask from agml.viz.labels import show_images_and_labels from agml.viz.tools import format_image, _inference_best_shape, convert_figure_to_image from agml.viz.display import display_image def show_sample(loader, image_only = False, **kwargs): """A simplified convenience method that visualizes a sample from a loader. This method works for all kind of annotations; it picks the appropriate visualization method and then calls it with a sample from the loader. If you want to customize the way the output looks, then you need to use the actual methods directly. Parameters ---------- loader : AgMLDataLoader An AgMLDataLoader of any annotation type. image_only : bool Whether to only display the image. Returns ------- The matplotlib figure. """ if kwargs.get('sample', None) is not None: sample = kwargs['sample'] else: sample = loader[0] if image_only: return show_images(sample[0]) if loader.task == 'object_detection': return show_image_and_boxes( sample, info = loader.info, no_show = kwargs.get('no_show', False)) elif loader.task == 'semantic_segmentation': return show_image_and_overlaid_mask( sample, no_show = kwargs.get('no_show', False)) elif loader.task == 'image_classification': return show_images_and_labels( sample, info = loader.info, no_show = kwargs.get('no_show', False)) def show_images(images, shape = None, **kwargs): """Shows multiple images in a grid format with the given shape. Given a set of images, this method will generate a grid for the images and display them as such. The shape of the grid will by default be inferenced to be the two closest factors of the number of images (to be as close to square as possible). If you don't want to display the image (and just get the output), pass `no_show` as true in order to bypass this. Parameters ---------- images : Any Either a list of images, a tuple of images and labels, or a list of image/label pairs (like you would get as the output of a dataset). shape : Any The shape of the display grid. Returns ------- The matplotlib figure with the plotted images. """ # If only one image is passed, then we can directly display that image. if not isinstance(images, (list, tuple)): if is_array_like(images): images = format_image(images) if not kwargs.get('no_show', False): display_image(images) return images # If a prime number is passed, e.g. 23, then the `_inference_best_shape` # method will return the shape of (23, 1). Likely, the user is expecting # a non-rectangular shape such as (6, 4), where the bottom right axis is # empty. This method does not support such computations (yet). if shape is None: shape = _inference_best_shape(len(images)) if max(shape) > 20: raise NotImplementedError( "Length of maximum shape length is greater than 20. " "This method does not support non-rectangular shapes.") fig, axes = plt.subplots( shape[0], shape[1], figsize = (shape[1] * 2, shape[0] * 2)) try: iter_ax = axes.flat except AttributeError: # If showing only a single image. iter_ax = [axes] for image, ax in zip(images, iter_ax): ax.imshow(format_image(image)) ax.set_aspect(1) ax.set_xticklabels([]) ax.set_yticklabels([]) ax.tick_params( axis = 'both', which = 'both', bottom = False, top = False, left = False, right = False ) plt.setp(ax.spines.values(), visible = False) fig.tight_layout() # Display and return the image. image = convert_figure_to_image() if not kwargs.get('no_show', False): display_image(image) return image
4,786
35.823077
87
py
AgML
AgML-main/agml/utils/general.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import re import cv2 import glob import math import matplotlib.pyplot as plt import matplotlib.patches as patches from math import pi, floor from scipy import signal import numpy as np # Represents an empty object, but allows passing `None` # as an independent object in certain cases. NoArgument = object() def placeholder(obj): """Equivalent of lambda x: x, but enables pickling.""" return obj def to_camel_case(s): """Converts a given string `s` to camel case.""" s = re.sub(r"(_|-)+", " ", s).title().replace(" ", "") # noqa return ''.join(s) def resolve_list_value(val): """Determines whether a list contains one or multiple values.""" if len(val) == 1: return val[0] return val def resolve_tuple_values(*inputs, custom_error = None): """Determines whether values are distributed amongst the values in `inputs`. """ if isinstance(inputs[0], (list, tuple)) and all(c is None for c in inputs[1:]): if len(inputs[0]) != len(inputs): # special case for COCO JSON if len(inputs) == 3 and len(inputs[0]) == 2 and isinstance(inputs[0][1], dict): try: return inputs[0][0], inputs[0][1]['bbox'], inputs[0][1]['category_id'] except KeyError: return inputs[0][0], inputs[0][1]['bboxes'], inputs[0][1]['labels'] if custom_error is not None: raise ValueError(custom_error) else: raise ValueError(f"Expected either a tuple with {len(inputs)} values " f"or {len(inputs)} values across two arguments.") else: return inputs[0] return inputs def resolve_tuple(sequence): """Resolves a sequence to a tuple.""" if isinstance(sequence, np.ndarray): sequence = sequence.tolist() return tuple(i for i in sequence) def has_nested_dicts(obj): """Returns whether a dictionary contains nested dicts.""" return any(isinstance(i, dict) for i in obj.values()) def as_scalar(inp): """Converts an input value to a scalar.""" if isinstance(inp, (int, float)): return inp if np.isscalar(inp): return inp.item() if isinstance(inp, np.ndarray): return inp.item() from agml.backend.tftorch import torch if isinstance(inp, torch.Tensor): return inp.item() from agml.backend.tftorch import tf if isinstance(inp, tf.Tensor): return inp.numpy() raise TypeError(f"Unsupported variable type {type(inp)}.") def scalar_unpack(inp): """Unpacks a 1-d array into a list of scalars.""" return [as_scalar(item) for item in inp] def is_array_like(inp): """Determines if an input is a np.ndarray, torch.Tensor, or tf.Tensor.""" if isinstance(inp, np.ndarray): return True from agml.backend.tftorch import torch if isinstance(inp, torch.Tensor): return True from agml.backend.tftorch import tf if isinstance(inp, tf.Tensor): return True return False def shapes(seq): """Returns the shapes (or lengths) of all of the objects in the sequence.""" try: return [getattr(obj, 'shape', len(obj)) for obj in seq] except: raise ValueError(f"One or more of the objects has no shape or length: {seq}.") def weak_squeeze(arr, ndims = 2): """Performs a 'weak squeeze', adding a dimension back if necessary.""" if isinstance(arr, np.ndarray): arr = np.squeeze(arr) while arr.ndim < ndims: arr = np.expand_dims(arr, axis = 0) return arr def is_float(num): """Determines if a number is a float.""" is_ = isinstance(num, float) or isinstance(num, np.float32) or isinstance(num, np.float64) if is_: return True try: float(num) except ValueError: return False return True def is_int(num): """Determines if a number is an int.""" is_ = isinstance(num, int) or isinstance(num, np.int32) or isinstance(num, np.int64) if is_: return True try: int(num) except ValueError: return False return True
4,734
29.352564
94
py
AgML
AgML-main/agml/data/object.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import abc from functools import wraps import cv2 import numpy as np from agml.framework import AgMLSerializable from agml.utils.image import imread_context from agml.backend.tftorch import scalar_unpack class DataObject(AgMLSerializable): """Stores a single piece of data and its corresponding annotation. This class stores an image and its corresponding annotation depending on the task, such as a label for a image classification task, a COCO JSON dictionary for an object detection task, or a mask for a semantic segmentation task. Fundamentally, it works as a two-object list, but the data is stored internally in a specific representation and images are only loaded lazily when necessary. This object is used internally in the `AgMLDataLoader`, and objects returned from it are returned as their expected contents (NumPy arrays/dictionaries/integers) when necessary. """ serializable = frozenset( ('image_object', 'annotation_obj', 'dataset_root')) _abstract = frozenset(('_load_image_input', '_parse_annotation')) def __init__(self, image, annotation, root): # The `image` parameter is constant among different tasks. self._image_object = image # The `annotation` parameter varies with the task. self._annotation_obj = annotation # The `root` is the local root of the dataset. This is used # for object detection datasets primarily, whose COCO JSON # dictionary doesn't contain the full path, only the base. self._dataset_root = root def __len__(self): return 2 def __getitem__(self, i): return self.get()[i] def __repr__(self): return f"<DataObject: {self._image_object}, {self._annotation_obj}>" @staticmethod def _parse_image(path): with imread_context(os.path.abspath(path)) as image: return cv2.cvtColor(image, cv2.COLOR_BGR2RGB) @staticmethod def _parse_depth_image(path): with imread_context(os.path.abspath(path), flags = -1) as image: return image.astype(np.int32) @staticmethod def _parse_spectral_image(path): None # noqa, prevents `all abstract methods must be implemented` raise NotImplementedError("Multi/Hyperspectral images are not yet supported.") def get(self): """Returns the image and annotation pair with applied transforms. This method is the main exposed method to process the data. It loads the image and processes the annotation, then applies transformations. """ return self._load() def _load(self): """Loads the image and annotation and returns them.""" image = self._load_image_input(self._image_object) annotation = self._parse_annotation(self._annotation_obj) return image, annotation def __init_subclass__(cls, **kwargs): # Wraps derived abstract methods with the corresponding # docstring from this `DataObject` class and updates # the derived class' dictionary. wrapped_updates = {} self = super(DataObject, cls).__thisclass__ # noqa for name, method in cls.__dict__.items(): if name in cls._abstract: wrapped = wraps(getattr(self, name))(method) wrapped_updates[name] = wrapped for name, method in wrapped_updates.items(): # noqa setattr(cls, name, method) @staticmethod def create(contents, task, root): """Creates a new `DataObject` for the corresponding task.""" if task == 'image_classification': cls = ImageClassificationDataObject elif task == 'image_regression': cls = ImageRegressionDataObject elif task == 'object_detection': cls = ObjectDetectionDataObject elif task == 'semantic_segmentation': cls = SemanticSegmentationDataObject else: raise ValueError(f"Unsupported task {task}.") return cls(*contents, root) # The following methods are used to load the image and annotation. @abc.abstractmethod def _load_image_input(self, path): """Loads image inputs based on the task. Derived by subclasses.""" raise NotImplementedError() @abc.abstractmethod def _parse_annotation(self, obj): """Parses an annotation based on the task. Derived by subclasses.""" raise NotImplementedError() # The following methods function independently whenever an # annotation is loaded. Depending on the task, the appropriate # method is called and a corresponding piece of data is returned. @staticmethod def _parse_label(obj): """Parses a label for an image classification task.""" try: obj = int(obj) except TypeError: raise Exception(f"Could not convert object {obj} of " f"type {type(obj)} to a scalar integer " f"for image classification.") return obj @staticmethod def _parse_coco(obj): """Parses a COCO JSON dictionary for an object detection task.""" annotation = {'bbox': [], 'category_id': [], 'area': [], 'image_id': "", 'iscrowd': [], 'segmentation': []} for a_set in obj: x, y, w, h = a_set['bbox'] x = int(np.clip(x, 0, None)) y = int(np.clip(y, 0, None)) a_set['bbox'] = [x, y, w, h] annotation['bbox'].append(a_set['bbox']) annotation['category_id'].append(a_set['category_id']) annotation['iscrowd'].append(a_set['iscrowd']) annotation['segmentation'].append(a_set['segmentation']) annotation['area'].append(a_set['area']) annotation['image_id'] = a_set['image_id'] for key, value in annotation.items(): # Creating nested sequences from ragged arrays (see numpy). if key in ['segmentation']: out = np.array(value, dtype = object) else: out = np.array(value) if np.isscalar(out): out = scalar_unpack(out) annotation[key] = out return annotation @staticmethod def _parse_mask(obj): """Parses a mask for a semantic segmentation task.""" with imread_context(os.path.realpath(obj)) as image: if image.ndim == 3: if not np.all(image[:, :, 0] == image[:, :, 1]): raise TypeError( f"Invalid annotation mask of shape {image.shape}.") image = image[:, :, 0] return np.squeeze(image) class ImageClassificationDataObject(DataObject): """Serves as a `DataObject` for image classification tasks.""" def _load_image_input(self, path): return self._parse_image(path) def _parse_annotation(self, obj): return self._parse_label(obj) class ImageRegressionDataObject(DataObject): """Serves as a `DataObject` for image regression tasks.""" def _load_image_input(self, contents): # The easy case, when there is only one input image. if isinstance(contents, str) and os.path.exists(contents): return self._parse_image(contents) # Otherwise, we have a dictionary containing multiple # input types, so we need to independently load those. images = dict.fromkeys(contents.keys(), None) for c_type, path in contents.items(): if c_type == 'image': images[c_type] = self._parse_image(path) elif c_type == 'depth_image': images[c_type] = self._parse_depth_image(path) else: images[c_type] = self._parse_spectral_image(path) return images def _parse_annotation(self, obj): return obj class ObjectDetectionDataObject(DataObject): """Serves as a `DataObject` from object detection tasks.""" def _load_image_input(self, path): path = os.path.join(self._dataset_root, 'images', path) return self._parse_image(path) def _parse_annotation(self, obj): return self._parse_coco(obj) class SemanticSegmentationDataObject(DataObject): """Serves as a `DataObject` for semantic segmentation tasks.""" def _load_image_input(self, path): return self._parse_image(path) def _parse_annotation(self, obj): return self._parse_mask(obj)
9,127
36.875519
86
py
AgML
AgML-main/agml/data/tools.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np from agml.backend.tftorch import is_array_like from agml.utils.general import resolve_list_value def _resolve_coco_annotations(annotations): """Resolves COCO annotations to a standard format. Generally, COCO annotations will either be represented as a dictionary which contains multiple elements for each of its keys, e.g. multiple bounding boxes and areas, or a list of multiple COCO dictionaries. This method resolves it into the former case. """ if is_array_like(annotations) and not isinstance(annotations, np.ndarray): return annotations.numpy() if isinstance(annotations, np.ndarray): return annotations if isinstance(annotations, list): if not isinstance(annotations[0], dict): return annotations if len(annotations) == 1: return annotations annotation = {'bboxes': [], 'labels': [], 'area': [], 'image_id': "", 'iscrowd': [], 'segmentation': []} for a_set in annotations: annotation['bboxes'].append(a_set['bbox']) annotation['labels'].append(a_set['category_id']) annotation['iscrowd'].append(a_set['iscrowd']) annotation['segmentation'].append(a_set['segmentation']) annotation['area'].append(a_set['area']) annotation['image_id'] = a_set['image_id'] for key, value in annotation.items(): out = np.array(value) if np.isscalar(out): out = out.item() annotation[key] = out return annotation elif isinstance(annotations, dict): return annotations else: raise TypeError( "Expected either a single COCO annotation " "dictionary or a list of multiple dictionaries.") def coco_to_bboxes(annotations): """Extracts the bounding boxes and labels from COCO JSON annotations. Given either a list of COCO JSON annotations, or a single dictionary with multiple values, this method will extract the bounding boxes and category labels from the annotations and return just the bounding boxes and category labels in two arrays. Parameters ---------- annotations : {list, dict} The COCO JSON annotations in list/dict format. Returns ------- Two arrays consisting of the bounding boxes and labels. """ annotations = _resolve_coco_annotations(annotations) return annotations['bboxes'], annotations['labels'] def convert_bbox_format(annotations_or_bboxes, fmt): """Converts bounding box formats for COCO JSON and others. This method converts the format of bounding boxes as specified in the 'fmt' argument, which describes the format of the bounding boxes passed to the 'annotations_or_bboxes' argument. From there, it will convert to the standard COCO JSON bounding box format, namely ('x1', 'y1' 'width', 'height'). This method supports the following conversions (note that 'x1' and 'y1' are the top-left coordinates, and 'x2' and 'y2' are top-right): 1. ('x1', 'x2', 'y1', 'y2') to COCO JSON. 2. ('x_min', 'y_min', 'x_max', 'y_max') to COCO JSON. This format can also be passed with the simple string `pascal_voc`. 3. ('x_min', 'y_min', 'width', 'height') to COCO JSON. Note that the variables in the bounding boxes in the above format can be in any order, this just needs to be reflected in the 'fmt' argument, and it should contain some combination of the above. Parameters ---------- annotations_or_bboxes : {np.ndarray, list, dict} Either a COCO JSON annotation dictionary, or a numpy array/list with all of the annotations. fmt : {list, tuple} A list or tuple with one of the above formats. Returns ------- The initial argument type (either dict or array) with the bounding boxes formatted in the COCO JSON format. """ annotations_or_bboxes = _resolve_coco_annotations(annotations_or_bboxes) if isinstance(annotations_or_bboxes, dict): annotations = annotations_or_bboxes['bboxes'] else: annotations = annotations_or_bboxes if isinstance(annotations[0], (int, float)): annotations = [annotations] if isinstance(fmt, str): if 'voc' in fmt or 'pascal' in fmt: fmt = 'x_min y_min x_max y_max' elif 'efficientdet' in fmt or 'effdet' in fmt: fmt = 'y_min x_min y_max x_max' if ',' in fmt: fmt = fmt.split(',') else: fmt = fmt.split(' ') if len(fmt) != 4: raise ValueError(f"Argument 'fmt' should contain 4 values, got {len(fmt)}.") # Define all of the intermediate conversion methods def _x1_x2_y1_y2_to_coco(annotation): # noqa x1, x2, y1, y2 = annotation width, height = abs(x2 - x1), abs(y2 - y1) return [x1, y1, width, height] def _xmin_ymin_xmax_ymax_to_coco(annotation): # noqa xmin, ymin, xmax, ymax = annotation width, height = abs(xmax - xmin), abs(ymax - ymin) return [xmin, ymin, width, height] def _xmin_ymin_width_height_to_coco(annotation): # noqa xmin, ymin, width, height = annotation x1, y1 = xmin, ymin - height return [x1, y1, width, height] def _x1_y1_width_height_to_coco(annotation): # noqa return annotation # This is just here for reordering. # Resolve the format fmt_bases = [['x1', 'x2', 'y1', 'y2'], ['x_min', 'y_min', 'x_max', 'y_max'], ['x_min', 'y_min', 'width', 'height'], ['x1', 'y1', 'width', 'height']] fmt_map = {0: _x1_x2_y1_y2_to_coco, 1: _xmin_ymin_xmax_ymax_to_coco, 2: _xmin_ymin_width_height_to_coco, 3: _x1_y1_width_height_to_coco} fmt_found = False map_fmt, select_order = None, None for indx, base in enumerate(fmt_bases): if all(i in base for i in fmt): fmt_found, map_fmt = True, fmt_map[indx] select_order = [base.index(i) for i in fmt] if not fmt_found: raise ValueError( f"Invalid format {fmt}, see `convert_bbox_format` " f"for information about valid formats.") # Convert the formats formatted_annotations = [] for bbox in annotations: sorted_bbox = np.array(bbox)[select_order] formatted_annotations.append(map_fmt(sorted_bbox)) formatted_annotations = np.array(formatted_annotations) if isinstance(annotations_or_bboxes, dict): res = annotations_or_bboxes.copy() res['bboxes'] = formatted_annotations return res return resolve_list_value(formatted_annotations)
7,351
39.395604
84
py
AgML
AgML-main/agml/data/experimental.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """ Experimental data utilities that are in development. """ import functools import numpy as np __all__ = ['generate_keras_segmentation_dual_transform'] def generate_keras_segmentation_dual_transform(*layers): """Generates a `dual_transform` pipeline from Keras preprocessing layers. This method takes in Keras preprocessing layers and generates a transformation pipeline for the `dual_transform` argument in *semantic segmentation* loaders, which applies the transform in the same fashion to both the image and annotation. This is due to the fact that TensorFlow has its operation-level random states different than its module-level random state, so the layers need to have their seeds manually set in order to work. In essence, for each of the preprocessing layers passed, this method conducts the following operations: > def preprocessing_transform(image, annotation): > layer = functools.partial(KerasPreprocessingLayer, **kwargs) > seed = np.random.randint(BUFFER_SIZE) # up to sys.maxsize > image = layer(image, seed = seed) > annotation = layer(annotation, seed = seed) > return image, annotation It then repeats this transform for all of the preprocessing layers passed, and returns a method which has this behavior wrapped into it and can perform it when the preprocessing is actually conducted. Parameters ---------- layers : Any Either a Sequential model with preprocessing layers, or a set of instantiated preprocessing layers. Returns ------- """ import tensorflow as tf if len(layers) == 1: if isinstance(layers[0], tf.keras.Sequential): layers = layers[0].layers # These methods perform the behavior indicated in the # code snippet above (for each of the layers given). def _single_preprocessing_layer_base(layer_, build_dict): def _internal(image, annotation, seed): instantiated_layer = functools.partial(layer_, **build_dict) seed_update = {} if seed is not None: seed_update['seed'] = seed image = instantiated_layer(**seed_update)(image) annotation = instantiated_layer(**seed_update)(annotation) return image, annotation return _internal preprocessing_methods, use_seeds = [], [] for layer in layers: config = layer.get_config() if 'seed' in config: config.pop('seed') use_seeds.append(True) else: use_seeds.append(False) preprocessing_methods.append( _single_preprocessing_layer_base(layer.__class__, config)) def _execute_preprocessing(layers_, use_seeds_): def _execute(image, annotation): for p_layer, seed_ in zip(layers_, use_seeds_): seed = np.random.randint(2147483647) if seed_ else None image, annotation = p_layer(image, annotation, seed = seed) return image, annotation return _execute return _execute_preprocessing(preprocessing_methods, use_seeds)
3,748
34.704762
77
py
AgML
AgML-main/agml/data/manager.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import numpy as np from agml.framework import AgMLSerializable from agml.data.object import DataObject from agml.data.builder import DataBuilder from agml.data.metadata import DatasetMetadata from agml.data.managers.transforms import TransformManager from agml.data.managers.resize import ImageResizeManager from agml.data.managers.training import TrainingManager from agml.utils.general import NoArgument from agml.backend.tftorch import convert_to_batch, is_array_like from agml.utils.random import seed_context from agml.utils.image import consistent_shapes from agml.utils.logging import log class DataManager(AgMLSerializable): """Manages the data for a dataset loaded in an `AgMLDataLoader`. The `DataManager` is the core internal object which, as its name suggests, manages the data being used in an `AgMLDataLoader`. Upon instantiation of an `AgMLDataLoader`, the internal data, which is represented generally as image paths and corresponding annotations, is placed into a `DataManager` container. Accessing data from an `AgMLDataLoader` calls the `DataManager`, which takes into account all of the potential splits, transforms, and batching that may have been applied to the loader. This enables more streamlined processing and transforming of data, as well as parallelizing data loading as desired (in the future). Another way to think about the `DataManager` is it being a list of `DataObjects`, with extra logic for processing. """ serializable = frozenset(( 'data_objects', 'resize_manager', 'accessors', 'task', 'dataset_name', 'shuffle', 'batch_size', 'dataset_root', 'transform_manager', 'builder', 'train_manager')) def __init__(self, builder, task, name, root, **kwargs): # Set the basic class information. self._task = task self._dataset_name = name self._dataset_root = root # Create the `DataObject`s from the `DataBuilder`. if not isinstance(builder, DataBuilder): builder = DataBuilder.from_data( contents = builder, info = DatasetMetadata(name), root = root) self._builder = builder self._create_objects(self._builder, task) # Set up the internal transform managers. These control # the application of transformations to the input data. self._transform_manager = TransformManager(task = task) self._resize_manager = ImageResizeManager( task = task, dataset = name, root = root ) # The transform and resize managers are wrapped inside of a # `TrainingManager`, which controls the application of # preprocessing to the data based on the class state. self._train_manager = TrainingManager( transform_manager = self._transform_manager, resize_manager = self._resize_manager, task = task ) # While we store data in the list of `DataObject`s, the actual # accessing of the data by index doesn't happen by directly # accessing that list, but instead by accessing the array of # indices below. This is useful for two reasons. # # It allows a much more simpler way to store and access the # state of the data. E.g., when splitting the data or accessing # it, we only need to check through this array rather than # searching through or shuffling the actual list of objects. # # Secondly, it makes batching data much more straightforward and # allows for storing representations of the data in different # formats, e.g., shuffling, to be done without having to interfere # with the actual `DataObject`s, which is much more convenient. self._accessors = np.arange(len(self._data_objects)) # The following parameters store various parameters which are # used internally or accessed by the `AgMLDataLoader` externally. self._batch_size = None self._shuffle = kwargs.get('shuffle', True) self._maybe_shuffle() def data_length(self): """Calculates the length of the data based on the batching state.""" return len(self._accessors) def _get_random_index(self): """Returns a random accessor index to get a specific sample.""" return np.random.choice(self._accessors) def _create_objects(self, builder, task): """Creates `DataObject`s from the provided content. Here, `content` is a dictionary mapping an an input data piece, an image, with its corresponding expected output, its annotation. """ self._data_objects = [] contents = builder.get_contents() if 'inputs' in contents.keys(): contents = zip(tuple(contents['inputs']), tuple(contents['outputs'])) else: contents = contents.items() for content in list(contents): self._data_objects.append(DataObject.create( contents = content, task = task, root = self._dataset_root)) def _maybe_shuffle(self, seed = None): """Wraps automatic shuffling to see if it is enabled or not.""" if self._shuffle: self.shuffle(seed = seed) def update_train_state(self, state): """Updates the training state in the `TrainingManager`.""" self._train_manager.update_state(state) def shuffle(self, seed = None): """Shuffles the contents of the `DataManager`. This method simply shuffles the order of the `DataObject`s which are stored inside this `DataManager`. Optionally, a seed can be provided to shuffle them inside of a specific context. """ if seed is None: np.random.shuffle(self._accessors) else: with seed_context(seed): np.random.shuffle(self._accessors) def generate_split_contents(self, splits): """Generates split contents given a dictionary of the split indexes. This method, given a set of data split indexes, applies the indexing to the original content and gets a mapping of images and annotations which are returned back to the `AgMLDataLoader` to be constructed into `DataBuilder`s and wrapped into new `DataManager`s. """ if self._task == 'object_detection': contents = np.array(list( self._builder.get_contents().items()), dtype = object) else: contents = np.array(list(self._builder.get_contents().items())) try: return {k: dict(contents[v]) for k, v in splits.items()} except IndexError: raise Exception( f"Could not generate split contents, likely due to an error " f"with the metadata for the dataset `{self._dataset_name}`. " f"Please raise this error with the AgML team.") def batch_data(self, batch_size): """Batches the data into consistent groups. The batched data is stored inside of this manager, as a set of indexes which are read and loaded when the data is accessed. See the information above the `_accessors` parameter above. """ # If the data is already batched and a new batch size is called, # then update the existing batch sizes. For unbatching the data, # update the batch state and then flatten the accessor array. if self._batch_size is not None: try: self._accessors = np.concatenate(self._accessors).ravel() except ValueError: # The array is currently 0-dimensional. pass if batch_size is None or batch_size == 0: self._batch_size = None return # If we have a batch size of `1`, then don't do anything # since this doesn't really mean to do anything. if batch_size == 1: return # Otherwise, calculate the actual batches and the overflow # of the contents, and then update the accessor. num_splits = len(self._accessors) // batch_size data_items = np.array(self._accessors) overflow = len(self._accessors) - num_splits * batch_size extra_items = data_items[-overflow:] try: batches = np.array_split( np.array(self._accessors [:num_splits * batch_size]), num_splits) except ValueError: log(f"There is less data ({len(self._accessors)}) than the provided " f"batch size ({batch_size}). Consider using a smaller batch size.") batches = [self._accessors] else: if len(extra_items) < batch_size: batches.append(extra_items) self._accessors = np.array(batches, dtype = object) self._batch_size = batch_size def assign_resize(self, image_size, method): """Assigns a resizing factor for the image and annotation data.""" if image_size is None: image_size = 'default' self._resize_manager.assign(image_size, method) def push_transforms(self, **transform_dict): """Pushes a transformation to the data transform pipeline.""" # Check if any transforms are being reset and assign them as such. if all(i is NoArgument for i in transform_dict): transform_dict = { 'transform': 'reset', 'target_transform': 'reset', 'dual_transform': 'reset' } else: empty_keys, reset_keys = [], [] for key, value in transform_dict.items(): if value is NoArgument: empty_keys.append(key) if value is None: reset_keys.append(key) if len(empty_keys) != 0: for key, value in transform_dict.items(): if value is None: transform_dict[key] = 'reset' elif value is NoArgument: transform_dict[key] = None # There is no `dual_transform` object for image classification. if self._task == 'image_classification': if transform_dict['dual_transform'] is None: transform_dict['dual_transform'] = 'reset' # Assign the transforms to the manager in order. for key in ['transform', 'target_transform', 'dual_transform']: self._transform_manager.assign(key, transform_dict[key]) def _load_one_image_and_annotation(self, obj): """Loads one image and annotation from a `DataObject`.""" return self._train_manager.apply( obj = obj, batch_state = self._batch_size is not None ) def _load_multiple_items(self, indexes): """Loads multiple images and annotations from a set of `DataObject`s.""" # Either we're getting multiple batches, or just multiple items. contents = [] if self._batch_size is not None: for i in indexes: contents.append(self._load_batch(self._accessors[i])) else: for i in indexes: contents.append(self._load_one_image_and_annotation( self._data_objects[self._accessors[i]])) return contents def _batch_multi_image_inputs(self, images): """Converts either a list of images or multiple input types into a batch.""" # If the input images are just a simple batch. if is_array_like(images[0]): return convert_to_batch(images) # Otherwise, convert all of them independently. keys = images[0].keys() batches = {k: [] for k in keys} for sample in images: for key in sample: batches[key].append(sample[key]) return {k: self._batch_multi_image_inputs(i) for k, i in batches.items()} def _batch_multi_output_annotations(self, annotations): """Converts either a list of annotations or multiple annotation types into a batch.""" # If the output annotations are simple objects. if (isinstance(annotations[0], (list, np.ndarray)) or isinstance(annotations, (list, np.ndarray)) and isinstance(annotations[0], (int, float))): if not consistent_shapes(annotations): annotations = np.array(annotations, dtype = object) else: annotations = np.array(annotations) return annotations # For object detection, just return the COCO JSON dictionaries. if self._task == 'object_detection': return annotations # Otherwise, convert all of them independently. keys = annotations[0].keys() batches = {k: [] for k in keys} for sample in annotations: for key in sample: batches[key].append(sample[key]) return {k: self._batch_multi_output_annotations(i) for k, i in batches.items()} def _load_batch(self, batch_indexes): """Gets a batch of data from the dataset. This differs from simply getting multiple pieces of data from the dataset, such as a slice, in that it also stacks the data together into a valid batch and returns it as such. """ # Get the images and annotations from the data objects. images, annotations = [], [] for index in batch_indexes: image, annotation = self._load_one_image_and_annotation( self._data_objects[index]) images.append(image) annotations.append(annotation) # Attempt to create batched image arrays. images = self._batch_multi_image_inputs(images) # Attempt the same for the annotation arrays. This is more complex # since there are many different types of annotations, namely labels, # annotation masks, COCO JSON dictionaries, etc. We need to properly # create a batch in each of these cases. annotations = self._batch_multi_output_annotations(annotations) # Return the batches. return self._train_manager.make_batch( images = images, annotations = annotations ) def get(self, indexes): """Loads and processes a piece (or pieces) of data from the dataset. This is the actual accessor method that performs the loading of data and the relevant processing as dictated by loading, image resizing, transform application, and other internal processing methods such as creating batches. This is called by the `AgMLDataLoader` to get data. """ # If there is only one index and the data is not batched, # then we just need to return a single `DataObject`. if isinstance(indexes, int) and self._batch_size is None: return self._load_one_image_and_annotation( self._data_objects[self._accessors[indexes]] ) # If we have a batch of images, then return the batch. if isinstance(indexes, int) and self._batch_size is not None: return self._load_batch(self._accessors[indexes]) # Otherwise, if there are multiple indexes (e.g., an unstacked # slice or just a tuple of integers), then we get multiple images. if isinstance(indexes, (list, tuple)): return self._load_multiple_items(indexes)
16,197
41.626316
94
py
AgML
AgML-main/agml/data/loader.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os import json import copy import glob from typing import Union from collections.abc import Sequence from decimal import getcontext, Decimal import numpy as np from agml.framework import AgMLSerializable from agml.data.manager import DataManager from agml.data.builder import DataBuilder from agml.data.metadata import DatasetMetadata, make_metadata from agml.utils.logging import log from agml.utils.io import get_file_list, get_dir_list from agml.utils.data import load_public_sources from agml.utils.general import NoArgument, resolve_list_value from agml.utils.random import inject_random_state from agml.backend.config import data_save_path, synthetic_data_save_path from agml.backend.experimental import AgMLExperimentalFeatureWrapper from agml.backend.tftorch import ( get_backend, set_backend, user_changed_backend, StrictBackendError, _add_dataset_to_mro, # noqa ) from agml.viz.general import show_sample class AgMLDataLoaderMeta(type): def __instancecheck__(self, instance): # This override allows for objects of type `AgMLMultiDatasetLoader` # to be treated as an `AgMLDataLoader` when the following command # is run: `isinstance(a, AgMLDataLoader)` (hacky fix, essentially). if isinstance(instance, self.__class__): return True from agml.data.multi_loader import AgMLMultiDatasetLoader if isinstance(instance, AgMLMultiDatasetLoader): return True return False class AgMLDataLoader(AgMLSerializable, metaclass = AgMLDataLoaderMeta): """Loads and provides a processing interface for a dataset. The `AgMLDataLoader` is the main interface to AgML's public dataset interface, and exposes an API which enables the downloading and subsequent local loading of a public dataset, as well as various preprocessing functions and hooks to integrate into existing pipelines. Methods provided include splitting the dataset into train/val/test sets, batching the data, applying transforms, and more. All of the processing code is contained internally, so all you need to do is instantiate the loader and call the relevant methods to apply the preprocessing methods. `AgMLDataLoader` supports both TensorFlow and PyTorch as backends, and can automatically perform tensor conversion and batching to enable seamless usage in training or inference pipelines. Data can also be exported into native TensorFlow and PyTorch objects. There is also support for using custom datasets outside of the AgML public data repository. To do this, you need to pass an extra argument containing metadata for the dataset, after which point the loader will work as normal (and all interfaces, except for the info parameters which are not provided, will also be available for standard use). Parameters ---------- dataset : str The name of the public dataset you want to load. See the helper method `agml.data.public_data_sources()` for a list of datasets. If using a custom dataset, this can be any valid string. kwargs : dict, optional dataset_path : str, optional A custom path to download and load the dataset from. overwrite : bool, optional Whether to rewrite and re-install the dataset. meta : dict, optional A dictionary consisting of metadata properties, if you want to create a custom loader. At minimum, this needs to contain two parameters: `task`, indicating the type of machine learning task that the dataset is for, and `classes`, a list of the classes that the dataset contains. Notes ----- See the methods for examples on how to use an `AgMLDataLoader` effectively. """ serializable = frozenset(( 'info', 'builder', 'manager', 'train_data', 'val_data', 'test_data', 'is_split', 'meta_properties')) def __new__(cls, dataset, **kwargs): # If a single dataset is passed, then we use the base `AgMLDataLoader`. # However, if an iterable of datasets is passed, then we need to # dispatch to the subclass `AgMLMultiDatasetLoader` for them. if isinstance(dataset, (str, DatasetMetadata)): return super(AgMLDataLoader, cls).__new__(cls) elif isinstance(dataset, Sequence): if len(dataset) == 1: log("Received a sequence with only one element when " "instantiating an `AgMLDataLoader`. Returning " "a regular, single-element data loader.") return super(AgMLDataLoader, cls).__new__(cls) from agml.data.multi_loader import AgMLMultiDatasetLoader return AgMLMultiDatasetLoader(dataset, **kwargs) raise TypeError( f"Expected either a single dataset name (or metadata), or" f"a list of dataset names/metadata when instantiating an " f"`AgMLDataLoader`. Got {dataset} of type {type(dataset)}.") def __getnewargs__(self): return self._info.name, def __init__(self, dataset, **kwargs): """Instantiates an `AgMLDataLoader` with the dataset.""" # Set up the dataset and its associated metadata. self._info = make_metadata(dataset, kwargs.get('meta', None)) # The data for the class is constructed in two stages. First, the # internal contents are constructed using a `DataBuilder`, which # finds and wraps the local data in a proper format. self._builder = DataBuilder( info = self._info, dataset_path = kwargs.get('dataset_path', None), overwrite = kwargs.get('overwrite', False) ) # These contents are then passed to a `DataManager`, which conducts # the actual loading and processing of the data when called. self._manager = DataManager( builder = self._builder, task = self._info.tasks.ml, name = self._info.name, root = self._builder.dataset_root ) # If the dataset is split, then the `AgMLDataLoader`s with the # split and reduced data are stored as accessible class properties. self._train_data = None self._val_data = None self._test_data = None self._is_split = False # Set the direct access metadata properties like `num_images` and # `classes`, since these can be modified depending on the state of # the loader, whilst the `info` parameter attributes cannot. self._meta_properties = { 'num_images': self._info.num_images, 'classes': self._info.classes, 'num_classes': self._info.num_classes, 'num_to_class': self._info.num_to_class, 'class_to_num': self._info.class_to_num, 'data_distributions': {self.name: self._info.num_images}} @classmethod def custom(cls, name, dataset_path = None, classes = None, **kwargs): """Creates an `AgMLDataLoader` with a set of custom data. If you have a custom dataset that you want to use in an `AgMLDataLoader`, this method constructs the loader using similar semantics to the regular loader instantiation. It is a wrapper around using the `meta` argument to provide dataset properties that provides additional convenience for some circumstances, as summarized below. Functionally, this method is equivalent to instantiating `AgMLDataLoader` with an extra argument `meta` that contains metadata for the dataset, with the `task` and `classes` keys required and the others not necessary. This would look like follows: > loader = AgMLDataLoader('name', meta = {'task': task, 'classes': classes}) This method replaces the meta dictionary with keyword arguments to allow for a more Pythonic construction of the custom loader. This method, however includes additional optimizations which allow for a more convenient way to instantiate the loader: 1. It automatically inferences the task from the structure which the data is in, so you don't need to provide the task at all to this method. 2. For image classification and object detection task, this method will attempt to automatically inference the classes in the loader (by searching for the image directories for image classification tasks, and searching in the COCO JSON file for object detection). Semantic segmentation tasks, however, still require the list of classes to be passed. This makes it so that in a variety of cases, the loader can be instantiated without even requiring any metadata, as most of it can be inferred directly by this method and thus streamlines the procedure for using custom data. If you want to cache the metadata, rather than constantly putting them as arguments, then create a file `.meta.json` at the path `/root/.meta.json` with the parameters that you want. Parameters ---------- name : str A name for the custom dataset (this can be any valid string). dataset_path : str, optional A custom path to load the dataset from. If this is not passed, we will assume that the dataset is at the traditional path: `~/.agml/datasets/<name>` (or the changed default data path). Otherwise, the dataset can be passed as a path such as `/root/name`, or `/root`; in the latter case the method will check for `/root/name`. classes : list, tuple A list of string-labels for the classes of the dataset, in order. This is not required for image classification/object detection. kwargs : dict Any other metadata for the dataset, this is not required. Returns ------- An `AgMLDataLoader` outfitted with the custom dataset. """ # Check the name and ensure that no dataset with that name exists. if name in load_public_sources().keys() or not isinstance(name, str): raise ValueError(f"Invalid name '{name}', the name should be " f"a string that is not an existing dataset in " f"the AgML public data source repository.") # Locate the path to the dataset. if dataset_path is None: dataset_path = os.path.abspath(os.path.join(data_save_path(), name)) if not os.path.exists(dataset_path): raise NotADirectoryError( f"Existing directory '{dataset_path}' for dataset of name " f"{name} not found, pass a custom path if you want to use " f"a custom dataset path for the dataset.") else: dataset_path = os.path.abspath(os.path.expanduser(dataset_path)) if not os.path.exists(dataset_path): if not os.path.exists(dataset_path): raise NotADirectoryError( f"Could not find a directory for dataset '{name}' at the " f"provided dataset path: {dataset_path}.") if not dataset_path.endswith(name): dataset_path = os.path.join(dataset_path, name) if not os.path.exists(dataset_path): raise NotADirectoryError( f"Could not find a directory for dataset '{name}' at the " f"provided dataset path: {dataset_path}.") # Infer the task based on the provided dataset path. if os.path.exists(os.path.join(dataset_path, 'annotations.json')): task = 'object_detection' elif os.path.exists(os.path.join(dataset_path, 'images')) and \ os.path.exists(os.path.join(dataset_path, 'annotations')): task = 'semantic_segmentation' elif len(get_file_list(dataset_path)) == 0 and \ len(get_dir_list(dataset_path)) != 0: task = 'image_classification' else: raise TypeError("Unrecognized dataset annotation format.") # Check if there is a metadata file. kwargs['classes'] = classes if os.path.exists(os.path.join(dataset_path, '.meta.json')): with open(os.path.join(dataset_path, '.meta.json'), 'r') as f: kwargs.update(json.load(f)) # Infer the classes for image classification/object detection. classes = kwargs.pop('classes') if classes is None: if task == 'semantic_segmentation': raise ValueError( "Classes are required for a semantic segmentation task.") elif task == 'image_classification': classes = get_dir_list(dataset_path) else: # object detection with open(os.path.join(dataset_path, 'annotations.json'), 'r') as f: classes = [c['name'] for c in json.load(f)['categories']] # Construct and return the `AgMLDataLoader`. return cls(name, dataset_path = dataset_path, meta = {'task': task, 'classes': classes, **kwargs}) @classmethod def helios(cls, name, dataset_path = None): """Creates an `AgMLDataLoader` from a Helios-generated dataset. Given the path to a Helios-generated (and converted) dataset, this method will generate an `AgMLDataLoader` which is constructed using similar semantics to the regular instantiation. This method is largely similar to `AgMLDataLoader.custom()`, but also takes into account the extra information which is provided in the `.metadata` directory of the Helios generated dataset, allowing it to contain potentially even more info. """ # Instantiate from a list of datasets. if isinstance(name, (list, tuple)): if dataset_path is None: dataset_path = [None] * len(name) elif isinstance(dataset_path, str): dataset_path = [dataset_path] * len(name) else: if not len(dataset_path) == len(name): raise ValueError("The number of dataset paths must be " "the same as the number of dataset names.") datasets = [cls.helios(n, dataset_path = dp) for n, dp in zip(name, dataset_path)] return cls.merge(*datasets) # Instantiate from a wildcard pattern. if isinstance(name, str) and '*' in name: if dataset_path is None: dataset_path = os.path.abspath(synthetic_data_save_path()) elif not os.path.exists(dataset_path): raise NotADirectoryError( f"Existing directory '{dataset_path}' for dataset of name " f"{name} not found, pass a custom path if you want to use " f"a custom dataset path for the dataset.") # Get the list of datasets. possible_datasets = glob.glob(os.path.join(dataset_path, name)) if len(possible_datasets) == 0: raise ValueError(f"No datasets found for pattern: {name}.") datasets = [cls.helios(os.path.basename(p), dataset_path = dataset_path) for p in sorted(possible_datasets)] return cls.merge(*datasets) # Locate the path to the dataset, using synthetic semantics. if dataset_path is None: dataset_path = os.path.abspath( os.path.join(synthetic_data_save_path(), name)) if not os.path.exists(dataset_path): raise NotADirectoryError( f"Existing directory '{dataset_path}' for dataset of name " f"{name} not found, pass a custom path if you want to use " f"a custom dataset path for the dataset.") else: dataset_path = os.path.abspath(os.path.expanduser(dataset_path)) if not os.path.exists(dataset_path): if not os.path.exists(dataset_path): raise NotADirectoryError( f"Could not find a directory for Helios dataset '{name}' " f"at the provided dataset path: {dataset_path}.") if not dataset_path.endswith(name): dataset_path = os.path.join(dataset_path, name) if not os.path.exists(dataset_path): raise NotADirectoryError( f"Could not find a directory for Helios dataset '{name}' " f"at the provided dataset path: {dataset_path}.") # Load the information file. info_file = os.path.join(dataset_path, '.metadata', 'agml_info.json') if not os.path.exists(info_file): raise FileNotFoundError(f"The information file at '{info_file}' for the " f"Helios dataset {name} could not be found.") with open(info_file, 'r') as f: meta = json.load(f) # Construct the loader. return cls.custom(name, dataset_path, **meta) @staticmethod def merge(*loaders, classes = None): """Merges a set of `AgMLDataLoader`s into a single loader. Given a set of input `AgMLDataLoader`s, this method will return a single `AgMLDataLoader` which is capable of returning data from any and every one of the input loaders. The resultant loader is functionally equivalent to the `AgMLDataLoader` returned by instantiating an `AgMLDataLoader` from a sequence of AgML public data sources, except that in this case, the input loaders may be subject to a number of input modifications before merging. This also allows the usage of both an AgML public data source and a custom dataset together in a single multi-dataset loader. As such, this method should be used with caution, as since input loaders may be allowed to have any modification, certain methods may not function as expected. For instance, if one of the passed loaders has already been split, then the overall new multi-loader cannot be split as a whole. Similarly, if also using a custom dataset, then any properties of the `info` parameter which are not passed to the dataset cannot be used, even if the other datasets have them. Parameters ---------- loaders : Tuple[AgMLDataLoader] A collection of `AgMLDataLoader`s (but not any `AgMLDataLoader`s which are already holding a collection of datasets). classes : list A list of classes in the new loader. This argument can be used to construct a custom ordering (non-alphabetical) of classes in the loader. Returns ------- A new `AgMLDataLoader` wrapping the input datasets. """ # Validate the input loaders. from agml.data.multi_loader import AgMLMultiDatasetLoader if len(loaders) == 1: raise ValueError("There should be at least two inputs to the `merge` method.") for loader in loaders: if isinstance(loader, AgMLMultiDatasetLoader): raise TypeError("Cannot merge datasets which already hold a " "collection of multiple datasets.") # Instantiate the `AgMLMultiDatasetLoader`. return AgMLMultiDatasetLoader._instantiate_from_collection( *loaders, classes = classes) def __add__(self, other): if not isinstance(other, AgMLDataLoader): return NotImplemented return AgMLDataLoader.merge(self, other) def __len__(self): return self._manager.data_length() def __getitem__(self, indexes: Union[list, int, slice]): if isinstance(indexes, slice): data = np.arange(self._manager.data_length()) indexes = data[indexes].tolist() if isinstance(indexes, int): indexes = [indexes] if np.isscalar(indexes): indexes = [indexes.item()] # noqa for idx in indexes: if idx not in range(len(self)): if idx not in [-i for i in range(1, len(self) + 1, 1)]: raise IndexError( f"Index {idx} out of range of " f"AgMLDataLoader length: {len(self)}.") return self._manager.get(resolve_list_value(indexes)) def __iter__(self): for indx in range(len(self)): yield self[indx] def __repr__(self): out = f"<AgMLDataLoader: (dataset={self.name}" out += f", task={self.task}" out += f", images={self.num_images}" out += f") at {hex(id(self))}>" return out def __str__(self): return repr(self) def __copy__(self): """Copies the loader and updates its state.""" cp = super(AgMLDataLoader, self).__copy__() cp.copy_state(self) return cp def copy(self): """Returns a deep copy of the data loader's contents.""" return self.__copy__() def copy_state(self, loader): """Copies the state of another `AgMLDataLoader` into this loader. This method copies the state of another `AgMLDataLoader` into this loader, including its transforms, resizing, and training state. Other general parameters such as batch size and shuffling are left intact. Parameters ---------- loader : AgMLDataLoader The data loader from which the state should be copied. Returns ------- This `AgMLDataLoader`. """ # Re-construct the training manager. new_train_manager = loader._manager._train_manager.__copy__() self._manager._train_manager = new_train_manager # Re-construct the transform manager. new_transform_manager = loader._manager._transform_manager.__copy__() self._manager._transform_manager = new_transform_manager self._manager._train_manager._transform_manager = new_transform_manager # Re-construct the resizing manager. new_resize_manager = loader._manager._resize_manager.__copy__() self._manager._resize_manager = new_resize_manager self._manager._train_manager._resize_manager = new_resize_manager @property def name(self): """Returns the name of the dataset in the loader.""" return self._info.name @property def dataset_root(self): """Returns the local path to the dataset being used.""" return self._builder.dataset_root @property def info(self): """Returns a `DatasetMetadata` object containing dataset info. The contents returned in the `DatasetMetadata` object can be used to inspect dataset metadata, such as the location the data was captured, the data formats, and the license/copyright information. See the `DatasetMetadata` class for more information. """ return self._info @property def task(self): """Returns the ML task that this dataset is constructed for.""" return self._info.tasks.ml @property def num_images(self): """Returns the number of images in the dataset.""" return self._meta_properties.get('num_images') @property def classes(self): """Returns the classes that the dataset is predicting.""" return self._meta_properties.get('classes') @property def num_classes(self): """Returns the number of classes in the dataset.""" return self._meta_properties.get('num_classes') @property def num_to_class(self): """Returns a mapping from a number to a class label.""" return self._meta_properties.get('num_to_class') @property def class_to_num(self): """Returns a mapping from a class label to a number.""" return self._meta_properties.get('class_to_num') @property def data_distributions(self): """Displays the distribution of images from each source.""" return self._meta_properties.get('data_distributions') @property def image_size(self): """Returns the determined image size for the loader. This is primarily useful when using auto shape inferencing, to access what the final result ends up being. Otherwise, it may just return `None` or the shape that the user has set. """ return self._manager._resize_manager.size def _generate_split_loader( self, contents, split, meta_properties = None, **kwargs): """Generates a split `AgMLDataLoader`.""" # Check if the data split exists. if contents is None: raise ValueError( f"Attempted to access '{split}' split when " f"the data has not been split for '{split}'.") # Load a new `DataManager` and update its internal managers # using the state of the existing loader's `DataManager`. builder = DataBuilder.from_data( contents = [contents, kwargs.get('labels_for_image', None)], info = self.info, root = self.dataset_root, builder = self._builder) current_manager = copy.deepcopy(self._manager.__getstate__()) current_manager.pop('builder') current_manager['builder'] = builder # Build the new accessors and construct the `DataManager`. accessors = np.arange(len(builder.get_contents())) if self._manager._shuffle: np.random.shuffle(accessors) current_manager['accessors'] = accessors batch_size = current_manager.pop('batch_size') current_manager['batch_size'] = None new_manager = DataManager.__new__(DataManager) new_manager.__setstate__(current_manager) # After the builder and accessors have been generated, we need # to generate a new list of `DataObject`s. new_manager._create_objects( new_manager._builder, self.task) # Update the `TransformManager` and `ResizeManager` of the # `TrainManager` in the `DataManager` (they need to be synchronized). new_manager._train_manager._transform_manager = \ new_manager._transform_manager new_manager._train_manager._resize_manager = \ new_manager._resize_manager # Batching data needs to be done independently. if batch_size is not None: new_manager.batch_data(batch_size = batch_size) # Update the metadata parameters. if meta_properties is None: meta_properties = self._meta_properties.copy() meta_properties['num_images'] = len(contents) meta_properties['data_distributions'] = { self.name: len(contents)} # Instantiate a new `AgMLDataLoader` from the contents. loader_state = self.copy().__getstate__() loader_state['builder'] = builder loader_state['manager'] = new_manager loader_state['meta_properties'] = meta_properties cls = super(AgMLDataLoader, self).__new__(AgMLDataLoader) cls.__setstate__(loader_state) for attr in ['train', 'val', 'test']: setattr(cls, f'_{attr}_data', None) cls._is_split = True return cls @property def train_data(self): """Stores the `train` split of the data in the loader.""" if isinstance(self._train_data, AgMLDataLoader): return self._train_data self._train_data = self._generate_split_loader( self._train_data, split = 'train') return self._train_data @property def val_data(self): """Stores the `val` split of the data in the loader.""" if isinstance(self._val_data, AgMLDataLoader): return self._val_data self._val_data = self._generate_split_loader( self._val_data, split = 'val') return self._val_data @property def test_data(self): """Stores the `test` split of the data in the loader.""" if isinstance(self._test_data, AgMLDataLoader): return self._test_data self._test_data = self._generate_split_loader( self._test_data, split = 'test') return self._test_data def eval(self) -> "AgMLDataLoader": """Sets the `AgMLDataLoader` in evaluation mode. Evaluation mode disables transforms, and only keeps the loader applying resizing to the contents. If the loader was previously set into TensorFlow or PyTorch mode, however, it will also keep up tensor conversion and potential batch adding (see `as_keras_sequence()` and `as_torch_dataset()` methods for more information on the exact operations). This method does not completely disable preprocessing, to completely disable preprocessing, use `loader.disable_preprocessing()`. Additionally, if you want to keep only the resizing but not the implicit tensor conversions based on the backend, then run: > loader.disable_preprocessing() # or loader.reset_preprocessing() > loader.eval() This will refresh the backend conversions and return it to `eval` mode. Returns ------- The `AgMLDataLoader` object. """ self._manager.update_train_state('eval') return self def disable_preprocessing(self) -> "AgMLDataLoader": """Disables all preprocessing on the `AgMLDataLoader`. This sets the loader in a no-preprocessing mode (represented internally as `False`), where only the raw data is returned: no transforms, resizing, or any conversion to any type of backend. This can be used to test or inspect the original data contents of the loader before processing. The loader can be set into any mode from here, for instance see `eval()`, `as_keras_sequence()`, and `as_torch_dataset()` for specific examples on the different potential training and evaluation states. If you just want to reset the loader to its default state, which applies only transforms and resizing, then use `loader.reset_preprocessing()`. Returns ------- The `AgMLDataLoader` object. """ self._manager.update_train_state(False) return self def reset_preprocessing(self) -> "AgMLDataLoader": """Re-enables preprocessing on the `AgMLDataLoader`. This resets the loader back to its default train state, namely where it applies just the given transforms and content resizing. This is a consistent method, meaning that regardless of the prior train state of the loader before running this method, it will hard reset it to its original state (similar to `disable_preprocessing()`, but it keeps some preprocessing). Returns ------- The `AgMLDataLoader` object. """ self._manager.update_train_state(None) return self def on_epoch_end(self): """Shuffles the dataset on the end of an epoch for a Keras sequence. If `as_keras_sequence()` is called and the `AgMLDataLoader` inherits from `tf.keras.utils.Sequence`, then this method will shuffle the dataset on the end of each epoch to improve training. """ self._manager._maybe_shuffle() def as_keras_sequence(self) -> "AgMLDataLoader": """Sets the `DataLoader` in TensorFlow mode. This TensorFlow extension converts the loader into a TensorFlow mode, adding inheritance from the superclass `keras.utils.Sequence` to enable it to be used directly in a Keras pipeline, and adding extra preprocessing to the images and annotations to make them compatible with TensorFlow. The main features added on enabling this include: 1. Conversion of output images and annotations to `tf.Tensor`s. 2. Adding an implicit batch size dimension to images even when the data is not batched (for compatibility in `Model.fit()`). 3. Adding inheritance from `keras.utils.Sequence` so that any `AgMLDataLoader` object can be used directly in `Model.fit()`. 4. Setting the data loader to use a constant image shape, namely `auto` (which will default to (512, 512) if none is found). This can be overridden by manually setting the image shape parameter back after running this method. Note that this may result in errors when attempting implicit tensor conversion. Returns ------- The `AgMLDataLoader` object. """ _add_dataset_to_mro(self, 'tf') self._manager.update_train_state('tf') return self def as_torch_dataset(self) -> "AgMLDataLoader": """Sets the `DataLoader` in PyTorch mode. This PyTorch extension converts the loader into a PyTorch mode, adding inheritance from th superclass `torch.utils.data.Dataset` to enable it to be used directly in a PyTorch pipeline, and adding extra preprocessing to the images and annotations to make them compatible with PyTorch. The main features added on enabling this include: 1. Conversion of output images and annotations to `torch.Tensor`s. 2. Converting the channel format of the input images from the default, channels_last, into channels_first (NHWC -> NCHW). 3. Adding inheritance from `torch.utils.data.Dataset` so that any `AgMLDataLoader` object can be used with a `torch.utils.data.DataLoader`. 4. Setting the data loader to use a constant image shape, namely `auto` (which will default to (512, 512) if none is found). This can be overridden by manually setting the image shape parameter back after running this method. Note that this may result in errors when attempting implicit tensor conversion. Returns ------- The `AgMLDataLoader` object. """ _add_dataset_to_mro(self, 'torch') self._manager.update_train_state('torch') return self @property def shuffle_data(self): """Returns whether the loader is set to shuffle data or not. By default, if no value is passed in initialization, this is set to `True`. It can be manually toggled to `False` using this property. """ return self._manager._shuffle @shuffle_data.setter def shuffle_data(self, value): """Set whether the loader should shuffle data or not. This can be used to enable/disable shuffling, by passing either `True` or `False`, respectively. """ if not isinstance(value, bool): raise TypeError("Expected either `True` or `False` for 'shuffle_data'.") self._manager._shuffle = value def shuffle(self, seed = None): """Potentially shuffles the contents of the loader. If shuffling is enabled on this loader (`shuffle = False` has not been passed to the instantiation), then this method will shuffle the order of contents in it. A seed can be provided to shuffle the dataset to an expected order. If the data is already batched, then the batch contents will be shuffled. For instance, if we have data batches [[1, 2], [3, 4]], then the shuffling result will be [[3, 4], [1, 2]]. If you want all of the contents to be shuffled, call `shuffle` before batching. Note that the data is automatically shuffled upon instantiation, unless the `shuffle = False` parameter is passed at instantiation. However, this disables automatic shuffling for the class permanently, and this method must be called to shuffle the data. Parameters ---------- seed : int, optional A pre-determined seed for shuffling. Returns ------- The `AgMLDataLoader` object. """ self._manager.shuffle(seed = seed) return self def take_dataset(self, name) -> "AgMLDataLoader": """Takes one of the datasets in a multi-dataset loader. This method selects one of the datasets (as denoted by `name`) in this multi-dataset collection and returns an `AgMLDataLoader` with its contents. These contents will be subject to any transforms and modifications as applied by the main loader, but the returned loader will be a copy, such that any new changes made to the main multi-dataset loader will not affect the new loader. Note that this method only works for multi-dataset collections. Parameters ---------- name : str The name of one of the sub-datasets of the loader. Returns ------- An `AgMLDataLoader`. """ raise ValueError( "The `loader.take_dataset` method only works for multi-dataset loaders.") def take_class(self, classes, reindex = True) -> "AgMLDataLoader": """Reduces the dataset to a subset of class labels. This method, given a set of either integer or string class labels, will return a new `AgMLDataLoader` containing a subset of the original dataset, where the only classes in the dataset are those specified in the `classes` argument. The new loader will have info parameters like `num_classes` and `class_to_num` updated for the new set of classes; however, the original `info` metadata will remain the same as the original. Note that if the dataset contains images which have bounding boxes corresponding to multiple classes, this method will not work. Parameters ---------- classes : list, int, str Either a single integer/string for a single class, or a list of integers or strings for multiple classes. Integers should be one-indexed for object detection. reindex : bool Re-indexes all of the new classes starting from 1, in ascending order based on their number in the original dataset. Notes ----- This method only works for object detection datasets. """ if self._info.tasks.ml != 'object_detection': raise RuntimeError("The `take_class` method can only be " "used for object detection datasets.") # Parse the provided classes and determine their numerical labels. if isinstance(classes, str): if classes not in self.classes: raise ValueError( f"Received a class '{classes}' for `loader.take_class`, " f"which is not in the classes for {self.name}: {self.classes}") classes = [self.class_to_num[classes]] elif isinstance(classes, int): try: self.num_to_class[classes] except IndexError: raise ValueError( f"The provided class number {classes} is out of " f"range for {self.num_classes} classes. Make sure " f"you are using zero-indexing.") classes = [classes] else: parsed_classes = [] if isinstance(classes[0], str): for cls in classes: if cls not in self.classes: raise ValueError( f"Received a class '{cls}' for `loader.take_class`, which " f"is not in the classes for {self.name}: {self.classes}") parsed_classes.append(self.class_to_num[cls]) elif isinstance(classes[0], int): for cls in classes: try: self.num_to_class[cls] except IndexError: raise ValueError( f"The provided class number {cls} is out of " f"range for {self.num_classes} classes. Make " f"sure you are using zero-indexing.") parsed_classes.append(cls) classes = parsed_classes.copy() # Ensure that there are no images with multi-category boxes. categories = self._builder._labels_for_image if not all(len(np.unique(c)) == 1 for c in categories.values()): raise ValueError( f"Dataset {self.name} has images with multiple categories for " f"bounding boxes, cannot take an individual set of classes.") # Get the new data which will go in the loader. The `DataBuilder` # stores a mapping of category IDs corresponding to the bounding # boxes in each image, so we use these to determine the new boxes. new_category_map = { k: v for k, v in categories.items() if v[0] in classes} new_coco_map = { k: v for k, v in self._builder._data.items() if k in new_category_map.keys()} # Create the new info parameters for the class. If reindexing # is requested, then we re-index the classes based on the order # in which they are given, and then create a new dictionary # to map the original annotations to the new ones (used later). if reindex: old_to_new = {cls: idx + 1 for idx, cls in enumerate(classes)} new_classes = [self.num_to_class[c] for c in classes] new_properties = { 'num_images': len(new_coco_map.keys()), 'classes': new_classes, 'num_classes': len(new_classes), 'num_to_class': {i + 1: c for i, c in enumerate(new_classes)}, 'class_to_num': {c: i + 1 for i, c in enumerate(new_classes)}} else: new_classes = [self.num_to_class[c] for c in classes] new_properties = { 'num_images': len(new_coco_map.keys()), 'classes': new_classes, 'num_classes': len(classes), 'num_to_class': {c: self.num_to_class[c] for c in classes}, 'class_to_num': {self.num_to_class[c]: c for c in classes}} # Create the new loader. obj = self._generate_split_loader( new_coco_map, 'train', meta_properties = new_properties, labels_for_image = new_category_map) obj._is_split = False # Re-index the loader if requested to. if reindex: class AnnotationRemap(AgMLSerializable): """A helper class to remap annotation labels for multiple datasets.""" serializable = frozenset(("map",)) def __init__(self, o2n): self._map = o2n def __call__(self, contents, name): """Re-maps the annotation for the new, multi-dataset mapping.""" image, annotations = contents # Re-map the annotation ID. category_ids = annotations['category_id'] category_ids[np.where(category_ids == 0)[0]] = 1 # fix new_ids = np.array([self._map[c] for c in category_ids]) annotations['category_id'] = new_ids return image, annotations # Maps the annotations. obj._manager._train_manager._set_annotation_remap_hook( AnnotationRemap(old_to_new)) # noqa # Return the loader. return obj @inject_random_state def take_random(self, k) -> "AgMLDataLoader": """Takes a random set of contents from the loader. This method selects a sub-sample of the contents in the loader, based on the provided number of (or proportion of) elements `k`. It then returns a new loader with just this reduced number of elements. The new loader is functionally similar to the original loader, and contains all of the transforms/batching/other settings which have been applied to it up until this method is called. Note that the data which is sampled as part of this new loader is not removed from the original loader; this simply serves as an interface to use a random set of images from the full dataset. Parameters ---------- k : int, float Either an integer specifying the number of samples or a float specifying the proportion of images from the total to take. {random_state} Returns ------- A reduced `AgMLDataLoader` with the new data. """ # Parse the input to an integer. if isinstance(k, float): # Check that 0.0 <= k <= 1.0. if not 0.0 <= k <= 1.0: raise ValueError( "If passing a proportion to `take_class`, " "it should be in range [0.0, 1.0].") # Convert the proportion float to an absolute int. Note that # the method used is rounding up to the nearest int for cases # where there is not an exact proportional equivalent. getcontext().prec = 4 # noqa proportion = Decimal(k) / Decimal(1) num_images = self.num_images k = int(proportion * num_images) # If the input is an integer (or the float is converted to an int # above), then select a random sampling of images from the dataset. if isinstance(k, int): # Check that `k` is valid for the number of images in the dataset. if not 0 <= k <= self.num_images: raise ValueError( f"Received a request to take a random sampling of " f"{k} images, when the dataset has {self.num_images}.") # We use a similar functionality to the `split` method here, # essentially choosing a random sampling up until `k` and then # using the `DataManager` to access the reduced data. split = np.arange(0, self.num_images) np.random.shuffle(split) indices = split[:k] content = list(self._manager.generate_split_contents( {'content': indices}).values())[0] # Create a new `AgMLDataLoader` from the new contents. obj = self._generate_split_loader(content, 'train') obj._is_split = False return obj # Otherwise, raise an error. else: raise TypeError( f"Expected only an int or a float when " f"taking a random split, got {type(k)}.") @inject_random_state def split(self, train = None, val = None, test = None, shuffle = True): """Splits the data into train, val and test splits. By default, this method does nothing (or if the data has been split into sets, it resets them all to one set). Setting the `train`, `val`, and `test` parameters randomly divides the data into train, validation, and/or test sets, depending on which ones are provided and their values. Values can either be passed as exact numbers or as proportions, e.g. either `train = 80, test = 20` in a 100-value dataset, or as `train = 0.8, test = 0.2`. Whichever value is not passed, e.g. `val` in this case, has no value in the loader. Parameters ---------- train : int, float The split for training data. val : int, float The split for validation data. test : int, float The split for testing data. shuffle : bool Whether to shuffle the split data. {random_state} Notes ----- Any processing applied to this `AgMLDataLoader` will also be present in the split loaders until they are accessed from the class. If you don't want these to be applied, access them right after splitting. """ # Check if the data is already split or batched. if not AgMLExperimentalFeatureWrapper.nested_splitting(): if self._is_split: raise ValueError("Cannot split already split data.") elif self._manager._batch_size is not None: raise ValueError("Cannot split already batched data. " "Split the data before batching.") # If no parameters are passed, then don't do anything. arg_dict = {'train': train, 'val': val, 'test': test} valid_args = {k: v for k, v in arg_dict.items() if v is not None} if all(i is None for i in arg_dict.values()): return None # There are two valid ways to pass splits. The first involves # passing the split values as floats, the second as ints. If we # receive the splits as floats, then we convert them to ints # in order to maintain maximum precision if we do manage to get # them as ints. Then the procedure is the same. if all(isinstance(i, float) for i in valid_args.values()): # To prevent potential precision errors, we need to convert the # splits to `Decimal` objects and then set the decimal precision. getcontext().prec = 4 # noqa valid_args = {k: Decimal(v) / Decimal(1) for k, v in valid_args.items()} if not sum(valid_args.values()) == Decimal(1): raise ValueError(f"Got floats for input splits and expected a sum " f"of 1, instead got {sum(valid_args.values())}.") # Convert the splits from floats to ints. If the sum of the int # splits are greater than the total number of data, then the largest # split is decreased in order to keep compatibility in usage. num_images = self.num_images proportions = {k: int(v * Decimal(num_images)) for k, v in valid_args.items()} if sum(proportions.values()) != num_images: diff = sum(proportions.values()) - num_images largest_split = list(proportions.keys())[ list(proportions.values()).index( max(proportions.values()))] proportions[largest_split] = proportions[largest_split] - diff valid_args = proportions.copy() # Create the actual data splits. if all(isinstance(i, int) for i in valid_args.values()): # Ensure that the sum of the splits is the length of the dataset. if not sum(valid_args.values()) == self.num_images: raise ValueError(f"Got ints for input splits and expected a sum " f"equal to the dataset length, {self.num_images}," f"but instead got {sum(valid_args.values())}.") # The splits will be generated as sequences of indices. generated_splits = {} split = np.arange(0, self.num_images) names, splits = list(valid_args.keys()), list(valid_args.values()) # Shuffling of the indexes will occur first, such that there is an even # shuffle across the whole sample and not just in the individual splits. if shuffle: np.random.shuffle(split) # Create the list of split indexes. if len(valid_args) == 1: generated_splits[names[0]] = split elif len(valid_args) == 2: generated_splits = {k: v for k, v in zip( names, [split[:splits[0]], split[splits[0]:]])} else: split_1 = split[:splits[0]] split_2 = split[splits[0]: splits[0] + splits[1]] split_3 = split[splits[0] + splits[1]:] generated_splits = {k: v for k, v in zip( names, [split_1, split_2, split_3])} # Get the actual split contents from the manager. These contents are # not `DataObject`s, rather they are simply the actual mapping of # data to be passed to a `DataBuilder` when constructing the splits. contents = self._manager.generate_split_contents(generated_splits) # Build new `DataBuilder`s and `DataManager`s for the split data. for split, content in contents.items(): setattr(self, f'_{split}_data', content) # Otherwise, raise an error for an invalid type. else: raise TypeError( "Expected either only ints or only floats when generating " f"a data split, got {[type(i) for i in arg_dict.values()]}.") def batch(self, batch_size = None): """Batches sets of image and annotation data according to a size. This method will group sets of data together into batches of size `batch_size`. In turn, items gathered from the loader will, rather than being an image and annotation, be an array of images and an array of annotations (not an array of image/annotation pairs). Batching data will include a `batch` dimension for the images and annotations that are returned (e.g., the image array will have dimensions NHWC instead of HWC). If the data is not batched, this dimension will not be present unless the loader is in training mode. The data can be un-batched by passing `None` to batch size (or calling the method with no arguments). Parameters ---------- batch_size : int, None The number of groups to batch data together into. Notes ----- The last batch will be of size <= `batch_size`. """ self._manager.batch_data( batch_size = batch_size ) def resize_images(self, image_size = None, method = 'bilinear'): """Resizes images within the loader to a specified size. This method applies a resizing parameter for images before they are returned from the data loader. The default starting point, if this method is never called, is to apply no resizing. If the loader is set in "training" mode and no size is specified, it defaults to (512, 512). Image resizing contains a few modes: 1. `default` or `None`: No resizing, leaves images in default size. This is the default parameter if nothing is passed. 2. `train`: This will set a default training size of (512, 512). 3. `imagenet`: This will set a default size of (224, 224). 4. custom size: Resizes the images to the provided size. 5. `auto`: Dynamically selects an image size based on a few factors. For example, if there are certain outliers in a dataset which are of a different size while the majority remain the same, then, the behavior of this method is chosen by a majority threshold of the sizes of all the images in the dataset. If no shape can be inferenced, it returns a default size of (512, 512). The resizing also applies to the annotation in certain cases, depending on the task and the actual content of the annotation: - For object detection, the bounding box coordinates will be resized and the area of the box will in turn be recomputed. - For semantic segmentation, the annotation mask will be resized, using a nearest-neighbor interpolation to keep it as similar as possible to the original mask (preventing data loss). Parameters ---------- image_size : optional The resizing parameter for the image. method : optional The method to resize the images. Should be one of 'nearest', 'bilinear', 'linear', or 'cubic'. Defaults to 'bilinear'. Notes ----- If a transform pipeline is provided, images will be resized *before* being passed into the transform pipeline. """ self._manager.assign_resize( image_size = image_size, method = method ) def transform(self, transform = NoArgument, target_transform = NoArgument, dual_transform = NoArgument): """Applies vision transforms to the input image and annotation data. This method applies transformations to the image and annotation data in the dataset. Transforms include augmentations and other processing methods, and can be applied independently to the image and annotation, or together to both (`transform`, `target_transform`, `dual_transform`). The hierarchy in which transforms are applied is: transform -> --------| |-----> dual_transform target_transform -> --| The `transform` and `target_transform` argument are used for methods which act independently on the image and the annotation, respectively. The values passed to these arguments can be: - An `albumentations` transform pipeline. - A `keras.Sequential` model (or preprocessing layer) or a set of `torchvision.transform`s. - A method which accepts one input and returns one output. The `dual_transform` argument is used for non-image-classification tasks. The following describe the types of arguments that can be passed to `dual_transform`, depending on the task: Object Detection: - An `albumentations` transform pipeline with `bbox_params` in to be applied to both the image and the bounding boxes. - A method (not a torchvision or Keras preprocessing pipeline) that accepts two inputs and returns two outputs. Semantic Segmentation: - An `albumentations` transform pipeline that may include spatial and/or visual augmentation. - A method to independently or dually apply transformations to the image and annotation mask. - A `torchvision.transforms` or `tf.keras.Sequential` pipeline which will be applied to the image and mask using the same random seed, for reproducibility. Use the provided method `generate_keras_segmentation_dual_transform` for this. If you want to reset the transforms, then simply call this method with no arguments. Alternatively, to reset just a single transform, pass the value of that argument as `None`. Parameters ---------- transform : optional A transform to be applied independently to the input image. target_transform : optional A transform to be applied independently to the annotation. dual_transform : optional A transform to be applied to both the input and annotation. Notes ----- - Image resizing takes place before any transformations are applied. After the transforms are applied in this order, they returned and if passed again, they will have a different transform applied to them. The state is independent of the images passed. - Albumentations transforms are special in that even transforms which would normally be passed to `dual_transform` (e.g., they act on the input image and the output annotation) can simply be passed to the `transform` argument and they will automatically be applied. """ self._manager.push_transforms( transform = transform, target_transform = target_transform, dual_transform = dual_transform ) def normalize_images(self, method = 'scale'): """Converts images from 0-255 integers to 0-1 floats and normalizes. This is a convenience method to convert all images from integer-valued arrays into float-valued arrays, and normalizes them (using shifting and scaling from mean and std). This is useful for training in order to reduce computational complexity (instead of large-valued integer multiplication, only float multiplication), and for extracting the most information out of different types of imagery. There are three different 'normalization' modes that can be initialized with this method, as described below: 1. `scale`: This simply scales images from the 0-255 pixel range to the 0-1 range (and converts them to floats as such). 2. `imagenet`: This performs normalization using the traditional ImageNet mean and standard deviation: (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]) This is useful when trying to conduct transfer learning, for instance. 3. `standard`: This performs normalization using a pre-calculated mean and standard deviation for the dataset (see the public sources JSON). To remove normalization altogether, pass `None` as a parameter. Parameters ---------- method : str The method by which to normalize the images. Notes ----- This method is not implicitly called when converting to PyTorch/TensorFlow mode, it needs to be manually called even if you just want 0-1 scaled images. """ if method not in ['scale', 'imagenet', 'standard', None]: raise ValueError(f"Received invalid normalization method: '{method}'.") if method == 'scale': normalization_params = 'scale' elif method == 'imagenet': normalization_params = 'imagenet' elif method == 'standard': normalization_params = self._info.image_stats else: normalization_params = None self.transform( transform = ('normalize', normalization_params) ) def labels_to_one_hot(self, add = True): """Converts image classification numerical labels to one-hot labels. This is a convenience method to apply one-hot vector transformations to the output labels for image classification. Essentially, if we have a set of labels, [1, 2], it will convert it to [[0, 1, 0], [0, 0, 1]]. This is a more commonly used format for image classification. Parameters ---------- add : bool If set to `None` or `False`, this will remove the one-hot label transformation from the manager. This is `True` by default, which adds the one-hot label transformation. """ if self._info.tasks.ml != 'image_classification': raise RuntimeError("The `one_hot` label transformation can only " "be used for image classification tasks.") self.transform( target_transform = ('one_hot', self._info.num_classes, add) ) def mask_to_channel_basis(self, add = True): """Converts semantic segmentation masks to channel-wise. This is a convenience method to convert integer-labeled semantic segmentation masks into channel-by-channel masks, essentially one-hot vector transformation but for semantic segmentation. Note that if the task is binary segmentation, e.g. there is only one class, then this method will do nothing. This method should traditionally be called *after* applying general transformations to the loader, in order to prevent any issues. Parameters ---------- add : bool If set to `None` or `False`, this will remove the one-hot label transformation from the manager. This is `True` by default, which adds the one-hot label transformation. """ if self._info.tasks.ml != 'semantic_segmentation': raise ValueError("The `mask_to_channel_basis` transformation " "can only be used for semantic segmentation tasks.") # Warn about binary segmentation tasks. if self._info.num_classes == 1: log(f"No mask-to-channel transformation will be applied for " f"a binary segmentation task (dataset {self.name}).") return self.transform( target_transform = ('channel_basis', self._info.num_classes, add) ) def generalize_class_detections(self): """Generalizes object detection classes to a single class. This is a convenience method for object detection tasks, and converts all of the individual class labels in the task into a single class, essentially allowing the model to purely focus on detection of objects and fine-tuning bounding boxes, with no focus on differentiating classes of different boxes. This method is intended to be used for multi-dataset loaders, and will raise an error if using with a single-dataset loader. """ raise ValueError( "This method can only be used with multi-dataset loaders.") def export_contents(self, export_format = None): """Exports the internal contents of the `AgMLDataLoader`. This method serves as a hook for high-level users who simply want to download and get the data, by exporting the unprocessed metadata of the actual dataset, with the following formats: Image Classification: A mapping between the local image paths and the numerical labels. Object Detection: A mapping between the local image paths (the full path, not just the file name), and the COCO JSON annotations corresponding to each of the images. To get the original COCO JSON annotation file contents, use `export_format = 'coco'`. Semantic Segmentation: A mapping between the local image paths and the local annotation mask paths. The `export_format` argument can be used to customize what this method returns. By default, it is set to `None`, and returns dictionaries with the above specified mappings. However, setting `export_format = 'arrays'` will return two arrays, with the first array containing the image paths and the second array containing the annotation data. Parameters ---------- export_format : optional The format to export the data in. Defaults to a mapping. Returns ------- The raw contents of the dataset. """ return self._builder.export_contents( export_format = export_format ) def export_tensorflow(self): """Exports the contents of the loader in a native TensorFlow dataset. This method constructs a `tf.data.Dataset` from the contents of the dataset. The dataset maintains the same loading and preprocessing pipeline as the actual `DataLoader`, but allows for faster computation and integration into a TensorFlow pipeline. When constructing the `tf.Data.Dataset`, the `AgMLDataLoader` uses the pre-set parameters of the class, including the transforms, image resizing, and the training mode. In particular, the tensor conversion and automatic batching is done inherently by the `tf.data.Dataset`, but transforms can be disabled in the same way they would be in a normal `AgMLDataLoader` by running > loader.disable_preprocessing() > loader.export_tensorflow() or potentially > loader.eval() > loader.export_tensorflow() Image resizing is an exception in that if a specific size is not set, then it will automatically be set to (512, 512) to prevent errors. The same behavior of transforms and image resizing applies also to batching. Calling the relevant methods before exporting the dataset will result in those methods being applied to the result. The only exception is shuffling, since the data is always shuffled upon being exported to a `tf.data.Dataset`. This enables better computation. Note that if the data is batched, then it is also prefetched. Please note, transforms will not be applied if exporting an object detection loader. This is due to the impossibility of applying transforms to COCO JSON dictionaries and passing them in TensorFlow's graph mode. Use `as_keras_sequence` if you want to use transforms. Returns ------- A `tf.data.Dataset` enabled to function like the `AgMLDataLoader`, but as a native TensorFlow object for TensorFlow pipelines. """ # Update the backend management system. from agml.data.exporters.tensorflow import TFExporter if get_backend() != 'tf': if user_changed_backend(): raise StrictBackendError( change = 'tf', obj = self.export_tensorflow) set_backend('tf') # Build the exporter. exporter = TFExporter( task = self.info.tasks.ml, builder = self._builder ) # Update the current state of the loader. exporter.assign_state(state = self._manager._train_manager.state) # Parse the transforms and resizing for the class. transforms = self._manager._transform_manager.get_transform_states() resizing = self._manager._resize_manager.size exporter.digest_transforms( transforms = transforms, resizing = resizing ) # Construct and return the loader. return exporter.build(batch_size = self._manager._batch_size) def export_torch(self, **loader_kwargs): """Exports the contents of the loader in a native PyTorch loader. This method wraps the contents of this data loader inside of a `torch.utils.data.DataLoader`. This method differs from the `export_tensorflow()` method in that there is no need to convert directly to a `tf.data.Dataset`, rather if this `AgMLDataLoader` inherits from `torch.utils.data.Dataset`, it can just be directly wrapped into a `torch.utils.data.DataLoader`. The returned `DataLoader` is functionally similar to the `AgMLDataLoader` in terms of preprocessing and transforming. You can pass arguments to the `DataLoader` instantiation as keyword arguments to this method. Note that the `AgMLDataLoader` which this method encloses is instead a copy of the instance the method is run on, so that any changes to the loader afterwards don't affect the exported loader. Parameters ---------- loader_kwargs : optional A set of keyword arguments for the `torch.utils.data.DataLoader`. See the documentation for the loader for more information. Returns ------- A `torch.utils.data.DataLoader` enclosing a copy of this loader. """ from agml.backend.tftorch import torch from torch.utils.data import DataLoader if get_backend() != 'torch': if user_changed_backend(): raise StrictBackendError( change = 'torch', obj = self.export_torch) set_backend('torch') # Make a copy of the `AgMLDataLoader` so the following changes # don't affect the original loader, just the new one. obj = self.copy() # Convert to a PyTorch dataset. obj.as_torch_dataset() # The `DataLoader` automatically batches objects using its # own mechanism, so we remove batching from this DataLoader. batch_size = loader_kwargs.pop( 'batch_size', obj._manager._batch_size) obj.batch(None) shuffle = loader_kwargs.pop( 'shuffle', obj._manager._shuffle) # The `collate_fn` for object detection is different because # the COCO JSON dictionaries each have different formats. So, # we need to replace it with a custom function. collate_fn = loader_kwargs.pop('collate_fn') if obj.task == 'object_detection' and collate_fn is None: def collate_fn(batch): images = torch.stack( [i[0] for i in batch], dim = 0) coco = tuple(zip(*[i[1] for i in batch])) return images, coco # Return the DataLoader with a copy of this AgMLDataLoader, so # that changes to this will not affect the returned loader. return DataLoader( obj, batch_size = batch_size, shuffle = shuffle, collate_fn = collate_fn, **loader_kwargs ) def show_sample(self, image_only = False, no_show = False): """Shows a single data sample from the dataset. This method generates a data sample from the dataset with an image and its corresponding annotation (or, if `image_only` is True, then only the image itself). This data sample is then displayed, unless `no_show` is True in which case the processed sample will simply be returned. Parameters ---------- image_only : optional Whether to show only the image or the image and the annotation. no_show : optional Whether to display the sample or not. Returns ------- The data sample with/without annotation. """ # Get the sample (and take only the first one in a batch if batched). image, annotations = self[self._manager._get_random_index()] if len(image.shape) == 4: image = image[0] annotations = annotations[0] # Show the sample. show_sample(self, image_only = image_only, no_show = no_show, sample = (image, annotations))
75,355
43.828079
90
py
AgML
AgML-main/agml/data/multi_loader.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import types import collections from typing import Union from decimal import Decimal, getcontext import numpy as np from agml.framework import AgMLSerializable from agml.data.metadata import DatasetMetadata from agml.data.loader import AgMLDataLoader from agml.utils.general import ( resolve_list_value, NoArgument ) from agml.utils.random import seed_context, inject_random_state from agml.utils.image import consistent_shapes from agml.utils.logging import log from agml.backend.tftorch import ( get_backend, set_backend, user_changed_backend, StrictBackendError, is_array_like, convert_to_batch ) class CollectionWrapper(AgMLSerializable): """Wraps a collection of items and calls their attributes and methods.""" serializable = frozenset(('collection', 'keys')) def __init__(self, collection, keys = None, ignore_types = False): self._collection = collection if not ignore_types: if not all(isinstance(c, type(collection[0])) for c in collection): raise TypeError( f"Items in a collection should all be of the " f"same type, got {[type(i) for i in collection]}") self._keys = keys def __len__(self): return len(self._collection) def __getitem__(self, item): if isinstance(item, str): return self._collection[self._keys.index(item)] return self._collection[item] @property def keys(self): return self._keys def get_attributes(self, attr): return [getattr(c, attr) for c in self._collection] def call_method(self, method, args = None, kwargs = None): if kwargs is None: kwargs = {} if args is None: args = () elif isinstance(args, tuple): if not len(args) == len(self._collection): raise IndexError( f"Got {len(args)} unique arguments for a " f"collection of length {len(self._collection)}.") return [getattr(c, method)(arg, **kwargs) for c, arg in zip(self._collection, args)] return [getattr(c, method)(*args, **kwargs) for c in self._collection] def apply(self, method, args = None): if args is None: for c in self._collection: method(c) else: for c, arg in zip(self._collection, args): method(c, arg) class MultiDatasetMetadata(AgMLSerializable): """Stores metadata for a collection of AgML datasets. Functionally, this class is just a wrapper around multiple `DatasetMetadata` objects, and for each of the traditional attributes in the original `DatasetMetadata`, returns a dictionary with all of the values for the corresponding datasets rather than just a single value. """ serializable = frozenset(("names", "metas", "task")) def __init__(self, datasets): # Build a collection of metadata objects. self._names = datasets self._metas = CollectionWrapper( [DatasetMetadata(d) for d in datasets]) self._validate_tasks() @classmethod def _from_collection(cls, metas): """Instantiates a `MultiDatasetMetadata` object from a collection.""" obj = MultiDatasetMetadata.__new__(MultiDatasetMetadata) obj._names = [meta.name for meta in metas] # We need to ignore types since some metadata objects might be # regular `DatasetMetadata`, but some might be `CustomDatasetMetadata`. obj._metas = CollectionWrapper(metas, ignore_types = True) obj._validate_tasks() return obj def _validate_tasks(self): # For a collection of datasets to work, they all need # to be of the same task. This is a check of this. tasks = self._metas.get_attributes('tasks') if not all(tasks[0].ml == t.ml for t in tasks): raise ValueError("To use a collection of datasets, all of them " f"must be of the same task. Got tasks {tasks} " f"for the provided datasets {self._names}.") self._task = tasks[0] def __getattr__(self, attr): # This is the main functionality of the `DatasetMetadata` # class. Rather than re-writing each of the attributes, # this method simply checks whether the requested attribute # or method exists in the original `DatasetMetadata` class, # then calls the relevant method on each of the stored # metadata objects and returns them. if hasattr(DatasetMetadata, attr): obj = getattr(DatasetMetadata, attr) # If it is a property, then we need to return something. if isinstance(obj, property): return {k: v for k, v in zip( self._names, self._metas.get_attributes(attr))} # If it is a method, then it is just printing. if isinstance(obj, types.FunctionType): self._metas.call_method(attr) return lambda: None # To act like a function. # Otherwise, raise an error for an invalid argument. else: raise AttributeError(f"Invalid attribute {attr} for `DatasetMetadata`.") class AnnotationRemap(AgMLSerializable): """A helper class to remap annotation labels for multiple datasets.""" serializable = frozenset(( "general_class_to_num", "num_to_class", "task", "generalize_class_detections")) def __init__(self, general_class_to_num, num_to_class, task, generalize_class_detections = False): self._task = task self._num_to_class = num_to_class self._general_class_to_num = general_class_to_num self._generalize_class_detections = generalize_class_detections def __call__(self, contents, name): """Re-maps the annotation for the new, multi-dataset mapping.""" image, annotations = contents # For image classification, simply re-map the label number. if self._task == 'image_classification': annotations = self._general_class_to_num[ self._num_to_class[name][annotations].lower()] # For semantic segmentation, re-map the mask IDs. if self._task == 'semantic_segmentation': unique_values = np.unique(annotations)[1:] # skip background new_values = np.array([self._general_class_to_num[ self._num_to_class[name][c].lower()] for c in unique_values]) for u, n in zip(unique_values, new_values): annotations[np.where(annotations == u)[0]] = n # For object detection, also just remap the annotation ID. if self._task == 'object_detection': if self._generalize_class_detections: annotations['category_id'] = np.ones_like( annotations['category_id']) else: category_ids = annotations['category_id'] category_ids[np.where(category_ids == 0)[0]] = 1 # fix new_ids = np.array([self._general_class_to_num[ self._num_to_class[name][c].lower()] for c in category_ids]) annotations['category_id'] = new_ids return image, annotations class AgMLMultiDatasetLoader(AgMLSerializable): """Loads and holds a collection of multiple datasets. This class serves as an interface for the `AgMLDataLoader` when using multiple datasets together, and enables similar seamless usage as with the traditional `AgMLDataLoader`, with additional features to support the usage of multiple datasets. Functionally, this class acts as a wrapper around multiple `AgMLDataLoader` objects, and draws upon similar functionality to the `DataManager` in order to access data from multiple objects. """ serializable = frozenset( ('info', 'loaders', 'loader_accessors', 'class_meta', 'set_to_keys', 'bounds', 'batch_size', 'shuffle_data', 'data_distributions', 'is_split', 'train_data', 'val_data', 'test_data')) def __init__(self, datasets, **kwargs): """Instantiates an `AgMLDataLoader` with multiple datasets.""" # The order of the dataset images and classes should be # agnostic to the order which they are passed in. So, the # same datasets but in a different order should yield the # same class mapping and internal representation. We sort # the datasets in order to ensure the same order every time. datasets = sorted(datasets) # Set up the datasets and their associated metadata. Since # there are multiple datasets, the `info` parameter here # will return a `MultipleDatasetMetadata` object which # is functionally similar to the `DatasetMetadata` object, # but returns dictionaries with the relevant parameters for # each of the different datasets, rather than just for one. self._info = MultiDatasetMetadata(datasets) # Create a set of `AgMLDataLoader` objects for each of the # datasets which are provided to the loader. These will # be the primary access interface for the objects. self._make_loaders(datasets, **kwargs) # Adapt all of the individual class types in the sub-datasets, # such that if there are two classes which are the same, then # the `class` that they represent should be the same. self._adapt_classes() # Similar to how the internal `DataManager` works, this wrapper # will access data from the internal `AgMLDataLoader` using an # accessor array, which will be the length of all of the datasets # in the loader combined. It will go the length of the datasets # in the order provided, where index 0 represents the first item # in the first dataset, and index 1 represents the last item in # the last dataset. Batching works in a similar way. self._loader_accessors = np.arange( sum(v for v in self._info.num_images.values())) sets = self._info.num_images.keys() bounds = np.cumsum(list(self._info.num_images.values())).tolist() bounds = (0, ) + (*bounds, ) self._set_to_keys = {} self._bounds = {s: b for s, b in zip(sets, bounds)} for i, set_ in enumerate(sets): value = 0 if i == 0 else 1 self._set_to_keys.update(dict.fromkeys( np.arange(bounds[i] - value, bounds[i + 1] + 1), set_)) # The batch size is modified similarly like the `DataManager`. # Since all of the data loaders should have the same properties, # transforms, and other values, we get the `make_batch` method # from the first loader and bound that to make batches. self._batch_size = None self._make_batch = self._loaders[0]._manager._train_manager.make_batch # Shuffling in a multi-dataset loader takes place on a high-level. # The contents of the actual datasets themselves are not shuffled. # Instead, the accessor array is shuffled (like the DataManager). self._shuffle_data = kwargs.get('shuffle', True) if self._shuffle_data: self.shuffle(self._shuffle_data) # We need to transform the class annotations before they are sent # through any other transforms, since those other transforms may # modify the value of the annotation and potentially cause issues # with the output annotation. However, we need this to happen # even when transforms are disabled. So, we modify each of the # loaders' `TrainingManager`s with a special argument to call an # extra helper method which in turn modifies the class annotation # before any other transforms are applied, and in any case. self._loaders.apply( lambda x: x._manager._train_manager._set_annotation_remap_hook( AnnotationRemap( self.class_to_num, self._info.num_to_class, self.task))) # We can't use the `auto` resizing mode for the resizing manager, # because of the complexity of trying to make it work with multiple # different data types. So, disable auto mode to prevent errors. self._loaders.apply( lambda x: x._manager._resize_manager.disable_auto()) # The data is not split to begin with. So, we set the split # parameter to false and store all of the split datasets themselves # as empty variables (which will be updated if and when it is split). self._is_split = False self._train_data = None self._val_data = None self._test_data = None # Calculate the distribution of the data. self._data_distributions = self._info.num_images # Total number of images in the entire dataset. self._num_images = sum(self._info.num_images.values()) @classmethod def _instantiate_from_collection(cls, *loaders, classes): """Instantiates an `AgMLMultiDatasetLoader` directly from a collection. This method is, in essence, a wrapper around the actual `__init__` method for the multi-loader, but one which takes into account the fact that the loaders are already instantiated, and thus works around those already-provided parameters, rather than starting from scratch. """ obj = AgMLMultiDatasetLoader.__new__(AgMLMultiDatasetLoader) # Create the custom dataset metadata wrapper. obj._info = MultiDatasetMetadata._from_collection([ loader.info for loader in loaders]) # Add the loaders and adapt classes. obj._loaders = CollectionWrapper( loaders, keys = [loader.info.name for loader in loaders]) obj._adapt_classes(cls = classes) # The remaining contents here are directly copied from the above # `__init__` method, without comments (see above for information): # Construct the accessor array. obj._loader_accessors = np.arange( sum(v for v in obj._info.num_images.values())) sets = obj._info.num_images.keys() bounds = np.cumsum(list(obj._info.num_images.values())).tolist() bounds = (0, ) + (*bounds, ) obj._set_to_keys = {} obj._bounds = {s: b for s, b in zip(sets, bounds)} for i, set_ in enumerate(sets): value = 0 if i == 0 else 1 obj._set_to_keys.update(dict.fromkeys( np.arange(bounds[i] - value, bounds[i + 1] + 1), set_)) # Set the batch size and shuffling. obj._batch_size = None obj._make_batch = obj._loaders[0]._manager._train_manager.make_batch obj._shuffle_data = loaders[0].shuffle_data if obj._shuffle_data: obj.shuffle(obj._shuffle_data) # Transform annotations and resizing. obj._loaders.apply( lambda x: x._manager._train_manager._set_annotation_remap_hook( AnnotationRemap( obj.class_to_num, obj._info.num_to_class, obj.task))) obj._loaders.apply( lambda x: x._manager._resize_manager.disable_auto()) # Finalize parameters. obj._is_split = False obj._train_data = None obj._val_data = None obj._test_data = None obj._data_distributions = obj._info.num_images obj._num_images = sum(obj._info.num_images.values()) # Return the object. return obj def __len__(self): # Return the length of the data, subject to batching. return self._data_length() def __getitem__(self, indexes: Union[int, slice, tuple, list]): # The `__getitem__` logic adopts the `DataManager` approach # towards getting multiple items, wrapped into this class. if isinstance(indexes, slice): data = np.arange(self._data_length()) indexes = data[indexes].tolist() if isinstance(indexes, int): indexes = [indexes] for idx in indexes: if idx not in range(len(self)): raise IndexError( f"Index {idx} out of range of " f"AgMLDataLoader length: {len(self)}.") return self._get_item_impl(resolve_list_value(indexes)) def __iter__(self): for indx in range(len(self)): yield self[indx] def __repr__(self): dsp = ", " out = f"<AgMLDataLoader: (datasets=[{dsp.join(self._info.name)}]" out += f", task={self.task}" out += f") at {hex(id(self))}>" return out def __str__(self): return repr(self) def copy(self): """Returns a deep copy of the data loader's contents.""" return self.__copy__() def __copy__(self): """Copies the loader and updates its state.""" cp = super(AgMLMultiDatasetLoader, self).__copy__() cp.copy_state(self) return cp def copy_state(self, loader): """Copies the state of another `AgMLDataLoader` into this loader. This method copies the state of another `AgMLDataLoader` into this loader, including its transforms, resizing, and training state. Other general parameters such as batch size and shuffling are left intact. Parameters ---------- loader : AgMLDataLoader The data loader from which the state should be copied. Returns ------- This `AgMLDataLoader`. """ # Copy the state for all the sub-loaders. If the loader state is of # a multi-loader, then only copy the state of its first loader. if isinstance(loader, AgMLMultiDatasetLoader): loader = loader._loaders[0] self._loaders.apply( lambda x: x.copy_state(loader) ) def _make_loaders(self, datasets, **kwargs): """Constructs the loaders for the datasets in the collection.""" # Get and validate the `dataset_path` argument. if 'dataset_path' in kwargs: dataset_path = kwargs.get('dataset_path') if isinstance(dataset_path, collections.Sequence): if not len(datasets) == len(dataset_path): raise IndexError( f"Got a sequence for the `dataset_path` of a " f"multi-dataset `AgMLDataLoader`, but it is not " f"the same length as the number of datasets: " f"{len(datasets)} datasets ({datasets}) but " f"{len(dataset_path)} paths ({dataset_path}).") elif isinstance(dataset_path, str): dataset_path = [dataset_path] * len(datasets) else: dataset_path = False kwargs.update({'dataset_path': dataset_path}) # Create all of the loaders. self._loaders = CollectionWrapper([ AgMLDataLoader(dataset, **kwargs) for dataset in datasets], keys = datasets) def _adapt_classes(self, cls = None): """Adapts the classes in the loader.""" # Get all of the unique classes in the loader. classes = self._info.classes.values() class_values = [[o.lower() for o in c] for c in classes] class_values = [i for s in class_values for i in s] unique_classes = np.unique(class_values).tolist() # Check that they match the given classes, if such a list is passed. if cls is not None: if not set(cls) == set(unique_classes): # noqa raise ValueError( f"Given list of classes {cls} to `AgMLDataLoader.merge`, " f"but calculated classes {unique_classes}. Check that the " f"given classes match the actual classes in the given datasets.") unique_classes = cls # Create a class metadata storing all of the unique # classes belonging to this loader and their mappings. self._class_meta = { 'classes': unique_classes, 'num_classes': len(unique_classes), 'class_to_num': { v: k + 1 for k, v in enumerate(unique_classes)}, 'num_to_class': { k + 1: v for k, v in enumerate(unique_classes)} } def _data_length(self): """Calculates the length of the data from the different datasets.""" return len(self._loader_accessors) @property def info(self): """Returns a `DatasetMetadata` object containing dataset info. The contents returned in the `DatasetMetadata` object can be used to inspect dataset metadata, such as the location the data was captured, the data formats, and the license/copyright information. See the `DatasetMetadata` class for more information. """ return self._info @property def task(self): """Returns the ML task that this dataset is constructed for.""" return self._info._task.ml @property def num_images(self): """Returns the number of images in the entire dataset.""" return sum(self.data_distributions.values()) @property def classes(self): """Returns the classes that the dataset is predicting.""" return self._class_meta['classes'] @property def num_classes(self): """Returns the number of classes in the dataset.""" return self._class_meta['num_classes'] @property def num_to_class(self): """Returns a mapping from a number to a class label.""" return self._class_meta['num_to_class'] @property def class_to_num(self): """Returns a mapping from a class label to a number.""" return self._class_meta['class_to_num'] @property def data_distributions(self): """Returns a distribution of data in the loader. This property returns a dictionary which provides information about the data within the loader; e.g., how many images are from each of the datasets in the collection. """ return self._data_distributions def _generate_split_loader(self, loaders, split): """Generates a split `AgMLDataLoader`.""" # Check if the data split exists. if any(loader is None for loader in loaders): raise ValueError( f"Attempted to access split loader {split} when " f"parent loader has not been split.") # Create a new `CollectionWrapper` around the datasets. new_collection = CollectionWrapper( loaders, keys = [l_.name for l_ in loaders]) # Get the state of the current loader and update it # with the new collection of loaders. Then, update # the accessors and number of images with the newly # reduced quantity (due to the splitting of data). loader_state = self.copy().__getstate__() loader_state['loaders'] = new_collection total_num_images = sum(len(loader) for loader in loaders) data_distributions = { loader.name: len(loader) for loader in loaders} loader_state['data_distributions'] = data_distributions accessors = np.arange(0, total_num_images) if self._shuffle_data: np.random.shuffle(accessors) loader_state['loader_accessors'] = accessors batch_size = loader_state.pop('batch_size') loader_state['batch_size'] = None # Re-generate the mapping for bounds. sets = self._info.num_images.keys() bound_ranges = np.cumsum([len(loader) for loader in loaders]).tolist() bound_ranges = (0, ) + (*bound_ranges, ) set_to_keys = {} bounds = {s: b for s, b in zip(sets, bound_ranges)} for i, set_ in enumerate(sets): value = 0 if i == 0 else 1 set_to_keys.update(dict.fromkeys( np.arange(bound_ranges[i], bound_ranges[i + 1] + value), set_)) loader_state['set_to_keys'] = set_to_keys loader_state['bounds'] = bounds # Create the new loader from the updated state. new_loader = AgMLMultiDatasetLoader.__new__(AgMLMultiDatasetLoader) new_loader.__setstate__(loader_state) # Batching data should be re-done independently. if batch_size is not None: new_loader.batch(batch_size = batch_size) # Block out all of the splits of the already split # loader and set the `_is_split` attribute to True, # preventing future splits, and return. for attr in ['train', 'val', 'test']: setattr(new_loader, f'_{attr}_data', None) new_loader._is_split = True return new_loader @property def train_data(self): """Stores the `train` split of the data in the loader.""" if isinstance(self._train_data, AgMLMultiDatasetLoader): return self._train_data self._train_data = self._generate_split_loader( self._loaders.get_attributes('train_data'), split = 'train') return self._train_data @property def val_data(self): """Stores the `val` split of the data in the loader.""" if isinstance(self._val_data, AgMLMultiDatasetLoader): return self._val_data self._val_data = self._generate_split_loader( self._loaders.get_attributes('val_data'), split = 'val') return self._val_data @property def test_data(self): """Stores the `test` split of the data in the loader.""" if isinstance(self._test_data, AgMLMultiDatasetLoader): return self._test_data self._test_data = self._generate_split_loader( self._loaders.get_attributes('test_data'), split = 'test') return self._test_data def eval(self): """Sets the `AgMLDataLoader` in evaluation mode. Evaluation mode disables transforms, and only keeps the loader applying resizing to the contents. If the loader was previously set into TensorFlow or PyTorch mode, however, it will also keep up tensor conversion and potential batch adding (see `as_keras_sequence()` and `as_torch_dataset()` methods for more information on the exact operations). This method does not completely disable preprocessing, to completely disable preprocessing, use `loader.disable_preprocessing()`. Additionally, if you want to keep only the resizing but not the implicit tensor conversions based on the backend, then run: > loader.disable_preprocessing() # or loader.reset_preprocessing() > loader.eval() This will refresh the backend conversions and return it to `eval` mode. Returns ------- The `AgMLDataLoader` object. """ self._loaders.call_method('eval') return self def disable_preprocessing(self): """Disables all preprocessing on the `AgMLDataLoader`. This sets the loader in a no-preprocessing mode (represented internally as `False`), where only the raw data is returned: no transforms, resizing, or any conversion to any type of backend. This can be used to test or inspect the original data contents of the loader before processing. The loader can be set into any mode from here, for instance see `eval()`, `as_keras_sequence()`, and `as_torch_dataset()` for specific examples on the different potential training and evaluation states. If you just want to reset the loader to its default state, which applies only transforms and resizing, then use `loader.reset_preprocessing()`. Returns ------- The `AgMLDataLoader` object. """ self._loaders.call_method('disable_preprocessing') return self def reset_preprocessing(self): """Re-enables preprocessing on the `AgMLDataLoader`. This resets the loader back to its default train state, namely where it applies just the given transforms and content resizing. This is a consistent method, meaning that regardless of the prior train state of the loader before running this method, it will hard reset it to its original state (similar to `disable_preprocessing()`, but it keeps some preprocessing). Returns ------- The `AgMLDataLoader` object. """ self._loaders.call_method('reset_preprocessing') return self def on_epoch_end(self): """Shuffles the dataset on the end of an epoch for a Keras sequence. If `as_keras_sequence()` is called and the `AgMLDataLoader` inherits from `tf.keras.utils.Sequence`, then this method will shuffle the dataset on the end of each epoch to improve training. """ self._loaders.call_method('on_epoch_end') def as_keras_sequence(self): """Sets the `DataLoader` in TensorFlow mode. This TensorFlow extension converts the loader into a TensorFlow mode, adding inheritance from the superclass `keras.utils.Sequence` to enable it to be used directly in a Keras pipeline, and adding extra preprocessing to the images and annotations to make them compatible with TensorFlow. The main features added on enabling this include: 1. Conversion of output images and annotations to `tf.Tensor`s. 2. Adding an implicit batch size dimension to images even when the data is not batched (for compatibility in `Model.fit()`). 3. Adding inheritance from `keras.utils.Sequence` so that any `AgMLDataLoader` object can be used directly in `Model.fit()`. 4. Setting the data loader to use a constant image shape, namely `auto` (which will default to (512, 512) if none is found). This can be overridden by manually setting the image shape parameter back after running this method. Note that this may result in errors when attempting implicit tensor conversion. Returns ------- The `AgMLDataLoader` object. """ self._loaders.call_method('as_keras_sequence') return self def as_torch_dataset(self): """Sets the `DataLoader` in PyTorch mode. This PyTorch extension converts the loader into a PyTorch mode, adding inheritance from th superclass `torch.utils.data.Dataset` to enable it to be used directly in a PyTorch pipeline, and adding extra preprocessing to the images and annotations to make them compatible with PyTorch. The main features added on enabling this include: 1. Conversion of output images and annotations to `torch.Tensor`s. 2. Converting the channel format of the input images from the default, channels_last, into channels_first (NHWC -> NCHW). 3. Adding inheritance from `torch.utils.data.Dataset` so that any `AgMLDataLoader` object can be used with a `torch.utils.data.DataLoader`. 4. Setting the data loader to use a constant image shape, namely `auto` (which will default to (512, 512) if none is found). This can be overridden by manually setting the image shape parameter back after running this method. Note that this may result in errors when attempting implicit tensor conversion. Returns ------- The `AgMLDataLoader` object. """ self._loaders.call_method('as_torch_dataset') @property def shuffle_data(self): """Returns whether the loader is set to shuffle data or not. By default, if no value is passed in initialization, this is set to `True`. It can be manually toggled to `False` using this property. """ return self._shuffle_data @shuffle_data.setter def shuffle_data(self, value): """Set whether the loader should shuffle data or not. This can be used to enable/disable shuffling, by passing either `True` or `False`, respectively. """ if not isinstance(value, bool): raise TypeError("Expected either `True` or `False` for 'shuffle_data'.") self._shuffle_data = value def shuffle(self, seed = None): """Potentially shuffles the contents of the loader. If shuffling is enabled on this loader (`shuffle = False` has not been passed to the instantiation), then this method will shuffle the order of contents in it. A seed can be provided to shuffle the dataset to an expected order. If the data is already batched, then the batch contents will be shuffled. For instance, if we have data batches [[1, 2], [3, 4]], then the shuffling result will be [[3, 4], [1, 2]]. If you want all of the contents to be shuffled, call `shuffle` before batching. Note that the data is automatically shuffled upon instantiation, unless the `shuffle = False` parameter is passed at instantiation. However, this disables automatic shuffling for the class permanently, and this method must be called to shuffle the data. Parameters ---------- seed : int, optional A pre-determined seed for shuffling. Returns ------- The `AgMLDataLoader` object. """ if seed is None: np.random.shuffle(self._loader_accessors) else: with seed_context(seed): np.random.shuffle(self._loader_accessors) return self def take_dataset(self, name) -> "AgMLDataLoader": """Takes one of the datasets in the multi-dataset collection. This method selects one of the datasets (as denoted by `name`) in this multi-dataset collection and returns an `AgMLDataLoader` with its contents. These contents will be subject to any transforms and modifications as applied by the main loader, but the returned loader will be a copy, such that any new changes made to the main multi-dataset loader will not affect the new loader. Parameters ---------- name : str The name of one of the sub-datasets of the loader. Returns ------- An `AgMLDataLoader`. """ return self._loaders[name].copy() @inject_random_state def take_random(self, k) -> "AgMLMultiDatasetLoader": """Takes a random set of contents from the loader. This method selects a sub-sample of the contents in the loader, based on the provided number of (or proportion of) elements `k`. It then returns a new loader with just this reduced number of elements. The new loader is functionally similar to the original loader, and contains all of the transforms/batching/other settings which have been applied to it up until this method is called. Note that the data which is sampled as part of this new loader is not removed from the original loader; this simply serves as an interface to use a random set of images from the full dataset. For a multi-dataset loader, data is sampled from each of the sub-datasets proportionally, e.g., the proportion of images in the new dataset (per each sub-dataset) will be the same as in the original dataset. Parameters ---------- k : {int, float} Either an integer specifying the number of samples or a float specifying the proportion of images from the total to take. Returns ------- A reduced `AgMLDataLoader` with the new data. """ # Parse the input to an integer. if isinstance(k, float): # Check that 0.0 <= k <= 1.0. if not 0.0 <= k <= 1.0: raise ValueError( "If passing a proportion to `take_class`, " "it should be in range [0.0, 1.0].") # Convert the proportion float to an absolute int. Note that # the method used is rounding up to the nearest int for cases # where there is not an exact proportional equivalent. getcontext().prec = 4 # noqa proportion = Decimal(k) / Decimal(1) num_images = self.num_images k = int(proportion * num_images) # If the input is an integer (or the float is converted to an int # above), then select a random sampling of images from the dataset. if isinstance(k, int): # Check that `k` is valid for the number of images in the dataset. if not 0 <= k <= self.num_images: raise ValueError( f"Received a request to take a random sampling of " f"{k} images, when the dataset has {self.num_images}.") # Calculate the proportions. If the total sum is less than `k`, # add 1 to the dataset with the lowest number of images. getcontext().prec = 4 # noqa num_images = self.num_images proportions = {key: int((Decimal(val) / Decimal(num_images)) * k) for key, val in self._data_distributions.items()} if sum(proportions.values()) != num_images: diff = sum(proportions.values()) - k smallest_split = list(proportions.keys())[ list(proportions.values()).index( min(proportions.values()))] proportions[smallest_split] = proportions[smallest_split] - diff return self._generate_split_loader( self._loaders.call_method( 'take_random', tuple(proportions.values())), 'train') # Otherwise, raise an error. else: raise TypeError( f"Expected only an int or a float when " f"taking a random split, got {type(k)}.") @inject_random_state def split(self, train = None, val = None, test = None, shuffle = True): """Splits the data into train, val and test splits. For this multi-dataset loader, an even split of data will be selected from each dataset. E.g., if you have a loader of two datasets, each with 100 images, and want a train/test split of 0.9/0.1, then 90 images from each dataset will form the new training set, and 10 images from each will form the new test set. This is to ensure consistency in training. By default, this method does nothing (or if the data has been split into sets, it resets them all to one set). Setting the `train`, `val`, and `test` parameters randomly divides the data into train, validation, and/or test sets, depending on which ones are provided and their values. Values can either be passed as exact numbers or as proportions, e.g. either `train = 80, test = 20` in a 100-value dataset, or as `train = 0.8, test = 0.2`. Whichever value is not passed, e.g. `val` in this case, has no value in the loader. Parameters ---------- train : {int, float} The split for training data. val : {int, float} The split for validation data. test : {int, float} The split for testing data. shuffle : bool Whether to shuffle the split data. Notes ----- Any processing applied to this `AgMLDataLoader` will also be present in the split loaders until they are accessed from the class. If you don't want these to be applied, access them right after splitting. """ # We run this by applying dataset splitting to all of the individual # loaders, which then create and get their own even split of data # for the dataset they are loading. This will first ensure an even # split of data from each of the different datasets. # # Similar to the `AgMLDataLoader` itself, we will not access the # `train/val/test_data` parameters instantly, as this will allow # any new transforms or other parameters which are applied to the # parent loader to be also applied to all of the child split loaders # until they are actually accessed in this overhead multi-dataset # loader class. Then, they will be unset and created. self._loaders.apply( lambda x: x.split( train = train, val = val, test = test, shuffle = shuffle)) def batch(self, batch_size = None): """Batches sets of image and annotation data according to a size. This method will group sets of data together into batches of size `batch_size`. In turn, items gathered from the loader will, rather than being an image and annotation, be an array of images and an array of annotations (not an array of image/annotation pairs). Batching data will include a `batch` dimension for the images and annotations that are returned (e.g., the image array will have dimensions NHWC instead of HWC). If the data is not batched, this dimension will not be present unless the loader is in training mode. The data can be un-batched by passing `None` to batch size (or calling the method with no arguments). Parameters ---------- batch_size : int, None The number of groups to batch data together into. Notes ----- The last batch will be of size <= `batch_size`. """ # If the data is already batched and a new batch size is called, # then update the existing batch sizes. For unbatching the data, # update the batch state and then flatten the accessor array. if self._batch_size is not None: try: self._loader_accessors = np.concatenate(self._loader_accessors).ravel() except ValueError: # The array is currently 0-dimensional. pass if batch_size is None or batch_size == 0: self._batch_size = None return # If we have a batch size of `1`, then don't do anything # since this doesn't really mean to do anything. if batch_size == 1: return # Otherwise, calculate the actual batches and the overflow # of the contents, and then update the accessor. num_splits = len(self._loader_accessors) // batch_size data_items = np.array(self._loader_accessors) overflow = len(self._loader_accessors) - num_splits * batch_size extra_items = data_items[-overflow:] try: batches = np.array_split( np.array(self._loader_accessors [:num_splits * batch_size]), num_splits) except ValueError: log(f"There is less data ({len(self._loader_accessors)}) than the provided " f"batch size ({batch_size}). Consider using a smaller batch size.") batches = [self._loader_accessors] else: if len(extra_items) < batch_size: batches.append(extra_items) self._loader_accessors = np.array(batches, dtype = object) self._batch_size = batch_size # Update the batch creation method. self._make_batch = self._loaders[0]._manager._train_manager.make_batch def resize_images(self, image_size = None, method = 'bilinear'): """Resizes images within the loader to a specified size. This method applies a resizing parameter for images before they are returned from the data loader. The default starting point, if this method is never called, is to apply no resizing. If the loader is set in "training" mode and no size is specified, it defaults to (512, 512). Image resizing contains a few modes: 1. `default` or `None`: No resizing, leaves images in default size. This is the default parameter if nothing is passed. 2. `train`: This will set a default training size of (512, 512). 3. `imagenet`: This will set a default size of (224, 224). 4. custom size: Resizes the images to the provided size. The resizing also applies to the annotation in certain cases, depending on the task and the actual content of the annotation: - For object detection, the bounding box coordinates will be resized and the area of the box will in turn be recomputed. - For semantic segmentation, the annotation mask will be resized, using a nearest-neighbor interpolation to keep it as similar as possible to the original mask (preventing data loss). Parameters ---------- image_size : optional The resizing parameter for the image. method : optional The method to resize the images. Should be one of 'nearest', 'bilinear', 'linear', or 'cubic'. Defaults to 'bilinear'. Notes ----- - There is no `auto` parameter for resizing images when using multiple data loaders. If auto is passed, it will warn the user and switch to `default`. - If a transform pipeline is provided, images will be resized *before* being passed into the transform pipeline. """ if image_size == 'auto': log("There is no `auto` parameter for resizing images when using" "multiple datasets in a loader. Switching to `default`.") self._loaders.apply( lambda x: x.resize_images( image_size = image_size, method = method ) ) def transform(self, transform = NoArgument, target_transform = NoArgument, dual_transform = NoArgument): """Applies vision transforms to the input image and annotation data. This method applies transformations to the image and annotation data in the dataset. Transforms include augmentations and other processing methods, and can be applied independently to the image and annotation, or together to both (`transform`, `target_transform`, `dual_transform`). The hierarchy in which transforms are applied is: transform -> --------| |-----> dual_transform target_transform -> --| The `transform` and `target_transform` argument are used for methods which act independently on the image and the annotation, respectively. The values passed to these arguments can be: - An `albumentations` transform pipeline. - A `keras.Sequential` model (or preprocessing layer) or a set of `torchvision.transform`s. - A method which accepts one input and returns one output. The `dual_transform` argument is used for non-image-classification tasks. The following describe the types of arguments that can be passed to `dual_transform`, depending on the task: Object Detection: - An `albumentations` transform pipeline with `bbox_params` in to be applied to both the image and the bounding boxes. - A method (not a torchvision or Keras preprocessing pipeline) that accepts two inputs and returns two outputs. Semantic Segmentation: - An `albumentations` transform pipeline that may include spatial and/or visual augmentation. - A method to independently or dually apply transformations to the image and annotation mask. - A `torchvision.transforms` or `tf.keras.Sequential` pipeline which will be applied to the image and mask using the same random seed, for reproducibility. Use the provided method `generate_keras_segmentation_dual_transform` for this. If you want to reset the transforms, then simply call this method with no arguments. Alternatively, to reset just a single transform, pass the value of that argument as `None`. Parameters ---------- transform : optional A transform to be applied independently to the input image. target_transform : optional A transform to be applied independently to the annotation. dual_transform : optional A transform to be applied to both the input and annotation. Notes ----- - Image resizing takes place before any transformations are applied. After the transforms are applied in this order, they returned and if passed again, they will have a different transform applied to them. The state is independent of the images passed. - Albumentations transforms are special in that even transforms which would normally be passed to `dual_transform` (e.g., they act on the input image and the output annotation) can simply be passed to the `transform` argument and they will automatically be applied. """ self._loaders.apply( lambda x: x._manager.push_transforms( transform = transform, target_transform = target_transform, dual_transform = dual_transform ) ) def normalize_images(self, method = 'scale'): """Converts images from 0-255 integers to 0-1 floats and normalizes. This is a convenience method to convert all images from integer-valued arrays into float-valued arrays, and normalizes them (using shifting and scaling from mean and std). This is useful for training in order to reduce computational complexity (instead of large-valued integer multiplication, only float multiplication), and for extracting the most information out of different types of imagery. There are three different 'normalization' modes that can be initialized with this method, as described below: 1. `scale`: This simply scales images from the 0-255 pixel range to the 0-1 range (and converts them to floats as such). 2. `imagenet`: This performs normalization using the traditional ImageNet mean and standard deviation: (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]) This is useful when trying to conduct transfer learning, for instance. 3. `standard`: This performs normalization using a pre-calculated mean and standard deviation for the dataset (see the public sources JSON). To remove normalization altogether, pass `None` as a parameter. Parameters ---------- method : str The method by which to normalize the images. Notes ----- This method is not implicitly called when converting to PyTorch/TensorFlow mode, it needs to be manually called even if you just want 0-1 scaled images. """ if method not in ['scale', 'imagenet', 'standard', None]: raise ValueError(f"Received invalid normalization method: '{method}'.") if method == 'scale': normalization_params = 'scale' elif method == 'imagenet': normalization_params = 'imagenet' elif method == 'standard': normalization_params = self._info.image_stats else: normalization_params = None # Normalization parameters may be specific to each dataset in # the loader, so we need to make sure we account for this. self._loaders.apply( lambda x, transform: x._manager.push_transforms( transform = transform, target_transform = NoArgument, dual_transform = NoArgument), args = [('normalize', normalization_params[key]) for key in self._loaders.keys] ) def labels_to_one_hot(self): """Converts image classification numerical labels to one-hot labels. This is a convenience method to apply one-hot vector transformations to the output labels for image classification. Essentially, if we have a set of labels, [1, 2], it will convert it to [[0, 1, 0], [0, 0, 1]]. This is a more commonly used format for image classification. """ if self.task != 'image_classification': raise RuntimeError("The `one_hot` label transformation can only " "be used for image classification tasks.") self.transform( target_transform = ('one_hot', self.num_classes) ) def mask_to_channel_basis(self): """Converts semantic segmentation masks to channel-wise. This is a convenience method to convert integer-labeled semantic segmentation masks into channel-by-channel masks, essentially one-hot vector transformation but for semantic segmentation. Note that if the task is binary segmentation, e.g. there is only one class, then this method will do nothing. This method should traditionally be called *after* applying general transformations to the loader, in order to prevent any issues. """ if self.task != 'semantic_segmentation': raise ValueError("The `mask_to_channel_basis` transformation " "can only be used for semantic segmentation tasks.") self.transform( target_transform = ('channel_basis', self._info.num_classes) ) def generalize_class_detections(self): """Generalizes object detection classes to a single class. This is a convenience method for object detection tasks, and converts all of the individual class labels in the task into a single class, essentially allowing the model to purely focus on detection of objects and fine-tuning bounding boxes, with no focus on differentiating classes of different boxes. """ if self.task != 'object_detection': raise ValueError("The `generalize_class_detections` transformation" "can only be used for object detection tasks.") self._loaders.apply( lambda x: x._manager._train_manager._set_annotation_remap_hook( AnnotationRemap( self.class_to_num, self._info.num_to_class, self.task, generalize_class_detections = True))) def export_contents(self, export_format = None): """Exports the internal contents of the `AgMLDataLoader`s. This method serves as a hook for high-level users who simply want to download and get the data, by exporting the unprocessed metadata of the actual dataset, with the following formats: Image Classification: A mapping between the local image paths and the numerical labels. Object Detection: A mapping between the local image paths (the full path, not just the file name), and the COCO JSON annotations corresponding to each of the images. To get the original COCO JSON annotation file contents, use `export_format = 'coco'`. Semantic Segmentation: A mapping between the local image paths and the local annotation mask paths. The `export_format` argument can be used to customize what this method returns. By default, it is set to `None`, and returns dictionaries with the above specified mappings. However, setting `export_format = 'arrays'` will return two arrays, with the first array containing the image paths and the second array containing the annotation data. Parameters ---------- export_format : optional The format to export the data in. Defaults to a mapping. Returns ------- The raw contents of the datasets. """ return { k: self._loaders[k].export_contents( export_format = export_format) for k in self._loaders.keys } @staticmethod def _calculate_data_and_loader_index(index, bound_map, set_map): loader_idx = set_map[index] is_equal = np.where(index == np.array(list(bound_map.values())))[0] if is_equal.size != 0: data_idx = 0 else: data_idx = int(index - list(bound_map.values())[int( np.searchsorted(np.array( list(bound_map.values())), index) - 1)]) if data_idx < 0: data_idx = 0 return loader_idx, data_idx def _load_one_image_and_annotation(self, index): """Loads one image and annotation from a `DataObject`.""" # Get the image and annotation from the corresponding loader. loader, data_idx = self._calculate_data_and_loader_index( index, self._bounds, self._set_to_keys) return self._loaders[loader][data_idx] def _load_multiple_items(self, indexes): """Loads multiple images and annotations from a set of `DataObject`s.""" # Either we're getting multiple batches, or just multiple items. contents = [] if self._batch_size is not None: for i in indexes: contents.append(self._load_batch(self._loader_accessors[i])) else: for i in indexes: contents.append(self._load_one_image_and_annotation( self._loader_accessors[i])) return contents def _batch_multi_image_inputs(self, images): """Converts either a list of images or multiple input types into a batch.""" # If the input images are just a simple batch. if is_array_like(images[0]): return convert_to_batch(images) # Otherwise, convert all of them independently. keys = images[0].keys() batches = {k: [] for k in keys} for sample in images: for key in sample: batches[key].append(sample[key]) return {k: self._batch_multi_image_inputs(i) for k, i in batches.items()} def _batch_multi_output_annotations(self, annotations): """Converts either a list of annotations or multiple annotation types into a batch.""" # If the output annotations are simple objects. if (isinstance(annotations[0], (list, np.ndarray)) or isinstance(annotations, (list, np.ndarray)) and isinstance(annotations[0], (int, float))): if not consistent_shapes(annotations): annotations = np.array(annotations, dtype = object) else: annotations = np.array(annotations) return annotations # For object detection, just return the COCO JSON dictionaries. if self.task == 'object_detection': return annotations # Otherwise, convert all of them independently. keys = annotations[0].keys() batches = {k: [] for k in keys} for sample in annotations: for key in sample: batches[key].append(sample[key]) return {k: self._batch_multi_output_annotations(i) for k, i in batches.items()} def _load_batch(self, batch_indexes): """Gets a batch of data from the dataset. This differs from simply getting multiple pieces of data from the dataset, such as a slice, in that it also stacks the data together into a valid batch and returns it as such. """ # Get the images and annotations from the data objects. images, annotations = [], [] for index in batch_indexes: image, annotation = self._load_one_image_and_annotation(index) images.append(image) annotations.append(annotation) # Attempt to create batched image arrays. images = self._batch_multi_image_inputs(images) # Attempt the same for the annotation arrays. This is more complex # since there are many different types of annotations, namely labels, # annotation masks, COCO JSON dictionaries, etc. We need to properly # create a batch in each of these cases. annotations = self._batch_multi_output_annotations(annotations) # Return the batches. return self._make_batch( images = images, annotations = annotations ) def _get_item_impl(self, indexes): """Loads and processes a piece (or pieces) of data from the dataset. This is the actual accessor method that performs the loading of data and the relevant processing as dictated by loading, image resizing, transform application, and other internal processing methods such as creating batches. This is called by the `AgMLDataLoader` to get data. """ # If there is only one index and the data is not batched, # then we just need to return a single `DataObject`. if isinstance(indexes, int) and self._batch_size is None: return self._load_one_image_and_annotation( self._loader_accessors[indexes]) # If we have a batch of images, then return the batch. if isinstance(indexes, int) and self._batch_size is not None: return self._load_batch(self._loader_accessors[indexes]) # Otherwise, if there are multiple indexes (e.g., an unstacked # slice or just a tuple of integers), then we get multiple images. if isinstance(indexes, (list, tuple)): return self._load_multiple_items(indexes) def export_torch(self, **loader_kwargs): """Exports the contents of the loader in a native PyTorch loader. This method wraps the contents of this data loader inside of a `torch.utils.data.DataLoader`. This method differs from the `export_tensorflow()` method in that there is no need to convert directly to a `tf.data.Dataset`, rather if this `AgMLDataLoader` inherits from `torch.utils.data.Dataset`, it can just be directly wrapped into a `torch.utils.data.DataLoader`. The returned `DataLoader` is functionally similar to the `AgMLDataLoader` in terms of preprocessing and transforming. You can pass arguments to the `DataLoader` instantiation as keyword arguments to this method. Note that the `AgMLDataLoader` which this method encloses is instead a copy of the instance the method is run on, so that any changes to the loader afterwards don't affect the exported loader. Parameters ---------- loader_kwargs : optional A set of keyword arguments for the `torch.utils.data.DataLoader`. See the documentation for the loader for more information. Returns ------- A `torch.utils.data.DataLoader` enclosing a copy of this loader. """ from agml.backend.tftorch import torch from torch.utils.data import DataLoader if get_backend() != 'torch': if user_changed_backend(): raise StrictBackendError( change = 'torch', obj = self.export_torch) set_backend('torch') # Make a copy of the `AgMLDataLoader` so the following changes # don't affect the original loader, just the new one. obj = self.copy() # Convert to a PyTorch dataset. obj.as_torch_dataset() # The `DataLoader` automatically batches objects using its # own mechanism, so we remove batching from this DataLoader. batch_size = loader_kwargs.pop( 'batch_size', obj._batch_size) obj.batch(None) shuffle = loader_kwargs.pop( 'shuffle', obj._shuffle_data) # The `collate_fn` for object detection is different because # the COCO JSON dictionaries each have different formats. So, # we need to replace it with a custom function. collate_fn = loader_kwargs.pop('collate_fn') if obj.task == 'object_detection' and collate_fn is None: def collate_fn(batch): images = torch.stack( [i[0] for i in batch], dim = 0) coco = tuple([i[1] for i in batch]) return images, coco # Return the DataLoader with a copy of this AgMLDataLoader, so # that changes to this will not affect the returned loader. return DataLoader( obj, batch_size = batch_size, shuffle = shuffle, collate_fn = collate_fn, **loader_kwargs )
66,025
42.639128
94
py
AgML
AgML-main/agml/data/managers/transform_helpers.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import abc import numpy as np from agml.framework import AgMLSerializable from agml.utils.random import seed_context class TransformApplierBase(AgMLSerializable): """Applies a transform to the input data. This is used as a wrapper class for applying transformations in more complex cases. Derived classes implement the `apply` method to wrap the input transformation. This is used in cases such as `albumentations` transforms, which require keyword arguments. This wrapper allows them to be used as a traditional method for consistency in application. """ serializable = frozenset(('transform', )) def __init__(self, transform): self._transform = transform @abc.abstractmethod def apply(self, *args, **kwargs): """Applies the transformation to the input data.""" return def __call__(self, *args): return self.apply(*args) def __str__(self): return self.__class__.__name__ + f": {self._transform}" class AlbumentationsTransformSingle(TransformApplierBase): def apply(self, image): transform = self._transform(image = image) return transform['image'] class AlbumentationsTransformMask(TransformApplierBase): def apply(self, image, mask): transform = self._transform(image = image, mask = mask) return transform['image'], transform['mask'] class AlbumentationsTransformCOCO(TransformApplierBase): def apply(self, image, coco): # This method is a bit more complex. We can't just apply it # to the COCO JSON dictionary as we need to extract the # bounding boxes/category IDs, do the transformation on those, # and then re-insert them into the COCO dictionary. coco_boxes = coco['bbox'] coco_labels = coco['category_id'] bboxes = np.c_[coco_boxes, coco_labels] transform = self._transform(image = image, bboxes = bboxes) image, bboxes = transform['image'], \ np.array(transform['bboxes'])[:, :-1] coco['bbox'], areas = bboxes, [] for box in bboxes: areas.append( (box[0] + box[2]) * (box[1] + box[3])) coco['area'] = areas.copy() return image, coco class SameStateImageMaskTransform(TransformApplierBase): def apply(self, image, mask): # This method applies a transformation to the image and mask. # Essentially, we need to set the random seed before applying # it to the image and mask to get the same results for both # the image and the mask. So, we change the seed each time, but # we also reset the seed so that it goes back to normal afterwards. seed = np.random.randint((1 << 31) - 1) # default is 32-bit systems with seed_context(seed) as context: image = self._transform(image) context.reset() mask = self._transform(mask) return image, mask class NormalizationTransformBase(TransformApplierBase, abc.ABC): """A subclass to mark transforms as normalizing transforms.""" pass class ScaleTransform(NormalizationTransformBase): def apply(self, image): if image.max() >= 1 or np.issubdtype(image.dtype, np.integer): image = (image / 255).astype(np.float32) return image class NormalizationTransform(NormalizationTransformBase): def apply(self, image): # This method applies normalization to input images, scaling them # to a 0-1 float range, and performing scaling normalization. if image.max() >= 1 or np.issubdtype(image.dtype, np.integer): image = (image / 255).astype(np.float32) mean, std = self._transform mean = np.array(mean, dtype = np.float32) std = np.array(std, dtype = np.float32) denominator = np.reciprocal(std, dtype = np.float32) image = (image - mean) * denominator return image class OneHotLabelTransform(TransformApplierBase): def apply(self, labels): # This applies a one-hot label transformation. The only argument # in the `_transform` parameter is simply the number of labels. one_hot = np.zeros(shape = (self._transform, )) one_hot[labels] = 1 return one_hot.astype(np.float32) class MaskToChannelBasisTransform(TransformApplierBase): def apply(self, mask): # Converts a 2-d mask of `n` labels, excluding background, to a # `n x h x w` 3-dimensional mask with each channel a label. input_shape = mask.shape mask = mask.ravel() n = mask.shape[0] mask = np.array(mask, dtype = np.int32) out = np.zeros(shape = (n, self._transform + 1)) out[np.arange(n), mask] = 1 out = np.reshape(out, input_shape + (self._transform + 1,)) return out[..., 1:].astype(np.int32) class ToTensorSeg(TransformApplierBase): def apply(self, image, mask): # This is essentially the same as `torchvision.transforms.ToTensor`, # except it doesn't scale the values of the segmentation mask from # 0-255 -> 0-1, as this would break the segmentaiotn pipeline. if image.ndim == 2: image = image[:, :, None] image = torch.from_numpy(image).permute(2, 0, 1).contiguous() if isinstance(image, torch.ByteTensor): image = image.to(dtype = torch.float32).div(255) mask = torch.from_numpy(mask) return mask
6,074
36.5
76
py
AgML
AgML-main/agml/data/managers/training.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from enum import Enum import numpy as np from agml.framework import AgMLSerializable from agml.utils.image import needs_batch_dim from agml.backend.tftorch import ( tf, torch, set_backend, get_backend, user_changed_backend, StrictBackendError, _convert_image_to_torch, is_array_like ) class TrainState(Enum): NONE = None TF = 'tf' TORCH = 'torch' EVAL = 'eval' EVAL_TF = 'eval-tf' EVAL_TORCH = 'eval-torch' FALSE = False # Shorthand form of the enum for testing explicit values. t_ = TrainState class TrainingManager(AgMLSerializable): """Controls the training state for the `AgMLDataLoader`. This manager is responsible for managing the backend system and training parameters of the `AgMLDataLoader`. In particular, it manages the data in different circumstances, such as when the data is in "train" and "eval" modes, or when it is set to be in a TensorFlow/PyTorch-compatible mode. This allows for more efficient compatibility management with the different backends, for instance dynamic tensor conversion and image formatting to account for channels_first vs. channels_last, as well as being able to manage whether certain preprocessing steps should be applied, allowing for independent train and eval modes. """ serializable = frozenset(( 'transform_manager', 'resize_manager', 'state', 'task', 'remap_hook', 'name')) def __init__(self, transform_manager, resize_manager, task = None): # Update the general parameters for the loader. self._task = task self._name = resize_manager._dataset_name # The `TrainingManager` is responsible for applying the # actual transforms, and thus controls the `TransformManager` # and the `ImageResizeManager` and their relevant states. self._transform_manager = transform_manager self._resize_manager = resize_manager # The `state` the the loader is in is variable, and can be # changed during program execution (e.g., between 'train' # and 'eval', or even into a specific backend). # # The `state` determines the actual preprocessing functions # that it should apply, as it controls the managers. # This variable tracks that state and uses it to determine # which steps that it should apply. # # See the `update_state()` method to see the valid states. self._state: "TrainState" = TrainState.NONE # A hook for multi-dataset loaders. self._remap_hook = False @property def state(self): """Exposes the internal state.""" return self._state def update_state(self, state): """Updates the state of the training manager. This updates the state which is used to determine what preprocessing steps, if any, are applied to the data. Valid states include: 1. `None`: This means that all of the transforms and image resizing will be applied to the data, but no automatic batching or tensor conversion. This is also the default state if none is specified. 2. 'tf': This is one of the two training modes. This applies transforms, image resizing, as well as automatic batching and tensor conversion. 3. 'torch': This is one of the two training modes. This applies transforms, image resizing, tensor conversion, as well as image formatting, but no automatic batching. 4. 'eval': By default, this only enables image resizing. However, if the loader is set to a 'tf' or a 'torch' state, and from here it is converted to 'eval', then it also keeps potential tensor conversion, automatic batching, and image formatting. 5. `False`: This disables all preprocessing and simply returns the raw loaded images and annotations. There are certain caveats here. Specifically, if you want to use 'eval' mode but maintain the tensor conversion and other related features, you need to use the following order of methods. > loader.as_keras_sequence() # or loader.as_torch_dataset() > loader.eval() If you want to completely disable preprocessing but keep the image resizing, then you need to first disable all preprocessing, then set the loader in 'eval' mode. > loader.as_keras_sequence() # or loader.as_torch_dataset() > loader.disable_preprocessing() > loader.eval() To re-enable preprocessing at this point, e.g., just the transforms and resizing but no tensor conversion or automatic batching, then use the following, which resets the train state of the loader. > loader.reset_processing() This enables the loader to track multiple states. """ # Fully disable all preprocessing. if t_(state) == TrainState.FALSE: self._state = TrainState.FALSE # Set the correct 'eval' mode, based on the prior state. elif t_(state) == TrainState.EVAL: if self._state == TrainState.TF: self._state = TrainState.EVAL_TF elif self._state == TrainState.TORCH: self._state = TrainState.EVAL_TORCH else: self._state = TrainState.EVAL # Apply a 'tf' or 'torch' backend conversion. elif t_(state) == TrainState.TORCH: self._state = TrainState.TORCH if get_backend() == 'tf': if user_changed_backend(): raise StrictBackendError( change = 'torch', obj = t_(state)) set_backend('torch') self._resize_manager.assign('train-auto') elif t_(state) == TrainState.TF: self._state = TrainState.TF if get_backend() == 'torch': if user_changed_backend(): raise StrictBackendError( change = 'tf', obj = t_(state)) set_backend('tf') self._resize_manager.assign('train-auto') # Set the default conversion (`None`). elif t_(state) == TrainState.NONE: self._state = TrainState.NONE def _set_annotation_remap_hook(self, hook): """Used to modify class annotations for multi-dataset loaders.""" self._remap_hook = hook def apply(self, obj, batch_state): """Applies preprocessing and conversions to the data contents. This method is responsible for actually loading and processing the data according to the training state, including loading the data, applying transforms and resizing, as well as the training management as described in the class. This is called by an enclosing `DataManager`. See the `TransformManager` and the `ImageResizeManager` for more information on the specific preprocessing applied there, and the `update_state()` method for more information on the training. """ # Extract the raw contents from the `DataObject`. contents = obj.get() # If there is a hook to apply (for multi-dataset loaders), # then apply the hook before doing anything else. if self._remap_hook: contents = self._remap_hook(contents, self._name) # noqa # If the state is set to `False`, then just return the raw contents. if self._state is TrainState.FALSE: return contents # In any other case other than `False`, we resize the images. contents = self._resize_manager.apply(contents) # If we are in a training state or `None`, (so not an evaluation # state or `False`), then we apply the transforms to the images. if self._state not in [TrainState.EVAL, TrainState.EVAL_TF, TrainState.EVAL_TORCH]: contents = self._transform_manager.apply(contents) # If the images are not in a batch, then we convert them to tensors # here, otherwise, they will be converted when the batch is created. if not batch_state: contents = self._train_state_apply(contents) # Return the processed contents. return contents def _train_state_apply(self, contents): """Preprocesses the data according to the class's training state.""" if self._state is TrainState.NONE: return contents elif self._state in [TrainState.TF, TrainState.EVAL_TF]: return self._tf_tensor_convert(contents, self._task) elif self._state in [TrainState.TORCH, TrainState.EVAL_TORCH]: return self._torch_tensor_convert(contents, self._task) return contents def make_batch(self, images, annotations): """Creates a batch of data out of processed images and annotations.""" if self._state in [TrainState.NONE, TrainState.FALSE, TrainState.EVAL]: return images, annotations elif self._state in [TrainState.TF, TrainState.EVAL_TF]: return self._tf_tensor_batch_convert( (images, annotations), self._task) elif self._state in [TrainState.TORCH, TrainState.EVAL_TORCH]: return self._torch_tensor_batch_convert( (images, annotations), self._task) return images, annotations @staticmethod def _tf_tensor_image_convert(image): """Converts potential multi-image input dicts to tensors.""" if isinstance(image, np.ndarray): return tf.constant(image) return {k: tf.constant(i) for k, i in image.items()} @staticmethod def _tf_tensor_image_batch_convert(batch): """Converts potential multi-image input batch dicts to tensor dicts.""" if isinstance(batch, np.ndarray): return tf.stack(batch) return {k: tf.stack(b) for k, b in batch.items()} @staticmethod def _tf_tensor_convert(contents, task): """Converts contents to `tf.Tensor`s where possible.""" # Convert the image and annotation to `tf.Tensor`s. image, annotation = contents image = TrainingManager._tf_tensor_image_convert(image) if task in ['image_classification', 'image_regression', 'semantic_segmentation']: if isinstance(annotation, (int, np.ndarray)): annotation = tf.constant(annotation) else: for k, v in annotation.items(): annotation[k] = tf.constant(v) elif task == 'object_detection': annotation = TrainingManager._tf_tensor_coco_convert( annotation) # Add a first-dimension batch to the image. if isinstance(image, tf.Tensor): if needs_batch_dim(image): image = tf.expand_dims(image, axis = 0) else: for k, v in image.items(): if needs_batch_dim(v): image[k] = tf.expand_dims(v, axis = 0) return image, annotation @staticmethod def _tf_tensor_batch_convert(contents, task): """Converts batch contents to `tf.Tensor`s where possible.""" # This stacks the images and annotations together. images, annotations = contents images = TrainingManager._tf_tensor_image_batch_convert(images) if task == 'image_classification': annotations = tf.constant(annotations) elif task == 'image_regression': if isinstance(annotations, np.ndarray): annotations = tf.constant(annotations) else: for k, v in annotations.items(): annotations[k] = tf.constant(v) elif task == 'semantic_segmentation': annotations = tf.stack(annotations, axis = 0) elif task == 'object_detection': annotations = [TrainingManager._tf_tensor_coco_convert( a_set) for a_set in annotations] return images, annotations @staticmethod def _tf_tensor_coco_convert(contents): """Converts a COCO JSON dictionary to a `tf.Tensor`.""" coco_tensor = {} for key, value in contents.items(): coco_tensor[key] = tf.constant(value) return coco_tensor @staticmethod def _torch_tensor_image_convert(image): """Converts potential multi-image input dicts to tensors.""" if is_array_like(image): return _convert_image_to_torch(image) return {k: _convert_image_to_torch(i) for k, i in image.items()} @staticmethod def _torch_tensor_image_batch_convert(batch): """Converts potential multi-image input batch dicts to tensor dicts.""" if is_array_like(batch): return torch.stack([ _convert_image_to_torch(image) for image in batch]) return {k: TrainingManager. _torch_tensor_image_batch_convert(b) for k, b in batch.items()} @staticmethod def _torch_tensor_convert(contents, task): """Converts contents to `torch.Tensor`s where possible.""" image, annotation = contents image = TrainingManager._torch_tensor_image_convert(image) if task in ['image_classification', 'image_regression']: if isinstance(annotation, (int, np.ndarray)): annotation = torch.tensor(annotation) else: for k, v in annotation.items(): annotation[k] = torch.tensor(v) elif task == 'semantic_segmentation': annotation = _convert_image_to_torch(annotation) elif task == 'object_detection': annotation = TrainingManager._torch_tensor_coco_convert( annotation) return image, annotation @staticmethod def _torch_tensor_batch_convert(contents, task): """Converts batch contents to `torch.Tensor`s where possible.""" images, annotations = contents images = TrainingManager._torch_tensor_image_batch_convert(images) if task in ['image_classification', 'image_regression']: if isinstance(annotations, (int, np.ndarray)): annotations = torch.tensor(annotations) else: for k, v in annotations.items(): annotations[k] = torch.tensor(v) elif task == 'semantic_segmentation': annotations = torch.stack([ _convert_image_to_torch(a) for a in annotations]) elif task == 'object_detection': annotations = [TrainingManager._torch_tensor_coco_convert( a_set) for a_set in annotations] return images, annotations @staticmethod def _torch_tensor_coco_convert(contents): """Converts a COCO JSON dictionary to a `torch.Tensor`.""" coco_tensor = {} for key, value in contents.items(): if key == 'segmentation': value = np.empty(0) if not isinstance(value, torch.Tensor): coco_tensor[key] = torch.tensor(value) else: coco_tensor[key] = value return coco_tensor
15,928
40.481771
82
py
AgML
AgML-main/agml/data/managers/transforms.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import types import inspect import functools from enum import Enum from agml.framework import AgMLSerializable from agml.backend.tftorch import ( get_backend, set_backend, user_changed_backend, StrictBackendError ) from agml.data.managers.transform_helpers import ( AlbumentationsTransformSingle, AlbumentationsTransformMask, AlbumentationsTransformCOCO, SameStateImageMaskTransform, NormalizationTransformBase, ScaleTransform, NormalizationTransform, OneHotLabelTransform, MaskToChannelBasisTransform, ToTensorSeg ) from agml.utils.logging import log class TransformKind(Enum): Transform = 'transform' TargetTransform = 'target_transform' DualTransform = 'dual_transform' # Shorthand form of the enum for testing explicit values. t_ = TransformKind class TransformManager(AgMLSerializable): """Manages the validation and application of transformations. This class serves as a helper for the `DataManager` class that focuses solely on applying transforms to the input images and annotations. There are three main transforms stored internally. 1. `transform`: A transform which is applied to only the input image. 2. `target_transform`: A transform which is applied only to the output. 3. `dual_transform`: A transform which is applied to both. Note that the values of these transforms don't always directly match the provided value to the corresponding argument in the method `AgMLDataLoader.transform()`. An example is an albumentations transform, which, while being passed to `transform`, ends up being stored as a `dual_transform` internally. The above case applies to semantic segmentation and object detection transformations, but there is no `dual_transform` for image classification. """ serializable = frozenset( ('task', 'transforms', 'time_inserted_transforms', 'contains_tf_transforms', 'contains_torch_transforms')) def __init__(self, task): self._task = task self._transforms = dict() # A user might want to apply a certain set of transforms first, # which are then followed by a different set of transforms. E.g., # calling `loader.transform()` with one set of transforms and then # following with another call to `loader.transform()` with the # intention that the transforms in the second call will only be # applied after all of the transforms in the first call are. # # So, while the transform types are tracked in the `_transforms` # attribute, we track the moment that transforms are inserted # using this attribute, which is a list of different transforms. # The `apply()` method loops sequentially through each transform # in this list and applies them as required. self._time_inserted_transforms = [] # Check if the loader already contains TensorFlow/PyTorch transforms. # This is to track whether there are issues when a new transform is added # or if there already exists one, so we whether to apply transforms. self._contains_tf_transforms = False self._contains_torch_transforms = False def get_transform_states(self): """Returns a copy of the existing transforms.""" transform_dict = {} for name, state in self._transforms.items(): transform_dict[name] = state.copy() return transform_dict def _pop_transform(self, t_type, search_param): """Removes a certain type of transform from the manager.""" for i, tfm in enumerate(self._transforms[search_param]): if isinstance(tfm, t_type): self._transforms[search_param].pop(i) break for i, tfm in enumerate(self._time_inserted_transforms): if isinstance(tfm[1], t_type): self._time_inserted_transforms.pop(i) break def assign(self, kind, transform): """Assigns a new transform to the manager.""" # Determine if the transform is being reset or unchanged. if transform == 'reset': self._transforms.pop(kind, None) new_time_transforms = [] for tfm in self._time_inserted_transforms: if tfm[0] == kind: continue new_time_transforms.append(tfm) self._time_inserted_transforms = new_time_transforms.copy() return elif transform is None: return # If an `albumentations` transform is passed to the `transform` # argument, then it is checked first to see if it is potentially # just processing the input image, and then passed as just a # `transform` (or else it will clash in object detection tasks). # Otherwise, it is stored internally as a `dual_transform`. if transform is not None: try: if 'albumentations' in transform.__module__: if t_(kind) == TransformKind.Transform: if len(transform.processors) != 0: kind = 'dual_transform' if self._task == 'semantic_segmentation': kind = 'dual_transform' except AttributeError: # Some type of object that doesn't have `__module__`. pass # We can only do this after the albumentations check, to ensure # that we are adding the transforms to the correct location. prev = self._transforms.get(kind, None) # Validate the transformation based on the task and kind. if self._task == 'image_classification': if t_(kind) == TransformKind.Transform: transform = self._maybe_normalization_or_regular_transform(transform) elif t_(kind) == TransformKind.TargetTransform: if isinstance(transform, tuple): # a special convenience case if transform[0] == 'one_hot': if transform[2] is not True: # removing the transform self._pop_transform(OneHotLabelTransform, kind) return transform = OneHotLabelTransform(transform[1]) else: raise ValueError("There is no `dual_transform` for image " "classification tasks. Please pass the " "input as a `transform` or `target_transform`.") elif self._task == 'image_regression': if t_(kind) == TransformKind.Transform: transform = self._maybe_normalization_or_regular_transform(transform) elif t_(kind) == TransformKind.TargetTransform: if isinstance(transform, tuple): # a special convenience case if transform[0] == 'one_hot': if transform[2] is not True: # removing the transform self._pop_transform(OneHotLabelTransform, kind) return transform = OneHotLabelTransform(transform[1]) else: pass elif self._task == 'semantic_segmentation': if t_(kind) == TransformKind.Transform: transform = self._maybe_normalization_or_regular_transform(transform) elif t_(kind) == TransformKind.TargetTransform: if isinstance(transform, tuple): # a special convenience case if transform[0] == 'channel_basis': if transform[2] is not True: # removing the transform self._pop_transform(MaskToChannelBasisTransform, kind) transform = MaskToChannelBasisTransform(transform[1]) else: transform = self._maybe_normalization_or_regular_transform(transform) else: transform = self._construct_image_and_mask_transform(transform) elif self._task == 'object_detection': if t_(kind) == TransformKind.Transform: transform = self._maybe_normalization_or_regular_transform(transform) elif t_(kind) == TransformKind.TargetTransform: pass else: transform = self._construct_image_and_coco_transform(transform) # Add the transformation to the internal storage. if transform is not None: if prev is not None: self._transforms[kind].append(transform) else: self._transforms[kind] = [transform] self._time_inserted_transforms.append((kind, transform)) def apply(self, contents): """Applies a transform to a set of input data. This method controls the application of the actual transforms. It does this inside of a context that can control the application of the transform and manage errors more effectively. The method of application is mostly similar across the different tasks, as the differentiation of the methods is taken care of in the `assign` method and each transform is converted to act like a generic method with simple input arguments. The hierarchy in which transforms are applied is: transform -> --------| |-----> dual_transform target_transform -> --| Furthermore, image resizing takes place before any transformations are applied. After the transforms are applied in this order, they returned and if passed again, they will have a different transform applied to them. The state is independent of the images passed. """ image, annotation = contents # Iterate through the different transforms. for (kind, transform) in self._time_inserted_transforms: if t_(kind) == TransformKind.Transform: image = self._apply_to_objects( transform, (image, ), kind) if t_(kind) == TransformKind.TargetTransform: annotation = self._apply_to_objects( transform, (annotation, ), kind) if t_(kind) == TransformKind.DualTransform: image, annotation = self._apply_to_objects( transform, (image, annotation), kind) # Return the processed image and annotation. return image, annotation @staticmethod def _apply_to_objects(transform, contents, kind): """Applies the actual transformations in a context.""" try: return transform(*contents) except Exception as e: default_msg = (f"Encountered an error when attempting to apply " f"a transform ({transform}) of kind '{kind}' to " f"objects: {contents}. See the above traceback.") # A specific case of exception where the image first needs # to be converted to a PIL image before being used in a # general `torchvision.transforms` pipeline. if "PIL" in str(e): raise TypeError("If using a `torchvision.transforms` pipeline " "when not in PyTorch training mode, you need " "to include `ToTensor()` in the pipeline.") # Otherwise, raise the default exception. raise Exception(default_msg) def _maybe_normalization_or_regular_transform(self, transform): """Dispatches to the correct single-image transform construction.""" if isinstance(transform, tuple): if transform[0] == 'normalize': return self._build_normalization_transform(transform) return self._construct_single_image_transform(transform) def _build_normalization_transform(self, transform): """Constructs a normalization transform if passed. This is a special case for transforms passed by the `normalize_images` 'method of the `AgMLDataLoader`, since these are treated almost as their own independent management system in terms of resetting or applying them in a different method. This is called by `assign`. """ # First, we check if a normalization transform already exists # within the transform dict, and then we get its location. norm_transform_index, norm_transform_index_time = -1, -1 try: for i, t in enumerate(self._transforms['transform']): if isinstance(t, NormalizationTransformBase): norm_transform_index = i break for i, (_, t) in enumerate(self._time_inserted_transforms): if isinstance(t, NormalizationTransformBase): norm_transform_index_time = i break except: self._transforms['transform'] = [] if transform[1] == 'scale': tfm = ScaleTransform(None) if norm_transform_index != -1: self._transforms['transform'][norm_transform_index] = tfm self._time_inserted_transforms[norm_transform_index_time] \ = ('transform', tfm) else: self._transforms['transform'].append(tfm) self._time_inserted_transforms.append(('transform', tfm)) elif hasattr(transform[1], 'mean') or transform[1] == 'imagenet': try: mean, std = transform[1].mean, transform[1].std except AttributeError: # Default ImageNet mean and std. mean = [0.485, 0.456, 0.406] std = [0.229, 0.224, 0.225] tfm = NormalizationTransform((mean, std)) if norm_transform_index != -1: self._transforms['transform'][norm_transform_index] = tfm self._time_inserted_transforms[norm_transform_index_time] \ = ('transform', tfm) else: self._transforms['transform'].append(tfm) self._time_inserted_transforms.append(('transform', tfm)) elif transform[1] == 'reset': if norm_transform_index != -1: self._transforms['transform'].pop(norm_transform_index) self._time_inserted_transforms.pop(norm_transform_index_time) return None def _transform_update_and_check(self, tfm_type): """Checks whether a TensorFlow/PyTorch transform has been added.""" # Check for transform conflicts. if tfm_type == 'torch' and self._contains_tf_transforms: raise TypeError("Received a PyTorch-type transform, yet the loader " "already contains TensorFlow/Keras transforms. This " "will cause an error, please only pass one format. If " "you want to remove a transform, pass a value of `None` " "to reset all of the transforms for a certain type.") if tfm_type == 'tf' and self._contains_torch_transforms: raise TypeError("Received a TensorFlow/Keras-type transform, yet the " "loader already contains PyTorch transforms. This " "will cause an error, please only pass one format. If " "you want to remove a transform, pass a value of `None` " "to reset all of the transforms for a certain type.") # Update the transform type. if tfm_type == 'torch': self._contains_torch_transforms = True if tfm_type == 'tf': self._contains_tf_transforms = True # The following methods implement different checks which validate # as well as process input transformations, and manage the backend. # The transforms here will be also checked to match a specific # backend. Alternatively, the backend will dynamically be switched. def _construct_single_image_transform(self, transform): """Validates a transform which is applied to a single image. This is used for image classification transforms, which only apply to the input image, as well as other general tasks whenever a transform that is only applied to the image is passed, e.g., a visual augmentation like random contrasting. """ # This case is used for clearing a transformation. if transform is None: return None # A general functional transformation. We don't verify what happens in # the function. The only check which occurs here that the signature of # the function is valid. Note that this also includes partial methods. elif isinstance(transform, (types.FunctionType, functools.partial)): sig = inspect.signature(transform).parameters if not len(sig) == 1: raise TypeError("Expected a single-image transform passed " "to `transform` to accept one input image, " f"instead got {len(sig)} parameters.") return transform # An `albumentations` transform to be applied to the image. This # wraps the transform into a method which treats it as a regular # functional transform, e.g. no keyword arguments (for easy use). elif 'albumentations' in transform.__module__: return AlbumentationsTransformSingle(transform) # A set of `torchvision` transforms wrapped into a `T.Compose` object # or just a single transformation. This simply confirms the backend. elif 'torchvision' in transform.__module__: if get_backend() != 'torch': if user_changed_backend(): raise StrictBackendError(change = 'tf', obj = transform) set_backend('torch') self._transform_update_and_check('torch') return transform # A `tf.keras.Sequential` preprocessing model or an individual # Keras preprocessing layer. This simply confirms the backend. elif 'keras' in transform.__module__ or 'tensorflow' in transform.__module__: if get_backend() != 'tf': if user_changed_backend(): raise StrictBackendError(change = 'torch', obj = transform) set_backend('tf') self._transform_update_and_check('tf') return transform # Otherwise, it may be a transform from a (lesser-known) third-party # library, in which case we just return it as a callable. Transforms # which are used in a more complex manner should be passed as decorators. return transform def _construct_image_and_mask_transform(self, transform): """Validates a transform for an image and annotation mask. This is used for a semantic segmentation transform. Such transformations should be passed as the following: - An `albumentations` transform pipeline that may include spatial and/or visual augmentation. - A method to independently or dually apply transformations to the image and annotation mask. - A `torchvision.transforms` or `tf.keras.Sequential` pipeline which will be applied to the image and mask using the same random seed, for reproducibility. Use the provided method `generate_keras_segmentation_dual_transform` for this. """ # This case is used for clearing a transformation. if transform is None: return None # A general functional transformation. We don't verify what happens in # the function. The only check which occurs here that the signature of # the function is valid. Note that this also includes partial methods. elif isinstance(transform, (types.FunctionType, functools.partial)): sig = inspect.signature(transform).parameters if not len(sig) == 2: raise TypeError(f"Expected a semantic segmentation transform " f"passed to `transform` to accept two args: " f"an input image and an annotation mask, " f"instead got {len(sig)} parameters.") return transform # An `albumentations` transform to be applied to the image. This # wraps the transform into a method which treats it as a regular # functional transform, e.g. no keyword arguments (for easy use). elif 'albumentations' in transform.__module__: return AlbumentationsTransformMask(transform) # If we have the case of a transform that needs to be applied to # both the input and the output mask simultaneously, then we wrap # that into a class which undertakes that behavior. This happens # when the signature of the input function accepts only one input # parameter or it belongs to `torchvision` transform (not Keras). if len(inspect.signature(transform).parameters) == 1: if 'torchvision' in transform.__module__: if get_backend() != 'torch': if user_changed_backend(): raise StrictBackendError( change = 'tf', obj = transform) set_backend('torch') # Update `torchvision.transforms.ToTensor` to a custom # updated class as this will modify the mask incorrectly. import torchvision if isinstance(transform, torchvision.transforms.ToTensor): transform = ToTensorSeg(None) log("Updated `ToTensor` transform in the provided pipeline " f"{transform} to an updated transform which does not " f"modify the mask. If you want to change this behaviour, " f"please raise an error with the AgML team.") elif isinstance(transform, torchvision.transforms.Compose): tfm_list = transform.transforms.copy() for i, compose_tfm in enumerate(transform.transforms): if isinstance(compose_tfm, torchvision.transforms.ToTensor): tfm_list[i] = ToTensorSeg(None) transform = torchvision.transforms.Compose(tfm_list) log("Updated `ToTensor` transform in the provided pipeline " f"{transform} to an updated transform which does not " f"modify the mask. If you want to change this behaviour, " f"please raise an error with the AgML team.") self._transform_update_and_check('torch') elif 'keras.layers' in transform.__module__: if get_backend() != 'tf': if user_changed_backend(): raise StrictBackendError( change = 'torch', obj = transform) set_backend('tf') log('Got a Keras transformation for a dual image and ' 'mask transform. If you are passing preprocessing ' 'layers to this method, then use `agml.data.experimental' '.generate_keras_segmentation_dual_transform` in order ' 'for the random state to be applied properly.', 'warning') self._transform_update_and_check('tf') return SameStateImageMaskTransform(transform) # Another type of transform, most likely some form of transform # class. No checks are applied here, since we can't account for # each of the potential cases of the transformations. return transform @staticmethod def _construct_image_and_coco_transform(transform): """Validates a transform for an image and COCO JSON dictionary. This is used for object detection transforms. Such transformations should be wrapped into a method (unless they are albumentations transforms). The method should accept two input arguments, the image and the COCO JSON dictionary, and return the two respectively. """ # This case is used for clearing a transformation. if transform is None: return None # A general functional transformation. We don't verify what happens in # the function. The only check which occurs here that the signature of # the function is valid. Note that this also includes partial methods. elif isinstance(transform, (types.FunctionType, functools.partial)): sig = inspect.signature(transform).parameters if not len(sig) == 2: raise TypeError(f"Expected a object detection transform passed " f"to `transform` to accept two args: an input " f"image and a COCO JSON dictionary, instead " f"got {len(sig)} parameters.") return transform # An `albumentations` transform to be applied to the image. This # wraps the transform into a method which treats it as a regular # functional transform, e.g. no keyword arguments (for easy use). elif 'albumentations' in transform.__module__: return AlbumentationsTransformCOCO(transform) # Another type of transform, most likely some form of transform # class. No checks are applied here, since we can't account for # each of the potential cases of the transformations. return transform
26,516
48.288104
89
py
AgML
AgML-main/agml/data/exporters/tensorflow.py
# Copyright 2021 UC Davis Plant AI and Biophysics Lab # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from agml.backend.tftorch import tf from agml.data.managers.training import TrainState from agml.data.object import DataObject from agml.utils.logging import log class TFExporter(object): """Exports an `AgMLDataLoader` as a `tf.data.Dataset`. This class manages the methods and conversions between `AgMLDataLoader` parameters, managers, and methods, into TensorFlow native methods. """ def __init__(self, task, builder): self._task = task self._builder = builder def digest_transforms(self, transforms, resizing): """Parses the transforms for the `AgMLDataLoader`.""" self._transforms = { k: state for k, state in transforms.items()} self._size = resizing if resizing is not None else (512, 512) def assign_state(self, state): """Updates the training state for the `tf.data.Dataset`.""" if state in [TrainState.EVAL, TrainState.EVAL_TF, TrainState.EVAL_TORCH, TrainState.FALSE]: self._state = 'eval' else: self._state = 'train' def _build_from_contents_by_type(self, builder): """Parses the provided mapping into a valid set of contents.""" if self._task == 'image_classification': ds = self._build_image_classification(builder) return ds.map(self._image_classification_load) elif self._task == 'semantic_segmentation': ds = self._build_semantic_segmentation(builder) return ds.map(self._semantic_segmentation_load) else: ds = self._build_object_detection(builder) return ds.map(self._object_detection_load) def _apply_resizing_by_type(self, ds): """Applies resizing based on the task.""" if self._task == 'image_classification': return ds.map(self._image_classification_resize) elif self._task == 'semantic_segmentation': return ds.map(self._semantic_segmentation_resize) else: return ds.map(self._object_detection_resize) def build(self, batch_size = None): """Builds the `tf.data.Dataset` using the provided parameters.""" # Construct the dataset from the contents. ds = self._build_from_contents_by_type(self._builder) # Apply the digested transforms and resizing. ds = self._apply_resizing_by_type(ds) if self._state != 'eval': # No transforms for object detection, since it is near impossible # for TensorFlow's graph mode to use COCO JSON dictionaries. if self._task == 'object_detection': if len(self._transforms) != 0: log("Got transforms when exporting an `AgMLDataLoader`" "to a `tf.data.Dataset`. These transforms will not be " "applied. To use transforms in TensorFlow, use the " "`as_keras_sequence()` method instead.") else: tfm = self._apply_transforms if len(self._transforms) != 0: ds = ds.map(tfm) # Apply batching and prefetching, then return the dataset. if batch_size is not None: ds = ds.batch(batch_size) ds = ds.prefetch(batch_size) return ds # The following methods are used to parse the input # contents into valid methods for the loaders. @staticmethod def _build_image_classification(builder): images, labels = builder.export_contents( export_format = 'arrays') images, labels = tf.constant(images), tf.constant(labels) ds = tf.data.Dataset.from_tensor_slices((images, labels)) return ds.shuffle(len(images)) @staticmethod def _build_semantic_segmentation(builder): images, masks = builder.export_contents( export_format = 'arrays') images, masks = tf.constant(images), tf.constant(masks) ds = tf.data.Dataset.from_tensor_slices((images, masks)) return ds.shuffle(len(images)) @staticmethod def _build_object_detection(builder): images, annotations = builder.export_contents( export_format = 'arrays') images = tf.constant(images) processed_annotations = [ DataObject._parse_coco(a) for a in annotations] features = {'bbox': [], 'category_id': [], 'area': [], 'image_id': [], 'iscrowd': [], 'segmentation': []} for a_set in processed_annotations: for feature in features.keys(): features[feature].append(a_set[feature]) # noqa for feature in features.keys(): features[feature] = tf.ragged.constant(features[feature]) feature_ds = tf.data.Dataset.from_tensor_slices(features) ds = tf.data.Dataset.zip(( tf.data.Dataset.from_tensor_slices(images), feature_ds)) return ds.shuffle(len(images)) # The following methods are used for loading images and # annotations for each of the different types of tasks. @staticmethod @tf.function def _image_classification_load(image, label): image = tf.cast(tf.image.decode_jpeg( tf.io.read_file(image)), tf.float32) / 255. return image, tf.convert_to_tensor(label) @staticmethod @tf.function def _semantic_segmentation_load(image, mask): image = tf.cast(tf.image.decode_jpeg( tf.io.read_file(image)), tf.float32) / 255. mask = tf.image.decode_jpeg(tf.io.read_file(mask)) return image, mask @staticmethod @tf.function def _object_detection_load(image, coco): image = tf.cast(tf.image.decode_jpeg( tf.io.read_file(image)), tf.float32) / 255. ret_coco = coco.copy() for key in coco.keys(): try: ret_coco[key] = coco[key].to_tensor() except: pass return image, ret_coco # The following methods apply resizing to the data. def _image_classification_resize(self, image, label): image = (tf.image.resize( image, self._size, method = 'nearest'), tf.float32) return image, label def _semantic_segmentation_resize(self, image, mask): image = tf.cast(tf.image.resize( image, self._size, method = 'nearest'), tf.float32) mask = tf.cast(tf.image.resize( mask, self._size, method = 'nearest'), tf.float32) return image, mask def _object_detection_resize(self, image, coco): # Helper for the `tf.py_function` for object detection. def _resize_image_and_bboxes(image, coco_boxes): nonlocal size y_scale, x_scale = image.shape[0:2] stack_boxes = tf.stack( [coco_boxes[:, 0] / x_scale, coco_boxes[:, 1] / y_scale, coco_boxes[:, 2] / x_scale, coco_boxes[:, 3] / y_scale], axis = -1) image = tf.cast(tf.image.resize( image, size), tf.float32) y_new, x_new = image.shape[0:2] new_stack = tf.cast(tf.stack( [stack_boxes[:, 0] * x_new, stack_boxes[:, 1] * y_new, stack_boxes[:, 2] * x_new, stack_boxes[:, 3] * y_new], axis = -1 ), tf.int32) areas = new_stack[:, 2] * new_stack[:, 3] return image, new_stack, areas # The actual resizing can't take place in graph mode, so we # dispatch to a `tf.py_function` to do the resizing, then # re-assign the values back to the COCO JSON dictionary. size = self._size image, ret_coco_boxes, ret_areas = tf.py_function( _resize_image_and_bboxes, [image, coco['bbox']], [tf.float32, tf.int32, tf.int32]) coco['bbox'] = ret_coco_boxes coco['area'] = ret_areas return image, coco # The following method manages the application of transforms. def _apply_transforms(self, image, annotation): # Helper for the `tf.py_function` for most transforms. def _py_apply(img, ann): nonlocal transforms img, ann = img.numpy(), ann.numpy() for key, state in transforms.items(): if key == 'transform': for t in state: img = t(img) elif key == 'target_transform': for t in state: ann = t(ann) else: for t in state: img, ann = t(img, ann) return img, ann # The actual transforming can't take place in graph mode # (in most cases), so we dispatch and reassign. transforms = self._transforms image, annotation = tf.py_function( _py_apply, [image, annotation], [tf.float32, tf.int32]) return image, annotation
9,639
38.02834
79
py
NeLLoC
NeLLoC-main/utils.py
''' Code by Hrituraj Singh Indian Institute of Technology Roorkee ''' from torchvision import datasets, transforms import configparser import os import torch.nn.functional as F import torch import numpy as np import matplotlib.pyplot as plt from torch.autograd import Variable from torch.nn.utils import weight_norm as wn from PIL import * import os import pickle normalize=lambda x: x/255. def unpickle(file): with open(file, 'rb') as fo: dict = pickle.load(fo) return dict class CustomDataSet(torch.utils.data.Dataset): def __init__(self, main_dir, transform): self.main_dir = main_dir self.transform = transform self.all_imgs = [] for (dirpath, dirnames, filenames) in os.walk(self.main_dir): self.all_imgs.extend([os.path.join(dirpath, file) for file in filenames]) self.all_imgs=sorted(self.all_imgs) def __len__(self): return len(self.all_imgs) def __getitem__(self, idx): img_loc = self.all_imgs[idx] # print(img_loc) image = Image.open(img_loc).convert("RGB") tensor_image = self.transform(image) return tensor_image def parse_value_from_string(val_str): if(is_int(val_str)): val = int(val_str) elif(is_float(val_str)): val = float(val_str) elif(is_list(val_str)): val = parse_list(val_str) elif(is_bool(val_str)): val = parse_bool(val_str) else: val = val_str return val def is_int(val_str): start_digit = 0 if(val_str[0] =='-'): start_digit = 1 flag = True for i in range(start_digit, len(val_str)): if(str(val_str[i]) < '0' or str(val_str[i]) > '9'): flag = False break return flag def is_float(val_str): flag = False if('.' in val_str and len(val_str.split('.'))==2): if(is_int(val_str.split('.')[0]) and is_int(val_str.split('.')[1])): flag = True else: flag = False elif('e' in val_str and len(val_str.split('e'))==2): if(is_int(val_str.split('e')[0]) and is_int(val_str.split('e')[1])): flag = True else: flag = False else: flag = False return flag def is_bool(var_str): if( var_str=='True' or var_str == 'true' or var_str =='False' or var_str=='false'): return True else: return False def parse_bool(var_str): if(var_str=='True' or var_str == 'true' ): return True else: return False def is_list(val_str): if(val_str[0] == '[' and val_str[-1] == ']'): return True else: return False def parse_list(val_str): sub_str = val_str[1:-1] splits = sub_str.split(',') output = [] for item in splits: item = item.strip() if(is_int(item)): output.append(int(item)) elif(is_float(item)): output.append(float(item)) elif(is_bool(item)): output.append(parse_bool(item)) else: output.append(item) return output def show_many(image,number_sqrt,dim=32, channels=3): if image.size(1)==3: image=image.permute(0,2,3,1) canvas_recon = np.empty((dim * number_sqrt, dim * number_sqrt, channels)) count=0 for i in range(number_sqrt): for j in range(number_sqrt): canvas_recon[i * dim:(i + 1) * dim, j * dim:(j + 1) * dim,:] = \ image[count] count+=1 plt.rcParams["axes.grid"] = False plt.figure(figsize=(number_sqrt, number_sqrt)) plt.axis('off') plt.imshow(canvas_recon) plt.show() def grey_show_many(image,number_sqrt): canvas_recon = np.empty((28 * number_sqrt, 28 * number_sqrt)) count=0 for i in range(number_sqrt): for j in range(number_sqrt): canvas_recon[i * 28:(i + 1) * 28, j * 28:(j + 1) * 28] = \ image[count].reshape([28, 28]) count+=1 plt.rcParams["axes.grid"] = False plt.figure(figsize=(number_sqrt, number_sqrt)) plt.axis('off') plt.imshow(canvas_recon, origin="upper", cmap="gray") plt.show() def color_sample(net, device,num_sqrt=10): sample = torch.Tensor(num_sqrt**2, 3, 32, 32).to(device) sample.fill_(0) with torch.no_grad(): #Generating images pixel by pixel for i in range(32): for j in range(32): out = net(sample).view(-1,256,3,32,32)[:,:,:,i,j].permute(0,2,1).contiguous() probs = F.softmax(out.view(-1,256), dim=-1).data sample[:,:,i,j] = (torch.multinomial(probs, 1).float() / 255.0).view(-1,3) #Saving images row wise show_many(sample.detach().cpu(),num_sqrt) def discretized_mix_logistic_uniform(x, l, alpha=0.0001): x = x.permute(0, 2, 3, 1) l = l.permute(0, 2, 3, 1) xs = [int(y) for y in x.size()] ls = [int(y) for y in l.size()] nr_mix = int(ls[-1] / 10) logit_probs = l[:, :, :, :nr_mix] l = l[:, :, :, nr_mix:].contiguous().view(xs + [nr_mix * 3]) means = l[:, :, :, :, :nr_mix] log_scales = torch.clamp(l[:, :, :, :, nr_mix:2 * nr_mix], min=-7.) coeffs = F.tanh(l[:, :, :, :, 2 * nr_mix:3 * nr_mix]) x = x.contiguous() x = x.unsqueeze(-1) + Variable(torch.zeros(xs + [nr_mix]).to(x.device), requires_grad=False) m2 = (means[:, :, :, 1, :] + coeffs[:, :, :, 0, :] * x[:, :, :, 0, :]).view(xs[0], xs[1], xs[2], 1, nr_mix) m3 = (means[:, :, :, 2, :] + coeffs[:, :, :, 1, :] * x[:, :, :, 0, :] + coeffs[:, :, :, 2, :] * x[:, :, :, 1, :]).view(xs[0], xs[1], xs[2], 1, nr_mix) means = torch.cat((means[:, :, :, 0, :].unsqueeze(3), m2, m3), dim=3) centered_x = x - means inv_stdv = torch.exp(-log_scales) plus_in = inv_stdv * (centered_x + 1. / 255.) cdf_plus = F.sigmoid(plus_in) min_in = inv_stdv * (centered_x - 1. / 255.) cdf_min = F.sigmoid(min_in) cdf_plus=torch.where(x > 0.999, torch.tensor(1.0).to(x.device),cdf_plus) cdf_min=torch.where(x <- 0.999, torch.tensor(0.0).to(x.device),cdf_min) uniform_cdf_min = ((x+1.)/2*255)/256. uniform_cdf_plus = ((x+1.)/2*255+1)/256. pi=torch.softmax(logit_probs,-1).unsqueeze(-2).repeat(1,1,1,3,1) mix_cdf_plus=((1-alpha)*pi*cdf_plus+(alpha/nr_mix)*uniform_cdf_plus).sum(-1) mix_cdf_min=((1-alpha)*pi*cdf_min+(alpha/nr_mix)*uniform_cdf_min).sum(-1) log_probs =torch.log(mix_cdf_plus-mix_cdf_min) return -log_probs.sum()
6,542
29.863208
96
py
NeLLoC
NeLLoC-main/model.py
import torch.nn as nn import torch import torch.nn.functional as F class MaskedCNN(nn.Conv2d): def __init__(self, mask_type, *args, **kwargs): self.mask_type = mask_type assert mask_type in ['A', 'B'], "Unknown Mask Type" super(MaskedCNN, self).__init__(*args, **kwargs) self.register_buffer('mask', self.weight.data.clone()) _, depth, height, width = self.weight.size() self.mask.fill_(1) if mask_type =='A': self.mask[:,:,height//2,width//2:] = torch.zeros(1) self.mask[:,:,height//2+1:,:] = torch.zeros(1) else: self.mask[:,:,height//2,width//2+1:] = torch.zeros(1) self.mask[:,:,height//2+1:,:] = torch.zeros(1) def forward(self, x): self.weight.data*=self.mask return super(MaskedCNN, self).forward(x) class LocalPixelCNN(nn.Module): """ Network of PixelCNN as described in A Oord et. al. """ def __init__(self, res_num=10, in_kernel = 7, in_channels=3, channels=256, out_channels=256, device=None): super(LocalPixelCNN, self).__init__() self.channels = channels self.layers = {} self.device = device self.res_num=res_num self.in_cnn=MaskedCNN('A',in_channels,channels, in_kernel, 1, in_kernel//2, bias=False) self.activation=nn.ReLU() self.resnet_cnn11=torch.nn.ModuleList([MaskedCNN('B',channels,channels, 1, 1, 0) for i in range(0,res_num)]) self.resnet_cnn3=torch.nn.ModuleList([MaskedCNN('B',channels,channels, 1, 1, 0) for i in range(0,res_num)]) self.resnet_cnn12=torch.nn.ModuleList([MaskedCNN('B',channels,channels, 1, 1, 0) for i in range(0,res_num)]) self.out_cnn1=nn.Conv2d(channels, channels, 1) self.out_cnn2=nn.Conv2d(channels, out_channels, 1) def forward(self, x): x=self.in_cnn(x) x=self.activation(x) for i in range(0, self.res_num): x_mid=self.resnet_cnn11[i](x) x_mid=self.activation(x_mid) x_mid=self.resnet_cnn3[i](x_mid) x_mid=self.activation(x_mid) x_mid=self.resnet_cnn12[i](x_mid) x_mid=self.activation(x_mid) x=x+x_mid x=self.out_cnn1(x) x=self.activation(x) x=self.out_cnn2(x) return x
2,366
31.875
116
py
NeLLoC
NeLLoC-main/train.py
import os import torch from torch import optim from torch.utils import data import torch.nn as nn from model import * import numpy as np import torchvision import torch.nn.functional as F from torch.optim import lr_scheduler from torchvision import datasets, transforms from torch.autograd import Variable from utils import * transform_train = transforms.Compose([ transforms.RandomHorizontalFlip(p=0.5), transforms.ToTensor() ]) trainset = torchvision.datasets.CIFAR10(root='../data/cifar10', train=True, download=False, transform=transform_train) train = data.DataLoader(trainset, batch_size=100, shuffle=True, num_workers=3) testset = torchvision.datasets.CIFAR10(root='../data/cifar10', train=False, download=False, transform=transform_train) test = data.DataLoader(testset, batch_size=1000, shuffle=True, num_workers=3) device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") net = LocalPixelCNN(res_num=0, in_kernel = 7, in_channels=3, channels=256, out_channels=100).to(device) optimizer = optim.Adam(net.parameters(), lr=3e-4) scheduler = lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.99995) criterion = lambda real, fake : discretized_mix_logistic_uniform(real, fake, alpha=0.0001) rescaling = lambda x : (x - .5) * 2. rescaling_inv = lambda x : .5 * x + .5 for e in range(1001): print('epoch',e) net.train() for images, labels in train: images = rescaling(images).to(device) optimizer.zero_grad() output = net(images) loss = criterion(images, output) loss.backward() optimizer.step() scheduler.step() with torch.no_grad(): net.eval() bpd_cifar_sum=0. for i, (images, labels) in enumerate(test): images = rescaling(images).to(device) output = net(images) loss = criterion(images, output).item() bpd_cifar_sum+=loss/(np.log(2.)*(1000*32*32*3)) bpd_cifar=bpd_cifar_sum/10 print('bpd_cifar',bpd_cifar) save_path='./model_save/' torch.save(net.state_dict(), save_path+'rs0_cifar_ks7.pt')
2,146
26.883117
118
py
NeLLoC
NeLLoC-main/coder/distributions.py
import torch import numpy as np rescaling = lambda x : (x - .5) * 2. rescaling_inv = lambda x : .5 * x + .5 def discretized_mix_logistic_cdftable(means, log_scales,pi, alpha=0.0001): x=rescaling(torch.arange(0,256)/255.).view(256,1).repeat(1,10).to(means.device) centered_x = x - means inv_stdv = torch.exp(-log_scales) plus_in = inv_stdv * (centered_x + 1. / 255.) cdf_plus = torch.sigmoid(plus_in) min_in = inv_stdv * (centered_x - 1. / 255.) cdf_min = torch.sigmoid(min_in) cdf_plus=torch.where(x > 0.999, torch.tensor(1.0).to(x.device),cdf_plus) cdf_min=torch.where(x <- 0.999, torch.tensor(0.0).to(x.device),cdf_min) uniform_cdf_min = ((x+1.)/2*255)/256. uniform_cdf_plus = ((x+1.)/2*255+1)/256. mix_cdf_plus=((1-alpha)*pi*cdf_plus+(alpha/10)*uniform_cdf_plus).sum(-1) mix_cdf_min=((1-alpha)*pi*cdf_min+(alpha/10)*uniform_cdf_min).sum(-1) return mix_cdf_plus,mix_cdf_min def compute_stats(l): nr_mix=10 pi=torch.softmax(l[:nr_mix],-1) l=l[nr_mix:].view(3,30) means=l[:,:nr_mix] log_scales = torch.clamp(l[:,nr_mix:2 * nr_mix], min=-7.) coeffs = torch.tanh(l[:,2 * nr_mix:3 * nr_mix]) return means,coeffs,log_scales, pi def cdf_table_processing(cdf_plus,cdf_min,p_prec): p_total=1<<p_prec cdf_min=np.rint(cdf_min.numpy()* p_total) cdf_plus=np.rint(cdf_plus.numpy()* p_total) probs=cdf_plus-cdf_min probs[probs==0]=1 probs[np.argmax(probs)]+=(p_total-np.sum(probs)) return np.concatenate(([0],np.cumsum(probs)[:-1])),probs
1,553
33.533333
83
py
NeLLoC
NeLLoC-main/coder/ans_coder.py
from coder.distributions import * import numpy as np class ANSStack(object): def __init__(self, s_prec , t_prec, p_prec): self.s_prec=s_prec self.t_prec=t_prec self.p_prec=p_prec self.t_mask = (1 << t_prec) - 1 self.s_min=1 << s_prec - t_prec self.s_max=1 << s_prec self.s, self.t_stack= self.s_min, [] def push(self,c_min,p): while self.s >= p << (self.s_prec - self.p_prec): self.t_stack.append(self.s & self.t_mask ) self.s=self.s>> self.t_prec self.s = (self.s//p << self.p_prec) + self.s%p + c_min assert self.s_min <= self.s < self.s_max def pop(self): return self.s & ((1 << self.p_prec) - 1) def update(self,s_bar,c_min,p): self.s = p * (self.s >> self.p_prec) + s_bar - c_min while self.s < self.s_min: t_top=self.t_stack.pop() self.s = (self.s << self.t_prec) + t_top assert self.s_min <= self.s < self.s_max def get_length(self): return len(self.t_stack)*self.t_prec+len(bin(self.s)) def get_length(s,t_stack): return ((len(t_stack))*16+len(bin(s)))/(32*32*3) def cpu_ans_compression(model,img,h,w,k,p_prec=16): c_list=[] p_list=[] rf=int(k/2) with torch.no_grad(): for i in range(0,h): for j in range(0,w): up=max(0,i-rf) left=max(0,j-rf) down=i+1 right=j+1+int(i>0)*rf m,n=min(rf,i),min(rf,j) patch_int=img[:,:,up:down,left:right] model_output=model(rescaling(patch_int/255.)) means,coeffs,log_scales, pi=compute_stats(model_output[0,:,m,n].view(-1)) c_0=rescaling(int(img[0,0,i,j])/255.) c_1=rescaling(int(img[0,1,i,j])/255.) for c in range(0,3): if c==0: mean=means[0:1, :] elif c==1: mean=means[1:2, :] + coeffs[0:1, :]* c_0 else: mean=means[2:3, :] + coeffs[1:2, :]* c_0 +coeffs[ 2:3, :] * c_1 cdf_min_table,probs_table= cdf_table_processing(*discretized_mix_logistic_cdftable(mean,log_scales[c:c+1],pi),p_prec) c_list.append(int(cdf_min_table[patch_int[0,c,m,n]])) p_list.append(int(probs_table[patch_int[0,c,m,n]])) ans_stack=ANSStack(s_prec = 32,t_prec = 16, p_prec=p_prec) for i in np.arange(len(c_list)-1,-1,-1): c_min,p=c_list[i],p_list[i] ans_stack.push(c_min,p) return ans_stack def cpu_ans_decompression(model,ans_stack,h,w,k,p_prec=16): with torch.no_grad(): rf=int(k/2) decode_img=torch.zeros([1,3,h,w]) for i in range(0,h): for j in range(0,w): up=max(0,i-rf) left=max(0,j-rf) down=i+1 right=j+1+int(i>0)*rf patch=decode_img[:,:,up:down,left:right] m,n=min(rf,i),min(rf,j) model_output=model(rescaling(patch/255.)) means,coeffs,log_scales, pi=compute_stats(model_output[0,:,m,n].view(-1)) c_vector=[0,0,0] for c in range(0,3): if c==0: mean=means[0:1, :] elif c==1: mean=means[1:2, :] + coeffs[0:1, :]* c_vector[0] else: mean=means[2:3, :] + coeffs[1:2, :]* c_vector[0] +coeffs[2:3, :] * c_vector[1] cdf_min_table,probs_table= cdf_table_processing(*discretized_mix_logistic_cdftable(mean,log_scales[c:c+1],pi),p_prec) s_bar = ans_stack.pop() pt=np.searchsorted(cdf_min_table, s_bar, side='right', sorter=None)-1 decode_img[0,c,i,j]=pt c_vector[c]=torch.tensor(rescaling(pt/255.)) # patch[0,c,m,n]=pt/255. c,p=int(cdf_min_table[pt]),int(probs_table[pt]) ans_stack.update(s_bar,c,p) return decode_img[0]
4,241
36.875
137
py
NeLLoC
NeLLoC-main/coder/ac_coder.py
from model import * from decimal import * from coder.distributions import * tensor2decimal= lambda x : Decimal(str(x.cpu().item())) def bin_2_float(binary): prob = Decimal('0.0') cur_prob=Decimal('0.5') for i in binary: prob=prob+cur_prob* int(i) cur_prob*=Decimal('0.5') return prob def range_2_bin(low, high): code = [] prob = Decimal('0.0') cur_prob=Decimal('0.5') while(prob < low): acc_prob=prob+cur_prob if acc_prob > high: code.append(0) else: code.append(1) prob = acc_prob cur_prob*=Decimal('0.5') return code def cpu_ac_compression(model,img,k): with torch.no_grad(): model.eval() device=next(model.parameters()).device rf=int(k/2) size=img.size() total_size=size[1]*size[2]*size[3] old_low = Decimal('0.0') old_high = Decimal('1.0') _range = Decimal('1.0') for i in range(0,size[2] ): for j in range(0,size[3]): up=max(0,i-rf) left=max(0,j-rf) down=i+1 right=j+1+int(i>0)*rf patch=rescaling(img[:,:,up:down,left:right]/255.) m,n=min(rf,i),min(rf,j) model_output=model(patch.to(device)) means,coeffs,log_scales, pi=compute_stats(model_output[0,:,m,n].view(-1)) c_0=rescaling(int(img[0,0,i,j])/255.) c_1=rescaling(int(img[0,1,i,j])/255.) for c in range(0,3): if c==0: mean=means[0:1, :] elif c==1: mean=means[1:2, :] + coeffs[0:1, :]* c_0 else: mean=means[2:3, :] + coeffs[1:2, :]* c_0 +coeffs[ 2:3, :] * c_1 cdf_plus,cdf_min= discretized_mix_logistic_cdftable(mean,log_scales[c:c+1],pi) low = old_low + _range * tensor2decimal(cdf_min[int(img[0,c,i,j])]) high = old_low + _range * tensor2decimal(cdf_plus[int(img[0,c,i,j])]) _range = high - low old_low = low old_high = high code=range_2_bin(low,high) return code def cpu_ac_decompression(model,code,h,w,k): model.eval() device=next(model.parameters()).device with torch.no_grad(): prob = bin_2_float(code) low = Decimal(0.0) high = Decimal(1.0) _range = Decimal(1.0) rf=int(k/2) decode_img=torch.zeros([1,3,h,w]) for i in range(0,h): for j in range(0,w): up=max(0,i-rf) left=max(0,j-rf) down=i+1 right=j+1+int(i>0)*rf patch=decode_img[:,:,up:down,left:right].clone() m,n=min(rf,i),min(rf,j) model_output=model(rescaling(patch/255.).to(device)) means,coeffs,log_scales, pi=compute_stats(model_output[0,:,m,n].view(-1)) c_vector=[0,0,0] for c in range(0,3): if c==0: mean=means[0:1, :] elif c==1: mean=means[1:2, :] + coeffs[0:1, :]* c_vector[0] else: mean=means[2:3, :] + coeffs[1:2, :]* c_vector[0] +coeffs[2:3, :] * c_vector[1] cdf_plus,cdf_min= discretized_mix_logistic_cdftable(mean,log_scales[c:c+1],pi) s=128 bl=0 br=256 for bs in range(0,9): if tensor2decimal(cdf_min[s])>prob: br=s s=int((s+bl)/2) elif tensor2decimal(cdf_plus[s])<prob: bl=s s=int((s+br)/2) else: decode_img[0,c,i,j]=s low=tensor2decimal(cdf_min[s]) high=tensor2decimal(cdf_plus[s]) c_vector[c]=torch.tensor(rescaling(s/255.)) _range=high-low prob=(prob-low)/_range break return decode_img[0]
4,462
33.596899
103
py
NeLLoC
NeLLoC-main/batch_coder/distributions.py
import torch import numpy as np rescaling = lambda x : (x - .5) * 2. rescaling_inv = lambda x : .5 * x + .5 def discretized_mix_logistic_cdftable(means, log_scales,pi, alpha=0.0001): bs=means.size(0) pi=pi.unsqueeze(1) x=rescaling(torch.arange(0,256)/255.).view(1,256,1).repeat(bs,1,10) centered_x = x - means inv_stdv = torch.exp(-log_scales) plus_in = inv_stdv * (centered_x + 1. / 255.) cdf_plus = torch.sigmoid(plus_in) min_in = inv_stdv * (centered_x - 1. / 255.) cdf_min = torch.sigmoid(min_in) cdf_plus=torch.where(x > 0.999, torch.tensor(1.0).to(x.device),cdf_plus) cdf_min=torch.where(x <- 0.999, torch.tensor(0.0).to(x.device),cdf_min) uniform_cdf_min = ((x+1.)/2*255)/256. uniform_cdf_plus = ((x+1.)/2*255+1)/256. mix_cdf_plus=((1-alpha)*pi*cdf_plus+(alpha/10)*uniform_cdf_plus).sum(-1) mix_cdf_min=((1-alpha)*pi*cdf_min+(alpha/10)*uniform_cdf_min).sum(-1) return mix_cdf_plus,mix_cdf_min def compute_stats(l): bs=l.size(0) nr_mix=10 pi=torch.softmax(l[:,:nr_mix],-1) l=l[:,nr_mix:].view(bs,3,30) means=l[:,:,:nr_mix] log_scales = torch.clamp(l[:,:,nr_mix:2 * nr_mix], min=-7.) coeffs = torch.tanh(l[:,:,2 * nr_mix:3 * nr_mix]) return means,coeffs,log_scales, pi def get_mean_c1(means,mean_linear,x): return means+x.unsqueeze(-1)*mean_linear def get_mean_c2(means,mean_linear,x): return means+torch.bmm(x.view(-1,1,2),mean_linear.view(-1,2,10)).view(-1,1,10) def cdf_table_processing(cdf_plus,cdf_min,p_prec): p_total=np.asarray((1 << p_prec),dtype='uint32') bs=cdf_plus.size(0) cdf_min=np.rint(cdf_min.numpy()* p_total).astype('uint32') cdf_plus=np.rint(cdf_plus.numpy()* p_total).astype('uint32') probs=cdf_plus-cdf_min probs[probs==0]=1 argmax_index=np.argmax(probs,axis=1).reshape(-1,1) diff=p_total-np.sum(probs,-1,keepdims=True) value=diff+np.take_along_axis(probs, argmax_index.reshape(-1,1), axis=-1) np.put_along_axis(probs, argmax_index,value , axis=-1) return np.concatenate((np.zeros((bs,1),dtype='uint32'),np.cumsum(probs[:,:-1],axis=-1,dtype='uint32')),1),probs
2,159
35
115
py
NeLLoC
NeLLoC-main/batch_coder/ians_coder.py
import torch import numpy as np from batch_coder.distributions import * s_prec = 64 t_prec = 32 t_mask = (1 << t_prec) - 1 s_min = 1 << s_prec - t_prec s_max = 1 << s_prec s_prec_u,t_prec_u=np.uint8(s_prec),np.uint8(t_prec) def get_length(s,t_stack): return ((len(t_stack))*t_prec+sum(len(bin(i)) for i in s))/(32*32*3) def cpu_ans_compression(model,img,h=32,w=32,k=7,p_prec=20): p_prec_u=np.uint8(p_prec) c_list=[] p_list=[] rf=int(k/2) with torch.no_grad(): for i in range(0,h): for j in range(0,w): up=max(0,i-rf) left=max(0,j-rf) down=i+1 right=j+1+int(i>0)*rf m,n=min(rf,i),min(rf,j) patch_int=img[:,:,up:down,left:right] patch=rescaling(patch_int/255.) model_output=model(patch) means,coeffs,log_scales, pi=compute_stats(model_output[:,:,m,n].view(img.size(0),-1)) c_0=rescaling(img[:,0:1,i,j]/255.) c_1=rescaling(img[:,0:2,i,j]/255.) for c in range(0,3): if c==0: mean=means[:,0:1, :] elif c==1: mean=get_mean_c1(means[:,1:2, :], coeffs[:,0:1, :],c_0) else: mean=get_mean_c2(means[:,2:3, :], coeffs[:,1:3, :],c_1) cdf_min_table,probs_table= cdf_table_processing(*discretized_mix_logistic_cdftable(mean,log_scales[:,c:c+1],pi),p_prec) c_list.append(np.take_along_axis(cdf_min_table,patch_int[:,c,m,n].numpy().reshape(-1,1),axis=-1).reshape(-1)) p_list.append(np.take_along_axis(probs_table,patch_int[:,c,m,n].numpy().reshape(-1,1),axis=-1).reshape(-1)) s, t_stack = np.asarray([s_min]*img.size(0),dtype='uint64'), [] for i in np.arange(len(c_list)-1,-1,-1): c_min,p=c_list[i],p_list[i] pos= s>>(s_prec - p_prec) >= p while True in pos: t_stack.extend(np.uint32(s[pos])) s[pos]>>= t_prec_u pos= s>>(s_prec_u - p_prec_u) >= p s = (s//p << p_prec_u) + s%p + c_min return s,t_stack def cpu_ans_decompression(model,s,t_stack,bs,h=32,w=32,k=7,p_prec=20): p_prec_u=np.uint8(p_prec) with torch.no_grad(): # device=next(model.parameters()).device rf=int(k/2) decode_img=torch.zeros([bs,3,h,w]) for i in range(0,h): for j in range(0,w): up=max(0,i-rf) left=max(0,j-rf) down=i+1 right=j+1+int(i>0)*rf patch=rescaling(decode_img[:,:,up:down,left:right]/255.) m,n=min(rf,i),min(rf,j) model_output=model(patch) means,coeffs,log_scales, pi=compute_stats(model_output[:,:,m,n].view(bs,-1)) for c in range(0,3): if c==0: mean=means[:,0:1, :] elif c==1: mean=get_mean_c1(means[:,1:2, :], coeffs[:,0:1, :],rescaling(decode_img[:,0:1,i,j]/255.)) else: mean=get_mean_c2(means[:,2:3, :], coeffs[:,1:3, :],rescaling(decode_img[:,0:2,i,j]/255.)) cdf_min_table,probs_table= cdf_table_processing(*discretized_mix_logistic_cdftable(mean,log_scales[:,c:c+1],pi),p_prec) s_bar = s & np.uint64(((1 << p_prec) - 1)) pt=np.asarray([np.searchsorted(cdf_min_table[i], s_bar[i], side='right', sorter=None)-1 for i in range(0,bs)]) decode_img[:,c,i,j]=torch.tensor(pt) patch[:,c,m,n]=torch.tensor(pt/255. ) c_min=np.take_along_axis(cdf_min_table,pt.reshape(-1,1),axis=-1).reshape(-1) p=np.take_along_axis(probs_table,pt.reshape(-1,1),axis=-1).reshape(-1) s = p * (s >> np.uint8(p_prec)) + s_bar - c_min # Renormalize pos= s < s_min while True in pos: t_top=t_stack[-sum(pos):] del t_stack[-sum(pos):] s[pos] = (s[pos] << t_prec_u) + t_top pos= s < s_min # for yo in s: # assert s_min <= yo < s_max return decode_img
4,562
37.025
150
py
onnx
onnx-main/docs/docsgen/source/conf.py
# Copyright (c) ONNX Project Contributors # # SPDX-License-Identifier: Apache-2.0 # pylint: disable=W0622 # type: ignore import os import sys import warnings import onnx sys.path.append(os.path.abspath(os.path.dirname(__file__))) # -- Project information ----------------------------------------------------- author = "ONNX" copyright = "2023" project = "ONNX" release = onnx.__version__ version = onnx.__version__ # define the latest opset to document, # this is meant to avoid documenting opset not released yet max_opset = onnx.helper.VERSION_TABLE[-1][2] # define the latest opset to document for every opset _opsets = [t for t in onnx.helper.VERSION_TABLE if t[2] == max_opset][-1] max_opsets = { '': max_opset, 'ai.onnx.ml': _opsets[3], 'ai.onnx.training': _opsets[4], } # -- General configuration --------------------------------------------------- extensions = [ "myst_parser", "onnx_sphinx", "sphinx_copybutton", "sphinx_exec_code", "sphinx_tabs.tabs", "sphinx.ext.autodoc", "sphinx.ext.autosummary", "sphinx.ext.coverage", "sphinx.ext.doctest", "sphinx.ext.githubpages", "sphinx.ext.graphviz", "sphinx.ext.ifconfig", "sphinx.ext.intersphinx", "sphinx.ext.mathjax", "sphinx.ext.napoleon", "sphinx.ext.viewcode", ] myst_enable_extensions = [ "amsmath", "attrs_inline", "colon_fence", "deflist", "dollarmath", "fieldlist", "html_admonition", "html_image", "linkify", "replacements", "smartquotes", "strikethrough", "substitution", "tasklist", ] coverage_show_missing_items = True exclude_patterns = [] graphviz_output_format = "svg" html_css_files = ["css/custom.css"] html_favicon = "onnx-favicon.png" html_sidebars = {} html_static_path = ["_static"] html_theme = "furo" language = "en" mathdef_link_only = True master_doc = "index" onnx_doc_folder = os.path.join(os.path.abspath(os.path.dirname(__file__)), "operators") pygments_style = "default" source_suffix = [".rst", ".md"] templates_path = ["_templates"] html_context = { "default_mode": "auto", # auto: the documentation theme will follow the system default that you have set (light or dark) } html_theme_options = { "light_logo": "onnx-horizontal-color.png", "dark_logo": "onnx-horizontal-white.png", } intersphinx_mapping = { "numpy": ("https://numpy.org/doc/stable/", None), "python": (f"https://docs.python.org/{sys.version_info.major}/", None), "scipy": ("https://docs.scipy.org/doc/scipy/", None), "torch": ("https://pytorch.org/docs/stable/", None), } sphinx_gallery_conf = { "examples_dirs": ["examples"], "gallery_dirs": ["auto_examples", "auto_tutorial"], "capture_repr": ("_repr_html_", "__repr__"), "ignore_repr_types": r"matplotlib.text|matplotlib.axes", "binder": { "org": "onnx", "repo": ".", "notebooks_dir": "auto_examples", "binderhub_url": "https://mybinder.org", "branch": "master", "dependencies": "./requirements.txt", }, } warnings.filterwarnings("ignore", category=FutureWarning)
3,121
24.590164
125
py
onnx
onnx-main/onnx/tools/net_drawer.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # A library and utility for drawing ONNX nets. Most of this implementation has # been borrowed from the caffe2 implementation # https://github.com/pytorch/pytorch/blob/master/caffe2/python/net_drawer.py # # The script takes two required arguments: # -input: a path to a serialized ModelProto .pb file. # -output: a path to write a dot file representation of the graph # # Given this dot file representation, you can-for example-export this to svg # with the graphviz `dot` utility, like so: # # $ dot -Tsvg my_output.dot -o my_output.svg import argparse import json from collections import defaultdict from typing import Any, Callable, Dict, Optional import pydot from onnx import GraphProto, ModelProto, NodeProto OP_STYLE = { "shape": "box", "color": "#0F9D58", "style": "filled", "fontcolor": "#FFFFFF", } BLOB_STYLE = {"shape": "octagon"} _NodeProducer = Callable[[NodeProto, int], pydot.Node] def _escape_label(name: str) -> str: # json.dumps is poor man's escaping return json.dumps(name) def _form_and_sanitize_docstring(s: str) -> str: url = "javascript:alert(" url += _escape_label(s).replace('"', "'").replace("<", "").replace(">", "") url += ")" return url def GetOpNodeProducer( # noqa: N802 embed_docstring: bool = False, **kwargs: Any ) -> _NodeProducer: def really_get_op_node(op: NodeProto, op_id: int) -> pydot.Node: if op.name: node_name = f"{op.name}/{op.op_type} (op#{op_id})" else: node_name = f"{op.op_type} (op#{op_id})" for i, input_ in enumerate(op.input): node_name += "\n input" + str(i) + " " + input_ for i, output in enumerate(op.output): node_name += "\n output" + str(i) + " " + output node = pydot.Node(node_name, **kwargs) if embed_docstring: url = _form_and_sanitize_docstring(op.doc_string) node.set_URL(url) return node return really_get_op_node def GetPydotGraph( # noqa: N802 graph: GraphProto, name: Optional[str] = None, rankdir: str = "LR", node_producer: Optional[_NodeProducer] = None, embed_docstring: bool = False, ) -> pydot.Dot: if node_producer is None: node_producer = GetOpNodeProducer(embed_docstring=embed_docstring, **OP_STYLE) pydot_graph = pydot.Dot(name, rankdir=rankdir) pydot_nodes: Dict[str, pydot.Node] = {} pydot_node_counts: Dict[str, int] = defaultdict(int) for op_id, op in enumerate(graph.node): op_node = node_producer(op, op_id) pydot_graph.add_node(op_node) for input_name in op.input: if input_name not in pydot_nodes: input_node = pydot.Node( _escape_label(input_name + str(pydot_node_counts[input_name])), label=_escape_label(input_name), **BLOB_STYLE, ) pydot_nodes[input_name] = input_node else: input_node = pydot_nodes[input_name] pydot_graph.add_node(input_node) pydot_graph.add_edge(pydot.Edge(input_node, op_node)) for output_name in op.output: if output_name in pydot_nodes: pydot_node_counts[output_name] += 1 output_node = pydot.Node( _escape_label(output_name + str(pydot_node_counts[output_name])), label=_escape_label(output_name), **BLOB_STYLE, ) pydot_nodes[output_name] = output_node pydot_graph.add_node(output_node) pydot_graph.add_edge(pydot.Edge(op_node, output_node)) return pydot_graph def main() -> None: parser = argparse.ArgumentParser(description="ONNX net drawer") parser.add_argument( "--input", type=str, required=True, help="The input protobuf file.", ) parser.add_argument( "--output", type=str, required=True, help="The output protobuf file.", ) parser.add_argument( "--rankdir", type=str, default="LR", help="The rank direction of the pydot graph.", ) parser.add_argument( "--embed_docstring", action="store_true", help="Embed docstring as javascript alert. Useful for SVG format.", ) args = parser.parse_args() model = ModelProto() with open(args.input, "rb") as fid: content = fid.read() model.ParseFromString(content) pydot_graph = GetPydotGraph( model.graph, name=model.graph.name, rankdir=args.rankdir, node_producer=GetOpNodeProducer( embed_docstring=args.embed_docstring, **OP_STYLE ), ) pydot_graph.write_dot(args.output) if __name__ == "__main__": main()
4,901
30.625806
86
py
onnx
onnx-main/onnx/reference/ops/op_lp_pool.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=W0221,R0913,R0914 import numpy as np from onnx.reference.ops._op_common_pool import CommonPool class LpPool(CommonPool): def _run( # type: ignore self, x, auto_pad=None, ceil_mode=None, dilations=None, kernel_shape=None, p=2, pads=None, strides=None, count_include_pad=None, ): # utilize AvgPool the same fashion Pytorch does. Note that there is a difference in computation. # it needs another PR to address. # https://github.com/pytorch/pytorch/blob/f58ba553b78db7f88477f9ba8c9333bd1590e30a/torch/nn/functional.py#L1015 power_average = CommonPool._run( self, "AVG", count_include_pad, np.power(np.absolute(x), p), auto_pad=auto_pad, ceil_mode=ceil_mode, dilations=dilations, kernel_shape=kernel_shape, pads=pads, strides=strides, ) kernel_element_count = np.prod(kernel_shape) return (np.power(kernel_element_count * power_average[0], 1.0 / p),)
1,216
27.97619
119
py
onnx
onnx-main/onnx/reference/ops/op_hamming_window.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=W0221 import numpy as np from onnx.reference.ops._op_common_window import _CommonWindow class HammingWindow(_CommonWindow): """ Returns :math:`\\omega_n = \\alpha - \\beta \\cos \\left( \\frac{\\pi n}{N-1} \\right)` where *N* is the window length. See `hamming_window <https://pytorch.org/docs/stable/generated/torch.hamming_window.html>`_. `alpha=0.54, beta=0.46` """ def _run(self, size, output_datatype=None, periodic=None): # type: ignore ni, N_1 = self._begin(size, periodic, output_datatype) alpha = 25.0 / 46.0 beta = 1 - alpha res = alpha - np.cos(ni * np.pi * 2 / N_1) * beta return self._end(size, res, output_datatype)
811
29.074074
83
py
onnx
onnx-main/onnx/reference/ops/op_hann_window.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=W0221 import numpy as np from onnx.reference.ops._op_common_window import _CommonWindow class HannWindow(_CommonWindow): """ Returns :math:`\\omega_n = \\sin^2\\left( \\frac{\\pi n}{N-1} \\right)` where *N* is the window length. See `hann_window <https://pytorch.org/docs/stable/generated/torch.hann_window.html>`_ """ def _run(self, size, output_datatype=None, periodic=None): # type: ignore ni, N_1 = self._begin(size, periodic, output_datatype) res = np.sin(ni * np.pi / N_1) ** 2 return self._end(size, res, output_datatype)
690
27.791667
78
py
onnx
onnx-main/onnx/reference/ops/op_col2im.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=R0913,R0914,W0221 import numpy as np from onnx.reference.op_run import OpRun from onnx.reference.ops._op_common_indices import _get_indices, _is_out def _col2im_shape_check_2d(X, output_shape, kernel_shape, dilations, pads, strides): # type: ignore output_height, output_width = output_shape kernel_height, kernel_width = kernel_shape dilation_height, dilation_width = dilations stride_height, stride_width = strides ndim = len(X.shape) if not ( (ndim == 2 and X.shape[0] != 0 and X.shape[1] != 0) or (ndim == 3 and X.shape[1] != 0 and X.shape[2] != 0) ): raise ValueError( "Expected 2D or 3D (batch mode) tensor for input with possibly 0 batch size and non-zero dimensions for input." ) batch_dim = 0 if len(X.shape) == 3 else -1 n_input_plane = X.shape[batch_dim + 1] if n_input_plane % (kernel_width * kernel_height) != 0: raise ValueError( f"Expected size of input's dimension 1 to be divisible by the " f"product of kernel_size, but got input.size(1)={n_input_plane} " f"and kernel_size={kernel_shape}." ) input_length = X.shape[batch_dim + 2] n_blocks_height = ( output_height + pads[0, :].sum() - dilation_height * (kernel_height - 1) - 1 ) // stride_height + 1 n_blocks_width = ( output_width + pads[1, :].sum() - dilation_width * (kernel_width - 1) - 1 ) // stride_width + 1 if input_length != (n_blocks_height * n_blocks_width): raise ValueError( f"Given batch_dim={batch_dim}, n_input_plane={n_input_plane}, X.shape={X.shape}, " f"output_shape={output_shape}, kernel_shape={kernel_shape}, " f"dilations={dilations}, pads={pads}, strides={strides}, " f"expected size of input's dimension 2 to match the calculated number of ", f"sliding blocks {n_blocks_height} * {n_blocks_width} = {n_blocks_height * n_blocks_width}, " f"but got input.size(2)={input_length}.", ) if not (n_blocks_height >= 1 and n_blocks_width >= 1): raise ValueError( f"Given batch_dim={batch_dim}, n_input_plane={n_input_plane}, X.shape={X.shape}, " f"output_shape={output_shape}, kernel_shape={kernel_shape}, " f"dilations={dilations}, pads={pads}, strides={strides}, " f"calculated shape of the array of sliding blocks as ({n_blocks_height}, {n_blocks_width}), " f"which is too small (non-positive)." ) def _col2im_naive_implementation_2d(res, image_shape, kernel_shape, dilations, pads, strides): # type: ignore # source: https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/im2col.h n_dims = len(pads) // 2 new_pads = np.array([(pads[i], pads[i + n_dims]) for i in range(n_dims)]) _col2im_shape_check_2d(res, image_shape, kernel_shape, dilations, new_pads, strides) data_col = res.ravel() data_im = np.zeros(image_shape, dtype=res.dtype).flatten() kernel_h, kernel_w = kernel_shape channels_col = kernel_h * kernel_w stride_h, stride_w = strides dilation_h, dilation_w = dilations pad_h, pad_w = new_pads[:, 0] height, width = image_shape output_height, output_width = image_shape height_col = ( output_height + new_pads[0, :].sum() - (dilation_h * (kernel_h - 1) + 1) ) // stride_h + 1 width_col = ( output_width + new_pads[1, :].sum() - (dilation_w * (kernel_w - 1) + 1) ) // stride_w + 1 for c_col in range(channels_col): w_offset = c_col % kernel_w h_offset = (c_col // kernel_w) % kernel_h c_im = c_col // (kernel_h * kernel_w) for h_col in range(height_col): h_im = h_col * stride_h - pad_h + h_offset * dilation_h for w_col in range(width_col): w_im = w_col * stride_w - pad_w + w_offset * dilation_w if 0 <= h_im < height and 0 <= w_im < width: i_im = (c_im * height + h_im) * width + w_im i_col = (c_col * height_col + h_col) * width_col + w_col if 0 <= i_col < data_col.shape[0]: data_im[i_im] += data_col[i_col] return data_im.reshape(image_shape) def _col2im_shape_check(X, output_shape, kernel_shape, dilations, pads, strides): # type: ignore n_input_plane = X.shape[0] kernel_size = np.prod(kernel_shape) if n_input_plane % kernel_size != 0: raise ValueError( f"Expected size of input's dimension 1 to be divisible by the " f"product of kernel_size={kernel_size}, " f"but got input.size(1)={n_input_plane} " f"and kernel_shape={kernel_shape}, X.shape={X.shape}, output_shape={output_shape}." ) input_length = X.shape[1] n_dims = len(output_shape) n_blocks = [] for i in range(n_dims): n_block = ( output_shape[i] + pads[i, :].sum() - dilations[i] * (kernel_shape[i] - 1) - 1 ) // strides[i] + 1 n_blocks.append(n_block) block_size = np.prod(n_blocks) if input_length != block_size: raise ValueError( f"Given n_input_plane={n_input_plane}, X.shape={X.shape}, " f"output_shape={output_shape}, kernel_shape={kernel_shape}, " f"dilations={dilations}, pads={pads}, strides={strides}, " f"expected size of input's dimension 2 to match the calculated number of " f"sliding blocks {n_blocks} = {block_size}, " f"but got input.size(2)={input_length}.", ) def col2im_naive_implementation(data, image_shape, kernel_shape, dilations, pads, strides): # type: ignore """ Naive implementation for `col2im`. """ n_dims = len(pads) // 2 new_pads = np.array([(pads[i], pads[i + n_dims]) for i in range(n_dims)]) _col2im_shape_check(data, image_shape, kernel_shape, dilations, new_pads, strides) data_col = data data_im = np.zeros(image_shape, dtype=data.dtype) dim_col = [] for i in range(n_dims): col = ( image_shape[i] + new_pads[i, :].sum() - (dilations[i] * (kernel_shape[i] - 1) + 1) ) // strides[i] + 1 dim_col.append(col) kernel_size = np.prod(kernel_shape) col_size = np.prod(dim_col) for c_col in range(kernel_size): offset = _get_indices(c_col, kernel_shape) for col in range(col_size): ind_col = _get_indices(col, dim_col) ind_im = [] for i in range(n_dims): ind = ( ind_col[i] * strides[i] - new_pads[i, 0] + offset[i] * dilations[i] ) ind_im.append(ind) if not _is_out(ind_im, data_im.shape): data_im[tuple(ind_im)] += data_col[c_col, col] return data_im class Col2Im(OpRun): def _run(self, data, image_shape, block_shape, dilations=None, pads=None, strides=None): # type: ignore if dilations is None: dilations = [1 for s in image_shape] if pads is None: pads = [0 for s in image_shape] * 2 if strides is None: strides = [1 for s in image_shape] bl = np.prod(block_shape) C = data.shape[1] // bl data = data.reshape(data.shape[:1] + (C,) + (bl,) + data.shape[2:]) ks = tuple(block_shape) res = None for n in range(data.shape[0]): for c in range(data.shape[1]): out = col2im_naive_implementation( data[n, c, ...], image_shape, ks, dilations, pads, strides ) if res is None: new_shape = data.shape[:2] + out.shape res = np.empty(new_shape, dtype=data.dtype) res[n, c, ...] = out return (res,) # type: ignore
8,046
37.319048
123
py
onnx
onnx-main/onnx/reference/ops/op_resize.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=C0123,C3001,R0912,R0913,R0914,R1730,W0221,W0613 from typing import Any, Callable, List, Optional, Tuple import numpy as np from onnx.reference.op_run import OpRun def _cartesian( arrays: List[np.ndarray], out: Optional[np.ndarray] = None ) -> np.ndarray: """ From https://stackoverflow.com/a/1235363 Generate a cartesian product of input arrays. Parameters ---------- arrays : list of array-like 1-D arrays to form the cartesian product of. out : ndarray Array to place the cartesian product in. Returns ------- out : ndarray 2-D array of shape (M, len(arrays)) containing cartesian products formed of input arrays. Examples -------- >>> cartesian(([1, 2, 3], [4, 5], [6, 7])) array([[1, 4, 6], [1, 4, 7], [1, 5, 6], [1, 5, 7], [2, 4, 6], [2, 4, 7], [2, 5, 6], [2, 5, 7], [3, 4, 6], [3, 4, 7], [3, 5, 6], [3, 5, 7]]) """ arrays = [np.asarray(x) for x in arrays] dtype = arrays[0].dtype n = np.prod([x.size for x in arrays]) if out is None: out = np.zeros([n, len(arrays)], dtype=dtype) m = n // arrays[0].size out[:, 0] = np.repeat(arrays[0], m) if arrays[1:]: _cartesian(arrays[1:], out=out[0:m, 1:]) for j in range(1, arrays[0].size): out[j * m : (j + 1) * m, 1:] = out[0:m, 1:] return out def _nearest_coeffs(ratio: float, mode: str = "round_prefer_floor") -> np.ndarray: if type(ratio) == int or ratio.is_integer(): return np.array([0, 1]) if mode == "round_prefer_floor": return np.array([ratio <= 0.5, ratio > 0.5]) if mode == "round_prefer_ceil": return np.array([ratio < 0.5, ratio >= 0.5]) if mode == "floor": return np.array([1, 0]) if mode == "ceil": return np.array([0, 1]) raise ValueError(f"Unexpected value {mode!r}.") def _cubic_coeffs( ratio: float, scale: Optional[float] = None, A: float = -0.75 ) -> np.ndarray: # scale is unused coeffs = [ ((A * (ratio + 1) - 5 * A) * (ratio + 1) + 8 * A) * (ratio + 1) - 4 * A, ((A + 2) * ratio - (A + 3)) * ratio * ratio + 1, ((A + 2) * (1 - ratio) - (A + 3)) * (1 - ratio) * (1 - ratio) + 1, ((A * ((1 - ratio) + 1) - 5 * A) * ((1 - ratio) + 1) + 8 * A) * ((1 - ratio) + 1) - 4 * A, ] return np.array(coeffs) def _cubic_coeffs_antialias(ratio: float, scale: float, A: float = -0.75) -> np.ndarray: if scale > 1.0: # Antialias is applied when downsampling scale = 1.0 def compute_coeff(x: float) -> float: x = abs(x) x_2 = x * x x_3 = x * x_2 if x <= 1: return (A + 2) * x_3 - (A + 3) * x_2 + 1 if x < 2: return A * x_3 - 5 * A * x_2 + 8 * A * x - 4 * A return 0.0 i_start = int(np.floor(-2 / scale) + 1) i_end = 2 - i_start args = [scale * (i - ratio) for i in range(i_start, i_end)] coeffs = [compute_coeff(x) for x in args] return np.array(coeffs) / sum(coeffs) def _linear_coeffs(ratio: float, scale: Optional[float] = None) -> np.ndarray: # scale is unused return np.array([1 - ratio, ratio]) def _linear_coeffs_antialias(ratio: float, scale: float) -> np.ndarray: if scale > 1.0: # Antialias is applied when downsampling scale = 1.0 start = int(np.floor(-1 / scale) + 1) footprint = 2 - 2 * start args = (np.arange(start, start + footprint) - ratio) * scale coeffs = np.clip(1 - np.abs(args), 0, 1) return np.array(coeffs) / sum(coeffs) # type: ignore[no-any-return] def _get_neighbor_idxes(x: float, n: int, limit: int) -> np.ndarray: """ Return the n nearest indexes to x among `[0, limit)`, prefer the indexes smaller than x. As a result, the ratio must be in `(0, 1]`. Examples:: get_neighbor_idxes(4, 2, 10) == [3, 4] get_neighbor_idxes(4, 3, 10) == [3, 4, 5] get_neighbor_idxes(4.4, 3, 10) == [3, 4, 5] get_neighbor_idxes(4.5, 3, 10) == [3, 4, 5] get_neighbor_idxes(4.6, 3, 10) == [4, 5, 6] get_neighbor_idxes(4.4, 1, 10) == [4] get_neighbor_idxes(4.6, 1, 10) == [5] :param x: :param n: the number of the wanted indexes :param limit: the maximum value of index :return: An np.array containing n nearest indexes in ascending order """ idxes = sorted(range(limit), key=lambda idx: (abs(x - idx), idx))[:n] idxes = sorted(idxes) return np.array(idxes) def _get_neighbor(x: float, n: int, data: np.ndarray) -> Tuple[np.ndarray, np.ndarray]: """ Pad `data` in 'edge' mode, and get n nearest elements in the padded array and their indexes in the original array. :param x: center index (in the unpadded coordinate system) of the found nearest elements. :param n: the number of neighbors. :param data: the array :return: A tuple containing the indexes of neighbor elements (the index can be smaller than 0 or higher than len(data)) and the value of these elements """ pad_width = np.ceil(n / 2).astype(int) padded = np.pad(data, pad_width, mode="edge") x += pad_width idxes = _get_neighbor_idxes(x, n, len(padded)) ret = padded[idxes] return idxes - pad_width, ret def _interpolate_1d_with_x( data: np.ndarray, scale_factor: float, output_width_int: int, x: float, get_coeffs: Callable[[float, float], np.ndarray], roi: Optional[np.ndarray] = None, extrapolation_value: float = 0.0, coordinate_transformation_mode: str = "half_pixel", exclude_outside: bool = False, ) -> np.ndarray: input_width = len(data) output_width = scale_factor * input_width if coordinate_transformation_mode == "align_corners": if output_width == 1: x_ori = 0.0 else: x_ori = x * (input_width - 1) / (output_width - 1) elif coordinate_transformation_mode == "asymmetric": x_ori = x / scale_factor elif coordinate_transformation_mode == "tf_crop_and_resize": if roi is None: raise ValueError("roi cannot be None.") if output_width == 1: x_ori = (roi[1] - roi[0]) * (input_width - 1) / 2 else: x_ori = x * (roi[1] - roi[0]) * (input_width - 1) / (output_width - 1) x_ori += roi[0] * (input_width - 1) # Return extrapolation_value directly as what TF CropAndResize does if x_ori < 0 or x_ori > input_width - 1: return np.array(extrapolation_value) elif coordinate_transformation_mode == "pytorch_half_pixel": if output_width == 1: x_ori = -0.5 else: x_ori = (x + 0.5) / scale_factor - 0.5 elif coordinate_transformation_mode == "half_pixel": x_ori = (x + 0.5) / scale_factor - 0.5 elif coordinate_transformation_mode == "half_pixel_symmetric": # Maps the center of the implicit ROI to the center of the output canvas. # The difference with `half_pixel` will be only relevant # when output_width_int != output_width adjustment = output_width_int / output_width center = input_width / 2 offset = center * (1 - adjustment) x_ori = offset + (x + 0.5) / scale_factor - 0.5 else: raise ValueError( f"Invalid coordinate_transformation_mode: {coordinate_transformation_mode!r}." ) x_ori_int = np.floor(x_ori).astype(int).item() # ratio must be in (0, 1] since we prefer the pixel on the left of `x_ori` if x_ori.is_integer(): ratio = 1 else: ratio = x_ori - x_ori_int coeffs = get_coeffs(ratio, scale_factor) n = len(coeffs) idxes, points = _get_neighbor(x_ori, n, data) if exclude_outside: for i, idx in enumerate(idxes): if idx < 0 or idx >= input_width: coeffs[i] = 0 coeffs /= sum(coeffs) return np.dot(coeffs, points).item() # type: ignore[no-any-return] def _interpolate_nd_with_x( data: np.ndarray, n: int, scale_factors: List[float], output_size: List[int], x: List[float], get_coeffs: Callable[[float, float], np.ndarray], roi: Optional[np.ndarray] = None, exclude_outside: bool = False, **kwargs: Any, ) -> np.ndarray: if n == 1: return _interpolate_1d_with_x( data, scale_factors[0], output_size[0], x[0], get_coeffs, roi=roi, exclude_outside=exclude_outside, **kwargs, ) res1d = [] for i in range(data.shape[0]): r = _interpolate_nd_with_x( data[i], n - 1, scale_factors[1:], output_size[1:], x[1:], get_coeffs, roi=None if roi is None else np.concatenate([roi[1:n], roi[n + 1 :]]), exclude_outside=exclude_outside, **kwargs, ) res1d.append(r) return _interpolate_1d_with_x( res1d, # type: ignore[arg-type] # FIXME scale_factors[0], output_size[0], x[0], get_coeffs, roi=None if roi is None else [roi[0], roi[n]], # type: ignore[arg-type] # FIXME exclude_outside=exclude_outside, **kwargs, ) def _get_all_coords(data: np.ndarray) -> np.ndarray: # FIXME: Fix input type return _cartesian( [list(range(data.shape[i])) for i in range(len(data.shape))] # type: ignore[arg-type,misc] ) def _interpolate_nd( data: np.ndarray, get_coeffs: Callable[[float, float], np.ndarray], output_size: Optional[List[int]] = None, scale_factors: Optional[List[float]] = None, axes: Optional[List[int]] = None, roi: Optional[np.ndarray] = None, keep_aspect_ratio_policy: Optional[str] = "stretch", exclude_outside: bool = False, **kwargs: Any, ) -> np.ndarray: if output_size is None and scale_factors is None: raise ValueError("output_size is None and scale_factors is None.") r = len(data.shape) if axes is not None: if scale_factors is not None: new_scale_factors = [1.0] * r for i, d in enumerate(axes): new_scale_factors[d] = scale_factors[i] scale_factors = new_scale_factors if output_size is not None: new_output_size = [data.shape[i] for i in range(r)] for i, d in enumerate(axes): new_output_size[d] = output_size[i] output_size = new_output_size if roi is not None: new_roi = ([0.0] * r) + ([1.0] * r) naxes = len(axes) for i, d in enumerate(axes): new_roi[d] = roi[i] new_roi[r + d] = roi[naxes + i] roi = new_roi # type: ignore[assignment] # FIXME else: axes = list(range(r)) if output_size is not None: scale_factors = [output_size[i] / data.shape[i] for i in range(r)] if keep_aspect_ratio_policy != "stretch": if keep_aspect_ratio_policy == "not_larger": scale = np.array(scale_factors)[axes].min() elif keep_aspect_ratio_policy == "not_smaller": scale = np.array(scale_factors)[axes].max() else: raise ValueError( f"Invalid keep_aspect_ratio_policy={keep_aspect_ratio_policy!r}" ) scale_factors = [scale if i in axes else 1.0 for i in range(r)] def round_half_up(x: float) -> int: return int(x + 0.5) output_size = [ round_half_up(scale * data.shape[i]) if i in axes else data.shape[i] for i in range(r) ] else: output_size = (scale_factors * np.array(data.shape)).astype(int) # type: ignore[union-attr] if scale_factors is None: raise ValueError("scale_factors is None.") if output_size is None: raise ValueError("output_size is None.") ret = np.zeros(output_size) for x in _get_all_coords(ret): ret[tuple(x)] = _interpolate_nd_with_x( data, len(data.shape), scale_factors, output_size, x, get_coeffs, roi=roi, exclude_outside=exclude_outside, **kwargs, ) return ret class Resize(OpRun): def _run( # type: ignore self, X, roi, scales=None, sizes=None, antialias=None, axes=None, coordinate_transformation_mode=None, cubic_coeff_a=None, exclude_outside=None, extrapolation_value=None, keep_aspect_ratio_policy=None, mode=None, nearest_mode=None, ): if mode == "nearest": # type: ignore if antialias: raise RuntimeError( f"antilias={antialias!r} is not supported for mode={mode!r}." ) if nearest_mode is not None: def fct(x, scale_factor): return _nearest_coeffs(x, mode=nearest_mode) else: fct = _nearest_coeffs elif mode == "cubic": fct_ = _cubic_coeffs_antialias if antialias else _cubic_coeffs def fct(x, scale): return fct_(x, scale, A=cubic_coeff_a) elif mode == "linear": fct = _linear_coeffs_antialias if antialias else _linear_coeffs else: raise ValueError(f"Unexpected value {mode!r} for mode.") if axes is None: output = _interpolate_nd( X, fct, scale_factors=scales, output_size=sizes, roi=roi, keep_aspect_ratio_policy=keep_aspect_ratio_policy, exclude_outside=exclude_outside, coordinate_transformation_mode=coordinate_transformation_mode, # type: ignore extrapolation_value=extrapolation_value, # type: ignore ).astype(X.dtype) return (output,) # axes is not None not_axes = [a for a in range(len(X.shape)) if a not in axes] perm = tuple(not_axes + axes) permuted = np.transpose(X, perm) new_shape = (-1, *tuple(X.shape[a] for a in axes)) reshaped = permuted.reshape(new_shape) res = None for i in range(reshaped.shape[0]): output = _interpolate_nd( reshaped[i], fct, scale_factors=scales, output_size=sizes, roi=roi, keep_aspect_ratio_policy=keep_aspect_ratio_policy, exclude_outside=exclude_outside, coordinate_transformation_mode=coordinate_transformation_mode, # type: ignore extrapolation_value=extrapolation_value, # type: ignore ).astype(X.dtype) if res is None: res = np.empty((reshaped.shape[0], *output.shape), dtype=output.dtype) res[i] = output res_reshaped = res.reshape(tuple(X.shape[a] for a in not_axes) + res[0].shape) # type: ignore new_perm = list(perm) for i, a in enumerate(perm): new_perm[a] = i final = np.transpose(res_reshaped, tuple(new_perm)) return (final,)
15,638
32.274468
102
py
onnx
onnx-main/onnx/reference/ops/op_stft.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=R0913,R0914,R0915,W0613,W0221 import numpy as np from onnx.reference.op_run import OpRun from onnx.reference.ops.op_concat_from_sequence import _concat_from_sequence from onnx.reference.ops.op_dft import _cfft as _dft from onnx.reference.ops.op_slice import _slice def _concat(*args, axis=0): # type: ignore return np.concatenate(args, axis=axis) def _unsqueeze(a, axis): # type: ignore try: return np.expand_dims(a, axis=axis) except TypeError: # numpy 1.18 supports axes as a tuple if len(axis) == 1: return np.expand_dims(a, axis=tuple(axis)[0]) for x in reversed(axis): a = np.expand_dims(a, axis=x) return a def _stft(x, fft_length, hop_length, n_frames, window, onesided=False): # type: ignore """ Applies one dimensional FFT with window weights. torch defines the number of frames as: `n_frames = 1 + (len - n_fft) // hop_length`. """ last_axis = len(x.shape) - 1 # op.Sub(op.Shape(op.Shape(x)), one) axis = [-2] axis2 = [-3] window_size = window.shape[0] # building frames seq = [] for fs in range(n_frames): begin = fs * hop_length end = begin + window_size sliced_x = _slice(x, np.array([begin]), np.array([end]), axis) # type: ignore # sliced_x may be smaller new_dim = sliced_x.shape[-2:-1] missing = (window_size - new_dim[0],) new_shape = sliced_x.shape[:-2] + missing + sliced_x.shape[-1:] cst = np.zeros(new_shape, dtype=x.dtype) pad_sliced_x = _concat(sliced_x, cst, axis=-2) # same size un_sliced_x = _unsqueeze(pad_sliced_x, axis2) seq.append(un_sliced_x) # concatenation new_x = _concat_from_sequence(seq, axis=-3, new_axis=0) # calling weighted dft with weights=window shape_x = new_x.shape shape_x_short = shape_x[:-2] shape_x_short_one = tuple(1 for _ in shape_x_short) window_shape = (*shape_x_short_one, window_size, 1) weights = np.reshape(window, window_shape) weighted_new_x = new_x * weights result = _dft( weighted_new_x, fft_length, last_axis, onesided=onesided ) # normalize=False return result def _istft(x, fft_length, hop_length, window, onesided=False): # type: ignore """ Reverses of `stft`. """ zero = [0] one = [1] two = [2] axisf = [-2] n_frames = x.shape[-2] expected_signal_len = fft_length[0] + hop_length * (n_frames - 1) # building frames seqr = [] seqi = [] seqc = [] for fs in range(n_frames): begin = fs end = fs + 1 frame_x = np.squeeze( _slice(x, np.array([begin]), np.array([end]), axisf), axis=axisf[0] # type: ignore ) # ifft ift = _dft(frame_x, fft_length, axis=-1, onesided=onesided, normalize=True) n_dims = len(ift.shape) # real part n_dims_1 = n_dims - 1 sliced = _slice(ift, np.array(zero), np.array(one), [n_dims_1]) # type: ignore ytmp = np.squeeze(sliced, axis=n_dims_1) ctmp = np.full(ytmp.shape, fill_value=1, dtype=x.dtype) * window shape_begin = ytmp.shape[:-1] n_left = fs * hop_length size = ytmp.shape[-1] n_right = expected_signal_len - (n_left + size) left_shape = (*shape_begin, n_left) right_shape = (*shape_begin, n_right) right = np.zeros(right_shape, dtype=x.dtype) left = np.zeros(left_shape, dtype=x.dtype) y = _concat(left, ytmp, right, axis=-1) yc = _concat(left, ctmp, right, axis=-1) # imaginary part sliced = _slice(ift, np.array(one), np.array(two), [n_dims_1]) # type: ignore itmp = np.squeeze(sliced, axis=n_dims_1) yi = _concat(left, itmp, right, axis=-1) # append seqr.append(_unsqueeze(y, axis=-1)) seqi.append(_unsqueeze(yi, axis=-1)) seqc.append(_unsqueeze(yc, axis=-1)) # concatenation redr = _concat_from_sequence(seqr, axis=-1, new_axis=0) redi = _concat_from_sequence(seqi, axis=-1, new_axis=0) redc = _concat_from_sequence(seqc, axis=-1, new_axis=0) # unweight resr = redr.sum(axis=-1, keepdims=0) # type: ignore resi = redi.sum(axis=-1, keepdims=0) # type: ignore resc = redc.sum(axis=-1, keepdims=0) # type: ignore rr = resr / resc ri = resi / resc # Make complex rr0 = np.expand_dims(rr, axis=0) ri0 = np.expand_dims(ri, axis=0) conc = _concat(rr0, ri0, axis=0) # rotation, bring first dimension to the last position result_shape = conc.shape reshaped_result = conc.reshape((2, -1)) transposed = np.transpose(reshaped_result, (1, 0)) other_dimensions = result_shape[1:] final_shape = _concat(other_dimensions, two, axis=0) final = transposed.reshape(final_shape) return final class STFT(OpRun): def _run(self, x, frame_step, window=None, frame_length=None, onesided=None): # type: ignore if frame_length is None: if window is None: frame_length = x.shape[-2] else: frame_length = window.shape[0] hop_length = frame_step if window is None: window = np.ones((frame_length,), dtype=x.dtype) n_frames = 1 + (x.shape[-2] - frame_length) // frame_step res = _stft(x, [frame_length], hop_length, n_frames, window, onesided=onesided) return (res.astype(x.dtype),)
5,591
31.511628
97
py
onnx
onnx-main/onnx/reference/ops/op_blackman_window.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=W0221 import numpy as np from onnx.reference.ops._op_common_window import _CommonWindow class BlackmanWindow(_CommonWindow): """ Returns :math:`\\omega_n = 0.42 - 0.5 \\cos \\left( \\frac{2\\pi n}{N-1} \\right) + 0.08 \\cos \\left( \\frac{4\\pi n}{N-1} \\right)` where *N* is the window length. See `blackman_window <https://pytorch.org/docs/stable/generated/torch.blackman_window.html>`_ """ def _run(self, size, output_datatype=None, periodic=None): # type: ignore ni, N_1 = np.arange(size), size if periodic == 0: N_1 = N_1 - 1 alpha = 0.42 beta = 0.08 pi = np.pi y = np.cos((ni * (pi * 2)) / N_1) * (-0.5) y += np.cos((ni * (pi * 4)) / N_1) * beta y += alpha return self._end(size, y, output_datatype)
932
27.272727
79
py
onnx
onnx-main/onnx/reference/ops/op_grid_sample.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # pylint: disable=R0912,R0913,R0914,R0915,R1702,R1716,W0221 import numbers from typing import List import numpy as np from onnx.reference.op_run import OpRun from onnx.reference.ops.op_resize import _get_all_coords class GridSample(OpRun): # https://github.com/pytorch/pytorch/blob/v2.0.0/aten/src/ATen/native/GridSampler.h#L26 def _gs_denormalize(self, n, length: int, align_corners: bool): # type: ignore # n is the normalized coordinate (float) # x is the unormalized coordinate (float) if align_corners: # Align to corners # x_min = 0 # x_max = d-1 # Linear mapping from [x_min, x_max] to [-1, 1] # Solving linear equation n = ax + b # a = 2/(d-1) # b = -1 # n = 2/(d-1) x - 1 # n(d-1) = 2x - (d-1) # x = (n+1)(d-1) / 2 x = (n + 1) / 2.0 * (length - 1) else: # Not align to corners # x_min = -0.5 # x_max = d-0.5 # Linear mapping from [x_min, x_max] to [-1, 1] # Solving linear equation n = ax + b # a = 2/d # b = 1/d - 1 # n = 2/d x + 1/d - 1 # nd = 2x + 1 - d # x = (nd + d - 1) / 2 # x = ((n + 1) d - 1) / 2 x = ((n + 1) * length - 1) / 2.0 return x def _gs_denormalize_coordinates(self, n, dims, align_corners: bool): x = np.zeros(len(n), dtype=np.float32) for i, (v, dim) in enumerate(zip(n, dims)): x[i] = self._gs_denormalize(n=v, length=dim, align_corners=align_corners) return x def _gs_reflect(self, x, x_min, x_max): # type: ignore """ Reflect by the near border till within the borders Use float for borders to avoid potential issues with integer T """ fx = x rng = x_max - x_min if fx < x_min: dx = x_min - fx n = int(dx / rng) r = dx - n * rng if n % 2 == 0: fx = x_min + r else: fx = x_max - r elif fx > x_max: dx = fx - x_max n = int(dx / rng) r = dx - n * rng if n % 2 == 0: fx = x_max - r else: fx = x_min + r return fx def _gs_get_cubic_coeffs(self, x, coeffs): # type: ignore """ Calculate cubic convolution interpolation coefficients ROBERT G. KEYS https://ieeexplore.ieee.org/document/1163711 Use float to avoid potential issues with integer. """ cubic_alpha = -0.75 x = abs(x) coeffs[0] = ( (cubic_alpha * (x + 1) - 5 * cubic_alpha) * (x + 1) + 8 * cubic_alpha ) * (x + 1) - 4 * cubic_alpha coeffs[1] = ((cubic_alpha + 2) * x - (cubic_alpha + 3)) * x * x + 1 coeffs[2] = ((cubic_alpha + 2) * (1 - x) - (cubic_alpha + 3)) * (1 - x) * ( 1 - x ) + 1 coeffs[3] = ( (cubic_alpha * (2 - x) - 5 * cubic_alpha) * (2 - x) + 8 * cubic_alpha ) * (2 - x) - 4 * cubic_alpha def _gs_get_linear_coeffs(self, x, coeffs): x = abs(x) coeffs[0] = 1 - x coeffs[1] = x def _gs_bicubic_interpolate(self, p, x, y): # type: ignore v = np.empty((4,), dtype=p.dtype) coeffs = np.empty((4,), dtype=p.dtype) self._gs_get_cubic_coeffs(x, coeffs) for i in range(4): v[i] = coeffs @ p[i, :] self._gs_get_cubic_coeffs(y, coeffs) return coeffs @ v def _gs_cubic_interpolation_1d_with_x(self, data, x, border, padding_mode): v = np.empty((4,), dtype=data.dtype) coeffs = np.empty((4,), dtype=data.dtype) x_0 = int(np.floor(x)) x_1 = x_0 + 1 x_2 = x_0 + 2 x_minus_1 = x_0 - 1 self._gs_get_cubic_coeffs(x - x_0, coeffs) v[0] = self._pixel_at_array( array=data, i=x_minus_1, border=border, padding_mode=padding_mode ) v[1] = self._pixel_at_array( array=data, i=x_0, border=border, padding_mode=padding_mode ) v[2] = self._pixel_at_array( array=data, i=x_1, border=border, padding_mode=padding_mode ) v[3] = self._pixel_at_array( array=data, i=x_2, border=border, padding_mode=padding_mode ) return coeffs @ v def _gs_linear_interpolation_1d_with_x(self, data, x, border, padding_mode): v = np.empty((2,), dtype=data.dtype) coeffs = np.empty((2,), dtype=data.dtype) x_0 = int(np.floor(x)) x_1 = x_0 + 1 self._gs_get_linear_coeffs(x - x_0, coeffs) v[0] = self._pixel_at_array( array=data, i=x_0, border=border, padding_mode=padding_mode ) v[1] = self._pixel_at_array( array=data, i=x_1, border=border, padding_mode=padding_mode ) return coeffs @ v def _gs_linear_interpolation_nd_with_x(self, data, x, border, padding_mode): num_dims = data.ndim assert num_dims == len(x) == int(len(border) / 2) if num_dims == 1: return self._gs_linear_interpolation_1d_with_x( data=data, x=x[0], border=border, padding_mode=padding_mode ) res1d = [] for i in range(data.shape[0]): r = self._gs_linear_interpolation_nd_with_x( data=data[i], x=x[1:], border=list(border[1:num_dims]) + list(border[1 + num_dims : 2 * num_dims]), padding_mode=padding_mode, ) res1d.append(r) res1d = np.array(res1d) return self._gs_linear_interpolation_1d_with_x( data=res1d, x=x[0], border=[border[0], border[num_dims]], padding_mode=padding_mode, ) def _gs_cubic_interpolation_nd_with_x(self, data, x, border, padding_mode): num_dims = data.ndim assert num_dims == len(x) == int(len(border) / 2) if num_dims == 1: return self._gs_cubic_interpolation_1d_with_x( data=data, x=x[0], border=border, padding_mode=padding_mode ) res1d = [] for i in range(data.shape[0]): r = self._gs_cubic_interpolation_nd_with_x( data=data[i], x=x[1:], border=list(border[1:num_dims]) + list(border[1 + num_dims : 2 * num_dims]), padding_mode=padding_mode, ) res1d.append(r) res1d = np.array(res1d) return self._gs_cubic_interpolation_1d_with_x( data=res1d, x=x[0], border=[border[0], border[num_dims]], padding_mode=padding_mode, ) def _clamp(self, val, lo, hi): # type: ignore if val < lo: return lo if val > hi: return hi return val def _pixel_at_ndarray(self, ndarray, x: List, border, padding_mode): # type: ignore # boarder: [x_1_min, x_2_min, ..., x_1_max, x_2_max, ...] num_dims = ndarray.ndim assert num_dims == len(x) == int(len(border) / 2) if num_dims == 1: return self._pixel_at_array( array=ndarray, i=x[0], border=border, padding_mode=padding_mode ) i = x[0] d = ndarray.shape[0] if padding_mode == "zeros": if i >= 0 and i < d: ndarray = ndarray[i] else: # Trick i = 0 ndarray = np.zeros_like(ndarray[i]) elif padding_mode == "border": i = self._clamp(i, 0, d - 1) ndarray = ndarray[i] else: # padding_mode == "reflection" i = int(self._gs_reflect(i, border[0], border[num_dims])) ndarray = ndarray[i] return self._pixel_at_ndarray( ndarray=ndarray, x=x[1:], border=list(border[1:num_dims]) + list(border[1 + num_dims : 2 * num_dims]), padding_mode=padding_mode, ) def _pixel_at_array(self, array, i: int, border, padding_mode): # type: ignore assert array.ndim == 1 d = array.shape[0] if padding_mode == "zeros": if i >= 0 and i < d: pixel = array[i] else: pixel = 0 elif padding_mode == "border": i = self._clamp(i, 0, d - 1) pixel = array[i] else: # padding_mode == "reflection" i = int(self._gs_reflect(i, border[0], border[1])) pixel = array[i] return pixel def _prepare_border(self, dims, align_corners: bool): # boarder: [x_1_min, x_2_min, ..., x_1_max, x_2_max, ...] num_dims = len(dims) borders = np.zeros(num_dims * 2) for i in range(num_dims): # min borders[i] = -0.5 # max borders[i + num_dims] = dims[i] - 0.5 if align_corners: # min borders[i] = 0.0 # max borders[i + num_dims] = dims[i] - 1.0 return borders def _cpp_std_round(self, x): # https://en.cppreference.com/w/cpp/numeric/math/round def round_single_value(v): if v >= 0.0: return np.floor(v + 0.5) else: return np.ceil(v - 0.5) if isinstance(x, numbers.Number): return round_single_value(x) else: assert x.ndim == 1 x_rounded = np.zeros_like(x) for i in range(x.shape[0]): x_rounded[i] = round_single_value(x[i]) x_rounded = x_rounded.astype(np.int32) return x_rounded def _run(self, X, grid, mode=None, padding_mode=None, align_corners=None): # This implementation supports GridSample arbitrary dimensions. mode = mode or self.mode # type: ignore padding_mode = padding_mode or self.padding_mode # type: ignore align_corners = align_corners or self.align_corners # type: ignore x_dims = X.shape grid_dims = grid.shape N = x_dims[0] C = x_dims[1] y_dims = (N, C, *grid_dims[1:-1]) if np.prod(y_dims) == 0: return np.array([], dtype=X.dtype) Y = np.empty(y_dims, dtype=X.dtype) for n in range(N): grid_data = grid[n] for c in range(C): # Because the indices in the grid_data are always in the "reverse" dimensional order. # To interpolate for certain positions, we either have to transpose the X_data or # reverse the indices. # In this implementation, we took the latter approach. X_data = X[n, c] num_dims = len(x_dims[2:]) dims = x_dims[2:] # Prepare borders. border = self._prepare_border(dims, align_corners=align_corners) for ox in _get_all_coords(Y[n, c]): # normalized coordinates. nx = grid_data[tuple(ox)] nx = nx[::-1] # denormalized coordinates. x = self._gs_denormalize_coordinates( n=nx, dims=dims, align_corners=align_corners ) if mode == "nearest": # PyTorch round the index to nearest even. # https://github.com/pytorch/pytorch/pull/97000 x = np.rint(x) # https://github.com/pytorch/pytorch/blob/v2.0.0/aten/src/ATen/native/GridSampler.h#L142 for i, v in enumerate(x): x_min = border[i] x_max = border[i + num_dims] if v < x_min or v > x_max: if padding_mode == "border": x[i] = self._clamp(v, 0, dims[i] - 1) elif padding_mode == "reflection": x[i] = self._gs_reflect(v, x_min, x_max) if mode == "nearest": x = x.astype(np.int32) Y[n][c][tuple(ox)] = self._pixel_at_ndarray( ndarray=X_data, x=x, border=border, padding_mode=padding_mode, ) elif mode == "linear": Y[n][c][tuple(ox)] = self._gs_linear_interpolation_nd_with_x( data=X_data, x=x, border=border, padding_mode=padding_mode ) elif mode == "cubic": Y[n][c][tuple(ox)] = self._gs_cubic_interpolation_nd_with_x( data=X_data, x=x, border=border, padding_mode=padding_mode ) else: raise RuntimeError( "GridSample interpolation only supports nearest, linear, and cubic modes." ) return (Y.astype(X.dtype),)
13,472
35.217742
108
py
onnx
onnx-main/onnx/backend/test/case/node/gridsample.py
# Copyright (c) ONNX Project Contributors # # SPDX-License-Identifier: Apache-2.0 import numpy as np import onnx from onnx.backend.test.case.base import Base from onnx.backend.test.case.node import expect class GridSample(Base): @staticmethod def export_gridsample() -> None: node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", padding_mode="zeros", align_corners=0, ) # X shape, [N, C, H, W] - [1, 1, 4, 4] X = np.array( [ [ [ [0.0, 1.0, 2.0, 3.0], [4.0, 5.0, 6.0, 7.0], [8.0, 9.0, 10.0, 11.0], [12.0, 13.0, 14.0, 15.0], ] ] ], dtype=np.float32, ) # Grid shape, [N, H_out, W_out, 2] - [1, 6, 6, 2] Grid = np.array( [ [ [ [-1.0000, -1.0000], [-0.6000, -1.0000], [-0.2000, -1.0000], [0.2000, -1.0000], [0.6000, -1.0000], [1.0000, -1.0000], ], [ [-1.0000, -0.6000], [-0.6000, -0.6000], [-0.2000, -0.6000], [0.2000, -0.6000], [0.6000, -0.6000], [1.0000, -0.6000], ], [ [-1.0000, -0.2000], [-0.6000, -0.2000], [-0.2000, -0.2000], [0.2000, -0.2000], [0.6000, -0.2000], [1.0000, -0.2000], ], [ [-1.0000, 0.2000], [-0.6000, 0.2000], [-0.2000, 0.2000], [0.2000, 0.2000], [0.6000, 0.2000], [1.0000, 0.2000], ], [ [-1.0000, 0.6000], [-0.6000, 0.6000], [-0.2000, 0.6000], [0.2000, 0.6000], [0.6000, 0.6000], [1.0000, 0.6000], ], [ [-1.0000, 1.0000], [-0.6000, 1.0000], [-0.2000, 1.0000], [0.2000, 1.0000], [0.6000, 1.0000], [1.0000, 1.0000], ], ] ], dtype=np.float32, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 6, 6] Y = np.array( [ [ [ [0.0000, 0.1500, 0.5500, 0.9500, 1.3500, 0.7500], [0.6000, 1.5000, 2.3000, 3.1000, 3.9000, 2.1000], [2.2000, 4.7000, 5.5000, 6.3000, 7.1000, 3.7000], [3.8000, 7.9000, 8.7000, 9.5000, 10.3000, 5.3000], [5.4000, 11.1000, 11.9000, 12.7000, 13.5000, 6.9000], [3.0000, 6.1500, 6.5500, 6.9500, 7.3500, 3.7500], ] ] ], dtype=np.float32, ) expect(node, inputs=[X, Grid], outputs=[Y], name="test_gridsample") @staticmethod def export_gridsample_paddingmode() -> None: # X shape, [N, C, H, W] - [1, 1, 3, 2] X = np.array( [[[[0.0, 1.0], [2.0, 3.0], [4.0, 5.0]]]], dtype=np.float32, ) # Grid shape, [N, H_out, W_out, 2] - [1, 2, 4, 2] Grid = np.array( [ [ [ [-10.0000, -10.0000], [-5.0000, -5.0000], [-0.2000, -0.2000], [10.0000, 10.0000], ], [ [10.0000, 10.0000], [-0.2000, -0.2000], [5.0000, 5.0000], [10.0000, 10.0000], ], ] ], dtype=np.float32, ) # setting padding_mode = 'zeros' node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], padding_mode="zeros", ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_zeros = np.array( [[[[0.0000, 0.0000, 1.7000, 0.0000], [0.0000, 1.7000, 0.0000, 0.0000]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_zeros], name="test_gridsample_zeros_padding", ) # setting padding_mode = 'border' node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], padding_mode="border", ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_border = np.array( [[[[0.0000, 0.0000, 1.7000, 5.0000], [5.0000, 1.7000, 5.0000, 5.0000]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_border], name="test_gridsample_border_padding", ) # setting padding_mode = 'reflection' node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], padding_mode="reflection", ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_reflection = np.array( [[[[2.5000, 0.0000, 1.7000, 2.5000], [2.5000, 1.7000, 5.0000, 2.5000]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_reflection], name="test_gridsample_reflection_padding", ) @staticmethod def export_gridsample_mode_aligncorners() -> None: # X shape, [N, C, H, W] - [1, 1, 3, 2] X = np.array( [[[[0.0, 1.0], [2.0, 3.0], [4.0, 5.0]]]], dtype=np.float32, ) # Grid shape, [N, H_out, W_out, 2] - [1, 2, 4, 2] Grid = np.array( [ [ [ [-1.0000, -1.0000], [-0.5000, -0.5000], [-0.2000, -0.2000], [0.0000, 0.0000], ], [ [0.0000, 0.0000], [-0.2000, -0.2000], [0.5000, 0.5000], [1.0000, 1.0000], ], ] ], dtype=np.float32, ) # setting mode = 'bilinear', default align_corners = 0 node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bilinear = np.array( [[[[0.0000, 0.5000, 1.7000, 2.5000], [2.5000, 1.7000, 4.5000, 1.2500]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bilinear], name="test_gridsample_bilinear", ) # setting mode = 'bilinear', align_corners = 1 node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", align_corners=1, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_align_corners = np.array( [[[[0.0000, 1.2500, 2.0000, 2.5000], [2.5000, 2.0000, 3.7500, 5.0000]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_align_corners], name="test_gridsample_aligncorners_true", ) # setting mode = 'nearest' node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="nearest", ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_nearest = np.array( [[[[0.0, 0.0, 2.0, 2.0], [2.0, 2.0, 5.0, 0.0]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_nearest], name="test_gridsample_nearest" ) # setting mode = 'bicubic' node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="cubic", ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bicubic = np.array( [[[[-0.1406, 0.3828, 1.7556, 2.9688], [2.9688, 1.7556, 5.1445, 1.3906]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bicubic], name="test_gridsample_bicubic" ) # ============================================================================ # Additional tests # The reference output tensors were generated using PyTorch 2.0. Grid = np.array( [ [ [[-1.0, -0.8], [-0.6, -0.5], [-0.1, -0.2], [0.7, 0.0]], [[0.0, 0.4], [0.2, -0.2], [-0.3, 0.5], [-1.0, 1.0]], ] ], dtype=np.float32, ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="nearest", align_corners=0, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_nearest = np.array( [[[[0.0, 0.0, 2.0, 3.0], [4.0, 3.0, 4.0, 4.0]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_nearest], name="test_gridsample_nearest_align_corners_0_additional_1", ) # setting mode = 'nearest' node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="nearest", align_corners=1, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_nearest = np.array( [[[[0.0, 0.0, 2.0, 3.0], [2.0, 3.0, 4.0, 4.0]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_nearest], name="test_gridsample_nearest_align_corners_1_additional_1", ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", align_corners=0, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bilinear = np.array( [[[[0.0000, 0.4500, 1.8000, 2.4000], [3.7000, 2.1000, 3.7000, 1.0000]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bilinear], name="test_gridsample_bilinear_align_corners_0_additional_1", ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", align_corners=1, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bilinear = np.array( [[[[0.4000, 1.2000, 2.0500, 2.8500], [3.3000, 2.2000, 3.3500, 4.0000]]]], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bilinear], name="test_gridsample_bilinear_align_corners_1_additional_1", ) # These two new bicubic tests produces slightly higher error ~5e-5 node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="cubic", align_corners=0, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bicubic = np.array( [ [ [ [-0.173250, 0.284265, 1.923106, 2.568000], [5.170375, 2.284414, 4.744844, 1.046875], ] ] ], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bicubic], name="test_gridsample_bicubic_align_corners_0_additional_1", ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="cubic", align_corners=1, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bicubic = np.array( [ [ [ [0.304001, 1.128750, 2.266270, 3.144844], [4.531500, 2.455360, 4.599819, 4.000000], ] ] ], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bicubic], name="test_gridsample_bicubic_align_corners_1_additional_1", ) @staticmethod def export_volumeetric_gridsample_mode_aligncorners() -> None: X = np.array( [ [ [ [[1.0, 2.0], [3.0, 4.0]], [[5.0, 6.0], [7.0, 8.0]], [[9.0, 10.0], [11.0, 12.0]], ] ] ], dtype=np.float32, ) Grid = np.array( [ [ [ [[-1.0, -1.0, -1.0], [-1.0, -0.5, 0.3]], [[-0.5, -0.5, -0.5], [1.0, -0.6, -1.0]], [[-0.2, -0.2, -0.2], [0.4, 0.2, 0.6]], [[0.0, 0.0, 0.0], [-1.0, 0.0, 0.0]], ], [ [[0.0, 0.0, 0.0], [-1.0, 1.0, 0.0]], [[-0.2, -0.2, -0.2], [1.0, 0.4, -0.2]], [[0.5, 0.5, 0.5], [-1.0, -0.8, 0.8]], [[1.0, 1.0, 1.0], [0.4, 0.6, -0.3]], ], ] ], dtype=np.float32, ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="nearest", align_corners=0, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_nearest = np.array( [ [ [ [[1.0, 5.0], [1.0, 0.0], [5.0, 12.0], [5.0, 5.0]], [[5.0, 0.0], [5.0, 0.0], [12.0, 9.0], [0.0, 8.0]], ] ] ], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_nearest], name="test_gridsample_volumetric_nearest_align_corners_0", ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="nearest", align_corners=1, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_nearest = np.array( [ [ [ [[1.0, 5.0], [1.0, 2.0], [5.0, 12.0], [5.0, 5.0]], [[5.0, 7.0], [5.0, 8.0], [12.0, 9.0], [12.0, 8.0]], ] ] ], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_nearest], name="test_gridsample_volumetric_nearest_align_corners_1", ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", align_corners=0, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bilinear = np.array( [ [ [ [ [0.1250, 3.4000], [2.0000, 0.4500], [4.7000, 10.9000], [6.5000, 3.0000], ], [ [6.5000, 1.7500], [4.7000, 3.3000], [11.0000, 2.5200], [1.5000, 5.4900], ], ] ] ], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bilinear], name="test_gridsample_volumetric_bilinear_align_corners_0", ) node = onnx.helper.make_node( "GridSample", inputs=["X", "Grid"], outputs=["Y"], mode="linear", align_corners=1, ) # Y shape, [N, C, H_out, W_out] - [1, 1, 2, 4] Y_bilinear = np.array( [ [ [ [ [1.0000, 6.7000], [3.7500, 2.4000], [5.4000, 9.3000], [6.5000, 6.0000], ], [ [6.5000, 7.0000], [5.4000, 6.6000], [9.2500, 8.4000], [12.0000, 6.1000], ], ] ] ], dtype=np.float32, ) expect( node, inputs=[X, Grid], outputs=[Y_bilinear], name="test_gridsample_volumetric_bilinear_align_corners_1", ) """ For someone who want to test by script. Comment it cause github ONNX CI do not have the torch python package. @staticmethod def export_gridsample_torch(): # type: () -> None node = onnx.helper.make_node( 'GridSample', inputs=['X', 'Grid'], outputs=['Y'], mode='bilinear', padding_mode='zeros', align_corners=0, ) # X shape, [N, C, H, W] - [1, 1, 4, 4] # Grid shape, [N, H_out, W_out, 2] - [1, 6, 6, 2] # Y shape, [N, C, H_out, W_out] - [1, 1, 6, 6] import torch X = torch.arange(3 * 3).view(1, 1, 3, 3).float() d = torch.linspace(-1, 1, 6) meshx, meshy = torch.meshgrid((d, d)) grid = torch.stack((meshy, meshx), 2) Grid = grid.unsqueeze(0) Y = torch.nn.functional.grid_sample(X, Grid, mode='bilinear', padding_mode='zeros', align_corners=False) expect(node, inputs=[X.numpy(), Grid.numpy()], outputs=[Y.numpy()], name='test_gridsample_torch') """
19,682
29.658879
87
py
onnx
onnx-main/onnx/backend/test/case/node/resize.py
# Copyright (c) ONNX Project Contributors # # SPDX-License-Identifier: Apache-2.0 import numpy as np import onnx from onnx.backend.test.case.base import Base from onnx.backend.test.case.node import expect from onnx.reference.ops.op_resize import _cubic_coeffs as cubic_coeffs from onnx.reference.ops.op_resize import ( _cubic_coeffs_antialias as cubic_coeffs_antialias, ) from onnx.reference.ops.op_resize import _interpolate_nd as interpolate_nd from onnx.reference.ops.op_resize import _linear_coeffs as linear_coeffs from onnx.reference.ops.op_resize import ( _linear_coeffs_antialias as linear_coeffs_antialias, ) from onnx.reference.ops.op_resize import _nearest_coeffs as nearest_coeffs class Resize(Base): @staticmethod def export_resize_upsample_scales_nearest() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="nearest", ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 3.0], dtype=np.float32) # [[[[1. 1. 1. 2. 2. 2.] # [1. 1. 1. 2. 2. 2.] # [3. 3. 3. 4. 4. 4.] # [3. 3. 3. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_nearest", ) @staticmethod def export_resize_downsample_scales_nearest() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="nearest", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32) # [[[[1. 3.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_nearest", ) @staticmethod def export_resize_upsample_sizes_nearest() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 7, 8], dtype=np.int64) # [[[[1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest", ) @staticmethod def export_resize_downsample_sizes_nearest() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 1, 3], dtype=np.int64) # [[[[1. 2. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_nearest", ) @staticmethod def export_resize_upsample_scales_linear() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32) # [[[[1. 1.25 1.75 2. ] # [1.5 1.75 2.25 2.5 ] # [2.5 2.75 3.25 3.5 ] # [3. 3.25 3.75 4. ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_linear", ) @staticmethod def export_resize_upsample_scales_linear_align_corners() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", coordinate_transformation_mode="align_corners", ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32) # [[[[1. 1.33333333 1.66666667 2. ] # [1.66666667 2. 2.33333333 2.66666667] # [2.33333333 2.66666667 3. 3.33333333] # [3. 3.33333333 3.66666667 4. ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), scale_factors=scales, coordinate_transformation_mode="align_corners", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_linear_align_corners", ) @staticmethod def export_resize_downsample_scales_linear() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32) # [[[[2.6666665 4.3333331]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_linear", ) @staticmethod def export_resize_downsample_scales_linear_align_corners() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", coordinate_transformation_mode="align_corners", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32) # [[[[1. 3.142857]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), scale_factors=scales, coordinate_transformation_mode="align_corners", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_linear_align_corners", ) @staticmethod def export_resize_upsample_scales_cubic() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32) # [[[[ 0.47265625 0.76953125 1.24609375 1.875 2.28125 # 2.91015625 3.38671875 3.68359375] # [ 1.66015625 1.95703125 2.43359375 3.0625 3.46875 # 4.09765625 4.57421875 4.87109375] # [ 3.56640625 3.86328125 4.33984375 4.96875 5.375 # 6.00390625 6.48046875 6.77734375] # [ 6.08203125 6.37890625 6.85546875 7.484375 7.890625 # 8.51953125 8.99609375 9.29296875] # [ 7.70703125 8.00390625 8.48046875 9.109375 9.515625 # 10.14453125 10.62109375 10.91796875] # [10.22265625 10.51953125 10.99609375 11.625 12.03125 # 12.66015625 13.13671875 13.43359375] # [12.12890625 12.42578125 12.90234375 13.53125 13.9375 # 14.56640625 15.04296875 15.33984375] # [13.31640625 13.61328125 14.08984375 14.71875 15.125 # 15.75390625 16.23046875 16.52734375]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x), scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_cubic", ) @staticmethod def export_resize_upsample_scales_cubic_align_corners() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", coordinate_transformation_mode="align_corners", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32) # [[[[ 1. 1.34110787 1.80029155 2.32944606 2.67055394 # 3.19970845 3.65889213 4. ] # [ 2.36443149 2.70553936 3.16472303 3.69387755 4.03498542 # 4.56413994 5.02332362 5.36443149] # [ 4.20116618 4.54227405 5.00145773 5.53061224 5.87172012 # 6.40087464 6.86005831 7.20116618] # [ 6.31778426 6.65889213 7.1180758 7.64723032 7.98833819 # 8.51749271 8.97667638 9.31778426] # [ 7.68221574 8.02332362 8.48250729 9.01166181 9.35276968 # 9.8819242 10.34110787 10.68221574] # [ 9.79883382 10.13994169 10.59912536 11.12827988 11.46938776 # 11.99854227 12.45772595 12.79883382] # [11.63556851 11.97667638 12.43586006 12.96501458 13.30612245 # 13.83527697 14.29446064 14.63556851] # [13. 13.34110787 13.80029155 14.32944606 14.67055394 # 15.19970845 15.65889213 16. ]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x), scale_factors=scales, coordinate_transformation_mode="align_corners", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_cubic_align_corners", ) @staticmethod def export_resize_downsample_scales_cubic() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.8, 0.8], dtype=np.float32) # [[[[ 1.47119141 2.78125 4.08251953] # [ 6.71142578 8.02148438 9.32275391] # [11.91650391 13.2265625 14.52783203]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x), scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_cubic", ) @staticmethod def export_resize_downsample_scales_cubic_align_corners() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", coordinate_transformation_mode="align_corners", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.8, 0.8], dtype=np.float32) # [[[[ 1. 2.39519159 3.79038317] # [ 6.58076634 7.97595793 9.37114951] # [12.16153268 13.55672427 14.95191585]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x), scale_factors=scales, coordinate_transformation_mode="align_corners", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_cubic_align_corners", ) @staticmethod def export_resize_upsample_sizes_cubic() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="cubic", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 9, 10], dtype=np.int64) # [[[[ 0.45507922 0.64057922 0.97157922 1.42257922 1.90732922 # 2.22332922 2.70807922 3.15907922 3.49007922 3.67557922] # [ 1.39437963 1.57987963 1.91087963 2.36187963 2.84662963 # 3.16262963 3.64737963 4.09837963 4.42937963 4.61487963] # [ 2.95130693 3.13680693 3.46780693 3.91880693 4.40355693 # 4.71955693 5.20430693 5.65530693 5.98630693 6.17180693] # [ 5.20525069 5.39075069 5.72175069 6.17275069 6.65750069 # 6.97350069 7.45825069 7.90925069 8.24025069 8.42575069] # [ 6.88975 7.07525 7.40625 7.85725 8.342 # 8.658 9.14275 9.59375 9.92475 10.11025 ] # [ 8.57424931 8.75974931 9.09074931 9.54174931 10.02649931 # 10.34249931 10.82724931 11.27824931 11.60924931 11.79474931] # [10.82819307 11.01369307 11.34469307 11.79569307 12.28044307 # 12.59644307 13.08119307 13.53219307 13.86319307 14.04869307] # [12.38512037 12.57062037 12.90162037 13.35262037 13.83737037 # 14.15337037 14.63812037 15.08912037 15.42012037 15.60562037] # [13.32442078 13.50992078 13.84092078 14.29192078 14.77667078 # 15.09267078 15.57742078 16.02842078 16.35942078 16.54492078]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x), output_size=sizes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_cubic", ) @staticmethod def export_resize_downsample_sizes_cubic() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="cubic", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 3, 3], dtype=np.int64) # [[[[ 1.63078704 3.00462963 4.37847222] # [ 7.12615741 8.5 9.87384259] # [12.62152778 13.99537037 15.36921296]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x), output_size=sizes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_cubic", ) # TensorFlow v1 bicubic with half_pixel_centers=True @staticmethod def export_resize_upsample_scales_cubic_A_n0p5_exclude_outside() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", cubic_coeff_a=-0.5, exclude_outside=True, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32) # [[[[ 0.55882353 0.81494204 1.35698249 1.89705882 2.39705882 # 2.93713516 3.47917561 3.73529412] # [ 1.58329755 1.83941606 2.38145651 2.92153285 3.42153285 # 3.96160918 4.50364964 4.75976814] # [ 3.75145936 4.00757787 4.54961832 5.08969466 5.58969466 # 6.12977099 6.67181144 6.92792995] # [ 5.91176471 6.16788321 6.70992366 7.25 7.75 # 8.29007634 8.83211679 9.08823529] # [ 7.91176471 8.16788321 8.70992366 9.25 9.75 # 10.29007634 10.83211679 11.08823529] # [10.07207005 10.32818856 10.87022901 11.41030534 11.91030534 # 12.45038168 12.99242213 13.24854064] # [12.24023186 12.49635036 13.03839082 13.57846715 14.07846715 # 14.61854349 15.16058394 15.41670245] # [13.26470588 13.52082439 14.06286484 14.60294118 15.10294118 # 15.64301751 16.18505796 16.44117647]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x, A=-0.5), scale_factors=scales, exclude_outside=True, ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_cubic_A_n0p5_exclude_outside", ) @staticmethod def export_resize_downsample_scales_cubic_A_n0p5_exclude_outside() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", cubic_coeff_a=-0.5, exclude_outside=True, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.8, 0.8], dtype=np.float32) # [[[[ 1.36812675 2.6695014 4.0133367 ] # [ 6.57362535 7.875 9.2188353 ] # [11.94896657 13.25034122 14.59417652]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x, A=-0.5), scale_factors=scales, exclude_outside=True, ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_cubic_A_n0p5_exclude_outside", ) # TensorFlow v1 bicubic with half_pixel_centers=False @staticmethod def export_resize_upsample_scales_cubic_asymmetric() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", coordinate_transformation_mode="asymmetric", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 2.0, 2.0], dtype=np.float32) # [[[[ 1. 1.40625 2. 2.5 3. 3.59375 4. # 4.09375] # [ 2.625 3.03125 3.625 4.125 4.625 5.21875 5.625 # 5.71875] # [ 5. 5.40625 6. 6.5 7. 7.59375 8. # 8.09375] # [ 7. 7.40625 8. 8.5 9. 9.59375 10. # 10.09375] # [ 9. 9.40625 10. 10.5 11. 11.59375 12. # 12.09375] # [11.375 11.78125 12.375 12.875 13.375 13.96875 14.375 # 14.46875] # [13. 13.40625 14. 14.5 15. 15.59375 16. # 16.09375] # [13.375 13.78125 14.375 14.875 15.375 15.96875 16.375 # 16.46875]]]] output = interpolate_nd( data, lambda x, _: cubic_coeffs(x, A=-0.75), scale_factors=scales, coordinate_transformation_mode="asymmetric", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_cubic_asymmetric", ) @staticmethod def export_resize_tf_crop_and_resize() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "roi", "", "sizes"], outputs=["Y"], mode="linear", coordinate_transformation_mode="tf_crop_and_resize", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) # Note: for some rois, the result may be different with that of TF for inaccurate floating point roi = np.array([0, 0, 0.4, 0.6, 1, 1, 0.6, 0.8], dtype=np.float32) sizes = np.array([1, 1, 3, 3], dtype=np.int64) # [[[[ 7.6000004 7.9 8.2 ] # [ 8.8 9.1 9.400001 ] # [10. 10.3 10.6 ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), output_size=sizes, roi=roi, coordinate_transformation_mode="tf_crop_and_resize", ).astype(np.float32) expect( node, inputs=[data, roi, sizes], outputs=[output], name="test_resize_tf_crop_and_resize", ) @staticmethod def export_resize_tf_crop_and_resize_extrapolation_value() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "roi", "", "sizes"], outputs=["Y"], mode="linear", coordinate_transformation_mode="tf_crop_and_resize", extrapolation_value=10.0, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) # Note: for some rois, the result may be different with that of TF for inaccurate floating point roi = np.array([0, 0, 0.4, 0.6, 1, 1, 1.2, 1.7], dtype=np.float32) sizes = np.array([1, 1, 3, 3], dtype=np.int64) # [[[[ 7.6000004 10. 10. ] # [12.400001 10. 10. ] # [10. 10. 10. ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), output_size=sizes, roi=roi, coordinate_transformation_mode="tf_crop_and_resize", extrapolation_value=10.0, ).astype(np.float32) expect( node, inputs=[data, roi, sizes], outputs=[output], name="test_resize_tf_crop_and_resize", ) @staticmethod def export_resize_downsample_sizes_linear_pytorch_half_pixel() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="linear", coordinate_transformation_mode="pytorch_half_pixel", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 3, 1], dtype=np.int64) # [[[[ 1.6666666] # [ 7. ] # [12.333333 ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), output_size=sizes, coordinate_transformation_mode="pytorch_half_pixel", ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_linear_pytorch_half_pixel", ) @staticmethod def export_resize_upsample_sizes_nearest_floor_align_corners() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", coordinate_transformation_mode="align_corners", nearest_mode="floor", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 8, 8], dtype=np.int64) # [[[[ 1. 1. 1. 2. 2. 3. 3. 4.] # [ 1. 1. 1. 2. 2. 3. 3. 4.] # [ 1. 1. 1. 2. 2. 3. 3. 4.] # [ 5. 5. 5. 6. 6. 7. 7. 8.] # [ 5. 5. 5. 6. 6. 7. 7. 8.] # [ 9. 9. 9. 10. 10. 11. 11. 12.] # [ 9. 9. 9. 10. 10. 11. 11. 12.] # [13. 13. 13. 14. 14. 15. 15. 16.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x, mode="floor"), output_size=sizes, coordinate_transformation_mode="align_corners", ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_floor_align_corners", ) @staticmethod def export_resize_upsample_sizes_nearest_round_prefer_ceil_asymmetric() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", coordinate_transformation_mode="asymmetric", nearest_mode="round_prefer_ceil", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 8, 8], dtype=np.int64) # [[[[ 1. 2. 2. 3. 3. 4. 4. 4.] # [ 5. 6. 6. 7. 7. 8. 8. 8.] # [ 5. 6. 6. 7. 7. 8. 8. 8.] # [ 9. 10. 10. 11. 11. 12. 12. 12.] # [ 9. 10. 10. 11. 11. 12. 12. 12.] # [13. 14. 14. 15. 15. 16. 16. 16.] # [13. 14. 14. 15. 15. 16. 16. 16.] # [13. 14. 14. 15. 15. 16. 16. 16.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x, mode="round_prefer_ceil"), output_size=sizes, coordinate_transformation_mode="asymmetric", ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_round_prefer_ceil_asymmetric", ) @staticmethod def export_resize_upsample_sizes_nearest_ceil_half_pixel() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", coordinate_transformation_mode="half_pixel", nearest_mode="ceil", ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 8, 8], dtype=np.int64) # [[[[ 1. 2. 2. 3. 3. 4. 4. 4.] # [ 5. 6. 6. 7. 7. 8. 8. 8.] # [ 5. 6. 6. 7. 7. 8. 8. 8.] # [ 9. 10. 10. 11. 11. 12. 12. 12.] # [ 9. 10. 10. 11. 11. 12. 12. 12.] # [13. 14. 14. 15. 15. 16. 16. 16.] # [13. 14. 14. 15. 15. 16. 16. 16.] # [13. 14. 14. 15. 15. 16. 16. 16.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x, mode="ceil"), output_size=sizes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_ceil_half_pixel", ) @staticmethod def export_resize_downsample_scales_linear_antialias() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", antialias=1, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32) # [[[[ 2.875 4.5 ] # [ 9.375 11. ]]]] output = interpolate_nd( data, linear_coeffs_antialias, scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_linear_antialias", ) @staticmethod def export_resize_downsample_sizes_linear_antialias() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="linear", antialias=1, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 3, 3], dtype=np.int64) # [[[[ 2.3636363 3.590909 4.818182 ] # [ 7.2727275 8.5 9.727273 ] # [12.181818 13.409091 14.636364 ]]]] output = interpolate_nd( data, linear_coeffs_antialias, output_size=sizes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_linear_antialias", ) @staticmethod def export_resize_downsample_scales_cubic_antialias() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="cubic", antialias=1, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) scales = np.array([1.0, 1.0, 0.6, 0.6], dtype=np.float32) # [[[[ 2.5180721 4.2858863] # [ 9.589329 11.357142 ]]]] output = interpolate_nd( data, cubic_coeffs_antialias, scale_factors=scales ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_cubic_antialias", ) @staticmethod def export_resize_downsample_sizes_cubic_antialias() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="cubic", antialias=1, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) sizes = np.array([1, 1, 3, 3], dtype=np.int64) # [[[[ 1.7750092 3.1200073 4.4650054] # [ 7.1550016 8.5 9.844998 ] # [12.534994 13.8799925 15.224991 ]]]] output = interpolate_nd(data, cubic_coeffs_antialias, output_size=sizes).astype( np.float32 ) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_cubic_antialias", ) @staticmethod def export_resize_upsample_scales_nearest_axes_2_3() -> None: axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="nearest", axes=axes, ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) scales = np.array([2.0, 3.0], dtype=np.float32) # [[[[1. 1. 1. 2. 2. 2.] # [1. 1. 1. 2. 2. 2.] # [3. 3. 3. 4. 4. 4.] # [3. 3. 3. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), scale_factors=scales, axes=axes ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_nearest_axes_2_3", ) @staticmethod def export_resize_upsample_scales_nearest_axes_3_2() -> None: axes = [3, 2] node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="nearest", axes=axes, ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) scales = np.array([3.0, 2.0], dtype=np.float32) # [[[[1. 1. 1. 2. 2. 2.] # [1. 1. 1. 2. 2. 2.] # [3. 3. 3. 4. 4. 4.] # [3. 3. 3. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), scale_factors=scales, axes=axes ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_nearest_axes_3_2", ) @staticmethod def export_resize_upsample_sizes_nearest_axes_2_3() -> None: axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", axes=axes, ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) sizes = np.array([7, 8], dtype=np.int64) # [[[[1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes, axes=axes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_axes_2_3", ) @staticmethod def export_resize_upsample_sizes_nearest_axes_3_2() -> None: axes = [3, 2] node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", axes=axes, ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) sizes = np.array([8, 7], dtype=np.int64) # [[[[1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes, axes=axes ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_axes_3_2", ) @staticmethod def export_resize_tf_crop_and_resize_axes_2_3() -> None: axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "roi", "", "sizes"], outputs=["Y"], mode="linear", coordinate_transformation_mode="tf_crop_and_resize", axes=axes, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) # Note: for some rois, the result may be different with that of TF for inaccurate floating point roi = np.array([0.4, 0.6, 0.6, 0.8], dtype=np.float32) sizes = np.array([3, 3], dtype=np.int64) # [[[[ 7.6000004 7.9 8.2 ] # [ 8.8 9.1 9.400001 ] # [10. 10.3 10.6 ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), output_size=sizes, roi=roi, axes=axes, coordinate_transformation_mode="tf_crop_and_resize", ).astype(np.float32) expect( node, inputs=[data, roi, sizes], outputs=[output], name="test_resize_tf_crop_and_resize_axes_2_3", ) @staticmethod def export_resize_tf_crop_and_resize_axes_3_2() -> None: axes = [3, 2] node = onnx.helper.make_node( "Resize", inputs=["X", "roi", "", "sizes"], outputs=["Y"], mode="linear", coordinate_transformation_mode="tf_crop_and_resize", axes=axes, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16], ] ] ], dtype=np.float32, ) # Note: for some rois, the result may be different with that of TF for inaccurate floating point roi = np.array([0.6, 0.4, 0.8, 0.6], dtype=np.float32) sizes = np.array([3, 3], dtype=np.int64) # [[[[ 7.6000004 7.9 8.2 ] # [ 8.8 9.1 9.400001 ] # [10. 10.3 10.6 ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), output_size=sizes, roi=roi, axes=axes, coordinate_transformation_mode="tf_crop_and_resize", ).astype(np.float32) expect( node, inputs=[data, roi, sizes], outputs=[output], name="test_resize_tf_crop_and_resize_axes_3_2", ) @staticmethod def export_resize_upsample_sizes_nearest_not_larger() -> None: keep_aspect_ratio_policy = "not_larger" axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) sizes = np.array([7, 8], dtype=np.int64) # Results in 7x7 # [[[[1. 1. 1. 1. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2.] # [3. 3. 3. 3. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes, axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_not_larger", ) @staticmethod def export_resize_upsample_sizes_nearest_not_smaller() -> None: keep_aspect_ratio_policy = "not_smaller" axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ) data = np.array( [ [ [ [1, 2], [3, 4], ] ] ], dtype=np.float32, ) sizes = np.array([7, 8], dtype=np.int64) # Results in 8x8 # [[[[1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [1. 1. 1. 1. 2. 2. 2. 2.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.] # [3. 3. 3. 3. 4. 4. 4. 4.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes, axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_upsample_sizes_nearest_not_larger", ) @staticmethod def export_resize_downsample_sizes_nearest_not_larger() -> None: keep_aspect_ratio_policy = "not_larger" axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], ] ] ], dtype=np.float32, ) sizes = np.array([1, 3], dtype=np.int64) # Results in 1x2 # [[[[1. 3.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes, axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_nearest_not_larger", ) @staticmethod def export_resize_downsample_sizes_nearest_not_smaller() -> None: keep_aspect_ratio_policy = "not_smaller" axes = [2, 3] node = onnx.helper.make_node( "Resize", inputs=["X", "", "", "sizes"], outputs=["Y"], mode="nearest", axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ) data = np.array( [ [ [ [1, 2, 3, 4], [5, 6, 7, 8], ] ] ], dtype=np.float32, ) sizes = np.array([1, 3], dtype=np.int64) # Results in 2x3 # [[[[1. 2. 4.] # [5. 6. 8.]]]] output = interpolate_nd( data, lambda x, _: nearest_coeffs(x), output_size=sizes, axes=axes, keep_aspect_ratio_policy=keep_aspect_ratio_policy, ).astype(np.float32) expect( node, inputs=[data, sizes], outputs=[output], name="test_resize_downsample_sizes_nearest_not_smaller", ) @staticmethod def export_resize_downsample_scales_linear_half_pixel_symmetric() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", coordinate_transformation_mode="half_pixel_symmetric", ) data = np.array([[[[1, 2, 3, 4]]]], dtype=np.float32) scales = np.array([1.0, 1.0, 1.0, 0.6], dtype=np.float32) # [[[[1.6666667, 3.3333333]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), scale_factors=scales, coordinate_transformation_mode="half_pixel_symmetric", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_downsample_scales_linear_half_pixel_symmetric", ) @staticmethod def export_resize_upsample_scales_linear_half_pixel_symmetric() -> None: node = onnx.helper.make_node( "Resize", inputs=["X", "", "scales"], outputs=["Y"], mode="linear", coordinate_transformation_mode="half_pixel_symmetric", ) data = np.array([[[[1, 2], [3, 4]]]], dtype=np.float32) scales = np.array([1.0, 1.0, 2.3, 2.94], dtype=np.float32) # [[[[1. , 1.15986395, 1.5 , 1.84013605, 2. ], # [1.56521738, 1.72508133, 2.06521738, 2.40535343, 2.56521738], # [2.43478262, 2.59464657, 2.93478262, 3.27491867, 3.43478262], # [3. , 3.15986395, 3.5 , 3.84013605, 4. ]]]] output = interpolate_nd( data, lambda x, _: linear_coeffs(x), scale_factors=scales, coordinate_transformation_mode="half_pixel_symmetric", ).astype(np.float32) expect( node, inputs=[data, scales], outputs=[output], name="test_resize_upsample_scales_linear_half_pixel_symmetric", )
51,616
29.114936
104
py
onnx
onnx-main/onnx/backend/test/runner/__init__.py
# Copyright (c) ONNX Project Contributors # # SPDX-License-Identifier: Apache-2.0 from __future__ import annotations import functools import glob import os import re import shutil import sys import tarfile import tempfile import time import unittest from collections import defaultdict from typing import Any, Callable, Iterable, Pattern, Sequence from urllib.request import urlretrieve import numpy as np import onnx import onnx.reference from onnx import ONNX_ML, ModelProto, NodeProto, TypeProto, ValueInfoProto, numpy_helper from onnx.backend.base import Backend from onnx.backend.test.case.test_case import TestCase from onnx.backend.test.loader import load_model_tests from onnx.backend.test.runner.item import TestItem class BackendIsNotSupposedToImplementIt(unittest.SkipTest): pass def retry_execute(times: int) -> Callable[[Callable[..., Any]], Callable[..., Any]]: assert times >= 1 def wrapper(func: Callable[..., Any]) -> Callable[..., Any]: @functools.wraps(func) def wrapped(*args: Any, **kwargs: Any) -> Any: for i in range(1, times + 1): try: return func(*args, **kwargs) except Exception: print(f"{i} times tried") if i == times: raise time.sleep(5 * i) return wrapped return wrapper class Runner: def __init__( self, backend: type[Backend], parent_module: str | None = None ) -> None: self.backend = backend self._parent_module = parent_module self._include_patterns: set[Pattern[str]] = set() self._exclude_patterns: set[Pattern[str]] = set() self._xfail_patterns: set[Pattern[str]] = set() # This is the source of the truth of all test functions. # Properties `test_cases`, `test_suite` and `tests` will be # derived from it. # {category: {name: func}} self._test_items: dict[str, dict[str, TestItem]] = defaultdict(dict) for rt in load_model_tests(kind="node"): self._add_model_test(rt, "Node") for rt in load_model_tests(kind="real"): self._add_model_test(rt, "Real") for rt in load_model_tests(kind="simple"): self._add_model_test(rt, "Simple") for ct in load_model_tests(kind="pytorch-converted"): self._add_model_test(ct, "PyTorchConverted") for ot in load_model_tests(kind="pytorch-operator"): self._add_model_test(ot, "PyTorchOperator") def _get_test_case(self, name: str) -> type[unittest.TestCase]: test_case = type(str(name), (unittest.TestCase,), {}) if self._parent_module: test_case.__module__ = self._parent_module return test_case def include(self, pattern: str) -> Runner: self._include_patterns.add(re.compile(pattern)) return self def exclude(self, pattern: str) -> Runner: self._exclude_patterns.add(re.compile(pattern)) return self def xfail(self, pattern: str) -> Runner: self._xfail_patterns.add(re.compile(pattern)) return self def enable_report(self) -> Runner: import pytest for category, items_map in self._test_items.items(): for item in items_map.values(): item.func = pytest.mark.onnx_coverage(item.proto, category)(item.func) return self @property def _filtered_test_items(self) -> dict[str, dict[str, TestItem]]: filtered: dict[str, dict[str, TestItem]] = {} for category, items_map in self._test_items.items(): filtered[category] = {} for name, item in items_map.items(): if self._include_patterns and ( not any(include.search(name) for include in self._include_patterns) ): item.func = unittest.skip("no matched include pattern")(item.func) for exclude in self._exclude_patterns: if exclude.search(name): item.func = unittest.skip( f'matched exclude pattern "{exclude.pattern}"' )(item.func) for xfail in self._xfail_patterns: if xfail.search(name): item.func = unittest.expectedFailure(item.func) filtered[category][name] = item return filtered @property def test_cases(self) -> dict[str, type[unittest.TestCase]]: """ List of test cases to be applied on the parent scope Example usage: globals().update(BackendTest(backend).test_cases) """ test_cases = {} for category, items_map in self._filtered_test_items.items(): test_case_name = f"OnnxBackend{category}Test" test_case = self._get_test_case(test_case_name) for name, item in sorted(items_map.items()): setattr(test_case, name, item.func) test_cases[test_case_name] = test_case return test_cases @property def test_suite(self) -> unittest.TestSuite: """ TestSuite that can be run by TestRunner Example usage: unittest.TextTestRunner().run(BackendTest(backend).test_suite) """ suite = unittest.TestSuite() for case in sorted( self.test_cases.values(), key=lambda cl: cl.__class__.__name__ ): suite.addTests(unittest.defaultTestLoader.loadTestsFromTestCase(case)) return suite # For backward compatibility (we used to expose `.tests`) @property def tests(self) -> type[unittest.TestCase]: """ One single unittest.TestCase that hosts all the test functions Example usage: onnx_backend_tests = BackendTest(backend).tests """ tests = self._get_test_case("OnnxBackendTest") for items_map in sorted( self._filtered_test_items.values(), key=lambda cl: cl.__class__.__name__ ): for name, item in sorted(items_map.items()): setattr(tests, name, item.func) return tests @classmethod def assert_similar_outputs( cls, ref_outputs: Sequence[Any], outputs: Sequence[Any], rtol: float, atol: float, ) -> None: np.testing.assert_equal(len(outputs), len(ref_outputs)) for i in range(len(outputs)): if isinstance(outputs[i], (list, tuple)): for j in range(len(outputs[i])): cls.assert_similar_outputs( ref_outputs[i][j], outputs[i][j], rtol, atol ) else: np.testing.assert_equal(outputs[i].dtype, ref_outputs[i].dtype) if ref_outputs[i].dtype == object: # type: ignore[attr-defined] np.testing.assert_array_equal(outputs[i], ref_outputs[i]) else: np.testing.assert_allclose( outputs[i], ref_outputs[i], rtol=rtol, atol=atol ) @classmethod @retry_execute(3) def download_model( cls, model_test: TestCase, model_dir: str, models_dir: str ) -> None: # On Windows, NamedTemporaryFile can not be opened for a # second time download_file = tempfile.NamedTemporaryFile(delete=False) try: download_file.close() assert model_test.url print( f"Start downloading model {model_test.model_name} from {model_test.url}" ) urlretrieve(model_test.url, download_file.name) print("Done") with tarfile.open(download_file.name) as t: t.extractall(models_dir) except Exception as e: print(f"Failed to prepare data for model {model_test.model_name}: {e}") raise finally: os.remove(download_file.name) @classmethod def prepare_model_data(cls, model_test: TestCase) -> str: onnx_home = os.path.expanduser( os.getenv("ONNX_HOME", os.path.join("~", ".onnx")) ) models_dir = os.getenv("ONNX_MODELS", os.path.join(onnx_home, "models")) model_dir: str = os.path.join(models_dir, model_test.model_name) if not os.path.exists(os.path.join(model_dir, "model.onnx")): if os.path.exists(model_dir): bi = 0 while True: dest = f"{model_dir}.old.{bi}" if os.path.exists(dest): bi += 1 continue shutil.move(model_dir, dest) break os.makedirs(model_dir) cls.download_model( model_test=model_test, model_dir=model_dir, models_dir=models_dir ) return model_dir def _add_test( self, category: str, test_name: str, test_func: Callable[..., Any], report_item: list[ModelProto | NodeProto | None], devices: Iterable[str] = ("CPU", "CUDA"), ) -> None: # We don't prepend the 'test_' prefix to improve greppability if not test_name.startswith("test_"): raise ValueError(f"Test name must start with test_: {test_name}") def add_device_test(device: str) -> None: device_test_name = f"{test_name}_{device.lower()}" if device_test_name in self._test_items[category]: raise ValueError( f'Duplicated test name "{device_test_name}" in category "{category}"' ) @unittest.skipIf( # type: ignore not self.backend.supports_device(device), f"Backend doesn't support device {device}", ) @functools.wraps(test_func) def device_test_func(*args: Any, **kwargs: Any) -> Any: try: return test_func(*args, device=device, **kwargs) except BackendIsNotSupposedToImplementIt as e: # hacky verbose reporting if "-v" in sys.argv or "--verbose" in sys.argv: print(f"Test {device_test_name} is effectively skipped: {e}") self._test_items[category][device_test_name] = TestItem( device_test_func, report_item ) for device in devices: add_device_test(device) @staticmethod def generate_dummy_data( x: ValueInfoProto, seed: int = 0, name: str = "", random: bool = False ) -> np.ndarray: """ Generates a random tensor based on the input definition. """ if not x.type.tensor_type: raise NotImplementedError( f"Input expected to have tensor type. " f"Unable to generate random data for model {name!r} and input {x}." ) if x.type.tensor_type.elem_type != 1: raise NotImplementedError( f"Currently limited to float tensors. " f"Unable to generate random data for model {name!r} and input {x}." ) shape = tuple( d.dim_value if d.HasField("dim_value") else 1 for d in x.type.tensor_type.shape.dim ) if random: gen = np.random.default_rng(seed=seed) return gen.random(shape, np.float32) n = np.prod(shape) return (np.arange(n).reshape(shape) / n).astype(np.float32) def _add_model_test(self, model_test: TestCase, kind: str) -> None: # model is loaded at runtime, note sometimes it could even # never loaded if the test skipped model_marker: list[ModelProto | NodeProto | None] = [None] def run(test_self: Any, device: str) -> None: if model_test.url is not None and model_test.url.startswith( "onnx/backend/test/data/light/" ): # testing local files model_pb_path = os.path.normpath( os.path.join( os.path.dirname(__file__), "..", "..", "..", "..", model_test.url, ) ) if not os.path.exists(model_pb_path): raise FileNotFoundError(f"Unable to find model {model_pb_path!r}.") onnx_home = os.path.expanduser( os.getenv("ONNX_HOME", os.path.join("~", ".onnx")) ) models_dir = os.getenv( "ONNX_MODELS", os.path.join(onnx_home, "models", "light") ) model_dir: str = os.path.join(models_dir, model_test.model_name) if not os.path.exists(model_dir): os.makedirs(model_dir) use_dummy = True else: if model_test.model_dir is None: model_dir = self.prepare_model_data(model_test) else: model_dir = model_test.model_dir model_pb_path = os.path.join(model_dir, "model.onnx") use_dummy = False if not ONNX_ML and "ai_onnx_ml" in model_dir: return model = onnx.load(model_pb_path) model_marker[0] = model if ( hasattr(self.backend, "is_compatible") and callable(self.backend.is_compatible) and not self.backend.is_compatible(model) ): raise unittest.SkipTest("Not compatible with backend") prepared_model = self.backend.prepare(model, device) assert prepared_model is not None if use_dummy: # When the backend test goes through a test involving a # model stored in onnx/backend/test/data/light, # this function generates expected output coming from # from ReferenceEvaluator run with random inputs. # A couple of models include many Conv operators and the # python implementation is slow (such as test_bvlc_alexnet). with open(model_pb_path, "rb") as f: onx = onnx.load(f) test_data_set = os.path.join(model_dir, "test_data_set_0") if not os.path.exists(test_data_set): os.mkdir(test_data_set) feeds = {} inits = {i.name for i in onx.graph.initializer} n_input = 0 inputs = [] for i in range(len(onx.graph.input)): if onx.graph.input[i].name in inits: continue name = os.path.join(test_data_set, f"input_{n_input}.pb") inputs.append(name) n_input += 1 x = onx.graph.input[i] value = self.generate_dummy_data( x, seed=0, name=model_test.model_name, random=False ) feeds[x.name] = value with open(name, "wb") as f: f.write(onnx.numpy_helper.from_array(value).SerializeToString()) # loads expected output if any available prefix = os.path.splitext(model_pb_path)[0] expected_outputs = [] for i in range(len(onx.graph.output)): name = f"{prefix}_output_{i}.pb" if os.path.exists(name): expected_outputs.append(name) continue expected_outputs = None break if expected_outputs is None: ref = onnx.reference.ReferenceEvaluator(onx) outputs = ref.run(None, feeds) for i, o in enumerate(outputs): name = os.path.join(test_data_set, f"output_{i}.pb") with open(name, "wb") as f: f.write(onnx.numpy_helper.from_array(o).SerializeToString()) else: for i, o in enumerate(expected_outputs): name = os.path.join(test_data_set, f"output_{i}.pb") shutil.copy(o, name) else: # TODO after converting all npz files to protobuf, we can delete this. for test_data_npz in glob.glob( os.path.join(model_dir, "test_data_*.npz") ): test_data = np.load(test_data_npz, encoding="bytes") inputs = list(test_data["inputs"]) outputs = list(prepared_model.run(inputs)) ref_outputs = test_data["outputs"] self.assert_similar_outputs( ref_outputs, outputs, rtol=model_test.rtol, atol=model_test.atol ) for test_data_dir in glob.glob(os.path.join(model_dir, "test_data_set*")): inputs = [] inputs_num = len(glob.glob(os.path.join(test_data_dir, "input_*.pb"))) for i in range(inputs_num): input_file = os.path.join(test_data_dir, f"input_{i}.pb") self._load_proto(input_file, inputs, model.graph.input[i].type) ref_outputs = [] ref_outputs_num = len( glob.glob(os.path.join(test_data_dir, "output_*.pb")) ) for i in range(ref_outputs_num): output_file = os.path.join(test_data_dir, f"output_{i}.pb") self._load_proto( output_file, ref_outputs, model.graph.output[i].type ) outputs = list(prepared_model.run(inputs)) self.assert_similar_outputs( ref_outputs, outputs, rtol=model_test.rtol, atol=model_test.atol ) self._add_test(kind + "Model", model_test.name, run, model_marker) def _load_proto( self, proto_filename: str, target_list: list[np.ndarray | list[Any]], model_type_proto: TypeProto, ) -> None: with open(proto_filename, "rb") as f: protobuf_content = f.read() if model_type_proto.HasField("sequence_type"): sequence = onnx.SequenceProto() sequence.ParseFromString(protobuf_content) target_list.append(numpy_helper.to_list(sequence)) elif model_type_proto.HasField("tensor_type"): tensor = onnx.TensorProto() tensor.ParseFromString(protobuf_content) target_list.append(numpy_helper.to_array(tensor)) elif model_type_proto.HasField("optional_type"): optional = onnx.OptionalProto() optional.ParseFromString(protobuf_content) target_list.append(numpy_helper.to_optional(optional)) # type: ignore[arg-type] else: print( "Loading proto of that specific type (Map/Sparse Tensor) is currently not supported" )
19,487
39.181443
104
py
onnx
onnx-main/onnx/test/reference_evaluator_test.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # type: ignore # pylint: disable=C3001,C0302,C0415,R0904,R0913,R0914,R0915,W0221,W0707 """ You can run a specific test by using the following syntax. :: python onnx/test/reference_evaluator_test.py TestReferenceEvaluator.test_function_attribute_nested_graph """ import itertools import math import unittest from contextlib import redirect_stdout from functools import wraps from io import StringIO from os import getenv from textwrap import dedent from typing import Sequence, Tuple import numpy as np import parameterized from numpy.testing import assert_allclose from onnx import AttributeProto, FunctionProto, ModelProto, TensorProto, checker, parser from onnx.backend.test.case.node.roialign import get_roi_align_input_values from onnx.checker import check_model from onnx.defs import onnx_opset_version from onnx.helper import ( float32_to_bfloat16, float32_to_float8e4m3, float32_to_float8e5m2, make_function, make_graph, make_model, make_model_gen_version, make_node, make_opsetid, make_sequence_type_proto, make_tensor, make_tensor_sequence_value_info, make_tensor_value_info, make_value_info, ) from onnx.numpy_helper import float8e4m3_to_float32, float8e5m2_to_float32, from_array from onnx.reference import ReferenceEvaluator from onnx.reference.op_run import OpRun, OpRunExpand from onnx.reference.ops import load_op from onnx.reference.ops._op_common_indices import _get_indices, _is_out from onnx.reference.ops._op_list import Celu from onnx.reference.ops.aionnx_preview_training._op_list import Adam from onnx.reference.ops.op_celu import _vcelu1 from onnx.reference.ops.op_col2im import ( _col2im_naive_implementation_2d, col2im_naive_implementation, ) from onnx.reference.ops.op_conv import Conv, _conv_implementation from onnx.reference.ops_optimized import Conv as ConvOptimized from onnx.reference.ops_optimized.op_conv_optimized import _conv_implementation_im2col # TODO (https://github.com/microsoft/onnxruntime/issues/14932): Get max supported version from onnxruntime directly # For now, bump the version in CIs whenever there is a new onnxruntime release ORT_MAX_IR_SUPPORTED_VERSION = int(getenv("ORT_MAX_IR_SUPPORTED_VERSION", "8")) ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION = int( getenv("ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION", "18") ) def skip_if_no_onnxruntime(fn): @wraps(fn) def wrapper(*args, **kwargs): try: import onnxruntime # pylint: disable=W0611 del onnxruntime except ImportError: raise unittest.SkipTest("onnxruntime not installed") from None fn(*args, **kwargs) return wrapper def skip_if_no_torch(fn): @wraps(fn) def wrapper(*args, **kwargs): try: import torch # pylint: disable=W0611 del torch except ImportError: raise unittest.SkipTest("torch not installed") from None fn(*args, **kwargs) return wrapper def skip_if_no_torchvision(fn): @wraps(fn) def wrapper(*args, **kwargs): try: import torchvision # pylint: disable=W0611 del torchvision except ImportError: raise unittest.SkipTest("torchvision not installed") from None fn(*args, **kwargs) return wrapper def make_sequence_value_info(name, elem_type, shape): if isinstance(elem_type, int): return make_tensor_sequence_value_info(name, elem_type, shape) s_type = make_sequence_type_proto(elem_type) return make_value_info(name, s_type, shape) def run_ort_inference(onnx_model): import onnxruntime as ort onnx_domain_opset = ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION for opset in onnx_model.opset_import: if opset.domain in ("", "ai.onnx"): onnx_domain_opset = opset.version break # The new IR or opset version is not supported by onnxruntime yet if ( onnx_model.ir_version > ORT_MAX_IR_SUPPORTED_VERSION or onnx_domain_opset > ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION ): return None return ort.InferenceSession( onnx_model.SerializeToString(), providers=["CPUExecutionProvider"] ) def im2col_naive_implementation(data, kernel_shape, dilations, pads, strides): # type: ignore """ Naive implementation for `im2col`. :param image: image (float) :param kernel_shape: kernel shape :param dilations: dilations :param pads: pads :param strides: strides :return: result """ if not isinstance(kernel_shape, tuple): raise TypeError(f"Unexpected type {type(kernel_shape)!r} for kernel_shape.") if len(data.shape) != len(kernel_shape): raise ValueError(f"Shape mismatch {data.shape!r} and {kernel_shape!r}.") n_dims = len(pads) // 2 new_pads = np.array([(pads[i], pads[i + n_dims]) for i in range(n_dims)]) list_output_shape = list(data.shape + kernel_shape) for d in range(n_dims): kd = kernel_shape[d] + (kernel_shape[d] - 1) * (dilations[d] - 1) nd = int( ((list_output_shape[d] - kd + new_pads[d][0] + new_pads[d][1]) / strides[d]) + 1 ) list_output_shape[d] = nd output_shape = tuple(list_output_shape) res = np.zeros(output_shape, dtype=data.dtype) kernel_size = np.prod(kernel_shape) res_size = np.prod(res.shape[:-n_dims]) for i in range(res_size): i_res = _get_indices(i, res.shape[:-n_dims]) t_res = tuple(i_res) for j in range(kernel_size): i_kernel = _get_indices(j, kernel_shape) t_kernel = tuple(i_kernel) i_img = i_res * strides - new_pads[:, 0] + i_kernel * dilations t_img = tuple(i_img) if _is_out(t_img, data.shape): res[t_res + t_kernel] = 0 else: res[t_res + t_kernel] = data[tuple(t_img)] return res def im2col( img: np.ndarray, kernel_shape: Tuple[int, ...], dilations: Sequence[int], pads: Sequence[int], strides: Sequence[int], ) -> np.ndarray: res = None for n in range(img.shape[0]): for c in range(img.shape[1]): out = im2col_naive_implementation( img[n, c, ...], kernel_shape, dilations, pads, strides ) if res is None: new_shape = img.shape[:2] + out.shape res = np.empty(new_shape, dtype=img.dtype) res[n, c, ...] = out new_shape = res.shape[: -len(kernel_shape)] + (-1,) # type: ignore return res.reshape(new_shape) # type: ignore class TestReferenceEvaluator(unittest.TestCase): m2_def = """ < ir_version: 7, opset_import: [ "": 10, "com.microsoft": 1] > agraph (float[N, M] B01, float[N, M] B11, float[N, M] B21) => (float[N, M] D0) { C0 = Add(B01, B11) C1 = Sub(B11, B21) D0 = Mul(C0, C1) } """ @staticmethod def _load_model(m_def: str) -> ModelProto: """ Parses a model from a string representation, including checking the model for correctness """ m = parser.parse_model(m_def) checker.check_model(m) return m @staticmethod def _linear_regression(clip=False, opset=None, min_value=-1.0, max_value=1.0): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.FLOAT, [None, None]) B = make_tensor_value_info("B", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("MatMul", ["X", "A"], ["XA"]) if clip: node2 = make_node("Add", ["XA", "B"], ["Y_clip"]) if opset is not None and opset < 11: if min_value: if max_value: node3 = make_node( "Clip", ["Y_clip"], ["Y"], min=min_value, max=max_value ) else: node3 = make_node("Clip", ["Y_clip"], ["Y"], min=min_value) elif max_value: node3 = make_node("Clip", ["Y_clip"], ["Y"], max=max_value) else: node3 = make_node("Clip", ["Y_clip"], ["Y"]) graph = make_graph([node1, node2, node3], "lr", [X, A, B], [Y]) else: mi = ( from_array(np.array([min_value], dtype=np.float32), name="mi") if min_value else None ) ma = ( from_array(np.array([max_value], dtype=np.float32), name="ma") if max_value else None ) inputs = ["Y_clip", "mi" if mi else "", "ma" if ma else ""] node3 = make_node("Clip", inputs, ["Y"]) initializer = [_ for _ in [mi, ma] if _] graph = make_graph( [node1, node2, node3], "lr", [X, A, B], [Y], initializer=initializer ) f = lambda x, a, b: np.clip(a @ a + b, min_value, max_value) # noqa: E731 else: node2 = make_node("Add", ["XA", "B"], ["Y"]) graph = make_graph([node1, node2], "lr", [X, A, B], [Y]) f = lambda x, a, b: a @ a + b # noqa: E731 if opset is None: onnx_model = make_model(graph) else: onnx_model = make_model(graph, opset_imports=[make_opsetid("", opset)]) try: check_model(onnx_model) except Exception as e: raise AssertionError(f"checker fails for\n{onnx_model}") from e return onnx_model, f def test_reference_evaluator_exceptions(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) with self.assertRaises(TypeError): ReferenceEvaluator(X) def test_reference_evaluator_no_attribute(self): m = TestReferenceEvaluator._load_model(TestReferenceEvaluator.m2_def) checker.check_model(m) sess = ReferenceEvaluator(m) self.assertEqual(sess.input_names, ["B01", "B11", "B21"]) self.assertEqual(sess.output_names, ["D0"]) self.assertEqual(sess.opsets, {"": 10, "com.microsoft": 1}) x = np.array([[0, 1], [2, 3]], dtype=np.float32) y = np.array([[4, 5], [6, 7]], dtype=np.float32) z = np.array([[-4, -5], [-6, -7]], dtype=np.float32) res = sess.run(None, {"B01": x, "B11": y, "B21": z})[0] expected = (x + y) * (y - z) assert_allclose(expected, res) def test_reference_evaluator_no_attribute_bytes(self): m = TestReferenceEvaluator._load_model(TestReferenceEvaluator.m2_def) checker.check_model(m) sess = ReferenceEvaluator(m.SerializeToString()) self.assertEqual(sess.input_names, ["B01", "B11", "B21"]) self.assertEqual(sess.output_names, ["D0"]) self.assertEqual(sess.opsets, {"": 10, "com.microsoft": 1}) x = np.array([[0, 1], [2, 3]], dtype=np.float32) y = np.array([[4, 5], [6, 7]], dtype=np.float32) z = np.array([[-4, -5], [-6, -7]], dtype=np.float32) res = sess.run(None, {"B01": x, "B11": y, "B21": z})[0] expected = (x + y) * (y - z) assert_allclose(expected, res) def test_reference_evaluator_no_attribute_verbose(self): m = TestReferenceEvaluator._load_model(TestReferenceEvaluator.m2_def) x = np.array([[0, 1], [2, 3]], dtype=np.float32) y = np.array([[4, 5], [6, 7]], dtype=np.float32) z = np.array([[-4, -5], [-6, -7]], dtype=np.float32) with self.subTest(level=2): sess = ReferenceEvaluator(m, verbose=2) stdout = StringIO() with redirect_stdout(stdout): sess.run(None, {"B01": x, "B11": y, "B21": z}) out = stdout.getvalue() log = "Add(B01, B11) -> C0\nSub(B11, B21) -> C1\nMul(C0, C1) -> D0\n" self.assertEqual(log, out) with self.subTest(level=3): sess = ReferenceEvaluator(m, verbose=3) stdout = StringIO() with redirect_stdout(stdout): sess.run(None, {"B01": x, "B11": y, "B21": z}) out = stdout.getvalue() log = dedent( """ +I B01: float32:(2, 2) in [0.0, 3.0] +I B11: float32:(2, 2) in [4.0, 7.0] +I B21: float32:(2, 2) in [-7.0, -4.0] Add(B01, B11) -> C0 + C0: float32:(2, 2) in [4.0, 10.0] Sub(B11, B21) -> C1 + C1: float32:(2, 2) in [8.0, 14.0] Mul(C0, C1) -> D0 + D0: float32:(2, 2) in [32.0, 140.0] """ ).lstrip("\n") self.assertEqual(log, out) with self.subTest(level=4): sess = ReferenceEvaluator(m, verbose=4) stdout = StringIO() with redirect_stdout(stdout): sess.run(None, {"B01": x, "B11": y, "B21": z}) out = stdout.getvalue() log = dedent( """ +I B01: float32:(2, 2):[0.0, 1.0, 2.0, 3.0] +I B11: float32:(2, 2):[4.0, 5.0, 6.0, 7.0] +I B21: float32:(2, 2):[-4.0, -5.0, -6.0, -7.0] Add(B01, B11) -> C0 + C0: float32:(2, 2):[4.0, 6.0, 8.0, 10.0] Sub(B11, B21) -> C1 + C1: float32:(2, 2):[8.0, 10.0, 12.0, 14.0] Mul(C0, C1) -> D0 + D0: float32:(2, 2):[32.0, 60.0, 96.0, 140.0] """ ).lstrip("\n") self.assertEqual(log, out) with self.subTest(level=15): sess = ReferenceEvaluator(m, verbose=15) stdout = StringIO() with redirect_stdout(stdout): sess.run(None, {"B01": x, "B11": y, "B21": z}) out = stdout.getvalue() log = dedent( """ +I B01: float32:(2, 2):[0.0, 1.0, 2.0, 3.0] +I B11: float32:(2, 2):[4.0, 5.0, 6.0, 7.0] +I B21: float32:(2, 2):[-4.0, -5.0, -6.0, -7.0] Add(B01, B11) -> C0 -- begin Add.run(2 inputs) -- done Add.run -> 1 outputs + C0: float32:(2, 2):[4.0, 6.0, 8.0, 10.0] Sub(B11, B21) -> C1 -- begin Sub.run(2 inputs) -- done Sub.run -> 1 outputs + C1: float32:(2, 2):[8.0, 10.0, 12.0, 14.0] Mul(C0, C1) -> D0 -- begin Mul.run(2 inputs) -- done Mul.run -> 1 outputs + D0: float32:(2, 2):[32.0, 60.0, 96.0, 140.0] """ ).lstrip("\n") self.assertEqual(log, out) def test_reference_evaluator_lr(self): lr, f = TestReferenceEvaluator._linear_regression() x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) def test_reference_evaluator_lr_clip(self): with self.subTest(opt="min+max"): lr, f = TestReferenceEvaluator._linear_regression(clip=True) x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) last_node = sess.rt_nodes_[-1] self.assertEqual(last_node.__class__.__name__, "Clip_11") got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) with self.subTest(opt="max"): lr, f = TestReferenceEvaluator._linear_regression(clip=True, min_value=None) x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) last_node = sess.rt_nodes_[-1] self.assertEqual(last_node.__class__.__name__, "Clip_11") got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) with self.subTest(opt="min"): lr, f = TestReferenceEvaluator._linear_regression(clip=True, max_value=None) x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) last_node = sess.rt_nodes_[-1] self.assertEqual(last_node.__class__.__name__, "Clip_11") got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) def test_reference_evaluator_lr_clip_6(self): with self.subTest(opt="min+max"): lr, f = TestReferenceEvaluator._linear_regression(clip=True, opset=10) x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) last_node = sess.rt_nodes_[-1] self.assertEqual(last_node.__class__.__name__, "Clip_6") self.assertEqual(last_node.min, -1) self.assertEqual(last_node.max, 1) got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) with self.subTest(opt="max"): lr, f = TestReferenceEvaluator._linear_regression( clip=True, opset=10, min_value=None ) x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) last_node = sess.rt_nodes_[-1] self.assertEqual(last_node.__class__.__name__, "Clip_6") self.assertEqual(last_node.max, 1) self.assertEqual(last_node.min, -3.4028234663852886e38) got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) with self.subTest(opt="min"): lr, f = TestReferenceEvaluator._linear_regression( clip=True, opset=10, max_value=None ) x = np.array([[0, 1], [2, 3]], dtype=np.float32) a = np.array([1, 1], dtype=np.float32) b = np.array([11], dtype=np.float32) expected = f(x, a, b) sess = ReferenceEvaluator(lr) last_node = sess.rt_nodes_[-1] self.assertEqual(last_node.__class__.__name__, "Clip_6") self.assertEqual(last_node.min, -1) self.assertEqual(last_node.max, 3.4028234663852886e38) got = sess.run(None, {"X": a, "A": a, "B": b})[0] assert_allclose(expected, got) def test_nested_local_functions(self): m = parser.parse_model( """ < ir_version: 8, opset_import: [ "" : 14, "local" : 1], producer_name: "test", producer_version: "1.0", model_version: 1, doc_string: "Test preprocessing model" > agraph (uint8[H, W, C] x) => (uint8[H, W, C] x_processed) { x_processed = local.func(x) } < opset_import: [ "" : 14 ], domain: "local", doc_string: "function 1" > f1 (x) => (y) { y = Identity(x) } < opset_import: [ "" : 14 ], domain: "local", doc_string: "function 2" > f2 (x) => (y) { y = Identity(x) } < opset_import: [ "" : 14, "local" : 1 ], domain: "local", doc_string: "Preprocessing function." > func (x) => (y) { x1 = local.f1(x) y = local.f2(x1) } """ ) sess = ReferenceEvaluator(m) x = np.array([0, 1, 3], dtype=np.uint8).reshape((1, 1, 3)) result = sess.run(None, {"x": x})[0] expected = x assert_allclose(expected, result) def test_reduce_sum_11(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSum", ["X"], ["Y"], axes=[1], keepdims=1) graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 11)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) expected = x.sum(axis=1, keepdims=1) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x})[0] assert_allclose(expected, got) def test_reduce_sum_square_11(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSumSquare", ["X"], ["Y"], axes=[1], keepdims=1) graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 11)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) expected = (x * x).sum(axis=1, keepdims=1) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x})[0] assert_allclose(expected, got) def test_reduce_sum_13(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.INT64, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSum", ["X", "A"], ["Y"], keepdims=1) graph = make_graph([node1], "rs", [X, A], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 13)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) a = np.array([1], dtype=np.int64) expected = x.sum(axis=1, keepdims=1) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x, "A": a})[0] assert_allclose(expected, got) def test_reduce_sum_attribute(self): opset = onnx_opset_version() new_domain = "custom" opset_imports = [make_opsetid("", opset), make_opsetid(new_domain, 1)] node = make_node("ReduceSum", ["X", "axis"], ["Y"]) att = AttributeProto() att.name = "keepdims" att.ref_attr_name = "keepdims" att.type = AttributeProto.INT node.attribute.append(att) my_reduce_sum = make_function( new_domain, "MyReduceSum", ["X", "axis"], ["Y"], [node], opset_imports, ["keepdims"], ) X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) axis = make_tensor_value_info("axis", TensorProto.INT64, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) graph = make_graph( [ make_node( "MyReduceSum", ["X", "axis"], ["Y"], domain=new_domain, keepdims=1, ), ], "example", [X, axis], [Y], ) onnx_model = make_model( graph, opset_imports=opset_imports, functions=[my_reduce_sum] ) sess = ReferenceEvaluator(onnx_model) x = np.arange(6).reshape((3, 2)).astype(np.float32) a = np.array([-1], dtype=np.int64) result = sess.run(None, {"X": x, "axis": a})[0] expected = x.sum(axis=-1, keepdims=1) assert_allclose(expected, result) def test_reduce_sum_square_18(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.INT64, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSumSquare", ["X", "A"], ["Y"], keepdims=1) graph = make_graph([node1], "rs", [X, A], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) a = np.array([1], dtype=np.int64) expected = (x * x).sum(axis=1, keepdims=1) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x, "A": a})[0] assert_allclose(expected, got) def test_reduce_sum_13_empty_axes(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.INT64, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSum", ["X", "A"], ["Y"], keepdims=1) graph = make_graph([node1], "rs", [X, A], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 13)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) a = np.array([], dtype=np.int64) expected = x.sum(keepdims=1) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x, "A": a})[0] assert_allclose(expected, got) def test_reduce_sum_square_18_empty_axes(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.INT64, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSumSquare", ["X", "A"], ["Y"], keepdims=1) graph = make_graph([node1], "rs", [X, A], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) a = np.array([], dtype=np.int64) expected = (x * x).sum(keepdims=1) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x, "A": a})[0] assert_allclose(expected, got) def test_reduce_sum_13_empty_axes_noop(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("ReduceSum", ["X"], ["Y"], keepdims=1, noop_with_empty_axes=1) graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 13)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x})[0] assert_allclose(x, got) def test_reduce_sum_square_18_empty_axes_noop(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node( "ReduceSumSquare", ["X"], ["Y"], keepdims=1, noop_with_empty_axes=1 ) graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) check_model(onnx_model) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x})[0] assert_allclose(x * x, got) def test_greater(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) Z = make_tensor_value_info("Z", TensorProto.FLOAT, [None]) node1 = make_node("Greater", ["X", "Y"], ["Z"]) graph = make_graph([node1], "g", [X, Y], [Z]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 13)]) check_model(onnx_model) x = np.arange(4).reshape((2, 2)).astype(np.float32) y = np.array([2], dtype=np.float32) expected = x > y sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x, "Y": y})[0] assert_allclose(expected, got) def test_node_proto(self): node1 = make_node("Greater", ["X", "Y"], ["Z"]) x = np.arange(4).reshape((2, 2)).astype(np.float32) y = np.array([2], dtype=np.float32) expected = x > y sess = ReferenceEvaluator(node1) got = sess.run(None, {"X": x, "Y": y})[0] assert_allclose(expected, got) def test_greater_or_equal(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) Z = make_tensor_value_info("Z", TensorProto.FLOAT, [None]) node1 = make_node("GreaterOrEqual", ["X", "Y"], ["Z"]) graph = make_graph([node1], "g", [X, Y], [Z]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 13)]) check_model(onnx_model) x = np.arange(4).reshape((2, 2)).astype(np.float32) y = np.array([2], dtype=np.float32) expected = x >= y sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": x, "Y": y})[0] assert_allclose(expected, got) def test_if(self): C = make_tensor_value_info("C", TensorProto.FLOAT, [None]) bthen = make_node( "Constant", [], ["C"], value_floats=from_array(np.array([1], dtype=np.float32)), ) bthen_body = make_graph([bthen], "gthen", [], [C]) C = make_tensor_value_info("C", TensorProto.FLOAT, [None]) belse = make_node( "Constant", [], ["C"], value_floats=from_array(np.array([0], dtype=np.float32)), ) belse_body = make_graph([belse], "gelse", [], [C]) zero = from_array(np.array([0], dtype=np.float32), name="zero") greater = make_node("Greater", ["X", "zero"], ["G"]) node_if = make_node( "If", ["G"], ["Z"], then_branch=bthen_body, else_branch=belse_body, ) X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Z = make_tensor_value_info("Z", TensorProto.FLOAT, [None]) graph = make_graph([greater, node_if], "g", [X], [Z], initializer=[zero]) model_def = make_model(graph) sess = ReferenceEvaluator(model_def) self.assertEqual(str(sess), "ReferenceEvaluator(X) -> Z") x = np.array([1, 2], dtype=np.float32) got = sess.run(None, {"X": x})[0] assert_allclose(np.array([1], dtype=np.float32), got) x = np.array([-1, -2], dtype=np.float32) got = sess.run(None, {"X": x})[0] assert_allclose(np.array([0], dtype=np.float32), got) def test_if_function(self): then_out = make_tensor_value_info("then_out", TensorProto.FLOAT, [5]) else_out = make_tensor_value_info("else_out", TensorProto.FLOAT, [5]) x = np.array([1, 2, 3, 4, 5]).astype(np.float32) y = np.array([5, 4, 3, 2, 1]).astype(np.float32) then_const_node = make_node( "Constant", inputs=[], outputs=["then_out"], value=from_array(x) ) else_const_node = make_node( "Constant", inputs=[], outputs=["else_out"], value=from_array(y) ) then_body = make_graph([then_const_node], "then_body", [], [then_out]) else_body = make_graph([else_const_node], "else_body", [], [else_out]) if_node = make_node( "If", inputs=["f_cond"], outputs=["f_res"], then_branch=then_body, else_branch=else_body, ) f = FunctionProto() f.domain = "custom" f.name = "fn" f.input.extend(["f_cond"]) f.output.extend(["f_res"]) f.node.extend([if_node]) opset = onnx_opset_version() f.opset_import.extend([make_opsetid("", opset)]) graph = make_graph( nodes=[make_node("fn", domain="custom", inputs=["cond"], outputs=["res"])], name="graph", inputs=[make_tensor_value_info("cond", TensorProto.BOOL, [])], outputs=[make_tensor_value_info("res", TensorProto.FLOAT, [5])], ) m = make_model( graph, producer_name="test", opset_imports=[make_opsetid("", opset), make_opsetid("custom", 1)], ) m.functions.extend([f]) sess = ReferenceEvaluator(m) result = sess.run(None, {"cond": np.array(True)}) expected = np.array([1, 2, 3, 4, 5], dtype=np.float32) assert_allclose(expected, result[0]) def test_function_attribute(self): opset = onnx_opset_version() new_domain = "custom" opset_imports = [make_opsetid("", opset), make_opsetid(new_domain, 1)] cst = make_node("Constant", [], ["B"]) att = AttributeProto() att.name = "value" att.ref_attr_name = "bias" att.type = AttributeProto.TENSOR cst.attribute.append(att) node1 = make_node("MatMul", ["X", "A"], ["XA"]) node2 = make_node("Add", ["XA", "B"], ["Y"]) linear_regression = make_function( new_domain, "LinearRegression", ["X", "A"], ["Y"], [cst, node1, node2], opset_imports, ["bias"], ) X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) graph = make_graph( [ make_node( "LinearRegression", ["X", "A"], ["Y1"], domain=new_domain, bias=make_tensor("former_B", TensorProto.FLOAT, [1], [0.67]), ), make_node("Abs", ["Y1"], ["Y"]), ], "example", [X, A], [Y], ) onnx_model = make_model( graph, opset_imports=opset_imports, functions=[linear_regression] ) sess = ReferenceEvaluator(onnx_model) x = np.arange(6).reshape((3, 2)).astype(np.float32) a = np.array([1, -1], dtype=np.float32) result = sess.run(None, {"X": x, "A": a})[0] expected = np.abs(x @ a + 0.67) assert_allclose(expected, result) def test_function_attribute_nested_graph(self): opset = onnx_opset_version() new_domain = "custom" opset_imports = [make_opsetid("", opset), make_opsetid(new_domain, 1)] cst1 = make_node("Constant", [], ["B1"]) att = AttributeProto() att.name = "value" att.ref_attr_name = "bias1" att.type = AttributeProto.TENSOR cst1.attribute.append(att) cst2 = make_node("Constant", [], ["B2"]) att = AttributeProto() att.name = "value" att.ref_attr_name = "bias2" att.type = AttributeProto.TENSOR cst2.attribute.append(att) then_out = make_tensor_value_info("B1", TensorProto.FLOAT, [None]) else_out = make_tensor_value_info("B2", TensorProto.FLOAT, [None]) then_body = make_graph([cst1], "then_body", [], [then_out]) else_body = make_graph([cst2], "else_body", [], [else_out]) zero = make_node( "Constant", inputs=[], outputs=["zero"], value=from_array(np.array([0], dtype=np.float32)), ) mini = make_node("ReduceMin", ["X"], ["Xmin"]) f_cond = make_node("Greater", ["Xmin", "zero"], ["f_cond"]) if_node = make_node( "If", inputs=["f_cond"], outputs=["B"], then_branch=then_body, else_branch=else_body, ) node1 = make_node("MatMul", ["X", "A"], ["XA"]) node2 = make_node("Add", ["XA", "B"], ["Y"]) linear_regression = make_function( new_domain, "LinearRegression", ["X", "A"], ["Y"], [zero, mini, f_cond, if_node, node1, node2], opset_imports, ["bias1", "bias2"], ) X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) graph = make_graph( [ make_node( "LinearRegression", ["X", "A"], ["Y1"], domain=new_domain, bias1=make_tensor("former_B1", TensorProto.FLOAT, [1], [0.67]), bias2=make_tensor("former_B2", TensorProto.FLOAT, [1], [777]), ), make_node("Abs", ["Y1"], ["Y"]), ], "example", [X, A], [Y], ) onnx_model = make_model( graph, opset_imports=opset_imports, functions=[linear_regression] ) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) self.assertEqual(sess.rt_nodes_[0].__class__.__name__, "OpFunction") self.assertEqual( sess.rt_nodes_[0].impl_.__class__.__name__, "ReferenceEvaluator" ) fct = sess.rt_nodes_[0].impl_ checked = False for node in fct.rt_nodes_: if node.__class__.__name__.startswith("If"): if not node.has_linked_attribute: raise AssertionError( f"Nested node {type(node)} declares no linked attribute " f"but a subgraph does." ) checked = True if not checked: raise AssertionError( "No node 'If' was found, has_linked_attribute could not be checked." ) x = np.arange(6).reshape((3, 2)).astype(np.float32) a = np.array([1, -1], dtype=np.float32) result = sess.run(None, {"X": x + 1, "A": a})[0] expected = np.abs(x @ a + 0.67) assert_allclose(expected, result) result = sess.run(None, {"X": x - 10, "A": a})[0] expected = np.abs(x @ a + 777) assert_allclose(expected, result) def test_function_attribute_nested_nested_graph(self): opset = onnx_opset_version() new_domain = "custom" opset_imports = [make_opsetid("", opset), make_opsetid(new_domain, 1)] # first If cst1 = make_node("Constant", [], ["B1"]) att = AttributeProto() att.name = "value" att.ref_attr_name = "bias1" att.type = AttributeProto.TENSOR cst1.attribute.append(att) cst2 = make_node("Constant", [], ["B2"]) att = AttributeProto() att.name = "value" att.ref_attr_name = "bias2" att.type = AttributeProto.TENSOR cst2.attribute.append(att) then_out = make_tensor_value_info("B1", TensorProto.FLOAT, [None]) else_out = make_tensor_value_info("B2", TensorProto.FLOAT, [None]) then_body1 = make_graph([cst1], "then_body", [], [then_out]) else_body1 = make_graph([cst2], "else_body", [], [else_out]) # sub graph 2 c100 = make_node( "Constant", inputs=[], outputs=["c100"], value=from_array(np.array([100], dtype=np.float32)), ) f_cond = make_node("Greater", ["Xmin", "c100"], ["f_cond_100"]) if_node = make_node( "If", inputs=["f_cond_100"], outputs=["B4"], then_branch=then_body1, else_branch=else_body1, ) # second If cst3 = make_node("Constant", [], ["B3"]) att = AttributeProto() att.name = "value" att.ref_attr_name = "bias3" att.type = AttributeProto.TENSOR cst3.attribute.append(att) then_out = make_tensor_value_info("B3", TensorProto.FLOAT, [None]) then_body2 = make_graph([cst3], "then_body", [], [then_out]) else_out = make_tensor_value_info("B4", TensorProto.FLOAT, [None]) else_body2 = make_graph([c100, f_cond, if_node], "else_body", [], [else_out]) # function zero = make_node( "Constant", inputs=[], outputs=["zero"], value=from_array(np.array([0], dtype=np.float32)), ) mini = make_node("ReduceMin", ["X"], ["Xmin"]) f_cond = make_node("Less", ["Xmin", "zero"], ["f_cond_zero"]) if_node = make_node( "If", inputs=["f_cond_zero"], outputs=["B"], then_branch=then_body2, else_branch=else_body2, ) node1 = make_node("MatMul", ["X", "A"], ["XA"]) node2 = make_node("Add", ["XA", "B"], ["Y"]) linear_regression = make_function( new_domain, "LinearRegression", ["X", "A"], ["Y"], [zero, mini, f_cond, if_node, node1, node2], opset_imports, ["bias1", "bias2", "bias3"], ) X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) A = make_tensor_value_info("A", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) graph = make_graph( [ make_node( "LinearRegression", ["X", "A"], ["Y1"], domain=new_domain, bias1=make_tensor("former_B1", TensorProto.FLOAT, [1], [0.67]), bias2=make_tensor("former_B2", TensorProto.FLOAT, [1], [777]), bias3=make_tensor("former_B3", TensorProto.FLOAT, [1], [-888]), ), make_node("Abs", ["Y1"], ["Y"]), ], "example", [X, A], [Y], ) onnx_model = make_model( graph, opset_imports=opset_imports, functions=[linear_regression] ) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) x = np.arange(6).reshape((3, 2)).astype(np.float32) a = np.array([1, -1], dtype=np.float32) result = sess.run(None, {"X": x + 1, "A": a})[0] expected = np.abs(x @ a + 777) assert_allclose(expected, result) result = sess.run(None, {"X": x - 10, "A": a})[0] expected = np.abs(x @ a - 888) assert_allclose(expected, result) result = sess.run(None, {"X": x + 1000, "A": a})[0] expected = np.abs(x @ a + 0.67) assert_allclose(expected, result) def test_custom_node(self): class _InvAlpha: op_domain = "custom" def __init__(self, onnx_node, run_params): # type: ignore self.onnx_node = onnx_node self.run_params = run_params def _run(self, x): # type: ignore return (1 / (x + self.alpha),) class InvAlpha2(OpRun): def _run(self, x): # type: ignore return (1 / (x + self.alpha),) class InvAlpha(OpRun): op_domain = "custom" def _run(self, x, alpha=None): # type: ignore alpha = alpha or self.alpha # type: ignore return (1 / (x + alpha),) X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("InvAlpha", ["X"], ["Y"], alpha=0.5, domain="custom") graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("custom", 1)]) x = np.arange(60).reshape((3, 4, 5)).astype(np.float32) + 1 with self.assertRaises(NotImplementedError): ReferenceEvaluator(onnx_model) node1 = make_node("_InvAlpha", ["X"], ["Y"], alpha=0.5, domain="custom") graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("custom", 1)]) with self.assertRaises(TypeError): ReferenceEvaluator(onnx_model, new_ops=[_InvAlpha]) node1 = make_node("InvAlpha2", ["X"], ["Y"], alpha=0.5, domain="custom") graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("custom", 1)]) with self.assertRaises(NotImplementedError): ReferenceEvaluator(onnx_model, new_ops=[InvAlpha2]) node1 = make_node("InvAlpha", ["X"], ["Y"], alpha=0.5, domain="custom") graph = make_graph([node1], "rs", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("custom", 1)]) sess = ReferenceEvaluator(onnx_model, new_ops=[InvAlpha, InvAlpha]) got = sess.run(None, {"X": x})[0] expected = 1 / (x + 0.5) assert_allclose(expected, got) def test_loop(self): # Given a tensor x of values [x1, ..., xN], # Return a sequence of tensors of # [[x1], [x1, x2], ..., [x1, ..., xN]] cond_in = make_tensor_value_info("cond_in", TensorProto.BOOL, []) cond_out = make_tensor_value_info("cond_out", TensorProto.BOOL, []) iter_count = make_tensor_value_info("iter_count", TensorProto.INT64, []) seq_in = make_tensor_sequence_value_info("seq_in", TensorProto.FLOAT, None) seq_out = make_tensor_sequence_value_info("seq_out", TensorProto.FLOAT, None) x = np.array([1, 2, 3, 4, 5]).astype(np.float32) x_const_node = make_node( "Constant", inputs=[], outputs=["x"], value=make_tensor( name="const_tensor_x", data_type=TensorProto.FLOAT, dims=x.shape, vals=x.flatten().astype(float), ), ) one_const_node = make_node( "Constant", inputs=[], outputs=["one"], value=make_tensor( name="const_tensor_one", data_type=TensorProto.INT64, dims=(), vals=[1], ), ) zero_const_node = make_node( "Constant", inputs=[], outputs=["slice_start"], value=make_tensor( name="const_tensor_zero", data_type=TensorProto.INT64, dims=(1,), vals=[0], ), ) axes_node = make_node( "Constant", inputs=[], outputs=["axes"], value=make_tensor( name="const_tensor_axes", data_type=TensorProto.INT64, dims=(), vals=[0], ), ) add_node = make_node("Add", inputs=["iter_count", "one"], outputs=["end"]) end_unsqueeze_node = make_node( "Unsqueeze", inputs=["end", "axes"], outputs=["slice_end"] ) slice_node = make_node( "Slice", inputs=["x", "slice_start", "slice_end"], outputs=["slice_out"] ) insert_node = make_node( "SequenceInsert", inputs=["seq_in", "slice_out"], outputs=["seq_out"] ) identity_node = make_node("Identity", inputs=["cond_in"], outputs=["cond_out"]) loop_body = make_graph( [ identity_node, x_const_node, one_const_node, zero_const_node, add_node, axes_node, end_unsqueeze_node, slice_node, insert_node, ], "loop_body", [iter_count, cond_in, seq_in], [cond_out, seq_out], ) node = make_node( "Loop", inputs=["trip_count", "cond", "seq_empty"], outputs=["seq_res"], body=loop_body, ) node_concat = make_node( "ConcatFromSequence", inputs=["seq_res"], outputs=["res"], axis=0, new_axis=0, ) trip_count = np.array(5).astype(np.int64) seq_empty = [] # type: List[Any] # seq_res = [x[:int(i)] for i in x] cond = np.array(1).astype(np.bool_) model_def = make_model( graph=make_graph( name="loop_test", inputs=[ make_tensor_value_info( "trip_count", TensorProto.INT64, trip_count.shape ), make_tensor_value_info("cond", TensorProto.BOOL, cond.shape), make_sequence_value_info("seq_empty", TensorProto.FLOAT, []), ], outputs=[make_tensor_value_info("res", TensorProto.FLOAT, None)], nodes=[node, node_concat], ) ) expected = np.array( [1.0, 1.0, 2.0, 1.0, 2.0, 3.0, 1.0, 2.0, 3.0, 4.0, 1.0, 2.0, 3.0, 4.0, 5.0], dtype=np.float32, ) oinf = ReferenceEvaluator(model_def) inputs = {"trip_count": trip_count, "cond": cond, "seq_empty": seq_empty} got = oinf.run(None, inputs) assert_allclose(expected, got[0]) def test_onnxt_runtime_bernoulli(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("Bernoulli", ["X"], ["Y"], seed=0.0) graph = make_graph([node1], "g", [X], [Y]) onnx_model = make_model(graph) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": np.zeros((2, 4), dtype=np.float32) + 0.5})[0] self.assertEqual(got.shape, (2, 4)) self.assertEqual(got.dtype, np.float32) self.assertGreater(got.min(), -1e-5) self.assertLess(got.max(), 1 + 1e-5) def test_onnxt_runtime_random_uniform(self): Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("RandomUniform", [], ["Y"], seed=0.0, shape=[2, 4]) graph = make_graph([node1], "g", [], [Y]) onnx_model = make_model(graph) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {})[0] self.assertEqual(got.shape, (2, 4)) self.assertEqual(got.dtype, np.float32) self.assertGreater(got.min(), 0) self.assertLess(got.max(), 1) def test_onnxt_runtime_random_uniform_like(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("RandomUniformLike", ["X"], ["Y"], seed=0.0) graph = make_graph([node1], "g", [X], [Y]) onnx_model = make_model(graph) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": np.zeros((2, 4), dtype=np.float32)})[0] self.assertEqual(got.shape, (2, 4)) self.assertEqual(got.dtype, np.float32) self.assertGreater(got.min(), 0) self.assertLess(got.max(), 1) def test_onnxt_runtime_random_normal(self): Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("RandomNormal", [], ["Y"], seed=0.0, shape=[2, 4]) graph = make_graph([node1], "g", [], [Y]) onnx_model = make_model(graph) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {})[0] self.assertEqual(got.shape, (2, 4)) self.assertEqual(got.dtype, np.float32) def test_onnxt_runtime_random_normal_like(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("RandomNormalLike", ["X"], ["Y"], seed=0.0) graph = make_graph([node1], "g", [X], [Y]) onnx_model = make_model(graph) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) got = sess.run(None, {"X": np.zeros((2, 4), dtype=np.float32)})[0] self.assertEqual(got.shape, (2, 4)) self.assertEqual(got.dtype, np.float32) def test_eval_celu(self): inst = Celu.create(alpha=0.5) self.assertEqual(inst.alpha, 0.5) x = np.array([[0, 1], [-1, 2]], dtype=np.float32) y = Celu.eval(x, alpha=0.5) expected = _vcelu1(x, alpha=0.5) assert_allclose(expected, y) def test_eval_celu_load_op(self): celu = load_op("", "Celu") self.assertEqual(celu.op_domain, "") inst = celu.create(alpha=0.5) self.assertEqual(inst.alpha, 0.5) x = np.array([[0, 1], [-1, 2]], dtype=np.float32) y = celu.eval(x, alpha=0.5) expected = _vcelu1(x, alpha=0.5) assert_allclose(expected, y) def test_create_adam(self): inst = Adam.create(alpha=0.5) self.assertEqual(inst.alpha, 0.5) @skip_if_no_onnxruntime def test_conv(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) B = make_tensor_value_info("B", TensorProto.FLOAT, [None, None, None, None]) W = make_tensor_value_info("W", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "Conv", ["X", "W", "B"], ["Y"], pads=[1, 1, 1, 1], dilations=[1, 1], strides=[2, 2], ) graph = make_graph([node], "g", [X, W, B], [Y]) onnx_model = make_model_gen_version(graph, opset_imports=[make_opsetid("", 16)]) sess1 = run_ort_inference(onnx_model) if sess1 is None: return sess2 = ReferenceEvaluator(onnx_model, optimized=False) self.assertIsInstance(sess2.rt_nodes_[0], Conv) sess3 = ReferenceEvaluator(onnx_model, new_ops=[ConvOptimized], optimized=False) self.assertIsInstance(sess3.rt_nodes_[0], ConvOptimized) sess4 = ReferenceEvaluator(onnx_model, optimized=True) self.assertIsInstance(sess4.rt_nodes_[0], ConvOptimized) sH, sW = 5, 6 for i in range(sH): for j in range(sW): X = np.zeros((1, 1, sH, sW), dtype=np.float32) X[0, 0, i, j] = 1.0 W = np.zeros((1, 1, 3, 3), dtype=np.float32) W[0, 0, :, :] = np.minimum(2 ** np.arange(9).reshape((3, -1)), 256) B = np.array([[[[0]]]], dtype=np.float32) expected = sess1.run(None, {"X": X, "W": W, "B": B})[0] got = sess2.run(None, {"X": X, "W": W, "B": B})[0] assert_allclose(expected, got) got3 = sess3.run(None, {"X": X, "W": W, "B": B})[0] assert_allclose(expected, got3) got4 = sess4.run(None, {"X": X, "W": W, "B": B})[0] assert_allclose(expected, got4) @skip_if_no_onnxruntime def test_qlinearconv(self): x = make_tensor_value_info("x", TensorProto.UINT8, [None, None, None, None]) w = make_tensor_value_info("w", TensorProto.UINT8, [None, None, None, None]) y = make_tensor_value_info("y", TensorProto.UINT8, [None, None, None, None]) x_scale = make_tensor_value_info("x_scale", TensorProto.FLOAT, [None]) w_scale = make_tensor_value_info("w_scale", TensorProto.FLOAT, [None]) y_scale = make_tensor_value_info("y_scale", TensorProto.FLOAT, [None]) x_zero_point = make_tensor_value_info("x_zero_point", TensorProto.UINT8, [None]) w_zero_point = make_tensor_value_info("w_zero_point", TensorProto.UINT8, [None]) y_zero_point = make_tensor_value_info("y_zero_point", TensorProto.UINT8, [None]) node = make_node( "QLinearConv", [ "x", "x_scale", "x_zero_point", "w", "w_scale", "w_zero_point", "y_scale", "y_zero_point", ], ["y"], ) graph = make_graph( [node], "g", [x, x_scale, x_zero_point, w, w_scale, w_zero_point, y_scale, y_zero_point], [y], ) onnx_model = make_model_gen_version(graph, opset_imports=[make_opsetid("", 16)]) sess1 = run_ort_inference(onnx_model) if sess1 is None: return sess2 = ReferenceEvaluator(onnx_model) sH, sW = 3, 3 for i in range(sH): for j in range(sW): x = np.zeros((1, 1, sH, sW), dtype=np.uint8) x[0, 0, i, j] = 1.0 with self.subTest(w="1x1", i=i, j=j): w = np.zeros((1, 1, 1, 1), dtype=np.uint8) w[0, 0, :, :] = 1 feeds = { "x": x, "x_scale": np.array([1], dtype=np.float32), "x_zero_point": np.array([0], dtype=np.uint8), "w": w, "w_scale": np.array([1], dtype=np.float32), "w_zero_point": np.array([0], dtype=np.uint8), "y_scale": np.array([1], dtype=np.float32), "y_zero_point": np.array([0], np.uint8), } expected = sess1.run(None, feeds)[0] got = sess2.run(None, feeds)[0] try: assert_allclose(expected, got) except AssertionError as e: raise e with self.subTest(w="3x3", i=i, j=j): w = np.zeros((1, 1, 3, 3), dtype=np.uint8) w[0, 0, :, :] = np.minimum(2 ** np.arange(9).reshape((3, -1)), 128) feeds = { "x": x, "x_scale": np.array([1], dtype=np.float32), "x_zero_point": np.array([0], dtype=np.uint8), "w": w, "w_scale": np.array([1], dtype=np.float32), "w_zero_point": np.array([0], dtype=np.uint8), "y_scale": np.array([1], dtype=np.float32), "y_zero_point": np.array([0], np.uint8), } expected = sess1.run(None, feeds)[0] got = sess2.run(None, feeds)[0] assert_allclose(expected, got) with self.subTest(w="1x1", i=i, j=j): w = np.zeros((1, 1, 1, 1), dtype=np.uint8) w[0, 0, :, :] = 0 feeds = { "x": x, "x_scale": np.array([0.00369204697], dtype=np.float32), "x_zero_point": np.array([132], dtype=np.uint8), "w": w, "w_scale": np.array([100.001727945750], dtype=np.float32), "w_zero_point": np.array([255], dtype=np.uint8), "y_scale": np.array([0.00162681262], dtype=np.float32), "y_zero_point": np.array([132], np.uint8), } expected = sess1.run(None, feeds)[0] got = sess2.run(None, feeds)[0] assert_allclose(expected, got) x = np.array( [ [255, 174, 162, 25, 203, 168, 58], [15, 59, 237, 95, 129, 0, 64], [56, 242, 153, 221, 168, 12, 166], [232, 178, 186, 195, 237, 162, 237], [188, 39, 124, 77, 80, 102, 43], [127, 230, 21, 83, 41, 40, 134], [255, 154, 92, 141, 42, 148, 247], ], dtype=np.uint8, ).reshape((1, 1, 7, 7)) x_scale = np.array([0.00369204697], dtype=np.float32) x_zero_point = np.array([132], dtype=np.uint8) w = np.array([0], dtype=np.uint8).reshape((1, 1, 1, 1)) w_scale = np.array([0.00172794575], dtype=np.float32) w_zero_point = np.array([255], dtype=np.uint8) y_scale = np.array([0.00162681262], dtype=np.float32) y_zero_point = np.array([123], dtype=np.uint8) feeds = { "x": x, "x_scale": x_scale, "x_zero_point": x_zero_point, "w": w, "w_scale": w_scale, "w_zero_point": w_zero_point, "y_scale": y_scale, "y_zero_point": y_zero_point, } expected = sess1.run(None, feeds)[0] got = sess2.run(None, feeds)[0] assert_allclose(expected, got) def common_test_im2col(self, kernel_shape, pads, strides, dilations): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) Y1 = make_tensor_value_info("Y1", TensorProto.FLOAT, [None, None, None, None]) Y2 = make_tensor_value_info("Y2", TensorProto.FLOAT, [None, None, None, None]) W = make_tensor_value_info("W", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "Conv", ["X", "W"], ["Y1"], pads=pads, strides=strides, dilations=dilations ) node_shape = make_node("Shape", ["W"], ["shape"]) node_im = make_node( "Im2Col", ["X", "shape"], ["xim"], pads=pads, strides=strides, dilations=dilations, domain="experimental", ) node_flat = make_node("Flatten", ["W"], ["wflat"]) node_gem = make_node("MatMul", ["wflat", "xim"], ["Y2"]) graph = make_graph( [node, node_shape, node_im, node_flat, node_gem], "g", [X, W], [Y1, Y2], ) onnx_model = make_model( graph, opset_imports=[make_opsetid("", 16), make_opsetid("experimental", 1)] ) graph_conv = make_graph([node], "g", [X, W], [Y1]) onnx_model_conv = make_model_gen_version( graph_conv, opset_imports=[make_opsetid("", 16)] ) sess = ReferenceEvaluator(onnx_model) try: sess_conv = run_ort_inference(onnx_model_conv) if sess_conv is None: return except ImportError: sess_conv = None sH, sW = 7, 7 nker = np.prod(kernel_shape) for i in range(sH): for j in range(sW): X = np.zeros((1, 1, sH, sW), dtype=np.float32) X[0, 0, i, j] = 1.0 W = np.zeros( (1, 1, *kernel_shape), dtype=np.float32, ) W[0, 0, :, :] = np.minimum( 2 ** np.arange(nker).reshape((kernel_shape[0], -1)), 256 ) got = sess.run(None, {"X": X, "W": W}) if sess_conv is not None: ort_res = sess_conv.run(None, {"X": X, "W": W})[0] assert_allclose(got[1].ravel(), ort_res.ravel()) try: assert_allclose(got[0].ravel(), got[1].ravel()) except AssertionError as e: raise AssertionError( f"Discrepancies: pads={pads}, dilations={dilations}, strides={strides}, " f"kernel_shape={kernel_shape}" f"\n{got[0]}\n!=\n{got[1]}" ) from e def test_im2col_1x1(self): self.common_test_im2col( (1, 1), pads=[1, 1, 1, 2], strides=[1, 1], dilations=[1, 1] ) def test_im2col_2x2(self): self.common_test_im2col( (2, 2), pads=[1, 1, 1, 2], strides=[1, 1], dilations=[1, 1] ) def test_im2col_3x3(self): self.common_test_im2col( (3, 3), pads=[1, 1, 1, 2], strides=[1, 1], dilations=[1, 1] ) def test_im2col_3x3_pads(self): self.common_test_im2col( (3, 3), pads=[0, 1, 2, 3], strides=[1, 1], dilations=[1, 1] ) def test_im2col_3x3_strides(self): self.common_test_im2col( (3, 3), pads=[0, 1, 1, 1], strides=[1, 2], dilations=[1, 1] ) def test_im2col_5x5(self): self.common_test_im2col( (5, 5), pads=[1, 1, 1, 2], strides=[1, 1], dilations=[1, 1] ) @skip_if_no_torch def test_col2im(self): import torch X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None]) IS = make_tensor_value_info("I", TensorProto.INT64, [None]) BS = make_tensor_value_info("B", TensorProto.INT64, [None]) node = make_node( "Col2Im", ["X", "I", "B"], ["Y"], pads=[0, 0, 0, 0], strides=[1, 1], dilations=[1, 1], ) graph = make_graph([node], "g", [X, IS, BS], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) sess = ReferenceEvaluator(onnx_model) X = np.array( [ [ [1.0, 6.0, 11.0, 16.0, 21.0], [2.0, 7.0, 12.0, 17.0, 22.0], [3.0, 8.0, 13.0, 18.0, 23.0], [4.0, 9.0, 14.0, 19.0, 24.0], [5.0, 0.0, 15.0, 20.0, 25.0], ] ] ).astype(np.float32) image_shape = np.array([5, 5]).astype(np.int64) block_shape = np.array([1, 5]).astype(np.int64) fold = torch.nn.Fold(output_size=tuple(image_shape), kernel_size=block_shape) got = sess.run(None, {"X": X, "B": block_shape, "I": image_shape}) output = fold(torch.from_numpy(X)).numpy() assert_allclose(output, got[0]) def common_test_col2im( self, size, image_shape, block_shape, pads, strides, dilations ): import torch X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None]) IS = make_tensor_value_info("I", TensorProto.INT64, [None]) BS = make_tensor_value_info("B", TensorProto.INT64, [None]) node = make_node( "Col2Im", ["X", "I", "B"], ["Y"], pads=pads, strides=strides, dilations=dilations, ) graph = make_graph([node], "g", [X, IS, BS], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) sess = ReferenceEvaluator(onnx_model) fold = torch.nn.Fold( output_size=tuple(image_shape), kernel_size=tuple(block_shape), dilation=tuple(dilations), padding=min(pads), stride=tuple(strides), ) nker = np.prod(block_shape) for i in range(nker): for j in range(size): X = np.zeros((1, nker, size), dtype=np.float32) X[0, i, j] = 1.0 i_shape = np.array(image_shape, dtype=np.int64) b_shape = np.array(block_shape, dtype=np.int64) output = fold(torch.from_numpy(X)).numpy() got = sess.run(None, {"X": X, "B": b_shape, "I": i_shape}) # print(output) # print(got) assert_allclose(output, got[0]) @skip_if_no_torch def test_col2im_2x3(self): self.common_test_col2im( 10, (6, 4), (2, 3), pads=[0, 0, 0, 0], strides=[1, 1], dilations=[1, 1] ) @skip_if_no_torch def test_col2im_2x3_pads(self): self.common_test_col2im( 28, (6, 4), (2, 3), pads=[1, 1, 1, 1], strides=[1, 1], dilations=[1, 1] ) def test_col2im_2d(self): data = np.zeros([6, 28], dtype=np.float32) data[0][0] = 1.0 image_shape, kernel_shape, dilations, pads, stride = ( np.array([6, 4]), (2, 3), np.array([1, 1]), np.array([1, 1, 1, 1]), np.array([1, 1]), ) r1 = _col2im_naive_implementation_2d( data, image_shape, kernel_shape, dilations, pads, stride ) r2 = col2im_naive_implementation( data, image_shape, kernel_shape, dilations, pads, stride ) assert_allclose(r1, r2) def test_conv_im2col_group4(self): # model 1 X = make_tensor_value_info("X", TensorProto.FLOAT, [2, 4, 6, 6]) W = make_tensor_value_info("W", TensorProto.FLOAT, [4, 1, 3, 3]) B = make_tensor_value_info("B", TensorProto.FLOAT, [4]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [2, 4, 6, 6]) node = make_node( "Conv", ["X", "W", "B"], ["Y"], group=4, dilations=[1, 1], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1], ) graph = make_graph([node], "g", [X, W, B], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.arange(2 * 4 * 6 * 6).reshape((2, 4, 6, 6)).astype(np.float32), "W": np.array( [ [ [ [ -0.026239916682243347, 0.07565222680568695, -0.03209298849105835, ], [ -0.08708783239126205, 0.0961190015077591, 0.13418219983577728, ], [ 0.1598859578371048, 0.03840477764606476, -0.13170936703681946, ], ] ], [ [ [ -0.0689004510641098, 0.1408083587884903, -0.03717087209224701, ], [ 0.030967697501182556, 0.0263785719871521, -0.0899493545293808, ], [ 0.07828782498836517, -0.06266771256923676, 0.10750330984592438, ], ] ], [ [ [ 0.020227551460266113, -0.04353883117437363, -0.10938453674316406, ], [ -0.14101561903953552, -0.03393106162548065, 0.12139306962490082, ], [ 0.02838282287120819, 0.13864465057849884, -0.06065710633993149, ], ] ], [ [ [ -0.06511610746383667, -0.05987360328435898, -0.008047685027122498, ], [ 0.07340313494205475, 0.0326494425535202, 0.012516498565673828, ], [ 0.13260947167873383, -0.022225692868232727, -0.11167611926794052, ], ] ], ], dtype=np.float32, ), "B": np.array( [ -0.1457933485507965, -0.07481209933757782, -0.05890338122844696, -0.11964251846075058, ], dtype=np.float32, ), } feeds["B"][:] = 0 # model 2 X = feeds["X"] W = feeds["W"] B = feeds["B"] Y = np.empty((2, 4, 6, 6), dtype=X.dtype) for b in range(X.shape[0]): for g in range(4): x = X[b : b + 1, g : g + 1] w = W[g] c2 = im2col(x, (3, 3), [1, 1], [1, 1, 1, 1], [1, 1]) mul = np.matmul(c2, w.flatten()) mul = mul + B[g] Y[b, g, :, :] = mul ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) assert_allclose(Y, got1[0], atol=1e-5) def test_conv_strides(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [1, 3, 6, 6]) W = make_tensor_value_info("W", TensorProto.FLOAT, [2, 3, 3, 3]) B = make_tensor_value_info("B", TensorProto.FLOAT, [2]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "Conv", ["X", "W", "B"], ["Y"], group=1, dilations=[1, 1], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2], ) graph = make_graph([node], "g", [X, W, B], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.arange(1 * 3 * 6 * 6).reshape((1, 3, 6, 6)).astype(np.float32) + 1, "W": np.zeros((2, 3, 3, 3), dtype=np.float32), "B": np.zeros((2,), dtype=np.float32), } feeds["W"][0, 0, 0, 1] = 1 ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array( [ [ [[0.0, 0.0, 0.0], [7.0, 9.0, 11.0], [19.0, 21.0, 23.0]], [[0.0, 0.0, 0.0], [0.0, 0.0, 0.0], [0.0, 0.0, 0.0]], ] ], dtype=np.float32, ) assert_allclose(expected, got1[0]) def test_max_pool_2d_1(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "MaxPool", ["X"], ["Y"], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2], ) graph = make_graph([node], "g", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = {"X": np.arange(49)[::-1].reshape((1, 1, 7, 7)).astype(np.float32)} expected = np.array( [ [ [ [48.0, 47.0, 45.0, 43.0], [41.0, 40.0, 38.0, 36.0], [27.0, 26.0, 24.0, 22.0], [13.0, 12.0, 10.0, 8.0], ] ] ], dtype=np.float32, ) ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) assert_allclose(expected, got1[0]) def test_max_pool_2d_2(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "MaxPool", ["X"], ["Y"], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2], ) graph = make_graph([node], "g", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.array( [ [ [ [683, 358, 726, 578, 650, 946, 200], [679, 260, 264, 5, 240, 255, 582], [322, 66, 687, 632, 852, 698, 428], [111, 452, 627, 332, 751, 842, 685], [472, 52, 956, 81, 807, 827, 360], [972, 574, 81, 799, 646, 499, 486], [892, 758, 75, 833, 972, 415, 736], ] ] ], dtype=np.float32, ) } expected = np.array( [ [ [ [683.0, 726.0, 946.0, 946.0], [679.0, 687.0, 852.0, 842.0], [972.0, 956.0, 842.0, 842.0], [972.0, 833.0, 972.0, 736.0], ] ] ], dtype=np.float32, ) ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) assert_allclose(expected, got1[0]) def test_scatter_elements(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Ind = make_tensor_value_info("I", TensorProto.INT64, [None, None]) U = make_tensor_value_info("U", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None]) node = make_node( "ScatterElements", ["X", "I", "U"], ["Y"], axis=1, reduction="min", ) graph = make_graph([node], "g", [X, Ind, U], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.array([[1.0, 2.0, 3.0, 4.0, 5.0]], dtype=np.float32), "I": np.array([[1, 1]]), "U": np.array([[1.1, 2.1]], dtype=np.float32), } ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array([[1.0, 1.1, 3.0, 4.0, 5.0]], dtype=np.float32) assert_allclose(expected, got1[0]) def test_scatternd(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Ind = make_tensor_value_info("I", TensorProto.INT64, [None, None]) U = make_tensor_value_info("U", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None]) node = make_node( "ScatterND", ["X", "I", "U"], ["Y"], ) graph = make_graph([node], "g", [X, Ind, U], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.array([[1.0, 2.0]], dtype=np.float32), "I": np.array([[0, 0]]), "U": np.array([3.0], dtype=np.float32), } ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array([[3.0, 2.0]], dtype=np.float32) assert_allclose(expected, got1[0]) def test_col2im_impl(self): def get_im2col_indices( x_shape, field_height, field_width, padding=None, stride=1 ): # source: https://stackoverflow.com/questions/51703367/col2im-implementation-in-convnet N, C, H, W = x_shape del N # Unused assert (H + padding[0] + padding[2] - field_height) % stride == 0 assert (W + padding[1] + padding[3] - field_height) % stride == 0 out_height = (H + padding[0] + padding[2] - field_height) // stride + 1 out_width = (W + padding[1] + padding[3] - field_width) // stride + 1 i0 = np.repeat(np.arange(field_height), field_width) i0 = np.tile(i0, C) i1 = stride * np.repeat(np.arange(out_height), out_width) j0 = np.tile(np.arange(field_width), field_height * C) j1 = stride * np.tile(np.arange(out_width), out_height) i = i0.reshape(-1, 1) + i1.reshape(1, -1) j = j0.reshape(-1, 1) + j1.reshape(1, -1) k = np.repeat(np.arange(C), field_height * field_width).reshape(-1, 1) return (k, i, j) def col2im_indices( cols, x_shape, field_height=3, field_width=3, padding=None, stride=1 ): # source: https://stackoverflow.com/questions/51703367/col2im-implementation-in-convnet N, C, H, W = x_shape H_padded, W_padded = ( H + padding[0] + padding[2], W + padding[1] + padding[3], ) x_padded = np.zeros((N, C, H_padded, W_padded), dtype=cols.dtype) k, i, j = get_im2col_indices( x_shape, field_height, field_width, padding, stride ) cols_reshaped = cols.reshape(C * field_height * field_width, -1, N) cols_reshaped = cols_reshaped.transpose(2, 0, 1) np.add.at(x_padded, (slice(None), k, i, j), cols_reshaped) padding = padding.copy() if padding[2] == 0: padding[2] += x_padded.shape[2] elif padding[2] > 0: padding[2] *= -1 if padding[3] == 0: padding[3] += x_padded.shape[3] elif padding[3] > 0: padding[3] *= -1 res = x_padded[:, :, padding[0] : padding[2], padding[1] : padding[3]] return res X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None]) IS = make_tensor_value_info("IS", TensorProto.INT64, [None]) BS = make_tensor_value_info("BS", TensorProto.INT64, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node("Col2Im", ["X", "IS", "BS"], ["Y"], pads=[0, 1, 0, 1]) graph = make_graph([node], "g", [X, IS, BS], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.arange(5 * 15).astype(np.float32).reshape((1, 5, 15)), "IS": np.array([5, 5]), "BS": np.array([1, 5]), } ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = col2im_indices( feeds["X"], (1, 1, 5, 5), field_height=1, field_width=5, padding=[0, 1, 0, 1], ) assert_allclose(expected, got1[0]) def test_conv_transpose_2d(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) W = make_tensor_value_info("W", TensorProto.FLOAT, [None, None, None, None]) B = make_tensor_value_info("B", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "ConvTranspose", ["X", "W", "B"], ["Y"], dilations=[1, 1], kernel_shape=[3, 3], output_padding=[0, 0], pads=[1, 1, 1, 1], strides=[1, 1], ) graph = make_graph([node], "g", [X, W, B], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.arange(1 * 3 * 5 * 4).reshape((1, 3, 5, 4)).astype(np.float32), "W": np.arange(3 * 1 * 3 * 3).reshape((3, 1, 3, 3)).astype(np.float32), "B": np.array([0, 0, 0, 0], dtype=np.float32), } # import torch # ex = torch.nn.functional.conv_transpose2d( # torch.Tensor(feeds["X"]), torch.Tensor(feeds["W"]), # bias=None, stride=1, padding=1, output_padding=0, groups=1, dilation=1) # print(ex) ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array( [ [ [ [4371, 6855, 7062, 4929], [7524, 11781, 12132, 8451], [8424, 13185, 13536, 9423], [9324, 14589, 14940, 10395], [7197, 11229, 11490, 7971], ], ] ], dtype=np.float32, ) assert_allclose(expected, got1[0]) feeds["X"] *= 0 feeds["X"][0, 0, 0, 0] = 1 ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array( [ [ [ [4, 5, 0, 0], [7, 8, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], [0, 0, 0, 0], ] ] ], dtype=np.float32, ) assert_allclose(expected, got1[0]) def test_conv_transpose_2d_upper(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) W = make_tensor_value_info("W", TensorProto.FLOAT, [None, None, None, None]) B = make_tensor_value_info("B", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "ConvTranspose", ["X", "W", "B"], ["Y"], auto_pad="SAME_UPPER", strides=[2, 2], # output_shape=[6, 6], ) graph = make_graph([node], "g", [X, W, B], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 16)]) feeds = { "X": np.arange(1 * 1 * 3 * 3).reshape((1, 1, 3, 3)).astype(np.float32), "W": np.arange(1 * 2 * 3 * 3).reshape((1, 2, 3, 3)).astype(np.float32), "B": np.array([0, 0, 0, 0], dtype=np.float32), } expected = np.array( [ [ [ [0, 0, 0, 1, 2, 2], [0, 0, 3, 4, 11, 8], [0, 3, 12, 11, 28, 19], [9, 12, 27, 16, 35, 20], [18, 27, 60, 35, 76, 43], [18, 24, 51, 28, 59, 32], ], [ [0, 0, 9, 10, 29, 20], [0, 0, 12, 13, 38, 26], [27, 30, 84, 56, 136, 82], [36, 39, 90, 52, 116, 65], [99, 108, 240, 134, 292, 160], [72, 78, 168, 91, 194, 104], ], ] ], dtype=np.float32, ) # import onnxruntime # ref0 = onnxruntime.InferenceSession(onnx_model.SerializeToString(), providers=["CPUExecutionProvider"]) # got0 = ref0.run(None, feeds) ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) assert_allclose(expected, got1[0]) def test_stft(self): signal = make_tensor_value_info("signal", TensorProto.FLOAT, [None, None, None]) frame_step = make_tensor_value_info("frame_step", TensorProto.INT64, [None]) frame_length = make_tensor_value_info("frame_length", TensorProto.INT64, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "STFT", ["signal", "frame_step", "", "frame_length"], ["Y"], ) graph = make_graph([node], "g", [signal, frame_step, frame_length], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 17)]) feeds = { "signal": np.arange(128).reshape((1, 128, 1)).astype(np.float32), "frame_step": np.array(8, dtype=np.int64), "frame_length": np.array(16, dtype=np.int64), } signal = feeds["signal"] frame_length = int(feeds["frame_length"]) frame_step = int(feeds["frame_step"]) onesided_length = (frame_length // 2) + 1 nstfts = ((feeds["signal"].shape[1] - frame_length) // frame_step) + 1 # [batch_size][frames][frame_length][2] expected = np.empty([1, nstfts, onesided_length, 2], dtype=np.float32) for i in range(nstfts): start = i * frame_step stop = i * frame_step + frame_length complex_out = np.fft.fft(signal[0, start:stop, 0]) c_out = complex_out[0:onesided_length] expected[0, i] = np.stack((c_out.real, c_out.imag), axis=1) # import torch # correspondance with torch # hop_length = frame_step # window = np.ones((frame_length,), dtype=np.float32) # ex = torch.stft( # torch.Tensor(feeds["signal"][:, :, 0]), # n_fft=frame_length, window=torch.Tensor(window), # hop_length=hop_length, win_length=frame_length, # onesided=True, return_complex=True, center=False, # normalized=False) # ex = np.transpose(ex.numpy(), [0, 2, 1]) ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) assert_allclose(expected, got1[0]) def test_stft_with_window(self): signal = make_tensor_value_info("signal", TensorProto.FLOAT, [None, None, None]) frame_step = make_tensor_value_info("frame_step", TensorProto.INT64, [None]) window = make_tensor_value_info("window", TensorProto.FLOAT, [None]) frame_length = make_tensor_value_info("frame_length", TensorProto.INT64, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) node = make_node( "STFT", ["signal", "frame_step", "window", "frame_length"], ["Y"], ) graph = make_graph([node], "g", [signal, frame_step, window, frame_length], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 17)]) feeds = { "signal": np.arange(128).reshape((1, 128, 1)).astype(np.float32), "frame_step": np.array(8, dtype=np.int64), "window": 0.5 + 0.5 * np.cos(2 * np.pi * np.arange(0, 16, 1, dtype=np.float32) / 16), "frame_length": np.array(16, dtype=np.int64), } signal = feeds["signal"] frame_length = int(feeds["frame_length"]) window = feeds["window"] frame_step = int(feeds["frame_step"]) onesided_length = (frame_length // 2) + 1 nstfts = 1 + (signal.shape[1] - window.shape[0]) // 8 # [batch_size][frames][frame_length][2] expected = np.empty([1, nstfts, onesided_length, 2], dtype=np.float32) for i in range(nstfts): start = i * frame_step stop = i * frame_step + frame_length complex_out = np.fft.fft(signal[0, start:stop, 0] * window)[ 0:onesided_length ] c_out = complex_out[0:onesided_length] expected[0, i] = np.stack((c_out.real, c_out.imag), axis=1) # import torch # hop_length = frame_step # ex = torch.stft( # torch.Tensor(feeds["signal"][:, :, 0]), # n_fft=frame_length, window=torch.Tensor(window), # hop_length=hop_length, win_length=frame_length, # onesided=True, return_complex=True, center=False, # normalized=False) # ex = np.transpose(ex.numpy(), [0, 2, 1]) ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) assert_allclose(expected, got1[0]) def get_roi_align_model(self, mode): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) rois = make_tensor_value_info("rois", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) IS = make_tensor_value_info("I", TensorProto.INT64, [None]) node = make_node( "RoiAlign", ["X", "rois", "I"], ["Y"], output_height=5, output_width=5, sampling_ratio=2, spatial_scale=1.0, coordinate_transformation_mode="output_half_pixel", mode=mode, ) graph = make_graph([node], "g", [X, rois, IS], [Y]) return make_model_gen_version(graph, opset_imports=[make_opsetid("", 17)]) def common_test_roi_align(self, mode): onnx_model = self.get_roi_align_model(mode) X, batch_indices, rois = get_roi_align_input_values() feeds = {"X": X, "rois": rois, "I": batch_indices} sess = run_ort_inference(onnx_model) if sess is None: return expected = sess.run(None, feeds) ref = ReferenceEvaluator(onnx_model) got = ref.run(None, feeds) assert_allclose(expected[0], got[0], atol=1e-5) @skip_if_no_onnxruntime def test_roi_align(self): with self.subTest(mode="avg"): self.common_test_roi_align("avg") # max does not have example in the backend with self.subTest(mode="max"): self.common_test_roi_align("max") def common_test_roi_align_torch(self, mode): import torch from torchvision.ops import RoIAlign onnx_model = self.get_roi_align_model(mode) sess = ReferenceEvaluator(onnx_model) X, batch_indices, rois = get_roi_align_input_values() got = sess.run(None, {"X": X, "rois": rois, "I": batch_indices}) a = RoIAlign((5, 5), spatial_scale=1.0, sampling_ratio=2) expected = a(torch.from_numpy(X), [torch.from_numpy(rois)]) assert_allclose(expected, got[0], atol=1e-5) @skip_if_no_torch @skip_if_no_torchvision def test_roi_align_torch(self): with self.subTest(mode="avg"): self.common_test_roi_align_torch("avg") # not implemented in torch # with self.subTest(mode="max"): # self.common_test_roi_align_torch("max") def test_split(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y1 = make_tensor_value_info("Y1", TensorProto.FLOAT, [None]) Y2 = make_tensor_value_info("Y2", TensorProto.FLOAT, [None]) Y3 = make_tensor_value_info("Y3", TensorProto.FLOAT, [None]) Y4 = make_tensor_value_info("Y4", TensorProto.FLOAT, [None]) node = make_node("Split", ["X"], ["Y1", "Y2", "Y3", "Y4"], num_outputs=4) graph = make_graph([node], "g", [X], [Y1, Y2, Y3, Y4]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) feeds = {"X": np.arange(10).astype(np.float32)} expected = [ np.array([0, 1, 2], dtype=np.float32), np.array([3, 4, 5], dtype=np.float32), np.array([6, 7, 8], dtype=np.float32), np.array([9], dtype=np.float32), ] ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) for i in range(4): assert_allclose(expected[i], got1[i]) def test_split_2(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y1 = make_tensor_value_info("Y1", TensorProto.FLOAT, [None]) Y2 = make_tensor_value_info("Y2", TensorProto.FLOAT, [None]) Y3 = make_tensor_value_info("Y3", TensorProto.FLOAT, [None]) Y4 = make_tensor_value_info("Y4", TensorProto.FLOAT, [None]) node = make_node("Split", ["X", "split"], ["Y1", "Y2", "Y3", "Y4"]) graph = make_graph([node], "g", [X], [Y1, Y2, Y3, Y4]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) feeds = { "X": np.arange(10).astype(np.float32), "split": np.array([3, 3, 2, 2], dtype=np.int64), } expected = [ np.array([0, 1, 2], dtype=np.float32), np.array([3, 4, 5], dtype=np.float32), np.array([6, 7], dtype=np.float32), np.array([8, 9], dtype=np.float32), ] ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) for i in range(4): assert_allclose(expected[i], got1[i]) def test_split_num_outputs_4(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y1 = make_tensor_value_info("Y1", TensorProto.FLOAT, [None]) Y2 = make_tensor_value_info("Y2", TensorProto.FLOAT, [None]) Y3 = make_tensor_value_info("Y3", TensorProto.FLOAT, [None]) Y4 = make_tensor_value_info("Y4", TensorProto.FLOAT, [None]) node = make_node("Split", ["X"], ["Y1", "Y2", "Y3", "Y4"], num_outputs=4) graph = make_graph([node], "g", [X], [Y1, Y2, Y3, Y4]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) # case 1 feeds = {"X": np.arange(10).astype(np.float32)} expected = [ np.array([0, 1, 2], dtype=np.float32), np.array([3, 4, 5], dtype=np.float32), np.array([6, 7, 8], dtype=np.float32), np.array([9], dtype=np.float32), ] ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) for i in range(4): assert_allclose(expected[i], got1[i]) # case 2 feeds = {"X": np.arange(9).astype(np.float32)} expected = [ np.array([0, 1, 2], dtype=np.float32), np.array([3, 4, 5], dtype=np.float32), np.array([6, 7, 8], dtype=np.float32), np.array([], dtype=np.float32), ] ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) for i in range(4): assert_allclose(expected[i], got1[i]) def test_argmin(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.INT64, [None]) node = make_node("ArgMin", ["X"], ["Y"], axis=1) graph = make_graph([node], "g", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) feeds = {"X": np.arange(12).reshape((3, 4)).astype(np.float32)} ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array([0, 0, 0], dtype=np.int64).reshape((-1, 1)) self.assertEqual(expected.tolist(), got1[0].tolist()) def test_argmax(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.INT64, [None]) node = make_node("ArgMax", ["X"], ["Y"], axis=1) graph = make_graph([node], "g", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) feeds = {"X": np.arange(12).reshape((3, 4)).astype(np.float32)} ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array([3, 3, 3], dtype=np.int64).reshape((-1, 1)) self.assertEqual(expected.tolist(), got1[0].tolist()) def test_slice_squeeze(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) starts = make_tensor_value_info("starts", TensorProto.INT64, [None]) ends = make_tensor_value_info("ends", TensorProto.INT64, [None]) axes = make_tensor_value_info("axes", TensorProto.INT64, [None]) Y = make_tensor_value_info("Y", TensorProto.INT64, [None]) nodes = [ make_node("Slice", ["X", "starts", "ends", "axes"], ["T"]), make_node("Squeeze", ["T", "axes"], ["Y"]), ] graph = make_graph(nodes, "g", [X, starts, ends, axes], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) feeds = { "X": np.array([[0]], dtype=np.int64), "starts": np.array([0], dtype=np.int64), "ends": np.array([1], dtype=np.int64), "axes": np.array([0], dtype=np.int64), } ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array([0], dtype=np.int64) self.assertEqual(expected.tolist(), got1[0].tolist()) def test_slice_squeeze_6(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.INT64, [None]) nodes = [ make_node("Slice", ["X"], ["T"], axes=[0], starts=[0], ends=[1]), make_node("Squeeze", ["T"], ["Y"], axes=[0]), ] graph = make_graph(nodes, "g", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 6)]) feeds = {"X": np.array([[0]], dtype=np.int64)} ref1 = ReferenceEvaluator(onnx_model) got1 = ref1.run(None, feeds) expected = np.array([0], dtype=np.int64) self.assertEqual(expected.tolist(), got1[0].tolist()) def test_onnxrt_reduce_mean(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None]) node1 = make_node("ReduceMean", ["X"], ["Y"]) graph = make_graph([node1], "g", [X], [Y]) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 17)]) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) cls = sess.rt_nodes_[0] self.assertEqual(cls.__class__.__name__, "ReduceMean_1") got = sess.run(None, {"X": np.ones((2, 4), dtype=np.float32)})[0] self.assertEqual(got.shape, (1, 1)) self.assertEqual(got[0, 0], 1) onnx_model = make_model(graph, opset_imports=[make_opsetid("", 18)]) check_model(onnx_model) sess = ReferenceEvaluator(onnx_model) cls = sess.rt_nodes_[0] self.assertEqual(cls.__class__.__name__, "ReduceMean_18") got = sess.run(None, {"X": np.ones((2, 4), dtype=np.float32)})[0] self.assertEqual(got.shape, (1, 1)) self.assertEqual(got[0, 0], 1) @staticmethod def _cdist_model(opset, reduce_op="ReduceSumSquare"): # subgraph initializers = [] inputs = [ make_tensor_value_info("next_in", TensorProto.FLOAT, [None, 4]), make_tensor_value_info("next", TensorProto.FLOAT, [None]), ] outputs = [ make_tensor_value_info("next_out", TensorProto.FLOAT, [None, None]), make_tensor_value_info("scan_out", TensorProto.FLOAT, [None]), ] if opset >= 18: initializers.append( from_array(np.array([1], dtype=np.int64), name="axis_red") ) node_reduce = make_node( reduce_op, ["cdistdf_17_C0", "axis_red"], ["cdistdf_17_reduced0"], name="cdistdf_17_ReduceSumSquare", keepdims=0, ) else: node_reduce = make_node( reduce_op, ["cdistdf_17_C0"], ["cdistdf_17_reduced0"], name="cdistdf_17_ReduceSumSquare", axes=[1], keepdims=0, ) nodes = [ make_node("Identity", ["next_in"], ["next_out"], name="cdistd_17_Identity"), make_node( "Sub", ["next_in", "next"], ["cdistdf_17_C0"], name="cdistdf_17_Sub" ), node_reduce, make_node( "Identity", ["cdistdf_17_reduced0"], ["scan_out"], name="cdistdf_17_Identity", ), ] graph = make_graph(nodes, "OnnxIdentity", inputs, outputs, initializers) # main graph initializers = [] list_value = [ 1.1394007205963135, -0.6848101019859314, -1.234825849533081, 0.4023416340351105, 0.17742614448070526, 0.46278226375579834, -0.4017809331417084, -1.630198359489441, -0.5096521973609924, 0.7774903774261475, -0.4380742907524109, -1.2527953386306763, -1.0485529899597168, 1.950775384902954, -1.420017957687378, -1.7062702178955078, 1.8675580024719238, -0.15135720372200012, -0.9772778749465942, 0.9500884413719177, -2.5529897212982178, -0.7421650290489197, 0.653618574142456, 0.8644362092018127, 1.5327792167663574, 0.37816253304481506, 1.4693588018417358, 0.154947429895401, -0.6724604368209839, -1.7262825965881348, -0.35955315828323364, -0.8131462931632996, -0.8707971572875977, 0.056165341287851334, -0.5788496732711792, -0.3115525245666504, 1.2302906513214111, -0.302302747964859, 1.202379822731018, -0.38732680678367615, 2.269754648208618, -0.18718385696411133, -1.4543657302856445, 0.04575851559638977, -0.9072983860969543, 0.12898291647434235, 0.05194539576768875, 0.7290905714035034, 1.4940791130065918, -0.8540957570075989, -0.2051582634449005, 0.3130677044391632, 1.764052391052246, 2.2408931255340576, 0.40015721321105957, 0.978738009929657, 0.06651721894741058, -0.3627411723136902, 0.30247190594673157, -0.6343221068382263, -0.5108051300048828, 0.4283318817615509, -1.18063223361969, -0.02818222902715206, -1.6138978004455566, 0.38690251111984253, -0.21274028718471527, -0.8954665660858154, 0.7610377073287964, 0.3336743414402008, 0.12167501449584961, 0.44386324286460876, -0.10321885347366333, 1.4542734622955322, 0.4105985164642334, 0.14404356479644775, -0.8877857327461243, 0.15634897351264954, -1.980796456336975, -0.34791216254234314, ] initializers.append( from_array( np.array(list_value, dtype=np.float32).reshape((20, 4)), name="Sc_Scancst", ) ) initializers.append( from_array(np.array([2], dtype=np.int64), name="To_TopKcst") ) inputs = [make_tensor_value_info("input", TensorProto.FLOAT, [None, 4])] outputs = [ make_tensor_value_info("values", TensorProto.FLOAT, [None, 2]), make_tensor_value_info("indices", TensorProto.INT64, [None, 2]), ] # nodes nodes = [ make_node( "Scan", ["input", "Sc_Scancst"], ["UU032UU", "UU033UU"], name="Sc_Scan", body=graph, num_scan_inputs=1, ), make_node( "Transpose", ["UU033UU"], ["Tr_transposed0"], name="Tr_Transpose", perm=[1, 0], ), make_node("Sqrt", ["Tr_transposed0"], ["Sq_Y0"], name="Sq_Sqrt"), make_node( "TopK", ["Sq_Y0", "To_TopKcst"], ["values", "indices"], name="To_TopK", largest=0, sorted=1, ), ] graph = make_graph(nodes, "dummy", inputs, outputs, initializers) # model onnx_model = make_model(graph, opset_imports=[make_opsetid("", opset)]) return onnx_model @parameterized.parameterized.expand( itertools.product( [ ( "ReduceMin", [ np.array( [[np.nan, np.nan], [14.422706, 18.80527]], dtype=np.float32 ), np.array([[2, 15], [10, 4]], dtype=np.int64), ], ), ( "ReduceL1", [ np.array( [[2.2367053, 2.3516612], [4.076292, 4.2970634]], dtype=np.float32, ), np.array([[18, 6], [13, 6]], dtype=np.int64), ], ), ( "ReduceL2", [ np.array( [[1.80155, 1.8169948], [2.9928076, 3.1205883]], dtype=np.float32, ), np.array([[11, 18], [13, 6]], dtype=np.int64), ], ), ( "ReduceLogSum", [ np.array( [[0.9497848, 1.1872643], [1.6764175, 1.70759]], dtype=np.float32, ), np.array([[6, 18], [13, 6]], dtype=np.int64), ], ), ( "ReduceLogSumExp", [ np.array( [[1.6005973, 1.7445935], [2.5616229, 2.6539795]], dtype=np.float32, ), np.array([[13, 6], [13, 6]], dtype=np.int64), ], ), ( "ReduceMax", [ np.array( [[1.4217108, 1.5069536], [2.453826, 2.5041783]], dtype=np.float32, ), np.array([[13, 11], [13, 11]], dtype=np.int64), ], ), ( "ReduceMean", [ np.array( [[0.39247903, 0.78497636], [2.038146, 2.1485317]], dtype=np.float32, ), np.array([[13, 6], [13, 6]], dtype=np.int64), ], ), ( "ReduceSumSquare", [ np.array( [[3.2455828, 3.3014696], [8.956896, 9.7380705]], dtype=np.float32, ), np.array([[11, 18], [13, 6]], dtype=np.int64), ], ), ( "ReduceProd", [ np.array( [[np.nan, np.nan], [14.422706, 18.80527]], dtype=np.float32 ), np.array([[2, 15], [13, 6]], dtype=np.int64), ], ), ], [17, 18], ) ) def test_op_reduce(self, reduce_op_expected, opset: int): reduce_op, expected = reduce_op_expected X = np.arange(8).reshape((-1, 4)).astype(np.float32) results = {} model = self._cdist_model(opset, reduce_op) sess = ReferenceEvaluator(model) got = sess.run(None, {"input": X}) results["ref", opset] = got cl = [ n for n in sess.rt_nodes_[0].body.rt_nodes_ if n.__class__.__name__.startswith(reduce_op) ] schema = cl[0]._schema # pylint: disable=protected-access new_cl = type(reduce_op, (cl[0].__class__,), {"op_schema": schema}) sess = ReferenceEvaluator(model, new_ops=[new_cl]) got = sess.run(None, {"input": X}) results["ref_cl", opset] = got baseline = "constant" for k, v in results.items(): for a, b in zip(reversed(expected), reversed(v)): if a.shape != b.shape: raise AssertionError( f"Shape mismatch for {reduce_op!r}, {baseline}:{a.shape} != {k}:{b.shape}." ) diff = np.abs(a - b).max() if diff > 1e-6: raise AssertionError( f"Discrepancies (max={diff}) for {reduce_op!r}, {baseline} != {k}\n{a}\n!=\n{b}" ) @parameterized.parameterized.expand( [ (13,), (17,), (18,), ] ) def test_mvn(self, opset: int, ref_opset: int = 13): X = make_tensor_value_info("X", TensorProto.FLOAT, [None, None, None, None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None, None, None, None]) nodes = [ make_node("MeanVarianceNormalization", ["X"], ["Y"]), ] graph = make_graph(nodes, "g", [X], [Y]) x = np.random.rand(3, 3, 3, 1).astype(np.float32) onnx_model = make_model(graph, opset_imports=[make_opsetid("", opset)]) ref = ReferenceEvaluator(onnx_model) got = ref.run(None, {"X": x})[0] ref_onnx_model = make_model(graph, opset_imports=[make_opsetid("", ref_opset)]) ref_expected = ReferenceEvaluator(ref_onnx_model) expected = ref_expected.run(None, {"X": x})[0] self.assertEqual(expected.shape, got.shape) assert_allclose(expected, got) def test_concat_in_a_function(self): def create_model(): nodes = [] inputs = [] outputs = [] functions = [] opsets = {"": onnx_opset_version(), "custom_domain": 1} nodes_fct = [] node = make_node("Concat", ["x:0", "x:1"], ["r__0"], axis=0, domain="") nodes_fct.append(node) opset_imports_fct = [ make_opsetid(domain, 1 if version is None else version) for domain, version in opsets.items() ] fct = make_function( "custom_domain", "concat_2", ["x:0", "x:1"], ["r__0"], nodes_fct, opset_imports_fct, ) functions.append(fct) inputs.append(make_tensor_value_info("I__0", TensorProto.DOUBLE, [])) inputs.append(make_tensor_value_info("I__1", TensorProto.DOUBLE, [])) inputs.append(make_tensor_value_info("I__2", TensorProto.DOUBLE, [])) outputs.append(make_tensor_value_info("r__4", TensorProto.DOUBLE, [])) node = make_node( "concat_2", ["I__0", "I__1"], ["r__3"], axis=0, domain="custom_domain" ) nodes.append(node) node = make_node( "concat_2", ["I__2", "r__3"], ["r__4"], axis=0, domain="custom_domain" ) nodes.append(node) opset_imports = [ make_opsetid(domain, 1 if version is None else version) for domain, version in opsets.items() ] graph = make_graph(nodes, "numpyx", inputs, outputs) onnx_model = make_model( graph, opset_imports=opset_imports, functions=functions ) return onnx_model onnx_model = create_model() x1 = np.array([[-5, 6], [15, 3]], dtype=np.float64) x2 = np.array([[1, 2]], dtype=np.float64) x3 = np.array([[-1, -2]], dtype=np.float64) z = np.vstack([x1, x2, x3]) ref = ReferenceEvaluator(onnx_model) feeds = {"I__2": x1, "I__0": x2, "I__1": x3} got = ref.run(None, feeds) assert_allclose(z, got[0]) def test_cast_float_to_string(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.STRING, [None]) model = make_model( make_graph( [ make_node("Cast", ["X"], ["Y"], to=TensorProto.STRING), ], "g", [X], [Y], ) ) ref = ReferenceEvaluator(model) data = np.array([1.152512, -0.152612, 0.0, np.nan]) got = ref.run(None, {"X": data})[0] self.assertTrue( (got == np.array([1.152512, -0.152612, 0.0, np.nan]).astype(np.str_)).all() ) def test_cast_float_to_string_and_back(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) model = make_model( make_graph( [ make_node("Cast", ["X"], ["Z"], to=TensorProto.STRING), make_node("Cast", ["Z"], ["Y"], to=TensorProto.FLOAT), ], "g", [X], [Y], ) ) ref = ReferenceEvaluator(model) data = np.array([1.152512, -0.152612, 0.0, np.nan]) got = ref.run(None, {"X": data})[0] assert_allclose(got, np.array([1.152512, -0.152612, 0.0, np.nan])) def test_split_to_sequence(self): X = make_tensor_value_info("X", TensorProto.FLOAT, None) Y = make_tensor_value_info("Y", TensorProto.INT64, None) Z = make_tensor_value_info("Z", TensorProto.UNDEFINED, None) nodes = [make_node("SplitToSequence", ["X", "Y"], ["Z"], axis=2)] model = make_model(make_graph(nodes, "g", [X, Y], [Z])) ref = ReferenceEvaluator(model) data = np.arange(18).reshape((1, 3, 6)).astype(np.float32) indices = np.array(2, dtype=np.int64) got = ref.run(None, {"X": data, "Y": indices}) expected = [ [ np.array([[[0.0, 1.0], [6.0, 7.0], [12.0, 13.0]]], dtype=np.float32), np.array([[[2.0, 3.0], [8.0, 9.0], [14.0, 15.0]]], dtype=np.float32), np.array([[[4.0, 5.0], [10.0, 11.0], [16.0, 17.0]]], dtype=np.float32), ] ] self.assertEqual(len(expected[0]), len(got[0])) for a, b in zip(expected[0], got[0]): assert_allclose(a, b) def test_split_to_sequence_1d(self): X = make_tensor_value_info("X", TensorProto.FLOAT, None) Y = make_tensor_value_info("Y", TensorProto.INT64, None) Z = make_tensor_value_info("Z", TensorProto.UNDEFINED, None) nodes = [make_node("SplitToSequence", ["X", "Y"], ["Z"], axis=2)] model = make_model(make_graph(nodes, "g", [X, Y], [Z])) ref = ReferenceEvaluator(model) data = np.arange(18).reshape((1, 3, 6)).astype(np.float32) indices = np.array([2, 2, 2], dtype=np.int64) got = ref.run(None, {"X": data, "Y": indices}) expected = [ [ np.array([[[0.0, 1.0], [6.0, 7.0], [12.0, 13.0]]], dtype=np.float32), np.array([[[2.0, 3.0], [8.0, 9.0], [14.0, 15.0]]], dtype=np.float32), np.array([[[4.0, 5.0], [10.0, 11.0], [16.0, 17.0]]], dtype=np.float32), ] ] self.assertEqual(len(expected[0]), len(got[0])) for a, b in zip(expected[0], got[0]): assert_allclose(a, b) def test_split_to_sequence_nokeepdims_noinput(self): # keepdims is ignored in that case X = make_tensor_value_info("X", TensorProto.FLOAT, None) Z = make_tensor_value_info("Z", TensorProto.UNDEFINED, None) nodes = [make_node("SplitToSequence", ["X"], ["Z"], axis=2, keepdims=0)] model = make_model(make_graph(nodes, "g", [X], [Z])) ref = ReferenceEvaluator(model) data = np.arange(18).reshape((1, 3, 6)).astype(np.float32) got = ref.run(None, {"X": data}) expected = [[data[:, :, i] for i in range(data.shape[2])]] self.assertEqual(len(expected[0]), len(got[0])) for a, b in zip(expected[0], got[0]): assert_allclose(a, b) def test_cast_float8(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) F1 = make_tensor_value_info("F1", TensorProto.FLOAT, [None]) F2 = make_tensor_value_info("F2", TensorProto.FLOAT, [None]) F3 = make_tensor_value_info("F3", TensorProto.FLOAT, [None]) F4 = make_tensor_value_info("F4", TensorProto.FLOAT, [None]) model = make_model( make_graph( [ make_node("Cast", ["X"], ["f81"], to=TensorProto.FLOAT8E4M3FN), make_node("Cast", ["X"], ["f82"], to=TensorProto.FLOAT8E5M2), make_node( "Constant", [], ["C1"], value=make_tensor( "C1", TensorProto.FLOAT8E4M3FN, [5], [0, 1, 2, 5e-2, 200] ), ), make_node( "Constant", [], ["C2"], value=make_tensor( "C2", TensorProto.FLOAT8E5M2, [5], [0, 1, 2, 5e-2, 200] ), ), make_node("Cast", ["f81"], ["F1"], to=TensorProto.FLOAT), make_node("Cast", ["f82"], ["F2"], to=TensorProto.FLOAT), make_node("Cast", ["C1"], ["F3"], to=TensorProto.FLOAT), make_node("Cast", ["C2"], ["F4"], to=TensorProto.FLOAT), ], "g", [X], [F1, F2, F3, F4], ) ) ref = ReferenceEvaluator(model) data = np.array([0, 1, 2, 5e-2, 200], dtype=np.float32) expected1 = np.array( [float8e4m3_to_float32(float32_to_float8e4m3(x)) for x in data] ) expected2 = np.array( [float8e5m2_to_float32(float32_to_float8e5m2(x)) for x in data] ) got = ref.run(None, {"X": data}) assert_allclose(got[0], expected1) assert_allclose(got[1], expected2) assert_allclose(got[2], expected1) assert_allclose(got[3], expected2) def test_cast_like_float8(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) model = make_model( make_graph( [ make_node("Cast", ["X"], ["f8"], to=TensorProto.FLOAT8E4M3FNUZ), make_node("CastLike", ["X", "f8"], ["f32"], saturate=0), make_node("Cast", ["f32"], ["Y"], to=TensorProto.FLOAT), ], "g", [X], [Y], ) ) data = np.array([0, 1e7], dtype=np.float32) expected = np.array( [ float8e4m3_to_float32( float32_to_float8e4m3(x, uz=True, saturate=False), uz=True ) for x in data ] ) ref = ReferenceEvaluator(model) got = ref.run(None, {"X": data}) assert_allclose(got[0], expected) # Forces ReferenceEvaluator to not use the associated implementation for CastLike # but its implementation as a function instead. class CastLike(OpRunExpand): op_domain = "" ref = ReferenceEvaluator(model, new_ops=[CastLike]) got = ref.run(None, {"X": data}) assert_allclose(got[0], expected) def test_cast_float8_output(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) F1 = make_tensor_value_info("F1", TensorProto.FLOAT8E4M3FN, [None]) F2 = make_tensor_value_info("F2", TensorProto.FLOAT8E5M2, [None]) model = make_model( make_graph( [ make_node("Cast", ["X"], ["F1"], to=TensorProto.FLOAT8E4M3FN), make_node("Cast", ["X"], ["F2"], to=TensorProto.FLOAT8E5M2), ], "g", [X], [F1, F2], ) ) ref = ReferenceEvaluator(model) data = np.array([0, 1, 2, 5e-2, 200], dtype=np.float32) expected1 = np.array([float32_to_float8e4m3(x) for x in data]) expected2 = np.array([float32_to_float8e5m2(x) for x in data]) got = ref.run(None, {"X": data}) self.assertEqual(expected1.tolist(), got[0].tolist()) self.assertEqual(expected2.tolist(), got[1].tolist()) def test_float8_4_types(self): x = np.array( [ 0.4068359375, 352, 416, 336, 304, 272, -248, -100, 1e-4, 1e-2, 416, 432, 1e5, np.inf, -np.inf, np.nan, ], dtype=np.float32, ) expected = { TensorProto.FLOAT8E4M3FN: np.array( [ 0.40625, 352.0, 416.0, 320.0, 320.0, 256.0, -256.0, -96.0, 0.0, 0.009765625, 416.0, 448.0, 448.0, 448.0, -448.0, np.nan, ], dtype=np.float32, ), TensorProto.FLOAT8E4M3FNUZ: np.array( [ 0.40625, 240.0, 240.0, 240.0, 240.0, 240.0, -240.0, -104.0, 0.0, 0.009765625, 240.0, 240.0, 240.0, 240.0, -240.0, np.nan, ], dtype=np.float32, ), TensorProto.FLOAT8E5M2: np.array( [ 0.4375, 384.0, 384.0, 320.0, 320.0, 256.0, -256.0, -96.0, 0.0001068115234375, 0.009765625, 384.0, 448.0, 57344.0, 57344.0, -57344.0, np.nan, ], dtype=np.float32, ), TensorProto.FLOAT8E5M2FNUZ: np.array( [ 4.3750000e-01, 3.8400000e02, 4.4800000e02, 3.2000000e02, 3.2000000e02, 2.5600000e02, -2.5600000e02, -9.6000000e01, 1.0681152e-04, 9.7656250e-03, 4.4800000e02, 4.4800000e02, 5.7344000e04, 5.7344000e04, -5.7344000e04, np.nan, ], dtype=np.float32, ), } def model_cast_cast(to): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) node1 = make_node("Cast", ["X"], ["T"], to=to) node2 = make_node("Cast", ["T"], ["Y"], to=TensorProto.FLOAT) graph = make_graph([node1, node2], "lr", [X], [Y]) onnx_model = make_model(graph) check_model(onnx_model) return onnx_model for to, expect in expected.items(): with self.subTest(to=to): onnx_model = model_cast_cast(to) ref = ReferenceEvaluator(onnx_model) y = ref.run(None, {"X": x})[0] assert_allclose(expect, y) self.assertEqual(expect.shape, y.shape) self.assertEqual(expect.dtype, y.dtype) def test_cast_bfloat16_output(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.BFLOAT16, [None]) model = make_model( make_graph( [ make_node("Cast", ["X"], ["Y"], to=TensorProto.BFLOAT16), ], "g", [X], [Y], ) ) ref = ReferenceEvaluator(model) data = np.array([0, 1, 2, 1e5, 200], dtype=np.float32) expected1 = np.array([float32_to_bfloat16(x) for x in data]) got = ref.run(None, {"X": data}) self.assertEqual(expected1.tolist(), got[0].tolist()) def test_quantize_linear_e4m3(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) model = make_model( make_graph( [ make_node( "Constant", [], ["scale"], value=make_tensor("scale", TensorProto.FLOAT, [1], [2.0]), ), make_node( "Constant", [], ["zero"], value=make_tensor("zero", TensorProto.FLOAT8E4M3FN, [1], [0.0]), ), make_node("QuantizeLinear", ["X", "scale", "zero"], ["T"]), make_node("DequantizeLinear", ["T", "scale"], ["Y"], axis=0), ], "g", [X], [Y], ) ) ref = ReferenceEvaluator(model) data = np.array([0, 1, 2, 1e5, 200], dtype=np.float32) expected = np.array([0, 1, 2, 896, 192], dtype=np.float32) got = ref.run(None, {"X": data}) assert_allclose(expected, got[0]) def test_quantize_linear_e4m3_initializer(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) model = make_model( make_graph( [ make_node("QuantizeLinear", ["X", "scale", "zero"], ["T"]), make_node("DequantizeLinear", ["T", "scale"], ["Y"], axis=0), ], "g", [X], [Y], [ make_tensor("scale", TensorProto.FLOAT, [1], [2.0]), make_tensor("zero", TensorProto.FLOAT8E4M3FN, [1], [0.0]), ], ) ) ref = ReferenceEvaluator(model) data = np.array([0, 1, 2, 1e5, 200], dtype=np.float32) expected = np.array([0, 1, 2, 896, 192], dtype=np.float32) got = ref.run(None, {"X": data}) assert_allclose(expected, got[0]) def test_quantize_linear_e5m2(self): X = make_tensor_value_info("X", TensorProto.FLOAT, [None]) Y = make_tensor_value_info("Y", TensorProto.FLOAT, [None]) model = make_model( make_graph( [ make_node( "Constant", [], ["scale"], value=make_tensor("scale", TensorProto.FLOAT, [1], [2.0]), ), make_node( "Constant", [], ["zero"], value=make_tensor("zero", TensorProto.FLOAT8E5M2, [1], [0.0]), ), make_node("QuantizeLinear", ["X", "scale", "zero"], ["T"]), make_node("DequantizeLinear", ["T", "scale"], ["Y"], axis=0), ], "g", [X], [Y], ) ) ref = ReferenceEvaluator(model) data = np.array([0, 1, 2, 1e5, 200], dtype=np.float32) expected = np.array([0, 1, 2, 98304, 192], dtype=np.float32) got = ref.run(None, {"X": data}) assert_allclose(expected, got[0]) def test_lrn(self): def _expected(x, alpha, beta, bias, size): square_sum = np.zeros((5, 5, 5, 5)).astype(np.float32) for n, c, h, w in np.ndindex(x.shape): square_sum[n, c, h, w] = sum( x[ n, max(0, c - int(math.floor((size - 1) / 2))) : min( 5, c + int(math.ceil((size - 1) / 2)) + 1 ), h, w, ] ** 2 ) y = x / ((bias + (alpha / size) * square_sum) ** beta) return y # keepdims is ignored in that case alpha = 0.0002 beta = 0.5 bias = 2.0 size = 3 X = make_tensor_value_info("X", TensorProto.FLOAT, [5, 5, 50, 50]) Z = make_tensor_value_info("Z", TensorProto.UNDEFINED, None) nodes = [ make_node("LRN", ["X"], ["Z"], alpha=alpha, beta=beta, bias=bias, size=size) ] model = make_model(make_graph(nodes, "g", [X], [Z])) ref = ReferenceEvaluator(model) data = np.random.rand(5, 5, 5, 5).astype(np.float32) got = ref.run(None, {"X": data}) expected = _expected(data, alpha, beta, bias, size) self.assertEqual(len(expected), len(got[0])) def test_conv_im2col_1d(self): feeds = { "X": np.arange(1 * 1 * 11).reshape((1, 1, 11)).astype(np.float32) + 1, "W": np.arange(3).reshape((1, 1, 3)).astype(np.float32), "B": np.zeros((1,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1], kernel_shape=[3], pads=[1, 1], strides=[1], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_1d_pad0(self): feeds = { "X": np.arange(2 * 4 * 3).reshape((2, 4, -1)).astype(np.float32) + 1, "W": np.arange(2 * 4 * 3).reshape((-1, 4, 3)).astype(np.float32), "B": np.zeros((1,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1], kernel_shape=[3], pads=[0, 0], strides=[1], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_2d(self): feeds = { "X": np.arange(1 * 1 * 11 * 23).reshape((1, 1, 11, 23)).astype(np.float32) + 1, "W": np.arange(9).reshape((1, 1, 3, 3)).astype(np.float32), "B": np.zeros((1,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1, 1], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[1, 1], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_2d_pad0(self): feeds = { "X": np.arange(2 * 3 * 5 * 2).reshape((2, 3, 5, -1)).astype(np.float32) + 1, "W": 2 ** np.arange(3 * 3 * 1 * 2).reshape((-1, 3, 1, 2)).astype(np.float32), "B": np.zeros((1,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1, 1], kernel_shape=[1, 2], pads=[0, 0, 0, 0], strides=[1, 1], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_2d_autopad(self): feeds = { "X": np.arange(5 * 5).reshape((1, 1, 5, -1)).astype(np.float32) + 1, "W": 2 ** np.arange(3 * 3).reshape((1, 1, 3, 3)).astype(np.float32), "B": np.zeros((1,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1, 1], kernel_shape=[3, 3], strides=[2, 2], pads=None, auto_pad="SAME_LOWER", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_3d(self): feeds = { "X": np.arange(1 * 1 * 11 * 5 * 13) .reshape((1, 1, 11, 5, 13)) .astype(np.float32) + 1, "W": np.arange(27).reshape((1, 1, 3, 3, 3)).astype(np.float32), "B": np.zeros((1,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1, 1, 1], kernel_shape=[3, 3, 3], pads=[1, 1, 1, 1, 1, 1], strides=[1, 1, 1], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_2d_strides(self): feeds = { "X": np.arange(1 * 3 * 6 * 6).reshape((1, 3, 6, 6)).astype(np.float32) + 1, "W": np.arange(2 * 3 * 3 * 3).reshape((2, 3, 3, 3)).astype(np.float32), "B": np.zeros((2,), dtype=np.float32), } kwargs = dict( group=1, dilations=[1, 1], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) def test_conv_im2col_2d_dilations(self): feeds = { "X": np.arange(1 * 3 * 6 * 6).reshape((1, 3, 6, 6)).astype(np.float32) + 1, "W": np.arange(2 * 3 * 3 * 3).reshape((2, 3, 3, 3)).astype(np.float32), "B": np.zeros((2,), dtype=np.float32), } kwargs = dict( group=1, dilations=[2, 1], kernel_shape=[3, 3], pads=[1, 1, 1, 1], strides=[2, 2], auto_pad="NOTSET", ) expected = _conv_implementation(**feeds, **kwargs) got = _conv_implementation_im2col(**feeds, **kwargs) assert_allclose(expected, got) @parameterized.parameterized.expand( [ ("ReduceSum",), ("ReduceL1",), ("ReduceL2",), ("ReduceMin",), ("ReduceMax",), ("ReduceProd",), ("ReduceSumSquare",), ] ) def test_reduce_op_no_axis(self, op): X = make_tensor_value_info("X", TensorProto.FLOAT, None) Y = make_tensor_value_info("Y", TensorProto.FLOAT, None) data = np.arange(6).reshape((1, 3, 2)).astype(np.float32) nodes = [make_node(op, ["X"], ["Y"], keepdims=0)] model = make_model(make_graph(nodes, "g", [X], [Y])) ref = ReferenceEvaluator(model) got = ref.run(None, {"X": data}) r = got[0] self.assertIsInstance(r, np.ndarray) self.assertEqual(r.shape, ()) @parameterized.parameterized.expand([(1,), (2,), (3,), (4,), (5,), (6,)]) def test_pad(self, dim): X = make_tensor_value_info("X", TensorProto.FLOAT, None) P = make_tensor_value_info("P", TensorProto.INT64, None) V = make_tensor_value_info("V", TensorProto.FLOAT, None) Y = make_tensor_value_info("Y", TensorProto.FLOAT, None) value = np.array([-5], dtype=np.float32) node = make_node("Pad", inputs=["X", "P", "V"], outputs=["Y"], mode="constant") model = make_model(make_graph([node], "g", [X, P, V], [Y])) ref = ReferenceEvaluator(model) x = np.array([1], dtype=np.float32).reshape((1,) * dim) p = np.array([1, 1] * dim, dtype=np.int64) got = ref.run(None, {"X": x, "P": p, "V": value})[0] self.assertEqual(got.shape, (3,) * dim) self.assertEqual(got.dtype, np.float32) p = np.repeat([7, 3], dim).astype(np.int64) got = ref.run(None, {"X": x, "P": p, "V": value})[0] self.assertEqual(got.shape, (11,) * dim) self.assertEqual(got.dtype, np.float32) def test_constant_of_shape(self): X = make_tensor_value_info("X", TensorProto.FLOAT, None) Y = make_tensor_value_info("Y", TensorProto.FLOAT, None) nodes = [ make_node("Shape", inputs=["X"], outputs=["shape"]), make_node( "ConstantOfShape", inputs=["shape"], outputs=["Y"], value=make_tensor("value", TensorProto.UINT16, [1], [1]), ), ] model = make_model(make_graph(nodes, "g", [X], [Y])) ref = ReferenceEvaluator(model) x = np.array(1, dtype=np.float32) got = ref.run(None, {"X": x})[0] self.assertEqual(got.shape, tuple()) self.assertEqual(got.dtype, np.uint16) assert_allclose(np.array(1, dtype=np.uint16), got) def test_constant_of_shape_castlike(self): X = make_tensor_value_info("X", TensorProto.FLOAT, None) Y = make_tensor_value_info("Y", TensorProto.FLOAT, None) nodes = [ make_node( "Constant", [], ["like"], value=make_tensor("c", TensorProto.UINT16, [1], [2]), ), make_node("Shape", inputs=["X"], outputs=["shape"]), make_node( "ConstantOfShape", inputs=["shape"], outputs=["cst"], value=make_tensor("value", TensorProto.INT64, [1], [1]), ), make_node("CastLike", ["cst", "like"], ["Y"]), ] model = make_model(make_graph(nodes, "g", [X], [Y])) ref = ReferenceEvaluator(model) x = np.array(1, dtype=np.float32) got = ref.run(None, {"X": x})[0] self.assertEqual(got.shape, tuple()) self.assertEqual(got.dtype, np.uint16) assert_allclose(np.array(1, dtype=np.uint16), got) @parameterized.parameterized.expand( [ (["abc", "def"], [".com", ".net"], ["abc.com", "def.net"], (2,)), (["cat", "dog", "snake"], ["s"], ["cats", "dogs", "snakes"], (3,)), ("cat", "s", "cats", ()), (["a", "ß", "y"], ["a", "ß", "y"], ["aa", "ßß", "yy"], (3,)), ] ) def test_string_concat(self, a, b, expected, expected_shape): A = make_tensor_value_info("A", TensorProto.STRING, None) B = make_tensor_value_info("B", TensorProto.STRING, None) Y = make_tensor_value_info("Y", TensorProto.STRING, None) node = make_node("StringConcat", inputs=["A", "B"], outputs=["Y"]) model = make_model(make_graph([node], "g", [A, B], [Y])) ref = ReferenceEvaluator(model) result, *_ = ref.run(None, {"A": np.array(a), "B": np.array(b)}) np.testing.assert_array_equal(result, expected) self.assertEqual(result.dtype.kind, "O") self.assertEqual(result.shape, expected_shape) if __name__ == "__main__": unittest.main(verbosity=2)
145,760
37.469517
115
py
onnx
onnx-main/onnx/test/reference_evaluator_backend_test.py
# Copyright (c) ONNX Project Contributors # SPDX-License-Identifier: Apache-2.0 # type: ignore # pylint: disable=C0415,R0912,R0913,R0914,R0915,W0613,W0640,W0703 """ These test evaluates the python runtime (class ReferenceEvaluator) against all the backend tests (in onnx/backend/test/case/node) and checks the runtime produces the expected outputs. You may run one specific test with following command line: :: python onnx/test/reference_evaluator_backend_test.py TestOnnxBackEndWithReferenceEvaluator.test_group_normalization_example You may bypass a test newly added by adding to the global variable `SKIP_TESTS`. You may refine the absolute or relative tolerance for a test by adding an item in method `setUpClass` and attributes `atol` or `rtol`. """ import os import pprint import unittest try: from packaging.version import parse as version except ImportError: from distutils.version import ( # noqa: N813 # pylint: disable=deprecated-module StrictVersion as version, ) from os import getenv import numpy as np from numpy import __version__ as npver from numpy import object_ as dtype_object from numpy.testing import assert_allclose # type: ignore from onnx import ONNX_ML, OptionalProto, SequenceProto, TensorProto, load from onnx.backend.test import __file__ as backend_folder from onnx.helper import __file__ as onnx_file from onnx.numpy_helper import bfloat16_to_float32, to_list, to_optional from onnx.reference import ReferenceEvaluator from onnx.reference.op_run import to_array_extended from onnx.reference.ops.op_cast import cast_to # TODO (https://github.com/microsoft/onnxruntime/issues/14932): Get max supported version from onnxruntime directly # For now, bump the version in CIs whenever there is a new onnxruntime release ORT_MAX_IR_SUPPORTED_VERSION = int(getenv("ORT_MAX_IR_SUPPORTED_VERSION", "8")) ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION = int( getenv("ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION", "18") ) # Number of tests expected to pass without raising an exception. MIN_PASSING_TESTS = 1235 # Update this list if one new operator does not have any implementation. SKIP_TESTS = { # mismatches # shapes (10, 9, 3), (10, 8, 3) shape mismatch unexpected as the operator is inlined "test_center_crop_pad_crop_axes_hwc_expanded", # deprecated "test_scan_sum", # deprecated, opset 8 -> not implemented "test_scatter_with_axis", # deprecated, scatter is removed "test_scatter_without_axis", # deprecated, scatter is removed # not implemented "test__simple_gradient_of_add", # gradient not implemented "test__simple_gradient_of_add_and_mul", # gradient not implemented "test_lppool_2d_dilations", # CommonPool._run returns incorrect output shape when dilations is set "test_averagepool_2d_dilations", # CommonPool._run returns incorrect output shape when dilations is set } if version(npver) < version("1.21.5"): SKIP_TESTS |= { "test_cast_FLOAT_to_BFLOAT16", "test_castlike_FLOAT_to_BFLOAT16", "test_castlike_FLOAT_to_BFLOAT16_expanded", } if version(npver) < version("1.21.5"): SKIP_TESTS |= { "test_cast_FLOAT_to_BFLOAT16", "test_castlike_FLOAT_to_BFLOAT16", "test_castlike_FLOAT_to_BFLOAT16_expanded", } def assert_allclose_string(expected, value): """ Compares two arrays knowing they contain strings. Raises an exception if the test fails. :param expected: expected array :param value: value """ def is_float(x): try: float(x) return True except ValueError: return False if all(map(is_float, expected.ravel())): expected_float = expected.astype(np.float32) value_float = value.astype(np.float32) assert_allclose(expected_float, value_float) else: if expected.tolist() != value.tolist(): raise AssertionError(f"Mismatches {expected} != {value}.") class OnnxBackendTest: """ Definition of a backend test. It starts with a folder, in this folder, one onnx file must be there, then a subfolder for each test to run with this model. :param folder: test folder :param onnx_path: onnx file :param onnx_model: loaded onnx file :param tests: list of test """ @staticmethod def _sort(filenames): temp = [] for f in filenames: name = os.path.splitext(f)[0] i = name.split("_")[-1] temp.append((int(i), f)) temp.sort() return [_[1] for _ in temp] @staticmethod def _read_proto_from_file(full): if not os.path.exists(full): raise FileNotFoundError(f"File not found: {full!r}.") with open(full, "rb") as f: serialized = f.read() return OnnxBackendTest._read_proto_from_serialized(serialized, full) @staticmethod def _read_proto_from_serialized(serialized, full): if not os.path.exists(full): raise FileNotFoundError(f"File not found: {full!r}.") with open(full, "rb") as f: serialized = f.read() proto_types = [ (TensorProto, to_array_extended), (SequenceProto, to_list), (OptionalProto, to_optional), ] exc = None for pt, cvt in proto_types: obj = pt() try: obj.ParseFromString(serialized) try: return cvt(obj) except ValueError as e: exc = e continue except Exception as e: exc = e raise RuntimeError( f"Unable to read {full!r}, error is {exc}, " f"content is {serialized[:100]!r}." ) from exc @staticmethod def _load(folder, names): res = [] for name in names: full = os.path.join(folder, name) obj = OnnxBackendTest._read_proto_from_file(full) res.append(obj) return res def __repr__(self): "usual" return f"{self.__class__.__name__}({self.folder!r})" def __init__(self, folder): if not os.path.exists(folder): raise FileNotFoundError(f"Unable to find folder {folder!r}.") content = os.listdir(folder) onx = [c for c in content if os.path.splitext(c)[-1] in {".onnx"}] if len(onx) != 1: raise ValueError( f"There is more than one onnx file in {folder!r} ({onx!r})." ) self.folder = folder self.onnx_path = os.path.join(folder, onx[0]) self.onnx_model = load(self.onnx_path) self.tests = [] for sub in content: full = os.path.join(folder, sub) if os.path.isdir(full): pb = [c for c in os.listdir(full) if os.path.splitext(c)[-1] in {".pb"}] inputs = OnnxBackendTest._sort(c for c in pb if c.startswith("input_")) outputs = OnnxBackendTest._sort( c for c in pb if c.startswith("output_") ) self.tests.append( { "inputs": OnnxBackendTest._load(full, inputs), "outputs": OnnxBackendTest._load(full, outputs), } ) @property def name(self): "Returns the test name." return os.path.split(self.folder)[-1] @property def fname(self): folder = self.folder.replace("\\", "/").split("/")[-2] if folder.endswith("node"): fname = self.name else: fname = f"test__{folder.replace('-', '_')}_{self.name[5:]}" if "/" in fname or fname == "test__test_AvgPool1d_AvgPool1d": raise AssertionError( f"name={self.name!r}, folder={folder!r}, self.folder={self.folder}." ) return fname def __len__(self): "Returns the number of tests." return len(self.tests) def _compare_results( self, index, i_output, desired, output, rtol=0, atol=0, comment="", inputs=None ): """ Compares the expected output and the output produced by the runtime. Raises an exception if not equal. :param index: test index :param i_output: output index :param desired: expected output :param output: output :param rtol: relative tolerance :param atol: absolute tolerance :param comment: addition text to give more insights to the user :param inputs: inputs to the model """ if comment == "": raise RuntimeError("Argument comment should be filled.") if atol is None: atol = 0 if rtol is None: rtol = 0 if isinstance(desired, np.ndarray): if isinstance(output, np.ndarray): if rtol == 0: if desired.dtype == np.float32: rtl = 1e-5 elif desired.dtype == np.float64: rtl = 1e-12 else: rtl = rtol else: rtl = rtol if desired.dtype == dtype_object: try: assert_allclose_string(desired, output) except AssertionError as ex: raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} failed, comment={comment}." ) from ex else: equal_nan = desired.dtype in (np.float16, np.float32, np.float64) if equal_nan: try: assert_allclose( desired, output, atol=atol, rtol=rtl, equal_nan=equal_nan, ) except AssertionError as ex: try: diff = output - desired except ValueError: diff = None raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} failed " f"(rtol={rtl}, atol={atol}), comment={comment}\n---\n{desired}\n----" f"\n{output}\n-----\n{diff}\n------INPUTS----\n{pprint.pformat(inputs)}." ) from ex else: # float 8 types if desired.dtype != output.dtype: raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} " f"has unexpected type {output.dtype} (expecting {desired.dtype}.)" ) if desired.tolist() != output.tolist(): raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} " f"has unexpected values {output} (expecting {desired}.)" ) if desired.shape != output.shape: raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} failed " f"(expected shape={desired.shape} but shape={output.shape}), " f"comment={comment}\n---\n{desired}\n----" f"\n{output}\n------INPUTS----\n{pprint.pformat(inputs)}." ) elif hasattr(output, "is_compatible"): # A shape if desired.dtype != output.dtype: raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} failed " f"(desired.dtype={desired.dtype!r}, output={output!r}), comment={comment}." ) if not output.is_compatible(desired.shape): raise AssertionError( f"Output {i_output} of test {index} in folder {self.folder!r} failed " f"(desired.shape={desired.shape}, output={output!r}), comment={comment}." ) elif isinstance(desired, list): if not isinstance(output, list): raise AssertionError( f"Expected result is 'list' but output type is {type(output)} for output {i_output}" f", comment={comment}\n--EXPECTED--\n{desired}\n--GOT--\n{output}." ) if len(desired) != len(output): raise AssertionError( f"Expected has {len(desired)} but output has {len(output)} for output {i_output}" f", comment={comment}\n--EXPECTED--\n{desired}\n--GOT--\n{output}." ) for a, b in zip(desired, output): self._compare_results( index, i_output, a, b, rtol=rtol, atol=atol, comment=comment ) else: raise NotImplementedError( f"Comparison not implemented for type {type(desired)} and output {i_output}, comment={comment}." ) def is_random(self): "Tells if a test is random or not." if "bernoulli" in self.folder: return True return False def run( self, load_fct, run_fct, index=None, rtol=1e-07, atol=0, comment="", print_io=False, ): """ Executes a tests or all tests if index is None. The function crashes if the tests fails. :param load_fct: loading function, takes a loaded onnx graph, and returns an object :param run_fct: running function, takes the result of previous function, the inputs, and returns the outputs :param index: index of the test to run or all. :param rtol: relative tolerance :param atol: absolute tolerance :param comment: additional information for the user :param print_io: prints out the input and output """ if index is None: res = [] for i in range(len(self)): res.append( self.run( load_fct, run_fct, index=i, atol=atol, rtol=rtol, comment=comment, print_io=print_io, ) ) return res if print_io: print("------ INPUTS") for k, v in enumerate(self.tests[index]["inputs"]): print(f"input {k!r}, shape={v.shape}, dtype={v.dtype}") print("------ EXPECTED OUTPUTS") for k, v in enumerate(self.tests[index]["outputs"]): print(f"output {k!r}, shape={v.shape}, dtype={v.dtype}") obj = load_fct(self.onnx_model) got = run_fct(obj, *self.tests[index]["inputs"]) expected = self.tests[index]["outputs"] if len(got) != len(expected): raise AssertionError( f"Unexpected number of output (test {index}, folder {self.folder!r}), " f"got {len(got)}, expected {len(expected)}." ) res = { "inputs": self.tests[index]["inputs"], "expected": self.tests[index]["outputs"], "results": got, } for i, (e, o) in enumerate(zip(expected, got)): if self.is_random(): if e.dtype != o.dtype: raise AssertionError( f"Output {i} of test {index} in folder {self.folder!r} failed " f"(type mismatch {e.dtype} != {o.dtype!r})." ) if e.shape != o.shape: raise AssertionError( f"Output {i} of test {index} in folder {self.folder!r} failed " f"(shape mismatch {e.shape} != {o.shape})." ) else: self._compare_results( index, i, e, o, atol=atol, rtol=rtol, comment=comment + "\n" + str(self.onnx_model), inputs=self.tests[index]["inputs"], ) return res def enumerate_onnx_tests(series, fct_filter=None): """ Collects test from a sub folder of `onnx/backend/test`. Works as an enumerator to start processing them without waiting or storing too much of them. :param series: which subfolder to load, possible values: (`'node'`, ...) :param fct_filter: function `lambda testname: boolean` to load or skip the test, None for all :return: list of @see cl OnnxBackendTest """ root = os.path.dirname(backend_folder) sub = os.path.join(root, "data", series) if not os.path.exists(sub): content = "\n".join(os.listdir(root)) raise FileNotFoundError( f"Unable to find series of tests in {root!r}, subfolders:\n{content}" ) tests = os.listdir(sub) for t in tests: if fct_filter is not None and not fct_filter(t): continue folder = os.path.join(sub, t) if not ONNX_ML and "ai_onnx_ml" in folder: continue content = os.listdir(folder) onx = [c for c in content if os.path.splitext(c)[-1] in {".onnx"}] if len(onx) == 1: yield OnnxBackendTest(folder) class TestOnnxBackEndWithReferenceEvaluator(unittest.TestCase): folder = os.path.join( os.path.abspath(os.path.dirname(__file__)), "onnx_backend_test_code" ) @classmethod def add_test_methods(cls): for folder in ["node", "pytorch-converted", "pytorch-operator", "simple"]: for te in enumerate_onnx_tests(folder): def _test_( self, te=te, check_other_runtime=None, verbose=0, print_io=False ): if te.fname in getattr(cls, "skip_test", set()): cls.skipped.append((te, None)) return rtol = getattr(cls, "rtol", {}) atol = getattr(cls, "atol", {}) if len(rtol) == 0 or len(atol) == 0: raise AssertionError("rtol or atol is empty.") self.common_test_onnx_test_run( te, getattr(cls, "successes", []), getattr(cls, "missed", []), getattr(cls, "skipped", []), getattr(cls, "load_failed", []), getattr(cls, "exec_failed", []), getattr(cls, "mismatch", []), verbose=verbose, rtol=rtol, atol=atol, check_other_runtime=check_other_runtime, print_io=print_io, ) setattr(TestOnnxBackEndWithReferenceEvaluator, te.fname, _test_) def test_onnx_backend_test_abs(self): name = "test_abs" code = [] for te in enumerate_onnx_tests("node", lambda folder: folder == name): code.append(te) self.assertEqual(len(code), 1) def test_onnx_backend_test_expand_shape_model1(self): name = "test_expand_shape_model1" code = [] for te in enumerate_onnx_tests("simple", lambda folder: folder == name): code.append(te) self.assertEqual(len(code), 1) @staticmethod def load_fct(obj, verbose=0): return ReferenceEvaluator(obj, verbose=verbose) @staticmethod def run_fct(obj, *inputs, verbose=0): # pylint: disable=W0613 if hasattr(obj, "input_names"): input_names = obj.input_names elif hasattr(obj, "get_inputs"): input_names = [_.name for _ in obj.get_inputs()] else: raise AttributeError( f"Unable to extract the number to guess the number of inputs for type {type(obj)}." ) if len(input_names) < len(inputs): raise AssertionError( f"Got {len(inputs)} inputs but expecting {len(obj.input_names)}." ) rewrite = False for i in range(len(inputs)): # pylint: disable=C0200 if ( isinstance(inputs[i], np.ndarray) and inputs[i].dtype == np.uint16 and obj.input_types[i].tensor_type.elem_type != TensorProto.UINT16 ): rewrite = True if rewrite: # bfloat16 does not exist for numpy. inputs = list(inputs) for i in range(len(inputs)): # pylint: disable=C0200 if ( isinstance(inputs[i], np.ndarray) and inputs[i].dtype == np.uint16 and obj.input_types[i].tensor_type.elem_type != TensorProto.UINT16 ): xr = inputs[i].ravel() xf = np.empty(xr.shape[0], dtype=np.float32) for ie in range(xr.shape[0]): el = bfloat16_to_float32(xr[ie]) xf[ie] = el inputs[i] = cast_to( xf.astype(np.float32).reshape(inputs[i].shape), TensorProto.BFLOAT16, True, ) feeds = {input_names[i]: inputs[i] for i in range(len(inputs))} got = obj.run(None, feeds) return got # def test_onnx_test_run_test_abs(self): # done = 0 # for te in enumerate_onnx_tests("node", lambda folder: folder == "test_abs"): # self.assertIn(te.name, repr(te)) # self.assertGreater(len(te), 0) # te.run( # TestOnnxBackEndWithReferenceEvaluator.load_fct, # TestOnnxBackEndWithReferenceEvaluator.run_fct, # comment="[runtime=ReferenceEvaluator]", # ) # done += 1 # self.assertEqual(done, 1) def common_test_onnx_test_run( self, te, successes, missed, skipped, load_failed, exec_failed, mismatch, verbose=0, rtol=None, atol=None, check_other_runtime=None, print_io=False, ): if verbose > 6: print("TEST:", te.name) if verbose > 7: print(" check runtime") self.assertIn(te.name, repr(te)) self.assertGreater(len(te), 0) try: if verbose > 7: print(" run") if verbose > 5: te.run( lambda *args, verbose=verbose: TestOnnxBackEndWithReferenceEvaluator.load_fct( *args, verbose ), TestOnnxBackEndWithReferenceEvaluator.run_fct, atol=atol.get(te.name, None), rtol=rtol.get(te.name, None), comment=f"[runtime=ReferenceEvaluator, verbose={verbose}]", print_io=print_io, ) else: te.run( TestOnnxBackEndWithReferenceEvaluator.load_fct, TestOnnxBackEndWithReferenceEvaluator.run_fct, atol=atol.get(te.fname, atol.get(te.name, None)), rtol=rtol.get(te.fname, rtol.get(te.name, None)), comment="[runtime=ReferenceEvaluator]", print_io=print_io, ) if verbose > 7: print(" end run") if verbose > 8: print(te.onnx_model) except NotImplementedError as e: if verbose > 7: print(" ", e, type(e)) missed.append((te, e)) with open(f"missed_{te.name}.onnx", "wb") as f: f.write(te.onnx_model.SerializeToString()) raise e except (AssertionError, ValueError) as e: if verbose > 7: print(" ", e, type(e)) mismatch.append((te, e)) with open(f"mismatch_{te.name}.onnx", "wb") as f: f.write(te.onnx_model.SerializeToString()) if check_other_runtime is None: raise e if "onnxruntime" in check_other_runtime: print("CHECK RUNTIME onnxruntime") from onnxruntime import InferenceSession onnx_domain_opset = ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION for opset in te.onnx_model.opset_import: if opset.domain in ("", "ai.onnx"): onnx_domain_opset = opset.version break # The new IR or opset version is not supported by onnxruntime yet if ( te.onnx_model.ir_version > ORT_MAX_IR_SUPPORTED_VERSION or onnx_domain_opset > ORT_MAX_ONNX_OPSET_SUPPORTED_VERSION ): print( "Skip test because of IR or opset version is not supported by onnxruntime yet" ) return te.run( lambda obj: InferenceSession( obj.SerializeToString(), providers=["CPUExecutionProvider"] ), lambda *a, **b: TestOnnxBackEndWithReferenceEvaluator.run_fct( *a, verbose=1, **b ), atol=1e-5, rtol=1e-3, comment="[runtime=onnxruntime]", ) print("done") raise e except Exception as e: if verbose > 7: print(" ", e, type(e)) with open(f"issue_{te.name}.onnx", "wb") as f: f.write(te.onnx_model.SerializeToString()) raise AssertionError( f"Unable to run test {te.name!r} due to {e}\n{te.onnx_model}" ) from e successes.append((te, atol.get(te.fname, None), rtol.get(te.fname, None))) if verbose > 7: print(" end example.") @staticmethod def _postprocess( successes, missed, skipped, load_failed, exec_failed, mismatch, verbose ): success = len(successes) failed = [ len(missed), len(skipped), len(load_failed), len(exec_failed), len(mismatch), ] coverage = success / (success + sum(failed)) if verbose: path = os.path.dirname(onnx_file) print("-----------") print( f"success={success}, skipped={len(skipped)}, missed={len(missed)}, load_failed={len(load_failed)}, " f"exec_failed={len(exec_failed)}, mismatch={len(mismatch)}" ) print( f"coverage {coverage * 100:.1f}% out of {success + sum(failed)} tests" ) if verbose > 3: def _print(s, path): return ( str(s) .replace("\\\\", "\\") .replace(path, "onnx") .replace("\\", "/") ) print("-----------") for t in sorted(load_failed, key=lambda m: m[0].fname): print("loading failed", t[0].fname, "---", _print(t[0], path)) for t in sorted(exec_failed, key=lambda m: m[0].fname): print("execution failed", t[0].fname, "---", _print(t[0], path)) for t in sorted(mismatch, key=lambda m: m[0].fname): print("mismatch", t[0].fname, "---", _print(t[0], path)) for t in sorted(missed, key=lambda m: m[0].fname): print("missed ", t[0].fname, "---", _print(t[0], path)) for t in sorted(skipped, key=lambda m: m[0].fname): print("skipped", t[0].fname, "---", _print(t[0], path)) if success > 30: print("-----------") print( f"success={success}, skipped={len(skipped)}, missed={len(missed)}, load_failed={len(load_failed)}, " f"exec_failed={len(exec_failed)}, mismatch={len(mismatch)}" ) print( f"coverage {coverage * 100:.1f}% out of {success + sum(failed)} tests" ) print("-----------") if len(mismatch) > 0: te, e = mismatch[0] raise AssertionError( f"Mismatch in test {te.name!r}\n{te.onnx_model}." ) from e if sum(failed) > len(SKIP_TESTS): raise AssertionError( f"Unexpected failures. {sum(failed)}/{success + sum(failed)} tests have failed." f"The coverage is {coverage * 100:.1f}%. " f"New operators were added with no corresponding runtime." ) @classmethod def setUpClass(cls, all_tests=False): # test not supported yet # not supported yet # see https://onnx.ai/backend-scoreboard/onnxruntime_details_stable.html # to compare with onnxruntime cls.rtol = { "test_adam_multiple": 1e-2, "test_blackmanwindow_expanded": 0, "test_blackmanwindow_symmetric_expanded": 0, "test_simple_rnn_batchwise": 0, "test__pytorch_converted_Conv1d_pad1": 1e-4, "test__pytorch_converted_Conv2d": 1e-5, "test__pytorch_converted_Conv2d_no_bias": 1e-3, "test__pytorch_converted_Conv2d_strided": 1e-4, "test_layer_normalization_4d_axis1_expanded_ver18": 1e-4, "test_layer_normalization_4d_axis_negative_1_expanded_ver18": 1e-4, "test_layer_normalization_4d_axis_negative_3_expanded_ver18": 1e-4, } cls.atol = { "test_blackmanwindow": 1e-7, "test_blackmanwindow_expanded": 1e-4, "test_blackmanwindow_symmetric": 1e-7, "test_blackmanwindow_symmetric_expanded": 1e-4, "test_Conv1d": 1e-6, "test_Conv2d_depthwise_padded": 1e-7, "test_Conv3d_dilated": 1e-6, "test_gridsample_bicubic": 1e-4, "test_gru_seq_length": 1e-7, "test_hammingwindow_expanded": 1e-4, "test_hammingwindow_symmetric_expanded": 1e-4, "test_hannwindow_expanded": 1e-4, "test_hannwindow_symmetric": 1e-7, "test_hannwindow_symmetric_expanded": 1e-4, "test_layer_normalization_4d_axis_negative_1_expanded": 1e-6, "test_layer_normalization_4d_axis1_expanded": 1e-6, "test_layer_normalization_4d_axis_negative_3_expanded": 1e-6, "test_mish": 1e-6, "test_mish_expanded": 1e-6, "test_roialign_aligned_false": 1e-4, "test_roialign_aligned_true": 1e-4, # extended list "test__pytorch_converted_ConvTranspose2d_no_bias": 1e-4, "test__pytorch_converted_Linear_no_bias": 1e-5, "test_Linear_no_bias": 1e-5, "test__pytorch_converted_Conv1d_pad1": 1e-6, "test__pytorch_converted_Conv2d": 1e-5, "test__pytorch_converted_Conv2d_depthwise": 1e-4, "test__pytorch_converted_Conv2d_depthwise_strided": 1e-4, "test__pytorch_converted_Conv2d_depthwise_with_multiplier": 1e-4, "test__pytorch_converted_Conv2d_depthwise_padded": 1e-4, "test__pytorch_converted_Conv2d_groups": 1e-4, "test__pytorch_converted_Conv2d_groups_thnn": 1e-4, "test__pytorch_converted_Conv2d_no_bias": 1e-5, "test__pytorch_converted_Conv2d_strided": 1e-4, "test__pytorch_operator_operator_symbolic_override": 1e-5, "test_operator_symbolic_override": 1e-4, "test__pytorch_converted_Conv3d_dilated_strided": 1e-4, "test__pytorch_converted_Conv3d_groups": 1e-4, } if version(npver) < version("1.21.5"): cls.atol.update( { "test_dft": 1e-11, "test_dft_axis": 1e-11, "test_dft_inverse": 1e-11, } ) cls.skip_test = SKIP_TESTS if all_tests: cls.skip_test = set() cls.successes = [] cls.missed = [] cls.skipped = [] cls.load_failed = [] cls.exec_failed = [] cls.mismatch = [] @classmethod def tearDownClass(cls): if len(cls.successes) == 0: failed = cls.mismatch + cls.missed + cls.load_failed + cls.exec_failed if len(failed) > 0: raise RuntimeError( f"No test was successful, {len(failed)} failed." ) from failed[0][1] raise RuntimeError("No test was successful.") cls._postprocess( cls.successes, cls.missed, cls.skipped, cls.load_failed, cls.exec_failed, cls.mismatch, 10, ) TestOnnxBackEndWithReferenceEvaluator.add_test_methods() if __name__ == "__main__": unittest.main(verbosity=2)
34,178
38.106407
127
py
adagrid
adagrid-main/uniform-negative-sampling/adagrid_uniform_negative_sampling.py
from pathlib import Path import torch.multiprocessing torch.multiprocessing.set_sharing_strategy('file_system') import argparse import copy import math import random import numpy as np import pandas as pd from sklearn.metrics import * import torch import torch.nn.functional as F from torch.utils.data import DataLoader from torch_geometric.datasets import Planetoid import torch_geometric.transforms as T import torch_geometric.nn as pyg_nn from deepsnap.dataset import GraphDataset from deepsnap.batch import Batch class Arguments: def __init__(self, dataset='cora', device='cpu', epochs=500, mode='all', model='GCN', edge_message_ratio=0.6, layers=2, hidden_dim=64, batch_size=1, data_split=[0.85, 0.05, 0.1], verbose=False, adapt=False, try_epochs=3, adapt_epochs=50, criterion='val', random=False): self.dataset = dataset self.device = device self.epochs = epochs self.mode = mode self.model = model self.edge_message_ratio = edge_message_ratio self.layers = layers self.hidden_dim = hidden_dim self.batch_size = batch_size self.data_split = data_split self.verbose = verbose self.adapt = adapt self.try_epochs = try_epochs self.adapt_epochs = adapt_epochs self.criterion = criterion self.random = random def arg_parse(): parser = argparse.ArgumentParser(description='Link prediction arguments.') parser.add_argument('--gpu', type=int, help='GPU device.') parser.add_argument('--dataset', type=str, help='Dataset.') parser.add_argument('--num_layers', type=int, help='Number of layers of GNN.') parser.add_argument('--hidden_dim', type=int, help='Hidden dimension of GNN.') parser.add_argument('--verbose', action='store_true', help='Whether output is verbose.') parser.set_defaults( gpu=0, dataset='cora', num_layers=2, hidden_dim=64, verbose=False ) return parser.parse_args() class Net(torch.nn.Module): def __init__(self, input_dim, args): super(Net, self).__init__() self.model = args.model if self.model == 'GCN': self.conv_first = pyg_nn.GCNConv(input_dim, args.hidden_dim) self.convs = torch.nn.ModuleList([pyg_nn.GCNConv(args.hidden_dim, args.hidden_dim) for i in range(args.layers - 2)]) self.conv_last = pyg_nn.GCNConv(args.hidden_dim, args.hidden_dim) else: raise ValueError('unknown conv') self.loss_fn = torch.nn.BCEWithLogitsLoss() def forward(self, graph): x = F.dropout(graph.node_feature, p=0.2, training=self.training) x = F.relu(self._conv_op(self.conv_first, x, graph)) for i in range(len(self.convs)): x = F.dropout(x, p=0.2, training=self.training) x = F.relu(self._conv_op(self.convs[i], x, graph)) x = F.dropout(x, p=0.2, training=self.training) x = self._conv_op(self.conv_last, x, graph) nodes_first = torch.index_select(x, 0, graph.edge_label_index[0,:].long()) nodes_second = torch.index_select(x, 0, graph.edge_label_index[1,:].long()) pred = torch.sum(nodes_first * nodes_second, dim=-1) return pred def _conv_op(self, conv, x, graph): if self.model == 'GCN': return conv(x, graph.edge_index) elif self.model == 'spline': return conv(x, graph.edge_index, graph.edge_feature) def loss(self, pred, link_label): return self.loss_fn(pred, link_label) def try_edge_message_ratio(edge_message_ratio, model, datasets, dataloaders, optimizer, args, scheduler=None): datasets['train'].edge_message_ratio = edge_message_ratio val_max = -math.inf best_model = model mean_accs = {mode: 0 for mode, dataloader in dataloaders.items()} for epoch in range(args.try_epochs): for iter_i, batch in enumerate(dataloaders['train']): batch.to(args.device) model.train() optimizer.zero_grad() pred = model(batch) loss = model.loss(pred, batch.edge_label.type(pred.dtype)) loss.backward() optimizer.step() if scheduler is not None: scheduler.step() accs, _ = test(model, dataloaders, args) for mode in accs: mean_accs[mode] += accs[mode] if val_max < accs['val']: val_max = accs['val'] best_model = copy.deepcopy(model) for mode in mean_accs: mean_accs[mode] /= args.try_epochs log = 'Edge message ratio: {:.4f}, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}' if(args.verbose): print(log.format(edge_message_ratio, mean_accs['train'], mean_accs['val'], mean_accs['test'])) return mean_accs, val_max, best_model def train(model, datasets, dataloaders, optimizer, args, scheduler=None): # training loop val_max = -math.inf best_model = model input_dim = datasets['train'].num_node_features num_classes = datasets['train'].num_edge_labels waiting = 0 edge_message_ratio_changes = [] for epoch in range(0, args.epochs): if(waiting > 0): waiting -= 1 continue # search to determine the next edge_message_ratio if(args.adapt and epoch % args.adapt_epochs == 0): best_ratio = 0.0 try_val_max = None if(args.criterion == 'val'): try_val_max = -math.inf if(args.criterion == 'gap'): try_val_max = math.inf best_try_model = model best_try_optimizer = optimizer best_try_scheduler = scheduler # best version of the best_try_model during training for try epochs (highest validation accuracy) best_val_model = model best_val = -math.inf edge_message_ratios = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] for edge_message_ratio in edge_message_ratios: new_model = Net(input_dim, args).to(args.device) new_model.load_state_dict(model.state_dict()) new_optimizer = torch.optim.SGD(new_model.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4) new_optimizer.load_state_dict(optimizer.state_dict()) new_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(new_optimizer, T_max=args.epochs) new_scheduler.load_state_dict(scheduler.state_dict()) try_accs, current_val, current_val_model = try_edge_message_ratio(edge_message_ratio, new_model, datasets, dataloaders, new_optimizer, args, new_scheduler) if (args.criterion == 'val' and try_val_max < try_accs['val']): try_val_max = try_accs['val'] best_ratio = edge_message_ratio best_try_model = new_model best_try_optimizer = new_optimizer best_try_scheduler = new_scheduler best_val = current_val best_val_model = current_val_model if (args.criterion == 'gap' and try_val_max > abs(try_accs['train'] - try_accs['val'])): try_val_max = abs(try_accs['train'] - try_accs['val']) best_ratio = edge_message_ratio best_try_model = new_model best_try_optimizer = new_optimizer best_try_scheduler = new_scheduler best_val = current_val best_val_model = current_val_model model = best_try_model optimizer = best_try_optimizer scheduler = best_try_scheduler datasets['train'].edge_message_ratio = best_ratio edge_message_ratio_changes.append(datasets['train'].edge_message_ratio) if val_max < best_val: val_max = best_val best_model = best_val_model waiting = args.try_epochs - 1 continue if(args.random): datasets['train'].edge_message_ratio = random.uniform(0.1, 0.9) for iter_i, batch in enumerate(dataloaders['train']): batch.to(args.device) model.train() optimizer.zero_grad() pred = model(batch) loss = model.loss(pred, batch.edge_label.type(pred.dtype)) loss.backward() optimizer.step() if scheduler is not None: scheduler.step() log = 'Epoch: {:03d}, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}' accs, _ = test(model, dataloaders, args) if(args.verbose): print(log.format(epoch, accs['train'], accs['val'], accs['test'])) if val_max < accs['val']: val_max = accs['val'] best_model = copy.deepcopy(model) log = 'Best, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}' accs, _ = test(best_model, dataloaders, args) print(log.format(accs['train'], accs['val'], accs['test'])) return np.array([accs['train'], accs['val'], accs['test']]), edge_message_ratio_changes def test(model, dataloaders, args, max_train_batches=1): model.eval() accs = {} losses = {} for mode, dataloader in dataloaders.items(): acc = 0 loss = 0 num_batches = 0 for batch in dataloader: batch.to(args.device) pred = model(batch) # only 1 graph in dataset. In general needs aggregation loss += model.loss(pred, batch.edge_label.type(pred.dtype)).cpu().data.numpy() acc += roc_auc_score(batch.edge_label.flatten().cpu().numpy(), pred.flatten().data.cpu().numpy()) num_batches += 1 if mode == 'train' and num_batches >= max_train_batches: # do not eval on the entire training set for efficiency break accs[mode] = acc / num_batches losses[mode] = loss / num_batches return accs, losses def run(args): pyg_dataset = None if(args.dataset == 'cora'): pyg_dataset = Planetoid('./datasets', 'Cora', transform=T.TargetIndegree()) if(args.dataset == 'citeseer'): pyg_dataset = Planetoid('./datasets', 'CiteSeer', transform=T.TargetIndegree()) if(args.dataset == 'pubmed'): pyg_dataset = Planetoid('./datasets', 'PubMed', transform=T.TargetIndegree()) # the input that we assume users have edge_train_mode = args.mode if(args.verbose): print('edge train mode: {}'.format(edge_train_mode)) graphs = GraphDataset.pyg_to_graphs(pyg_dataset) dataset = GraphDataset(graphs, task='link_pred', edge_message_ratio=args.edge_message_ratio, edge_train_mode=edge_train_mode, resample_disjoint=True, resample_disjoint_period=1) if(args.verbose): print('Initial dataset: {}'.format(dataset)) # split dataset datasets = {} datasets['train'], datasets['val'], datasets['test'] = dataset.split( transductive=True, split_ratio=args.data_split) if(args.verbose): print('after split') print('Train message-passing graph: {} nodes; {} edges.'.format( datasets['train'][0].G.number_of_nodes(), datasets['train'][0].G.number_of_edges())) print('Val message-passing graph: {} nodes; {} edges.'.format( datasets['val'][0].G.number_of_nodes(), datasets['val'][0].G.number_of_edges())) print('Test message-passing graph: {} nodes; {} edges.'.format( datasets['test'][0].G.number_of_nodes(), datasets['test'][0].G.number_of_edges())) # node feature dimension input_dim = datasets['train'].num_node_features # link prediction needs 2 classes (0, 1) num_classes = datasets['train'].num_edge_labels model = Net(input_dim, args).to(args.device) optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4) scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs) follow_batch = [] dataloaders = {split: DataLoader( ds, collate_fn=Batch.collate(follow_batch), batch_size=args.batch_size, shuffle=(split=='train')) for split, ds in datasets.items()} if(args.verbose): print('Graphs after split: ') for key, dataloader in dataloaders.items(): for batch in dataloader: print(key, ': ', batch) return train(model, datasets, dataloaders, optimizer, args, scheduler=scheduler) def write_file(file_name, total_acc): print(total_acc) total_acc = total_acc.reshape(1, -1) frame = pd.DataFrame(data=total_acc, columns=['Train', 'Validation', 'Test']) frame.to_csv(file_name, index=False) print(frame) def experiment(iterations, args, file_name_results=None, file_name_edge_message_ratio_changes=None): edge_message_ratio_changes_iterations = [] total_acc = np.zeros(3) for seed in range(1, iterations + 1): random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) acc, edge_message_ratio_changes = run(args) total_acc = np.add(total_acc, acc) edge_message_ratio_changes_iterations.append(edge_message_ratio_changes) total_acc /= iterations if(file_name_results != None): write_file(file_name_results, total_acc) if(file_name_edge_message_ratio_changes != None): edge_message_ratio_changes_iterations = np.array(edge_message_ratio_changes_iterations) data = {"Iteration {}".format(i) : edge_message_ratio_changes_iterations[i - 1, :] for i in range(1, iterations + 1)} frame = pd.DataFrame(data) frame.to_csv(file_name_edge_message_ratio_changes, index=False) def main(): global_args = arg_parse() iterations = 3 device = torch.device('cuda:{}'.format(global_args.gpu) if torch.cuda.is_available() else 'cpu') dataset_name = global_args.dataset verbose = global_args.verbose layer = global_args.num_layers hidden_dim = global_args.hidden_dim data_splits = [[0.8, 0.1, 0.1], [0.2, 0.4, 0.4], [0.5, 0.25, 0.25]] adapt_epochs = [100, 50, 10] try_epochs = [1, 5, -1] criterions = ['val', 'gap'] edge_message_ratios = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] for data_split in data_splits: folder_dataset = "{}-{}-{}".format(dataset_name, layer, hidden_dim) folder_split = "/{}-{}-{}".format(int(100 * data_split[0]), int(100 * data_split[1]), int(100 * data_split[2])) folder_name = folder_dataset + "/results/" + folder_split Path(folder_name).mkdir(parents=True, exist_ok=True) # complete search - constant edge message ratio for edge_message_ratio in edge_message_ratios: args = Arguments(device=device, mode="disjoint", verbose=verbose, dataset=dataset_name, adapt=False, edge_message_ratio=edge_message_ratio, hidden_dim=hidden_dim, layers=layer, data_split=data_split) file_name = folder_name + "/normal_{}.csv".format(int(100 * edge_message_ratio)) experiment(iterations, args, file_name) # random search - changes edge message randomly after every epoch args = Arguments(device=device, mode="disjoint", verbose=verbose, dataset=dataset_name, adapt=False, random=True, hidden_dim=hidden_dim, layers=layer, data_split=data_split) file_name = folder_name + "/random.csv" experiment(iterations, args, file_name) # AdaGrid for criterion in criterions: for adapt_epoch in adapt_epochs: for try_epoch in try_epochs: if(try_epoch == -1): try_epoch = adapt_epoch args = Arguments(device=device, mode="disjoint", verbose=verbose, dataset=dataset_name, adapt=True, adapt_epochs=adapt_epoch, try_epochs=try_epoch, criterion=criterion, hidden_dim=hidden_dim, layers=layer, data_split=data_split) file_name_results = folder_name + "/adapt_{}_{}_{}.csv".format(criterion, adapt_epoch, try_epoch) file_name_edge_message_ratio = folder_name + "/edge_message_ratio_adapt_{}_{}_{}.csv".format(criterion, adapt_epoch, try_epoch) experiment(iterations, args, file_name_results, file_name_edge_message_ratio) if __name__ == '__main__': main()
14,995
30.241667
132
py
adagrid
adagrid-main/community-ratio-based-negative-sampling/adagrid_community_ratio_based_negative_sampling.py
from pathlib import Path import torch.multiprocessing torch.multiprocessing.set_sharing_strategy('file_system') import argparse import copy import math import random import numpy as np import pandas as pd from sklearn.metrics import * import torch import torch.nn.functional as F from torch.utils.data import DataLoader from networkx.algorithms.community import greedy_modularity_communities import networkx as nx from torch_geometric.datasets import Planetoid import torch_geometric.transforms as T import torch_geometric.nn as pyg_nn from deepsnap.dataset import GraphDataset from deepsnap.batch import Batch class Arguments: def __init__(self, dataset='cora', device='cpu', epochs=500, mode='all', model='GCN', edge_message_ratio=0.6, layers=2, hidden_dim=64, batch_size=1, data_split=[0.85, 0.05, 0.1], verbose=False, adapt=False, try_epochs=3, adapt_epochs=50, criterion='val', random=False): self.dataset = dataset self.device = device self.epochs = epochs self.mode = mode self.model = model self.edge_message_ratio = edge_message_ratio self.layers = layers self.hidden_dim = hidden_dim self.batch_size = batch_size self.data_split = data_split self.verbose = verbose self.adapt = adapt self.try_epochs = try_epochs self.adapt_epochs = adapt_epochs self.criterion = criterion self.random = random def arg_parse(): parser = argparse.ArgumentParser(description='Link prediction arguments.') parser.add_argument('--gpu', type=int, help='GPU device.') parser.add_argument('--dataset', type=str, help='Dataset.') parser.add_argument('--num_layers', type=int, help='Number of layers of GNN.') parser.add_argument('--hidden_dim', type=int, help='Hidden dimension of GNN.') parser.add_argument('--verbose', action='store_true', help='Whether output is verbose.') parser.set_defaults( gpu=0, dataset='cora', num_layers=2, hidden_dim=64, verbose=False ) return parser.parse_args() class Net(torch.nn.Module): def __init__(self, input_dim, args): super(Net, self).__init__() self.model = args.model if self.model == 'GCN': self.conv_first = pyg_nn.GCNConv(input_dim, args.hidden_dim) self.convs = torch.nn.ModuleList([pyg_nn.GCNConv(args.hidden_dim, args.hidden_dim) for i in range(args.layers - 2)]) self.conv_last = pyg_nn.GCNConv(args.hidden_dim, args.hidden_dim) else: raise ValueError('unknown conv') self.loss_fn = torch.nn.BCEWithLogitsLoss() def forward(self, graph): x = F.dropout(graph.node_feature, p=0.2, training=self.training) x = F.relu(self._conv_op(self.conv_first, x, graph)) for i in range(len(self.convs)): x = F.dropout(x, p=0.2, training=self.training) x = F.relu(self._conv_op(self.convs[i], x, graph)) x = F.dropout(x, p=0.2, training=self.training) x = self._conv_op(self.conv_last, x, graph) nodes_first = torch.index_select(x, 0, graph.edge_label_index[0,:].long()) nodes_second = torch.index_select(x, 0, graph.edge_label_index[1,:].long()) pred = torch.sum(nodes_first * nodes_second, dim=-1) return pred def _conv_op(self, conv, x, graph): if self.model == 'GCN': return conv(x, graph.edge_index) elif self.model == 'spline': return conv(x, graph.edge_index, graph.edge_feature) def loss(self, pred, link_label): return self.loss_fn(pred, link_label) def sample_negative_community(edge_index, num_nodes, num_neg_edges, community_ratio, communities, from_community): num_within_communities = int(num_neg_edges * community_ratio) num_between_communities = num_neg_edges - num_within_communities # idx = N * i + j idx = (edge_index[0] * num_nodes + edge_index[1]).to("cpu") edges_within = [] edges_between = [] while(len(edges_within) < num_within_communities): x = random.randint(0, num_nodes - 1) comm_x = from_community[x] y_ind = random.randint(0, len(communities[comm_x]) - 1) y = communities[comm_x][y_ind] if(from_community[x] == from_community[y] and (x * num_nodes + y) not in idx): edges_within.append([x, y]) while(len(edges_between) < num_between_communities): x = random.randint(0, num_nodes - 1) y = random.randint(0, num_nodes - 1) if(from_community[x] != from_community[y] and ((x * num_nodes + y) not in idx)): edges_between.append([x, y]) edges_within = np.array(edges_within) edges_between = np.array(edges_between) row = torch.tensor(np.append(edges_within[:, 0], edges_between[:, 0])) col = torch.tensor(np.append(edges_within[:, 1], edges_between[:, 1])) neg_edge_index = torch.stack([row, col], dim=0).long() return neg_edge_index.to(edge_index.device) def sample_community_ratio(graph, community_ratio, communities, from_community): if graph._num_positive_examples is not None: # remove previous negative samples first # if self._num_positive_examples is None then no previous sampling was done graph.edge_label_index = graph.edge_label_index[:, : graph._num_positive_examples] num_pos_edges = graph.edge_label_index.shape[-1] num_neg_edges = num_pos_edges if graph.edge_index.size() == graph.edge_label_index.size() and ( torch.sum(graph.edge_index - graph.edge_label_index) == 0 ): edge_index_all = graph.edge_index else: edge_index_all = ( torch.cat((graph.edge_index, graph.edge_label_index), -1) ) if len(edge_index_all) > 0: negative_edges = sample_negative_community(edge_index_all, graph.num_nodes, num_neg_edges, community_ratio, communities, from_community) else: return torch.LongTensor([]) # label for negative edges is 0 negative_label = torch.zeros(num_neg_edges, dtype=torch.long) # positive edges if graph.edge_label is not None: # when resampling, get the positive portion of labels positive_label = graph.edge_label[:num_pos_edges] elif graph.edge_label is None: # if label is not yet specified, use all ones for positives positive_label = torch.ones(num_pos_edges, dtype=torch.long) else: # reserve class 0 for negatives; increment other edge labels positive_label = graph.edge_label + 1 graph._num_positive_examples = num_pos_edges # append to edge_label_index graph.edge_label_index = ( torch.cat((graph.edge_label_index, negative_edges), -1) ) graph.edge_label = ( torch.cat((positive_label, negative_label), -1).type(torch.long) ) def try_edge_message_ratio(edge_message_ratio, model, datasets, dataloaders, optimizer, args, communities, from_community, scheduler=None): datasets['train'].edge_message_ratio = edge_message_ratio val_max = -math.inf best_model = model mean_accs = {mode: 0 for mode, dataloader in dataloaders.items()} for epoch in range(args.try_epochs): for i in range(len(datasets['train'].graphs)): sample_community_ratio(datasets['train'].graphs[i], datasets['train'].community_ratio, communities, from_community) for iter_i, batch in enumerate(dataloaders['train']): batch.to(args.device) model.train() optimizer.zero_grad() pred = model(batch) loss = model.loss(pred, batch.edge_label.type(pred.dtype)) loss.backward() optimizer.step() if scheduler is not None: scheduler.step() accs, _ = test(model, dataloaders, args) for mode in accs: mean_accs[mode] += accs[mode] if val_max < accs['val']: val_max = accs['val'] best_model = copy.deepcopy(model) for mode in mean_accs: mean_accs[mode] /= args.try_epochs log = 'Edge message ratio: {:.4f}, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}' if(args.verbose): print(log.format(edge_message_ratio, mean_accs['train'], mean_accs['val'], mean_accs['test'])) return mean_accs, val_max, best_model def train(model, datasets, dataloaders, optimizer, args, scheduler=None): graph = datasets['val'][0] G = nx.Graph() G.add_nodes_from(range(graph.num_nodes)) G.add_edges_from([(int(graph.edge_index[0][i]), int(graph.edge_index[1][i])) for i in range(graph.num_edges)]) communities = [list(community) for community in greedy_modularity_communities(G)] from_community = {node: ind for ind in range(len(communities)) for node in communities[ind]} # get community ratio for positive edges in training dataset within = 0 between = 0 for iter_i, batch in enumerate(dataloaders['val']): for i in range(len(batch.edge_label_index[0])): if(batch.edge_label[i] > 0.5): x = batch.edge_label_index[0][i].item() y = batch.edge_label_index[1][i].item() if(from_community[x] == from_community[y]): within += 1 else: between += 1 community_ratio = within / (within + between) datasets['train'].community_ratio = community_ratio datasets['val'].community_ratio = community_ratio datasets['test'].community_ratio = community_ratio for i in range(len(datasets['train'].graphs)): sample_community_ratio(datasets['train'].graphs[i], community_ratio, communities, from_community) for i in range(len(datasets['val'].graphs)): sample_community_ratio(datasets['val'].graphs[i], community_ratio, communities, from_community) for i in range(len(datasets['test'].graphs)): sample_community_ratio(datasets['test'].graphs[i], community_ratio, communities, from_community) # training loop val_max = -math.inf best_model = model input_dim = datasets['train'].num_node_features num_classes = datasets['train'].num_edge_labels waiting = 0 edge_message_ratio_changes = [] for epoch in range(0, args.epochs): if(waiting > 0): waiting -= 1 continue # search to determine the next edge_message_ratio if(args.adapt and epoch % args.adapt_epochs == 0): best_ratio = 0.0 try_val_max = None if(args.criterion == 'val'): try_val_max = -math.inf if(args.criterion == 'gap'): try_val_max = math.inf best_try_model = model best_try_optimizer = optimizer best_try_scheduler = scheduler # best version of the best_try_model during training for try epochs (highest validation accuracy) best_val_model = model best_val = -math.inf edge_message_ratios = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] for edge_message_ratio in edge_message_ratios: new_model = Net(input_dim, args).to(args.device) new_model.load_state_dict(model.state_dict()) new_optimizer = torch.optim.SGD(new_model.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4) new_optimizer.load_state_dict(optimizer.state_dict()) new_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(new_optimizer, T_max=args.epochs) new_scheduler.load_state_dict(scheduler.state_dict()) try_accs, current_val, current_val_model = try_edge_message_ratio(edge_message_ratio, new_model, datasets, dataloaders, new_optimizer, args, communities, from_community, new_scheduler) if (args.criterion == 'val' and try_val_max < try_accs['val']): try_val_max = try_accs['val'] best_ratio = edge_message_ratio best_try_model = new_model best_try_optimizer = new_optimizer best_try_scheduler = new_scheduler best_val = current_val best_val_model = current_val_model if (args.criterion == 'gap' and try_val_max > abs(try_accs['train'] - try_accs['val'])): try_val_max = abs(try_accs['train'] - try_accs['val']) best_ratio = edge_message_ratio best_try_model = new_model best_try_optimizer = new_optimizer best_try_scheduler = new_scheduler best_val = current_val best_val_model = current_val_model model = best_try_model optimizer = best_try_optimizer scheduler = best_try_scheduler datasets['train'].edge_message_ratio = best_ratio edge_message_ratio_changes.append(datasets['train'].edge_message_ratio) if val_max < best_val: val_max = best_val best_model = best_val_model waiting = args.try_epochs - 1 continue if(args.random): datasets['train'].edge_message_ratio = random.uniform(0.1, 0.9) for i in range(len(datasets['train'].graphs)): sample_community_ratio(datasets['train'].graphs[i], datasets['train'].community_ratio, communities, from_community) for iter_i, batch in enumerate(dataloaders['train']): batch.to(args.device) model.train() optimizer.zero_grad() pred = model(batch) loss = model.loss(pred, batch.edge_label.type(pred.dtype)) loss.backward() optimizer.step() if scheduler is not None: scheduler.step() log = 'Epoch: {:03d}, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}' accs, _ = test(model, dataloaders, args) if(args.verbose): print(log.format(epoch, accs['train'], accs['val'], accs['test'])) if val_max < accs['val']: val_max = accs['val'] best_model = copy.deepcopy(model) log = 'Best, Train: {:.4f}, Val: {:.4f}, Test: {:.4f}' accs, _ = test(best_model, dataloaders, args) print(log.format(accs['train'], accs['val'], accs['test'])) return np.array([accs['train'], accs['val'], accs['test']]), edge_message_ratio_changes def test(model, dataloaders, args, max_train_batches=1): model.eval() accs = {} losses = {} for mode, dataloader in dataloaders.items(): acc = 0 loss = 0 num_batches = 0 for batch in dataloader: batch.to(args.device) pred = model(batch) # only 1 graph in dataset. In general needs aggregation loss += model.loss(pred, batch.edge_label.type(pred.dtype)).cpu().data.numpy() acc += roc_auc_score(batch.edge_label.flatten().cpu().numpy(), pred.flatten().data.cpu().numpy()) num_batches += 1 if mode == 'train' and num_batches >= max_train_batches: # do not eval on the entire training set for efficiency break accs[mode] = acc / num_batches losses[mode] = loss / num_batches return accs, losses def run(args): pyg_dataset = None if(args.dataset == 'cora'): pyg_dataset = Planetoid('./datasets', 'Cora', transform=T.TargetIndegree()) if(args.dataset == 'citeseer'): pyg_dataset = Planetoid('./datasets', 'CiteSeer', transform=T.TargetIndegree()) if(args.dataset == 'pubmed'): pyg_dataset = Planetoid('./datasets', 'PubMed', transform=T.TargetIndegree()) edge_train_mode = args.mode if(args.verbose): print('edge train mode: {}'.format(edge_train_mode)) graphs = GraphDataset.pyg_to_graphs(pyg_dataset) dataset = GraphDataset(graphs, task='link_pred', edge_message_ratio=args.edge_message_ratio, edge_train_mode=edge_train_mode, resample_disjoint=True, resample_disjoint_period=1, resample_negatives=False) if(args.verbose): print('Initial dataset: {}'.format(dataset)) # split dataset datasets = {} datasets['train'], datasets['val'], datasets['test'] = dataset.split( transductive=True, split_ratio=args.data_split) if(args.verbose): print('after split') print('Train message-passing graph: {} nodes; {} edges.'.format( datasets['train'][0].G.number_of_nodes(), datasets['train'][0].G.number_of_edges())) print('Val message-passing graph: {} nodes; {} edges.'.format( datasets['val'][0].G.number_of_nodes(), datasets['val'][0].G.number_of_edges())) print('Test message-passing graph: {} nodes; {} edges.'.format( datasets['test'][0].G.number_of_nodes(), datasets['test'][0].G.number_of_edges())) # node feature dimension input_dim = datasets['train'].num_node_features # link prediction needs 2 classes (0, 1) num_classes = datasets['train'].num_edge_labels model = Net(input_dim, args).to(args.device) optimizer = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9, weight_decay=5e-4) scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=args.epochs) follow_batch = [] dataloaders = {split: DataLoader( ds, collate_fn=Batch.collate(follow_batch), batch_size=args.batch_size, shuffle=(split=='train')) for split, ds in datasets.items()} if(args.verbose): print('Graphs after split: ') for key, dataloader in dataloaders.items(): for batch in dataloader: print(key, ': ', batch) return train(model, datasets, dataloaders, optimizer, args, scheduler=scheduler) def write_file(file_name, total_acc): print(total_acc) total_acc = total_acc.reshape(1, -1) frame = pd.DataFrame(data=total_acc, columns=['Train', 'Validation', 'Test']) frame.to_csv(file_name, index=False) print(frame) def experiment(iterations, args, file_name_results=None, file_name_edge_message_ratio_changes=None): edge_message_ratio_changes_iterations = [] total_acc = np.zeros(3) for seed in range(1, iterations + 1): random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) acc, edge_message_ratio_changes = run(args) total_acc = np.add(total_acc, acc) edge_message_ratio_changes_iterations.append(edge_message_ratio_changes) total_acc /= iterations if(file_name_results != None): write_file(file_name_results, total_acc) if(file_name_edge_message_ratio_changes != None): edge_message_ratio_changes_iterations = np.array(edge_message_ratio_changes_iterations) data = {"Iteration {}".format(i) : edge_message_ratio_changes_iterations[i - 1, :] for i in range(1, iterations + 1)} frame = pd.DataFrame(data) frame.to_csv(file_name_edge_message_ratio_changes, index=False) def main(): global_args = arg_parse() iterations = 3 device = torch.device('cuda:{}'.format(global_args.gpu) if torch.cuda.is_available() else 'cpu') dataset_name = global_args.dataset verbose = global_args.verbose layer = global_args.num_layers hidden_dim = global_args.hidden_dim data_splits = [[0.8, 0.1, 0.1], [0.5, 0.25, 0.25], [0.2, 0.4, 0.4]] adapt_epochs = [100, 50, 10] try_epochs = [1, 5, -1] criterions = ['val', 'gap'] edge_message_ratios = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9] for data_split in data_splits: folder_dataset = "{}-{}-{}".format(dataset_name, layer, hidden_dim) folder_split = "/{}-{}-{}".format(int(100 * data_split[0]), int(100 * data_split[1]), int(100 * data_split[2])) folder_name = folder_dataset + "/results/" + folder_split Path(folder_name).mkdir(parents=True, exist_ok=True) # complete search - constant edge message ratio for edge_message_ratio in edge_message_ratios: args = Arguments(device=device, mode="disjoint", verbose=verbose, dataset=dataset_name, adapt=False, edge_message_ratio=edge_message_ratio, hidden_dim=hidden_dim, layers=layer, data_split=data_split) file_name = folder_name + "/normal_{}.csv".format(int(100 * edge_message_ratio)) experiment(iterations, args, file_name) # random search - changes edge message randomly after every epoch args = Arguments(device=device, mode="disjoint", verbose=verbose, dataset=dataset_name, adapt=False, random=True, hidden_dim=hidden_dim, layers=layer, data_split=data_split) file_name = folder_name + "/random.csv" experiment(iterations, args, file_name) # AdaGrid for criterion in criterions: for adapt_epoch in adapt_epochs: for try_epoch in try_epochs: if(try_epoch == -1): try_epoch = adapt_epoch args = Arguments(device=device, mode="disjoint", verbose=verbose, dataset=dataset_name, adapt=True, adapt_epochs=adapt_epoch, try_epochs=try_epoch, criterion=criterion, hidden_dim=hidden_dim, layers=layer, data_split=data_split) file_name_results = folder_name + "/adapt_{}_{}_{}.csv".format(criterion, adapt_epoch, try_epoch) file_name_edge_message_ratio = folder_name + "/edge_message_ratio_adapt_{}_{}_{}.csv".format(criterion, adapt_epoch, try_epoch) experiment(iterations, args, file_name_results, file_name_edge_message_ratio) if __name__ == '__main__': main()
19,954
30.979167
132
py
SMERF
SMERF-main/scripts/adv_train.py
""" This is an example of how to use ART and Keras to perform adversarial training using data generators for CIFAR10 """ import keras import numpy as np from keras.layers import Conv2D, Dense, Flatten, MaxPooling2D, Input, BatchNormalization from keras.models import Model from keras.preprocessing.image import ImageDataGenerator from keras.regularizers import l2 from keras.utils.np_utils import to_categorical import os import keras.backend as K from keras.callbacks import EarlyStopping, ModelCheckpoint from keras.preprocessing import image from keras.preprocessing.image import ImageDataGenerator from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPool2D, ZeroPadding2D from keras.models import Sequential, Model from keras.layers.core import Flatten, Dense, Dropout from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D import sys sys.path.append('../') import smerf import argparse from art.attacks.evasion import ProjectedGradientDescent, FastGradientMethod from art.classifiers import KerasClassifier from art.data_generators import KerasDataGenerator from art.defences.trainer import AdversarialTrainer import time # Directory to save the data DATA_DIR = '../data' # Directory to save intermediary/final results CACHE_DIR = '../outputs/cache' if not os.path.exists(CACHE_DIR): os.mkdir(CACHE_DIR) # Directory to save plots PLOT_DIR = '../outputs/plots' if not os.path.exists(PLOT_DIR): os.mkdir(PLOT_DIR) def build_simple_model(input_shape=(64,64,3), interm_dim=200, lr=0.0001): model = keras.Sequential([ keras.layers.Conv2D(filters=32, kernel_size=3, strides=(2, 2), activation='relu', input_shape=input_shape), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Flatten(), keras.layers.Dense(interm_dim, activation='relu'), keras.layers.Dense(2) ]) opt = keras.optimizers.Adam(lr=lr) model.compile(optimizer=opt, loss='binary_crossentropy', metrics=['accuracy']) return model def build_complex_model(input_shape=(64,64,3), interm_dim=200, lr=0.0001): model = keras.Sequential([ keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu', input_shape=input_shape), keras.layers.Conv2D(filters=128, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=256, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Flatten(), keras.layers.Dense(interm_dim, activation='relu'), keras.layers.Dense(interm_dim, activation='relu'), keras.layers.Dense(2) ]) opt = keras.optimizers.Adam(lr=lr) model.compile(optimizer=opt, loss='binary_crossentropy', metrics=['accuracy']) return model def main(args): exp_no = args.exp # experiment number attack_type = args.attack # attack type epochs = args.ep batch_size = args.batch print('EXP_NO = %f'%exp_no) lr = args.lr # learning rate used to train the model model_name = 'w-adv2-%0.2f.pt'%exp_no # model file to save/load train_type = args.type if exp_no == 1.11: import smerf.simple_fr as textbox_exp no_data = 2000 elif exp_no == 2.11: import smerf.simple_nr as textbox_exp no_data = 5000 elif exp_no == 1.2: import smerf.complex_fr as textbox_exp no_data = 2000 elif exp_no == 3.71: # Complex-CR1 import smerf.complex_cr1 as textbox_exp no_data = 15000 elif exp_no == 3.72: # Complex-CR2 import smerf.complex_cr2 as textbox_exp no_data = 15000 elif exp_no == 3.73: # Complex-CR3 import smerf.complex_cr3 as textbox_exp no_data = 15000 elif exp_no == 3.74: # Complex-CR4 import smerf.complex_cr4 as textbox_exp no_data = 15000 ### Generate (or load) datasets train_data, test_data, train_primary, test_primary, train_secondary, test_secondary = \ textbox_exp.generate_textbox_data(n=no_data, save=True, save_dir='../data', exp_no=exp_no, random_bg=0) x_train = train_data.X x_test = test_data.X y_train = train_data.y y_test = test_data.y original_name = 'w%0.2f.pt'%exp_no y_train_oh = to_categorical(y_train, 2) y_test_oh = to_categorical(y_test, 2) print('data loaded') datagen = ImageDataGenerator() datagen.fit(x_train) art_datagen = KerasDataGenerator(datagen.flow(x=x_train, y=y_train_oh, batch_size=batch_size, shuffle=True), size=x_train.shape[0], batch_size=batch_size) print('generator fit') if exp_no == 1.11 or exp_no == 2.11: model = build_simple_model(lr=lr) #if train_type != 'scratch': model.load_weights(os.path.join(CACHE_DIR, original_name)) classifier = KerasClassifier(model, clip_values=(0, 1), use_logits=False) else: model = build_complex_model(lr=lr) #if train_type != 'scratch': model.load_weights(os.path.join(CACHE_DIR, original_name)) classifier = KerasClassifier(model, clip_values=(0, 1), use_logits=False) # Create attack for adversarial trainer; here, we use 2 attacks, both crafting adv examples on the target model print('Creating Attack') if attack_type == 'pgd': attacker = ProjectedGradientDescent(classifier, eps=0.3, eps_step=0.1, max_iter=10, num_random_init=1) elif attack_type == 'fgsm': attacker = FastGradientMethod(classifier, eps=0.2) else: raise ValueError() # Create advareasial samples if os.path.exists('x_train2_%s_%0.2f.npy'%(attack_type, exp_no)): x_train_pgd = np.load('x_train2_%s_%0.2f.npy'%(attack_type, exp_no)) x_test_pgd = np.load('x_test2_%s_%0.2f.npy'%(attack_type, exp_no)) else: x_test_pgd = attacker.generate(x_test) x_train_pgd = attacker.generate(x_train) np.save(open('x_train2_%s_%0.2f.npy'%(attack_type, exp_no), 'wb'), x_train_pgd) np.save(open('x_test2_%s_%0.2f.npy'%(attack_type, exp_no), 'wb'), x_test_pgd) print(x_test_pgd.shape) preds = np.argmax(classifier.predict(x_test_pgd), axis=1) acc = np.sum(preds == np.argmax(y_test_oh, axis=1)) / y_test.shape[0] print("Classifier before adversarial training") print("Accuracy on adversarial samples: %.2f%%", (acc * 100)) # Create adversarial trainer and perform adversarial training print('Training') if train_type == 'aug': # augment the data x_train = np.append(x_train, x_train_pgd, axis=0) y_train = np.append(y_train, y_train, axis=0) y_train_oh = np.append(y_train_oh, y_train_oh, axis=0) #retrain the model with augmented data model.compile(optimizer=keras.optimizers.Adam(lr=lr), loss='binary_crossentropy', metrics=['accuracy']) classifier.fit(x_train, y_train_oh, nb_epochs=epochs, batch_size=batch_size, verbose=True) classifier._model.save_weights(os.path.join(CACHE_DIR, model_name)) elif train_type == 'scratch': adv_trainer = AdversarialTrainer(classifier, attacks=attacker, ratio=1.) adv_trainer.fit_generator(art_datagen, nb_epochs=epochs) model = classifier._model model.save_weights(os.path.join(CACHE_DIR, model_name)) # load the pretrained model #model.load_weights(os.path.join(CACHE_DIR, model_name)) #classifier = KerasClassifier(model, clip_values=(0, 1), use_logits=False) # adv_trainer = AdversarialTrainer(classifier, attacks=attacker, ratio=1.) # adv_trainer.fit_generator(art_datagen, nb_epochs=20) # model = classifier._model # model.save_weights(os.path.join(CACHE_DIR, model_name)) # Evaluate the adversarially trained model on clean test set labels_true = np.argmax(y_test_oh, axis=1) labels_test = np.argmax(classifier.predict(x_test), axis=1) print('Accuracy test set: %.2f%%' % (np.sum(labels_test == labels_true) / x_test.shape[0] * 100)) # Evaluate the adversarially trained model on original adversarial samples labels_pgd = np.argmax(classifier.predict(x_test_pgd), axis=1) print('Accuracy on original PGD adversarial samples: %.2f%%' % (np.sum(labels_pgd == labels_true) / x_test.shape[0] * 100)) # Evaluate the adversarially trained model on fresh adversarial samples produced on the adversarially trained model x_test_pgd = attacker.generate(x_test) labels_pgd = np.argmax(classifier.predict(x_test_pgd), axis=1) print('Accuracy on new PGD adversarial samples: %.2f%%' % (np.sum(labels_pgd == labels_true) / x_test.shape[0] * 100)) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--exp', type=float, help='experiment number. 1-FR, 2-NR, 3-CR') parser.add_argument('--ep', type=int, default=10, help='max epoch') parser.add_argument('--lr', type=float, default=0.0001, help='learning rate') parser.add_argument('--attack', type=str, default='fgsm', help='attack method') parser.add_argument('--batch', type=int, default=256, help='batch size') parser.add_argument('--type', type=str, default='scratch', help='train type') args = parser.parse_args() print(args) main(args)
9,861
43.224215
128
py
SMERF
SMERF-main/smerf/explanations.py
import numpy as np import imp textcolorutils = imp.load_source('textcolor_utils', '../smerf/textcolor_utils.py') import innvestigate import keras import keras.backend as K import tensorflow as tf import cv2 import pickle import os from tqdm import tqdm # NOTE Helper functions for saliency methods that are not supported by iNNvestigate libary # We recommend specifying these methods on a separate file to avoid clutter. from .grad_cam_utils import * from .shap_utils import * from .lime_utils import * def softmax(x): return np.exp(x) / np.sum(np.exp(x), axis=1)[:, None] def run_methods(model, x_data, y_data, x_train, no_images=10, exp_no=0, load=True, f_name=None, directory='../outputs/cache', split=None, model_type=0): """ Given a trained model, run saliency methods and return the results. :param model: keras sequntial model to be explained :param x_data: (N, H, W, C) image dataset :param y_data: (N,) image labels :param no_images: number of sample images to run methods on (n) :param exp_no: experiment number for reference :param load: if True, load from a cached results :param directory: directory to save or load the results :param split: if not None, specifies the bucket index the methods are run on :param model_type: 0 (default) for simple CNNs. 1 for VGG16. 2 for ResNet50. :return (result, method, text, idx) tuple where result is (no_images, M, H, C, W) saliency output for M methods and no_images images; method is the list of configs for the saliency methods; text is the prediction information on the images; idx is the original index of the sampled images from the dataset """ model_in = keras.models.Model(inputs=model.inputs, outputs=model.outputs) if split is None: result_name = os.path.join(directory, 'result_%0.2f.pkl'%exp_no) idx_name = os.path.join(directory, 'idx_%0.2f.pkl'%exp_no) methods_name = os.path.join(directory, 'methods_%0.2f.pkl'%exp_no) text_name = os.path.join(directory, 'text_%0.2f.pkl'%exp_no) else: result_name = os.path.join(directory, 'result_%0.2f_%d.pkl'%(exp_no, split)) idx_name = os.path.join(directory, 'idx_%0.2f_%d.pkl'%(exp_no, split)) methods_name = os.path.join(directory, 'methods_%0.2f_%d.pkl'%(exp_no, split)) text_name = os.path.join(directory, 'text_%0.2f_%d.pkl'%(exp_no, split)) if os.path.exists(result_name) and os.path.exists(idx_name) and load==True: loaded = True print('loading results from cache in {}'.format(directory)) result = pickle.load(open(result_name, 'rb')) idx = pickle.load(open(idx_name, 'rb')) methods = pickle.load(open(methods_name, 'rb')) text = pickle.load(open(text_name, 'rb')) else: loaded = False if f_name is not None: result_name = os.path.join(directory, f_name+'.pkl') idx_name = os.path.join(directory, f_name+'_idx.pkl') methods_name = os.path.join(directory, f_name+'_methods.pkl') text_name = os.path.join(directory, f_name+'_text.pkl') print('cache not found') noise_scale = 0.1 input_range = (0,1) ## NOTE Methods defined here is using the iNNvestigate library. ## Other methods should be added manually via separately defined helper functions at the end. methods = [ # NAME OPT.PARAMS POSTPROC FXN TITLE # Show input. ("input", {}, textcolorutils.identity, "Input"), # Function ("gradient", {"postprocess": "abs"}, textcolorutils.graymap, "Gradient"), ("smoothgrad", {"augment_by_n": 64, "noise_scale": noise_scale, "postprocess": "square"},textcolorutils.graymap, "SmoothGrad"), # Signal ("deconvnet", {}, textcolorutils.bk_proj, "Deconvnet"), ("guided_backprop", {}, textcolorutils.bk_proj, "Guided Backprop",), # Interaction ("deep_taylor.bounded", {"low": input_range[0], "high": input_range[1]}, textcolorutils.heatmap, "DeepTaylor"), ("input_t_gradient", {}, textcolorutils.heatmap, "Input * Gradient"), ("integrated_gradients", {"reference_inputs": input_range[0], "steps": 64}, textcolorutils.heatmap, "Integrated Gradients"), ("lrp.z", {}, textcolorutils.heatmap, "LRP-Z"), ("lrp.epsilon", {"epsilon": 1}, textcolorutils.heatmap, "LRP-Epsilon"), ("lrp.sequential_preset_a_flat",{"epsilon": 1}, textcolorutils.heatmap, "LRP-PresetAFlat"), ("lrp.sequential_preset_b_flat",{"epsilon": 1}, textcolorutils.heatmap, "LRP-PresetBFlat"), ("deep_lift.wrapper", {"nonlinear_mode":"reveal_cancel", "reference_inputs": 0, "verbose": 0}, textcolorutils.heatmap, "DeepLIFT-RevealCancel"), ("deep_lift.wrapper", {"nonlinear_mode":"rescale", "reference_inputs": 0, "verbose":0}, textcolorutils.heatmap, "DeepLIFT-Rescale"), ] # Create analyzers. analyzers = [] for method in methods: try: analyzer = innvestigate.create_analyzer(method[0], # analysis method identifier model_in, # model without softmax output **method[1]) # optional analysis parameters except innvestigate.NotAnalyzeableModelException: # Not all methods work with all models. analyzer = None analyzers.append(analyzer) # Run the saliency methods text = [] label_to_class_name = {0: 'Neg', 1: 'Pos'} color_conversion = None channels_first = keras.backend.image_data_format() == "channels_first" # random set of images to test np.random.seed(1) idx = np.random.choice(len(x_data), no_images, replace=False) # index of the test data selected images = x_data[idx] labels = y_data[idx] h, w, c = images[0].shape result = np.zeros((len(images), len(analyzers), h, w, c)) # Run methods on batch for aidx, analyzer in enumerate(tqdm(analyzers)): if methods[aidx][0] == "input": # Do not analyze, but keep not preprocessed input. a = images elif analyzer: if model_type == 1 or model_type ==2: # NOTE for fine-tuned model, deep lift does not support maxpooling2d if 'deep_lift' not in methods[aidx][0]: # Analyze. a = analyzer.analyze(images) # Apply common postprocessing, e.g., re-ordering the channels for plotting. a = textcolorutils.postprocess(a, color_conversion, channels_first) # Apply analysis postprocessing, e.g., creating a heatmap. a = methods[aidx][2](a) else: print('deepLift does not support specific layers') a = np.zeros_like(images) else: # Analyze. a = analyzer.analyze(images) # Apply common postprocessing, e.g., re-ordering the channels for plotting. a = textcolorutils.postprocess(a, color_conversion, channels_first) # Apply analysis postprocessing, e.g., creating a heatmap. a = methods[aidx][2](a) else: a = np.zeros_like(images) # Store the analysis. result[:, aidx] = a # Predict final activations, probabilites, and label. presm = model.predict(images) prob = softmax(presm) y_hat = prob.argmax(axis=1) for i, y in enumerate(labels): # Save prediction info: text.append(("%s" % label_to_class_name[y], # ground truth label "%.2f" % presm.max(axis=1)[i], # pre-softmax logits "%.2f" % prob.max(axis=1)[i], # probabilistic softmax output "%s" % label_to_class_name[y_hat[i]] # predicted label )) ####### NOTE Add additional methods to run below (those that are not supported ####### in the iNNvestigate library) # Add Grad-CAM print(' Running Grad-CAM') result, methods = add_grad_cam(result, methods, model, images, exp_no, directory, model_type) # Add SHAP print(' Running SHAP') result, methods = add_shap(result, methods, model, images, labels, x_train, exp_no, model_type) # # Add LIME # print(' Running LIME') # result, methods = add_lime(result, methods, model, images, labels, x_train, exp_no) # random baseline print(' Running random') random_results = np.random.random((len(images), 1, h, w, c)) methods.append(('random',{}, textcolorutils.heatmap, "Random")) # edge detector print(' Running edge detection') edge_results = np.zeros((len(images), 1, h, w, c)) for i, x in enumerate(images): ed = cv2.Sobel(x, cv2.CV_64F,1,0,ksize=5) ed = (ed - np.min(ed)) / (np.max(ed) - np.min(ed)) edge_results[i,0] = ed methods.append(('edge', {}, textcolorutils.heatmap, "Edge-detection")) result = np.concatenate((result, random_results, edge_results), axis=1) ###### Adding new methods should be completed above ###### if not loaded: # Save results pickle.dump(result, open(result_name, 'wb')) pickle.dump(idx, open(idx_name, 'wb')) pickle.dump(methods, open(methods_name, 'wb')) pickle.dump(text, open(text_name, 'wb')) ## saliency results: size (no_samples, methods, img_size) print(' No-images: %d \t No-methods: %d finished.'%(result.shape[0], result.shape[1])) return result, methods, text, idx #### Helper functions for additional explanation methods to include #### #### NOTE Add additional explanation methods below #### Define methods that will take in the result matrix containing outputs from other methods #### and add the new method's result to return. #### The result matrix has a shape (no_samples, no_methods, h, w, c), so the output of the new methods #### defined below should respect this format and be concatenated to the existing result matrix #### along axis=1 (no_methods). #### Below shows examples of adding two new methods (GradCAM and DeepSHAP) and basic steps that #### should be followed within these functions, what should be returned in the end. # Function for adding GradCAM results to the existing results. def add_grad_cam(result, methods, model, images, exp_no, directory, model_type): # compute attributions if model_type == -1: # adversarial model_name = 'w-adv-%0.2f.pt'%exp_no elif model_type == 0: model_name = 'w%0.2f.pt'%exp_no elif model_type == 1: model_name = 'w_vgg%0.2f.pt'%exp_no elif model_type == 2: model_name = 'w_alex%0.2f.pt'%exp_no else: raise ValueError('model_type not supported') c, h, g = grad_cam_run(model, images, os.path.join(directory, model_name), exp_no, model_type) # add the new results to the existing results added_result = np.expand_dims(h, 1) result = np.concatenate((result, added_result), axis=1) # add the new method information to the existing method information methods.append(('grad-cam', {}, textcolorutils.graymap, "Grad-CAM")) # return both result and method information return result, methods # Function for adding DeepSHAP results to the existing results. def add_shap(result, methods, model, images, labels, x_train, exp_no, model_type): # compute attributions output = shap_run(model, images, labels, x_train, exp_no, model_type) h, w, c = images[0].shape assert(output.shape == (images.shape[0], 1, h, w, c)) # add the new results to the existing results result = np.concatenate((result, output), axis=1) # add the new method information to the existing method information methods.append(('deep-shap', {}, textcolorutils.graymap, "DeepSHAP")) # return both result and method information return result, methods def add_lime(result, methods, model, images, labels, x_train, exp_no): # compute attributions output = lime_run(model, images, labels, x_train, exp_no) h, w, c = images[0].shape assert(output.shape == (images.shape[0], 1, h, w, c)) # add the new results to the existing results result = np.concatenate((result, output), axis=1) # add the new method information to the existing method information methods.append(('lime', {}, textcolorutils.graymap, "LIME")) # return both result and method information return result, methods # New method to be added to the pipeline def add_your_new_method(result, methods, model, images, **kwargs): # TODO fill in the code for your new method to be added. return result, methods
13,998
45.97651
156
py
SMERF
SMERF-main/smerf/grad_cam_utils.py
""" Source code adapted from https://github.com/wawaku/grad-cam-keras """ from keras.preprocessing import image from tensorflow.python.framework import ops import keras.backend as K import tensorflow as tf import numpy as np import keras import cv2 import os, gc from .models import TextBoxCNN as TextBoxCNN from .models import TextBoxCNN_adv as TextBoxCNN_adv from .models import VGG16_model as VGG16_model from .models import AlexNet_model as AlexNet_model def register_gradient(): if "GuidedBackProp" not in ops._gradient_registry._registry: @ops.RegisterGradient("GuidedBackProp") def _GuidedBackProp(op, grad): dtype = op.inputs[0].dtype return grad * tf.cast(grad > 0., dtype) * \ tf.cast(op.inputs[0] > 0., dtype) def compile_saliency_function(model, activation_layer='block5_pool'): input_img = model.input layer_dict = dict([(layer.name, layer) for layer in model.layers[1:]]) layer_output = layer_dict[activation_layer].output max_output = K.max(layer_output, axis=3) saliency = K.gradients(K.sum(max_output), input_img)[0] return K.function([input_img, K.learning_phase()], [saliency]) def modify_backprop(model, name, model_file, exp_no, model_type): g = tf.get_default_graph() with g.gradient_override_map({'Relu': name}): # get layers that have an activation layer_dict = [layer for layer in model.layers[1:] if hasattr(layer, 'activation')] # replace relu activation for layer in layer_dict: if layer.activation == keras.activations.relu: layer.activation = tf.nn.relu # re-instanciate a new model if model_type == 0 or model_type == -1: # adversarial model if exp_no >= 3.5 or exp_no == 1.2: new_model = TextBoxCNN_adv().model else: new_model = TextBoxCNN().model elif model_type == 1: new_model = VGG16_model().model elif model_type == 2: new_model = AlexNet_model().model else: raise ValueError('model_type must be 0, 1, or 2') new_model.load_weights(model_file) return new_model def deprocess_image(x): ''' Same normalization as in: https://github.com/fchollet/keras/blob/master/examples/conv_filter_visualization.py ''' if np.ndim(x) > 3: x = np.squeeze(x) # normalize tensor: center on 0., ensure std is 0.1 x -= x.mean() x /= (x.std() + 1e-5) x *= 0.1 # clip to [0, 1] x += 0.5 x = np.clip(x, 0, 1) # convert to RGB array x *= 255 if K.image_dim_ordering() == 'th': x = x.transpose((1, 2, 0)) x = np.clip(x, 0, 255).astype('uint8') return x def grad_cam(model, x, category_index, layer_name): """ Args: model: model x: image input category_index: category index layer_name: last convolution layer name """ # get category loss class_output = model.output[:, category_index] # layer output convolution_output = model.get_layer(layer_name).output # get gradients grads = K.gradients(class_output, convolution_output)[0] # get convolution output and gradients for input gradient_function = K.function([model.input], [convolution_output, grads]) output, grads_val = gradient_function([x]) output, grads_val = output[0], grads_val[0] # avg weights = np.mean(grads_val, axis=(0, 1)) cam = np.dot(output, weights) # create heat map cam = cv2.resize(cam, (x.shape[1], x.shape[2]), cv2.INTER_LINEAR) cam = np.maximum(cam, 0) heatmap = cam / np.max(cam) # Return to BGR [0..255] from the preprocessed image image_rgb = x[0, :] image_rgb -= np.min(image_rgb) image_rgb = np.minimum(image_rgb, 255) cam = cv2.applyColorMap(np.uint8(255 * heatmap), cv2.COLORMAP_JET) cam = np.float32(cam) + np.float32(image_rgb) cam = 255 * cam / np.max(cam) return np.uint8(cam), heatmap def grad_cam_run(model, x_sample, model_file, exp_no, model_type): last_conv_layer_name = [x for x in model.layers if type(x) == keras.layers.convolutional.Conv2D][-1].name cam_imgs = np.zeros(x_sample.shape) heat_maps = np.zeros(x_sample.shape) grad_cam_imgs = np.zeros(x_sample.shape) for i in range(x_sample.shape[0]): img = x_sample[i][None,:,:,:] predictions = model.predict(img) pred_class = predictions.argmax(axis=1)[0] cam_image, heat_map = grad_cam(model, img, pred_class, last_conv_layer_name) # guided grad_cam img register_gradient() guided_model = modify_backprop(model, 'GuidedBackProp', model_file, exp_no, model_type) guided_model_name = [x for x in guided_model.layers if type(x) == keras.layers.convolutional.Conv2D][-1].name saliency_fn = compile_saliency_function(guided_model, activation_layer=guided_model_name) saliency = saliency_fn([img, 0]) grad_cam_img = saliency[0] * heat_map[..., np.newaxis] if np.max(grad_cam_img) - np.min(grad_cam_img) != 0: grad_cam_img = (grad_cam_img - np.min(grad_cam_img)) / (np.max(grad_cam_img) - np.min(grad_cam_img)) cam_imgs[i] = cam_image heat_maps[i] = np.repeat(heat_map[:, :, np.newaxis], 3, axis=2) grad_cam_imgs[i] = grad_cam_img[0] del guided_model gc.collect() return cam_imgs, heat_maps, grad_cam_imgs
5,486
33.727848
117
py
SMERF
SMERF-main/smerf/shap_utils.py
import shap import tensorflow as tf import keras.backend as K import numpy as np from smerf.models import * import gc # Uses SHAP library to obtain feature attributions def shap_run(model, x_sample, y_sample, x_train, exp_no, model_type): ## NOTE due to complications in keras and TF versions (this code works only in TF1), ## a separate model should be redefined on a separate session where SHAP is run. model.save_weights('/tmp/model.pt') background = x_train[np.random.choice(x_train.shape[0], 1000, replace=False)] n, h, w, c = x_sample.shape output = np.zeros((n, 1, h, w, c)) with tf.Session() as sess: sess.run(tf.global_variables_initializer()) # redefine the model if model_type == 0 or model_type ==-1: # adversarial model if exp_no >= 3.5 or exp_no==1.2: model_obj = TextBoxCNN_adv(lr=0.0001, max_epoch=10) else: model_obj = TextBoxCNN(lr=0.0001, max_epoch=10) elif model_type == 1: model_obj = VGG16_model() elif model_type == 2: model_obj = AlexNet_model() else: raise ValueError('model_type must be 0, 1, or 2') model_obj.model.load_weights('/tmp/model.pt') model_sess = model_obj.model # DeepSHAP e = shap.DeepExplainer(model_sess, background) shap_vals_deep = e.shap_values(x_sample) #shap_vals_deep = np.array([shap_vals_deep[y_sample[i]][i] for i in range(n)]) shap_vals_deep = np.array([scale(shap_vals_deep[y_sample[i]][i]) for i in range(n)]) shap_vals_deep = np.max(shap_vals_deep, axis=3) shap_vals_deep = np.expand_dims(shap_vals_deep, 3) shap_vals_deep = np.concatenate((shap_vals_deep, shap_vals_deep, shap_vals_deep), axis=3) output[:, 0, :, :, :] = shap_vals_deep return output def scale(x): if x.max() - x.min() != 0: return (x - x.min()) / (x.max() - x.min()) return x
1,980
40.270833
97
py
SMERF
SMERF-main/smerf/textbox_data.py
import numpy as np import PIL from PIL import Image, ImageDraw, ImageFont, ImageEnhance from torch.utils.data import Dataset, DataLoader import os from smerf.eval import setup_bboxes import pickle DATA_DIR = '../data/' class TextBoxDataset(Dataset): def __init__(self, X, y): self.X = X self.y = y def __len__(self): return len(self.y) def __getitem__(self, idx): return self.X[idx], self.y[idx] # Interpolate between blue (v = 0) and red (v = 1) def shade(im, v): if v == -1: im[:, :, :] = 255 #plain white background elif v == -2: im[:,:,:] = 0 # plain black background elif v == -3: im[:,:,:] = np.asarray(np.random.random((64,64,3)) * 100, dtype=int) # random gray background elif v == -4: # natural image background places_img_file = pickle.load(open(os.path.join(DATA_DIR, 'places_img_file.pkl'), 'rb')) choices = places_img_file['stadium/baseball'] img_ids = [0, 9, 10, 12, 15, 16, 17, 19, 20, 21, 24, 25, 26, 27, 28, 33, 34, 35, 37, 39, 41, 42, 43, 45, 46, 47, 49, 50, 51, 52, 55, 56, 58, 64, 65, 58, 68, 71, 74, 78, 86, 88, 90, 91, 92, 93] #choices = places_img_file['bamboo_forest'] #img_ids = [0, 1, 2, 3, 4, 6, 7, 8, 10, 11, 12, 13, 14, 15, 16, 17, 19, 20, 21, 25, 28, 37, 44, 47, 57, 59, 65, 68, 69, 72, 75, 77, 85, 93, 96, 98, 99] img_dir = os.path.join(DATA_DIR, 'val_256') img = Image.open(os.path.join(img_dir, choices[np.random.choice(img_ids)])) img = img.resize((64,64)) enhancer = ImageEnhance.Brightness(img) img = enhancer.enhance(0.7) img = np.array(img) im[:,:,:] = img else: im[:, :, 0] = 255 * v im[:, :, 2] = 255 * (1 - v) return im # Add a square def sticker(im, x_start = 0, y_start = 0, delta = 10, color = [0, 0, 0]): im[y_start:y_start + delta, x_start:x_start + delta, :] = color return im # Add text def text(im, text, x, y, color = (0,0,0), size = 20): im = Image.fromarray(im) draw = ImageDraw.Draw(im) font = ImageFont.truetype("../data/arial.ttf", size) w, h = font.getsize(text) draw.text((x, y), text, color, font = font) im = np.array(im) return im, (w,h) def vec2im(features, **kwargs): """ Convert a feature into an image with certain features. :return image, text location, small box location """ im = np.zeros((64, 64, 3), dtype = np.uint8) im = shade(im, features[4]) if features[5] == 1: # place large box in the image if 'x_start' in kwargs.keys(): patch_x = kwargs['x_start'] else: patch_x = 0 if 'y_start' in kwargs.keys(): patch_y = kwargs['y_start'] else: patch_y = 0 if 'p_color' in kwargs.keys(): p_color = kwargs['p_color'] else: p_color = [0,0,0] if 'p_size' in kwargs.keys(): p_delta = kwargs['p_size'] else: p_delta = 10 # Add a large box in the image at the location set in the argument. im = sticker(im, x_start=patch_x, y_start=patch_y, color=p_color, delta=p_delta) else: # no large box in the image so the locations are set to None patch_x = None patch_y = None # Determine the character to be included in the image if features[0] == 0: char = "A" elif features[0] == 1: char = "B" elif features[0] == -1: char = None # no character # Determine the color of the character to be included in the image if features[3] == 0: # set text color as black color = (0, 0, 0) elif features[3] == 1: # set text color as green color = (0, 255, 0) elif features[3] == 3: # set text color as white color = (255, 255, 255) elif features[3] == 2: # set manual text color from (R, G, B) input color = kwargs['color'] textloc = (None, None, None, None) # Add text if character is not None if char != None: xstart = int(36 * features[1] + 6) ystart = int(36 * features[2] + 6) im, dim = text(im, char, xstart, ystart, color = color) # keep the location of the character to return textloc = (ystart, ystart+dim[1], xstart, xstart+dim[0]) # Add a small box if switch argument is turned on dist_ = 300 if kwargs['switch'] == 1: # small box at a fixed location switch_x = 58 switch_y = 58 im = sticker(im, x_start=58, y_start=58, delta=4, color=kwargs['s_color']) elif kwargs['switch'] == 2: # small box at a random location while True: switch_x = np.random.random_integers(0, 53) switch_y = np.random.random_integers(0, 53) # prevent overlap of switch with text and larger box if patch_x is not None and features[0] != -1: # yes patch, yes character if (int(36*features[1]) - switch_x)**2 + (int(36*features[2]) - switch_y)**2 > dist_+100 and \ (patch_x - switch_x)**2 + (patch_y - switch_y)**2 > dist_: break elif patch_x is None and features[0] != -1: # no patch, yes character if (int(36*features[1]) - switch_x)**2 + (int(36*features[2]) - switch_y)**2 > dist_+100: break elif patch_x is not None and features[0] == -1: # yes patch, no character if (patch_x - switch_x)**2 + (patch_y - switch_y)**2 > dist_: break elif patch_x is None and features[0] == -1: # neither break im = sticker(im, x_start=switch_x, y_start=switch_y, delta=4, color=kwargs['s_color']) else: switch_x = None switch_y = None # keep the small box location to return if switch_x is not None: switch_loc = (switch_y, switch_y+4, switch_x, switch_x+4) else: switch_loc = (None, None, None, None) # return the image generated, character location, and small box location return im, textloc, switch_loc def save_data(exp_no, save_dir, train_data, test_data, train_coord, train_avoid, train_avoid2, test_coord, test_avoid, test_avoid2, save=True): # setup bbox info to save to the file fname = os.path.join(save_dir, 'textbox_%0.2f.npz'%exp_no) # NOTE need to specify below based on different type of experiments if exp_no in [1.11, 2.11]: # for simple FR and NR, only one object to include gt_flag = [1,0,0] elif exp_no == 1.2: # for complex-FR, there are two ground-truth objects to include gt_flag = [1,0,1] elif exp_no >= 3.7: # for complex-CR, there are two ground-truth objects to inlcude gt_flag = [1,0,1] train_primary, train_secondary = setup_bboxes(train_coord, train_avoid, train_avoid2, np.array(range(train_data.X.shape[0])), gt_flag=gt_flag) test_primary, test_secondary = setup_bboxes(test_coord, test_avoid, test_avoid2, np.array(range(test_data.X.shape[0])), gt_flag=gt_flag) if save: np.savez(open(fname, 'wb'), x_train=train_data.X, y_train=train_data.y, x_test=test_data.X, y_test=test_data.y, train_primary=train_primary, test_primary=test_primary, train_secondary=train_secondary, test_secondary=test_secondary) return train_data, test_data, train_primary, train_secondary, test_primary, test_secondary def load_data(exp_no, load_dir): fname = os.path.join(load_dir, 'textbox_%0.2f.npz'%exp_no) tmp = np.load(open(fname, 'rb'), allow_pickle=True) train_data = TextBoxDataset(tmp['x_train'], tmp['y_train']) test_data = TextBoxDataset(tmp['x_test'], tmp['y_test']) train_primary = tmp['train_primary'] test_primary = tmp['test_primary'] train_secondary = tmp['train_secondary'] test_secondary = tmp['test_secondary'] return train_data, test_data, train_primary, train_secondary, test_primary, test_secondary # Generate text data with spurious features # make the labels to be correlated with color, not the digit itself # or the other way def sample_uniform(): feature = np.zeros((6)) feature[0] = np.random.randint(2) #character feature[1] = np.random.uniform() #x feature[2] = np.random.uniform() #y feature[3] = 0 feature[4] = np.random.uniform() # shade feature[5] = 0 return feature def generate_data(n=10000): #plain data rep = np.zeros((n, 6)) labels = np.zeros(n) im = np.zeros((n, 64, 64, 3)) for i in range(n): rep[i] = sample_uniform() im[i] = vec2im(rep[i]) labels[i] = int(rep[i][0]) im = np.float32(im / 255) return im, labels, rep def original_textbox_data(n=10000, save=True, save_dir='data'): if not os.path.exists(save_dir): os.mkdir(save_dir) fname = os.path.join(save_dir, 'textbox_original.npz') if os.path.exists(fname): tmp = np.load(open(fname, 'rb')) im = tmp['x_train'] labels = tmp['y_train'] im_test = tmp['x_test'] labels_test = tmp['y_test'] else: # train data im, labels, rep = generate_data(n=n) # val data test_n = int(n * 0.3) im_test, labels_test, rep_test = generate_data(n=test_n) train_data = TextBoxDataset(im, labels) test_data = TextBoxDataset(im_test, labels_test) if save: np.savez(open(fname, 'wb'), x_train=train_data.X, y_train=train_data.y, x_test=test_data.X, y_test=test_data.y) np.savez(open(os.path.join(save_dir, 'textbox_original_meta.npz'), 'wb'), rep=rep, rep_test=rep_test) return train_data, test_data
9,823
38.773279
200
py
SMERF
SMERF-main/smerf/models.py
import keras from keras.layers.pooling import GlobalAveragePooling1D, GlobalAveragePooling2D import numpy as np from keras.utils.np_utils import to_categorical import os import keras.backend as K from keras.callbacks import EarlyStopping, ModelCheckpoint from keras.metrics import categorical_accuracy from keras.applications.resnet50 import ResNet50 from keras.applications.vgg16 import VGG16 from keras.applications.densenet import DenseNet121 from keras.applications.mobilenet_v2 import MobileNetV2 from keras.preprocessing import image from keras.applications.resnet50 import preprocess_input, decode_predictions from keras.preprocessing.image import ImageDataGenerator from keras.layers import Dense, Dropout, Flatten, Conv2D, MaxPool2D, ZeroPadding2D, MaxPooling2D, BatchNormalization, Activation from keras.models import Sequential, Model from keras.layers.core import Flatten, Dense, Dropout from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D class EarlyStoppingByLossVal(keras.callbacks.Callback): def __init__(self, monitor='loss', value=0.1, verbose=0): super(keras.callbacks.Callback, self).__init__() self.monitor = monitor self.value = value self.verbose = verbose def on_batch_end(self, batch, logs={}): current = logs.get(self.monitor) if current < self.value: print('stopping with %f'%current) self.model.stop_training = True def on_epoch_end(self, epoch, logs={}): current = logs.get(self.monitor) if current is None: print("Early stopping requires %s available!" % self.monitor) if current < self.value: print('stopping with %f'%current) self.model.stop_training = True if current > 7.00: self.model.stop_training = True class TextBoxCNN: def __init__(self, lr=0.0001, batch=128, max_epoch=10, interm_dim=200, input_shape=(64, 64, 3), model_name='w.pt', output_dir='../outputs'): self.input_shape = input_shape self.model = keras.Sequential([ #keras.layers.InputLayer(input_shape=(64, 64, 3)), keras.layers.Conv2D(filters=32, kernel_size=3, strides=(2, 2), activation='relu', input_shape=input_shape), #keras.layers.Conv2D(filters=32, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Flatten(), keras.layers.Dense(interm_dim, activation='relu'), keras.layers.Dense(2) ]) self.lr = lr self.opt = keras.optimizers.Adam(lr=self.lr) #self.opt = keras.optimizers.SGD(lr=self.lr) self.batch = batch self.max_epoch = max_epoch self.model.compile(optimizer=self.opt, loss='binary_crossentropy', metrics=['accuracy']) self.modelfile = os.path.join(output_dir, model_name) def train(self, x_train, y_train, retrain=False, validate=False, earlystop=False, verbose=True, adversarial=False): if not adversarial: if earlystop: #cb = [EarlyStopping(monitor='accuracy', mode='min', verbose=1)] #cb = [EarlyStopping(monitor='loss', patience=1, mode='min'), ModelCheckpoint(self.modelfile, monitor='loss', mode='min', save_best_only=True)] cb = [EarlyStoppingByLossVal(monitor='loss', value=0.005)] else: cb = [] if os.path.exists(self.modelfile) and not retrain: self.model.load_weights(self.modelfile) elif not os.path.exists(self.modelfile) and retrain: raise ValueError('modelfile not found') else: y_train_oh = to_categorical(y_train, 2) if validate: self.model.fit(x_train, y_train_oh, batch_size=self.batch, epochs=self.max_epoch, validation_split=0.1, shuffle=True, callbacks=cb) else: self.model.fit(x_train, y_train_oh, batch_size=self.batch, epochs=self.max_epoch, validation_split=0, shuffle=True, callbacks=cb) self.model.save_weights(self.modelfile) else: # add adversarial training: load from pretrained if not os.path.exists(self.modelfile): raise ValueError('modelfile %s not found. Make sure to run script/adv_train.py to train the model first.'%self.modelfile) self.model.load_weights(self.modelfile) if verbose: print(self.model.summary()) def test(self, x_test, y_test): pred = self.model.predict_classes(x_test) score = (pred == y_test).sum() / y_test.shape[0] print('Accuracy=%f'%score) return score class TextBoxCNN_adv(TextBoxCNN): def __init__(self, lr=0.0001, batch=128, max_epoch=10, interm_dim=200, input_shape=(64, 64, 3), model_name='w.pt', output_dir='../outputs'): self.input_shape = input_shape self.model = keras.Sequential([ #keras.layers.InputLayer(input_shape=(64, 64, 3)), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu', input_shape=input_shape), keras.layers.Conv2D(filters=128, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=256, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Conv2D(filters=64, kernel_size=3, strides=(2, 2), activation='relu'), keras.layers.Flatten(), keras.layers.Dense(interm_dim, activation='relu'), keras.layers.Dense(interm_dim, activation='relu'), keras.layers.Dense(2) ]) self.lr = lr self.opt = keras.optimizers.Adam(lr=self.lr) self.batch = batch self.max_epoch = max_epoch self.model.compile(optimizer=self.opt, loss='binary_crossentropy', metrics=['accuracy']) self.modelfile = os.path.join(output_dir, model_name) ## CNNs class AlexNet_model(TextBoxCNN): def __init__(self, lr=0.005, batch=1024, max_epoch=10, interm_dim=200, input_shape=(64, 64, 3), model_name='w_alex.pt', output_dir='../outputs'): self.input_shape = input_shape #Instantiation self.model = keras.Sequential() #1st Convolutional Layer self.model.add(Conv2D(filters=96, input_shape=self.input_shape, kernel_size=(11,11), strides=(4,4), padding='same')) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) self.model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2), padding='same')) #2nd Convolutional Layer self.model.add(Conv2D(filters=256, kernel_size=(5, 5), strides=(1,1), padding='same')) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) self.model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2), padding='same')) #3rd Convolutional Layer self.model.add(Conv2D(filters=384, kernel_size=(3,3), strides=(1,1), padding='same')) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) #4th Convolutional Layer self.model.add(Conv2D(filters=384, kernel_size=(3,3), strides=(1,1), padding='same')) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) #5th Convolutional Layer self.model.add(Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), padding='same')) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) self.model.add(MaxPooling2D(pool_size=(2,2), strides=(2,2), padding='same')) #Passing it to a Fully Connected layer self.model.add(Flatten()) # 1st Fully Connected Layer self.model.add(Dense(4096, input_shape=(32,32,3,))) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) # Add Dropout to prevent overfitting self.model.add(Dropout(0.4)) #2nd Fully Connected Layer self.model.add(Dense(4096)) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) #Add Dropout self.model.add(Dropout(0.4)) #3rd Fully Connected Layer self.model.add(Dense(1000)) self.model.add(BatchNormalization()) self.model.add(Activation('relu')) #Add Dropout self.model.add(Dropout(0.4)) #Output Layer self.model.add(Dense(2)) self.model.add(BatchNormalization()) # set up other hyperparmeters self.lr = lr self.opt = keras.optimizers.Adam(lr=self.lr) #self.opt = keras.optimizers.SGD(lr=self.lr) self.batch = batch self.max_epoch = max_epoch self.model.compile(optimizer=self.opt, loss='binary_crossentropy', metrics=['accuracy']) self.modelfile = os.path.join(output_dir, model_name) class VGG16_model(TextBoxCNN): def __init__(self, lr=0.005, batch=1024, max_epoch=10, interm_dim=200, input_shape=(64, 64, 3), model_name='w_res.pt', output_dir='../outputs'): self.input_shape = input_shape # define the base model to fine-tuen from base_model = VGG16(weights='imagenet', include_top=False, input_shape=input_shape) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(1024, activation='relu')(x) x = Dense(512, activation='relu')(x) x = Dense(256, activation='relu')(x) predictions = Dense(2)(x) self.model = Model(inputs=base_model.input, outputs=predictions) for layer in base_model.layers: layer.trainable = False # set up other hyperparmeters self.lr = lr self.opt = keras.optimizers.Adam(lr=self.lr) #self.opt = keras.optimizers.SGD(lr=self.lr) self.batch = batch self.max_epoch = max_epoch self.model.compile(optimizer=self.opt, loss='binary_crossentropy', metrics=['accuracy']) self.modelfile = os.path.join(output_dir, model_name) def test(self, x_test, y_test): pred = self.softmax(self.model.predict(x_test)) pred = np.argmax(pred, axis=1) score = (pred == y_test).sum() / y_test.shape[0] print('Accuracy=%f'%score) return score @staticmethod def softmax(x): return np.exp(x) / np.sum(np.exp(x), axis=1, keepdims=True)
10,558
45.928889
159
py
SMERF
SMERF-main/smerf/lime_utils.py
import lime from lime import lime_image from lime.wrappers.scikit_image import SegmentationAlgorithm import tensorflow as tf import keras.backend as K import numpy as np from smerf.models import * # Uses LIME library to obtain feature attributions def lime_run(model, x_sample, y_sample, x_train, exp_no): n, h, w, c = x_sample.shape output = np.zeros((n, 1, h, w, c)) explainer = lime_image.LimeImageExplainer(verbose=False) for n_i in range(n): explanation = explainer.explain_instance(x_sample[n_i].astype('double'), model.predict, top_labels=2, hide_color=(0,0,0), num_samples=1000, segmentation_fn=SegmentationAlgorithm('quickshift', kernel_size=4, max_dist=10, ratio=0.2) ) ind = explanation.top_labels[0] dict_heatmap = dict(explanation.local_exp[ind]) heatmap = np.vectorize(dict_heatmap.get)(explanation.segments) heatmap = scale(heatmap) output[n_i, 0, :, :, 0] = heatmap output[n_i, 0, :, :, 1] = heatmap output[n_i, 0, :, :, 2] = heatmap return output def scale(x): if x.max() - x.min() != 0: return (x - x.min()) / (x.max() - x.min()) return x
1,723
45.594595
101
py
cpnest
cpnest-master/docs/conf.py
#!/usr/bin/env python3 # -*- coding: utf-8 -*- # # CPNest documentation build configuration file, created by # sphinx-quickstart on Thu Dec 7 13:57:03 2017. # # This file is execfile()d with the current directory set to its # containing dir. # # Note that not all possible configuration values are present in this # autogenerated file. # # All configuration values have a default; values that are commented out # serve to show the default. # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. # #import os #import sys #sys.path.insert(0, os.path.abspath('../cpnest')) import cpnest # -- General configuration ------------------------------------------------ # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # Add any Sphinx extension module names here, as strings. They can be # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = ['sphinx.ext.autodoc', 'sphinx.ext.doctest', 'sphinx.ext.intersphinx', 'sphinx.ext.mathjax', 'sphinx.ext.githubpages', 'sphinx.ext.ifconfig', 'sphinx.ext.viewcode', 'sphinx.ext.napoleon'] # Add any paths that contain templates here, relative to this directory. templates_path = ['_templates'] # The suffix(es) of source filenames. # You can specify multiple suffix as a list of string: # # source_suffix = ['.rst', '.md'] source_suffix = '.rst' # The master toctree document. master_doc = 'index' # General information about the project. project = 'CPNest' copyright = '2017-2021, W. Del Pozzo, J. Veitch' author = 'W. Del Pozzo, J. Veitch' # The version info for the project you're documenting, acts as replacement for # |version| and |release|, also used in various other places throughout the # built documents. # # The short X.Y version. version = cpnest.__version__ # The full version, including alpha/beta/rc tags. release = cpnest.__version__ # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. # # This is also used if you do content translation via gettext catalogs. # Usually you set "language" from the command line for these cases. language = None # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This patterns also effect to html_static_path and html_extra_path exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store'] # The name of the Pygments (syntax highlighting) style to use. pygments_style = 'sphinx' # If true, `todo` and `todoList` produce output, else they produce nothing. todo_include_todos = False # -- Options for HTML output ---------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = 'classic' # Theme options are theme-specific and customize the look and feel of a theme # further. For a list of options available for each theme, see the # documentation. # # html_theme_options = {} # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". html_static_path = ['_static'] # Custom sidebar templates, must be a dictionary that maps document names # to template names. # # This is required for the alabaster theme # refs: http://alabaster.readthedocs.io/en/latest/installation.html#sidebars html_sidebars = { '**': [ 'relations.html', # needs 'show_related': True theme option to display 'searchbox.html', ] } # -- Options for HTMLHelp output ------------------------------------------ # Output file base name for HTML help builder. htmlhelp_basename = 'CPNestdoc' # -- Options for LaTeX output --------------------------------------------- latex_elements = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # # 'pointsize': '10pt', # Additional stuff for the LaTeX preamble. # # 'preamble': '', # Latex figure (float) alignment # # 'figure_align': 'htbp', } # Grouping the document tree into LaTeX files. List of tuples # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ (master_doc, 'CPNest.tex', 'CPNest Documentation', 'W. Del Pozzo, J. Veitch', 'manual'), ] # -- Options for manual page output --------------------------------------- # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ (master_doc, 'cpnest', 'CPNest Documentation', [author], 1) ] # -- Options for Texinfo output ------------------------------------------- # Grouping the document tree into Texinfo files. List of tuples # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ (master_doc, 'CPNest', 'CPNest Documentation', author, 'CPNest', 'One line description of project.', 'Miscellaneous'), ] # Example configuration for intersphinx: refer to the Python standard library. intersphinx_mapping = {'https://docs.python.org/': None, 'numpy': ('https://docs.scipy.org/doc/numpy', None), 'scipy': ('https://docs.scipy.org/doc/scipy/reference', None), 'matplotlib': ('https://matplotlib.sourceforge.net', None) }
5,836
30.896175
85
py
CR-VAE
CR-VAE-main/CRVAE_demo.py
# -*- coding: utf-8 -*- """ Created on Sat Aug 6 20:00:04 2022 @author: 61995 """ import torch import numpy as np import matplotlib.pyplot as plt from models.cgru_error import CRVAE, VRAE4E, train_phase1, train_phase2 import scipy.io device = torch.device('cuda') X_np = np.load('henon.npy').T dim = X_np.shape[-1] GC = np.zeros([dim,dim]) for i in range(dim): GC[i,i] = 1 if i!=0: GC[i,i-1] = 1 X = torch.tensor(X_np[np.newaxis], dtype=torch.float32, device=device) full_connect = np.ones(GC.shape) cgru = CRVAE(X.shape[-1], full_connect, hidden=64).cuda(device=device) vrae = VRAE4E(X.shape[-1], hidden=64).cuda(device=device) #%% train_loss_list = train_phase1( cgru, X, context=20, lam=0.1, lam_ridge=0, lr=5e-2, max_iter=1000, check_every=50)#0.1 #%% GC_est = cgru.GC().cpu().data.numpy() print('True variable usage = %.2f%%' % (100 * np.mean(GC))) print('Estimated variable usage = %.2f%%' % (100 * np.mean(GC_est))) print('Accuracy = %.2f%%' % (100 * np.mean(GC == GC_est))) # Make figures fig, axarr = plt.subplots(1, 2, figsize=(10, 5)) axarr[0].imshow(GC, cmap='Blues') axarr[0].set_title('Causal-effect matrix') axarr[0].set_ylabel('Effect series') axarr[0].set_xlabel('Causal series') axarr[0].set_xticks([]) axarr[0].set_yticks([]) axarr[1].imshow(GC_est, cmap='Blues', vmin=0, vmax=1, extent=(0, len(GC_est), len(GC_est), 0)) axarr[1].set_ylabel('Effect series') axarr[1].set_xlabel('Causal series') axarr[1].set_xticks([]) axarr[1].set_yticks([]) # Mark disagreements for i in range(len(GC_est)): for j in range(len(GC_est)): if GC[i, j] != GC_est[i, j]: rect = plt.Rectangle((j, i-0.05), 1, 1, facecolor='none', edgecolor='red', linewidth=1) axarr[1].add_patch(rect) plt.show() #np.save('GC_henon.npy', GC_est) full_connect = np.load('GC_henon.npy') #%% cgru = CRVAE(X.shape[-1], full_connect, hidden=64).cuda(device=device) vrae = VRAE4E(X.shape[-1], hidden=64).cuda(device=device) train_loss_list = train_phase2( cgru, vrae, X, context=20, lam=0., lam_ridge=0, lr=5e-2, max_iter=10000, check_every=50)
2,116
23.333333
99
py
CR-VAE
CR-VAE-main/models/cgru_error.py
# -*- coding: utf-8 -*- """ Created on Sat Aug 6 20:01:33 2022 @author: 61995 """ import torch import torch.nn as nn import numpy as np from copy import deepcopy import matplotlib.pyplot as plt from metrics.visualization_metrics import visualization import torch.optim as optim class GRU(nn.Module): def __init__(self, num_series, hidden): super(GRU, self).__init__() self.p = num_series self.hidden = hidden # Set up network. self.gru = nn.GRU(num_series, hidden, batch_first=True) self.gru.flatten_parameters() self.linear = nn.Linear(hidden, 1) self.sigmoid = nn.Sigmoid() def init_hidden(self, batch): #Initialize hidden states device = self.gru.weight_ih_l0.device return torch.zeros(1, batch, self.hidden, device=device) def forward(self, X, z, connection, mode = 'train'): X=X[:,:,np.where(connection!=0)[0]] device = self.gru.weight_ih_l0.device tau = 0 if mode == 'train': X_right, hidden_out = self.gru(torch.cat((X[:,0:1,:],X[:,11:-1,:]),1), z) X_right = self.linear(X_right) return X_right, hidden_out class VRAE4E(nn.Module): def __init__(self, num_series, hidden): ''' Error VAE ''' super(VRAE4E, self).__init__() self.device = torch.device('cuda') self.p = num_series self.hidden = hidden self.gru_left = nn.GRU(num_series, hidden, batch_first=True) self.gru_left.flatten_parameters() self.fc_mu = nn.Linear(hidden, hidden)#nn.Linear(hidden, 1) self.fc_std = nn.Linear(hidden, hidden) self.linear_hidden = nn.Linear(hidden, hidden) self.tanh = nn.Tanh() self.gru = nn.GRU(num_series, hidden, batch_first=True) self.gru.flatten_parameters() self.linear = nn.Linear(hidden, num_series) def init_hidden(self, batch): '''Initialize hidden states for GRU cell.''' device = self.gru.weight_ih_l0.device return torch.zeros(1, batch, self.hidden, device=device) def forward(self, X, mode = 'train'): X = torch.cat((torch.zeros(X.shape,device = self.device)[:,0:1,:],X),1) if mode == 'train': hidden_0 = torch.zeros(1, X.shape[0], self.hidden, device=self.device) out, h_t = self.gru_left(X[:,1:,:], hidden_0.detach()) mu = self.fc_mu(h_t) log_var = self.fc_std(h_t) sigma = torch.exp(0.5*log_var) z = torch.randn(size = mu.size()) z = z.type_as(mu) z = mu + sigma*z z = self.tanh(self.linear_hidden(z)) X_right, hidden_out = self.gru(X[:,:-1,:], z) pred = self.linear(X_right) return pred, log_var, mu if mode == 'test': X_seq = torch.zeros(X[:,:1,:].shape).to(self.device) h_t = torch.randn(size = (1, X_seq[:,-2:-1,:].size(0),self.hidden)).to(self.device) for i in range(int(20/1)+1): out, h_t = self.gru(X_seq[:,-1:,:], h_t) out = self.linear(out) #out = self.sigmoid(out) X_seq = torch.cat([X_seq,out],dim = 1) return X_seq class CRVAE(nn.Module): def __init__(self, num_series, connection, hidden): ''' connection: pruned networks ''' super(CRVAE, self).__init__() self.device = torch.device('cuda') self.p = num_series self.hidden = hidden self.gru_left = nn.GRU(num_series, hidden, batch_first=True) self.gru_left.flatten_parameters() self.fc_mu = nn.Linear(hidden, hidden) self.fc_std = nn.Linear(hidden, hidden) self.connection = connection # Set up networks. self.networks = nn.ModuleList([ GRU(int(connection[:,i].sum()), hidden) for i in range(num_series)]) def forward(self, X, noise = None, mode = 'train', phase = 0): if phase == 0: X = torch.cat((torch.zeros(X.shape,device = self.device)[:,0:1,:],X),1) if mode == 'train': hidden_0 = torch.zeros(1, X.shape[0], self.hidden, device=self.device) out, h_t = self.gru_left(X[:,1:11,:], hidden_0.detach()) mu = self.fc_mu(h_t) log_var = self.fc_std(h_t) sigma = torch.exp(0.5*log_var) z = torch.randn(size = mu.size()) z = z.type_as(mu) z = mu + sigma*z pred = [self.networks[i](X, z, self.connection[:,i])[0] for i in range(self.p)] return pred, log_var, mu if mode == 'test': X_seq = torch.zeros(X[:,:1,:].shape).to(self.device) h_0 = torch.randn(size = (1, X_seq[:,-2:-1,:].size(0),self.hidden)).to(self.device) ht_last =[] for i in range(self.p): ht_last.append(h_0) for i in range(int(20/1)+1):#int(20/2)+1 ht_new = [] for j in range(self.p): # out, h_t = self.gru_out[j](X_seq[:,-1:,:], ht_last[j]) # out = self.fc[j](out) out, h_t = self.networks[j](X_seq[:,-1:,:], ht_last[j], self.connection[:,j]) if j == 0: X_t = out else: X_t = torch.cat((X_t,out),-1) ht_new.append(h_t) ht_last = ht_new if i ==0: X_seq = X_t else: X_seq = torch.cat([X_seq,X_t],dim = 1) #out = self.sigmoid(out) return X_seq if phase == 1: X = torch.cat((torch.zeros(X.shape,device = self.device)[:,0:1,:],X),1) if mode == 'train': hidden_0 = torch.zeros(1, X.shape[0], self.hidden, device=self.device) out, h_t = self.gru_left(X[:,1:11,:], hidden_0.detach()) mu = self.fc_mu(h_t) log_var = self.fc_std(h_t) sigma = torch.exp(0.5*log_var) z = torch.randn(size = mu.size()) z = z.type_as(mu) # Setting z to be .cuda when using GPU training z = mu + sigma*z pred = [self.networks[i](X, z, self.connection[:,i])[0] for i in range(self.p)] return pred, log_var, mu if mode == 'test': X_seq = torch.zeros(X[:,:1,:].shape).to(self.device) h_0 = torch.randn(size = (1, X_seq[:,-2:-1,:].size(0),self.hidden)).to(self.device) ht_last =[] for i in range(self.p): ht_last.append(h_0) for i in range(int(20/1)+1):#int(20/2)+1 ht_new = [] for j in range(self.p): # out, h_t = self.gru_out[j](X_seq[:,-1:,:], ht_last[j]) # out = self.fc[j](out) out, h_t = self.networks[j](X_seq[:,-1:,:], ht_last[j], self.connection[:,j]) if j == 0: X_t = out else: X_t = torch.cat((X_t,out),-1) ht_new.append(h_t) ht_last = ht_new if i ==0: X_seq = X_t + 0.1*noise[:,i:i+1,:] else: X_seq = torch.cat([X_seq,X_t+0.1*noise[:,i:i+1,:]],dim = 1) #out = self.sigmoid(out) return X_seq def GC(self, threshold=True): ''' Extract learned Granger causality. Args: threshold: return norm of weights, or whether norm is nonzero. Returns: GC: (p x p) matrix. Entry (i, j) indicates whether variable j is Granger causal of variable i. ''' GC = [torch.norm(net.gru.weight_ih_l0, dim=0) for net in self.networks] GC = torch.stack(GC) #print(GC) if threshold: return (torch.abs(GC) > 0).int() else: return GC def prox_update(network, lam, lr): '''Perform in place proximal update on first layer weight matrix.''' W = network.gru.weight_ih_l0 norm = torch.norm(W, dim=0, keepdim=True) W.data = ((W / torch.clamp(norm, min=(lam * lr))) * torch.clamp(norm - (lr * lam), min=0.0)) network.gru.flatten_parameters() def regularize(network, lam): '''Calculate regularization term for first layer weight matrix.''' W = network.gru.weight_ih_l0 return lam * torch.sum(torch.norm(W, dim=0)) def ridge_regularize(network, lam): '''Apply ridge penalty at linear layer and hidden-hidden weights.''' return lam * ( torch.sum(network.linear.weight ** 2) + torch.sum(network.gru.weight_hh_l0 ** 2))# + #torch.sum(network.fc_std.weight ** 2) + #torch.sum(network.fc_mu.weight ** 2) + #torch.sum(network.fc_std.weight ** 2)) def restore_parameters(model, best_model): '''Move parameter values from best_model to model.''' for params, best_params in zip(model.parameters(), best_model.parameters()): params.data = best_params def arrange_input(data, context): ''' Arrange a single time series into overlapping short sequences. Args: data: time series of shape (T, dim). context: length of short sequences. ''' assert context >= 1 and isinstance(context, int) input = torch.zeros(len(data) - context, context, data.shape[1], dtype=torch.float32, device=data.device) target = torch.zeros(len(data) - context, context, data.shape[1], dtype=torch.float32, device=data.device) for i in range(context): start = i end = len(data) - context + i input[:, i, :] = data[start:end] target[:, i, :] = data[start+1:end+1] return input.detach(), target.detach() def MinMaxScaler(data): """Min-Max Normalizer. Args: - data: raw data Returns: - norm_data: normalized data - min_val: minimum values (for renormalization) - max_val: maximum values (for renormalization) """ min_val = np.min(np.min(data, axis = 0), axis = 0) data = data - min_val max_val = np.max(np.max(data, axis = 0), axis = 0) norm_data = data / (max_val + 1e-7) return norm_data def train_phase2(crvae, vrae, X, context, lr, max_iter, lam=0, lam_ridge=0, lookback=5, check_every=50, verbose=1,sparsity = 100, batch_size = 1024): '''Train model with Adam.''' optimizer = optim.Adam(vrae.parameters(), lr=1e-3) p = X.shape[-1] device = crvae.networks[0].gru.weight_ih_l0.device loss_fn = nn.MSELoss() train_loss_list = [] batch_size = batch_size # Set up data. X, Y = zip(*[arrange_input(x, context) for x in X]) X_all = torch.cat(X, dim=0) Y_all = torch.cat(Y, dim=0) idx = np.random.randint(len(X_all), size=(batch_size,)) X = X_all[idx] Y = Y_all[idx] X_v = X_all[batch_size:] start_point = 0#context-10-1 beta = 1#0.001 beta_e = 1 # For early stopping. best_it = None best_loss = np.inf best_model = None # Calculate smooth error. pred,mu,log_var = crvae(X)# loss = sum([loss_fn(pred[i][:, :, 0], X[:, 10:, i]) for i in range(p)]) mmd = (-0.5*(1+log_var - mu**2- torch.exp(log_var)).sum(dim = -1).sum(dim = 0)).mean(dim =0) #mmd = sum([MMD(torch.randn(200, Y[:, :, 0].shape[-1], requires_grad = False).to(device), latent[i][:,:,0]) for i in range(p)]) ridge = sum([ridge_regularize(net, lam_ridge) for net in crvae.networks]) smooth = loss + ridge + beta*mmd error = (-torch.stack(pred)[:, :, :, 0].permute(1,2,0) + X[:, 10:, :]).detach() pred_e,mu_e,log_var_e = vrae(error) loss_e = loss_fn(pred_e, error) mmd_e = (-0.5*(1+log_var_e - mu_e**2- torch.exp(log_var_e)).sum(dim = -1).sum(dim = 0)).mean(dim =0) smooth_e = loss_e + beta_e*mmd_e best_mmd = np.inf ######################################################################## #lr = 1e-3 for it in range(max_iter): # Take gradient step. smooth_e.backward() if lam == 0: optimizer.step() optimizer.zero_grad() smooth.backward() for param in crvae.parameters(): param.data -= lr * param.grad # Take prox step. if lam > 0: for net in crvae.networks: prox_update(net, lam, lr) crvae.zero_grad() # Calculate loss for next iteration. idx = np.random.randint(len(X_all), size=(batch_size,)) #X = X_all[idx] #Y = Y_all[idx] pred,mu,log_var = crvae(X)# loss = sum([loss_fn(pred[i][:, :, 0], X[:, 10:, i]) for i in range(p)]) mmd = (-0.5*(1+log_var - mu**2- torch.exp(log_var)).sum(dim = -1).sum(dim = 0)).mean(dim =0) ridge = sum([ridge_regularize(net, lam_ridge) for net in crvae.networks]) smooth = loss + ridge + beta*mmd error = (-torch.stack(pred)[:, :, :, 0].permute(1,2,0) + X[:, 10:, :]).detach() pred_e,mu_e,log_var_e = vrae(error) loss_e = loss_fn(pred_e, error) mmd_e = (-0.5*(1+log_var_e - mu_e**2- torch.exp(log_var_e)).sum(dim = -1).sum(dim = 0)).mean(dim =0) smooth_e = loss_e + beta_e*mmd_e # Check progress. if (it) % check_every == 0: X_t = X pred_t,mu_t ,log_var_t= crvae(X_t) loss_t = sum([loss_fn(pred_t[i][:, :, 0], X_t[:, 10:, i]) for i in range(p)]) mmd_t = (-0.5*(1+log_var_t - mu_t**2- torch.exp(log_var_t)).sum(dim = -1).sum(dim = 0)).mean(dim =0) ridge_t = sum([ridge_regularize(net, lam_ridge) for net in crvae.networks]) smooth_t = loss_t + ridge_t# + beta*mmd_t nonsmooth = sum([regularize(net, lam) for net in crvae.networks]) mean_loss = (smooth_t) / p if verbose > 0: print(('-' * 10 + 'Iter = %d' + '-' * 10) % (it )) print('Loss = %f' % mean_loss) print('KL = %f' % mmd) print('Loss_e = %f' % smooth_e) print('KL_e = %f' % mmd_e) if lam>0: print('Variable usage = %.2f%%' % (100 * torch.mean(crvae.GC().float()))) if mean_loss < best_loss: best_loss = mean_loss best_it = it best_model = deepcopy(crvae) start_point = 0 predicted_error = vrae(error, mode = 'test').detach() predicted_data = crvae(X_t, predicted_error, mode = 'test', phase = 1) syn = predicted_data[:,:-1,:].cpu().detach().numpy() ori= X_t[:,start_point:,:].cpu().detach().numpy() if it % 1000 ==0: plt.plot(ori[0,:,1]) plt.plot(syn[0,:,1]) plt.show() visualization(ori, syn, 'pca') visualization(ori, syn, 'tsne') np.save('ori_henon.npy',ori) np.save('syn_henon.npy',syn) # Restore best model. restore_parameters(crvae, best_model) return train_loss_list def train_phase1(crvae, X, context, lr, max_iter, lam=0, lam_ridge=0, lookback=5, check_every=50, verbose=1,sparsity = 100, batch_size = 2048): '''Train model with Adam.''' p = X.shape[-1] device = crvae.networks[0].gru.weight_ih_l0.device loss_fn = nn.MSELoss() train_loss_list = [] batch_size = batch_size # Set up data. X, Y = zip(*[arrange_input(x, context) for x in X]) X_all = torch.cat(X, dim=0) Y_all = torch.cat(Y, dim=0) idx = np.random.randint(len(X_all), size=(batch_size,)) X = X_all[idx] Y = Y_all[idx] start_point = 0 beta = 0.1 # For early stopping. best_it = None best_loss = np.inf best_model = None # Calculate crvae error. pred,mu,log_var = crvae(X) loss = sum([loss_fn(pred[i][:, :, 0], X[:, 10:, i]) for i in range(p)]) mmd = (-0.5*(1+log_var - mu**2- torch.exp(log_var)).sum(dim = -1).sum(dim = 0)).mean(dim =0) #mmd = sum([MMD(torch.randn(200, Y[:, :, 0].shape[-1], requires_grad = False).to(device), latent[i][:,:,0]) for i in range(p)]) ridge = sum([ridge_regularize(net, lam_ridge) for net in crvae.networks]) smooth = loss + ridge + beta*mmd best_mmd = np.inf ######################################################################## #lr = 1e-3 for it in range(max_iter): # Take gradient step. smooth.backward() for param in crvae.parameters(): param.data -= lr * param.grad # Take prox step. if lam > 0: for net in crvae.networks: prox_update(net, lam, lr) crvae.zero_grad() pred,mu,log_var = crvae(X) loss = sum([loss_fn(pred[i][:, :, 0], X[:, 10:, i]) for i in range(p)]) mmd = (-0.5*(1+log_var - mu**2- torch.exp(log_var)).sum(dim = -1).sum(dim = 0)).mean(dim =0) ridge = sum([ridge_regularize(net, lam_ridge) for net in crvae.networks]) smooth = loss + ridge + beta*mmd # Check progress. if (it) % check_every == 0: X_t = X Y_t = Y pred_t,mu_t ,log_var_t= crvae(X_t) loss_t = sum([loss_fn(pred_t[i][:, :, 0], X_t[:, 10:, i]) for i in range(p)]) mmd_t = (-0.5*(1+log_var_t - mu_t**2- torch.exp(log_var_t)).sum(dim = -1).sum(dim = 0)).mean(dim =0) ridge_t = sum([ridge_regularize(net, lam_ridge) for net in crvae.networks]) smooth_t = loss_t + ridge_t# + beta*mmd_t nonsmooth = sum([regularize(net, lam) for net in crvae.networks]) mean_loss = (smooth_t) / p if verbose > 0: print(('-' * 10 + 'Iter = %d' + '-' * 10) % (it )) print('Loss = %f' % mean_loss) print('KL = %f' % mmd) if lam>0: print('Variable usage = %.2f%%' % (100 * torch.mean(crvae.GC().float()))) if mean_loss < best_loss: best_loss = mean_loss best_it = it best_model = deepcopy(crvae) start_point = 0 predicted_data = crvae(X_t,mode = 'test') syn = predicted_data[:,:-1,:].cpu().detach().numpy() ori= X_t[:,start_point:,:].cpu().detach().numpy() syn = MinMaxScaler(syn) ori = MinMaxScaler(ori) # Restore best model. restore_parameters(crvae, best_model) return train_loss_list
20,309
30.734375
132
py
ulysses
ulysses-master/doc/conf.py
# -*- coding: utf-8 -*- # # Configuration file for the Sphinx documentation builder. # # This file does only contain a selection of the most common options. For a # full list see the documentation: # http://www.sphinx-doc.org/en/master/config # -- Path setup -------------------------------------------------------------- # If extensions (or modules to document with autodoc) are in another directory, # add these directories to sys.path here. If the directory is relative to the # documentation root, use os.path.abspath to make it absolute, like shown here. # # import os # import sys # sys.path.insert(0, os.path.abspath('.')) # -- Project information ----------------------------------------------------- project = 'ulysses' copyright = '2019, Kris Moffat, Holger Schulz, Jessica Turner' author = 'Kris Moffat, Holger Schulz, Jessica Turner' # The short X.Y version version = '' # The full version, including alpha/beta/rc tags release = '0.2' # -- General configuration --------------------------------------------------- # If your documentation needs a minimal Sphinx version, state it here. # # needs_sphinx = '1.0' # Add any Sphinx extension module names here, as strings. They can be # extensions coming with Sphinx (named 'sphinx.ext.*') or your custom # ones. extensions = [ 'sphinx.ext.autodoc', 'sphinx.ext.coverage', 'sphinx.ext.mathjax', 'sphinx.ext.viewcode', 'sphinx.ext.githubpages', ] # Add any paths that contain templates here, relative to this directory. templates_path = ['_templates'] # The suffix(es) of source filenames. # You can specify multiple suffix as a list of string: # # source_suffix = ['.rst', '.md'] source_suffix = '.rst' # The master toctree document. master_doc = 'index' # The language for content autogenerated by Sphinx. Refer to documentation # for a list of supported languages. # # This is also used if you do content translation via gettext catalogs. # Usually you set "language" from the command line for these cases. language = None # List of patterns, relative to source directory, that match files and # directories to ignore when looking for source files. # This pattern also affects html_static_path and html_extra_path. exclude_patterns = ['_build', 'Thumbs.db', '.DS_Store'] # The name of the Pygments (syntax highlighting) style to use. pygments_style = None # -- Options for HTML output ------------------------------------------------- # The theme to use for HTML and HTML Help pages. See the documentation for # a list of builtin themes. # html_theme = 'alabaster' # Theme options are theme-specific and customize the look and feel of a theme # further. For a list of options available for each theme, see the # documentation. # # html_theme_options = {} # Add any paths that contain custom static files (such as style sheets) here, # relative to this directory. They are copied after the builtin static files, # so a file named "default.css" will overwrite the builtin "default.css". html_static_path = ['_static'] # Custom sidebar templates, must be a dictionary that maps document names # to template names. # # The default sidebars (for documents that don't match any pattern) are # defined by theme itself. Builtin themes are using these templates by # default: ``['localtoc.html', 'relations.html', 'sourcelink.html', # 'searchbox.html']``. # # html_sidebars = {} # -- Options for HTMLHelp output --------------------------------------------- # Output file base name for HTML help builder. htmlhelp_basename = 'ulyssesdoc' # -- Options for LaTeX output ------------------------------------------------ latex_elements = { # The paper size ('letterpaper' or 'a4paper'). # # 'papersize': 'letterpaper', # The font size ('10pt', '11pt' or '12pt'). # # 'pointsize': '10pt', # Additional stuff for the LaTeX preamble. # # 'preamble': '', # Latex figure (float) alignment # # 'figure_align': 'htbp', } # Grouping the document tree into LaTeX files. List of tuples # (source start file, target name, title, # author, documentclass [howto, manual, or own class]). latex_documents = [ (master_doc, 'ulysses.tex', 'ulysses Documentation', 'Kris Moffat, Holger Schulz, Jessica Turner', 'manual'), ] # -- Options for manual page output ------------------------------------------ # One entry per manual page. List of tuples # (source start file, name, description, authors, manual section). man_pages = [ (master_doc, 'ulysses', 'ulysses Documentation', [author], 1) ] # -- Options for Texinfo output ---------------------------------------------- # Grouping the document tree into Texinfo files. List of tuples # (source start file, target name, title, author, # dir menu entry, description, category) texinfo_documents = [ (master_doc, 'ulysses', 'ulysses Documentation', author, 'ulysses', 'One line description of project.', 'Miscellaneous'), ] # -- Options for Epub output ------------------------------------------------- # Bibliographic Dublin Core info. epub_title = project # The unique identifier of the text. This can be a ISBN number # or the project homepage. # # epub_identifier = '' # A unique identification for the text. # # epub_uid = '' # A list of files that should not be packed into the epub file. epub_exclude_files = ['search.html'] # -- Extension configuration ------------------------------------------------- B
5,441
28.737705
79
py
Tatoeba-Challenge
Tatoeba-Challenge-master/scripts/convert_marian_tatoeba_to_pytorch.py
# Copyright 2020 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import argparse import os from pathlib import Path from typing import List, Tuple from transformers.models.marian.convert_marian_to_pytorch import ( FRONT_MATTER_TEMPLATE, _parse_readme, convert_all_sentencepiece_models, get_system_metadata, remove_prefix, remove_suffix, ) try: import pandas as pd except ImportError: pass DEFAULT_REPO = "Tatoeba-Challenge" DEFAULT_MODEL_DIR = os.path.join(DEFAULT_REPO, "models") LANG_CODE_URL = "https://datahub.io/core/language-codes/r/language-codes-3b2.csv" ISO_URL = "https://cdn-datasets.huggingface.co/language_codes/iso-639-3.csv" ISO_PATH = "lang_code_data/iso-639-3.csv" LANG_CODE_PATH = "lang_code_data/language-codes-3b2.csv" class TatoebaConverter: """ Convert Tatoeba-Challenge models to huggingface format. Steps: 1. convert numpy state dict to hf format (same code as OPUS-MT-Train conversion). 2. rename opus model to huggingface format. This means replace each alpha3 code with an alpha2 code if a unique one exists. e.g. aav-eng -> aav-en, heb-eng -> he-en 3. write a model card containing the original Tatoeba-Challenge/README.md and extra info about alpha3 group members. """ def __init__(self, save_dir="marian_converted"): assert Path(DEFAULT_REPO).exists(), "need git clone git@github.com:Helsinki-NLP/Tatoeba-Challenge.git" reg = self.make_tatoeba_registry() self.download_metadata() self.registry = reg reg_df = pd.DataFrame(reg, columns=["id", "prepro", "url_model", "url_test_set"]) assert reg_df.id.value_counts().max() == 1 reg_df = reg_df.set_index("id") reg_df["src"] = reg_df.reset_index().id.apply(lambda x: x.split("-")[0]).values reg_df["tgt"] = reg_df.reset_index().id.apply(lambda x: x.split("-")[1]).values released_cols = [ "url_base", "pair", # (ISO639-3/ISO639-5 codes), "short_pair", # (reduced codes), "chrF2_score", "bleu", #"brevity_penalty", #"ref_len", "src_name", "tgt_name", ] released = pd.read_csv("Tatoeba-Challenge/models/released-models.txt", sep="\t", header=None).iloc[:-1] released.columns = released_cols released["fname"] = released["url_base"].apply( lambda x: remove_prefix( remove_suffix(x.split("/")[5], ".zip"), "opus") ) # lambda x: remove_suffix(remove_prefix(x, "https://object.pouta.csc.fi/Tatoeba-Challenge/opus"), ".zip") released["2m"] = released.fname.str.startswith("2m") # released["date"] = pd.to_datetime( # released["fname"].apply(lambda x: remove_prefix(remove_prefix(x, "2m-"), "-")) # ) # attempted fix: released["date"] = pd.to_datetime( released["url_base"].apply( lambda x: (remove_suffix(x, ".zip")[-10:])) # extract date: last 10 chars before ".zip" ) # released["fname"].apply(lambda x: x.split("-")[1] ) released["base_ext"] = released.url_base.apply(lambda x: Path(x).name) reg_df["base_ext"] = reg_df.url_model.apply(lambda x: Path(x).name) metadata_new = reg_df.reset_index().merge(released.rename(columns={"pair": "id"}), on=["base_ext", "id"]) metadata_renamer = {"src": "src_alpha3", "tgt": "tgt_alpha3", "id": "long_pair", "date": "train_date"} metadata_new = metadata_new.rename(columns=metadata_renamer) metadata_new["src_alpha2"] = metadata_new.short_pair.apply(lambda x: x.split("-")[0]) metadata_new["tgt_alpha2"] = metadata_new.short_pair.apply(lambda x: x.split("-")[1]) DROP_COLS_BOTH = ["url_base", "base_ext", "fname"] metadata_new = metadata_new.drop(DROP_COLS_BOTH, 1) metadata_new["prefer_old"] = metadata_new.long_pair.isin([]) metadata_new = metadata_new.drop_duplicates(subset=['short_pair']) # added row self.metadata = metadata_new assert self.metadata.short_pair.value_counts().max() == 1, "Multiple metadata entries for a short pair" self.metadata = self.metadata.set_index("short_pair") # wget.download(LANG_CODE_URL) mapper = pd.read_csv(LANG_CODE_PATH) mapper.columns = ["a3", "a2", "ref"] self.iso_table = pd.read_csv(ISO_PATH, sep="\t").rename(columns=lambda x: x.lower()) more_3_to_2 = self.iso_table.set_index("id").part1.dropna().to_dict() more_3_to_2.update(mapper.set_index("a3").a2.to_dict()) self.alpha3_to_alpha2 = more_3_to_2 self.model_card_dir = Path(save_dir) self.constituents = GROUP_MEMBERS def convert_models(self, tatoeba_ids, dry_run=False): entries_to_convert = [x for x in self.registry if x[0] in tatoeba_ids] converted_paths = convert_all_sentencepiece_models(entries_to_convert, dest_dir=self.model_card_dir) for path in converted_paths: long_pair = remove_prefix(path.name, "opus-mt-").split("-") # eg. heb-eng assert len(long_pair) == 2 new_p_src = self.get_two_letter_code(long_pair[0]) new_p_tgt = self.get_two_letter_code(long_pair[1]) hf_model_id = f"opus-mt-{new_p_src}-{new_p_tgt}" new_path = path.parent.joinpath(hf_model_id) # opus-mt-he-en os.rename(str(path), str(new_path)) self.write_model_card(hf_model_id, dry_run=dry_run) def get_two_letter_code(self, three_letter_code): return self.alpha3_to_alpha2.get(three_letter_code, three_letter_code) def expand_group_to_two_letter_codes(self, grp_name): return [self.get_two_letter_code(x) for x in self.constituents[grp_name]] def get_tags(self, code, ref_name): if len(code) == 2: assert "languages" not in ref_name, f"{code}: {ref_name}" return [code], False elif "languages" in ref_name or len(self.constituents.get(code, [])) > 1: group = self.expand_group_to_two_letter_codes(code) group.append(code) return group, True else: # zho-> zh print(f"Three letter monolingual code: {code}") return [code], False def resolve_lang_code(self, r) -> Tuple[List[str], str, str]: """R is a row in ported""" short_pair = r.short_pair src, tgt = short_pair.split("-") src_tags, src_multilingual = self.get_tags(src, r.src_name) assert isinstance(src_tags, list) tgt_tags, tgt_multilingual = self.get_tags(tgt, r.tgt_name) assert isinstance(tgt_tags, list) return dedup(src_tags + tgt_tags), src_multilingual, tgt_multilingual def write_model_card( self, hf_model_id: str, repo_root=DEFAULT_REPO, dry_run=False, ) -> str: """ Copy the most recent model's readme section from opus, and add metadata. upload command: aws s3 sync model_card_dir s3://models.huggingface.co/bert/Helsinki-NLP/ --dryrun """ short_pair = remove_prefix(hf_model_id, "opus-mt-") extra_metadata = self.metadata.loc[short_pair].drop("2m") extra_metadata["short_pair"] = short_pair lang_tags, src_multilingual, tgt_multilingual = self.resolve_lang_code(extra_metadata) opus_name = f"{extra_metadata.src_alpha3}-{extra_metadata.tgt_alpha3}" # opus_name: str = self.convert_hf_name_to_opus_name(hf_model_name) assert repo_root in ("OPUS-MT-train", "Tatoeba-Challenge") opus_readme_path = Path(repo_root).joinpath("models", opus_name, "README.md") assert opus_readme_path.exists(), f"Readme file {opus_readme_path} not found" opus_src, opus_tgt = [x.split("+") for x in opus_name.split("-")] readme_url = f"https://github.com/Helsinki-NLP/{repo_root}/tree/master/models/{opus_name}/README.md" s, t = ",".join(opus_src), ",".join(opus_tgt) metadata = { "hf_name": short_pair, "source_languages": s, "target_languages": t, "opus_readme_url": readme_url, "original_repo": repo_root, "tags": ["translation"], "languages": lang_tags, } lang_tags = l2front_matter(lang_tags) metadata["src_constituents"] = self.constituents[s] metadata["tgt_constituents"] = self.constituents[t] metadata["src_multilingual"] = src_multilingual metadata["tgt_multilingual"] = tgt_multilingual metadata.update(extra_metadata) metadata.update(get_system_metadata(repo_root)) # combine with Tatoeba markdown extra_markdown = f"### {short_pair}\n\n* source group: {metadata['src_name']} \n* target group: {metadata['tgt_name']} \n* OPUS readme: [{opus_name}]({readme_url})\n" content = opus_readme_path.open().read() content = content.split("\n# ")[-1] # Get the lowest level 1 header in the README -- the most recent model. splat = content.split("*")[2:] content = "*".join(splat) # BETTER FRONT MATTER LOGIC content = ( FRONT_MATTER_TEMPLATE.format(lang_tags) + extra_markdown + "\n* " + content.replace("download", "download original " "weights") ) items = "\n".join([f"- {k}: {v}" for k, v in metadata.items()]) sec3 = "\n### System Info: \n" + items content += sec3 if dry_run: return content, metadata sub_dir = self.model_card_dir / hf_model_id sub_dir.mkdir(exist_ok=True) dest = sub_dir / "README.md" dest.open("w").write(content) pd.Series(metadata).to_json(sub_dir / "metadata.json") return content, metadata def download_metadata(self): Path(LANG_CODE_PATH).parent.mkdir(exist_ok=True) import wget if not os.path.exists(ISO_PATH): wget.download(ISO_URL, ISO_PATH) if not os.path.exists(LANG_CODE_PATH): wget.download(LANG_CODE_URL, LANG_CODE_PATH) @staticmethod def make_tatoeba_registry(repo_path=DEFAULT_MODEL_DIR): if not (Path(repo_path) / "zho-eng" / "README.md").exists(): raise ValueError( f"repo_path:{repo_path} does not exist: " "You must run: git clone git@github.com:Helsinki-NLP/Tatoeba-Challenge.git before calling." ) results = {} for p in Path(repo_path).iterdir(): if len(p.name) != 7: continue lns = list(open(p / "README.md").readlines()) results[p.name] = _parse_readme(lns) new_list = [(k, v) for k, v in results.items()] new_list_2 = [] for element in new_list: try: new_list_2.append((element[0], element[1]["pre-processing"], element[1]["download"], element[1]["download"][:-4] + ".test.txt")) except: print("Skipped element " + element[0]) return new_list_2 GROUP_MEMBERS = { # three letter code -> (group/language name, {constituents...} # if this language is on the target side the constituents can be used as target language codes. # if the language is on the source side they are supported natively without special codes. "aav": ("Austro-Asiatic languages", {"hoc", "hoc_Latn", "kha", "khm", "khm_Latn", "mnw", "vie", "vie_Hani"}), "afa": ( "Afro-Asiatic languages", { "acm", "afb", "amh", "apc", "ara", "arq", "ary", "arz", "hau_Latn", "heb", "kab", "mlt", "rif_Latn", "shy_Latn", "som", "thv", "tir", }, ), "afr": ("Afrikaans", {"afr"}), "alv": ( "Atlantic-Congo languages", { "ewe", "fuc", "fuv", "ibo", "kin", "lin", "lug", "nya", "run", "sag", "sna", "swh", "toi_Latn", "tso", "umb", "wol", "xho", "yor", "zul", }, ), "ara": ("Arabic", {"afb", "apc", "apc_Latn", "ara", "ara_Latn", "arq", "arq_Latn", "arz"}), "art": ( "Artificial languages", { "afh_Latn", "avk_Latn", "dws_Latn", "epo", "ido", "ido_Latn", "ile_Latn", "ina_Latn", "jbo", "jbo_Cyrl", "jbo_Latn", "ldn_Latn", "lfn_Cyrl", "lfn_Latn", "nov_Latn", "qya", "qya_Latn", "sjn_Latn", "tlh_Latn", "tzl", "tzl_Latn", "vol_Latn", }, ), "aze": ("Azerbaijani", {"aze_Latn"}), "bat": ("Baltic languages", {"lit", "lav", "prg_Latn", "ltg", "sgs"}), "bel": ("Belarusian", {"bel", "bel_Latn"}), "ben": ("Bengali", {"ben"}), "bnt": ( "Bantu languages", {"kin", "lin", "lug", "nya", "run", "sna", "swh", "toi_Latn", "tso", "umb", "xho", "zul"}, ), "bul": ("Bulgarian", {"bul", "bul_Latn"}), "cat": ("Catalan", {"cat"}), "cau": ("Caucasian languages", {"abk", "kat", "che", "ady"}), "ccs": ("South Caucasian languages", {"kat"}), "ceb": ("Cebuano", {"ceb"}), "cel": ("Celtic languages", {"gla", "gle", "bre", "cor", "glv", "cym"}), "ces": ("Czech", {"ces"}), "cpf": ("Creoles and pidgins, French‑based", {"gcf_Latn", "hat", "mfe"}), "cpp": ( "Creoles and pidgins, Portuguese-based", {"zsm_Latn", "ind", "pap", "min", "tmw_Latn", "max_Latn", "zlm_Latn"}, ), "cus": ("Cushitic languages", {"som"}), "dan": ("Danish", {"dan"}), "deu": ("German", {"deu"}), "dra": ("Dravidian languages", {"tam", "kan", "mal", "tel"}), "ell": ("Modern Greek (1453-)", {"ell"}), "eng": ("English", {"eng"}), "epo": ("Esperanto", {"epo"}), "est": ("Estonian", {"est"}), "euq": ("Basque (family)", {"eus"}), "eus": ("Basque", {"eus"}), "fin": ("Finnish", {"fin"}), "fiu": ( "Finno-Ugrian languages", { "est", "fin", "fkv_Latn", "hun", "izh", "kpv", "krl", "liv_Latn", "mdf", "mhr", "myv", "sma", "sme", "udm", "vep", "vro", }, ), "fra": ("French", {"fra"}), "gem": ( "Germanic languages", { "afr", "ang_Latn", "dan", "deu", "eng", "enm_Latn", "fao", "frr", "fry", "gos", "got_Goth", "gsw", "isl", "ksh", "ltz", "nds", "nld", "nno", "nob", "nob_Hebr", "non_Latn", "pdc", "sco", "stq", "swe", "swg", "yid", }, ), "gle": ("Irish", {"gle"}), "glg": ("Galician", {"glg"}), "gmq": ("North Germanic languages", {"dan", "nob", "nob_Hebr", "swe", "isl", "nno", "non_Latn", "fao"}), "gmw": ( "West Germanic languages", { "afr", "ang_Latn", "deu", "eng", "enm_Latn", "frr", "fry", "gos", "gsw", "ksh", "ltz", "nds", "nld", "pdc", "sco", "stq", "swg", "yid", }, ), "grk": ("Greek languages", {"grc_Grek", "ell"}), "hbs": ("Serbo-Croatian", {"hrv", "srp_Cyrl", "bos_Latn", "srp_Latn"}), "heb": ("Hebrew", {"heb"}), "hin": ("Hindi", {"hin"}), "hun": ("Hungarian", {"hun"}), "hye": ("Armenian", {"hye", "hye_Latn"}), "iir": ( "Indo-Iranian languages", { "asm", "awa", "ben", "bho", "gom", "guj", "hif_Latn", "hin", "jdt_Cyrl", "kur_Arab", "kur_Latn", "mai", "mar", "npi", "ori", "oss", "pan_Guru", "pes", "pes_Latn", "pes_Thaa", "pnb", "pus", "rom", "san_Deva", "sin", "snd_Arab", "tgk_Cyrl", "tly_Latn", "urd", "zza", }, ), "ilo": ("Iloko", {"ilo"}), "inc": ( "Indic languages", { "asm", "awa", "ben", "bho", "gom", "guj", "hif_Latn", "hin", "mai", "mar", "npi", "ori", "pan_Guru", "pnb", "rom", "san_Deva", "sin", "snd_Arab", "urd", }, ), "ine": ( "Indo-European languages", { "afr", "afr_Arab", "aln", "ang_Latn", "arg", "asm", "ast", "awa", "bel", "bel_Latn", "ben", "bho", "bjn", "bos_Latn", "bre", "bul", "bul_Latn", "cat", "ces", "cor", "cos", "csb_Latn", "cym", "dan", "deu", "dsb", "egl", "ell", "eng", "enm_Latn", "ext", "fao", "fra", "frm_Latn", "frr", "fry", "gcf_Latn", "gla", "gle", "glg", "glv", "gom", "gos", "got_Goth", "grc_Grek", "gsw", "guj", "hat", "hif_Latn", "hin", "hrv", "hsb", "hye", "hye_Latn", "ind", "isl", "ita", "jdt_Cyrl", "ksh", "kur_Arab", "kur_Latn", "lad", "lad_Latn", "lat_Grek", "lat_Latn", "lav", "lij", "lit", "lld_Latn", "lmo", "ltg", "ltz", "mai", "mar", "max_Latn", "mfe", "min", "mkd", "mwl", "nds", "nld", "nno", "nob", "nob_Hebr", "non_Latn", "npi", "oci", "ori", "orv_Cyrl", "oss", "pan_Guru", "pap", "pcd", "pdc", "pes", "pes_Latn", "pes_Thaa", "pms", "pnb", "pol", "por", "prg_Latn", "pus", "roh", "rom", "ron", "rue", "rus", "rus_Latn", "san_Deva", "scn", "sco", "sgs", "sin", "slv", "snd_Arab", "spa", "sqi", "srd", "srp_Cyrl", "srp_Latn", "stq", "swe", "swg", "tgk_Cyrl", "tly_Latn", "tmw_Latn", "ukr", "urd", "vec", "wln", "yid", "zlm_Latn", "zsm_Latn", "zza", }, ), "isl": ("Icelandic", {"isl"}), "ita": ("Italian", {"ita"}), "itc": ( "Italic languages", { "arg", "ast", "bjn", "cat", "cos", "egl", "ext", "fra", "frm_Latn", "gcf_Latn", "glg", "hat", "ind", "ita", "lad", "lad_Latn", "lat_Grek", "lat_Latn", "lij", "lld_Latn", "lmo", "max_Latn", "mfe", "min", "mwl", "oci", "pap", "pcd", "pms", "por", "roh", "ron", "scn", "spa", "srd", "tmw_Latn", "vec", "wln", "zlm_Latn", "zsm_Latn", }, ), "jpn": ("Japanese", {"jpn", "jpn_Bopo", "jpn_Hang", "jpn_Hani", "jpn_Hira", "jpn_Kana", "jpn_Latn", "jpn_Yiii"}), "jpx": ("Japanese (family)", {"jpn"}), "kat": ("Georgian", {"kat"}), "kor": ("Korean", {"kor_Hani", "kor_Hang", "kor_Latn", "kor"}), "lav": ("Latvian", {"lav"}), "lit": ("Lithuanian", {"lit"}), "mkd": ("Macedonian", {"mkd"}), "mkh": ("Mon-Khmer languages", {"vie_Hani", "mnw", "vie", "kha", "khm_Latn", "khm"}), "msa": ("Malay (macrolanguage)", {"zsm_Latn", "ind", "max_Latn", "zlm_Latn", "min"}), "mul": ( "Multiple languages", { "abk", "acm", "ady", "afb", "afh_Latn", "afr", "akl_Latn", "aln", "amh", "ang_Latn", "apc", "ara", "arg", "arq", "ary", "arz", "asm", "ast", "avk_Latn", "awa", "aze_Latn", "bak", "bam_Latn", "bel", "bel_Latn", "ben", "bho", "bod", "bos_Latn", "bre", "brx", "brx_Latn", "bul", "bul_Latn", "cat", "ceb", "ces", "cha", "che", "chr", "chv", "cjy_Hans", "cjy_Hant", "cmn", "cmn_Hans", "cmn_Hant", "cor", "cos", "crh", "crh_Latn", "csb_Latn", "cym", "dan", "deu", "dsb", "dtp", "dws_Latn", "egl", "ell", "enm_Latn", "epo", "est", "eus", "ewe", "ext", "fao", "fij", "fin", "fkv_Latn", "fra", "frm_Latn", "frr", "fry", "fuc", "fuv", "gan", "gcf_Latn", "gil", "gla", "gle", "glg", "glv", "gom", "gos", "got_Goth", "grc_Grek", "grn", "gsw", "guj", "hat", "hau_Latn", "haw", "heb", "hif_Latn", "hil", "hin", "hnj_Latn", "hoc", "hoc_Latn", "hrv", "hsb", "hun", "hye", "iba", "ibo", "ido", "ido_Latn", "ike_Latn", "ile_Latn", "ilo", "ina_Latn", "ind", "isl", "ita", "izh", "jav", "jav_Java", "jbo", "jbo_Cyrl", "jbo_Latn", "jdt_Cyrl", "jpn", "kab", "kal", "kan", "kat", "kaz_Cyrl", "kaz_Latn", "kek_Latn", "kha", "khm", "khm_Latn", "kin", "kir_Cyrl", "kjh", "kpv", "krl", "ksh", "kum", "kur_Arab", "kur_Latn", "lad", "lad_Latn", "lao", "lat_Latn", "lav", "ldn_Latn", "lfn_Cyrl", "lfn_Latn", "lij", "lin", "lit", "liv_Latn", "lkt", "lld_Latn", "lmo", "ltg", "ltz", "lug", "lzh", "lzh_Hans", "mad", "mah", "mai", "mal", "mar", "max_Latn", "mdf", "mfe", "mhr", "mic", "min", "mkd", "mlg", "mlt", "mnw", "moh", "mon", "mri", "mwl", "mww", "mya", "myv", "nan", "nau", "nav", "nds", "niu", "nld", "nno", "nob", "nob_Hebr", "nog", "non_Latn", "nov_Latn", "npi", "nya", "oci", "ori", "orv_Cyrl", "oss", "ota_Arab", "ota_Latn", "pag", "pan_Guru", "pap", "pau", "pdc", "pes", "pes_Latn", "pes_Thaa", "pms", "pnb", "pol", "por", "ppl_Latn", "prg_Latn", "pus", "quc", "qya", "qya_Latn", "rap", "rif_Latn", "roh", "rom", "ron", "rue", "run", "rus", "sag", "sah", "san_Deva", "scn", "sco", "sgs", "shs_Latn", "shy_Latn", "sin", "sjn_Latn", "slv", "sma", "sme", "smo", "sna", "snd_Arab", "som", "spa", "sqi", "srp_Cyrl", "srp_Latn", "stq", "sun", "swe", "swg", "swh", "tah", "tam", "tat", "tat_Arab", "tat_Latn", "tel", "tet", "tgk_Cyrl", "tha", "tir", "tlh_Latn", "tly_Latn", "tmw_Latn", "toi_Latn", "ton", "tpw_Latn", "tso", "tuk", "tuk_Latn", "tur", "tvl", "tyv", "tzl", "tzl_Latn", "udm", "uig_Arab", "uig_Cyrl", "ukr", "umb", "urd", "uzb_Cyrl", "uzb_Latn", "vec", "vie", "vie_Hani", "vol_Latn", "vro", "war", "wln", "wol", "wuu", "xal", "xho", "yid", "yor", "yue", "yue_Hans", "yue_Hant", "zho", "zho_Hans", "zho_Hant", "zlm_Latn", "zsm_Latn", "zul", "zza", }, ), "nic": ( "Niger-Kordofanian languages", { "bam_Latn", "ewe", "fuc", "fuv", "ibo", "kin", "lin", "lug", "nya", "run", "sag", "sna", "swh", "toi_Latn", "tso", "umb", "wol", "xho", "yor", "zul", }, ), "nld": ("Dutch", {"nld"}), "nor": ("Norwegian", {"nob", "nno"}), "phi": ("Philippine languages", {"ilo", "akl_Latn", "war", "hil", "pag", "ceb"}), "pol": ("Polish", {"pol"}), "por": ("Portuguese", {"por"}), "pqe": ( "Eastern Malayo-Polynesian languages", {"fij", "gil", "haw", "mah", "mri", "nau", "niu", "rap", "smo", "tah", "ton", "tvl"}, ), "roa": ( "Romance languages", { "arg", "ast", "cat", "cos", "egl", "ext", "fra", "frm_Latn", "gcf_Latn", "glg", "hat", "ind", "ita", "lad", "lad_Latn", "lij", "lld_Latn", "lmo", "max_Latn", "mfe", "min", "mwl", "oci", "pap", "pms", "por", "roh", "ron", "scn", "spa", "tmw_Latn", "vec", "wln", "zlm_Latn", "zsm_Latn", }, ), "ron": ("Romanian", {"ron"}), "run": ("Rundi", {"run"}), "rus": ("Russian", {"rus"}), "sal": ("Salishan languages", {"shs_Latn"}), "sem": ("Semitic languages", {"acm", "afb", "amh", "apc", "ara", "arq", "ary", "arz", "heb", "mlt", "tir"}), "sla": ( "Slavic languages", { "bel", "bel_Latn", "bos_Latn", "bul", "bul_Latn", "ces", "csb_Latn", "dsb", "hrv", "hsb", "mkd", "orv_Cyrl", "pol", "rue", "rus", "slv", "srp_Cyrl", "srp_Latn", "ukr", }, ), "slv": ("Slovenian", {"slv"}), "spa": ("Spanish", {"spa"}), "swe": ("Swedish", {"swe"}), "taw": ("Tai", {"lao", "tha"}), "tgl": ("Tagalog", {"tgl_Latn"}), "tha": ("Thai", {"tha"}), "trk": ( "Turkic languages", { "aze_Latn", "bak", "chv", "crh", "crh_Latn", "kaz_Cyrl", "kaz_Latn", "kir_Cyrl", "kjh", "kum", "ota_Arab", "ota_Latn", "sah", "tat", "tat_Arab", "tat_Latn", "tuk", "tuk_Latn", "tur", "tyv", "uig_Arab", "uig_Cyrl", "uzb_Cyrl", "uzb_Latn", }, ), "tur": ("Turkish", {"tur"}), "ukr": ("Ukrainian", {"ukr"}), "urd": ("Urdu", {"urd"}), "urj": ( "Uralic languages", { "est", "fin", "fkv_Latn", "hun", "izh", "kpv", "krl", "liv_Latn", "mdf", "mhr", "myv", "sma", "sme", "udm", "vep", "vro", }, ), "vie": ("Vietnamese", {"vie", "vie_Hani"}), "war": ("Waray (Philippines)", {"war"}), "zho": ( "Chinese", { "cjy_Hans", "cjy_Hant", "cmn", "cmn_Bopo", "cmn_Hang", "cmn_Hani", "cmn_Hans", "cmn_Hant", "cmn_Hira", "cmn_Kana", "cmn_Latn", "cmn_Yiii", "gan", "hak_Hani", "lzh", "lzh_Bopo", "lzh_Hang", "lzh_Hani", "lzh_Hans", "lzh_Hira", "lzh_Kana", "lzh_Yiii", "nan", "nan_Hani", "wuu", "wuu_Bopo", "wuu_Hani", "wuu_Latn", "yue", "yue_Bopo", "yue_Hang", "yue_Hani", "yue_Hans", "yue_Hant", "yue_Hira", "yue_Kana", "zho", "zho_Hans", "zho_Hant", }, ), "zle": ("East Slavic languages", {"bel", "orv_Cyrl", "bel_Latn", "rus", "ukr", "rue"}), "zls": ("South Slavic languages", {"bos_Latn", "bul", "bul_Latn", "hrv", "mkd", "slv", "srp_Cyrl", "srp_Latn"}), "zlw": ("West Slavic languages", {"csb_Latn", "dsb", "hsb", "pol", "ces"}), } def l2front_matter(langs): return "".join(f"- {l}\n" for l in langs) def dedup(lst): """Preservers order""" new_lst = [] for item in lst: if not item: continue elif item in new_lst: continue else: new_lst.append(item) return new_lst if __name__ == "__main__": parser = argparse.ArgumentParser() parser.add_argument( "-m", "--models", action="append", help="<Required> Set flag", required=True, nargs="+", dest="models" ) parser.add_argument("-save_dir", "--save_dir", default="marian_converted", help="where to save converted models") args = parser.parse_args() resolver = TatoebaConverter(save_dir=args.save_dir) resolver.convert_models(args.models[0])
34,591
25.94081
175
py
defend_framework
defend_framework-main/DPA_certified_train.py
import torch import os import json import time import numpy as np from utils.data_processing import MNIST17DataPreprocessor, MNISTDataPreprocessor, IMDBDataPreprocessor, \ EmberDataPreProcessor, EMBER_DATASET, EmberPoisonDataPreProcessor, MNIST01DataPreprocessor, \ MNIST17LimitedDataPreprocessor, FMNISTDataPreprocessor, CIFARDataPreprocessor from models import MNISTLiRPAModel, EmberLiRPAModel from utils.train_utils import train_many from utils.cert_train_argments import get_arguments if __name__ == "__main__": parser = get_arguments() # Set random seeds args = parser.parse_args() os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # random.seed(args.seed) # np.random.seed(args.seed) # tf.random.set_seed(args.seed) # make dirs if args.exp_name is None: args.exp_name = time.strftime("%Y%m%d-%H%M%S") if args.model_save_dir is not None: if not os.path.exists(args.model_save_dir): os.mkdir(args.model_save_dir) os.mkdir(os.path.join(args.model_save_dir, args.exp_name)) res = None res_noise = None if args.res_save_dir is not None: if not os.path.exists(args.res_save_dir): os.mkdir(args.res_save_dir) if os.path.exists(os.path.join(args.res_save_dir, args.exp_name)): respond = input("Experiment already exists, type [O] to overwrite, type [R] to resume") if respond == "O": pass elif respond == "R": res, res_noise = np.load(os.path.join(args.res_save_dir, args.exp_name, "aggre_res.npy")) else: exit(0) else: os.mkdir(os.path.join(args.res_save_dir, args.exp_name)) assert args.res_save_dir is not None and args.exp_name is not None if args.dataset == "mnist": if args.load_poison_dir is not None: data_loader = MNISTDataPreprocessor.load(os.path.join(args.load_poison_dir, "data"), args) else: data_loader = MNISTDataPreprocessor(args) model = MNISTLiRPAModel.MNISTModel(data_loader.n_features, data_loader.n_classes, args, device, lr=args.lr) elif args.dataset == "ember": if args.load_poison_dir is not None: data_loader = EmberPoisonDataPreProcessor(args) else: data_loader = EmberDataPreProcessor(args) model = EmberLiRPAModel.EmberModel(data_loader.n_features, data_loader.n_classes, args, device, lr=args.lr) else: raise NotImplementedError with open(os.path.join(args.res_save_dir, args.exp_name, "commandline_args.txt"), 'w') as f: json.dump(args.__dict__, f, indent=2) aggregate_results = train_many(data_loader, model, args, res, res_noise)
2,833
39.485714
115
py
defend_framework
defend_framework-main/gen_attack_dataset.py
import random import os import json import time import numpy as np import tensorflow as tf from tensorflow import keras from utils.data_processing import MNISTDataPreprocessor, MNIST17DataPreprocessor, MNIST01DataPreprocessor, \ CIFAR02DataPreprocessor from models.MNISTModel import MNISTModel, MNIST17Model, MNIST01Model from models.CIFAR10Model import CIFAR10Model from utils.train_utils import train_single from attack.BadNetAttack import BadNetAttackLabel, BadNetAttackNoLabel from utils.cert_train_argments import get_arguments if __name__ == "__main__": parser = get_arguments() # attack parameters parser.add_argument("--attack", choices=["badnet"], help="attack algorithms", required=True) parser.add_argument("--consecutive", action="store_true", help="Whether the poisoned features need to be inside a block") parser.add_argument("--attack_label", action="store_true", help="Whether to attack the label of the training image") parser.add_argument("--poisoned_feat_num", type=int, required=True, help="poisoned feature number") parser.add_argument("--poisoned_ins_rate", default=0.1, type=float, help="the rate of instances to be poisoned") parser.add_argument("--attack_targets", type=str, help="A list of ints of length n_classes, attacking label i to its target attack_targets[i], " "attack_targets[i] can be None.") # dirs and files parser.add_argument("--save_poison_dir", type=str, help="dir for save poisoned dataset" ) parser.add_argument("--load", action="store_true", help="whether to load the saved file") # Set random seeds args = parser.parse_args() os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_id gpus = tf.config.list_physical_devices('GPU') if gpus: try: # Currently, memory growth needs to be the same across GPUs for gpu in gpus: tf.config.experimental.set_memory_growth(gpu, True) logical_gpus = tf.config.list_logical_devices('GPU') print(len(gpus), "Physical GPUs,", len(logical_gpus), "Logical GPUs") except RuntimeError as e: # Memory growth must be set before GPUs have been initialized print(e) # random.seed(args.seed) # np.random.seed(args.seed) # tf.random.set_seed(args.seed) # make dirs if not os.path.exists(args.save_poison_dir): os.mkdir(args.save_poison_dir) if args.exp_name is None: args.exp_name = time.strftime("%Y%m%d-%H%M%S") filepath = os.path.join(args.save_poison_dir, args.exp_name) if args.load: assert os.path.exists(filepath) and os.path.exists(os.path.join(filepath, "data")) if os.path.exists(filepath) and not args.load: respond = input("Experiment already exists, type [Y] to overwrite") if respond != "Y": exit(0) elif not os.path.exists(filepath): os.mkdir(filepath) if args.dataset == "mnist": DataPreprocessor_type = MNISTDataPreprocessor Model_type = MNISTModel elif args.dataset == "mnist17": DataPreprocessor_type = MNIST17DataPreprocessor Model_type = MNIST17Model elif args.dataset == "mnist01": DataPreprocessor_type = MNIST01DataPreprocessor Model_type = MNIST01Model elif args.dataset == "cifar10-02": DataPreprocessor_type = CIFAR02DataPreprocessor Model_type = CIFAR10Model else: raise NotImplementedError with open(os.path.join(filepath, "commandline_args.txt"), 'w') as f: json.dump(args.__dict__, f, indent=2) attack_targets = eval(args.attack_targets) if not args.load: data_loader = DataPreprocessor_type(args) if args.attack_label: attack = BadNetAttackLabel(data_loader, attack_targets, args.poisoned_feat_num, consecutive=args.consecutive, poisoned_ins_rate=args.poisoned_ins_rate) else: attack = BadNetAttackNoLabel(data_loader, attack_targets, args.poisoned_feat_num, consecutive=args.consecutive, poisoned_ins_rate=args.poisoned_ins_rate) attack.attack() attack.save(os.path.join(filepath, "data")) else: if args.attack_label: attack = BadNetAttackLabel.load(os.path.join(filepath, "data")) else: attack = BadNetAttackNoLabel.load(os.path.join(filepath, "data")) data_loader = attack.data_processor model = Model_type(data_loader.n_features, data_loader.n_classes) train_single(data_loader, model, args) print("Clean Test Set:") model.evaluate(data_loader.x_test, keras.utils.to_categorical(data_loader.y_test, data_loader.n_classes)) print("Poisoned Test Set:") for i in range(data_loader.n_classes): idx = np.where(data_loader.y_test == i)[0] if attack_targets[i] is None: print(f"class {i} is not poisoned:") model.evaluate(data_loader.x_test_poisoned[idx], keras.utils.to_categorical(data_loader.y_test_poisoned[idx], data_loader.n_classes)) else: print(f"class {i} is poisoned:") model.evaluate(data_loader.x_test_poisoned[idx], keras.utils.to_categorical(data_loader.y_test_poisoned[idx], data_loader.n_classes))
5,498
42.642857
118
py
defend_framework
defend_framework-main/models/LiRPAModel.py
from abc import ABC, abstractmethod from torch.utils.data import TensorDataset, DataLoader import torch.optim as optim from torch.nn import CrossEntropyLoss from auto_LiRPA import BoundedModule, BoundedTensor from auto_LiRPA.perturbations import * from auto_LiRPA.eps_scheduler import LinearScheduler, AdaptiveScheduler, SmoothedScheduler, FixedScheduler from auto_LiRPA.utils import MultiAverageMeter class LiRPAModel(ABC): def __init__(self, n_features, n_classes, args, device, model_ori, lr=1e-3): self.input_shape = n_features self.n_classes = n_classes self.lr = lr self.args = args self.device = device self.init_model = model_ori self.init() def init(self): ## Step 1: Initial original model as usual self.model_ori = self.init_model(in_ch=self.input_shape[-1], in_dim=self.input_shape[0]) ## Step 3: wrap model with auto_LiRPA # The second parameter dummy_input is for constructing the trace of the computational graph. dummy_input = torch.randn(*([2] + list(self.input_shape))) self.model = BoundedModule(self.model_ori, dummy_input, bound_opts={'relu': self.args.bound_opts, 'conv_mode': self.args.conv_mode}, device=self.device) # print("Model structure: \n", str(self.model_ori)) def fit(self, X, y, batch_size, epochs, data_aug=None): data = TensorDataset(torch.Tensor(X), torch.Tensor(y).long().max(dim=-1)[1]) loader = DataLoader(data, batch_size=batch_size, shuffle=True, pin_memory=True) ## Step 4 prepare optimizer, epsilon scheduler and learning rate scheduler eps_scheduler = eval(self.args.scheduler_name)(self.args.eps, self.args.scheduler_opts) opt = optim.Adam(self.model.parameters(), lr=self.lr) lr_scheduler = optim.lr_scheduler.StepLR(opt, step_size=10, gamma=0.5) for t in range(1, epochs + 1): if eps_scheduler.reached_max_eps(): # Only decay learning rate after reaching the maximum eps lr_scheduler.step() self.train(eps_scheduler, opt, loader, data_aug) def evaluate(self, x_test, y_test): data = TensorDataset(torch.Tensor(x_test), torch.Tensor(y_test).long().max(dim=-1)[1]) loader = DataLoader(data, batch_size=256, shuffle=False, pin_memory=True) eps_scheduler = FixedScheduler(self.args.eps) with torch.no_grad(): predictions, verified = self.test(eps_scheduler, loader) print("Test accuracy: ", np.mean(predictions == np.argmax(y_test, axis=-1))) predictions_cert = predictions * verified + (1 - verified) * self.n_classes print("Certified accuracy: ", np.mean(predictions_cert == np.argmax(y_test, axis=-1))) return predictions, predictions_cert def test(self, eps_scheduler, loader): norm = float(self.args.norm) self.model.eval() eps_scheduler.eval() predictions = np.array([], dtype=np.int) verified = np.array([], dtype=np.bool) for i, (data, _) in enumerate(loader): eps_scheduler.step_batch() eps = eps_scheduler.get_eps() # bound input for Linf norm used only data_ub = data_lb = data if list(self.model.parameters())[0].is_cuda: data = data.cuda() data_lb, data_ub = data_lb.cuda(), data_ub.cuda() # Specify Lp norm perturbation. # When using Linf perturbation, we manually set element-wise bound x_L and x_U. eps is not used for Linf norm. if norm > 0: ptb = PerturbationLpNorm(norm=norm, eps=eps, x_L=data_lb, x_U=data_ub) elif norm == 0: ptb = PerturbationL0Norm(eps=eps_scheduler.get_max_eps(), ratio=eps_scheduler.get_eps() / eps_scheduler.get_max_eps()) x = BoundedTensor(data, ptb) output = self.model(x) batch_predictions = torch.argmax(output, dim=1) # meter.update('CE', regular_ce.item(), x.size(0)) # meter.update('Err', torch.sum(torch.argmax(output, dim=1) != labels).cpu().detach().numpy() / x.size(0), # x.size(0)) # generate specifications c = torch.eye(self.n_classes).type_as(data)[batch_predictions].unsqueeze(1) - torch.eye( self.n_classes).type_as( data).unsqueeze(0) # remove specifications to self I = (~(batch_predictions.data.unsqueeze(1) == torch.arange(self.n_classes).type_as( batch_predictions.data).unsqueeze(0))) c = (c[I].view(data.size(0), self.n_classes - 1, self.n_classes)) if list(self.model.parameters())[0].is_cuda: c = c.cuda() if self.args.bound_type == "IBP": lb, ub = self.model.compute_bounds(IBP=True, C=c, method=None) elif self.args.bound_type == "CROWN": lb, ub = self.model.compute_bounds(IBP=False, C=c, method="backward", bound_upper=False) elif self.args.bound_type == "CROWN-IBP": # lb, ub = model.compute_bounds(ptb=ptb, IBP=True, x=data, C=c, method="backward") # pure IBP bound # we use a mixed IBP and CROWN-IBP bounds, leading to better performance (Zhang et al., ICLR 2020) factor = (eps_scheduler.get_max_eps() - eps) / eps_scheduler.get_max_eps() ilb, iub = self.model.compute_bounds(IBP=True, C=c, method=None) if factor < 1e-5: lb = ilb else: clb, cub = self.model.compute_bounds(IBP=False, C=c, method="backward", bound_upper=False) lb = clb * factor + ilb * (1 - factor) elif self.args.bound_type == "CROWN-FAST": # Similar to CROWN-IBP but no mix between IBP and CROWN bounds. lb, ub = self.model.compute_bounds(IBP=True, C=c, method=None) lb, ub = self.model.compute_bounds(IBP=False, C=c, method="backward", bound_upper=False) # Pad zero at the beginning for each example, and use fake label "0" for all examples # lb_padded = torch.cat((torch.zeros(size=(lb.size(0), 1), dtype=lb.dtype, device=lb.device), lb), dim=1) # fake_labels = torch.zeros(size=(lb.size(0),), dtype=torch.int64, device=lb.device) # robust_ce = CrossEntropyLoss()(-lb_padded, fake_labels) # meter.update('Loss', loss.item(), data.size(0)) # if batch_method != "natural": # meter.update('Robust_CE', robust_ce.item(), data.size(0)) # # For an example, if lower bounds of margins is >0 for all classes, the output is verifiably correct. # # If any margin is < 0 this example is counted as an error # meter.update('Verified_Err', torch.sum((lb < 0).any(dim=1)).item() / data.size(0), data.size(0)) # if i % 50 == 0 and train: # print('[{:4d}]: eps={:.8f} {}'.format(i, eps, meter)) batch_predictions = batch_predictions.cpu().numpy() batch_verified = (lb > 0).all(dim=1).cpu().numpy() predictions = np.append(predictions, batch_predictions) verified = np.append(verified, batch_verified) # print('[{:4d}]: eps={:.8f} {}'.format(i, eps, meter)) return predictions, verified def train(self, eps_scheduler, opt, loader, data_aug): norm = float(self.args.norm) meter = MultiAverageMeter() self.model.train() eps_scheduler.train() eps_scheduler.step_epoch() eps_scheduler.set_epoch_length(int((len(loader.dataset) + loader.batch_size - 1) / loader.batch_size)) for i, (data, labels) in enumerate(loader): eps_scheduler.step_batch() eps = eps_scheduler.get_eps() # For small eps just use natural training, no need to compute LiRPA bounds batch_method = "robust" if eps < 1e-20: batch_method = "natural" opt.zero_grad() # generate specifications c = torch.eye(self.n_classes).type_as(data)[labels].unsqueeze(1) - torch.eye(self.n_classes).type_as( data).unsqueeze(0) # remove specifications to self I = (~(labels.data.unsqueeze(1) == torch.arange(self.n_classes).type_as(labels.data).unsqueeze(0))) c = (c[I].view(data.size(0), self.n_classes - 1, self.n_classes)) # bound input for Linf norm used only # data_ub = data_lb = data if list(self.model.parameters())[0].is_cuda: data, labels, c = data.cuda(), labels.cuda(), c.cuda() # data_lb, data_ub = data_lb.cuda(), data_ub.cuda() if data_aug is not None: data = data_aug(data) # Specify Lp norm perturbation. # When using Linf perturbation, we manually set element-wise bound x_L and x_U. eps is not used for Linf norm. # if norm > 0: # ptb = PerturbationLpNorm(norm=norm, eps=eps, x_L=data_lb, x_U=data_ub) # elif norm == 0: ptb = PerturbationL0Norm(eps=eps_scheduler.get_max_eps(), ratio=eps_scheduler.get_eps() / eps_scheduler.get_max_eps()) x = BoundedTensor(data, ptb) output = self.model(x) regular_ce = CrossEntropyLoss()(output, labels) # regular CrossEntropyLoss used for warming up meter.update('CE', regular_ce.item(), x.size(0)) meter.update('Err', torch.sum(torch.argmax(output, dim=1) != labels).cpu().detach().numpy() / x.size(0), x.size(0)) if batch_method == "robust": if self.args.bound_type == "IBP": lb, ub = self.model.compute_bounds(IBP=True, C=c, method=None) elif self.args.bound_type == "CROWN": lb, ub = self.model.compute_bounds(IBP=False, C=c, method="backward", bound_upper=False) elif self.args.bound_type == "CROWN-IBP": # lb, ub = model.compute_bounds(ptb=ptb, IBP=True, x=data, C=c, method="backward") # pure IBP bound # we use a mixed IBP and CROWN-IBP bounds, leading to better performance (Zhang et al., ICLR 2020) factor = (eps_scheduler.get_max_eps() - eps) / eps_scheduler.get_max_eps() ilb, iub = self.model.compute_bounds(IBP=True, C=c, method=None) if factor < 1e-5: lb = ilb else: clb, cub = self.model.compute_bounds(IBP=False, C=c, method="backward", bound_upper=False) lb = clb * factor + ilb * (1 - factor) elif self.args.bound_type == "CROWN-FAST": # Similar to CROWN-IBP but no mix between IBP and CROWN bounds. lb, ub = self.model.compute_bounds(IBP=True, C=c, method=None) lb, ub = self.model.compute_bounds(IBP=False, C=c, method="backward", bound_upper=False) # Pad zero at the beginning for each example, and use fake label "0" for all examples lb_padded = torch.cat((torch.zeros(size=(lb.size(0), 1), dtype=lb.dtype, device=lb.device), lb), dim=1) fake_labels = torch.zeros(size=(lb.size(0),), dtype=torch.int64, device=lb.device) robust_ce = CrossEntropyLoss()(-lb_padded, fake_labels) if batch_method == "robust": loss = robust_ce elif batch_method == "natural": loss = regular_ce loss.backward() eps_scheduler.update_loss(loss.item() - regular_ce.item()) opt.step() meter.update('Loss', loss.item(), data.size(0)) if batch_method != "natural": meter.update('Robust_CE', robust_ce.item(), data.size(0)) # # For an example, if lower bounds of margins is >0 for all classes, the output is verifiably correct. # # If any margin is < 0 this example is counted as an error meter.update('Verified_Err', torch.sum((lb < 0).any(dim=1)).item() / data.size(0), data.size(0)) # # if i % 50 == 0 and train: # print('[{:4d}]: eps={:.8f} {}'.format(i, eps, meter)) # print('[{:4d}]: eps={:.8f} {}'.format(i, eps, meter))
12,734
53.191489
123
py
defend_framework
defend_framework-main/models/EmberLiRPAModel.py
import torch import torch.nn as nn import torch.nn.functional as F import numpy as np from models.LiRPAModel import LiRPAModel class EmberModel(LiRPAModel): def __init__(self, n_features, n_classes, args, device, lr=1e-3): super(EmberModel, self).__init__([1, 1, n_features], n_classes, args, device, mlp_4layer, lr) def fit(self, X, y, batch_size, epochs, dummy=None): X = np.expand_dims(X, axis=[1, 2]) def data_aug(data): data += torch.clamp(torch.normal(torch.zeros_like(data), torch.ones_like(data) * 0.1), min=0, max=1) return data super(EmberModel, self).fit(X, y, batch_size, epochs, data_aug) def evaluate(self, x_test, y_test): x_test = np.expand_dims(x_test, axis=[1, 2]) return super(EmberModel, self).evaluate(x_test, y_test) class mlp_4layer(nn.Module): def __init__(self, in_ch, in_dim): super(mlp_4layer, self).__init__() self.fc1 = nn.Linear(in_ch * in_dim, 2000) self.fc2 = nn.Linear(2000, 1000) self.fc3 = nn.Linear(1000, 100) self.fc4 = nn.Linear(100, 2) def forward(self, x): x = x.view(x.size(0), -1) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = F.relu(self.fc3(x)) x = self.fc4(x) return x
1,310
30.214286
112
py
defend_framework
defend_framework-main/models/MNISTModel.py
from tensorflow import keras from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, Flatten, Dropout from tensorflow.keras.layers import Conv2D, AveragePooling2D, MaxPooling2D from tensorflow.keras.regularizers import l2 from models.Model import Model class MNISTModel(Model): def __init__(self, n_features, n_classes, lr=1e-3): super(MNISTModel, self).__init__(n_features, n_classes, lr) def build_model(self): def scheduler(epoch, lr): if epoch > 0 and epoch % 100 == 0: lr *= 0.5 # print(lr) return lr reg = l2(1e-3) model = Sequential() model.add(Conv2D(16, kernel_size=(5, 5), activation='relu', input_shape=self.input_shape)) model.add(AveragePooling2D(pool_size=(2, 2))) model.add(Conv2D(32, (5, 5), activation='relu')) model.add(AveragePooling2D(pool_size=(2, 2))) model.add(Flatten()) model.add(Dropout(0.25)) model.add(Dense(32, activation='relu', kernel_regularizer=reg, bias_regularizer=reg)) model.add(Dropout(0.25)) model.add(Dense(512, activation='relu', kernel_regularizer=reg, bias_regularizer=reg)) model.add(Dense(self.n_classes, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=self.lr), metrics=['accuracy']) # self.callback = [keras.callbacks.LearningRateScheduler(scheduler)] return model class MNIST01Model(Model): def __init__(self, n_features, n_classes, lr=1e-3): super(MNIST01Model, self).__init__(n_features, n_classes, lr) def build_model(self): model = Sequential() model.add(Conv2D(16, kernel_size=(5, 5), activation='relu', input_shape=self.input_shape)) model.add(MaxPooling2D(pool_size=(2, 2), strides=2)) model.add(Conv2D(32, (5, 5), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2), strides=2)) model.add(Flatten()) model.add(Dense(128, activation='relu')) model.add(Dense(self.n_classes, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=self.lr), metrics=['accuracy']) return model class MNIST17Model(Model): def __init__(self, n_features, n_classes, lr=1e-3): super(MNIST17Model, self).__init__(n_features, n_classes, lr) def build_model(self): model = Sequential() model.add(Flatten(input_shape=self.input_shape)) model.add(Dense(256, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(128, activation='relu')) model.add(Dropout(0.5)) model.add(Dense(self.n_classes, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=self.lr), metrics=['accuracy']) return model class FMNISTModel(Model): def __init__(self, n_features, n_classes, lr=1e-3): super(FMNISTModel, self).__init__(n_features, n_classes, lr) def build_model(self): reg = l2(1e-3) model = Sequential() model.add(Conv2D(16, kernel_size=(3, 3), activation='relu', input_shape=self.input_shape)) model.add(Conv2D(32, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2), strides=2)) model.add(Conv2D(64, (3, 3), activation='relu')) model.add(MaxPooling2D(pool_size=(2, 2), strides=2)) model.add(Flatten()) model.add(Dropout(0.6)) model.add(Dense(128, activation='relu', kernel_regularizer=reg, bias_regularizer=reg)) model.add(Dropout(0.5)) model.add(Dense(self.n_classes, activation='softmax')) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=self.lr), metrics=['accuracy']) return model
4,251
38.37037
94
py
defend_framework
defend_framework-main/models/MNISTLiRPAModel.py
from torchvision import transforms import torch.nn as nn import torch.nn.functional as F import numpy as np from models.LiRPAModel import LiRPAModel class MNISTModel(LiRPAModel): def __init__(self, n_features, n_classes, args, device, lr=1e-3): super(MNISTModel, self).__init__(n_features, n_classes, args, device, mlp_3layer, lr) def fit(self, X, y, batch_size, epochs, dummy=None): X = np.transpose(X, (0, 3, 1, 2)) def data_aug(data): data = transforms.RandomCrop(28, 3)(data) data = transforms.RandomRotation(10)(data) return data super(MNISTModel, self).fit(X, y, batch_size, epochs, data_aug) def evaluate(self, x_test, y_test): x_test = np.transpose(x_test, (0, 3, 1, 2)) return super(MNISTModel, self).evaluate(x_test, y_test) class mlp_3layer(nn.Module): def __init__(self, in_ch, in_dim, width=1): super(mlp_3layer, self).__init__() self.fc1 = nn.Linear(in_ch * in_dim * in_dim, 256 * width) self.fc2 = nn.Linear(256 * width, 128 * width) self.fc3 = nn.Linear(128 * width, 10) def forward(self, x): x = x.view(x.size(0), -1) x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) x = self.fc3(x) return x
1,294
30.585366
93
py
defend_framework
defend_framework-main/models/Model.py
from abc import ABC, abstractmethod import os import numpy as np from tensorflow.keras.models import load_model class Model(ABC): def __init__(self, input_shape, n_classes, lr): self.input_shape = input_shape self.n_classes = n_classes self.lr = lr self.callback = None self.model = self.build_model() @abstractmethod def build_model(self): pass def init(self): self.model = self.build_model() def save(self, save_path, file_name='mnist_nn'): save_model = self.model save_model.save(os.path.join(save_path, file_name + '.h5')) def load(self, save_path, file_name): self.model = load_model(os.path.join(save_path, file_name + '.h5')) def fit_generator(self, datagen, epochs): self.model.fit(datagen, epochs=epochs, verbose=0, workers=4, callbacks=self.callback) def fit(self, X, y, batch_size, epochs): self.model.fit(X, y, batch_size=batch_size, epochs=epochs, verbose=0, workers=4, callbacks=self.callback) def evaluate(self, x_test, y_test): score = self.model.evaluate(x_test, y_test, verbose=0, batch_size=512) print('Test loss:', score[0]) print('Test accuracy:', score[1]) prediction_label = np.argmax(self.model.predict(x_test), axis=1) return prediction_label
1,351
30.44186
113
py
defend_framework
defend_framework-main/models/IMDBTransformerModel.py
import tensorflow as tf from tensorflow import keras from tensorflow.keras import layers from models.Model import Model class TransformerBlock(layers.Layer): def __init__(self, embed_dim, num_heads, ff_dim, rate=0.1): super(TransformerBlock, self).__init__() self.att = layers.MultiHeadAttention(num_heads=num_heads, key_dim=embed_dim) self.ffn = keras.Sequential( [layers.Dense(ff_dim, activation="relu"), layers.Dense(embed_dim), ] ) self.layernorm1 = layers.LayerNormalization(epsilon=1e-6) self.layernorm2 = layers.LayerNormalization(epsilon=1e-6) self.dropout1 = layers.Dropout(rate) self.dropout2 = layers.Dropout(rate) def call(self, inputs, training): attn_output = self.att(inputs, inputs) attn_output = self.dropout1(attn_output, training=training) out1 = self.layernorm1(inputs + attn_output) ffn_output = self.ffn(out1) ffn_output = self.dropout2(ffn_output, training=training) return self.layernorm2(out1 + ffn_output) class TokenAndPositionEmbedding(layers.Layer): def __init__(self, maxlen, embed_dim): super(TokenAndPositionEmbedding, self).__init__() self.token_emb = layers.Embedding(input_dim=10000, output_dim=embed_dim) self.pos_emb = layers.Embedding(input_dim=maxlen, output_dim=embed_dim) def call(self, x): maxlen = tf.shape(x)[-1] positions = tf.range(start=0, limit=maxlen, delta=1) positions = self.pos_emb(positions) x = self.token_emb(x) return x + positions class IMDBTransformerModel(Model): def __init__(self, maxlen, n_classes, lr=1e-3): super(IMDBTransformerModel, self).__init__(maxlen, n_classes, lr) def build_model(self): inputs = layers.Input(shape=(self.input_shape,)) embedding_layer = TokenAndPositionEmbedding(self.input_shape, 50) x = embedding_layer(inputs) transformer_block = TransformerBlock(50, 2, 100) x = transformer_block(x) x = layers.GlobalAveragePooling1D()(x) x = layers.Dropout(0.1)(x) x = layers.Dense(20, activation="relu")(x) x = layers.Dropout(0.1)(x) outputs = layers.Dense(2, activation="softmax")(x) model = keras.Model(inputs=inputs, outputs=outputs) model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=self.lr), metrics=['accuracy']) return model
2,535
38.015385
84
py
defend_framework
defend_framework-main/models/EmberModel.py
from tensorflow import keras from tensorflow.keras.models import Sequential from tensorflow.keras.layers import Dense, Input, Dropout, BatchNormalization, Activation from tensorflow.keras.layers import Conv2D, AveragePooling2D from tensorflow.keras.regularizers import l2 from models.Model import Model class EmberModel(Model): def __init__(self, n_features, n_classes, lr=1e-3): super(EmberModel, self).__init__(n_features, n_classes, lr) def build_model(self): def scheduler(epoch, lr): if epoch > 0 and epoch % 10 == 0: lr *= 0.9 # print(lr) return lr reg = l2(1e-5) model = Sequential() model.add(Input(shape=(self.input_shape,))) model.add(Dense(2000, activation='relu', kernel_regularizer=reg, bias_regularizer=reg)) # model.add(BatchNormalization()) model.add(Dropout(0.25)) model.add(Dense(1000, activation='relu', kernel_regularizer=reg, bias_regularizer=reg)) # model.add(BatchNormalization()) model.add(Dropout(0.25)) model.add(Dense(100, activation='relu', kernel_regularizer=reg, bias_regularizer=reg)) # model.add(BatchNormalization()) model.add(Dropout(0.25)) model.add(Dense(2, activation='softmax')) momentum = 0.9 decay = 0.000001 opt = keras.optimizers.SGD(lr=self.lr, momentum=momentum, decay=decay) self.callback = [] self.callback += [keras.callbacks.LearningRateScheduler(scheduler)] # self.callback += [keras.callbacks.EarlyStopping( # monitor="loss", # min_delta=0.01, # patience=20, # verbose=0, # baseline=None, # restore_best_weights=True, # )] model.compile(loss=keras.losses.categorical_crossentropy, optimizer=keras.optimizers.Adam(lr=self.lr), metrics=['accuracy'], ) return model
2,009
36.924528
95
py
defend_framework
defend_framework-main/models/CIFAR10Model.py
from tensorflow import keras from tensorflow.keras.layers import BatchNormalization, Activation, Conv2D, Input, AveragePooling2D, Flatten, Dense from tensorflow.keras.regularizers import l2 from tensorflow.keras.optimizers import Adam import numpy as np from tensorflow.keras.callbacks import LearningRateScheduler, ReduceLROnPlateau from models.Model import Model def lr_schedule(epoch): """Learning Rate Schedule Learning rate is scheduled to be reduced after 80, 120, 160, 180 epochs. Called automatically every epoch as part of callbacks during training. # Arguments epoch (int): The number of epochs # Returns lr (float32): learning rate """ lr = 1e-3 if epoch > 180: lr *= 0.5e-3 elif epoch > 160: lr *= 1e-3 elif epoch > 120: lr *= 1e-2 elif epoch > 80: lr *= 1e-1 # print('Learning rate: ', lr) return lr def resnet_layer(inputs, num_filters=16, kernel_size=3, strides=1, activation='relu', batch_normalization=True, conv_first=True): """2D Convolution-Batch Normalization-Activation stack builder # Arguments inputs (tensor): input tensor from input image or previous layer num_filters (int): Conv2D number of filters kernel_size (int): Conv2D square kernel dimensions strides (int): Conv2D square stride dimensions activation (string): activation name batch_normalization (bool): whether to include batch normalization conv_first (bool): conv-bn-activation (True) or bn-activation-conv (False) # Returns x (tensor): tensor as input to the next layer """ conv = Conv2D(num_filters, kernel_size=kernel_size, strides=strides, padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(1e-4)) x = inputs if conv_first: x = conv(x) if batch_normalization: x = BatchNormalization()(x) if activation is not None: x = Activation(activation)(x) else: if batch_normalization: x = BatchNormalization()(x) if activation is not None: x = Activation(activation)(x) x = conv(x) return x def resnet_v1(input_shape, depth, num_classes=10): """ResNet Version 1 Model builder [a] Stacks of 2 x (3 x 3) Conv2D-BN-ReLU Last ReLU is after the shortcut connection. At the beginning of each stage, the feature map size is halved (downsampled) by a convolutional layer with strides=2, while the number of filters is doubled. Within each stage, the layers have the same number filters and the same number of filters. Features maps sizes: stage 0: 32x32, 16 stage 1: 16x16, 32 stage 2: 8x8, 64 The Number of parameters is approx the same as Table 6 of [a]: ResNet20 0.27M ResNet32 0.46M ResNet44 0.66M ResNet56 0.85M ResNet110 1.7M # Arguments input_shape (tensor): shape of input image tensor depth (int): number of core convolutional layers num_classes (int): number of classes (CIFAR10 has 10) # Returns model (Model): Keras model instance """ if (depth - 2) % 6 != 0: raise ValueError('depth should be 6n+2 (eg 20, 32, 44 in [a])') # Start model definition. num_filters = 16 num_res_blocks = int((depth - 2) / 6) inputs = Input(shape=input_shape) x = resnet_layer(inputs=inputs) # Instantiate the stack of residual units for stack in range(3): for res_block in range(num_res_blocks): strides = 1 if stack > 0 and res_block == 0: # first layer but not first stack strides = 2 # downsample y = resnet_layer(inputs=x, num_filters=num_filters, strides=strides) y = resnet_layer(inputs=y, num_filters=num_filters, activation=None) if stack > 0 and res_block == 0: # first layer but not first stack # linear projection residual shortcut connection to match # changed dims x = resnet_layer(inputs=x, num_filters=num_filters, kernel_size=1, strides=strides, activation=None, batch_normalization=False) x = keras.layers.add([x, y]) x = Activation('relu')(x) num_filters *= 2 # Add classifier on top. # v1 does not use BN after last shortcut connection-ReLU x = AveragePooling2D(pool_size=8)(x) y = Flatten()(x) outputs = Dense(num_classes, activation='softmax', kernel_initializer='he_normal')(y) # Instantiate model. model = keras.Model(inputs=inputs, outputs=outputs) return model class CIFAR10Model(Model): def __init__(self, n_features, n_classes, lr=1e-3): super(CIFAR10Model, self).__init__(n_features, n_classes, lr) def build_model(self): model = resnet_v1(self.input_shape, 20, num_classes=self.n_classes) model.compile(loss='categorical_crossentropy', optimizer=Adam(lr=lr_schedule(0)), metrics=['accuracy']) # model.summary() lr_scheduler = LearningRateScheduler(lr_schedule) lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1), monitor='loss', cooldown=0, patience=5, min_lr=0.5e-6) self.callback = [lr_reducer, lr_scheduler] return model
6,009
33.342857
115
py
defend_framework
defend_framework-main/utils/dataaug.py
from tensorflow.keras.preprocessing.image import ImageDataGenerator from tensorflow import keras import numpy as np import torch from torchvision import transforms def DataGeneratorForMNIST(): datagen = ImageDataGenerator( # set input mean to 0 over the dataset featurewise_center=False, # set each sample mean to 0 samplewise_center=False, # divide inputs by std of dataset featurewise_std_normalization=False, # divide each input by its std samplewise_std_normalization=False, # apply ZCA whitening zca_whitening=False, # epsilon for ZCA whitening zca_epsilon=1e-06, # randomly rotate images in the range (deg 0 to 180) rotation_range=10, # randomly shift images horizontally width_shift_range=0.1, # randomly shift images vertically height_shift_range=0.1, # set range for random shear shear_range=0.0, # set range for random zoom zoom_range=0.0, # set range for random channel shifts channel_shift_range=0., # set mode for filling points outside the input boundaries fill_mode='nearest', # value used for fill_mode = "constant" cval=0., # randomly flip images horizontal_flip=False, # randomly flip images vertical_flip=False, # set rescaling factor (applied before any other transformation) rescale=None, # set function that will be applied on each input preprocessing_function=None, # image data format, either "channels_first" or "channels_last" data_format="channels_last", # fraction of images reserved for validation (strictly between 0 and 1) validation_split=0.0) return datagen class DataGenerator(keras.utils.Sequence): 'Generates data for Keras' def __init__(self, X, y, batch_size, shuffle=True): 'Initialization' self.batch_size = batch_size self.X = X self.y = y self.shuffle = shuffle self.on_epoch_end() def __len__(self): 'Denotes the number of batches per epoch' return int(np.ceil(len(self.X) / self.batch_size)) def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch indexes = self.indexes[index * self.batch_size:(index + 1) * self.batch_size] return self.X[indexes], self.y[indexes] def on_epoch_end(self): 'Updates indexes after each epoch' self.indexes = np.arange(len(self.X)) if self.shuffle: np.random.shuffle(self.indexes) class MNISTDataGenerator(DataGenerator): def __init__(self, X, y, batch_size, data_processor, no_eval_noise, shuffle=True): super(MNISTDataGenerator, self).__init__(X, y, batch_size, shuffle=shuffle) self.data_processor = data_processor self.no_eval_noise = no_eval_noise def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch X, y = super(MNISTDataGenerator, self).__getitem__(index) data = torch.Tensor(np.transpose(X, (0, 3, 1, 2))) data = transforms.RandomCrop(28, 2)(data) data = transforms.RandomRotation(10)(data) data = np.transpose(data.numpy(), (0, 2, 3, 1)) if self.no_eval_noise: return np.vstack([data, X]), np.vstack([y, y]) # return np.vstack([X + np.random.normal(0, 0.1, X.shape), X]), np.vstack([y, y]) else: return np.vstack([self.data_processor.noise_data(X), data, X]), np.vstack([y, y, y]) class CIFARDataGenerator(DataGenerator): def __init__(self, X, y, batch_size, data_processor, no_eval_noise, shuffle=True): super(CIFARDataGenerator, self).__init__(X, y, batch_size, shuffle=shuffle) self.data_processor = data_processor self.no_eval_noise = no_eval_noise def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch X, y = super(CIFARDataGenerator, self).__getitem__(index) data = torch.Tensor(np.transpose(X, (0, 3, 1, 2))) data = transforms.RandomCrop(32, 3)(data) data = transforms.RandomRotation(10)(data) data = np.transpose(data.numpy(), (0, 2, 3, 1)) if self.no_eval_noise: return np.vstack([data, X]), np.vstack([y, y]) # return np.vstack([X + np.random.normal(0, 0.1, X.shape), X]), np.vstack([y, y]) else: return np.vstack([self.data_processor.noise_data(X), data, X]), np.vstack([y, y, y]) class EmberDataGenerator(DataGenerator): def __init__(self, X, y, batch_size, data_processor, no_eval_noise, shuffle=True): super(EmberDataGenerator, self).__init__(X, y, batch_size, shuffle=shuffle) self.data_processor = data_processor self.no_eval_noise = no_eval_noise def __getitem__(self, index): 'Generate one batch of data' # Generate indexes of the batch X, y = super(EmberDataGenerator, self).__getitem__(index) if self.no_eval_noise: return np.vstack([X + np.random.normal(0, 0.1, X.shape), X]), np.vstack([y, y]) else: return np.vstack([self.data_processor.noise_data(X), X + np.random.normal(0, 0.1, X.shape), X]), np.vstack( [y, y, y])
5,425
38.318841
119
py
defend_framework
defend_framework-main/utils/data_processing.py
import os import pickle import numpy as np from tensorflow.keras.datasets import mnist, imdb, cifar10, fashion_mnist from tensorflow import keras import ember import pandas as pd from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler, KBinsDiscretizer, MinMaxScaler from utils.ember_feature_utils import load_features from utils import EMBER_DATASET, FEATURE_DATASET, LANGUAGE_DATASET, CONTAGIO_DATASET class DataProcessor: def __init__(self, X, y, select_strategy=None, k=None, noise_strategy=None, dataset=None, **kwargs): """ The initializer of data processor :param X: the training data (features) :param y: the training labels :param select_strategy: ["bagging_replace", "bagging_wo_replace", "binomial"] bagging_replace: bagging with replacement (the original bagging paper) bagging_wo_replace: bagging without replacement binomial: select each instance with probability p = k / |X| :param k: the size of the (expected) bag :param noise_strategy: ["feature_flipping", "label_flipping", "RAB_gaussian", "RAB_uniform"] feature_flipping / label_flipping: each feature / label remains with alpha, flipped with 1 - alpha RAB_gaussian: add gaussian noise of mu=0, sigma RAB_uniform: add uniform noise of U[a, b] :param dataset: the name of the dataset :param kwargs: the parameters for each noise strategy """ self.select_strategy = select_strategy self.noise_strategy = noise_strategy self.dataset = dataset self.X = X self.y = y self.DPA_partition_cnt = 0 if select_strategy is not None: assert select_strategy in ["bagging_replace", "bagging_wo_replace", "binomial", "DPA"] assert 0 < k <= len(X) self.k = k if select_strategy == "DPA": self.ids = np.arange(self.X.shape[0]) np.random.shuffle(self.ids) assert self.k * kwargs["N"] <= self.X.shape[0] and noise_strategy is None if dataset in EMBER_DATASET: self.minmax = MinMaxScaler() self.minmax.fit(self.X) if noise_strategy is not None: assert noise_strategy in ["feature_flipping", "label_flipping", "all_flipping", "RAB_gaussian", "RAB_uniform", "sentence_select"] if dataset in FEATURE_DATASET: if noise_strategy in ["feature_flipping", "label_flipping", "all_flipping"]: self.K = kwargs["K"] self.alpha = kwargs["alpha"] self.test_alpha = kwargs["test_alpha"] if self.test_alpha is None: self.test_alpha = self.alpha if noise_strategy in ["feature_flipping", "all_flipping"]: if dataset in EMBER_DATASET or dataset in CONTAGIO_DATASET: self.kbin = KBinsDiscretizer(n_bins=self.K + 1, strategy='uniform', encode='ordinal') self.kbin.fit(self.X) if dataset == "ember_limited": self.limit_id, _, _, _ = load_features(False) self.limit_id = self.limit_id['feasible'] self.limit_mask = np.zeros_like(self.X[0]).astype(np.bool) self.limit_mask[self.limit_id] = True else: assert (self.X >= 0).all() and (self.X <= 1).all() if noise_strategy in ["label_flipping", "all_flipping"]: assert (self.y >= 0).all() and (self.y <= self.K).all() elif noise_strategy == "RAB_gaussian": self.sigma = kwargs["sigma"] self.minmax = MinMaxScaler() self.minmax.fit(self.X) elif noise_strategy == "RAB_uniform": self.a = kwargs["a"] self.b = kwargs["b"] self.minmax = MinMaxScaler() self.minmax.fit(self.X) else: raise NotImplementedError elif dataset in LANGUAGE_DATASET: assert noise_strategy in ["sentence_select", "label_flipping", "all_flipping"] if noise_strategy in ["sentence_select", "all_flipping"]: self.l = kwargs["l"] if noise_strategy in ["label_flipping", "all_flipping"]: self.K = kwargs["K"] self.alpha = kwargs["alpha"] assert (self.y >= 0).all() and (self.y <= self.K).all() else: raise NotImplementedError def noise_data(self, ret_X, alpha=None, shape=None): if alpha is None: alpha = self.alpha if shape is None: shape = ret_X.shape mask = np.random.random(shape) < alpha delta = np.random.randint(1, self.K + 1, shape) / self.K ret_X = ret_X + (1 - mask) * delta ret_X[ret_X > 1 + 1e-4] -= (1 + self.K) / self.K return ret_X def process_train(self, key_dict): ret_X = self.X.copy() ret_y = self.y.copy() # make sure the original data is not modified if self.select_strategy is not None: if self.select_strategy in ["bagging_replace", "bagging_wo_replace"]: indices = np.random.choice(np.arange(len(self.X)), self.k, replace=self.select_strategy == "bagging_replace") ret_X = ret_X[indices] ret_y = ret_y[indices] elif self.select_strategy == "binomial": pred = np.random.random(len(self.X)) * len(self.X) < self.k ret_X = ret_X[pred] ret_y = ret_y[pred] elif self.select_strategy == "DPA": ids = self.ids[self.DPA_partition_cnt * self.k:(self.DPA_partition_cnt + 1) * self.k] ret_X = ret_X[ids] ret_y = ret_y[ids] if self.dataset in EMBER_DATASET: ret_X = self.minmax.transform(ret_X) if self.noise_strategy is not None: if self.dataset in FEATURE_DATASET: if self.noise_strategy in ["feature_flipping", "all_flipping"]: if self.dataset in EMBER_DATASET or self.dataset in CONTAGIO_DATASET: categorized = self.kbin.transform(ret_X) / self.K if self.dataset == "ember_limited": ret_X[:, self.limit_id] = categorized[:, self.limit_id] else: ret_X = categorized pre_ret_X = ret_X if self.dataset == "cifar10-02": ret_X = self.noise_data(ret_X, shape=list(ret_X.shape[:-1]) + [1]) else: ret_X = self.noise_data(ret_X) if self.dataset == "ember_limited": # protect other features ret_X = ret_X * self.limit_mask + pre_ret_X * (1 - self.limit_mask) if self.noise_strategy in ["label_flipping", "all_flipping"]: mask = np.random.random(ret_y.shape) < self.alpha delta = np.random.randint(1, self.K + 1, ret_y.shape) ret_y = ret_y * mask + (1 - mask) * (ret_y + delta) ret_y[ret_y > self.K] -= self.K + 1 if self.noise_strategy == "RAB_gaussian": ret_X = self.minmax.transform(ret_X) ret_X += np.random.normal(0, self.sigma, ret_X.shape) if self.noise_strategy == "RAB_uniform": ret_X = self.minmax.transform(ret_X) ret_X += np.random.uniform(self.a, self.b, ret_X.shape) elif self.dataset in LANGUAGE_DATASET: if self.noise_strategy in ["sentence_select", "all_flipping"]: maxlen = ret_X.shape[1] ret_X_new = [] for x in ret_X: indices = sorted(np.random.choice(np.arange(maxlen), self.l, replace=False)) ret_X_new.append( np.pad(x[indices], (0, maxlen - self.l), 'constant', constant_values=(0, 0))) ret_X = np.array(ret_X_new) if self.noise_strategy in ["label_flipping", "all_flipping"]: mask = np.random.random(ret_y.shape) < self.alpha delta = np.random.randint(1, self.K + 1, ret_y.shape) ret_y = ret_y * mask + (1 - mask) * (ret_y + delta) ret_y[ret_y > self.K] -= self.K + 1 if (self.noise_strategy is not None or self.select_strategy is not None) and self.dataset == "imdb": for x in ret_X: for i in range(len(x)): if x[i] not in key_dict: key_dict[x[i]] = len(key_dict) x[i] = key_dict[x[i]] if self.dataset in EMBER_DATASET and self.noise_strategy is None: self.normal = StandardScaler() ret_X = self.normal.fit_transform(ret_X) return ret_X, ret_y def process_test(self, X, fix_noise): ret_X = X.copy() if fix_noise: if self.noise_strategy is not None: if self.dataset in FEATURE_DATASET: if self.noise_strategy in ["feature_flipping", "all_flipping"]: pre_ret_X = ret_X if self.dataset == "cifar10-02": ret_X = self.noise_data(ret_X, alpha=self.test_alpha, shape=list(ret_X.shape[1:-1]) + [1]) else: ret_X = self.noise_data(ret_X, alpha=self.test_alpha, shape=ret_X.shape[1:]) if self.dataset == "ember_limited": # protect other features ret_X = ret_X * self.limit_mask + pre_ret_X * (1 - self.limit_mask) if self.noise_strategy == "RAB_gaussian": ret_X += np.random.normal(0, self.sigma, ret_X.shape[1:]) # fix the noise for each example if self.noise_strategy == "RAB_uniform": ret_X += np.random.uniform(self.a, self.b, ret_X.shape[1:]) # fix the noise for each example elif self.dataset in LANGUAGE_DATASET: if self.noise_strategy in ["sentence_select", "all_flipping"]: maxlen = ret_X.shape[1] ret_X_new = np.zeros_like(ret_X) indices = sorted( np.random.choice(np.arange(maxlen), self.l, replace=False)) # fix the noise for each example ret_X_new[:, :self.l] = ret_X[:, indices] ret_X = ret_X_new else: if self.noise_strategy is not None: if self.dataset in FEATURE_DATASET: if self.noise_strategy in ["feature_flipping", "all_flipping"]: pre_ret_X = ret_X if self.dataset == "cifar10-02": ret_X = self.noise_data(ret_X, alpha=self.test_alpha, shape=list(ret_X.shape[:-1]) + [1]) else: ret_X = self.noise_data(ret_X, alpha=self.test_alpha) if self.dataset == "ember_limited": # protect other features ret_X = ret_X * self.limit_mask + pre_ret_X * (1 - self.limit_mask) if self.noise_strategy == "RAB_gaussian": ret_X += np.random.normal(0, self.sigma, ret_X.shape) if self.noise_strategy == "RAB_uniform": ret_X += np.random.uniform(self.a, self.b, ret_X.shape) elif self.dataset in LANGUAGE_DATASET: if self.noise_strategy in ["sentence_select", "all_flipping"]: maxlen = ret_X.shape[1] ret_X_new = [] for x in ret_X: indices = sorted(np.random.choice(np.arange(maxlen), self.l, replace=False)) ret_X_new.append( np.pad(x[indices], (0, maxlen - self.l), 'constant', constant_values=(0, 0))) ret_X = np.array(ret_X_new) if self.dataset in EMBER_DATASET and self.noise_strategy is None: ret_X = self.normal.transform(ret_X) return ret_X class DataPreprocessor: def __init__(self): pass @classmethod def load(cls, filename, args): with open(filename, "rb") as f: attack = pickle.load(f) this = attack.data_processor this.attack = attack this.data_processor = DataPreprocessor.build_processor(this.x_train, this.y_train, args) return this @staticmethod def build_processor(x_train, y_train, args): return DataProcessor(x_train, y_train, select_strategy=args.select_strategy, k=args.k, noise_strategy=args.noise_strategy, K=args.K, alpha=args.alpha, sigma=args.sigma, a=args.a, b=args.b, dataset=args.dataset, l=args.l, test_alpha=args.test_alpha, N=args.N) class MNIST17DataPreprocessor(DataPreprocessor): def __init__(self, args): super(MNIST17DataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 28, 28 self.n_classes = 2 self.n_features = (img_rows, img_cols, 1) (x_train, self.y_train), (x_test, self.y_test) = mnist.load_data() x_train = x_train[(self.y_train == 1) | (self.y_train == 7)] self.y_train = self.y_train[(self.y_train == 1) | (self.y_train == 7)] self.y_train = self.y_train > 1 x_test = x_test[(self.y_test == 1) | (self.y_test == 7)] self.y_test = self.y_test[(self.y_test == 1) | (self.y_test == 7)] self.y_test = self.y_test > 1 x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.K != 1 and args.noise_strategy in ["label_flipping", "all_flipping"]: raise NotImplementedError("K != 1 not implemented for MNIST17DataPreprocessor.") if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = self.x_train >= 0.5 self.x_test = self.x_test >= 0.5 self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', x_train.shape, self.y_train.shape) print(x_train.shape[0], 'train samples') print(x_test.shape[0], 'test samples') class MNIST17LimitedDataPreprocessor(DataPreprocessor): def __init__(self, args): super(MNIST17LimitedDataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 28, 28 self.n_classes = 2 self.n_features = (img_rows, img_cols, 1) (x_train, self.y_train), (x_test, self.y_test) = mnist.load_data() x_train = x_train[(self.y_train == 1) | (self.y_train == 7)] self.y_train = self.y_train[(self.y_train == 1) | (self.y_train == 7)] self.y_train = self.y_train > 1 x_test = x_test[(self.y_test == 1) | (self.y_test == 7)] self.y_test = self.y_test[(self.y_test == 1) | (self.y_test == 7)] self.y_test = self.y_test > 1 x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.K != 1 and args.noise_strategy in ["label_flipping", "all_flipping"]: raise NotImplementedError("K != 1 not implemented for MNIST17DataPreprocessor.") if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = self.x_train >= 0.5 self.x_test = self.x_test >= 0.5 train_ids = np.random.choice(np.arange(x_train.shape[0]), 100, replace=False) # test_ids = np.random.choice(np.arange(x_test.shape[0]), 1000, replace=False) self.x_train = self.x_train[train_ids] self.y_train = self.y_train[train_ids] # self.x_test = self.x_test[test_ids] # self.y_test = self.y_test[test_ids] self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', self.x_train.shape, self.y_train.shape) print(self.x_train.shape[0], 'train samples') print(self.x_test.shape[0], 'test samples') class MNIST01DataPreprocessor(DataPreprocessor): def __init__(self, args): super(MNIST01DataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 28, 28 self.n_classes = 2 self.n_features = (img_rows, img_cols, 1) (x_train, self.y_train), (x_test, self.y_test) = mnist.load_data() x_train = x_train[(self.y_train == 0) | (self.y_train == 1)] self.y_train = self.y_train[(self.y_train == 0) | (self.y_train == 1)] self.y_train = self.y_train > 0 x_test = x_test[(self.y_test == 0) | (self.y_test == 1)] self.y_test = self.y_test[(self.y_test == 0) | (self.y_test == 1)] self.y_test = self.y_test > 0 x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.K != 1 and args.noise_strategy in ["label_flipping", "all_flipping"]: raise NotImplementedError("K != 1 not implemented for MNIST17DataPreprocessor.") if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = self.x_train >= 0.5 self.x_test = self.x_test >= 0.5 self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', x_train.shape, self.y_train.shape) print(x_train.shape[0], 'train samples') print(x_test.shape[0], 'test samples') class MNISTDataPreprocessor(DataPreprocessor): def __init__(self, args): super(MNISTDataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 28, 28 self.n_classes = 10 self.n_features = (img_rows, img_cols, 1) (x_train, self.y_train), (x_test, self.y_test) = mnist.load_data() x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.noise_strategy in ["label_flipping", "all_flipping"]: assert args.K == 9 if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = np.minimum(np.floor(self.x_train * (args.K + 1)) / args.K, 1) self.x_test = np.minimum(np.floor(self.x_test * (args.K + 1)) / args.K, 1) self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', x_train.shape, self.y_train.shape) print(x_train.shape[0], 'train samples') print(x_test.shape[0], 'test samples') class FMNISTDataPreprocessor(DataPreprocessor): def __init__(self, args): super(FMNISTDataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 28, 28 self.n_classes = 10 self.n_features = (img_rows, img_cols, 1) (x_train, self.y_train), (x_test, self.y_test) = fashion_mnist.load_data() x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1) x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1) x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.noise_strategy in ["label_flipping", "all_flipping"]: assert args.K == 9 if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = np.minimum(np.floor(self.x_train * (args.K + 1)) / args.K, 1) self.x_test = np.minimum(np.floor(self.x_test * (args.K + 1)) / args.K, 1) self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', x_train.shape, self.y_train.shape) print(x_train.shape[0], 'train samples') print(x_test.shape[0], 'test samples') class CIFARDataPreprocessor(DataPreprocessor): def __init__(self, args): super(CIFARDataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 32, 32 self.n_classes = 10 self.n_features = (img_rows, img_cols, 3) (x_train, self.y_train), (x_test, self.y_test) = cifar10.load_data() self.y_test = np.reshape(self.y_test, -1) self.y_train = np.reshape(self.y_train, -1) x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.noise_strategy in ["label_flipping", "all_flipping"]: assert args.K == 9 if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = np.minimum(np.floor(self.x_train * (args.K + 1)) / args.K, 1) self.x_test = np.minimum(np.floor(self.x_test * (args.K + 1)) / args.K, 1) self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', x_train.shape, self.y_train.shape) print(x_train.shape[0], 'train samples') print(x_test.shape[0], 'test samples') class CIFAR02DataPreprocessor(DataPreprocessor): def __init__(self, args): super(CIFAR02DataPreprocessor, self).__init__() # input image dimensions img_rows, img_cols = 32, 32 self.n_classes = 2 self.n_features = (img_rows, img_cols, 3) (x_train, self.y_train), (x_test, self.y_test) = cifar10.load_data() self.y_test = np.reshape(self.y_test, -1) self.y_train = np.reshape(self.y_train, -1) x_train = x_train[(self.y_train == 0) | (self.y_train == 2)] self.y_train = self.y_train[(self.y_train == 0) | (self.y_train == 2)] self.y_train = self.y_train > 0 x_test = x_test[(self.y_test == 0) | (self.y_test == 2)] self.y_test = self.y_test[(self.y_test == 0) | (self.y_test == 2)] self.y_test = self.y_test > 0 x_train = x_train.astype('float32') x_test = x_test.astype('float32') self.x_train = x_train / 255 self.x_test = x_test / 255 if args.noise_strategy in ["label_flipping", "all_flipping"]: assert args.K == 9 if args.noise_strategy in ["feature_flipping", "all_flipping"]: self.x_train = np.minimum(np.floor(self.x_train * (args.K + 1)) / args.K, 1) self.x_test = np.minimum(np.floor(self.x_test * (args.K + 1)) / args.K, 1) self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', x_train.shape, self.y_train.shape) print(x_train.shape[0], 'train samples') print(x_test.shape[0], 'test samples') class IMDBDataPreprocessor(DataPreprocessor): def __init__(self, args): super(IMDBDataPreprocessor, self).__init__() vocab_size = 10000 # Only consider the top 20k words self.n_features = args.L # Only consider the first 200 words of each movie review self.n_classes = 2 (x_train, self.y_train), (x_test, self.y_test) = imdb.load_data(num_words=vocab_size) self.x_train = keras.preprocessing.sequence.pad_sequences(x_train, maxlen=self.n_features) self.x_test = keras.preprocessing.sequence.pad_sequences(x_test, maxlen=self.n_features) self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', self.x_train.shape, self.y_train.shape) print(self.x_train.shape[0], 'train samples') print(self.x_test.shape[0], 'test samples') class EmberDataPreProcessor(DataPreprocessor): def __init__(self, args): super(EmberDataPreProcessor, self).__init__() try: x_train, y_train, x_test, y_test = ember.read_vectorized_features( args.ember_data_dir, feature_version=1 ) except: ember.create_vectorized_features( args.ember_data_dir, feature_version=1 ) x_train, y_train, x_test, y_test = ember.read_vectorized_features( args.ember_data_dir, feature_version=1 ) x_train = x_train.astype(dtype='float64') x_test = x_test.astype(dtype='float64') if args.K != 1 and args.noise_strategy in ["all_flipping"]: raise NotImplementedError("K != 1 not implemented for EmberDataPreProcessor with all_flipping.") # Get rid of unknown labels self.x_train = x_train[y_train != -1] self.y_train = y_train[y_train != -1] self.x_test = x_test[y_test != -1] self.y_test = y_test[y_test != -1] self.n_features = x_train.shape[1] self.n_classes = 2 self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', self.x_train.shape, self.y_train.shape) print(self.x_train.shape[0], 'train samples') print(self.x_test.shape[0], 'test samples') class EmberPoisonDataPreProcessor(DataPreprocessor): def __init__(self, args): super(EmberPoisonDataPreProcessor, self).__init__() self.x_train = np.load(os.path.join(args.load_poison_dir, "watermarked_X.npy")) self.y_train = np.load(os.path.join(args.load_poison_dir, "watermarked_y.npy")) self.x_test = np.load(os.path.join(args.load_poison_dir, "watermarked_X_test.npy")) self.y_test = np.ones(self.x_test.shape[0]) if args.K != 1 and args.noise_strategy in ["all_flipping", "label_flipping"]: raise NotImplementedError("K != 1 not implemented for EmberDataPreProcessor with all_flipping.") self.n_features = self.x_train.shape[1] self.n_classes = 2 self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', self.x_train.shape, self.y_train.shape) print(self.x_train.shape[0], 'train samples') print(self.x_test.shape[0], 'test samples') class ContagioDataPreProcessor(DataPreprocessor): def __init__(self, args): super(ContagioDataPreProcessor, self).__init__() mw_file = 'ogcontagio_mw.npy' gw_file = 'ogcontagio_gw.npy' # Load malicious mw = np.load( # os.path.join(constants.SAVE_FILES_DIR, mw_file), os.path.join(args.contagio_data_dir, mw_file), allow_pickle=True ).item() mwdf = pd.DataFrame(mw) mwdf = mwdf.transpose() mwdf['class'] = [True] * mwdf.shape[0] mwdf.index.name = 'filename' mwdf = mwdf.reset_index() train_mw, test_mw = train_test_split(mwdf, test_size=0.4, random_state=42) # Load benign gw = np.load( # os.path.join(constants.SAVE_FILES_DIR, gw_file), os.path.join(args.contagio_data_dir, gw_file), allow_pickle=True ).item() gwdf = pd.DataFrame(gw) gwdf = gwdf.transpose() gwdf['class'] = [False] * gwdf.shape[0] gwdf.index.name = 'filename' gwdf = gwdf.reset_index() train_gw, test_gw = train_test_split(gwdf, test_size=0.4, random_state=42) # Merge dataframes train_df = pd.concat([train_mw, train_gw]) test_df = pd.concat([test_mw, test_gw]) # Transform to numpy self.y_train = train_df['class'].to_numpy() self.y_test = test_df['class'].to_numpy() # x_train_filename = train_df['filename'].to_numpy() # x_test_filename = test_df['filename'].to_numpy() self.x_train = train_df.drop(columns=['class', 'filename']).to_numpy() self.x_test = test_df.drop(columns=['class', 'filename']).to_numpy() self.x_train = self.x_train.astype(dtype='float64') self.x_test = self.x_test.astype(dtype='float64') self.n_features = self.x_train.shape[1] self.n_classes = 2 self.data_processor = self.build_processor(self.x_train, self.y_train, args) print('x_train shape:', self.x_train.shape, self.y_train.shape) print(self.x_train.shape[0], 'train samples') print(self.x_test.shape[0], 'test samples')
29,834
46.132701
118
py
defend_framework
defend_framework-main/utils/train_utils.py
import numpy as np from tensorflow import keras from tqdm import trange import os from utils.dataaug import DataGeneratorForMNIST, MNISTDataGenerator, EmberDataGenerator, CIFARDataGenerator from utils import EMBER_DATASET, IMAGE_DATASET def train_many(data_loader, model, args, aggregate_result, aggregate_noise_result): test_size = data_loader.x_test.shape[0] if aggregate_result is None: aggregate_result = np.zeros([test_size, data_loader.n_classes + 1 + int(args.select_strategy == "DPA")], dtype=np.int) aggregate_noise_result = np.zeros([test_size, data_loader.n_classes + 1 + int(args.select_strategy == "DPA")], dtype=np.int) aggregate_result[np.arange(0, test_size), -1] = data_loader.y_test aggregate_noise_result[np.arange(0, test_size), -1] = data_loader.y_test remaining = args.N - np.sum(aggregate_result[0, :-1]) datagen = None for i in trange(remaining): key_dict = {0: 0, 1: 1, 2: 2} # used for imdb dataset to get word idx X, y = data_loader.data_processor.process_train(key_dict) # using the last index for the ground truth label y = keras.utils.to_categorical(y, data_loader.n_classes) if args.data_aug: if args.dataset in IMAGE_DATASET: if args.dataset == "cifar10": datagen = CIFARDataGenerator(X, y, args.batch_size, data_loader.data_processor, args.no_eval_noise) else: datagen = MNISTDataGenerator(X, y, args.batch_size, data_loader.data_processor, args.no_eval_noise) elif args.dataset in EMBER_DATASET: datagen = EmberDataGenerator(X, y, args.batch_size, data_loader.data_processor, args.no_eval_noise) y_test = keras.utils.to_categorical(data_loader.y_test, data_loader.n_classes) x_test = data_loader.x_test.copy() if args.dataset == "imdb": for x in x_test: for i in range(len(x)): if x[i] in key_dict: x[i] = key_dict[x[i]] else: x[i] = 2 elif args.dataset in EMBER_DATASET and args.noise_strategy in ["feature_flipping", "all_flipping"]: categorized = data_loader.data_processor.kbin.transform(x_test) / args.K if args.dataset == "ember_limited": x_test[:, data_loader.data_processor.limit_id] = categorized[:, data_loader.data_processor.limit_id] else: x_test = categorized elif args.noise_strategy in ["RAB_gaussian", "RAB_uniform"] or ( args.dataset in EMBER_DATASET and args.select_strategy == "DPA"): x_test = data_loader.data_processor.minmax.transform(x_test) if datagen is not None: model.fit_generator(datagen, args.epochs) else: model.fit(X, y, args.batch_size, args.epochs) if args.select_strategy != "DPA": if args.dataset in EMBER_DATASET and args.noise_strategy is None: prediction_label = model.evaluate(data_loader.data_processor.normal.transform(x_test), y_test) else: prediction_label = model.evaluate(x_test, y_test) aggregate_result[np.arange(0, test_size), prediction_label] += 1 if args.noise_strategy is None or args.no_eval_noise: aggregate_noise_result[np.arange(0, test_size), prediction_label] += 1 else: X_test = data_loader.data_processor.process_test(x_test, args.fix_noise) prediction_label = model.evaluate(X_test, y_test) aggregate_noise_result[np.arange(0, test_size), prediction_label] += 1 else: prediction_label, prediction_label_cert = model.evaluate(x_test, y_test) aggregate_result[np.arange(0, test_size), prediction_label] += 1 aggregate_noise_result[np.arange(0, test_size), prediction_label_cert] += 1 model.init() np.save(os.path.join(args.res_save_dir, args.exp_name, "aggre_res"), (aggregate_result, aggregate_noise_result)) print(aggregate_result, aggregate_noise_result) def train_single(data_loader, model, args): # train single classifier for attacking model.fit(data_loader.x_train, keras.utils.to_categorical(data_loader.y_train, data_loader.n_classes), args.batch_size, args.epochs) # model.save(args.model_save_dir)
4,535
50.545455
120
py
secml-zoo
secml-zoo-master/models/mnist/mnist_cnn.py
""" .. module:: MNISTCNN :synopsis: A CNN model for MNIST .. moduleauthor:: Ambra Demontis <ambra.demontis@diee.unica.it> """ from collections import OrderedDict import torch from torch import nn, optim from secml.ml.classifiers import CClassifierPyTorch class Flatten(nn.Module): def __init__(self): super(Flatten, self).__init__() def forward(self, x): x = x.view(x.shape[0], -1) return x class MNISTCNN(nn.Module): """ The CNN learned on the MNIST dataset in the paper: N. Carlini and D. A. Wagner, "Adversarial examples are not easily detected: Bypassing ten detection methods" Original code: - https://github.com/carlini/nn_breaking_detection/blob/master/setup_mnist.py Parameters ---------- num_classes : int Number of classes (outputs). Default 10. init_strategy : str If 'default', use the default initialization strategy for all the layers, If `fan_out`, use the default init strategy for the linear layer and the `kaiming_normal` init strategy with the option fan_out for the convolutional layers. If `uniform_scaling` use the uniform scaling strategy for all the layers """ def __init__(self, num_classes=10, init_strategy='default'): nb_filters = 64 super(MNISTCNN, self).__init__() self.features = nn.Sequential(OrderedDict([ ('conv1', nn.Conv2d(1, nb_filters, kernel_size=5, stride=2)), ('relu1', nn.ReLU(inplace=True)), ('conv2', nn.Conv2d(in_channels=nb_filters, out_channels=nb_filters, kernel_size=3, stride=2)), ('relu2', nn.ReLU(inplace=True)), ('conv3', nn.Conv2d(in_channels=nb_filters, out_channels=nb_filters, kernel_size=3, stride=1)), ('relu3', nn.ReLU(inplace=True)), ('flatten', Flatten()), ('fc1', nn.Linear(576, out_features=32)), ('relu4', nn.ReLU(inplace=True)), ('dropout', nn.Dropout(p=.5, inplace=False)), ('flatten2', Flatten()), ])) self.classifier = nn.Linear(32, num_classes) if init_strategy == "fan_out": # https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') elif init_strategy == 'uniform_scaling': for m in self.modules(): if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear): uniform_scaling_(m.weight) uniform_scaling_(m.bias) elif init_strategy == "default": # Delving deep into rectifiers: Surpassing human - level # performance on ImageNet classification - He, K. et al. (2015) pass else: raise ValueError("Unknown initialization strategy!") def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.classifier(x) return x def mnist_cnn(lr=0.1, momentum=0.9, weight_decay=0, preprocess=None, softmax_outputs=False, random_state=None, epochs=30, batch_size=128, **kwargs): if random_state is not None: torch.manual_seed(random_state) model = MNISTCNN(**kwargs) loss = nn.CrossEntropyLoss() optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum, weight_decay=weight_decay) cnn_mnist = CClassifierPyTorch(model=model, loss=loss, optimizer=optimizer, input_shape=(1, 28, 28), random_state=random_state, preprocess=preprocess, epochs=epochs, batch_size=batch_size, softmax_outputs=softmax_outputs) return cnn_mnist def uniform_scaling_(tensor, factor=1.0): """Initialization with random values from uniform distribution without scaling variance. When initializing a deep network, it is in principle advantageous to keep the scale of the input variance constant, so it does not explode or diminish by reaching the final layer. If the input is `x` and the operation `x * W`, and we want to initialize `W` uniformly at random, we need to pick `W` from [-sqrt(3) / sqrt(dim), sqrt(3) / sqrt(dim)] to keep the scale intact, where `dim = W.shape[0]` (the size of the input). A similar calculation for convolutional networks gives an analogous result with `dim` equal to the product of the first 3 dimensions. When nonlinearities are present, we need to multiply this by a constant `factor`. See [Sussillo et al., 2014](https://arxiv.org/abs/1412.6558) ([pdf](http://arxiv.org/pdf/1412.6558.pdf)) for deeper motivation, experiments and the calculation of constants. In section 2.3 there, the constants were numerically computed: for a linear layer it's 1.0, relu: ~1.43, tanh: ~1.15. Arguments: factor: `float`. A multiplicative factor by which the values will be scaled. dtype: The tensor data type. Only float are supported. seed: `int`. Used to create a random seed for the distribution. Returns: The Initializer, or an initialized `Tensor` if shape is specified. Args: tensor: an n-dimensional `torch.Tensor` a: the lower bound of the uniform distribution b: the upper bound of the uniform distribution """ import torch import math with torch.no_grad(): shape = tensor.shape input_size = 1.0 for dim in shape[:-1]: input_size *= dim max_val = math.sqrt(3 / input_size) * factor return torch.FloatTensor(shape).uniform_(-max_val, max_val)
6,209
36.409639
88
py
secml-zoo
secml-zoo-master/models/mnist/mnist3c_cnn.py
""" .. module:: MNIST3cCNN :synopsis: CNN to be trained on MNIST 3-classes dataset .. moduleauthor:: Maura Pintor <maura.pintor@unica.it> .. moduleauthor:: Marco Melis <marco.melis@unica.it> """ import torch from torch import nn, optim from secml.ml.classifiers import CClassifierPyTorch class MNIST3cCNN(nn.Module): """Model with input size (-1, 28, 28) for MNIST 3-classes dataset.""" def __init__(self): super(MNIST3cCNN, self).__init__() self.conv1 = nn.Conv2d(1, 10, kernel_size=5) self.conv2 = nn.Conv2d(10, 20, kernel_size=5) self.conv2_drop = nn.Dropout2d() self.fc1 = nn.Linear(320, 50) self.fc2 = nn.Linear(50, 3) def forward(self, x): x = torch.relu(torch.max_pool2d(self.conv1(x), 2)) x = torch.relu(torch.max_pool2d(self.conv2_drop(self.conv2(x)), 2)) x = x.view(-1, 320) x = torch.relu(self.fc1(x)) return self.fc2(x) def mnist3c_cnn(): # Random seed torch.manual_seed(0) net = MNIST3cCNN() criterion = nn.CrossEntropyLoss() optimizer = optim.SGD(net.parameters(), lr=0.01, momentum=0.9) return CClassifierPyTorch(model=net, loss=criterion, optimizer=optimizer, epochs=20, batch_size=20, input_shape=(1, 28, 28), random_state=0)
1,486
29.346939
75
py
secml-zoo
secml-zoo-master/models/iCubWorld/utils.py
import numpy as np from secml.array import CArray from secml.ml import CNormalizerDNN from .alexnet import alexnet def attach_alexnet(clf): """Attach AlexNet (as a preprocessor) to input CClassifier. The output of `classifier:4` layer is used as input for the classifier. """ clf.preprocess = CNormalizerDNN(net=alexnet(), out_layer='classifier:4') def ds_numpy_to_pytorch(ds): """Converts ds samples from numpy flatten to pytorch flatten.""" imgs = ds.X # Pytorch networks expects images in the tensor format (C x H x W) # Our images have been flatten from the numpy format (H x W x C) # We firstly need to get back to (n_samples x H x W x C) imgs = imgs.tondarray().reshape( (imgs.shape[0], ds.header.img_h, ds.header.img_w, 3)) # Then we move the "C" axis to the correct position, # and finally ravel the rows again, done. imgs = np.moveaxis(imgs, 3, 1).reshape( imgs.shape[0], 3 * ds.header.img_h * ds.header.img_w) ds.X = CArray(imgs)
1,021
30.9375
76
py
secml-zoo
secml-zoo-master/models/iCubWorld/alexnet.py
""" .. module:: AlexNet :synopsis: AlexNet Convolutional Neural Network .. moduleauthor:: Marco Melis <marco.melis@unica.it> """ from torchvision import models from secml.ml.classifiers import CClassifierPyTorch from secml.ml.features.normalization import CNormalizerMeanStd def alexnet(): """CClassifierPyTorch with AlexNet Convolutional Neural Network.""" model = models.alexnet(pretrained=True) norm_rgb = CNormalizerMeanStd((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) net = CClassifierPyTorch(model=model, input_shape=(3, 224, 224), pretrained=True, softmax_outputs=False, preprocess=norm_rgb, random_state=0) return net
803
27.714286
79
py
secml-zoo
secml-zoo-master/models/iCubWorld/iCubWorld28/tests/test_models_icubworld28.py
from models.iCubWorld.tests import CICubWorldTestCases from secml.ml.classifiers.multiclass import CClassifierMulticlassOVA from secml.utils import fm, pickle_utils from models.iCubWorld.utils import ds_numpy_to_pytorch class TestModelsICubWorld28(CICubWorldTestCases): """Unittests for iCubWorld28 models.""" @classmethod def setUpClass(cls): CICubWorldTestCases.setUpClass() cls.ds = pickle_utils.load( fm.join(fm.abspath(__file__), 'iCubWorld28_red.gz')) ds_numpy_to_pytorch(cls.ds) def test_icubworld28_day4_svm(self): model_id = 'icubworld28-day4-svm' expected_class = CClassifierMulticlassOVA expected_acc = 0.7 self._test_model(self.ds, model_id, expected_class, expected_acc) def test_icubworld28_day4_svm_rbf(self): model_id = 'icubworld28-day4-svm-rbf' expected_class = CClassifierMulticlassOVA expected_acc = 0.69 self._test_model(self.ds, model_id, expected_class, expected_acc) if __name__ == '__main__': CICubWorldTestCases.main()
1,080
29.027778
73
py
secml-zoo
secml-zoo-master/models/iCubWorld/iCubWorld28/_exporters/icubworld28-day4-svm.py
import sys sys.path.insert(0, '../../../') from svm_ova import svm_ova from iCubWorld.utils import attach_alexnet, ds_numpy_to_pytorch from secml.data.loader import CDataLoaderICubWorld28 from secml.data.splitter import CDataSplitter from secml.ml.peval.metrics import CMetricAccuracy dl = CDataLoaderICubWorld28() dl.verbose = 2 tr = dl.load( ds_type='train', day='day4', resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(tr) clf = svm_ova() clf.verbose = 1 clf.set_params({ 'C': 0.01, 'class_weight': 'balanced' }) attach_alexnet(clf) # xval_parameters = {'C': [1e-4, 1e-3, 1e-2, 1e-1, 1e0, 1e1, 1e2]} # # xval_splitter = CDataSplitter.create('kfold', num_folds=3, random_state=0) # # clf.estimate_parameters( # tr, xval_parameters, xval_splitter, 'accuracy') clf.fit(tr.X, tr.Y) ts = dl.load( ds_type='test', day='day4', resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(ts) print("Accuracy: {:}".format( CMetricAccuracy().performance_score(ts.Y, clf.predict(ts.X)))) clf.preprocess = None state_path = '../icubworld28-day4-svm.gz' clf.save_state(state_path) import os print("Model stored into: " + os.path.abspath(state_path)) from hashlib import md5 md5_hash = md5() a_file = open(state_path, "rb") content = a_file.read() md5_hash.update(content) print('md5: ' + md5_hash.hexdigest())
1,418
20.5
76
py
secml-zoo
secml-zoo-master/models/iCubWorld/iCubWorld28/_exporters/icubworld28-day4-svm-rbf.py
import sys sys.path.insert(0, '../../../') from svm_rbf_ova import svm_rbf_ova from iCubWorld.utils import attach_alexnet, ds_numpy_to_pytorch from secml.data.loader import CDataLoaderICubWorld28 from secml.data.splitter import CDataSplitter from secml.ml.peval.metrics import CMetricAccuracy dl = CDataLoaderICubWorld28() dl.verbose = 2 tr = dl.load( ds_type='train', day='day4', resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(tr) clf = svm_rbf_ova() clf.verbose = 1 clf.set_params({ 'C': 1e3, 'kernel.gamma': 2e-4, 'class_weight': 'balanced' }) attach_alexnet(clf) # xval_parameters = {'C': [1e-2, 1e-1, 1e0, 1e1, 1e2, 1e3, 1e4], # 'kernel.gamma': [1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1e0, 1e1]} # # xval_splitter = CDataSplitter.create('kfold', num_folds=3, random_state=0) # # clf.estimate_parameters( # tr, xval_parameters, xval_splitter, 'accuracy') clf.fit(tr.X, tr.Y) ts = dl.load( ds_type='test', day='day4', resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(ts) print("Accuracy: {:}".format( CMetricAccuracy().performance_score(ts.Y, clf.predict(ts.X)))) clf.preprocess = None state_path = '../icubworld28-day4-svm-rbf.gz' clf.save_state(state_path) import os print("Model stored into: " + os.path.abspath(state_path)) from hashlib import md5 md5_hash = md5() a_file = open(state_path, "rb") content = a_file.read() md5_hash.update(content) print('md5: ' + md5_hash.hexdigest())
1,536
21.602941
78
py
secml-zoo
secml-zoo-master/models/iCubWorld/iCubWorld7/tests/test_models_icubworld7.py
from models.iCubWorld.tests import CICubWorldTestCases from secml.ml.classifiers.multiclass import CClassifierMulticlassOVA from secml.utils import fm, pickle_utils from models.iCubWorld.utils import ds_numpy_to_pytorch class TestModelsICubWorld7(CICubWorldTestCases): """Unittests for iCubWorld7 models.""" @classmethod def setUpClass(cls): CICubWorldTestCases.setUpClass() cls.ds = pickle_utils.load( fm.join(fm.abspath(__file__), 'iCubWorld7_red.gz')) ds_numpy_to_pytorch(cls.ds) def test_icubworld7_day4_svm(self): model_id = 'icubworld7-day4-svm' expected_class = CClassifierMulticlassOVA expected_acc = 0.93 self._test_model(self.ds, model_id, expected_class, expected_acc) def test_icubworld7_day4_svm_rbf(self): model_id = 'icubworld7-day4-svm-rbf' expected_class = CClassifierMulticlassOVA expected_acc = 0.94 self._test_model(self.ds, model_id, expected_class, expected_acc) if __name__ == '__main__': CICubWorldTestCases.main()
1,074
28.861111
73
py
secml-zoo
secml-zoo-master/models/iCubWorld/iCubWorld7/_exporters/icubworld7-day4-svm.py
import sys sys.path.insert(0, '../../../') from svm_ova import svm_ova from iCubWorld.utils import attach_alexnet, ds_numpy_to_pytorch from secml.data.loader import CDataLoaderICubWorld28 from secml.data.splitter import CDataSplitter from secml.ml.peval.metrics import CMetricAccuracy dl = CDataLoaderICubWorld28() dl.verbose = 2 tr = dl.load( ds_type='train', day='day4', icub7=True, resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(tr) clf = svm_ova() clf.verbose = 1 clf.set_params({ 'C': 0.1, 'class_weight': 'balanced' }) attach_alexnet(clf) # xval_parameters = {'C': [1e-3, 1e-2, 1e-1, 1e0, 1e1, 1e2, 1e3]} # # xval_splitter = CDataSplitter.create('kfold', num_folds=3, random_state=0) # # clf.estimate_parameters( # tr, xval_parameters, xval_splitter, 'accuracy') clf.fit(tr.X, tr.Y) ts = dl.load( ds_type='test', day='day4', icub7=True, resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(ts) print("Accuracy: {:}".format( CMetricAccuracy().performance_score(ts.Y, clf.predict(ts.X)))) clf.preprocess = None state_path = '../icubworld7-day4-svm.gz' clf.save_state(state_path) import os print("Model stored into: " + os.path.abspath(state_path)) from hashlib import md5 md5_hash = md5() a_file = open(state_path, "rb") content = a_file.read() md5_hash.update(content) print('md5: ' + md5_hash.hexdigest())
1,439
20.818182
76
py
secml-zoo
secml-zoo-master/models/iCubWorld/iCubWorld7/_exporters/icubworld7-day4-svm-rbf.py
import sys sys.path.insert(0, '../../../') from svm_rbf_ova import svm_rbf_ova from iCubWorld.utils import attach_alexnet, ds_numpy_to_pytorch from secml.data.loader import CDataLoaderICubWorld28 from secml.ml.peval.metrics import CMetricAccuracy dl = CDataLoaderICubWorld28() dl.verbose = 2 tr = dl.load( ds_type='train', day='day4', icub7=True, resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(tr) clf = svm_rbf_ova() clf.verbose = 1 clf.set_params({ 'C': 1e2, 'kernel.gamma': 2e-4, 'class_weight': 'balanced' }) attach_alexnet(clf) # xval_parameters = {'C': [1e-2, 1e-1, 1e0, 1e1, 1e2, 1e3, 1e4], # 'kernel.gamma': [1e-5, 1e-4, 1e-3, 1e-2, 1e-1, 1e0, 1e1]} # # xval_splitter = CDataSplitter.create('kfold', num_folds=3, random_state=0) # # clf.estimate_parameters( # tr, xval_parameters, xval_splitter, 'accuracy') clf.fit(tr.X, tr.Y) ts = dl.load( ds_type='test', day='day4', icub7=True, resize_shape=(256, 256), crop_shape=(224, 224), normalize=True ) ds_numpy_to_pytorch(ts) print("Accuracy: {:}".format( CMetricAccuracy().performance_score(ts.Y, clf.predict(ts.X)))) clf.preprocess = None state_path = '../icubworld7-day4-svm-rbf.gz' clf.save_state(state_path) import os print("Model stored into: " + os.path.abspath(state_path)) from hashlib import md5 md5_hash = md5() a_file = open(state_path, "rb") content = a_file.read() md5_hash.update(content) print('md5: ' + md5_hash.hexdigest())
1,513
21.597015
78
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/main.py
import os import copy import logging import cv2 import numpy as np import pandas as pd from tqdm import tqdm from matplotlib import pyplot as plt import torch import torch.utils.data from torch.utils.tensorboard import SummaryWriter from torch.optim.lr_scheduler import ReduceLROnPlateau as ReduceLROnPlateau from config import ParseConfig from utils import fast_argmax from utils import visualize_image from utils import heatmap_loss from utils import count_parameters from utils import get_pairwise_joint_distances from activelearning import ActiveLearning from activelearning_viz import ActiveLearning_Visualization from dataloader import load_hp_dataset from dataloader import HumanPoseDataLoader from evaluation import PercentageCorrectKeypoint from models.auxiliary.AuxiliaryNet import AuxNet from models.hrnet.pose_hrnet import PoseHighResolutionNet as HRNet from models.stacked_hourglass.StackedHourglass import PoseNet as Hourglass # Global declarations torch.backends.cudnn.benchmark = True torch.backends.cudnn.enabled = True logging.getLogger().setLevel(logging.INFO) os.chdir(os.path.dirname(os.path.realpath(__file__))) class Train(object): def __init__(self, pose_model, aux_net, hyperparameters, dataset_obj, conf, tb_writer): """ Class for training the human pose and aux_net model :param pose_model: (torch.nn) Human pose model :param aux_net: (torch.nn) Auxiliary network :param hyperparameters: (dict) Various hyperparameters used in training :param dataset_obj: (torch.utils.data.Dataset) :param conf: (Object of ParseConfig) Contains the configurations for the model :param tb_writer: (Object of SummaryWriter) Tensorboard writer to log values """ self.conf = conf self.aux_net = aux_net self.network = pose_model self.tb_writer = tb_writer self.dataset_obj = dataset_obj self.hyperparameters = hyperparameters # Experiment Settings self.batch_size = conf.experiment_settings['batch_size'] self.epoch = hyperparameters['num_epochs'] self.optimizer = hyperparameters['optimizer'] # Adam / SGD self.loss_fn = hyperparameters['loss_fn'] # MSE self.learning_rate = hyperparameters['optimizer_config']['lr'] self.start_epoch = hyperparameters['start_epoch'] # Used in case of resume training self.num_hm = conf.experiment_settings['num_hm'] # Number of heatmaps self.joint_names = self.dataset_obj.ind_to_jnt self.model_save_path = conf.model['save_path'] self.train_aux_net = conf.model['aux_net']['train'] # Stacked Hourglass scheduling if self.train_aux_net: min_lr = [0.000003, conf.experiment_settings['lr']] else: min_lr = 0.000003 self.scheduler = ReduceLROnPlateau(self.optimizer, factor=0.5, patience=5, cooldown=2, min_lr=min_lr, verbose=True) self.torch_dataloader = torch.utils.data.DataLoader(self.dataset_obj, self.batch_size, shuffle=True, num_workers=2, drop_last=True) def train_model(self): """ Training loop :return: """ print("Initializing training: Epochs - {}\tBatch Size - {}".format( self.hyperparameters['num_epochs'], self.batch_size)) # 'mean_loss_validation': {'Pose': validation_loss_pose, 'AuxNet': validation_aux_net} if self.conf.resume_training: best_val_pose = self.hyperparameters['mean_loss_validation']['Pose'] best_val_auxnet = self.hyperparameters['mean_loss_validation']['AuxNet'] best_epoch_pose = self.hyperparameters['start_epoch'] best_epoch_auxnet = -1 global_step = 0 else: best_val_pose = np.inf best_val_auxnet = np.inf best_epoch_pose = -1 best_epoch_auxnet = -1 global_step = 0 # Variable to store all the loss values for logging loss_across_epochs = [] validation_across_epochs = [] for e in range(self.start_epoch, self.epoch): epoch_loss = [] epoch_loss_aux_net = [] # Network alternates between train() and validate() self.network.train() if self.train_aux_net: self.aux_net.train() self.dataset_obj.input_dataset(train=True) # Training loop logging.info('Training for epoch: {}'.format(e+1)) for (images, heatmaps, _, _, _, gt_per_image, split, _, _, _, joint_exist) in tqdm(self.torch_dataloader): assert split[0] == 0, "Training split should be 0." self.optimizer.zero_grad() outputs, pose_features = self.network(images) # images.cuda() done internally within the model loss = heatmap_loss(outputs, heatmaps) # heatmaps transferred to GPU within the function if self.conf.model['aux_net']['train'] and self.conf.model['aux_net']['method'] == 'learning_loss': learning_loss_ = loss.clone().detach().to('cuda:{}'.format(torch.cuda.device_count() - 1)) learning_loss_ = torch.mean(learning_loss_, dim=[1]) loss_learnloss = self.learning_loss(pose_encodings=pose_features, true_loss=learning_loss_, gt_per_img=gt_per_image, epoch=e) loss_learnloss.backward() epoch_loss_aux_net.append(loss_learnloss.item()) if self.conf.model['aux_net']['train'] and self.conf.model['aux_net']['method'] == 'aleatoric': loss_aleatoric = self.aleatoric_uncertainty(pose_encodings=pose_features, outputs=outputs, heatmaps=heatmaps, joint_exist=joint_exist, epoch=e) loss_aleatoric.backward() epoch_loss_aux_net.append(loss_aleatoric.item()) if self.conf.model['aux_net']['train'] and self.conf.model['aux_net']['method'] == 'vl4pose': loss_vl4pose = self.vl4pose(pose_encodings=pose_features, heatmaps=heatmaps, joint_exist=joint_exist, epoch=e) loss_vl4pose.backward() epoch_loss_aux_net.append(loss_vl4pose.item()) if self.conf.model['aux_net']['train_auxnet_only']: loss = torch.mean(loss) * 0 else: loss = torch.mean(loss) loss.backward() if self.conf.tensorboard: self.tb_writer.add_scalar('Train/Loss_batch', torch.mean(loss), global_step) epoch_loss.append(loss.item()) # Weight update self.optimizer.step() global_step += 1 # Epoch training ends ------------------------------------------------------------------------------------- epoch_loss = np.mean(epoch_loss) if self.conf.model['aux_net']['train']: epoch_loss_aux_net = np.mean(epoch_loss_aux_net) validation_loss_pose, validation_aux_net = self.validation(e) else: validation_loss_pose = self.validation(e) validation_aux_net = 0.0 # Learning rate scheduler on the Human Pose validation loss self.scheduler.step(validation_loss_pose) # TensorBoard Summaries if self.conf.tensorboard: self.tb_writer.add_scalar('Train', torch.tensor([epoch_loss]), global_step) self.tb_writer.add_scalar('Validation/HG_Loss', torch.tensor([validation_loss_pose]), global_step) if self.conf.model['aux_net']['train']: self.tb_writer.add_scalar('Validation/Learning_Loss', torch.tensor([validation_aux_net]), global_step) # Save if best model if best_val_pose > validation_loss_pose: torch.save(self.network.state_dict(), os.path.join(self.model_save_path, 'model_checkpoints/pose_net.pth')) if self.conf.model['aux_net']['train']: torch.save( self.aux_net.state_dict(), os.path.join(self.model_save_path, 'model_checkpoints/aux_net_{}_BestPose.pth'.format(self.conf.model['aux_net']['method']))) best_val_pose = validation_loss_pose best_epoch_pose = e + 1 torch.save({'epoch': e + 1, 'optimizer_load_state_dict': self.optimizer.state_dict(), 'mean_loss_train': epoch_loss, 'mean_loss_validation': {'Pose': validation_loss_pose, 'AuxNet': validation_aux_net}, 'aux_net': self.conf.model['aux_net']['train']}, os.path.join(self.model_save_path, 'model_checkpoints/optim_best_model.tar')) if self.conf.model['aux_net']['train']: if (best_val_auxnet > validation_aux_net) and (validation_aux_net != 0.0): torch.save(self.aux_net.state_dict(), os.path.join(self.model_save_path, 'model_checkpoints/aux_net_{}.pth'.format(self.conf.model['aux_net']['method']))) best_val_auxnet = validation_aux_net best_epoch_auxnet = e + 1 print("Loss at epoch {}/{}: (train:Pose) {}\t" "(train:AuxNet) {}\t" "(validation:Pose) {}\t" "(Validation:AuxNet) {}\t" "(Best Model) {}".format( e+1, self.epoch, epoch_loss, epoch_loss_aux_net, validation_loss_pose, validation_aux_net, best_epoch_pose)) loss_across_epochs.append(epoch_loss) validation_across_epochs.append(validation_loss_pose) # Save the loss values f = open(os.path.join(self.model_save_path, 'model_checkpoints/loss_data.txt'), "w") f_ = open(os.path.join(self.model_save_path, 'model_checkpoints/validation_data.txt'), "w") f.write("\n".join([str(lsx) for lsx in loss_across_epochs])) f_.write("\n".join([str(lsx) for lsx in validation_across_epochs])) f.close() f_.close() if self.conf.tensorboard: self.tb_writer.close() logging.info("Model training completed\nBest validation loss (Pose): {}\tBest Epoch: {}" "\nBest validation loss (AuxNet): {}\tBest Epoch: {}".format( best_val_pose, best_epoch_pose, best_val_auxnet, best_epoch_auxnet)) def validation(self, e): """ :param e: Epoch :return: """ with torch.no_grad(): # Stores the loss for all batches epoch_val_pose = [] self.network.eval() if self.conf.model['aux_net']['train']: epoch_val_auxnet = [] self.aux_net.eval() # Augmentation only needed in Training self.dataset_obj.input_dataset(validate=True) # Compute and store batch-wise validation loss in a list logging.info('Validation for epoch: {}'.format(e+1)) for (images, heatmaps, _, _, _, gt_per_img, split, _, _, _, joint_exist) in tqdm(self.torch_dataloader): assert split[0] == 1, "Validation split should be 1." outputs, pose_features = self.network(images) loss_val_pose = heatmap_loss(outputs, heatmaps) if self.conf.model['aux_net']['train'] and self.conf.model['aux_net']['method'] == 'learning_loss': learning_loss_val = loss_val_pose.clone().detach().to('cuda:{}'.format(torch.cuda.device_count() - 1)) learning_loss_val = torch.mean(learning_loss_val, dim=[1]) loss_val_auxnet = self.learning_loss(pose_features, learning_loss_val, gt_per_img, e) epoch_val_auxnet.append(loss_val_auxnet.item()) if self.conf.model['aux_net']['train'] and self.conf.model['aux_net']['method'] == 'aleatoric': loss_val_aleatoric = self.aleatoric_uncertainty(pose_encodings=pose_features, outputs=outputs, heatmaps=heatmaps, joint_exist=joint_exist, epoch=e) epoch_val_auxnet.append(loss_val_aleatoric.item()) if self.conf.model['aux_net']['train'] and self.conf.model['aux_net']['method'] == 'vl4pose': vl4pose_loss = self.vl4pose(pose_encodings=pose_features, heatmaps=heatmaps, joint_exist=joint_exist, epoch=e) epoch_val_auxnet.append(vl4pose_loss.item()) loss_val_pose = torch.mean(loss_val_pose) epoch_val_pose.append(loss_val_pose.item()) if self.conf.model['aux_net']['train']: return np.mean(epoch_val_pose), np.mean(epoch_val_auxnet) else: return np.mean(epoch_val_pose) def learning_loss(self, pose_encodings, true_loss, gt_per_img, epoch): ''' Learning Loss module Based on the paper: "Learning Loss For Active Learning, CVPR 2019" and "A Mathematical Analysis of Learning Loss for Active Learning in Regression, CVPR-W 2021" :param pose_encodings: (Dict of tensors) Intermediate (Hourglass) and penultimate layer output of the M10 network :param true_loss: (Tensor of shape [Batch Size]) Loss computed from M10 prediction and ground truth :param gt_per_img: (Tensor, shape [Batch Size]) Number of ground truth per image :param epoch: (scalar) Epoch, used in learning loss warm start-up :return: (Torch scalar tensor) Learning Loss ''' learnloss_margin = self.conf.active_learning['learningLoss']['margin'] learnloss_objective = self.conf.active_learning['learningLoss']['objective'] learnloss_warmup = self.conf.model['aux_net']['warmup'] emperical_loss = self._aux_net_inference(pose_encodings) emperical_loss = emperical_loss.squeeze() assert emperical_loss.shape == true_loss.shape, "Mismatch in Batch size for true and emperical loss" with torch.no_grad(): # Scale the images as per the number of joints # To prevent DivideByZero. PyTorch does not throw an exception to DivideByZero gt_per_img = torch.sum(gt_per_img, dim=1) gt_per_img += 0.1 true_loss = true_loss / gt_per_img.to(true_loss.device) # Splitting into pairs: (i, i+half) half_split = true_loss.shape[0] // 2 true_loss_i = true_loss[: half_split] true_loss_j = true_loss[half_split: 2 * half_split] emp_loss_i = emperical_loss[: (emperical_loss.shape[0] // 2)] emp_loss_j = emperical_loss[(emperical_loss.shape[0] // 2): 2 * (emperical_loss.shape[0] // 2)] # Pair wise loss as mentioned in the original paper if learnloss_objective == 'YooAndKweon': loss_sign = torch.sign(true_loss_i - true_loss_j) loss_emp = (emp_loss_i - emp_loss_j) # Learning Loss objective llal_loss = torch.max(torch.zeros(half_split, device=loss_sign.device), (-1 * (loss_sign * loss_emp)) + learnloss_margin) # Computing loss over the entire batch using softmax. elif learnloss_objective == 'KLdivergence': # Removed the standardization-KL Divergence parts with torch.no_grad(): true_loss_ = torch.cat([true_loss_i.reshape(-1, 1), true_loss_j.reshape(-1, 1)], dim=1) true_loss_scaled = true_loss_ / torch.sum(true_loss_, dim=1, keepdim=True) emp_loss_ = torch.cat([emp_loss_i.reshape(-1, 1), emp_loss_j.reshape(-1, 1)], dim=1) emp_loss_logsftmx = torch.nn.LogSoftmax(dim=1)(emp_loss_) # Scaling the cross entropy loss with respect to true loss values llal_loss = torch.nn.KLDivLoss(reduction='batchmean')(input=emp_loss_logsftmx, target=true_loss_scaled) #llal_loss = torch.sum((-true_loss_scaled * torch.log(emp_loss_logsftmx)), dim=1, keepdim=True) else: raise NotImplementedError('Currently only "YooAndKweon" or "KLdivergence" supported. ') if learnloss_warmup <= epoch: return torch.mean(llal_loss) else: return 0.0 * torch.mean(llal_loss) def aleatoric_uncertainty(self, pose_encodings, outputs, heatmaps, joint_exist, epoch): """ Extension of Kendall and Gal's method for calculating uncertainty to human pose estimation Auxiliary Network module to train sigmas for the HG/HRN network's heatmaps directly :param pose_encodings: (Dict of tensors) Intermediate and penultimate layer outputs of the pose model :param outputs: Tensor of size (batch_size, num_joints, hm_size, hm_size) Outputs of the main HG/HRN network :param heatmaps: Tensor of size (batch_size, num_joints, hm_size, hm_size) Ground truth heatmaps :param joint_exist: Tensor of size (batch_size, num_joints) Joint status (1=Present, 0=Absent) Present=1 may include occluded joints depending on configuration.yml setting :param epoch: (scalar) Epoch, used in auxiliary network loss to compare warmp-up :return: auxnet_loss: (Torch scalar tensor) auxiliary network Loss """ parameters = self._aux_net_inference(pose_encodings) # Final stack of hourglass / output of HRNet outputs = outputs[:, -1].clone().detach().to('cuda:{}'.format(torch.cuda.device_count() - 1)) assert heatmaps.shape == outputs.shape # Batch Size x num_joints x 64 x 64 joint_exist = joint_exist.float().to(device=parameters.device) residual = torch.sum((fast_argmax(outputs) - fast_argmax(heatmaps).to(outputs.device))**2, dim=-1) # along axis representing u,v residual = 0.5 * residual * torch.exp(-parameters) neg_log_likelihood = residual + (0.5 * parameters) neg_log_likelihood = neg_log_likelihood * joint_exist if self.conf.model['aux_net']['warmup'] <= epoch: return torch.mean(neg_log_likelihood) else: return 0.0 * torch.mean(neg_log_likelihood) def vl4pose(self, pose_encodings, heatmaps, joint_exist, epoch): ''' Train the auxiliary network for VL4Pose. - param pose_encodings: (Dict of tensors) output from the pose network - param heatmaps: Tensor of size (batch_size, num_joints, hm_size, hm_size) - param joint_exist: Tensor of size (batch_size, num_joints) - Joint status (1=Present, 0=Absent) - Present=1 may include occluded joints depending on configuration.yml setting - param epoch: (scalar) Epoch, used in auxiliary network loss warm start-up - return auxnet_loss: (Torch scalar tensor) auxiliary network Loss ''' assert joint_exist.dim() == 2, "joint_exist should be BS x num_hm, received: {}".format(joint_exist.shape) j2i = self.dataset_obj.jnt_to_ind if self.conf.dataset['load'] == 'mpii': links = [[j2i['head'], j2i['neck']], [j2i['neck'], j2i['thorax']], [j2i['thorax'], j2i['pelvis']], [j2i['thorax'], j2i['lsho']], [j2i['lsho'], j2i['lelb']], [j2i['lelb'], j2i['lwri']], [j2i['thorax'], j2i['rsho']], [j2i['rsho'], j2i['relb']], [j2i['relb'], j2i['rwri']], [j2i['pelvis'], j2i['lhip']], [j2i['lhip'], j2i['lknee']], [j2i['lknee'], j2i['lankl']], [j2i['pelvis'], j2i['rhip']], [j2i['rhip'], j2i['rknee']], [j2i['rknee'], j2i['rankl']]] else: links = [[j2i['head'], j2i['neck']], [j2i['neck'], j2i['lsho']], [j2i['lsho'], j2i['lelb']], [j2i['lelb'], j2i['lwri']], [j2i['neck'], j2i['rsho']], [j2i['rsho'], j2i['relb']], [j2i['relb'], j2i['rwri']], [j2i['lsho'], j2i['lhip']], [j2i['lhip'], j2i['lknee']], [j2i['lknee'], j2i['lankl']], [j2i['rsho'], j2i['rhip']], [j2i['rhip'], j2i['rknee']], [j2i['rknee'], j2i['rankl']]] parameters = self._aux_net_inference(pose_encodings) parameters = parameters.reshape(self.batch_size, len(links), 2) joint_exist = joint_exist.to(parameters.device) heatmaps = heatmaps.to(parameters.device) with torch.no_grad(): joint_distances = get_pairwise_joint_distances(heatmaps) joint_exist = torch.matmul(joint_exist.unsqueeze(2).type(torch.float16), joint_exist.unsqueeze(1).type(torch.float16)) # Batch Size x num_links skeleton_exist = torch.stack([joint_exist[:, u, v] for u,v in links], dim=1) skeleton_distances = torch.stack([joint_distances[:, u, v] for u,v in links], dim=1) #### residual = (parameters[:, :, 0].squeeze() - skeleton_distances)**2 residual = 0.5 * residual * torch.exp(-parameters[:, :, 1]).squeeze() neg_log_likelihood = residual + (0.5 * parameters[:, :, 1].squeeze()) neg_log_likelihood = neg_log_likelihood * skeleton_exist if self.conf.model['aux_net']['warmup'] <= epoch: return torch.mean(neg_log_likelihood) else: return 0.0 * torch.mean(neg_log_likelihood) def _aux_net_inference(self, pose_features): """ Common to VL4Pose, LearningLoss++ and Aleatoric which all use an auxiliary network """ extractor = self.conf.architecture['aux_net']['conv_or_avg_pooling'] with torch.no_grad(): if extractor == 'avg': # Transfer to GPU where auxiliary network is stored encodings = pose_features['penultimate'] else: depth = len(self.conf.architecture['aux_net']['spatial_dim']) encodings = torch.cat( [pose_features['feature_{}'.format(i)].reshape( self.batch_size, pose_features['feature_{}'.format(i)].shape[1], -1) for i in range(depth, 0, -1)], dim=2) aux_out = self.aux_net(encodings) return aux_out class Metric(object): def __init__(self, network, dataset_obj, conf): ''' Class for Testing the model: 1. Compute ground truth and predictions 2. Computing metrics: PCK@0.x :param network: (torch.nn) Hourglass network to compute predictions :param dataset_obj: (Dataset object) Handles data to be fed to PyTorch DataLoader :param conf: (Object of ParseConfig) Configuration for the experiment ''' self.dataset_obj = dataset_obj self.dataset_obj.input_dataset(validate=True) self.network = network self.viz=conf.viz # Controls visualization self.conf = conf self.batch_size = conf.experiment_settings['batch_size'] self.ind_to_jnt = self.dataset_obj.ind_to_jnt self.torch_dataloader = torch.utils.data.DataLoader(self.dataset_obj, batch_size=self.batch_size, shuffle=False, num_workers=2) def inference(self): ''' Obtains model inference :return: None ''' self.network.eval() logging.info("Starting model inference") outputs_ = None scale_ = None num_gt_ = None dataset_ = None name_ = None gt_ = None normalizer_ = None with torch.no_grad(): for (images, _, gt, name, dataset, num_gt, split, _, scale_params, normalizer, joint_exist) in tqdm( self.torch_dataloader): assert split[0] == 1, "Validation split should be 1." outputs, pose_features = self.network(images) outputs = outputs[:, -1] try: outputs_ = torch.cat((outputs_, outputs.cpu().clone()), dim=0) scale_['scale_factor'] = torch.cat((scale_['scale_factor'], scale_params['scale_factor']), dim=0) scale_['padding_u'] = torch.cat((scale_['padding_u'], scale_params['padding_u']), dim=0) scale_['padding_v'] = torch.cat((scale_['padding_v'], scale_params['padding_v']), dim=0) num_gt_ = torch.cat((num_gt_, num_gt), dim=0) dataset_ = dataset_ + dataset name_ = name_ + name gt_ = torch.cat((gt_, gt), dim=0) normalizer_ = torch.cat((normalizer_, normalizer), dim=0) except TypeError: outputs_ = outputs.cpu().clone() scale_ = copy.deepcopy(scale_params) num_gt_ = num_gt dataset_ = dataset name_ = name gt_ = gt normalizer_ = normalizer # Generate visualizations (256x256) for that batch of images if self.conf.viz: hm_uv_stack = [] # Compute u,v values from heatmap for i in range(images.shape[0]): hm_uv = self.dataset_obj.estimate_uv(hm_array=outputs.cpu().numpy()[i], pred_placeholder=-np.ones_like(gt[i].numpy())) hm_uv_stack.append(hm_uv) hm_uv = np.stack(hm_uv_stack, axis=0) self.visualize_predictions(image=images.numpy(), name=name, dataset=dataset, gt=gt.numpy(), pred=hm_uv) scale_['scale_factor'] = scale_['scale_factor'].numpy() scale_['padding_u'] = scale_['padding_u'].numpy() scale_['padding_v'] = scale_['padding_v'].numpy() model_inference = {'heatmap': outputs_.numpy(), 'scale': scale_, 'dataset': dataset_, 'name': name_, 'gt': gt_.numpy(), 'normalizer': normalizer_.numpy()} return model_inference def keypoint(self, infer): ''' :param infer: :return: ''' heatmap = infer['heatmap'] scale = infer['scale'] dataset = infer['dataset'] name = infer['name'] gt = infer['gt'] normalizer = infer['normalizer'] hm_uv_stack = [] csv_columns = ['name', 'dataset', 'normalizer', 'joint', 'uv'] gt_csv = [] pred_csv = [] # Iterate over all heatmaps to obtain predictions for i in range(gt.shape[0]): heatmap_ = heatmap[i] gt_uv = gt[i] hm_uv = self.dataset_obj.estimate_uv(hm_array=heatmap_, pred_placeholder=-np.ones_like(gt_uv)) hm_uv_stack.append(hm_uv) # Scaling the point ensures that the distance between gt and pred is same as the scale of normalization scale_factor = scale['scale_factor'][i] padding_u = scale['padding_u'][i] padding_v = scale['padding_v'][i] # Scaling ground truth gt_uv_correct = np.copy(gt_uv) hm_uv_correct = np.copy(hm_uv) gt_uv_correct[:, :, 1] -= padding_v gt_uv_correct[:, :, 0] -= padding_u gt_uv_correct /= np.array([scale_factor, scale_factor, 1]).reshape(1, 1, 3) # Scaling predictions hm_uv_correct[:, :, 1] -= padding_v hm_uv_correct[:, :, 0] -= padding_u hm_uv_correct /= np.array([scale_factor, scale_factor, 1]).reshape(1, 1, 3) assert gt_uv_correct.shape == hm_uv_correct.shape, "Mismatch in gt ({}) and prediction ({}) shape".format( gt_uv_correct.shape, hm_uv_correct.shape) # Iterate over joints for jnt in range(gt_uv_correct.shape[1]): gt_entry = { 'name': name[i], 'dataset': dataset[i], 'normalizer': normalizer[i], 'joint': self.ind_to_jnt[jnt], 'uv': gt_uv_correct[:, jnt, :].astype(np.float32) } pred_entry = { 'name': name[i], 'dataset': dataset[i], 'normalizer': normalizer[i], 'joint': self.ind_to_jnt[jnt], 'uv': hm_uv_correct[:, jnt, :].astype(np.float32) } gt_csv.append(gt_entry) pred_csv.append(pred_entry) pred_csv = pd.DataFrame(pred_csv, columns=csv_columns) gt_csv = pd.DataFrame(gt_csv, columns=csv_columns) pred_csv.sort_values(by='dataset', ascending=True, inplace=True) gt_csv.sort_values(by='dataset', ascending=True, inplace=True) assert len(pred_csv.index) == len(gt_csv.index), "Mismatch in number of entries in pred and gt dataframes." pred_csv.to_csv(os.path.join(self.conf.model['save_path'], 'model_checkpoints/pred.csv'), index=False) gt_csv.to_csv(os.path.join(self.conf.model['save_path'], 'model_checkpoints/gt.csv'), index=False) logging.info('Pandas dataframe saved successfully.') return gt_csv, pred_csv def visualize_predictions(self, image=None, name=None, dataset=None, gt=None, pred=None): dataset_viz = {} dataset_viz['img'] = image dataset_viz['name'] = name dataset_viz['display_string'] = name dataset_viz['split'] = np.ones(image.shape[0]) dataset_viz['dataset'] = dataset dataset_viz['bbox_coords'] = np.zeros([image.shape[0], 4, 4]) dataset_viz['num_persons'] = np.ones([image.shape[0], 1]) dataset_viz['gt'] = gt dataset_viz['pred'] = pred dataset_viz = self.dataset_obj.recreate_images(gt=True, pred=True, external=True, ext_data=dataset_viz) visualize_image(dataset_viz, save_dir=self.conf.model['save_path'], bbox=False) def compute_metrics(self, gt_df=None, pred_df=None): ''' Loads the ground truth and prediction CSVs into memory. Evaluates Precision, FPFN metrics for the prediction and stores them into memory. :return: None ''' # Ensure that same datasets have been loaded assert all(pred_df['dataset'].unique() == gt_df['dataset'].unique()), \ "Mismatch in dataset column for gt and pred" logging.info('Generating evaluation metrics for dataset:') # Iterate over unique datasets for dataset_ in gt_df['dataset'].unique(): logging.info(str(dataset_)) # Separate out images based on dataset pred_ = pred_df.loc[pred_df['dataset'] == dataset_] gt_ = gt_df.loc[gt_df['dataset'] == dataset_] # Compute scores pck_df = PercentageCorrectKeypoint( pred_df=pred_, gt_df=gt_, config=self.conf, jnts=list(self.ind_to_jnt.values())) # Save the tables if dataset_ == 'mpii': metric_ = 'PCKh' else: metric_ = 'PCK' pck_df.to_csv(os.path.join(self.conf.model['save_path'], 'model_checkpoints/{}_{}.csv'.format(metric_, dataset_)), index=False) print("Metrics computation completed.") def eval(self): ''' :return: ''' model_inference = self.inference() gt_csv, pred_csv = self.keypoint(model_inference) self.compute_metrics(gt_df=gt_csv, pred_df=pred_csv) def load_models(conf, load_pose, load_aux, model_dir): """ :param conf: :param load_pose: :param load_aux: :param model_dir: :return: """ # Initialize AuxNet, Hourglass/HRNet # Elsewhere, resume training ensures the code creates a copy of the best models from the interrupted run. if conf.use_auxnet: logging.info('Initializing Auxiliary Network') aux_net = AuxNet(arch=conf.architecture['aux_net']) else: aux_net = None if conf.model['type'] == 'hourglass': logging.info('Initializing Hourglass Network') pose_net = Hourglass(arch=conf.architecture['hourglass'], auxnet=conf.use_auxnet, intermediate_features=conf.architecture['aux_net']['conv_or_avg_pooling']) print('Number of parameters (Hourglass): {}\n'.format(count_parameters(pose_net))) else: logging.info('Initializing HRNet') assert conf.model['type'] == 'hrnet', "Currently support 'hourglass' and 'hrnet'." pose_net = HRNet(arch=conf.architecture['hrnet'], auxnet=conf.use_auxnet, intermediate_features=conf.architecture['aux_net']['conv_or_avg_pooling']) print('Number of parameters (HRNet): {}\n'.format(count_parameters(pose_net))) # Load AuxNet modules (Best Model / resume training) if load_aux: if conf.resume_training: logging.info('\n-------------- Resuming training (Loading AuxNet) --------------\n') # Load and save the previous best model aux_net.load_state_dict(torch.load( os.path.join(model_dir, 'model_checkpoints/aux_net_{}.pth'.format(conf.active_learning['algorithm'])), map_location='cpu')) torch.save( aux_net.state_dict(), os.path.join( conf.model['save_path'], 'model_checkpoints/aux_net_{}.pth'.format(conf.active_learning['algorithm']))) # Load the model corresponding to best pose model aux_net.load_state_dict( torch.load( os.path.join( model_dir, 'model_checkpoints/aux_net_{}_BestPose.pth'.format(conf.active_learning['algorithm'])), map_location='cpu')) else: logging.info('Loading AuxNet Best Model') aux_net.load_state_dict( torch.load(os.path.join( model_dir, 'model_checkpoints/aux_net_{}.pth'.format(conf.active_learning['algorithm'])), map_location='cpu')) # Load Pose model (code is independent of architecture) if load_pose: # Load model logging.info('Loading Pose model from: ' + model_dir) pose_net.load_state_dict(torch.load(os.path.join(model_dir, 'model_checkpoints/pose_net.pth'), map_location='cpu')) logging.info("Successfully loaded Pose model.") if conf.resume_training: logging.info('\n-------------- Resuming training (Loading PoseNet) --------------\n') torch.save(pose_net.state_dict(), os.path.join(conf.model['save_path'], 'model_checkpoints/pose_net.pth')) # CUDA support: Single/Multi-GPU # Hourglass net and HRNet have CUDA definitions inside __init__(), specify only for aux_net if conf.model['aux_net']['train'] or load_aux: aux_net.cuda(torch.device('cuda:{}'.format(torch.cuda.device_count()-1))) logging.info('Successful: Model transferred to GPUs.\n') return pose_net, aux_net def define_hyperparams(conf, pose_model, aux_net):#(conf, net, learnloss): """ :param conf: :param pose_model: :param aux_net: :return: """ logging.info('Initializing the hyperparameters for the experiment.') hyperparameters = dict() hyperparameters['optimizer_config'] = { 'lr': conf.experiment_settings['lr'], 'weight_decay': conf.experiment_settings['weight_decay'] } hyperparameters['loss_params'] = {'size_average': True} hyperparameters['num_epochs'] = conf.experiment_settings['epochs'] hyperparameters['start_epoch'] = 0 # Used for resume training # Parameters declared to the optimizer if conf.model['aux_net']['train']: logging.info('Parameters of AuxNet and PoseNet passed to Optimizer.') params_list = [{'params': pose_model.parameters()}, {'params': aux_net.parameters()}] else: logging.info('Parameters of PoseNet passed to Optimizer') params_list = [{'params': pose_model.parameters()}] hyperparameters['optimizer'] = torch.optim.Adam(params_list, **hyperparameters['optimizer_config']) if conf.resume_training: logging.info('Loading optimizer state dictionary') optim_dict = torch.load(os.path.join(conf.model['load_path'], 'model_checkpoints/optim_best_model.tar')) # If the previous experiment trained aux_net, ensure the flag is true for the current experiment assert optim_dict['aux_net'] == conf.model['aux_net']['train'], "AuxNet model needed to resume training" hyperparameters['optimizer'].load_state_dict(optim_dict['optimizer_load_state_dict']) logging.info('Optimizer state loaded successfully.\n') logging.info('Optimizer and Training parameters:\n') for key in optim_dict: if key == 'optimizer_load_state_dict': logging.info('Param group length: {}'.format(len(optim_dict[key]['param_groups']))) else: logging.info('Key: {}\tValue: {}'.format(key, optim_dict[key])) logging.info('\n') hyperparameters['start_epoch'] = optim_dict['epoch'] hyperparameters['mean_loss_validation'] = optim_dict['mean_loss_validation'] hyperparameters['loss_fn'] = torch.nn.MSELoss(reduction='none') return hyperparameters def main(): """ Control flow for the code """ # 1. Load configuration file -------------------------------------------------------------------------------------- logging.info('Loading configurations.\n') conf = ParseConfig() # 2. Loading datasets --------------------------------------------------------------------------------------------- logging.info('Loading pose dataset(s)\n') dataset_dict = load_hp_dataset(dataset_conf=conf.dataset, model_conf=conf.model) # 3. Defining the network ----------------------------------------------------------------------------------------- logging.info('Initializing (and loading) human pose network and auxiliary network for Active Learning.\n') pose_model, aux_net = load_models(conf=conf, load_pose=conf.model['load'], load_aux=conf.model['aux_net']['load'], model_dir=conf.model['load_path']) # 4. Defining the Active Learning library -------------------------------------------------------------------------- logging.info('Importing active learning object.\n') if conf.activelearning_viz: activelearning = ActiveLearning_Visualization(conf=conf, pose_net=pose_model, aux_net=aux_net) else: activelearning = ActiveLearning(conf=conf, pose_net=pose_model, aux_net=aux_net) # 5. Defining DataLoader ------------------------------------------------------------------------------------------- logging.info('Defining DataLoader.\n') datasets = HumanPoseDataLoader(dataset_dict=dataset_dict, activelearning=activelearning, conf=conf) # 5.a: Delete models, activelearning object to remove stray computational graphs (esp. for EGL) if conf.activelearning_viz: exit() del activelearning del pose_model, aux_net torch.cuda.empty_cache() logging.info('Re-Initializing (and loading) human pose network and auxiliary network.\n') pose_model, aux_net = load_models(conf=conf, load_pose=conf.model['load'], load_aux=conf.model['aux_net']['load'], model_dir=conf.model['load_path']) # 6. Defining Hyperparameters, TensorBoard directory --------------------------------------------------------------- logging.info('Initializing experiment settings.') hyperparameters = define_hyperparams(conf=conf, pose_model=pose_model, aux_net=aux_net) if conf.tensorboard: writer = SummaryWriter(log_dir=os.path.join(conf.model['save_path'], 'tensorboard')) else: writer = None # 7. Train the model if conf.train: train_obj = Train(pose_model=pose_model, aux_net=aux_net, hyperparameters=hyperparameters, dataset_obj=datasets, conf=conf, tb_writer=writer) train_obj.train_model() del train_obj # Reload the best model for metric evaluation conf.resume_training = False pose_model, _ = load_models(conf=conf, load_pose=True, load_aux=False, model_dir=conf.model['save_path']) if conf.metric: metric_obj = Metric(network=pose_model, dataset_obj=datasets, conf=conf) metric_obj.eval() if __name__ == "__main__": main()
41,564
42.342023
168
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/activelearning.py
import os import cv2 import torch import torch.utils.data import numpy as np import logging from tqdm import tqdm from pathlib import Path from sklearn.metrics import pairwise_distances from skimage.feature import peak_local_max from scipy.special import softmax as softmax_fn from scipy.stats import entropy as entropy_fn from matplotlib import pyplot as plt # EGL sampling import autograd_hacks from utils import heatmap_loss from utils import shannon_entropy from utils import heatmap_generator class ActiveLearning(object): """ Contains collection of active learning algorithms for human joint localization """ def __init__(self, conf, pose_net, aux_net): #(self, conf, hg_network, learnloss_network): self.conf = conf self.pose_model = pose_net self.aux_net = aux_net self.num_images = conf.active_learning['num_images'] self.pose_model.eval() if conf.active_learning['algorithm'] in ['learning_loss', 'aleatoric', 'vl4pose']: self.aux_net.eval() def base(self, train, dataset_size): """ :param train: :param dataset_size: :return: """ logging.info('Initializing base dataset.') # Set of indices not annotated unlabelled_idx = train['index'] # Determine if per dataset sampling or overall if self.conf.dataset['load'] == 'mpii': selection = np.random.choice(unlabelled_idx, size=self.num_images, replace=False).astype(np.int32) elif self.conf.dataset['load'] == 'lsp': # First sample from lsp dataset only lsp_images = min(dataset_size['lsp']['train'], self.num_images) lspet_images = self.num_images - lsp_images selection_lsp = np.random.choice(np.arange(dataset_size['lsp']['train']), size=lsp_images, replace=False) selection_lspet = np.random.choice(np.arange(dataset_size['lsp']['train'], dataset_size['lsp']['train'] + dataset_size['lspet']['train']), size=lspet_images, replace=False) selection = np.concatenate([selection_lsp, selection_lspet], axis=0).astype(np.int32) else: # Merged dataset selection = np.random.choice(unlabelled_idx, size=self.num_images, replace=False).astype(np.int32) self._uniquecounts(dataset=train, selection=selection, method='base') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def random(self, train, dataset_size): """ :param train: :param dataset_size: Maintain same method signature across all sampling methods, not used :return: """ logging.info('Performing random sampling.') if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Load previously annotated images indices assert self.conf.model['load'], "Use 'base' to train model from scratch." annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) selection = np.random.choice(unlabelled_idx, size=self.num_images, replace=False) selection = np.concatenate([annotated_idx, selection], axis=0).astype(np.int32) self._uniquecounts(dataset=train, selection=selection, method='random') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def coreset_sampling(self, train, dataset_size): ''' :return: ''' logging.info('Performing Core-Set sampling.') def update_distances(cluster_centers, encoding, min_distances=None): ''' Based on: https://github.com/google/active-learning/blob/master/sampling_methods/kcenter_greedy.py Update min distances given cluster centers. Args: cluster_centers: indices of cluster centers only_new: only calculate distance for newly selected points and update min_distances. rest_dist: whether to reset min_distances. ''' if len(cluster_centers) != 0: # Update min_distances for all examples given new cluster center. x = encoding[cluster_centers] dist = pairwise_distances(encoding, x, metric='euclidean') if min_distances is None: min_distances = np.min(dist, axis=1).reshape(-1, 1) else: min_distances = np.minimum(min_distances, dist) return min_distances if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "Core-set requires a pretrained model." annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert np.all(train['index'] == np.arange(train['name'].shape[0])) dataset_ = ActiveLearningDataset(train, indices=np.arange(train['name'].shape[0]), conf=self.conf) coreset_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) pose_encoding = None # Part 1: Obtain embeddings for all (labelled, unlabelled images) # Disable autograd to speed up inference with torch.no_grad(): for images in tqdm(coreset_dataloader): _, pose_features = self.pose_model(images) try: pose_encoding = torch.cat((pose_encoding, pose_features['penultimate'].cpu()), dim=0) except TypeError: pose_encoding = pose_features['penultimate'].cpu() pose_encoding = pose_encoding.squeeze().numpy() logging.info('Core-Set encodings computed.') # Part 2: k-Centre Greedy core_set_budget = self.num_images min_distances = None assert len(annotated_idx) != 0, "No annotations for previous model found, core-set cannot proceeed." min_distances = update_distances(cluster_centers=annotated_idx, encoding=pose_encoding, min_distances=None) for _ in tqdm(range(core_set_budget)): if len(annotated_idx) == 0: # Initial choice of point # Initialize center with a randomly selected datapoint ind = np.random.choice(np.arange(pose_encoding.shape[0])) else: ind = np.argmax(min_distances) # New examples should not be in already selected since those points # should have min_distance of zero to a cluster center. min_distances = update_distances(cluster_centers=[ind], encoding=pose_encoding, min_distances=min_distances) annotated_idx = np.concatenate([annotated_idx, [ind]], axis=0).astype(np.int32) selection = annotated_idx self._uniquecounts(dataset=train, selection=selection, method='coreset') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def learning_loss_sampling(self, train, dataset_size): """ :param train: :param dataset_size: :param hg_depth: :return: """ logging.info('Performing learning loss sampling.') if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "Learning loss requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) dataset_ = ActiveLearningDataset(train, indices=unlabelled_idx, conf=self.conf) learnloss_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) learnloss_pred = None # Prediction and concatenation of the learning loss network outputs with torch.no_grad(): for images in tqdm(learnloss_dataloader): _, pose_features = self.pose_model(images) learnloss_pred_ = self._aux_net_inference(pose_features) learnloss_pred_ = learnloss_pred_.squeeze() try: learnloss_pred = torch.cat([learnloss_pred, learnloss_pred_.cpu()], dim=0) except TypeError: learnloss_pred = learnloss_pred_.cpu() # argsort defaults to ascending pred_with_index = np.concatenate([learnloss_pred.numpy().reshape(-1, 1), unlabelled_idx.reshape(-1, 1)], axis=-1) pred_with_index = pred_with_index[pred_with_index[:, 0].argsort()] indices = pred_with_index[-self.num_images:, 1] selection = np.concatenate([annotated_idx, indices], axis=0).astype(np.int32) self._uniquecounts(dataset=train, selection=selection, method='learning_loss') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def expected_gradient_length_sampling(self, train, dataset_size): ''' :return: ''' def probability(pair_dist): ''' Computes P(j|i) using Binary Search :param pairwise_dist: (2D Tensor) pairwise distances between samples --> actual dist, not squared :return: 2D Tensor containing conditional probabilities ''' def calc_probs_perp(lower_bound, upper_bound, pair_dist): sigmas = (lower_bound + upper_bound) / 2 variance = (sigmas ** 2).reshape(-1, 1) scaled_pair_dist_neg = -pair_dist / (2 * variance) probs_unnormalized = torch.exp(scaled_pair_dist_neg) probs_unnormalized = torch.clamp(probs_unnormalized, min=1e-20, max=1.) softmax = probs_unnormalized / torch.sum(probs_unnormalized, dim=1, keepdim=True) softmax = torch.clamp(softmax, min=1e-30, max=1.) entropy = shannon_entropy(softmax) perplexity_hat = torch.pow(2 * torch.ones(n_samples), entropy) return perplexity_hat, softmax def condition(perplexity_hat, perplexity): mask = torch.lt(torch.abs(perplexity_hat - perplexity), TOLERANCE) return False in mask global PERPLEXITY, TOLERANCE, n_samples tries = 100 n_samples = pair_dist.shape[0] PERPLEXITY = self.conf.active_learning['egl']['perplexity'] TOLERANCE = self.conf.active_learning['egl']['tolerance'] * torch.ones(n_samples) pair_dist = pair_dist ** 2 lower = torch.zeros(n_samples) upper = (torch.max(torch.max(pair_dist), torch.max(pair_dist**0.5))) * torch.ones(n_samples) * 5 perplexity = PERPLEXITY * torch.ones(n_samples) perplexity_hat, probs = calc_probs_perp(lower, upper, pair_dist) while condition(perplexity_hat, perplexity): if tries < 0: break tries -= 1 mask_gt = torch.gt(perplexity_hat - perplexity, TOLERANCE).type(torch.float32) upper_update = upper - torch.mul(mask_gt, (upper - lower) / 2) mask_lt = torch.lt(perplexity_hat - perplexity, -TOLERANCE).type(torch.float32) lower_update = lower + torch.mul(mask_lt, (upper - lower) / 2) upper = upper_update lower = lower_update perplexity_hat, probs = calc_probs_perp(lower, upper, pair_dist) del PERPLEXITY, TOLERANCE, n_samples return probs logging.info('Performing expected gradient length sampling.') # Setup -------------------------------------------------------------------------------------------------------- if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Load indices of previously annotated data assert self.conf.model['load'], "Expected Gradient Length requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index']) - set(annotated_idx))) # Part 1: Obtain embeddings and heatmaps for LABELLED data ---------------------------------------------------- dataset_ = EGLpp_Dataset(dataset_dict=train, conf=self.conf, indices=annotated_idx) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) logging.info('Computing heatmaps, embedding for labelled images.') # Disable autograd to speed up inference with torch.no_grad(): pose_encoding_L = None pose_heatmap_L = None for images, og_heatmaps in tqdm(egl_dataloader): heatmaps, pose_features = self.pose_model(images) if self.conf.active_learning['egl']['og_heatmap']: heatmaps = torch.stack([og_heatmaps, og_heatmaps], dim=1).to(heatmaps.device) try: pose_encoding_L = torch.cat((pose_encoding_L, pose_features['penultimate'].cpu()), dim=0) # GAP over the 4x4 lyr pose_heatmap_L = torch.cat((pose_heatmap_L, heatmaps.cpu()), dim=0) except TypeError: pose_encoding_L = pose_features['penultimate'].cpu() pose_heatmap_L = heatmaps.cpu() # Part 2: Obtain embeddings ONLY for UNLABELLED data ----------------------------------------------------------- logging.info('Computing embeddings (not heatmaps) for unlabelled data') dataset_ = ActiveLearningDataset(dataset_dict=train, indices=unlabelled_idx, conf=self.conf) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Disable autograd to speed up inference with torch.no_grad(): pose_encoding_U = None for images in tqdm(egl_dataloader): _, pose_features = self.pose_model(images) try: pose_encoding_U = torch.cat((pose_encoding_U, pose_features['penultimate'].cpu()), dim=0) # GAP over the 4x4 lyr except TypeError: pose_encoding_U = pose_features['penultimate'].cpu() # Part 3: Compute the heatmap error between the unlabelled images and its neighbors ---------------------------- with torch.no_grad(): pair_dist = torch.cdist(pose_encoding_U, pose_encoding_L, p=2) # Unlabelled[i] to Labelled[j] p_i_given_j = probability(pair_dist) k = self.conf.active_learning['egl']['k'] assert len(p_i_given_j.shape) == 2, "Not a 2-dimensional tensor" vals, idx = torch.topk(p_i_given_j, k=k, dim=1, sorted=True, largest=True) logging.info('Computing the gradient between the unlabelled and labelled images.') pose_gradients_nbrs = torch.zeros(size=(unlabelled_idx.shape[0], k), dtype=torch.float32).to(vals.device) assert vals.shape == pose_gradients_nbrs.shape autograd_hacks.add_hooks(self.pose_model) dataset_ = ActiveLearningDataset(dataset_dict=train, indices=unlabelled_idx, conf=self.conf) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) i_unlabelled = 0 # Obtain images in batches: for unlabelled_images in tqdm(egl_dataloader): # Iterate over each unlabelled image for i_ in range(unlabelled_images.shape[0]): self.pose_model.zero_grad() i_unlabelled_copies = torch.cat(k * [unlabelled_images[i_].unsqueeze(0)], dim=0)#.cuda() i_heatmaps, _ = self.pose_model(i_unlabelled_copies) loss = heatmap_loss(i_heatmaps, pose_heatmap_L[idx[i_unlabelled]], egl=True).mean() loss.backward() autograd_hacks.compute_grad1(model=self.pose_model, loss_type='mean') with torch.no_grad(): grads = torch.zeros((i_heatmaps.shape[0],), dtype=torch.float32) for param in self.pose_model.parameters(): try: # Sum of squared gradients for each batch element grads = grads.to(param.grad1.device) grads += (param.grad1 ** 2).sum(dim=list(range(len(param.grad1.shape)))[1:]) except AttributeError: continue pose_gradients_nbrs[i_unlabelled] = grads.to(pose_gradients_nbrs.device) # Removing gradients due to previous image self.pose_model.zero_grad() autograd_hacks.clear_backprops(self.pose_model) i_unlabelled += 1 autograd_hacks.remove_hooks(self.pose_model) egl = (vals * pose_gradients_nbrs).sum(dim=1).squeeze() vals, idx = torch.topk(egl, k=self.num_images, sorted=False, largest=True) assert idx.dim() == 1, "'idx' should be a single dimensional array" selection = np.concatenate([annotated_idx, unlabelled_idx[idx.cpu().numpy()]], axis=0).astype(np.int32) self._uniquecounts(dataset=train, selection=selection, method='egl') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def multipeak_entropy(self, train, dataset_size): """ :param train: :param dataset_size: :return: """ logging.info('Performing multi-peak entropy sampling.') if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "Multipeak entropy was called without a pretrained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) unlabelled_idx = np.array(list(set(train['index']) - set(annotated_idx))) # Multi-peak entropy only over the unlabelled set of images dataset_ = ActiveLearningDataset(dataset_dict=train, indices=unlabelled_idx, conf=self.conf) mpe_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) pose_heatmaps = None # Part 1: Obtain set of heatmaps # Disable autograd to speed up inference with torch.no_grad(): for images in tqdm(mpe_dataloader): pose_heatmaps_, _ = self.pose_model(images) try: pose_heatmaps = torch.cat((pose_heatmaps, pose_heatmaps_[:, -1, :, :, :].cpu()), dim=0) except TypeError: pose_heatmaps = pose_heatmaps_[:, -1, :, :, :].cpu() pose_heatmaps = pose_heatmaps.squeeze().numpy() logging.info('Heatmaps computed. Calculating multi-peak entropy') # Part 2: Multi-peak entropy mpe_budget = self.num_images mpe_value_per_img = np.zeros(pose_heatmaps.shape[0], dtype=np.float32) # e.g. shape of heatmap final is BS x 14 x 64 x 64 for i in tqdm(range(pose_heatmaps.shape[0])): normalizer = 0 entropy = 0 for hm in range(pose_heatmaps.shape[1]): loc = peak_local_max(pose_heatmaps[i, hm], min_distance=5, num_peaks=5) peaks = pose_heatmaps[i, hm][loc[:, 0], loc[:, 1]] if peaks.shape[0] > 0: normalizer += 1 peaks = softmax_fn(peaks) entropy += entropy_fn(peaks) mpe_value_per_img[i] = entropy mpe_value_per_img = torch.from_numpy(mpe_value_per_img) vals, idx = torch.topk(mpe_value_per_img, k=mpe_budget, sorted=False, largest=True) assert idx.dim() == 1, "'idx' should be a single dimensional array" annotated_idx = np.concatenate([annotated_idx, unlabelled_idx[idx.numpy()]], axis=0).astype(np.int32) selection = annotated_idx self._uniquecounts(dataset=train, selection=selection, method='multipeak_entropy') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def aleatoric_uncertainty(self, train, dataset_size): """ :param train: :param dataset_size: :param hg_depth: :return: """ logging.info('Performing Uncertainty: Kendall and Gal sampling.') if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "Aleatoric uncertainty requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) dataset_ = ActiveLearningDataset(train, indices=unlabelled_idx, conf=self.conf) aleatoric_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) aleatoric_pred = None # Prediction and concatenation of the aleatoric predictions with torch.no_grad(): for images in tqdm(aleatoric_dataloader): _, pose_features = self.pose_model(images) aleatoric_pred_ = self._aux_net_inference(pose_features) aleatoric_pred_ = aleatoric_pred_.squeeze() try: aleatoric_pred = torch.cat([aleatoric_pred, aleatoric_pred_.cpu()], dim=0) except TypeError: aleatoric_pred = aleatoric_pred_.cpu() aleatoric_pred = aleatoric_pred.mean(dim=-1) # argsort defaults to ascending pred_with_index = np.concatenate([aleatoric_pred.numpy().reshape(-1, 1), unlabelled_idx.reshape(-1, 1)], axis=-1) pred_with_index = pred_with_index[pred_with_index[:, 0].argsort()] indices = pred_with_index[-self.num_images:, 1] selection = np.concatenate([annotated_idx, indices], axis=0).astype(np.int32) self._uniquecounts(dataset=train, selection=selection, method='aleatoric') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def vl4pose(self, train, dataset_size): """ :param train: :param dataset_size: :param hg_depth: :return: """ logging.info('Performing VL4Pose sampling.') if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "VL4Pose requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) j2i = {'head': 0, 'neck': 1, 'lsho': 2, 'lelb': 3, 'lwri': 4, 'rsho': 5, 'relb': 6, 'rwri': 7, 'lhip': 8, 'lknee': 9, 'lankl': 10, 'rhip': 11, 'rknee': 12, 'rankl': 13} i2j = {0: 'head', 1: 'neck', 2: 'lsho', 3: 'lelb', 4: 'lwri', 5: 'rsho', 6: 'relb', 7: 'rwri', 8: 'lhip', 9: 'lknee', 10: 'lankl', 11: 'rhip', 12: 'rknee', 13: 'rankl'} if self.conf.dataset['load'] == 'mpii' or self.conf.dataset['load'] == 'merged': j2i['pelvis'] = 14 j2i['thorax'] = 15 i2j[14] = 'pelvis' i2j[15] = 'thorax' if self.conf.dataset['load'] == 'mpii': links = [[j2i['head'], j2i['neck']], [j2i['neck'], j2i['thorax']], [j2i['thorax'], j2i['pelvis']], [j2i['thorax'], j2i['lsho']], [j2i['lsho'], j2i['lelb']], [j2i['lelb'], j2i['lwri']], [j2i['thorax'], j2i['rsho']], [j2i['rsho'], j2i['relb']], [j2i['relb'], j2i['rwri']], [j2i['pelvis'], j2i['lhip']], [j2i['lhip'], j2i['lknee']], [j2i['lknee'], j2i['lankl']], [j2i['pelvis'], j2i['rhip']], [j2i['rhip'], j2i['rknee']], [j2i['rknee'], j2i['rankl']]] else: links = [[j2i['head'], j2i['neck']], [j2i['neck'], j2i['lsho']], [j2i['lsho'], j2i['lelb']], [j2i['lelb'], j2i['lwri']], [j2i['neck'], j2i['rsho']], [j2i['rsho'], j2i['relb']], [j2i['relb'], j2i['rwri']], [j2i['lsho'], j2i['lhip']], [j2i['lhip'], j2i['lknee']], [j2i['lknee'], j2i['lankl']], [j2i['rsho'], j2i['rhip']], [j2i['rhip'], j2i['rknee']], [j2i['rknee'], j2i['rankl']]] dataset_ = ActiveLearningDataset(train, indices=unlabelled_idx, conf=self.conf) vl4pose_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) pose_heatmaps = None likelihood_params = None # Part 1: Obtain set of heatmaps # Disable autograd to speed up inference with torch.no_grad(): for images in tqdm(vl4pose_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) likelihood_pred_ = self._aux_net_inference(pose_features_) try: pose_heatmaps = torch.cat((pose_heatmaps, pose_heatmaps_[:, -1, :, :, :].cpu()), dim=0) likelihood_params = torch.cat([likelihood_params, likelihood_pred_.cpu().reshape(images.shape[0], len(links), 2)], dim=0) except TypeError: pose_heatmaps = pose_heatmaps_[:, -1, :, :, :].cpu() likelihood_params = likelihood_pred_.cpu().reshape(images.shape[0], len(links), 2) pose_heatmaps = pose_heatmaps.squeeze().numpy() likelihood_params = likelihood_params.numpy() del vl4pose_dataloader logging.info('Heatmaps computed. Calculating likelihood of pose.') keypoint_compute = Keypoint_ParallelWrapper(hm=pose_heatmaps, param=likelihood_params, j2i=j2i, i2j=i2j, links=links, vl4pose_config=self.conf.active_learning['vl4pose']) vl4pose_dataloader = torch.utils.data.DataLoader(keypoint_compute, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) max_likelihood = np.zeros(shape=pose_heatmaps.shape[0]) ptr = 0 for likelihoods, trace in tqdm(vl4pose_dataloader): max_likelihood[ptr: ptr + likelihoods.shape[0]] = likelihoods.squeeze().numpy() ptr += likelihoods.shape[0] loglikelihood_with_index = np.concatenate([max_likelihood.reshape(-1, 1), unlabelled_idx.reshape(-1, 1)], axis=-1) loglikelihood_with_index = loglikelihood_with_index[loglikelihood_with_index[:, 0].argsort()] # Select the images with the lowest likelihood indices = loglikelihood_with_index[:self.num_images, 1] selection = np.concatenate([annotated_idx, indices], axis=0).astype(np.int32) self._uniquecounts(dataset=train, selection=selection, method='vl4pose') np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/annotation.npy'), arr=selection) return selection def _uniquecounts(self, dataset, selection, method): """ :param dataset: :param selection: :return: """ # ['dataset'] is the dataset name such as mpii, lsp, lspet for an image unique, counts = np.unique(dataset['dataset'][selection], return_counts=True) proportion = {key: value for (key, value) in zip(unique, counts)} with open(os.path.join(self.conf.model['save_path'], 'model_checkpoints/sampling_proportion.txt'), "x") as file: file.write('{} Sampling\n'.format(method)) [file.write("{}: {}\n".format(key, proportion[key])) for key in proportion.keys()] def _aux_net_inference(self, pose_features): extractor = self.conf.architecture['aux_net']['conv_or_avg_pooling'] with torch.no_grad(): if extractor == 'avg': # Transfer to GPU where auxiliary network is stored encodings = pose_features['penultimate'] else: depth = len(self.conf.architecture['aux_net']['spatial_dim']) encodings = torch.cat( [pose_features['feature_{}'.format(i)].reshape( pose_features['feature_{}'.format(i)].shape[0], pose_features['feature_{}'.format(i)].shape[1], -1) for i in range(depth, 0, -1)], dim=2) aux_out = self.aux_net(encodings) return aux_out class ActiveLearningDataset(torch.utils.data.Dataset): def __init__(self, dataset_dict, indices, conf): ''' :param dataset_dict: ''' self.names = dataset_dict['name'][indices] self.bounding_box = dataset_dict['bbox_coords'][indices] self.dataset = dataset_dict['dataset'][indices] self.xy_to_uv = lambda xy: (xy[1], xy[0]) def __len__(self): return self.dataset.shape[0] def __getitem__(self, item): ''' :param item: :return: ''' root = Path(os.getcwd()).parent mpii_path = os.path.join(root, 'data', 'mpii') lsp_path = os.path.join(root, 'data', 'lsp') lspet_path = os.path.join(root, 'data', 'lspet') bounding_box = self.bounding_box[item] dataset = self.dataset[item] name = self.names[item] if dataset == 'mpii': image = plt.imread(os.path.join(mpii_path, 'images', '{}.jpg'.format(name.split('_')[0]))) elif dataset == 'lsp': image = plt.imread(os.path.join(lsp_path, 'images', name)) else: image = plt.imread(os.path.join(lspet_path, 'images', name)) # Determine crop img_shape = np.array(image.shape) # Bounding box for the first person [min_x, min_y, max_x, max_y] = bounding_box[0] tl_uv = self.xy_to_uv(np.array([min_x, min_y])) br_uv = self.xy_to_uv(np.array([max_x, max_y])) min_u = tl_uv[0] min_v = tl_uv[1] max_u = br_uv[0] max_v = br_uv[1] centre = np.array([(min_u + max_u) / 2, (min_v + max_v) / 2]) height = max_u - min_u width = max_v - min_v scale = 1.75 top_left = np.array([centre[0] - (scale * height / 2), centre[1] - (scale * width / 2)]) bottom_right = np.array([centre[0] + (scale * height / 2), centre[1] + (scale * width / 2)]) top_left = np.maximum(np.array([0, 0], dtype=np.int16), top_left.astype(np.int16)) bottom_right = np.minimum(img_shape.astype(np.int16)[:-1], bottom_right.astype(np.int16)) # Cropping the image image = image[top_left[0]: bottom_right[0], top_left[1]: bottom_right[1], :] # Resize the image image = self.resize_image(image, target_size=[256, 256, 3]) return torch.tensor(data=image / 256.0, dtype=torch.float32, device='cpu') def resize_image(self, image_=None, target_size=None): ''' :return: ''' # Compute the aspect ratios image_aspect_ratio = image_.shape[0] / image_.shape[1] tgt_aspect_ratio = target_size[0] / target_size[1] # Compare the original and target aspect ratio if image_aspect_ratio > tgt_aspect_ratio: # If target aspect ratio is smaller, scale the first dim scale_factor = target_size[0] / image_.shape[0] else: # If target aspect ratio is bigger or equal, scale the second dim scale_factor = target_size[1] / image_.shape[1] # Compute the padding to fit the target size pad_u = (target_size[0] - int(image_.shape[0] * scale_factor)) pad_v = (target_size[1] - int(image_.shape[1] * scale_factor)) output_img = np.zeros(target_size, dtype=image_.dtype) # Write scaled size in reverse order because opencv resize scaled_size = (int(image_.shape[1] * scale_factor), int(image_.shape[0] * scale_factor)) padding_u = int(pad_u / 2) padding_v = int(pad_v / 2) im_scaled = cv2.resize(image_, scaled_size) # logging.debug('Scaled, pre-padding size: {}'.format(im_scaled.shape)) output_img[padding_u : im_scaled.shape[0] + padding_u, padding_v : im_scaled.shape[1] + padding_v, :] = im_scaled return output_img class EGLpp_Dataset(torch.utils.data.Dataset): def __init__(self, dataset_dict, conf, indices=None): """ :param dataset_dict: :param conf: :param indices: """ self.names = dataset_dict['name'][indices] self.bounding_box = dataset_dict['bbox_coords'][indices] self.dataset = dataset_dict['dataset'][indices] self.gt = dataset_dict['gt'][indices] self.occlusion = conf.experiment_settings['occlusion'] self.hm_shape = [64, 64] self.hm_peak = conf.experiment_settings['hm_peak'] self.xy_to_uv = lambda xy: (xy[1], xy[0]) def __len__(self): return self.dataset.shape[0] def __getitem__(self, item): ''' :param item: :return: ''' root = Path(os.getcwd()).parent mpii_path = os.path.join(root, 'data', 'mpii') lsp_path = os.path.join(root, 'data', 'lsp') lspet_path = os.path.join(root, 'data', 'lspet') dataset = self.dataset[item] name = self.names[item] if dataset == 'mpii': image = plt.imread(os.path.join(mpii_path, 'images', '{}.jpg'.format(name.split('_')[0]))) elif dataset == 'lsp': image = plt.imread(os.path.join(lsp_path, 'images', name)) else: image = plt.imread(os.path.join(lspet_path, 'images', name)) bounding_box = self.bounding_box[item] gt = self.gt[item] # Determine crop img_shape = np.array(image.shape) # Bounding box for the first person [min_x, min_y, max_x, max_y] = bounding_box[0] tl_uv = self.xy_to_uv(np.array([min_x, min_y])) br_uv = self.xy_to_uv(np.array([max_x, max_y])) min_u = tl_uv[0] min_v = tl_uv[1] max_u = br_uv[0] max_v = br_uv[1] centre = np.array([(min_u + max_u) / 2, (min_v + max_v) / 2]) height = max_u - min_u width = max_v - min_v scale = 2.0 window = max(scale * height, scale * width) top_left = np.array([centre[0] - (window / 2), centre[1] - (window / 2)]) bottom_right = np.array([centre[0] + (window / 2), centre[1] + (window / 2)]) top_left = np.maximum(np.array([0, 0], dtype=np.int16), top_left.astype(np.int16)) bottom_right = np.minimum(img_shape.astype(np.int16)[:-1], bottom_right.astype(np.int16)) # Cropping the image and adjusting the ground truth image = image[top_left[0]: bottom_right[0], top_left[1]: bottom_right[1], :] for person in range(gt.shape[0]): for joint in range(gt.shape[1]): gt_uv = self.xy_to_uv(gt[person][joint]) gt_uv = gt_uv - top_left gt[person][joint] = np.concatenate([gt_uv, np.array([gt[person][joint][2]])], axis=0) # Resize the image image, gt = self.resize_image(image, gt, target_size=[256, 256, 3]) heatmaps, joint_exist = heatmap_generator( joints=np.copy(gt), occlusion=self.occlusion, hm_shape=self.hm_shape, img_shape=image.shape) heatmaps = self.hm_peak * heatmaps return torch.tensor(data=image / 256.0, dtype=torch.float32, device='cpu'),\ torch.tensor(data=heatmaps, dtype=torch.float32, device='cpu') def resize_image(self, image_=None, gt=None, target_size=None): ''' :return: ''' # Compute the aspect ratios image_aspect_ratio = image_.shape[0] / image_.shape[1] tgt_aspect_ratio = target_size[0] / target_size[1] # Compare the original and target aspect ratio if image_aspect_ratio > tgt_aspect_ratio: # If target aspect ratio is smaller, scale the first dim scale_factor = target_size[0] / image_.shape[0] else: # If target aspect ratio is bigger or equal, scale the second dim scale_factor = target_size[1] / image_.shape[1] # Compute the padding to fit the target size pad_u = (target_size[0] - int(image_.shape[0] * scale_factor)) pad_v = (target_size[1] - int(image_.shape[1] * scale_factor)) output_img = np.zeros(target_size, dtype=image_.dtype) # Write scaled size in reverse order because opencv resize scaled_size = (int(image_.shape[1] * scale_factor), int(image_.shape[0] * scale_factor)) padding_u = int(pad_u / 2) padding_v = int(pad_v / 2) im_scaled = cv2.resize(image_, scaled_size) # logging.debug('Scaled, pre-padding size: {}'.format(im_scaled.shape)) output_img[padding_u : im_scaled.shape[0] + padding_u, padding_v : im_scaled.shape[1] + padding_v, :] = im_scaled gt *= np.array([scale_factor, scale_factor, 1]).reshape(1, 1, 3) gt[:, :, 0] += padding_u gt[:, :, 1] += padding_v return output_img, gt class Keypoint_ParallelWrapper(torch.utils.data.Dataset): def __init__(self, hm, param, j2i, i2j, links, vl4pose_config): self.hm = hm self.param = param self.j2i = j2i self.i2j = i2j self.links = links self.config = vl4pose_config def __len__(self): return self.hm.shape[0] def __getitem__(self, i): joints = {} heatmaps = self.hm[i] parameters = self.param[i] # Initialize keypoints for each node for key in self.j2i.keys(): heatmap = heatmaps[self.j2i[key]] loc = peak_local_max(heatmap, min_distance=self.config['min_distance'], num_peaks=self.config['num_peaks']) peaks = heatmap[loc[:, 0], loc[:, 1]] peaks = softmax_fn(peaks) joints[key] = Keypoint(name=key, loc=loc, peaks=peaks) # Initialize parent-child relations for k, l in enumerate(self.links): joints[self.i2j[l[0]]].parameters.append(parameters[k]) joints[self.i2j[l[0]]].children.append(joints[self.i2j[l[1]]]) max_ll, trace = joints['head'].run_likelihood() return max_ll, trace class Keypoint(object): def __init__(self, name, loc, peaks): self.name = name self.loc = loc self.peaks = peaks self.children = [] self.parameters = [] def run_likelihood(self): """ :return: """ assert self.name == 'head' likelihood_per_location = [] per_location_trace = [] for location in range(self.loc.shape[0]): log_ll = np.log(self.peaks[location]) per_child_trace = [] for child in range(len(self.children)): child_ll, joint_trace = self.children[child].compute_likelihood_given_parent(self.loc[location], self.parameters[child]) log_ll += child_ll per_child_trace.append(joint_trace) likelihood_per_location.append(log_ll) per_location_trace.append(per_child_trace) likelihood_per_location = np.array(likelihood_per_location) return_trace = {} for child_trace in per_location_trace[np.argmax(likelihood_per_location)]: return_trace.update(child_trace) return_trace[self.name] = np.argmax(likelihood_per_location) return_trace['{}_uv'.format(self.name)] = self.loc[np.argmax(likelihood_per_location)] return np.sum(likelihood_per_location), return_trace def compute_likelihood_given_parent(self, parent_location, gaussian_params): """ :param parent_location: :param gaussian_params: :return: """ likelihood_per_location = [] per_location_trace = [] for location in range(self.loc.shape[0]): log_ll = np.log(2 * np.pi) + gaussian_params[1] log_ll += (gaussian_params[0] - np.linalg.norm(parent_location - self.loc[location]))**2 * np.exp(-gaussian_params[1]) log_ll *= -0.5 log_ll += np.log(self.peaks[location]) if len(self.children) == 0: likelihood_per_location.append(log_ll) else: per_child_trace = [] for child in range(len(self.children)): child_ll, joint_trace = self.children[child].compute_likelihood_given_parent(self.loc[location], self.parameters[child]) log_ll += child_ll per_child_trace.append(joint_trace) likelihood_per_location.append(log_ll) per_location_trace.append(per_child_trace) likelihood_per_location = np.array(likelihood_per_location) if len(self.children) == 0: return np.sum(likelihood_per_location), {self.name: np.argmax(likelihood_per_location), '{}_uv'.format(self.name): self.loc[np.argmax(likelihood_per_location)]} return_trace = {} for child_trace in per_location_trace[np.argmax(likelihood_per_location)]: return_trace.update(child_trace) return_trace[self.name] = np.argmax(likelihood_per_location) return_trace['{}_uv'.format(self.name)] = self.loc[np.argmax(likelihood_per_location)] return np.sum(likelihood_per_location), return_trace
43,791
39.774674
160
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/dataloader.py
import os import copy import logging from pathlib import Path import cv2 import scipy.io import numpy as np from tqdm import tqdm from matplotlib import pyplot as plt from matplotlib.patches import Circle import torch import torch.utils.data import albumentations as albu from utils import heatmap_generator from utils import uv_from_heatmap jnt_to_ind = {'head': 0, 'neck': 1, 'lsho': 2, 'lelb': 3, 'lwri': 4, 'rsho': 5, 'relb': 6, 'rwri': 7, 'lhip': 8, 'lknee': 9, 'lankl': 10, 'rhip': 11, 'rknee': 12, 'rankl': 13} ind_to_jnt = {0: 'head', 1: 'neck', 2: 'lsho', 3: 'lelb', 4: 'lwri', 5: 'rsho', 6: 'relb', 7: 'rwri', 8: 'lhip', 9: 'lknee', 10: 'lankl', 11: 'rhip', 12: 'rknee', 13: 'rankl'} def load_mpii(mpii_conf, del_extra_jnts): """ Converts Matlab structure .mat file into a more intuitive dictionary object. :param mpii_conf: :return: """ # Lambda_head brings the head keypoint annotation closer to the neck lambda_head = mpii_conf['lambda_head'] precached_mpii = mpii_conf['precached'] mpii_only = False if del_extra_jnts else True root = Path(os.getcwd()).parent dataset_path = os.path.join(root, 'data', 'mpii') # Adds unique joints corresponding to MPII if not del_extra_jnts: global jnt_to_ind, ind_to_jnt jnt_to_ind['pelvis'] = 14 jnt_to_ind['thorax'] = 15 ind_to_jnt[14] = 'pelvis' ind_to_jnt[15] = 'thorax' # Load MPII images directly into memory if precached_mpii: logging.info('Loading precached MPII.') if mpii_only: string = '16jnts' else: string = '14jnts' img_dict = np.load(os.path.join(root, 'cached', 'mpii_cache_{}.npy'.format(string)), allow_pickle=True) img_dict = img_dict[()] try: assert img_dict['mpii']['del_extra_jnts'] == del_extra_jnts assert img_dict['mpii']['lambda_head'] == lambda_head # Why am I deleting this? :'( ANSWER - TO MAINTAIN COMPATIBILITY IF DATASETS ARE MERGED del img_dict['mpii']['del_extra_jnts'] del img_dict['mpii']['lambda_head'] return img_dict except AssertionError: logging.warning('Cannot load MPII due to different configurations.') logging.warning('Loading MPII from scratch.\n') mpii_idx_to_jnt = {0: 'rankl', 1: 'rknee', 2: 'rhip', 5: 'lankl', 4: 'lknee', 3: 'lhip', 6: 'pelvis', 7: 'thorax', 8: 'neck', 11: 'relb', 10: 'rwri', 9: 'head', 12: 'rsho', 13: 'lsho', 14: 'lelb', 15: 'lwri'} max_person_in_img = 0 # Create a template for GT and Pred to follow mpii_template = dict([(mpii_idx_to_jnt[i], []) for i in range(16)]) img_dict = {'mpii': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'normalizer': [], 'dataset': [], 'num_gt': [], 'split': [], 'scale': [], 'objpos': [], 'num_ppl': []}} # Load MPII matlab_mpii = scipy.io.loadmat(os.path.join(dataset_path, 'joints.mat'), struct_as_record=False)['RELEASE'][0, 0] # Iterate over all images # matlab_mpii.__dict__['annolist'][0].shape[0] for img_idx in tqdm(range(matlab_mpii.__dict__['annolist'][0].shape[0])): # Load annotation data per image annotation_mpii = matlab_mpii.__dict__['annolist'][0, img_idx] train_test_mpii = matlab_mpii.__dict__['img_train'][0, img_idx].flatten()[0] person_id = matlab_mpii.__dict__['single_person'][img_idx][0].flatten() num_people = len(person_id) max_person_in_img = max(max_person_in_img, len(person_id)) # Read image img_name = annotation_mpii.__dict__['image'][0, 0].__dict__['name'][0] try: image = plt.imread(os.path.join(dataset_path, 'images', img_name)) except FileNotFoundError: logging.warning('Could not load filename: {}'.format(img_name)) continue # Create a deepcopy of the template to avoid overwriting the original gt_per_image = copy.deepcopy(mpii_template) num_joints_persons = [] normalizer_persons = [] scale = [] objpos = [] # Default is that there are no annotated people in the image annotated_person_flag = False # Iterate over each person for person in (person_id - 1): try: per_person_jnts = [] # If annopoints not present, then annotations for that person absent. Throw exception and skip to next annopoints_img_mpii = annotation_mpii.__dict__['annorect'][0, person].__dict__['annopoints'][0, 0] scale_img_mpii = annotation_mpii.__dict__['annorect'][0, person].__dict__['scale'][0][0] objpose_img_mpii = annotation_mpii.__dict__['annorect'][0, person].__dict__['objpos'][0][0] objpose_img_mpii = [objpose_img_mpii.__dict__['x'][0][0], objpose_img_mpii.__dict__['y'][0][0]] num_joints = annopoints_img_mpii.__dict__['point'][0].shape[0] remove_pelvis_thorax_from_num_joints = 0 # PCKh@0.x: Head bounding box normalizer head_x1 = annotation_mpii.__dict__['annorect'][0, person].__dict__['x1'][0][0] head_y1 = annotation_mpii.__dict__['annorect'][0, person].__dict__['y1'][0][0] head_x2 = annotation_mpii.__dict__['annorect'][0, person].__dict__['x2'][0][0] head_y2 = annotation_mpii.__dict__['annorect'][0, person].__dict__['y2'][0][0] xy_1 = np.array([head_x1, head_y1], dtype=np.float32) xy_2 = np.array([head_x2, head_y2], dtype=np.float32) normalizer_persons.append(np.linalg.norm(xy_1 - xy_2, ord=2)) # If both are true, pulls the head joint closer to the neck, and body head_jt, neck_jt = False, False # MPII does not have a [-1, -1] or absent GT, hence the number of gt differ for each image for i in range(num_joints): x = annopoints_img_mpii.__dict__['point'][0, i].__dict__['x'].flatten()[0] y = annopoints_img_mpii.__dict__['point'][0, i].__dict__['y'].flatten()[0] id_ = annopoints_img_mpii.__dict__['point'][0, i].__dict__['id'][0][0] vis = annopoints_img_mpii.__dict__['point'][0, i].__dict__['is_visible'].flatten() # No entry corresponding to visible, mostly head vis is missing. if vis.size == 0: vis = 1 else: vis = vis.item() if id_ == 9: head_jt = True if id_ == 8: neck_jt = True if ((id_ == 6) or (id_ == 7)) and del_extra_jnts: remove_pelvis_thorax_from_num_joints += 1 # Arrange ground truth in form {jnt: [[person1], [person2]]} gt_per_joint = np.array([x, y, vis]).astype(np.float16) gt_per_image[mpii_idx_to_jnt[id_]].append(gt_per_joint) per_person_jnts.append(mpii_idx_to_jnt[id_]) # If person 1 does not have rankl and person 2 has rankl, then prevent rankl being associated with p1 # If jnt absent in person, then we append np.array([-1, -1, -1]) all_jnts = set(list(mpii_idx_to_jnt.values())) per_person_jnts = set(per_person_jnts) jnt_absent_person = all_jnts - per_person_jnts for abs_joint in jnt_absent_person: gt_per_image[abs_joint].append(np.array([-1, -1, -1])) num_joints_persons.append(num_joints - remove_pelvis_thorax_from_num_joints) scale.append(scale_img_mpii) objpos.append(objpose_img_mpii) # If both head and neck joint present, then move the head joint linearly towards the neck joint. if head_jt and neck_jt: gt_per_image['head'][-1] = (lambda_head * gt_per_image['head'][-1])\ + ((1 - lambda_head) * gt_per_image['neck'][-1]) # Since annotation for atleast on person in image present, this flag will add GT to the dataset annotated_person_flag = True except KeyError: # Person 'x' could not have annotated joints, hence move to person 'y' continue if not annotated_person_flag: continue # Maintain compatibility with MPII and LSPET if del_extra_jnts: del gt_per_image['pelvis'] del gt_per_image['thorax'] # Add image, name, pred placeholder and gt img_dict['mpii']['img_name'].append(img_name) img_dict['mpii']['img_pred'].append(mpii_template.copy()) img_dict['mpii']['img_gt'].append(gt_per_image) img_dict['mpii']['normalizer'].append(normalizer_persons) img_dict['mpii']['dataset'].append('mpii') img_dict['mpii']['num_gt'].append(num_joints_persons) img_dict['mpii']['split'].append(train_test_mpii) img_dict['mpii']['scale'].append(scale) img_dict['mpii']['objpos'].append(objpos) img_dict['mpii']['num_ppl'].append(num_people) img_dict['mpii']['del_extra_jnts'] = del_extra_jnts img_dict['mpii']['lambda_head'] = lambda_head img_dict['mpii']['max_person_in_img'] = max_person_in_img if mpii_only: string = '16jnts' else: string = '14jnts' np.save(file=os.path.join(root, 'cached', 'mpii_cache_{}.npy'.format(string)), arr=img_dict, allow_pickle=True) del img_dict['mpii']['del_extra_jnts'] del img_dict['mpii']['lambda_head'] return img_dict def load_lsp(lsp_conf, model_conf):#shuffle=False, train_ratio=0.7, conf=None): """ :param lsp_conf: :param model_conf: :return: """ root = Path(os.getcwd()).parent dataset_path = os.path.join(root, 'data', 'lsp') with open(os.path.join(dataset_path, 'lsp_filenames.txt'), 'r') as f: filenames = f.read().split() lsp_idx_to_jnt = {0: 'rankl', 1: 'rknee', 2: 'rhip', 5: 'lankl', 4: 'lknee', 3: 'lhip', 6: 'rwri', 7: 'relb', 8: 'rsho', 11: 'lwri', 10: 'lelb', 9: 'lsho', 12: 'neck', 13: 'head'} lsp_template = dict([(lsp_idx_to_jnt[i], []) for i in range(14)]) img_dict = {'lsp': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'normalizer': [], 'dataset': [], 'num_gt': [], 'split': []}} annotation_lsp = scipy.io.loadmat(os.path.join(dataset_path, 'joints.mat'))['joints'] # Shape: 3,14,2000 # 0: Train; 1: Validate # Load Train/Test split if conf.model_load_hg == True if model_conf['load']: train_test_split = np.load(os.path.join(model_conf['load_path'], 'model_checkpoints/lsp_split.npy')) else: train_test_split = np.concatenate( [np.zeros((int(annotation_lsp.shape[2]*lsp_conf['train_ratio']),), dtype=np.int8), np.ones((annotation_lsp.shape[2] - int(annotation_lsp.shape[2]*lsp_conf['train_ratio']),), dtype=np.int8)], axis=0) if lsp_conf['shuffle']: logging.info('Shuffling LSP') np.random.shuffle(train_test_split) np.save(file=os.path.join(model_conf['save_path'], 'model_checkpoints', 'lsp_split.npy'), arr=train_test_split) for index in tqdm(range(annotation_lsp.shape[2])): image = plt.imread(os.path.join(dataset_path, 'images', filenames[index])) # Broadcasting rules apply: Toggle visibility of ground truth gt = abs(np.array([[0], [0], [1]]) - annotation_lsp[:, :, index]) gt_dict = dict([(lsp_idx_to_jnt[i], [gt[:, i]]) for i in range(gt.shape[1])]) num_gt = sum([1 for i in range(gt.shape[1]) if gt[:, i][2]]) # PCK@0.x : Normalizer lsho = gt[:2, 9] rsho = gt[:2, 8] lhip = gt[:2, 3] rhip = gt[:2, 2] torso_1 = np.linalg.norm(lsho - rhip) torso_2 = np.linalg.norm(rsho - lhip) torso = max(torso_1, torso_2) img_dict['lsp']['img_name'].append(filenames[index]) img_dict['lsp']['img_pred'].append(copy.deepcopy(lsp_template)) img_dict['lsp']['img_gt'].append(gt_dict) img_dict['lsp']['normalizer'].append([torso]) img_dict['lsp']['dataset'].append('lsp') img_dict['lsp']['num_gt'].append([num_gt]) img_dict['lsp']['split'].append(train_test_split[index]) return img_dict def load_lspet(lspet_conf, model_conf): """ :param lspet_conf: :param model_conf: :return: """ root = Path(os.getcwd()).parent dataset_path = os.path.join(root, 'data', 'lspet') with open(os.path.join(dataset_path, 'lspet_filenames.txt'), 'r') as f: filenames = f.read().split() lspet_idx_to_jnt = {0: 'rankl', 1: 'rknee', 2: 'rhip', 5: 'lankl', 4: 'lknee', 3: 'lhip', 6: 'rwri', 7: 'relb', 8: 'rsho', 11: 'lwri', 10: 'lelb', 9: 'lsho', 12: 'neck', 13: 'head'} lspet_template = dict([(lspet_idx_to_jnt[i], []) for i in range(14)]) img_dict = {'lspet': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'normalizer': [], 'dataset': [], 'num_gt': [], 'split': []}} annotation_lspet = scipy.io.loadmat(os.path.join(dataset_path, 'joints.mat'))['joints'] # Shape: 14,3,10000 # 0: Train; 1: Validate # Load Train/Test split if conf.model_load_hg == True if model_conf['load']: train_test_split = np.load(os.path.join(model_conf['load_path'], 'model_checkpoints/lspet_split.npy')) else: train_test_split = np.concatenate( [np.zeros((int(annotation_lspet.shape[2] * lspet_conf['train_ratio']),), dtype=np.int8), np.ones((annotation_lspet.shape[2] - int(annotation_lspet.shape[2] * lspet_conf['train_ratio']),), dtype=np.int8)], axis=0) if lspet_conf['shuffle']: logging.info('Shuffling LSPET') np.random.shuffle(train_test_split) np.save(file=os.path.join(model_conf['save_path'], 'model_checkpoints', 'lspet_split.npy'), arr=train_test_split) for index in tqdm(range(annotation_lspet.shape[2])): image = plt.imread(os.path.join(dataset_path, 'images', filenames[index])) gt = annotation_lspet[:, :, index] gt_dict = dict([(lspet_idx_to_jnt[i], [gt[i]]) for i in range(gt.shape[0])]) num_gt = sum([1 for i in range(gt.shape[0]) if gt[i][2]]) # PCK@0.x : Normalizer lsho = gt[9, :2] rsho = gt[8, :2] lhip = gt[3, :2] rhip = gt[2, :2] torso_1 = np.linalg.norm(lsho - rhip) torso_2 = np.linalg.norm(rsho - lhip) torso = max(torso_1, torso_2) img_dict['lspet']['img_name'].append(filenames[index]) img_dict['lspet']['img_pred'].append(copy.deepcopy(lspet_template)) img_dict['lspet']['img_gt'].append(gt_dict) img_dict['lspet']['normalizer'].append([torso]) img_dict['lspet']['dataset'].append('lspet') img_dict['lspet']['num_gt'].append([num_gt]) img_dict['lspet']['split'].append(train_test_split[index]) return img_dict def load_hp_dataset(dataset_conf, model_conf): """ :param dataset_conf: :return: """ dataset_dict = dict() if dataset_conf['load'] == 'mpii' or dataset_conf['load'] == 'merged': logging.info('Loading MPII dataset') del_extra_jnts = False if dataset_conf['load'] == 'mpii' else True dataset_dict.update(load_mpii(mpii_conf=dataset_conf['mpii_params'], del_extra_jnts=del_extra_jnts)) if dataset_conf['load'] == 'lsp' or dataset_conf['load'] == 'merged': logging.info('Loading LSP dataset') dataset_dict.update(load_lsp(lsp_conf=dataset_conf['lsp_params'], model_conf=model_conf)) logging.info('Loading LSPET dataset') dataset_dict.update(load_lspet(lspet_conf=dataset_conf['lspet_params'], model_conf=model_conf)) return dataset_dict class HumanPoseDataLoader(torch.utils.data.Dataset): def __init__(self, dataset_dict, activelearning, conf): #(self, mpii_dict=None, lspet_dict=None, lsp_dict=None, activelearning_obj=None, #getitem_dump=None, conf=None, **kwargs): """ :param dataset_dict: :param activelearning: :param conf: """ self.conf = conf self.viz = conf.viz self.occlusion = conf.experiment_settings['occlusion'] self.hm_shape = [64, 64] # Hardcoded self.hm_peak = conf.experiment_settings['hm_peak'] self.threshold = conf.experiment_settings['threshold'] * self.hm_peak self.model_save_path = conf.model['save_path'] self.jnt_to_ind = jnt_to_ind self.ind_to_jnt = ind_to_jnt # Training specific attributes: self.train_flag = False self.validate_flag = False self.model_input_dataset = None # Define active learning functions activelearning_samplers = { 'base': activelearning.base, 'random': activelearning.random, 'coreset': activelearning.coreset_sampling, 'learning_loss': activelearning.learning_loss_sampling, 'egl': activelearning.expected_gradient_length_sampling, 'multipeak_entropy': activelearning.multipeak_entropy, 'aleatoric': activelearning.aleatoric_uncertainty, 'vl4pose': activelearning.vl4pose } self.dataset_size = dict() # Create dataset by converting into numpy compatible types if conf.dataset['load'] in ['mpii', 'merged']: logging.info('Creating MPII dataset\n') self.mpii = dataset_dict['mpii'] self.mpii_dataset = self.create_mpii() self.mpii_train, self.mpii_validate = self.create_train_validate(dataset=self.mpii_dataset) if self.conf.experiment_settings['all_joints']: logging.info('Creating single person patches\n') self.mpii_train = self.mpii_single_person_extractor(mpii_dataset=self.mpii_train) self.mpii_validate = self.mpii_single_person_extractor(mpii_dataset=self.mpii_validate) logging.info('Selecting train and validation images where all joints are present.\n') self.mpii_train = self.mpii_all_joints(mpii_dataset=self.mpii_train) self.mpii_validate = self.mpii_all_joints(mpii_dataset=self.mpii_validate) else: logging.info('Creating single person patches\n') self.mpii_train = self.mpii_single_person_extractor(mpii_dataset=self.mpii_train) self.mpii_validate = self.mpii_single_person_extractor(mpii_dataset=self.mpii_validate) logging.info('Size of MPII processed dataset: ') logging.info('Train: {}'.format(self.mpii_train['name'].shape[0])) logging.info('Validate: {}\n'.format(self.mpii_validate['name'].shape[0])) self.dataset_size.update( {'mpii': {'train': self.mpii_train['name'].shape[0], 'validation': self.mpii_validate['name'].shape[0]}}) if conf.dataset['load'] in ['lsp', 'merged']: self.lspet = dataset_dict['lspet'] self.lsp = dataset_dict['lsp'] logging.info('Creating LSPET dataset\n') self.lspet_dataset = self.create_lspet() logging.info('Creating LSP dataset\n') self.lsp_dataset = self.create_lsp() self.lspet_train, self.lspet_validate = self.create_train_validate(dataset=self.lspet_dataset) self.lsp_train, self.lsp_validate = self.create_train_validate(dataset=self.lsp_dataset) self.dataset_size.update({ 'lspet': {'train': self.lspet_train['name'].shape[0], 'validation': self.lspet_validate['name'].shape[0]}, 'lsp': {'train': self.lsp_train['name'].shape[0], 'validation': self.lsp_validate['name'].shape[0]}}) # Create train / validation data by combining individual components logging.info('Creating train and validation splits\n') if self.conf.dataset['load'] == 'mpii': self.train = self.merge_dataset(datasets=[self.mpii_train], indices=[np.arange(self.mpii_train['name'].shape[0])]) self.validate = self.merge_dataset(datasets=[self.mpii_validate], indices=[np.arange(self.mpii_validate['name'].shape[0])]) elif self.conf.dataset['load'] == 'lsp': self.train = self.merge_dataset(datasets=[self.lsp_train, self.lspet_train], indices=[np.arange(self.lsp_train['name'].shape[0]), np.arange(self.lspet_train['name'].shape[0])]) self.validate = self.merge_dataset(datasets=[self.lsp_validate, self.lspet_validate], indices=[np.arange(self.lsp_validate['name'].shape[0]), np.arange(self.lspet_validate['name'].shape[0])]) else: assert self.conf.dataset['load'] == 'merged', "dataset['load'] should be mpii, lsp or merged" self.train = self.merge_dataset(datasets=[self.mpii_train, self.lsp_train, self.lspet_train], indices=[np.arange(self.mpii_train['name'].shape[0]), np.arange(self.lsp_train['name'].shape[0]), np.arange(self.lspet_train['name'].shape[0])]) self.validate = self.merge_dataset(datasets=[self.mpii_validate, self.lsp_validate, self.lspet_validate], indices=[np.arange(self.mpii_validate['name'].shape[0]), np.arange(self.lsp_validate['name'].shape[0]), np.arange(self.lspet_validate['name'].shape[0])]) # Clearing RAM if conf.dataset['load'] in ['mpii', 'merged']: del self.mpii_train, self.mpii_validate, self.mpii_dataset, self.mpii if conf.dataset['load'] in ['lsp', 'merged']: del self.lspet_train, self.lspet_validate, self.lspet_dataset, self.lspet,\ self.lsp_train, self.lsp_validate, self.lsp_dataset, self.lsp, indices = activelearning_samplers[conf.active_learning['algorithm']]( train=self.train, dataset_size=self.dataset_size) self.train = self.merge_dataset(datasets=[self.train], indices=[indices]) logging.info('\nFinal size of Training Data: {}'.format(self.train['name'].shape[0])) logging.info('Final size of Validation Data: {}\n'.format(self.validate['name'].shape[0])) # Decide which dataset is input to the model self.input_dataset(train=True) # Deciding augmentation techniques self.shift_scale_rotate = self.augmentation([albu.ShiftScaleRotate(p=1, shift_limit=0.2, scale_limit=0.25, rotate_limit=45, interpolation=cv2.INTER_LINEAR, border_mode=cv2.BORDER_CONSTANT, value=0)]) self.flip_prob = 0.5 self.horizontal_flip = self.augmentation([albu.HorizontalFlip(p=1)]) logging.info('\nDataloader Initialized.\n') def __len__(self): """ Returns the length of the dataset. Duh! :return: """ return self.model_input_dataset['gt'].shape[0] def __getitem__(self, i): """ :param i: :return: """ root = Path(os.getcwd()).parent mpii_path = os.path.join(root, 'data', 'mpii') lsp_path = os.path.join(root, 'data', 'lsp') lspet_path = os.path.join(root, 'data', 'lspet') name = self.model_input_dataset['name'][i] gt = self.model_input_dataset['gt'][i] dataset = self.model_input_dataset['dataset'][i] num_gt = self.model_input_dataset['num_gt'][i] split = self.model_input_dataset['split'][i] num_persons = self.model_input_dataset['num_persons'][i] bbox_coords = self.model_input_dataset['bbox_coords'][i] normalizer = self.model_input_dataset['normalizer'][i] if dataset == 'mpii': image = plt.imread(os.path.join(mpii_path, 'images', '{}.jpg'.format(name.split('_')[0]))) elif dataset == 'lsp': image = plt.imread(os.path.join(lsp_path, 'images', name)) else: image = plt.imread(os.path.join(lspet_path, 'images', name)) # Convert from XY cartesian to UV image coordinates xy_to_uv = lambda xy: (xy[1], xy[0]) uv_to_xy = lambda uv: (uv[1], uv[0]) # Determine crop img_shape = np.array(image.shape) [min_x, min_y, max_x, max_y] = bbox_coords[0] tl_uv = xy_to_uv(np.array([min_x, min_y])) br_uv = xy_to_uv(np.array([max_x, max_y])) min_u = tl_uv[0] min_v = tl_uv[1] max_u = br_uv[0] max_v = br_uv[1] centre = np.array([(min_u + max_u) / 2, (min_v + max_v) / 2]) height = max_u - min_u width = max_v - min_v if self.train_flag: scale = np.random.uniform(low=1.5,high=2) else: scale = 1.75 window = max(scale * height, scale * width) top_left = np.array([centre[0] - (window / 2), centre[1] - (window / 2)]) bottom_right = np.array([centre[0] + (window / 2), centre[1] + (window / 2)]) top_left = np.maximum(np.array([0, 0], dtype=np.int16), top_left.astype(np.int16)) bottom_right = np.minimum(img_shape.astype(np.int16)[:-1], bottom_right.astype(np.int16)) # Cropping the image and adjusting the ground truth image = image[top_left[0]: bottom_right[0], top_left[1]: bottom_right[1], :] for person in range(gt.shape[0]): for joint in range(gt.shape[1]): gt_uv = xy_to_uv(gt[person][joint]) gt_uv = gt_uv - top_left gt[person][joint] = np.concatenate([gt_uv, np.array([gt[person][joint][2]])], axis=0) # Resize the image image, gt, scale_params = self.resize_image(image, gt, target_size=[256, 256, 3]) # Augmentation if self.train_flag: # Horizontal flip can't be done using albu's probability if torch.rand(1) < self.flip_prob: # Augment image and keypoints augmented = self.horizontal_flip(image=image, keypoints=gt.reshape(-1, 3)[:, :2]) image = augmented['image'] gt[:, :, :2] = np.stack(augmented['keypoints'], axis=0).reshape(-1, self.conf.experiment_settings['num_hm'], 2) # Flip ground truth to match horizontal flip gt[:, [jnt_to_ind['lsho'], jnt_to_ind['rsho']], :] = gt[:, [jnt_to_ind['rsho'], jnt_to_ind['lsho']], :] gt[:, [jnt_to_ind['lelb'], jnt_to_ind['relb']], :] = gt[:, [jnt_to_ind['relb'], jnt_to_ind['lelb']], :] gt[:, [jnt_to_ind['lwri'], jnt_to_ind['rwri']], :] = gt[:, [jnt_to_ind['rwri'], jnt_to_ind['lwri']], :] gt[:, [jnt_to_ind['lhip'], jnt_to_ind['rhip']], :] = gt[:, [jnt_to_ind['rhip'], jnt_to_ind['lhip']], :] gt[:, [jnt_to_ind['lknee'], jnt_to_ind['rknee']], :] = gt[:, [jnt_to_ind['rknee'], jnt_to_ind['lknee']], :] gt[:, [jnt_to_ind['lankl'], jnt_to_ind['rankl']], :] = gt[:, [jnt_to_ind['rankl'], jnt_to_ind['lankl']], :] # Ensure shift scale rotate augmentation retains all joints if self.conf.experiment_settings['all_joints']: tries = 5 augment_ok = False image_, gt_ = None, None while tries > 0: tries -= 1 augmented = self.shift_scale_rotate(image=image, keypoints=gt.reshape(-1, 3)[:, :2]) image_ = augmented['image'] gt_ = np.stack(augmented['keypoints'], axis=0).reshape( -1, self.conf.experiment_settings['num_hm'], 2) if (np.all(gt_[0]) > -5) and (np.all(gt_[0]) < 261): # 0 index single person augment_ok = True break if augment_ok: image = image_ gt[:, :, :2] = gt_ # Else augmentation does not need to keep all joints else: augmented = self.shift_scale_rotate(image=image, keypoints=gt.reshape(-1, 3)[:, :2]) image = augmented['image'] gt[:, :, :2] = np.stack(augmented['keypoints'], axis=0).reshape(-1, self.conf.num_hm, 2) heatmaps, joint_exist = heatmap_generator( joints=np.copy(gt),occlusion=self.occlusion, hm_shape=self.hm_shape, img_shape=image.shape) heatmaps = self.hm_peak * heatmaps # Reproduce GT overlayed on Image if False:# self.viz: fig, ax = plt.subplots(1, 1) ax.imshow(image) for person in range(gt.shape[0]): for joint in range(gt.shape[1]): gt_xy = uv_to_xy(gt[person][joint]) if gt[person][joint][2] >= 0 and gt[person][joint][0] > 0 and gt[person][joint][1] > 0: ax.add_patch(Circle(gt_xy, radius=2.5, color='green', fill=True)) for k in range(joint_exist.shape[0]): print(ind_to_jnt[k], joint_exist[k]) print('\n') plt.show() #os.makedirs(os.path.join(self.model_save_path[:-20], 'viz_getitem'), exist_ok=True) #plt.savefig(os.path.join(self.model_save_path[:-20], 'viz_getitem', '{}.jpg'.format(i)), dpi=350) return torch.tensor(data=image/256.0, dtype=torch.float32, device='cpu'), \ torch.tensor(data=heatmaps, dtype=torch.float32, device='cpu'),\ gt, name, dataset, num_gt.astype(np.float32), split, num_persons,\ scale_params, normalizer, joint_exist def create_mpii(self):#(self, train_ratio=0.7, max_persons=0, shuffle=False): ''' :param train_ratio: :param max_persons: :param shuffle: :return: ''' mpii = self.mpii max_persons = mpii['max_person_in_img'] len_dataset = len(mpii['img_name']) assert len(mpii['img_name']) == len(mpii['img_gt']), "MPII dataset image and labels mismatched." # What the create mpii dataset will look like. dataset = {'img': [], 'name': [], 'gt': -np.ones(shape=(len_dataset, max_persons, self.conf.experiment_settings['num_hm'], 3)), 'dataset': [], 'num_gt': np.zeros(shape=(len_dataset, max_persons)), 'split': [], 'num_persons': np.zeros(shape=(len_dataset, 1)), 'normalizer': np.zeros(shape=(len_dataset, max_persons)), 'bbox_coords': -np.ones(shape=(len_dataset, max_persons, 4))} # Converting the dataset to numpy arrays for i in range(len_dataset): image_name = mpii['img_name'][i] ground_truth = mpii['img_gt'][i] dataset_ = mpii['dataset'][i] num_gt = mpii['num_gt'][i] split = mpii['split'][i] normalizer = mpii['normalizer'][i] # Calculating the number of people num_ppl = 0 for key in ground_truth.keys(): num_ppl = max(num_ppl, len(ground_truth[key])) break # All keys have same length, as if jnt absent in person, then we append np.array([-1, -1, -1]) dataset['num_persons'][i] = num_ppl # Split 0 indicates testing dataset assert split == 1, "All annotated images should have split == 1" # Assigning to a Numpy Ground truth array for jnt in ground_truth.keys(): for person_id in range(len(ground_truth[jnt])): dataset['gt'][i, person_id, jnt_to_ind[jnt]] = ground_truth[jnt][person_id] # Assigning Bounding Box coordinates per person for person_id in range(num_ppl): x_coord = dataset['gt'][i, person_id, :, 0][np.where(dataset['gt'][i, person_id, :, 0] > -1)] y_coord = dataset['gt'][i, person_id, :, 1][np.where(dataset['gt'][i, person_id, :, 1] > -1)] min_x = np.min(x_coord) max_x = np.max(x_coord) min_y = np.min(y_coord) max_y = np.max(y_coord) dataset['bbox_coords'][i, person_id] = np.array([min_x, min_y, max_x, max_y]) # Number of joints scaling factor for person_id in range(len(num_gt)): dataset['num_gt'][i, person_id] = num_gt[person_id] # PCK normalizer for person_id in range(len(normalizer)): dataset['normalizer'][i, person_id] = normalizer[person_id] dataset['name'].append(image_name) dataset['dataset'].append(dataset_) dataset['split'].append(split) # Load Train/Test split if conf.model_load_hg = True if self.conf.model['load']: dataset['split'] = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/mpii_split.npy')) else: # Create our own train/validation split for multi person dataset # 0: Train; 1: Validate if self.conf.dataset['mpii_params']['newell_validation']: root = Path(os.getcwd()).parent logging.info('\nCreating the Newell validation split.\n') with open(os.path.join(root, 'cached', 'Stacked_HG_ValidationImageNames.txt')) as valNames: valNames_ = [x.strip('\n') for x in valNames.readlines()] dataset['split'] = np.array([1 if x in valNames_ else 0 for x in dataset['name']]) # assert np.sum(dataset['split']) == len(valNames_) # "THIS ASSERTION WILL NOT HOLD TRUE. Newell list has duplicates." else: train_ratio = self.conf.dataset['mpii_params']['train_ratio'] dataset['split'] = np.concatenate([np.zeros(int(len_dataset*train_ratio),), np.ones((len_dataset - int(len_dataset*train_ratio)),)], axis=0) if self.conf.dataset['mpii_params']['shuffle']: np.random.shuffle(dataset['split']) np.save(file=os.path.join(self.conf.model['save_path'], 'model_checkpoints/mpii_split.npy'), arr=dataset['split']) dataset['img'] = np.array(dataset['img']) # array of shape == 0, dataset['img'] exists for legacy reasons only dataset['name'] = np.array(dataset['name']) dataset['dataset'] = np.array(dataset['dataset']) logging.info('MPII dataset description:') logging.info('Length (#images): {}'.format(dataset['gt'].shape[0])) return dataset def create_lspet(self): """ :return: """ try: max_persons = self.mpii['max_person_in_img'] except AttributeError: max_persons = 1 lspet = copy.deepcopy(self.lspet) assert len(lspet['img_name']) == len(lspet['img_gt']), "LSPET dataset image and labels mismatched." dataset = {'img': [], 'name': [], 'gt': -np.ones(shape=(len(lspet['img_name']), max_persons, self.conf.experiment_settings['num_hm'], 3)), 'dataset': [], 'num_gt': np.zeros(shape=(len(lspet['img_name']), max_persons)), 'split': [], 'num_persons': np.ones(shape=(len(lspet['img_name']), 1)), 'normalizer': np.zeros(shape=(len(lspet['img_name']), max_persons)), 'bbox_coords': -np.ones(shape=(len(lspet['img_name']), max_persons, 4))} # max_persons is always 1 for lsp* len_dataset = len(lspet['img_name']) for i in range(len_dataset): image_name = lspet['img_name'][i] ground_truth = lspet['img_gt'][i] dataset_ = lspet['dataset'][i] num_gt = lspet['num_gt'][i] split = lspet['split'][i] normalizer = lspet['normalizer'][i] for jnt in ground_truth.keys(): dataset['gt'][i, 0, jnt_to_ind[jnt]] = ground_truth[jnt][0] # Assigning Bounding Box coordinates per person x_coord = dataset['gt'][i, 0, :, 0][np.where(dataset['gt'][i, 0, :, 2] == 1)] y_coord = dataset['gt'][i, 0, :, 1][np.where(dataset['gt'][i, 0, :, 2] == 1)] x_coord = x_coord[np.where(x_coord > -1)] y_coord = y_coord[np.where(y_coord > -1)] min_x = np.min(x_coord) max_x = np.max(x_coord) min_y = np.min(y_coord) max_y = np.max(y_coord) dataset['bbox_coords'][i, 0] = np.array([min_x, min_y, max_x, max_y]) # Assigning number of GT to person 0 dataset['num_gt'][i, 0] = num_gt[0] dataset['normalizer'][i, 0] = normalizer[0] dataset['name'].append(image_name) dataset['dataset'].append(dataset_) dataset['split'].append(split) dataset['img'] = np.array(dataset['img']) # Empty array if load_all_images_in_memory = False dataset['name'] = np.array(dataset['name']) dataset['dataset'] = np.array(dataset['dataset']) dataset['split'] = np.array(dataset['split']) logging.info('LSPET dataset description:') logging.info('Length (#images): {}'.format(dataset['gt'].shape[0])) return dataset def create_lsp(self): """ :param max_persons: :return: """ try: max_persons = self.mpii['max_person_in_img'] except AttributeError: max_persons = 1 lsp = copy.deepcopy(self.lsp) assert len(lsp['img_name']) == len(lsp['img_gt']), "LSP dataset image and labels mismatched." dataset = {'img': [], 'name': [], 'gt': -np.ones(shape=(len(lsp['img_name']), max_persons, self.conf.experiment_settings['num_hm'], 3)), 'dataset': [], 'num_gt': np.zeros(shape=(len(lsp['img_name']), max_persons)), 'split': [], 'num_persons': np.ones(shape=(len(lsp['img_name']), 1)), 'normalizer': np.zeros(shape=(len(lsp['img_name']), max_persons)), 'bbox_coords': -np.ones(shape=(len(lsp['img_name']), max_persons, 4))} # max_persons is always 1 for lsp* len_dataset = len(lsp['img_name']) for i in range(len_dataset): image_name = lsp['img_name'][i] ground_truth = lsp['img_gt'][i] dataset_ = lsp['dataset'][i] num_gt = lsp['num_gt'][i] split = lsp['split'][i] normalizer = lsp['normalizer'][i] for jnt in ground_truth.keys(): dataset['gt'][i, 0, jnt_to_ind[jnt]] = ground_truth[jnt][0] # Assigning Bounding Box coordinates per person x_coord = dataset['gt'][i, 0, :, 0][np.where(dataset['gt'][i, 0, :, 0] > -1)] y_coord = dataset['gt'][i, 0, :, 1][np.where(dataset['gt'][i, 0, :, 1] > -1)] min_x = np.min(x_coord) max_x = np.max(x_coord) min_y = np.min(y_coord) max_y = np.max(y_coord) dataset['bbox_coords'][i, 0] = np.array([min_x, min_y, max_x, max_y]) dataset['num_gt'][i, 0] = num_gt[0] dataset['normalizer'][i, 0] = normalizer[0] dataset['name'].append(image_name) dataset['dataset'].append(dataset_) dataset['split'].append(split) dataset['img'] = np.array(dataset['img']) dataset['name'] = np.array(dataset['name']) dataset['dataset'] = np.array(dataset['dataset']) dataset['split'] = np.array(dataset['split']) logging.info('LSP dataset description:') logging.info('Length (#images): {}'.format(dataset['gt'].shape[0])) return dataset def create_train_validate(self, dataset=None): """ :param dataset: :return: """ # Separate train and validate train_idx = [] val_idx = [] for i in range(len(dataset['name'])): if dataset['split'][i] == 0: train_idx.append(i) else: assert dataset['split'][i] == 1, \ "Split has value: {}, should be either 0 or 1".format(dataset['split'][i]) val_idx.append(i) train_dataset = {} val_dataset = {} for key in dataset.keys(): if key == 'img': train_dataset[key] = dataset[key] # Empty numpy array val_dataset[key] = dataset[key] continue train_dataset[key] = dataset[key][train_idx] val_dataset[key] = dataset[key][val_idx] return train_dataset, val_dataset def merge_dataset(self, datasets=None, indices=None): """ :param datasets: :param indices: :return: """ assert type(datasets) == list and len(datasets) != 0 assert len(datasets) == len(indices) for i in range(len(datasets) - 1): assert datasets[i].keys() == datasets[i+1].keys(), "Dataset keys do not match" # Merge datasets merged_dataset = {} for key in datasets[0].keys(): if key == 'img': merged_dataset['img'] = np.array([]) continue merged_dataset[key] = np.concatenate([data[key][index_] for index_, data in zip(indices, datasets)], axis=0) # Sampling based on indices merged_dataset['index'] = np.arange(merged_dataset['name'].shape[0]) return merged_dataset def recreate_images(self, gt=False, pred=False, train=False, validate=False, external=False, ext_data=None): ''' :return: ''' assert gt + pred != 0, "Specify atleast one of GT or Pred" assert train + validate + external == 1,\ "Can create visualize_image compatible arrays only for train/val in one function call." if external: assert ext_data, "ext_dataset can't be none to recreate external datasets" data_split = ext_data elif train: data_split = self.train else: data_split = self.validate # Along with the below entries, we also pass bbox coordinates for each dataset img_dict = {'mpii': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'split': [], 'dataset': [], 'display_string': []}, 'lspet': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'split': [], 'dataset': [], 'display_string': []}, 'lsp': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'split': [], 'dataset': [], 'display_string': []}} for i in range(len(data_split['img'])): dataset = data_split['dataset'][i] img_dict[dataset]['img'].append(data_split['img'][i]) img_dict[dataset]['img_name'].append(data_split['name'][i]) img_dict[dataset]['split'].append(data_split['split'][i]) img_dict[dataset]['dataset'].append(data_split['dataset'][i]) img_dict[dataset]['display_string'].append(data_split['name'][i]) try: img_dict[dataset]['bbox_coords'] = np.concatenate([img_dict[dataset]['bbox_coords'], data_split['bbox_coords'][i].reshape(1, -1, 4)], axis=0) except KeyError: img_dict[dataset]['bbox_coords'] = data_split['bbox_coords'][i].reshape(1, -1, 4) joint_dict = dict([(ind_to_jnt[i], []) for i in range(self.conf.experiment_settings['num_hm'])]) gt_dict = copy.deepcopy(joint_dict) pred_dict = copy.deepcopy(joint_dict) if gt: for person in range(int(data_split['num_persons'][i, 0])): for joint in range(self.conf.experiment_settings['num_hm']): gt_dict[ind_to_jnt[joint]].append(data_split['gt'][i, person, joint]) if pred: for person in range(int(data_split['num_persons'][i, 0])): for joint in range(self.conf.experiment_settings['num_hm']): pred_dict[ind_to_jnt[joint]].append(data_split['pred'][i, person, joint]) img_dict[dataset]['img_gt'].append(gt_dict) img_dict[dataset]['img_pred'].append(pred_dict) return img_dict def input_dataset(self, train=False, validate=False): ''' :param train: :param validate_entire: :return: ''' assert train + validate == 1, "Either one of train or validate_entire needs to be chosen" if train: self.model_input_dataset = self.train self.train_flag = True self.validate_flag = False else: self.model_input_dataset = self.validate self.validate_flag = True self.train_flag = False return None def mpii_single_person_extractor(self, mpii_dataset): """ :param train: :param validate: :param max_persons: :return: """ max_persons = self.mpii['max_person_in_img'] dataset = {'img': [], 'name': [], 'gt': np.empty(shape=(0, max_persons, self.conf.experiment_settings['num_hm'], 3)), 'dataset': [], 'num_gt': np.empty(shape=(0, max_persons)), 'split': [], 'num_persons': np.empty(shape=(0, 1)), 'normalizer': np.empty(shape=(0, max_persons)), 'bbox_coords': np.empty(shape=(0, max_persons, 4))} for i in range(len(mpii_dataset['name'])): for p in range(int(mpii_dataset['num_persons'][i][0])): dataset['name'].append(mpii_dataset['name'][i][:-4] + '_{}.jpg'.format(p)) dataset['dataset'].append(mpii_dataset['dataset'][i]) dataset['split'].append(mpii_dataset['split'][i]) gt_ = -1 * np.ones_like(mpii_dataset['gt'][i]) gt_[0] = mpii_dataset['gt'][i, p] dataset['gt'] = np.concatenate([dataset['gt'], gt_.reshape(1, max_persons, self.conf.experiment_settings['num_hm'], 3)], axis=0) num_gt_ = np.zeros_like(mpii_dataset['num_gt'][i]) num_gt_[0] = mpii_dataset['num_gt'][i, p] dataset['num_gt'] = np.concatenate([dataset['num_gt'], num_gt_.reshape(1, max_persons)], axis=0) normalizer_ = np.zeros_like(mpii_dataset['normalizer'][i]) normalizer_[0] = mpii_dataset['normalizer'][i, p] dataset['normalizer'] = np.concatenate([dataset['normalizer'], normalizer_.reshape(1, max_persons)], axis=0) dataset['num_persons'] = np.concatenate([dataset['num_persons'], np.array([1]).reshape(1, 1)], axis=0) bbox_ = np.zeros_like(mpii_dataset['bbox_coords'][i]) bbox_[0] = mpii_dataset['bbox_coords'][i, p] dataset['bbox_coords'] = np.concatenate([dataset['bbox_coords'],bbox_.reshape(1, max_persons, 4)], axis=0) dataset['img'] = np.array(dataset['img']) dataset['name'] = np.array(dataset['name']) dataset['dataset'] = np.array(dataset['dataset']) dataset['split'] = np.array(dataset['split']) return dataset def mpii_all_joints(self, mpii_dataset): """ :param mpii_dataset: :return: """ # [:, 0, :, 2] --> all images, first person, all joints, visibility all_joint_indices = mpii_dataset['gt'][:, 0, :, 2] > -0.5 # Identify indices: occluded and visible joints all_joint_indices = np.all(all_joint_indices, axis=1) for key in mpii_dataset.keys(): if key == 'img': mpii_dataset['img'] = mpii_dataset['img'] continue mpii_dataset[key] = mpii_dataset[key][all_joint_indices] return mpii_dataset def resize_image(self, image_=None, gt=None, target_size=None): ''' :return: ''' # Compute the aspect ratios image_aspect_ratio = image_.shape[0] / image_.shape[1] tgt_aspect_ratio = target_size[0] / target_size[1] # Compare the original and target aspect ratio if image_aspect_ratio > tgt_aspect_ratio: # If target aspect ratio is smaller, scale the first dim scale_factor = target_size[0] / image_.shape[0] else: # If target aspect ratio is bigger or equal, scale the second dim scale_factor = target_size[1] / image_.shape[1] # Compute the padding to fit the target size pad_u = (target_size[0] - int(image_.shape[0] * scale_factor)) pad_v = (target_size[1] - int(image_.shape[1] * scale_factor)) output_img = np.zeros(target_size, dtype=image_.dtype) # Write scaled size in reverse order because opencv resize scaled_size = (int(image_.shape[1] * scale_factor), int(image_.shape[0] * scale_factor)) padding_u = int(pad_u / 2) padding_v = int(pad_v / 2) im_scaled = cv2.resize(image_, scaled_size) # logging.debug('Scaled, pre-padding size: {}'.format(im_scaled.shape)) output_img[padding_u : im_scaled.shape[0] + padding_u, padding_v : im_scaled.shape[1] + padding_v, :] = im_scaled gt *= np.array([scale_factor, scale_factor, 1]).reshape(1, 1, 3) gt[:, :, 0] += padding_u gt[:, :, 1] += padding_v scale_params = {'scale_factor': scale_factor, 'padding_u': padding_u, 'padding_v': padding_v} return output_img, gt, scale_params def augmentation(self, transform): ''' :param transformation: :return: ''' return albu.Compose(transform, p=1, keypoint_params=albu.KeypointParams(format='yx', remove_invisible=False)) def estimate_uv(self, hm_array, pred_placeholder): ''' :param hm_array: :param pred_placeholder: :return: ''' if self.conf.experiment_settings['all_joints']: threshold = 0 else: threshold = self.threshold # Iterate over each heatmap for jnt_id in range(hm_array.shape[0]): pred_placeholder[0, jnt_id, :] = uv_from_heatmap(hm=hm_array[jnt_id], threshold=threshold) return pred_placeholder def upscale(self, joints, scale_params): ''' :return: ''' joints[:, :, 0] -= scale_params['padding_u'] joints[:, :, 1] -= scale_params['padding_v'] joints /= np.array([scale_params['scale_factor'], scale_params['scale_factor'], 1]).reshape(1, 1, 3) return joints
52,432
41.182623
137
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/utils.py
import os import sys import copy import math from pathlib import Path import torch import scipy.io import numpy as np from tqdm import tqdm from adjustText import adjust_text from matplotlib import pyplot as plt from matplotlib.patches import Circle, Rectangle import umap from sklearn.decomposition import PCA plt.style.use('ggplot') plt.switch_backend('agg') def visualize_image(image_info, save_dir, bbox=False, adjusttext=False): ''' :param image_info: (dict) ''' uv_to_xy = lambda uv: (uv[1], uv[0]) mpii_skelton = [['head', 'neck', (0, 0, 1)], ['neck', 'thorax', (0, 0, 1)], ['thorax', 'lsho', (0, 1, 0)], ['lsho', 'lelb', (0, 1, 1)], ['lelb', 'lwri', (0, 1, 1)], ['thorax', 'rsho', (0, 1, 0)], ['rsho', 'relb', (1, 0, 0)], ['relb', 'rwri', (1, 0, 0)], ['lsho', 'lhip', (0, 1, 0)], ['rsho', 'rhip', (0, 1, 0)], ['pelvis', 'lhip', (0, 1, 0)], ['lhip', 'lknee', (1, 0, 1)], ['lknee', 'lankl', (1, 0, 1)], ['pelvis', 'rhip', (0, 1, 0)], ['rhip', 'rknee', (1, 1, 0)], ['rknee', 'rankl', (1, 1, 0)]] lsp_skeleton = [['head', 'neck', (0, 0, 1)], ['lhip', 'rhip', (0, 1, 0)], ['neck', 'lsho', (0, 1, 0)], ['lsho', 'lelb', (0, 1, 1)], ['lelb', 'lwri', (0, 1, 1)], ['neck', 'rsho', (0, 1, 0)], ['rsho', 'relb', (1, 0, 0)], ['relb', 'rwri', (1, 0, 0)], ['lsho', 'lhip', (0, 1, 0)], ['lhip', 'lknee', (1, 0, 1)], ['lknee', 'lankl', (1, 0, 1)], ['rsho', 'rhip', (0, 1, 0)], ['rhip', 'rknee', (1, 1, 0)], ['rknee', 'rankl', (1, 1, 0)]] colour = {'rankl': (1, 1, 0), 'rknee': (1, 1, 0), 'rhip': (1, 1, 0), 'lankl': (1, 0, 1), 'lknee': (1, 0, 1), 'lhip': (1, 0, 1), 'pelvis': (0, 1, 0), 'rwri': (1, 0, 0), 'relb': (1, 0, 0), 'rsho': (1, 0, 0), 'lwri': (0, 1, 1), 'lelb': (0, 1, 1), 'lsho': (0, 1, 1), 'head': (0, 1, 1), 'neck': (0, 1, 1), 'thorax': (0, 1, 0)} os.makedirs(os.path.join(save_dir, 'skeleton_visualizations'), exist_ok=True) img_dump = os.path.join(save_dir, 'skeleton_visualizations') # Currently will iterate over MPII and LSPET and LSP for dataset_name_ in image_info.keys(): # Iterate over all images for i in tqdm(range(len(image_info[dataset_name_]['img']))): fig, ax = plt.subplots(nrows=1, ncols=1, frameon=False) ax.set_axis_off() img = image_info[dataset_name_]['img'][i] img_name = image_info[dataset_name_]['img_name'][i] img_pred = image_info[dataset_name_]['img_pred'][i] img_gt = image_info[dataset_name_]['img_gt'][i] img_dataset = image_info[dataset_name_]['dataset'][i] img_string = image_info[dataset_name_]['display_string'][i] # One list for each, ground truth and predictions text_overlay = [] ax.set_title('Name: {}, Shape: {}, Dataset: {}'.format(img_string, str(img.shape), img_dataset), color='orange') ax.imshow(img) if dataset_name_ == 'mpii': skeleton = mpii_skelton else: assert dataset_name_ == 'lsp' or dataset_name_ == 'lspet' skeleton = lsp_skeleton for link in skeleton: joint_1_name = link[0] joint_2_name = link[1] color = link[2] joint_1 = img_pred[joint_1_name][0] joint_2 = img_pred[joint_2_name][0] if joint_1[2] == 1 and joint_2[2] == 1: joint_1 = uv_to_xy(joint_1) joint_2 = uv_to_xy(joint_2) ax.plot([joint_1[0], joint_2[0]], [joint_1[1], joint_2[1]], color=color) joint_names = list(colour.keys()) for jnt in joint_names: if (dataset_name_ in ['lsp', 'lspet']) and (jnt in ['pelvis', 'thorax']): continue for jnt_gt in img_gt[jnt]: if jnt_gt[2] >= 0 and jnt_gt[1] >= 0 and jnt_gt[0] >= 0: jnt_gt = uv_to_xy(jnt_gt) text_overlay.append(ax.text(x=jnt_gt[0], y=jnt_gt[1], s=jnt, color=colour[jnt], fontsize=6)) ax.add_patch(Circle(jnt_gt[:2], radius=1.5, color=colour[jnt], fill=False)) if bbox: for person_patch in range(image_info[dataset_name_]['bbox_coords'].shape[1]): coords = image_info[dataset_name_]['bbox_coords'][i, person_patch] ax.add_patch(Rectangle(xy=(coords[0], coords[1]), height=(coords[3] - coords[1]), width=(coords[2]-coords[0]), linewidth=1, edgecolor='r', fill=False)) if adjusttext: adjust_text(text_overlay) plt.savefig(fname=os.path.join(img_dump, '{}'.format(img_string)), facecolor='black', edgecolor='black', bbox_inches='tight', dpi=500) del fig, ax plt.close() def heatmap_loss(combined_hm_preds, heatmaps, egl=False): ''' :param combined_hm_preds: :param heatmaps: :param nstack: :return: ''' calc_loss = lambda pred, gt: ((pred - gt) ** 2).mean(dim=[1, 2, 3]) combined_loss = [] nstack = combined_hm_preds.shape[1] for i in range(nstack): if egl: combined_loss.append(calc_loss(combined_hm_preds[:, i], heatmaps[:, i].to(combined_hm_preds[:, i].device))) else: combined_loss.append(calc_loss(combined_hm_preds[:, i], heatmaps.to(combined_hm_preds[:, i].device))) combined_loss = torch.stack(combined_loss, dim=1) return combined_loss def heatmap_generator(joints, occlusion, hm_shape=(0, 0), img_shape=(0, 0)): ''' :param joints: :return: ''' def draw_heatmap(pt_uv, use_occlusion, hm_shape, sigma=1.75): ''' 2D gaussian (exponential term only) centred at given point. No constraints on point to be integer only. :param im: (Numpy array of size=64x64) Heatmap :param pt: (Numpy array of size=2) Float values denoting point on the heatmap :param sigma: (Float) self.joint_size which determines the standard deviation of the gaussian :return: (Numpy array of size=64x64) Heatmap with gaussian centred around point. ''' im = np.zeros(hm_shape, dtype=np.float32) # If joint is absent if pt_uv[2] == -1: return im, 0 elif pt_uv[2] == 0: if not use_occlusion: return im, 0 else: assert pt_uv[2] == 1, "joint[2] should be (-1, 0, 1), but got {}".format(pt_uv[2]) # Point around which Gaussian will be centred. pt_uv = pt_uv[:2] pt_uv_rint = np.rint(pt_uv).astype(int) # Size of 2D Gaussian window. size = int(math.ceil(6 * sigma)) # Ensuring that size remains an odd number if not size % 2: size += 1 # Check whether gaussian intersects with im: if (pt_uv_rint[0] - (size//2) >= hm_shape[0]) or (pt_uv_rint[0] + (size//2) <= 0) \ or (pt_uv_rint[1] - (size//2) > hm_shape[1]) or (pt_uv_rint[1] + (size//2) < 0): return im, 0 else: # Generate gaussian, with window=size and variance=sigma u = np.arange(pt_uv_rint[0] - (size // 2), pt_uv_rint[0] + (size // 2) + 1) v = np.arange(pt_uv_rint[1] - (size // 2), pt_uv_rint[1] + (size // 2) + 1) uu, vv = np.meshgrid(u, v, sparse=True) z = np.exp(-((uu - pt_uv[0]) ** 2 + (vv - pt_uv[1]) ** 2) / (2 * (sigma ** 2))) z = z.T # Identify indices in im that will define the crop area top = max(0, pt_uv_rint[0] - (size//2)) bottom = min(hm_shape[0], pt_uv_rint[0] + (size//2) + 1) left = max(0, pt_uv_rint[1] - (size//2)) right = min(hm_shape[1], pt_uv_rint[1] + (size//2) + 1) im[top:bottom, left:right] = \ z[top - (pt_uv_rint[0] - (size//2)): top - (pt_uv_rint[0] - (size//2)) + (bottom - top), left - (pt_uv_rint[1] - (size//2)): left - (pt_uv_rint[1] - (size//2)) + (right - left)] return im, 1 # heatmap, joint_exist assert len(joints.shape) == 3, 'Joints should be rank 3:' \ '(num_person, num_joints, [u,v,vis]), but is instead {}'.format(joints.shape) heatmaps = np.zeros([joints.shape[1], hm_shape[0], hm_shape[1]], dtype=np.float32) joints_exist = np.zeros([joints.shape[1]], dtype=np.uint8) # Downscale downscale = [(img_shape[0] - 1)/(hm_shape[0] - 1), ((img_shape[1] - 1)/(hm_shape[1] - 1))] joints /= np.array([downscale[0], downscale[1], 1]).reshape(1, 1, 3) # Iterate over number of heatmaps for i in range(joints.shape[1]): # Create new heatmap for joint hm_i = np.zeros(hm_shape, dtype=np.float32) # Iterate over persons for p in range(joints.shape[0]): hm_, joint_present = draw_heatmap(pt_uv=joints[p, i, :], use_occlusion=occlusion, hm_shape=hm_shape) joints_exist[i] = max(joints_exist[i], joint_present) hm_i = np.maximum(hm_i, hm_) heatmaps[i] = hm_i return heatmaps, joints_exist def uv_from_heatmap(hm=None, threshold=None, img_shape=(256, 256)): ''' :param hm: :param threshold: :param img_shape: :return: ''' max_uv = arg_max(hm) corrected_uv = weight_avg_centre(hm=hm, max_uv=max_uv) if hm[int(corrected_uv[0]), int(corrected_uv[1])] < threshold: return np.array([-1, -1, -1]) else: joints = np.array([corrected_uv[0], corrected_uv[1], 1]) hm_shape = hm.shape upscale = [(img_shape[0] - 1) / (hm_shape[0] - 1), ((img_shape[1] - 1) / (hm_shape[1] - 1))] joints *= np.array([upscale[0], upscale[1], 1]) return joints def arg_max(img): ''' Find the indices corresponding to maximum values in the heatmap :param img: (Numpy array of size=64x64) Heatmap :return: (Torch tensor of size=2) argmax of the image ''' img = torch.tensor(img) assert img.dim() == 2, 'Expected img.dim() == 2, got {}'.format(img.dim()) h = img.shape[0] w = img.shape[1] rawmaxidx = img.flatten().argmax() max_u = int(rawmaxidx) // int(w) max_v = int(rawmaxidx) % int(w) return torch.FloatTensor([max_u, max_v]) def fast_argmax(_heatmaps): """ Direct argmax from the heatmap, does not perform smoothing of heatmaps :param _heatmaps: :return: """ batch_size = _heatmaps.shape[0] num_jnts = _heatmaps.shape[1] spatial_dim = _heatmaps.shape[3] assert _heatmaps.shape[2] == _heatmaps.shape[3] assert len(_heatmaps.shape) == 4, "Heatmaps should be of shape: BatchSize x num_joints x 64 x64" _heatmaps = _heatmaps.reshape(batch_size, num_jnts, -1) indices = torch.argmax(_heatmaps, dim=2) indices = torch.cat(((indices // spatial_dim).view(batch_size, num_jnts, 1), (indices % spatial_dim).view(batch_size, num_jnts, 1)), dim=2) return indices.type(torch.float32) def weight_avg_centre(hm, max_uv=None, jnt_size=1.75): ''' Weighted average of points around the maxima. Weighted average avoids solitary spikes being identified. :param hm: (Numpy array of size 64x64) :param jnt_size: (Float) Windows size around the maxima to compute weighted average. :return: (Numpy array of size=2) ''' hm = torch.clamp(torch.from_numpy(hm), min=0.0) mx = max_uv # Dimension of the heatmap siz = torch.Tensor([hm.shape[0], hm.shape[1]]).float() # Clip indices if needed so that start and end indices are valid points. st_idx = torch.max(torch.zeros(2), mx - np.ceil(jnt_size)) end_idx = torch.min(siz - 1, mx + np.ceil(jnt_size)) # Crop around the maxima. img_crop = hm[int(st_idx[0]):int(end_idx[0] + 1), int(st_idx[1]):int(end_idx[1] + 1)].clone() img_crop = img_crop.type(torch.FloatTensor) img_sum = img_crop.sum() if img_sum == 0: img_sum = img_sum + 0.000001 # Weighted average along column/row u = img_crop.sum(1).mul(torch.arange(st_idx[0], end_idx[0] + 1)).div(img_sum).sum() v = img_crop.sum(0).mul(torch.arange(st_idx[1], end_idx[1] + 1)).div(img_sum).sum() return np.array([u, v]) def principal_component_analysis(encodings, n_components=2): ''' Compute the principal component transform of the encodings :param encodings: Encodings generated by LLAL network :param n_components: Number of components to retain :return: Principal Component Transform of encodings ''' pca = PCA(n_components=n_components) pca.fit(encodings) pca_encodings = pca.transform(encodings) return pca_encodings def umap_fn(encodings, n_components=2): ''' NUMPY https://umap-learn.readthedocs.io/en/latest/how_umap_works.html Compute the UMAP transform of the encodings :param encodings: Encodings generated by LLAL network :param n_components: Number of components to retain :return: UMAP transform of the encodings ''' # Number of neighbours balances the local versus the global structure of the data umap_transform = umap.UMAP(n_neighbors=30, min_dist=0.0, n_components=n_components).fit(encodings) umap_encodings = umap_transform.transform(encodings) return umap_encodings def shannon_entropy(probs): ''' Computes the Shannon Entropy for a distribution :param probs_array: 2D-Tensor; Probability distribution along axis=1 :return: Scalar; H(p) ''' return torch.sum(-probs * torch.log(probs), dim=1) def count_parameters(model): return sum(p.numel() for p in model.parameters() if p.requires_grad) def get_pairwise_joint_distances(heatmaps): ''' Computes pairwise distances between joints, given a batch of heatmaps - param heatmaps: Tensor of size (batch_size, num_joints, hm_size, hm_size) - return pairwise_dists: List[2D lists] - length = batch_size - Each 2D List (pairwise distances) of dim [num_joints,num_joints] ''' # Batch Size, num_jnts, 64, 64 assert heatmaps.dim() == 4, 'Dimension of input heatmaps must be 4 but was {}'.format(heatmaps.shape) # Batch Size, num_jnts, 2 joint_uv = fast_argmax(heatmaps) pairwise_dists = torch.cdist(joint_uv, joint_uv) return pairwise_dists
14,742
36.513995
130
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/activelearning_viz.py
import os import cv2 import math import torch import torch.utils.data import numpy as np import logging from tqdm import tqdm from pathlib import Path from sklearn.metrics import pairwise_distances from skimage.feature import peak_local_max from scipy.special import softmax as softmax_fn from scipy.stats import entropy as entropy_fn from matplotlib import pyplot as plt import copy # EGL sampling import autograd_hacks from utils import heatmap_loss from utils import shannon_entropy from utils import heatmap_generator from utils import uv_from_heatmap from utils import visualize_image plt.style.use('ggplot') class ActiveLearning_Visualization(object): """ Contains collection of active learning algorithms for human joint localization """ def __init__(self, conf, pose_net, aux_net): self.conf = conf self.pose_model = pose_net self.aux_net = aux_net self.pose_model.eval() if conf.active_learning['algorithm'] in ['learning_loss', 'aleatoric', 'vl4pose']: self.aux_net.eval() self.j2i = {'head': 0, 'neck': 1, 'lsho': 2, 'lelb': 3, 'lwri': 4, 'rsho': 5, 'relb': 6, 'rwri': 7, 'lhip': 8, 'lknee': 9, 'lankl': 10, 'rhip': 11, 'rknee': 12, 'rankl': 13} self.i2j = {0: 'head', 1: 'neck', 2: 'lsho', 3: 'lelb', 4: 'lwri', 5: 'rsho', 6: 'relb', 7: 'rwri', 8: 'lhip', 9: 'lknee', 10: 'lankl', 11: 'rhip', 12: 'rknee', 13: 'rankl'} # update j2i, i2j with new joints if conf.dataset['load'] == 'mpii' or conf.dataset['load'] == 'merged': self.j2i['pelvis'] = 14 self.j2i['thorax'] = 15 self.i2j[14] = 'pelvis' self.i2j[15] = 'thorax' # I don't know why I have this assertion, need to check assert self.conf.dataset['load'] != 'merged' def base(self, train, dataset_size): """ Do no visualization """ raise Exception('Base method cannot be visualized') def random(self, train, dataset_size): """ Do no visualization """ raise Exception('Random method cannot be visualized') def coreset_sampling(self, train, dataset_size): """ Sener and Savarese, "Active Learning for Convolutional Neural Networks: A Core-Set Approach" ICLR 2018 https://arxiv.org/abs/1708.00489 """ logging.info('\nVisualizing Core-Set.') def update_distances(cluster_centers, encoding, min_distances=None): ''' Based on: https://github.com/google/active-learning/blob/master/sampling_methods/kcenter_greedy.py Update min distances given cluster centers. Args: cluster_centers: indices of cluster centers only_new: only calculate distance for newly selected points and update min_distances. rest_dist: whether to reset min_distances. ''' if len(cluster_centers) != 0: # Update min_distances for all examples given new cluster center. x = encoding[cluster_centers] dist = pairwise_distances(encoding, x, metric='euclidean') if min_distances is None: min_distances = np.min(dist, axis=1).reshape(-1, 1) else: min_distances = np.minimum(min_distances, dist) return min_distances if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "Core-set requires a pretrained model." annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) assert np.all(train['index'] == np.arange(train['name'].shape[0])) dataset_ = ActiveLearningVizDataLoader(train, indices=np.arange(train['name'].shape[0]), conf=self.conf) coreset_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) ################################################################################################################ # Part 1: Obtain Pose Embeddings ################################################################################################################ pose_encoding = None logging.info('\nComputing Core-Set embeddings.') with torch.no_grad(): for images, _, names, gts, datasets in tqdm(coreset_dataloader): pose_heatmaps_, pose_features = self.pose_model(images) try: pose_encoding = torch.cat((pose_encoding, pose_features['penultimate'].cpu()), dim=0) except TypeError: pose_encoding = pose_features['penultimate'].cpu() pose_encoding = pose_encoding.squeeze().numpy() logging.info('Core-Set embeddings computed.\n') ################################################################################################################ # Part 2: K-Centre Greedy ################################################################################################################ logging.info('\nComputing k-Centre Greedy') core_set_budget = 15 # Hardcoded, we want to see the first 15 diverse samples min_distances = None assert len(annotated_idx) != 0, "No annotations for previous model found, core-set cannot proceeed." min_distances = update_distances(cluster_centers=annotated_idx, encoding=pose_encoding, min_distances=None) display_idx = [] distances_over_time = [] for _ in tqdm(range(unlabelled_idx.shape[0])): ind = np.argmax(min_distances) distances_over_time.append(np.max(min_distances)) # New examples should not be in already selected since those points # should have min_distance of zero to a cluster center. min_distances = update_distances(cluster_centers=[ind], encoding=pose_encoding, min_distances=min_distances) annotated_idx = np.concatenate([annotated_idx, [ind]], axis=0).astype(np.int32) display_idx.append(ind) logging.info('Computed k-Centre Greedy.\n') del pose_encoding ################################################################################################################ # Part 3: Plot max distances over time ################################################################################################################ plt.plot(np.arange(unlabelled_idx.shape[0]), distances_over_time, label='Maximum distance') plt.title('Core-Set: Maximum Distances over n-selections') os.makedirs(os.path.join(self.conf.model['save_path'], 'coreset_visualizations'), exist_ok=True) plt.savefig(fname=os.path.join(self.conf.model['save_path'], 'coreset_visualizations/Distances_CoreSet.jpg'), facecolor='black', edgecolor='black', bbox_inches='tight', dpi=300) plt.close() ################################################################################################################ # Part 4: Visualize Core-Set images ################################################################################################################ dataset_ = ActiveLearningVizDataLoader(train, indices=display_idx[:core_set_budget], conf=self.conf) coreset_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] logging.info('Obtaining images and data for selected samples.') # Disable autograd to speed up inference with torch.no_grad(): for images, _, names, gts, datasets in tqdm(coreset_dataloader): pose_heatmaps_, _ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() logging.info('Images loaded.\n') logging.info('Visualizing images (GT and Prediction)') self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_all) def learning_loss_sampling(self, train, dataset_size): """ Yoo and Kweon, "Learning Loss for Active Learning" CVPR 2019 https://openaccess.thecvf.com/content_CVPR_2019/papers/Yoo_Learning_Loss_for_Active_Learning_CVPR_2019_paper.pdf Shukla and Ahmed, "A Mathematical Analysis of Learning Loss for Active Learning in Regression" CVPR-W 2021 https://openaccess.thecvf.com/content/CVPR2021W/TCV/papers/Shukla_A_Mathematical_Analysis_of_Learning_Loss_for_Active_Learning_in_CVPRW_2021_paper.pdf """ logging.info('Visualizing learning loss.') assert self.conf.model['load'], "Learning loss requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx, conf=self.conf) learnloss_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) ################################################################################################################ # Part 1: Obtain Learning Loss predictions ################################################################################################################ learnloss_pred = None with torch.no_grad(): for images, _, names, gts, datasets in tqdm(learnloss_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) learnloss_pred_ = self._aux_net_inference(pose_features_).squeeze() try: learnloss_pred = torch.cat([learnloss_pred, learnloss_pred_.cpu()], dim=0) except TypeError: learnloss_pred = learnloss_pred_.cpu() learnloss_pred = learnloss_pred.squeeze().numpy() # argsort defaults to ascending new_index = np.arange(learnloss_pred.shape[0]).reshape(-1, 1) learnloss_with_index = np.concatenate([learnloss_pred.reshape(-1, 1), new_index], axis=-1) learnloss_with_index = learnloss_with_index[learnloss_with_index[:, 0].argsort()] # Slice aleatoric for top-5 and bottom-5 images min_learnloss_idx = learnloss_with_index[:15, 1].astype(np.int32) max_learnloss_idx = learnloss_with_index[-15:, 1].astype(np.int32) ################################################################################################################ # Part 2: Visualize images with low loss ################################################################################################################ idx = min_learnloss_idx # Part 2.a: Loading images dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx[idx], conf=self.conf) learnloss_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] with torch.no_grad(): for images, _, names, gts, datasets in tqdm(learnloss_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() names_modified = [] for i in range(15): temp = 'LearningLoss_{}_andName_{}'.format(learnloss_pred[min_learnloss_idx[i]], names_all[i]) names_modified.append(temp) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) ################################################################################################################ # Part 3: Visualize images with high loss ################################################################################################################ idx = max_learnloss_idx # Part 3.a: Loading images dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx[idx], conf=self.conf) learnloss_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] with torch.no_grad(): for images, _, names, gts, datasets in tqdm(learnloss_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() names_modified = [] for i in range(15): temp = 'LearningLoss_{}_andName_{}'.format(learnloss_pred[max_learnloss_idx[i]], names_all[i]) names_modified.append(temp) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) def expected_gradient_length_sampling(self, train, dataset_size): """ Megh Shukla, "Bayesian Uncertainty and Expected Gradient Length - Regression: Two Sides Of The Same Coin?" WACV 2022 https://openaccess.thecvf.com/content/WACV2022/papers/Shukla_Bayesian_Uncertainty_and_Expected_Gradient_Length_-_Regression_Two_Sides_WACV_2022_paper.pdf """ def _probability(pair_dist): ''' Computes P(j|i) using Binary Search :param pairwise_dist: (2D Tensor) pairwise distances between samples --> actual dist, not squared :return: 2D Tensor containing conditional probabilities ''' def calc_probs_perp(lower_bound, upper_bound, pair_dist): sigmas = (lower_bound + upper_bound) / 2 variance = (sigmas ** 2).reshape(-1, 1) scaled_pair_dist_neg = -pair_dist / (2 * variance) probs_unnormalized = torch.exp(scaled_pair_dist_neg) probs_unnormalized = torch.clamp(probs_unnormalized, min=1e-20, max=1.) softmax = probs_unnormalized / torch.sum(probs_unnormalized, dim=1, keepdim=True) softmax = torch.clamp(softmax, min=1e-30, max=1.) entropy = shannon_entropy(softmax) perplexity_hat = torch.pow(2 * torch.ones(n_samples), entropy) return perplexity_hat, softmax def condition(perplexity_hat, perplexity): mask = torch.lt(torch.abs(perplexity_hat - perplexity), TOLERANCE) return False in mask global PERPLEXITY, TOLERANCE, n_samples tries = 100 n_samples = pair_dist.shape[0] PERPLEXITY = self.conf.active_learning['egl']['perplexity'] TOLERANCE = self.conf.active_learning['egl']['tolerance'] * torch.ones(n_samples) pair_dist = pair_dist ** 2 lower = torch.zeros(n_samples) upper = (torch.max(torch.max(pair_dist), torch.max(pair_dist**0.5))) * torch.ones(n_samples) * 5 perplexity = PERPLEXITY * torch.ones(n_samples) perplexity_hat, probs = calc_probs_perp(lower, upper, pair_dist) while condition(perplexity_hat, perplexity): if tries < 0: break tries -= 1 mask_gt = torch.gt(perplexity_hat - perplexity, TOLERANCE).type(torch.float32) upper_update = upper - torch.mul(mask_gt, (upper - lower) / 2) mask_lt = torch.lt(perplexity_hat - perplexity, -TOLERANCE).type(torch.float32) lower_update = lower + torch.mul(mask_lt, (upper - lower) / 2) upper = upper_update lower = lower_update perplexity_hat, probs = calc_probs_perp(lower, upper, pair_dist) del PERPLEXITY, TOLERANCE, n_samples return probs logging.info('VISUALIZING expected gradient length sampling.') # Setup -------------------------------------------------------------------------------------------------------- # Load indices of previously annotated data assert self.conf.model['load'], "Expected Gradient Length requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index']) - set(annotated_idx))) ################################################################################################################ # Part 1: Obtain embeddings for labelled data ################################################################################################################ dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, conf=self.conf, indices=annotated_idx) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) logging.info('Computing embeddings for labelled images.') # Disable autograd to speed up inference with torch.no_grad(): pose_encoding_L = None for images, _, names, gts, datasets in tqdm(egl_dataloader): _, pose_features_ = self.pose_model(images) try: pose_encoding_L = torch.cat((pose_encoding_L, pose_features_['penultimate'].cpu()), dim=0) except TypeError: pose_encoding_L = pose_features_['penultimate'].cpu() ################################################################################################################ # Part 2: Obtain embeddings for unlabelled data ################################################################################################################ logging.info('Computing embeddings for unlabelled data') dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, conf=self.conf, indices=unlabelled_idx) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Disable autograd to speed up inference with torch.no_grad(): pose_encoding_U = None names_U = [] gts_U = None hm_uv_stack_U = [] datasets_U = [] for images, _, names, gts, datasets in tqdm(egl_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack_U.append(hm_uv) names_U.append(names[i]) datasets_U.append(datasets[i]) try: pose_encoding_U = torch.cat((pose_encoding_U, pose_features_['penultimate'].cpu()), dim=0) # GAP over the 4x4 lyr gts_U = torch.cat([gts_U, gts], dim=0) except TypeError: pose_encoding_U = pose_features_['penultimate'].cpu() gts_U = gts hm_uv_stack_U = np.stack(hm_uv_stack_U, axis=0) names_U = np.array(names_U) datasets_U = np.array(datasets_U) gts_U = gts_U.numpy() ################################################################################################################ # Part 3: Compute pairwise distances ################################################################################################################ logging.info('Computing pairwise probabilities data') with torch.no_grad(): pair_dist = torch.cdist(pose_encoding_U, pose_encoding_L, p=2) # Unlabelled[i] to Labelled[j] p_i_given_j = _probability(pair_dist) k = self.conf.active_learning['egl']['k'] assert len(p_i_given_j.shape) == 2, "Not a 2-dimensional tensor" vals_L, idx_L = torch.topk(p_i_given_j, k=k, dim=1, sorted=True, largest=True) del pose_encoding_L, pose_encoding_U, pair_dist, p_i_given_j ################################################################################################################ # Part 4: Compute expected gradient length ################################################################################################################ logging.info('Computing the gradient between the unlabelled and labelled images.') pose_gradients_nbrs = torch.zeros(size=(unlabelled_idx.shape[0], k), dtype=torch.float32).to(vals_L.device) assert vals_L.shape == pose_gradients_nbrs.shape autograd_hacks.add_hooks(self.pose_model) unlabelled_dataset = ActiveLearningVizDataLoader(dataset_dict=train, indices=unlabelled_idx, conf=self.conf) for i in tqdm(range(len(unlabelled_dataset))): neighbors_ = ActiveLearningVizDataLoader(dataset_dict=train, indices=annotated_idx[idx_L[i]], conf=self.conf) neighbors_dataloader = torch.utils.data.DataLoader(neighbors_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Keep on GPU, collect heatmaps for the neighbours with torch.no_grad(): hm_L = None for images_l, _, _, _, _ in neighbors_dataloader: hm_l, _ = self.pose_model(images_l) try: hm_L = torch.cat((hm_L, hm_l), dim=0) except TypeError: hm_L = hm_l # Compute gradient wrt these neighbors image, _, _, _, _ = unlabelled_dataset.__getitem__(i) images = torch.cat(k * [image.unsqueeze(0)], dim=0) hm_U, _ = self.pose_model(images) loss = heatmap_loss(hm_U, hm_L, egl=True).mean() loss.backward() autograd_hacks.compute_grad1(model=self.pose_model, loss_type='mean') with torch.no_grad(): grads = torch.zeros((k,), dtype=torch.float32) for param in self.pose_model.parameters(): try: # Sum of squared gradients for each batch element grads = grads.to(param.grad1.device) grads += (param.grad1 ** 2).sum(dim=list(range(len(param.grad1.shape)))[1:]) except AttributeError: continue pose_gradients_nbrs[i] = grads.to(pose_gradients_nbrs.device) # Removing gradients due to previous image self.pose_model.zero_grad() autograd_hacks.clear_backprops(self.pose_model) autograd_hacks.remove_hooks(self.pose_model) egl = (vals_L * pose_gradients_nbrs).sum(dim=1).squeeze() ################################################################################################################ # Part 5: Visualize top-K images and their nearest neighbors ################################################################################################################ vals_topK, idx_topK = torch.topk(egl, k=15, sorted=True, largest=True) idx_topK = idx_topK.numpy() vals_topK = vals_topK.numpy() # Part 5.a: topk actual imgs visualization dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, conf=self.conf, indices=unlabelled_idx[idx_topK]) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) logging.info('Generating top-K images') images_U = None # Disable autograd to speed up inference with torch.no_grad(): for images, _, _, _, datasets in tqdm(egl_dataloader): try: images_U = torch.cat([images_U, images], dim=0) except TypeError: images_U = images images_U = images_U.numpy() name_modified = [] for i in range(15): #name_ = train['name'][unlabelled_idx_topK[i]] name_ = names_U[idx_topK[i]] name_modified.append('top_{}_egl_{}_name_{}.png'.format(i + 1, vals_topK[i], name_)) self._visualize_predictions(image=images_U, name=names_U[idx_topK], dataset=datasets_U[idx_topK], gt=gts_U[idx_topK], pred=hm_uv_stack_U[idx_topK], string=np.array(name_modified)) # Part 5.b: topk nbrs logging.info('Generating neighbors for top-K images') images_L = None gts_L = None names_L = [] hm_uv_stack_L = [] datasets_L = [] name_modified = [] for i in tqdm(range(15)): # Top fifteen samples with highest EGL # Top 5 neighbors for each joint dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, conf=self.conf, indices=annotated_idx[idx_L[idx_topK[i], :5]]) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2, drop_last=False) nbr = 0 with torch.no_grad(): for images, _, names, gts, datasets in egl_dataloader: pose_heatmaps_, pose_features_ = self.pose_model(images) for j in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[j]) hm_uv_stack_L.append(hm_uv) names_L.append(names[j]) datasets_L.append(datasets[j]) name_modified.append( 'top_{}_nbr_{}_pji_{}_name_{}.png'.format(i + 1, nbr + j, vals_L[idx_topK[i], nbr+j], names[j])) nbr += images.shape[0] try: images_L = torch.cat([images_L, images], dim=0) gts_L = torch.cat([gts_L, gts], dim=0) except TypeError: images_L = images gts_L = gts images_L = images_L.numpy() hm_uv_stack_L = np.stack(hm_uv_stack_L, axis=0) names_L = np.array(names_L) datasets_L = np.array(datasets_L) gts_L = gts_L.numpy() name_modified = np.array(name_modified) self._visualize_predictions(image=images_L, name=names_L, dataset=datasets_L, gt=gts_L, pred=hm_uv_stack_L, string=name_modified) ################################################################################################################ # Part 6: Visualize bottom-K images and their nearest neighbors ################################################################################################################ vals_topK, idx_topK = torch.topk(-egl, k=15, sorted=True, largest=True) idx_topK = idx_topK.numpy() vals_topK = vals_topK.numpy() # Generate images dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, conf=self.conf, indices=unlabelled_idx[idx_topK]) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) logging.info('Generating Bottom-K images') images_U = None # Disable autograd to speed up inference with torch.no_grad(): for images, _, _, _, datasets in tqdm(egl_dataloader): try: images_U = torch.cat([images_U, images], dim=0) except TypeError: images_U = images images_U = images_U.numpy() name_modified = [] for i in range(15): #name_ = train['name'][unlabelled_idx_topK[i]] name_ = names_U[idx_topK[i]] name_modified.append('bottom_{}_egl_{}_name_{}.png'.format(i + 1, vals_topK[i], name_)) self._visualize_predictions(image=images_U, name=names_U[idx_topK], dataset=datasets_U[idx_topK], gt=gts_U[idx_topK], pred=hm_uv_stack_U[idx_topK], string=np.array(name_modified)) # Part 5.d: bottom-K nbrs logging.info('Generating neighbors for bottom-K images') images_L = None gts_L = None names_L = [] hm_uv_stack_L = [] datasets_L = [] name_modified = [] for i in tqdm(range(15)): # Top fifteen samples with highest EGL # Top 5 neighbors for each joint dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, conf=self.conf, indices=annotated_idx[idx_L[idx_topK[i], :5]]) egl_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2, drop_last=False) nbr = 0 with torch.no_grad(): for images, _, names, gts, datasets in egl_dataloader: pose_heatmaps_, pose_features_ = self.pose_model(images) for j in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[j]) hm_uv_stack_L.append(hm_uv) names_L.append(names[j]) datasets_L.append(datasets[j]) name_modified.append( 'bottom_{}_nbr_{}_pji_{}_name_{}.png'.format(i + 1, nbr + j, vals_L[idx_topK[i], nbr+j], names[j])) nbr += images.shape[0] try: images_L = torch.cat([images_L, images], dim=0) gts_L = torch.cat([gts_L, gts], dim=0) except TypeError: images_L = images gts_L = gts images_L = images_L.numpy() hm_uv_stack_L = np.stack(hm_uv_stack_L, axis=0) names_L = np.array(names_L) datasets_L = np.array(datasets_L) gts_L = gts_L.numpy() name_modified = np.array(name_modified) self._visualize_predictions(image=images_L, name=names_L, dataset=datasets_L, gt=gts_L, pred=hm_uv_stack_L, string=name_modified) def multipeak_entropy(self, train, dataset_size): """ Liu and Ferrari, "Active Learning for Human Pose Estimation" ICCV 2017 https://openaccess.thecvf.com/content_ICCV_2017/papers/Liu_Active_Learning_for_ICCV_2017_paper.pdf """ logging.info('VISUALIZING multi-peak entropy sampling.') if self.conf.resume_training: return np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) assert self.conf.model['load'], "Multipeak entropy was called without a base model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) unlabelled_idx = np.array(list(set(train['index']) - set(annotated_idx))) # Multi-peak entropy only over the unlabelled set of images dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, indices=unlabelled_idx, conf=self.conf) mpe_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) ################################################################################################################ # Part 1: Computing entropy for unlabelled data ################################################################################################################ mpe_value_per_img = [] logging.info('Computing entropy.') with torch.no_grad(): for images, _, _, _, _ in tqdm(mpe_dataloader): pose_heatmaps_, _ = self.pose_model(images) pose_heatmaps_ = pose_heatmaps_.detach().cpu().numpy()[:, -1, :, :, :] for i in range(pose_heatmaps_.shape[0]): entropy = 0 normalize = 0 for hm in range(pose_heatmaps_.shape[1]): loc = peak_local_max(pose_heatmaps_[i, hm], min_distance=5, num_peaks=5, exclude_border=False) peaks = pose_heatmaps_[i, hm][loc[:, 0], loc[:, 1]] if peaks.shape[0] > 0: peaks = softmax_fn(peaks) entropy += entropy_fn(peaks) normalize += 1 mpe_value_per_img.append(entropy / normalize) ################################################################################################################ # Part 2: Finding images with lowest and highest entropy ################################################################################################################ mpe_value_per_img = np.array(mpe_value_per_img) mpe_with_index = np.concatenate([mpe_value_per_img.reshape(-1, 1), unlabelled_idx.reshape(-1, 1)], axis=-1) mpe_with_index = mpe_with_index[mpe_with_index[:, 0].argsort()] # Slice multipeak entropy for top-15 and bottom-15 images min_mpe_idx = mpe_with_index[:15, 1].astype(np.int32) max_mpe_idx = mpe_with_index[-15:, 1].astype(np.int32) ################################################################################################################ # Part 3: Visualizing images with lowest entropy ################################################################################################################ logging.info('Visualizing samples with low entropy') idx = min_mpe_idx dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, indices=idx, conf=self.conf) mpe_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) pose_heatmaps = None images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] # Disable autograd to speed up inference with torch.no_grad(): for images, _, names, gts, datasets in mpe_dataloader: pose_heatmaps_, _ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: pose_heatmaps = torch.cat((pose_heatmaps, pose_heatmaps_[:, -1, :, :, :].cpu()), dim=0) images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: pose_heatmaps = pose_heatmaps_[:, -1, :, :, :].cpu() images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() pose_heatmaps = pose_heatmaps.squeeze().numpy() for i in range(15): img = images_all[i] hm = pose_heatmaps[i] name = names_all[i] for j in range(hm.shape[0]): plt.imshow(img) plt.imshow(cv2.resize(hm[j], dsize=(256, 256), interpolation=cv2.INTER_CUBIC), alpha=.5) plt.title('{}'.format(self.i2j[j])) os.makedirs(os.path.join(self.conf.model['save_path'], 'entropy_visualizations/images_entropy/minimum'), exist_ok=True) plt.savefig(fname=os.path.join(self.conf.model['save_path'], 'entropy_visualizations/images_entropy/minimum/{}_{}.jpg'.format(name, self.i2j[j])), facecolor='black', edgecolor='black', bbox_inches='tight', dpi=300) plt.close() names_modified = [] for i in range(15): names_modified.append('MPE_{}_andName_{}'.format(mpe_with_index[i, 0], names_all[i])) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) ################################################################################################################ # Part 4: Visualizing images with highest entropy ################################################################################################################ logging.info('Visualizing samples with high entropy') idx = max_mpe_idx dataset_ = ActiveLearningVizDataLoader(dataset_dict=train, indices=idx, conf=self.conf) mpe_dataloader = torch.utils.data.DataLoader(dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) pose_heatmaps = None images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] # Disable autograd to speed up inference with torch.no_grad(): for images, _, names, gts, datasets in mpe_dataloader: pose_heatmaps_, _ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: pose_heatmaps = torch.cat((pose_heatmaps, pose_heatmaps_[:, -1, :, :, :].cpu()), dim=0) images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: pose_heatmaps = pose_heatmaps_[:, -1, :, :, :].cpu() images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() pose_heatmaps = pose_heatmaps.squeeze().numpy() for i in range(15): img = images_all[i] hm = pose_heatmaps[i] name = names_all[i] for j in range(hm.shape[0]): plt.imshow(img) plt.imshow(cv2.resize(hm[j], dsize=(256, 256), interpolation=cv2.INTER_CUBIC), alpha=.5) plt.title('{}'.format(self.i2j[j])) os.makedirs(os.path.join(self.conf.model['save_path'], 'entropy_visualizations/images_entropy/maximum'), exist_ok=True) plt.savefig(fname=os.path.join(self.conf.model['save_path'], 'entropy_visualizations/images_entropy/maximum/{}_{}.jpg'.format(name, self.i2j[j])), facecolor='black', edgecolor='black', bbox_inches='tight', dpi=300) plt.close() names_modified = [] for i in range(15): names_modified.append('MPE_{}_andName_{}'.format(mpe_with_index[-15 + i, 0], names_all[i])) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) def aleatoric_uncertainty(self, train, dataset_size): """ Kendall and Gal, "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?" NeurIPS 2017 https://proceedings.neurips.cc/paper/2017/hash/2650d6089a6d640c5e85b2b88265dc2b-Abstract.html """ logging.info('VISUALIZING Uncertainty: Kendall and Gal sampling.') assert self.conf.model['load'], "Aleatoric uncertainty requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx, conf=self.conf) aleatoric_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) ################################################################################################################ # Part 1: Computing aleatoric uncertainty for unlabelled data ################################################################################################################ aleatoric_pred = None # Part 1: Active Learning logging.info('Computing aleatoric uncertainty.') with torch.no_grad(): for images, _, names, gts, datasets in tqdm(aleatoric_dataloader): _, pose_features_ = self.pose_model(images) aleatoric_pred_ = self._aux_net_inference(pose_features_) aleatoric_pred_ = aleatoric_pred_.squeeze() try: aleatoric_pred = torch.cat([aleatoric_pred, aleatoric_pred_.cpu()], dim=0) except TypeError: aleatoric_pred = aleatoric_pred_.cpu() aleatoric_pred_copy = aleatoric_pred.mean(dim=-1) ################################################################################################################ # Part 2: Sort images based on aleatoric uncertainty ################################################################################################################ # argsort defaults to ascending aleatoric_with_index = np.concatenate([aleatoric_pred_copy.numpy().reshape(-1, 1), unlabelled_idx.reshape(-1, 1)], axis=-1) aleatoric_with_index = aleatoric_with_index[aleatoric_with_index[:, 0].argsort()] # Slice aleatoric for top-5 and bottom-5 images min_aleatoric_idx = aleatoric_with_index[:15, 1].astype(np.int32) max_aleatoric_idx = aleatoric_with_index[-15:, 1].astype(np.int32) ################################################################################################################ # Part 3: Visualize skeletons for samples with low aleatoric uncertainty ################################################################################################################ logging.info('Visualizing samples with low aleatoric uncertainty') idx = min_aleatoric_idx dataset_ = ActiveLearningVizDataLoader(train, indices=idx, conf=self.conf) aleatoric_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Compile all together aleatoric_pred = None images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] # Disable autograd to speed up inference with torch.no_grad(): for images, _, names, gts, datasets in aleatoric_dataloader: pose_heatmaps_, pose_features_ = self.pose_model(images) aleatoric_pred_ = self._aux_net_inference(pose_features_) aleatoric_pred_ = aleatoric_pred_.squeeze() for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: aleatoric_pred = torch.cat([aleatoric_pred, aleatoric_pred_.cpu()], dim=0) images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: aleatoric_pred = aleatoric_pred_.cpu() images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() aleatoric_pred_copy = aleatoric_pred.mean(dim=-1) names_modified = [] for i in range(15): temp = 'Aleatoric_{}andName_{}'.format(aleatoric_pred_copy[i], names_all[i]) names_modified.append(temp) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) ################################################################################################################ # Part 4: Visualize heatmaps for samples with low aleatoric uncertainty ################################################################################################################ aleatoric_hm_all = [] for i in range(15): aleatoric_hm = np.zeros([64, 64]) for j in range(self.conf.experiment_settings['num_hm']): # after idx slicing: hm_uv_stack = 5 x 1 x 14 x 3 pred = hm_uv_stack[i, 0, j] log_var = aleatoric_pred[i, j] # print('Mean: {}'.format(mean)) # print('Var: {}'.format(np.exp(log_var))) target = pred[:2] / np.array([(256 - 1) / (64 - 1), ((256 - 1) / (64 - 1))]) # print('Target: {}'.format(target)) aleatoric_hm_jnt = self._draw_gaussian(target, log_var) aleatoric_hm = np.maximum(aleatoric_hm, aleatoric_hm_jnt) # print() # RESIZE HM aleatoric_hm_all.append(cv2.resize(aleatoric_hm, dsize=(256, 256), interpolation=cv2.INTER_CUBIC)) aleatoric_hm_all = np.stack(aleatoric_hm_all, axis=0) names_modified = [] for i in range(15): names_modified.append('SkeletonViz_Name_{}'.format(names_all[i])) names_modified = np.array(names_modified) self._visualize_predictions(image=aleatoric_hm_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) ################################################################################################################ # Part 5: Visualize skeletons for samples with high aleatoric uncertainty ################################################################################################################ logging.info('Visualizing samples with high aleatoric uncertainty') idx = max_aleatoric_idx dataset_ = ActiveLearningVizDataLoader(train, indices=idx, conf=self.conf) aleatoric_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Compile all together aleatoric_pred = None images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] # Disable autograd to speed up inference with torch.no_grad(): for images, _, names, gts, datasets in aleatoric_dataloader: pose_heatmaps_, pose_features_ = self.pose_model(images) aleatoric_pred_ = self._aux_net_inference(pose_features_) aleatoric_pred_ = aleatoric_pred_.squeeze() for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: aleatoric_pred = torch.cat([aleatoric_pred, aleatoric_pred_.cpu()], dim=0) images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: aleatoric_pred = aleatoric_pred_.cpu() images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() aleatoric_pred_copy = aleatoric_pred.mean(dim=-1) names_modified = [] for i in range(15): temp = 'Aleatoric_{}andName_{}'.format(aleatoric_pred_copy[i], names_all[i]) names_modified.append(temp) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) ################################################################################################################ # Part 4: Visualize heatmaps for samples with high aleatoric uncertainty ################################################################################################################ aleatoric_hm_all = [] for i in range(15): aleatoric_hm = np.zeros([64, 64]) for j in range(self.conf.experiment_settings['num_hm']): # after idx slicing: hm_uv_stack = 5 x 1 x 14 x 3 pred = hm_uv_stack[i, 0, j] log_var = aleatoric_pred[i, j] # print('Mean: {}'.format(mean)) # print('Var: {}'.format(np.exp(log_var))) target = pred[:2] / np.array([(256 - 1) / (64 - 1), ((256 - 1) / (64 - 1))]) # print('Target: {}'.format(target)) aleatoric_hm_jnt = self._draw_gaussian(target, log_var) aleatoric_hm = np.maximum(aleatoric_hm, aleatoric_hm_jnt) # print() # RESIZE HM aleatoric_hm_all.append(cv2.resize(aleatoric_hm, dsize=(256, 256), interpolation=cv2.INTER_CUBIC)) aleatoric_hm_all = np.stack(aleatoric_hm_all, axis=0) names_modified = [] for i in range(15): names_modified.append('SkeletonViz_Name_{}'.format(names_all[i])) names_modified = np.array(names_modified) self._visualize_predictions(image=aleatoric_hm_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) def vl4pose(self, train, dataset_size): """ Shukla et al., "VL4Pose: Active Learning Through Out-Of-Distribution Detection For Pose Estimation" BMVC 2022 https://bmvc2022.mpi-inf.mpg.de/610/ """ logging.info('Visualizing: VL4Pose Sampling.') assert self.conf.model['load'], "VL4Pose requires a previously trained model" annotated_idx = np.load(os.path.join(self.conf.model['load_path'], 'model_checkpoints/annotation.npy')) # Set of indices not annotated unlabelled_idx = np.array(list(set(train['index'])-set(annotated_idx))) # links definition if self.conf.dataset['load'] == 'mpii': links = [[self.j2i['head'], self.j2i['neck']], [self.j2i['neck'], self.j2i['thorax']], [self.j2i['thorax'], self.j2i['pelvis']], [self.j2i['thorax'], self.j2i['lsho']], [self.j2i['lsho'], self.j2i['lelb']], [self.j2i['lelb'], self.j2i['lwri']], [self.j2i['thorax'], self.j2i['rsho']], [self.j2i['rsho'], self.j2i['relb']], [self.j2i['relb'], self.j2i['rwri']], [self.j2i['pelvis'], self.j2i['lhip']], [self.j2i['lhip'], self.j2i['lknee']], [self.j2i['lknee'], self.j2i['lankl']], [self.j2i['pelvis'], self.j2i['rhip']], [self.j2i['rhip'], self.j2i['rknee']], [self.j2i['rknee'], self.j2i['rankl']]] else: links = [[self.j2i['head'], self.j2i['neck']], [self.j2i['neck'], self.j2i['lsho']], [self.j2i['lsho'], self.j2i['lelb']], [self.j2i['lelb'], self.j2i['lwri']], [self.j2i['neck'], self.j2i['rsho']], [self.j2i['rsho'], self.j2i['relb']], [self.j2i['relb'], self.j2i['rwri']], [self.j2i['lsho'], self.j2i['lhip']], [self.j2i['lhip'], self.j2i['lknee']], [self.j2i['lknee'], self.j2i['lankl']], [self.j2i['rsho'], self.j2i['rhip']], [self.j2i['rhip'], self.j2i['rknee']], [self.j2i['rknee'], self.j2i['rankl']]] dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx, conf=self.conf) vl4pose_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=1) ################################################################################################################ # Part 1: Computing images where VL4Pose has changed the pose estimator's prediction ################################################################################################################ logging.info('Computing images where poses have changed.') with torch.no_grad(): has_pose_changed = None vl4pose_refinement = None for images, _, names, gts, datasets in tqdm(vl4pose_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) likelihood_pred_ = self._aux_net_inference(pose_features_).reshape(images.shape[0], len(links), 2) keypoint_compute = Keypoint_ParallelWrapper( hm=pose_heatmaps_[:, -1, :, :, :].cpu().numpy(), param=likelihood_pred_.cpu().numpy(), j2i=self.j2i, i2j=self.i2j, links=links, vl4pose_config=self.conf.active_learning['vl4pose'], function=np.max) keypoint_dataloader = torch.utils.data.DataLoader( keypoint_compute, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Poses evaluated according to VL4Pose for likelihoods, vl4pose_refinement_, has_pose_changed_ in keypoint_dataloader: try: vl4pose_refinement = torch.cat((vl4pose_refinement, vl4pose_refinement_), dim=0) has_pose_changed.extend(has_pose_changed_.tolist()) except TypeError: vl4pose_refinement = vl4pose_refinement_ has_pose_changed = has_pose_changed_.tolist() vl4pose_refinement = vl4pose_refinement.numpy() vl4pose_refinement *= np.array([4.0476, 4.0476, 1]) # Part 1.b: Collect the corresponding images logging.info('Collecting data where {} poses have changed.'.format(sum(has_pose_changed))) dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx[has_pose_changed], conf=self.conf) vl4pose_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=1) images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] for images, _, names, gts, datasets in tqdm(vl4pose_dataloader): pose_heatmaps_, _ = self.pose_model(images) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) except TypeError: images_all = images gts_all = gts images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() # Part 1.c: For these images visualize the poses logging.info('Visualizing images where pose has changed.') self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=hm_uv_stack, pred=vl4pose_refinement[has_pose_changed], string=names_all) del vl4pose_refinement, has_pose_changed, images_all, names_all, datasets_all, hm_uv_stack ################################################################################################################ # Part 2: Computing minimum and maximum expected likelihood ################################################################################################################ logging.info('Computing images with maximum and minimum likelihood.') dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx, conf=self.conf) vl4pose_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=1) with torch.no_grad(): max_likelihood = None for images, _, names, gts, datasets in tqdm(vl4pose_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) likelihood_pred_ = self._aux_net_inference(pose_features_).reshape(images.shape[0], len(links), 2) keypoint_compute = Keypoint_ParallelWrapper( hm=pose_heatmaps_[:, -1, :, :, :].cpu().numpy(), param=likelihood_pred_.cpu().numpy(), j2i=self.j2i, i2j=self.i2j, links=links, vl4pose_config=self.conf.active_learning['vl4pose'], function=np.sum) keypoint_dataloader = torch.utils.data.DataLoader( keypoint_compute, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=2) # Poses evaluated according to VL4Pose for likelihoods, _, _ in keypoint_dataloader: try: max_likelihood = torch.cat((max_likelihood, likelihoods.squeeze()), dim=0) except TypeError: max_likelihood = likelihoods.squeeze() max_likelihood = max_likelihood.numpy() new_index = np.arange(max_likelihood.shape[0]) loglikelihood_with_index = np.concatenate([max_likelihood.reshape(-1, 1), new_index.reshape(-1, 1)], axis=-1) loglikelihood_with_index = loglikelihood_with_index[loglikelihood_with_index[:, 0].argsort()] # Slice images, heatmaps, likelihood and parameters for top-5 and bottom-5 images min_likelihood_idx = loglikelihood_with_index[:15, 1].astype(np.int32) max_likelihood_idx = loglikelihood_with_index[-15:, 1].astype(np.int32) ################################################################################################################ # Part 3: Visualizing images with minimum expected likelihood ################################################################################################################ logging.info('Visualizing images with minimum likelihood.') idx = min_likelihood_idx # Collect images for these indices dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx[idx], conf=self.conf) vl4pose_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=1) with torch.no_grad(): images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] likelihood_params = None for images, _, names, gts, datasets in tqdm(vl4pose_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) likelihood_pred_ = self._aux_net_inference(pose_features_) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) likelihood_params = torch.cat([likelihood_params, likelihood_pred_.cpu().reshape(images.shape[0], len(links), 2)], dim=0) except TypeError: images_all = images gts_all = gts likelihood_params = likelihood_pred_.cpu().reshape(images.shape[0], len(links), 2) images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() likelihood_params = likelihood_params.numpy() names_modified = [] for i in range(15): temp = 'LogLikelihood_{}_andName_{}'.format(max_likelihood[min_likelihood_idx[i]], names_all[i]) names_modified.append(temp) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) del images_all # Part 3.b: Next, prepare conditional heatmaps logging.info('Visualizing heatmaps for minimum likelihood.') conditional_hm_all = [] for i in range(15): conditional_hm = np.zeros([64, 64]) for j, link in enumerate(links): pred1 = hm_uv_stack[i, 0, link[0]] pred2 = hm_uv_stack[i, 0, link[1]] mean = likelihood_params[i, j, 0] log_var = likelihood_params[i, j, 1] target = self._find_point_along_line(source_pt=pred1, dest_pt=pred2, magnitude=mean) conditional_hm_link = self._draw_gaussian(target, log_var) conditional_hm = np.maximum(conditional_hm, conditional_hm_link) #RESIZE HM conditional_hm_all.append(cv2.resize(conditional_hm, dsize=(256, 256), interpolation=cv2.INTER_CUBIC)) conditional_hm_all = np.stack(conditional_hm_all, axis=0) names_modified = [] for i in range(15): names_modified.append('SkeletonVizandName{}'.format(names_all[i])) names_modified = np.array(names_modified) self._visualize_predictions(image=conditional_hm_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) ################################################################################################################ # Part 4: Visualizing images with maximum expected likelihood ################################################################################################################ logging.info('Visualizing images with maximum likelihood.') idx = max_likelihood_idx # Collect images for these indices dataset_ = ActiveLearningVizDataLoader(train, indices=unlabelled_idx[idx], conf=self.conf) vl4pose_dataloader = torch.utils.data.DataLoader( dataset_, batch_size=self.conf.experiment_settings['batch_size'], shuffle=False, num_workers=1) with torch.no_grad(): images_all = None gts_all = None names_all = [] hm_uv_stack = [] datasets_all = [] likelihood_params = None for images, _, names, gts, datasets in tqdm(vl4pose_dataloader): pose_heatmaps_, pose_features_ = self.pose_model(images) likelihood_pred_ = self._aux_net_inference(pose_features_) for i in range(images.shape[0]): hm_uv = self._estimate_uv(hm_array=pose_heatmaps_[:, -1].cpu().numpy()[i]) hm_uv_stack.append(hm_uv) names_all.append(names[i]) datasets_all.append(datasets[i]) try: images_all = torch.cat([images_all, images], dim=0) gts_all = torch.cat([gts_all, gts], dim=0) likelihood_params = torch.cat([likelihood_params, likelihood_pred_.cpu().reshape(images.shape[0], len(links), 2)], dim=0) except TypeError: images_all = images gts_all = gts likelihood_params = likelihood_pred_.cpu().reshape(images.shape[0], len(links), 2) images_all = images_all.numpy() hm_uv_stack = np.stack(hm_uv_stack, axis=0) names_all = np.array(names_all) datasets_all = np.array(datasets_all) gts_all = gts_all.numpy() likelihood_params = likelihood_params.numpy() names_modified = [] for i in range(15): temp = 'LogLikelihood_{}_andName_{}'.format(max_likelihood[max_likelihood_idx[i]], names_all[i]) names_modified.append(temp) names_modified = np.array(names_modified) self._visualize_predictions(image=images_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) del images_all # Part 4.b: Next, prepare conditional heatmaps logging.info('Visualizing heatmaps for maximum likelihood.') conditional_hm_all = [] for i in range(15): conditional_hm = np.zeros([64, 64]) for j, link in enumerate(links): pred1 = hm_uv_stack[i, 0, link[0]] pred2 = hm_uv_stack[i, 0, link[1]] mean = likelihood_params[i, j, 0] log_var = likelihood_params[i, j, 1] target = self._find_point_along_line(source_pt=pred1, dest_pt=pred2, magnitude=mean) conditional_hm_link = self._draw_gaussian(target, log_var) conditional_hm = np.maximum(conditional_hm, conditional_hm_link) #RESIZE HM conditional_hm_all.append(cv2.resize(conditional_hm, dsize=(256, 256), interpolation=cv2.INTER_CUBIC)) conditional_hm_all = np.stack(conditional_hm_all, axis=0) names_modified = [] for i in range(15): names_modified.append('SkeletonVizandName{}'.format(names_all[i])) names_modified = np.array(names_modified) self._visualize_predictions(image=conditional_hm_all, name=names_all, dataset=datasets_all, gt=gts_all, pred=hm_uv_stack, string=names_modified) def _draw_gaussian(self, pt, log_var, hm_shape=(64, 64)): """ :param pt: :param log_var: :param hm_shape: :return: """ im = np.zeros(hm_shape, dtype=np.float32) pt_rint = np.rint(pt).astype(int) sigma = (np.exp(log_var))**0.5 # Size of 2D Gaussian window. size = int(math.ceil(6 * sigma)) # Ensuring that size remains an odd number if not size % 2: size += 1 # Generate gaussian, with window=size and variance=sigma u = np.arange(pt_rint[0] - (size // 2), pt_rint[0] + (size // 2) + 1) v = np.arange(pt_rint[1] - (size // 2), pt_rint[1] + (size // 2) + 1) uu, vv = np.meshgrid(u, v, sparse=True) z = (np.exp(-((uu - pt[0]) ** 2 + (vv - pt[1]) ** 2) / (2 * (sigma ** 2)))) * (1/((2*np.pi*sigma*sigma)**0.5)) z = z.T # Identify indices in im that will define the crop area top = max(0, pt_rint[0] - (size//2)) bottom = min(hm_shape[0], pt_rint[0] + (size//2) + 1) left = max(0, pt_rint[1] - (size//2)) right = min(hm_shape[1], pt_rint[1] + (size//2) + 1) im[top:bottom, left:right] = \ z[top - (pt_rint[0] - (size//2)): top - (pt_rint[0] - (size//2)) + (bottom - top), left - (pt_rint[1] - (size//2)): left - (pt_rint[1] - (size//2)) + (right - left)] return im def _visualize_predictions(self, image=None, name=None, dataset=None, gt=None, pred=None, string=None): dataset_viz = {} dataset_viz['img'] = image dataset_viz['name'] = name dataset_viz['dataset'] = dataset dataset_viz['gt'] = gt dataset_viz['pred'] = pred dataset_viz['string'] = string dataset_viz = self._recreate_images(gt=True, pred=True, external=True, ext_data=dataset_viz) visualize_image(dataset_viz, save_dir=self.conf.model['save_path'], bbox=False) def _recreate_images(self, gt=False, pred=False, external=False, ext_data=None): ''' :return: ''' assert gt + pred != 0, "Specify atleast one of GT or Pred" assert external assert ext_data, "ext_dataset can't be none to recreate external datasets" data_split = ext_data # Along with the below entries, we also pass bbox coordinates for each dataset img_dict = {'mpii': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'dataset': [], 'display_string': []}, 'lspet': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'dataset': [], 'display_string': []}, 'lsp': {'img': [], 'img_name': [], 'img_pred': [], 'img_gt': [], 'dataset': [], 'display_string': []}} for i in range(len(data_split['img'])): dataset = data_split['dataset'][i] img_dict[dataset]['img'].append(data_split['img'][i]) img_dict[dataset]['img_name'].append(data_split['name'][i]) img_dict[dataset]['dataset'].append(data_split['dataset'][i]) img_dict[dataset]['display_string'].append(data_split['string'][i]) joint_dict = dict([(self.i2j[i], []) for i in range(self.conf.experiment_settings['num_hm'])]) gt_dict = copy.deepcopy(joint_dict) pred_dict = copy.deepcopy(joint_dict) if gt: for person in range(1): for joint in range(self.conf.experiment_settings['num_hm']): gt_dict[self.i2j[joint]].append(data_split['gt'][i, person, joint]) if pred: for person in range(1): for joint in range(self.conf.experiment_settings['num_hm']): pred_dict[self.i2j[joint]].append(data_split['pred'][i, person, joint]) img_dict[dataset]['img_gt'].append(gt_dict) img_dict[dataset]['img_pred'].append(pred_dict) return img_dict def _aux_net_inference(self, pose_features): extractor = self.conf.architecture['aux_net']['conv_or_avg_pooling'] with torch.no_grad(): if extractor == 'avg': # Transfer to GPU where auxiliary network is stored encodings = pose_features['penultimate'] else: depth = len(self.conf.architecture['aux_net']['spatial_dim']) encodings = torch.cat( [pose_features['feature_{}'.format(i)].reshape( pose_features['feature_{}'.format(i)].shape[0], pose_features['feature_{}'.format(i)].shape[1], -1) for i in range(depth, 0, -1)], dim=2) aux_out = self.aux_net(encodings) return aux_out def _estimate_uv(self, hm_array): ''' Assumes single person :param hm_array: :param pred_placeholder: :return: ''' threshold = 0 joint = np.empty(shape=[1, hm_array.shape[0], 3], dtype=np.float32) # Iterate over each heatmap for jnt_id in range(hm_array.shape[0]): joint[0, jnt_id, :] = uv_from_heatmap(hm=hm_array[jnt_id], threshold=threshold) return joint def _find_point_along_line(self, source_pt, dest_pt, magnitude): downscale = [(256 - 1) / (64 - 1), ((256 - 1) / (64 - 1))] # Ignore visibility flag source_pt = source_pt[:2] / np.array(downscale) dest_pt = dest_pt[:2] / np.array(downscale) direction = dest_pt - source_pt t = magnitude / np.linalg.norm(source_pt - dest_pt) return source_pt + (t * direction) class ActiveLearningVizDataLoader(torch.utils.data.Dataset): def __init__(self, dataset_dict, conf, indices=None): """ :param dataset_dict: :param conf: :param indices: """ self.names = dataset_dict['name'][indices] self.bounding_box = dataset_dict['bbox_coords'][indices] self.dataset = dataset_dict['dataset'][indices] self.gt = dataset_dict['gt'][indices] self.occlusion = conf.experiment_settings['occlusion'] self.hm_shape = [64, 64] self.hm_peak = conf.experiment_settings['hm_peak'] self.xy_to_uv = lambda xy: (xy[1], xy[0]) def __len__(self): return self.dataset.shape[0] def __getitem__(self, item): ''' :param item: :return: ''' root = Path(os.getcwd()).parent mpii_path = os.path.join(root, 'data', 'mpii') lsp_path = os.path.join(root, 'data', 'lsp') lspet_path = os.path.join(root, 'data', 'lspet') dataset = self.dataset[item] name = self.names[item] bounding_box = self.bounding_box[item] gt = self.gt[item] if self.load_all_imgs: image = self.images[item] else: if dataset == 'mpii': image = plt.imread(os.path.join(mpii_path, 'images', '{}.jpg'.format(name.split('_')[0]))) elif dataset == 'lsp': image = plt.imread(os.path.join(lsp_path, 'images', name)) else: image = plt.imread(os.path.join(lspet_path, 'images', name)) # Determine crop img_shape = np.array(image.shape) # Bounding box for the first person [min_x, min_y, max_x, max_y] = bounding_box[0] tl_uv = self.xy_to_uv(np.array([min_x, min_y])) br_uv = self.xy_to_uv(np.array([max_x, max_y])) min_u = tl_uv[0] min_v = tl_uv[1] max_u = br_uv[0] max_v = br_uv[1] centre = np.array([(min_u + max_u) / 2, (min_v + max_v) / 2]) height = max_u - min_u width = max_v - min_v scale = 1.75 window = max(scale * height, scale * width) top_left = np.array([centre[0] - (window / 2), centre[1] - (window / 2)]) bottom_right = np.array([centre[0] + (window / 2), centre[1] + (window / 2)]) top_left = np.maximum(np.array([0, 0], dtype=np.int16), top_left.astype(np.int16)) bottom_right = np.minimum(img_shape.astype(np.int16)[:-1], bottom_right.astype(np.int16)) # Cropping the image and adjusting the ground truth image = image[top_left[0]: bottom_right[0], top_left[1]: bottom_right[1], :] for person in range(gt.shape[0]): for joint in range(gt.shape[1]): gt_uv = self.xy_to_uv(gt[person][joint]) gt_uv = gt_uv - top_left gt[person][joint] = np.concatenate([gt_uv, np.array([gt[person][joint][2]])], axis=0) # Resize the image image, gt = self.resize_image(image, gt, target_size=[256, 256, 3]) heatmaps, joint_exist = heatmap_generator( joints=np.copy(gt), occlusion=self.occlusion, hm_shape=self.hm_shape, img_shape=image.shape) heatmaps = self.hm_peak * heatmaps return torch.tensor(data=image / 256.0, dtype=torch.float32, device='cpu'),\ torch.tensor(data=heatmaps, dtype=torch.float32, device='cpu'), name, gt, dataset def resize_image(self, image_=None, gt=None, target_size=None): ''' :return: ''' # Compute the aspect ratios image_aspect_ratio = image_.shape[0] / image_.shape[1] tgt_aspect_ratio = target_size[0] / target_size[1] # Compare the original and target aspect ratio if image_aspect_ratio > tgt_aspect_ratio: # If target aspect ratio is smaller, scale the first dim scale_factor = target_size[0] / image_.shape[0] else: # If target aspect ratio is bigger or equal, scale the second dim scale_factor = target_size[1] / image_.shape[1] # Compute the padding to fit the target size pad_u = (target_size[0] - int(image_.shape[0] * scale_factor)) pad_v = (target_size[1] - int(image_.shape[1] * scale_factor)) output_img = np.zeros(target_size, dtype=image_.dtype) # Write scaled size in reverse order because opencv resize scaled_size = (int(image_.shape[1] * scale_factor), int(image_.shape[0] * scale_factor)) padding_u = int(pad_u / 2) padding_v = int(pad_v / 2) im_scaled = cv2.resize(image_, scaled_size) # logging.debug('Scaled, pre-padding size: {}'.format(im_scaled.shape)) output_img[padding_u : im_scaled.shape[0] + padding_u, padding_v : im_scaled.shape[1] + padding_v, :] = im_scaled gt *= np.array([scale_factor, scale_factor, 1]).reshape(1, 1, 3) gt[:, :, 0] += padding_u gt[:, :, 1] += padding_v return output_img, gt class Keypoint_ParallelWrapper(torch.utils.data.Dataset): def __init__(self, hm, param, j2i, i2j, links, vl4pose_config, function): self.hm = hm self.param = param self.j2i = j2i self.i2j = i2j self.links = links self.config = vl4pose_config self.function = function def __len__(self): return self.hm.shape[0] def __getitem__(self, i): joints = {} heatmaps = self.hm[i] parameters = self.param[i] # Initialize keypoints for each node for key in self.j2i.keys(): heatmap = heatmaps[self.j2i[key]] loc = peak_local_max(heatmap, min_distance=self.config['min_distance'], num_peaks=self.config['num_peaks'], exclude_border=False) peaks = heatmap[loc[:, 0], loc[:, 1]] peaks = softmax_fn(peaks) joints[key] = Keypoint(name=key, loc=loc, peaks=peaks, function=self.function) # Initialize parent-child relations for k, l in enumerate(self.links): joints[self.i2j[l[0]]].parameters.append(parameters[k]) joints[self.i2j[l[0]]].children.append(joints[self.i2j[l[1]]]) max_ll, trace = joints['head'].run_likelihood() vl4pose_image = [] for j in range(heatmaps.shape[0]): vl4pose_image.append(torch.from_numpy(trace['{}_uv'.format(self.i2j[j])])) vl4pose_image = torch.stack(vl4pose_image, dim=0) vl4pose_image = torch.cat([vl4pose_image, torch.ones(heatmaps.shape[0]).view(-1, 1)], dim=1) vl4pose_image = vl4pose_image.unsqueeze(0) string = '' for jnt in self.j2i.keys(): string += str(trace[jnt].item()) if string != ('0'*len(self.j2i.keys())): has_pose_changed = True else: has_pose_changed = False return max_ll, vl4pose_image, has_pose_changed class Keypoint(object): def __init__(self, name, loc, peaks, function): self.name = name self.loc = loc self.peaks = peaks self.children = [] self.parameters = [] self.function = function def run_likelihood(self): """ :return: """ assert self.name == 'head' likelihood_per_location = [] per_location_trace = [] for location in range(self.loc.shape[0]): log_ll = np.log(self.peaks[location]) per_child_trace = [] for child in range(len(self.children)): child_ll, joint_trace = self.children[child].compute_likelihood_given_parent(self.loc[location], self.parameters[child]) log_ll += child_ll per_child_trace.append(joint_trace) likelihood_per_location.append(log_ll) per_location_trace.append(per_child_trace) likelihood_per_location = np.array(likelihood_per_location) return_trace = {} for child_trace in per_location_trace[np.argmax(likelihood_per_location)]: return_trace.update(child_trace) return_trace[self.name] = np.argmax(likelihood_per_location) return_trace['{}_uv'.format(self.name)] = self.loc[np.argmax(likelihood_per_location)] return self.function(likelihood_per_location), return_trace def compute_likelihood_given_parent(self, parent_location, gaussian_params): """ :param parent_location: :param gaussian_params: :return: """ likelihood_per_location = [] per_location_trace = [] for location in range(self.loc.shape[0]): log_ll = np.log(2 * np.pi) + gaussian_params[1] log_ll += (gaussian_params[0] - np.linalg.norm(parent_location - self.loc[location]))**2 * np.exp(-gaussian_params[1]) log_ll *= -0.5 log_ll += np.log(self.peaks[location]) if len(self.children) == 0: likelihood_per_location.append(log_ll) else: per_child_trace = [] for child in range(len(self.children)): child_ll, joint_trace = self.children[child].compute_likelihood_given_parent(self.loc[location], self.parameters[child]) log_ll += child_ll per_child_trace.append(joint_trace) likelihood_per_location.append(log_ll) per_location_trace.append(per_child_trace) likelihood_per_location = np.array(likelihood_per_location) if len(self.children) == 0: return self.function(likelihood_per_location), {self.name: np.argmax(likelihood_per_location), '{}_uv'.format(self.name): self.loc[np.argmax(likelihood_per_location)]} return_trace = {} for child_trace in per_location_trace[np.argmax(likelihood_per_location)]: return_trace.update(child_trace) return_trace[self.name] = np.argmax(likelihood_per_location) return_trace['{}_uv'.format(self.name)] = self.loc[np.argmax(likelihood_per_location)] return self.function(likelihood_per_location), return_trace
89,213
43.056296
162
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/autograd_hacks.py
""" Credits: https://github.com/cybertronai/autograd-hacks Library for extracting interesting quantites from autograd, see README.md Not thread-safe because of module-level variables Notation: o: number of output classes (exact Hessian), number of Hessian samples (sampled Hessian) n: batch-size do: output dimension (output channels for convolution) di: input dimension (input channels for convolution) Hi: per-example Hessian of matmul, shaped as matrix of [dim, dim], indices have been row-vectorized Hi_bias: per-example Hessian of bias Oh, Ow: output height, output width (convolution) Kh, Kw: kernel height, kernel width (convolution) Jb: batch output Jacobian of matmul, output sensitivity for example,class pair, [o, n, ....] Jb_bias: as above, but for bias A, activations: inputs into current layer B, backprops: backprop values (aka Lop aka Jacobian-vector product) observed at current layer """ from typing import List import torch import torch.nn as nn import torch.nn.functional as F _supported_layers = ['Linear', 'Conv2d'] # Supported layer class types _hooks_disabled: bool = False # work-around for https://github.com/pytorch/pytorch/issues/25723 _enforce_fresh_backprop: bool = False # global switch to catch double backprop errors on Hessian computation def add_hooks(model: nn.Module) -> None: """ Adds hooks to model to save activations and backprop values. The hooks will 1. save activations into param.activations during forward pass 2. append backprops to params.backprops_list during backward pass. Call "remove_hooks(model)" to disable this. Args: model: """ global _hooks_disabled _hooks_disabled = False handles = [] for layer in model.modules(): if _layer_type(layer) in _supported_layers: handles.append(layer.register_forward_hook(_capture_activations)) handles.append(layer.register_backward_hook(_capture_backprops)) # handles.append(layer.register_full_backward_hook(_capture_backprops)) model.__dict__.setdefault('autograd_hacks_hooks', []).extend(handles) def remove_hooks(model: nn.Module) -> None: """ Remove hooks added by add_hooks(model) """ #assert model == 0, "not working, remove this after fix to https://github.com/pytorch/pytorch/issues/25723" if not hasattr(model, 'autograd_hacks_hooks'): print("Warning, asked to remove hooks, but no hooks found") else: for handle in model.autograd_hacks_hooks: handle.remove() del model.autograd_hacks_hooks def disable_hooks() -> None: """ Globally disable all hooks installed by this library. """ global _hooks_disabled _hooks_disabled = True def enable_hooks() -> None: """the opposite of disable_hooks()""" global _hooks_disabled _hooks_disabled = False def is_supported(layer: nn.Module) -> bool: """Check if this layer is supported""" return _layer_type(layer) in _supported_layers def _layer_type(layer: nn.Module) -> str: return layer.__class__.__name__ def _capture_activations(layer: nn.Module, input: List[torch.Tensor], output: torch.Tensor): """Save activations into layer.activations in forward pass""" if _hooks_disabled: return assert _layer_type(layer) in _supported_layers, "Hook installed on unsupported layer, this shouldn't happen" setattr(layer, "activations", input[0].detach()) def _capture_backprops(layer: nn.Module, _input, output): """Append backprop to layer.backprops_list in backward pass.""" global _enforce_fresh_backprop if _hooks_disabled: return if _enforce_fresh_backprop: assert not hasattr(layer, 'backprops_list'), "Seeing result of previous backprop, use clear_backprops(model) to clear" _enforce_fresh_backprop = False if not hasattr(layer, 'backprops_list'): setattr(layer, 'backprops_list', []) layer.backprops_list.append(output[0].detach()) def clear_backprops(model: nn.Module) -> None: """Delete layer.backprops_list in every layer.""" for layer in model.modules(): if hasattr(layer, 'backprops_list'): del layer.backprops_list def compute_grad1(model: nn.Module, loss_type: str = 'mean') -> None: """ Compute per-example gradients and save them under 'param.grad1'. Must be called after loss.backprop() Args: model: loss_type: either "mean" or "sum" depending whether backpropped loss was averaged or summed over batch """ assert loss_type in ('sum', 'mean') for layer in model.modules(): layer_type = _layer_type(layer) if layer_type not in _supported_layers: continue if not hasattr(layer, 'activations'): continue #"No activations detected, run forward after add_hooks(model)" assert hasattr(layer, 'backprops_list'), "No backprops detected, run backward after add_hooks(model)" assert len(layer.backprops_list) == 1, "Multiple backprops detected, make sure to call clear_backprops(model)" A = layer.activations n = A.shape[0] if loss_type == 'mean': B = layer.backprops_list[0] * n else: # loss_type == 'sum': B = layer.backprops_list[0] if layer_type == 'Linear': setattr(layer.weight, 'grad1', torch.einsum('ni,nj->nij', B, A)) if layer.bias is not None: setattr(layer.bias, 'grad1', B) elif layer_type == 'Conv2d': A = torch.nn.functional.unfold(A, layer.kernel_size, dilation=layer.dilation, padding=layer.padding, stride=layer.stride) B = B.reshape(n, -1, A.shape[-1]) grad1 = torch.einsum('ijk,ilk->ijl', B, A) shape = [n] + list(layer.weight.shape) setattr(layer.weight, 'grad1', grad1.reshape(shape)) if layer.bias is not None: setattr(layer.bias, 'grad1', torch.sum(B, dim=2)) def compute_hess(model: nn.Module,) -> None: """Save Hessian under param.hess for each param in the model""" for layer in model.modules(): layer_type = _layer_type(layer) if layer_type not in _supported_layers: continue assert hasattr(layer, 'activations'), "No activations detected, run forward after add_hooks(model)" assert hasattr(layer, 'backprops_list'), "No backprops detected, run backward after add_hooks(model)" if layer_type == 'Linear': A = layer.activations B = torch.stack(layer.backprops_list) n = A.shape[0] o = B.shape[0] A = torch.stack([A] * o) Jb = torch.einsum("oni,onj->onij", B, A).reshape(n*o, -1) H = torch.einsum('ni,nj->ij', Jb, Jb) / n setattr(layer.weight, 'hess', H) if layer.bias is not None: setattr(layer.bias, 'hess', torch.einsum('oni,onj->ij', B, B)/n) elif layer_type == 'Conv2d': Kh, Kw = layer.kernel_size di, do = layer.in_channels, layer.out_channels A = layer.activations.detach() A = torch.nn.functional.unfold(A, (Kh, Kw)) # n, di * Kh * Kw, Oh * Ow n = A.shape[0] B = torch.stack([Bt.reshape(n, do, -1) for Bt in layer.backprops_list]) # o, n, do, Oh*Ow o = B.shape[0] A = torch.stack([A] * o) # o, n, di * Kh * Kw, Oh*Ow Jb = torch.einsum('onij,onkj->onik', B, A) # o, n, do, di * Kh * Kw Hi = torch.einsum('onij,onkl->nijkl', Jb, Jb) # n, do, di*Kh*Kw, do, di*Kh*Kw Jb_bias = torch.einsum('onij->oni', B) Hi_bias = torch.einsum('oni,onj->nij', Jb_bias, Jb_bias) setattr(layer.weight, 'hess', Hi.mean(dim=0)) if layer.bias is not None: setattr(layer.bias, 'hess', Hi_bias.mean(dim=0)) def backprop_hess(output: torch.Tensor, hess_type: str) -> None: """ Call backprop 1 or more times to get values needed for Hessian computation. Args: output: prediction of neural network (ie, input of nn.CrossEntropyLoss()) hess_type: type of Hessian propagation, "CrossEntropy" results in exact Hessian for CrossEntropy Returns: """ assert hess_type in ('LeastSquares', 'CrossEntropy') global _enforce_fresh_backprop n, o = output.shape _enforce_fresh_backprop = True if hess_type == 'CrossEntropy': batch = F.softmax(output, dim=1) mask = torch.eye(o).expand(n, o, o) diag_part = batch.unsqueeze(2).expand(n, o, o) * mask outer_prod_part = torch.einsum('ij,ik->ijk', batch, batch) hess = diag_part - outer_prod_part assert hess.shape == (n, o, o) for i in range(n): hess[i, :, :] = symsqrt(hess[i, :, :]) hess = hess.transpose(0, 1) elif hess_type == 'LeastSquares': hess = [] assert len(output.shape) == 2 batch_size, output_size = output.shape id_mat = torch.eye(output_size) for out_idx in range(output_size): hess.append(torch.stack([id_mat[out_idx]] * batch_size)) for o in range(o): output.backward(hess[o], retain_graph=True) def symsqrt(a, cond=None, return_rank=False, dtype=torch.float32): """Symmetric square root of a positive semi-definite matrix. See https://github.com/pytorch/pytorch/issues/25481""" s, u = torch.symeig(a, eigenvectors=True) cond_dict = {torch.float32: 1e3 * 1.1920929e-07, torch.float64: 1E6 * 2.220446049250313e-16} if cond in [None, -1]: cond = cond_dict[dtype] above_cutoff = (abs(s) > cond * torch.max(abs(s))) psigma_diag = torch.sqrt(s[above_cutoff]) u = u[:, above_cutoff] B = u @ torch.diag(psigma_diag) @ u.t() if return_rank: return B, len(psigma_diag) else: return B
9,972
33.628472
133
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/models/stacked_hourglass/StackedHourglass.py
''' Baseline Architecture: Stacked Hourglass https://github.com/princeton-vl/pytorch_stacked_hourglass ''' import torch from torch import nn from .layers import Conv, Hourglass, Pool, Residual class Merge(nn.Module): ''' ''' def __init__(self, x_dim, y_dim): super(Merge, self).__init__() self.conv = Conv(x_dim, y_dim, 1, relu=False, bn=False) def forward(self, x): return self.conv(x) class PoseNet(nn.Module): def __init__(self, arch, auxnet, intermediate_features):#, nstack, inp_dim, oup_dim, bn=False, increase=0, **kwargs): ''' :param nstack: (int) Number of stacks :param inp_dim: (int) Number of input channels for the Stacked Hourglass :param oup_dim: (int) Number of output channels for the Stacked Hourglass :param bn: (bool) Whether to perform Batch Normalization :param increase: :param kwargs: ''' super(PoseNet, self).__init__() self.auxnet = auxnet # Whether to compute features for auxnet self.intermediate_features = intermediate_features # Whether extractor is conv or avg self.nstack = arch['nstack'] inp_dim = arch['channels'] oup_dim = arch['num_hm'] if torch.cuda.device_count() > 1: # We don't need 8 GPUs for 3 stacks lol. n_gpus = min(torch.cuda.device_count(), self.nstack) stacks_per_gpu = torch.zeros(size=(n_gpus,), dtype=torch.int16) # 1. Equal allocation to all stacks_per_gpu += (self.nstack // n_gpus) # 2. Distribute the remaining (max 1) among all the GPUs temp_tensor = torch.zeros(size=(n_gpus,), dtype=torch.int16) for i in range(self.nstack % n_gpus): temp_tensor[i] = 1 stacks_per_gpu += temp_tensor cuda_devices = [] for i in range(stacks_per_gpu.shape[0]): for _ in range(stacks_per_gpu[i]): cuda_devices.append(torch.device('cuda:{}'.format(i))) else: cuda_devices = [torch.device('cuda:0')] * self.nstack self.cuda_devices = cuda_devices self.pre = nn.Sequential( Conv(inp_dim=3, out_dim=64, kernel_size=7, stride=2, bn=True, relu=True), Residual(inp_dim=64, out_dim=128), Pool(2, 2), Residual(inp_dim=128, out_dim=128), Residual(inp_dim=128, out_dim=inp_dim)).cuda(cuda_devices[0]) self.hgs = nn.ModuleList( [nn.Sequential(Hourglass(n=4, f=inp_dim, bn=False, increase=0, intermediate_features=intermediate_features, auxnet=auxnet) ).cuda(cuda_devices[i]) for i in range(self.nstack)]) self.features = nn.ModuleList([nn.Sequential(Residual(inp_dim, inp_dim), Conv(inp_dim, inp_dim, 1, bn=True, relu=True) ).cuda(cuda_devices[i]) for i in range(self.nstack)]) self.outs = nn.ModuleList( [Conv(inp_dim=inp_dim, out_dim=oup_dim, kernel_size=1, relu=False, bn=False).cuda(cuda_devices[i]) for i in range(self.nstack)]) self.merge_features = nn.ModuleList([Merge(inp_dim, inp_dim).cuda(cuda_devices[i]) for i in range(self.nstack-1)]) self.merge_preds = nn.ModuleList([Merge(oup_dim, inp_dim).cuda(cuda_devices[i]) for i in range(self.nstack-1)]) self.global_avg_pool = nn.ModuleList([nn.AvgPool2d(kernel_size=(64, 64), stride=1).cuda(cuda_devices[i]) for i in range(self.nstack)]) def forward(self, imgs): ''' Constructing the Stacked Hourglass Posenet Model :param imgs: :return: ''' # x is of shape: (BatchSize, #channels == 3, input_dim1, input_dim2) x = imgs.permute(0, 3, 1, 2).cuda(self.cuda_devices[0]) x = self.pre(x) combined_hm_preds = [] hourglass_dict= {} for i in range(self.nstack): x = x.cuda(self.cuda_devices[i]) hourglass_dict = self.hgs[i](x) x = hourglass_dict['out'] # Hourglass parameters if self.intermediate_features == 'conv' and self.auxnet: hourglass_dict['feature_5'] = x.clone().detach().to( 'cuda:{}'.format(torch.cuda.device_count() - 1)) x = self.features[i](x) hourglass_dict['penultimate'] = self.global_avg_pool[i](x).clone().detach().to( 'cuda:{}'.format(torch.cuda.device_count() - 1)).reshape(x.shape[0], -1) preds = self.outs[i](x) combined_hm_preds.append(preds.cuda(self.cuda_devices[-1])) if i < self.nstack - 1: x = x + self.merge_preds[i](preds) + self.merge_features[i](x) del hourglass_dict['out'] # ip_learn_loss_dict is a dictionary containing intermediate outputs of hourglass return torch.stack(combined_hm_preds, 1), hourglass_dict
5,127
40.024
122
py
ActiveLearningForHumanPose
ActiveLearningForHumanPose-main/code/models/stacked_hourglass/layers.py
import torch from torch import nn Pool = nn.MaxPool2d def batchnorm(x): return nn.BatchNorm2d(x.size()[1])(x) class Conv(nn.Module): ''' Initializes: Conv, Conv-Relu or Conv-Relu-BN combinaton ''' def __init__(self, inp_dim, out_dim, kernel_size, stride=1, bn=False, relu=True): ''' :param inp_dim: (int) Number of input channels :param out_dim: (int) Number of output channels :param kernel_size: (int) Kernel size :param stride: (int) Convolution stride :param bn: (bool) Whether to perform Batch Normalization :param relu: (bool) Whether to perform ReLU ''' assert type(inp_dim) == type(out_dim) == type(kernel_size) == type(stride) == int, "[Conv]: Wrong typing" assert type(bn) == type(relu) == bool, "[Conv]: Wrong typing" super(Conv, self).__init__() self.bn = None self.relu = None self.inp_dim = inp_dim # Input spatial dim is same as output spatial dim for stride == 1 self.conv = nn.Conv2d(inp_dim, out_dim, kernel_size, stride=stride, padding=(kernel_size-1)//2, bias=True) if relu: self.relu = nn.ReLU() if bn: self.bn = nn.BatchNorm2d(num_features=out_dim) def forward(self, x): assert x.size()[1] == self.inp_dim, "Passed: {}\tExpected: {}".format(x.size()[1], self.inp_dim) x = self.conv(x) if self.bn: x = self.bn(x) if self.relu: x = self.relu(x) return x class Residual(nn.Module): ''' ''' def __init__(self, inp_dim, out_dim): super(Residual, self).__init__() self.relu = nn.ReLU() self.bn1 = nn.BatchNorm2d(num_features=inp_dim) self.conv1 = Conv(inp_dim=inp_dim, out_dim=int(out_dim/2), kernel_size=1, relu=False) self.bn2 = nn.BatchNorm2d(num_features=int(out_dim/2)) self.conv2 = Conv(inp_dim=int(out_dim/2), out_dim=int(out_dim/2), kernel_size=3, relu=False) self.bn3 = nn.BatchNorm2d(num_features=int(out_dim/2)) self.conv3 = Conv(inp_dim=int(out_dim/2), out_dim=out_dim, kernel_size=1, relu=False) self.skip_layer = Conv(inp_dim=inp_dim, out_dim=out_dim, kernel_size=1, relu=False) if inp_dim == out_dim: self.need_skip = False else: self.need_skip = True def forward(self, x): if self.need_skip: residual = self.skip_layer(x) else: residual = x out = self.bn1(x) out = self.relu(out) out = self.conv1(out) out = self.bn2(out) out = self.relu(out) out = self.conv2(out) out = self.bn3(out) out = self.relu(out) out = self.conv3(out) out += residual return out class Hourglass(nn.Module): ''' ''' def __init__(self, n, f, bn=None, increase=0, intermediate_features=None, auxnet=None): super(Hourglass, self).__init__() self.auxnet = auxnet self.intermediate_features = intermediate_features nf = f + increase self.up1 = Residual(f, f) # Lower branch self.pool1 = Pool(2, 2) self.low1 = Residual(f, nf) self.n = n # Recursive hourglass if self.n > 1: self.low2 = Hourglass(n=n-1, f=nf, bn=bn, increase=0, intermediate_features=intermediate_features, auxnet=auxnet) else: self.low2 = Residual(nf, nf) self.low3 = Residual(nf, f) self.up2 = nn.Upsample(scale_factor=2, mode='nearest') self.avg_pool = nn.AvgPool2d(kernel_size=2 ** (self.n + 1), stride=1) def forward(self, x): upper_1 = self.up1(x) x = self.pool1(x) x = self.low1(x) if self.n > 1: hourglass_dict = self.low2(x) x = hourglass_dict['out'] else: x = self.low2(x) x = self.low3(x) if self.intermediate_features == 'conv' and self.auxnet: hourglass_feature_map = x.clone().detach().to('cuda:{}'.format(torch.cuda.device_count() - 1)) upper_2 = self.up2(x) if self.n > 1: hourglass_dict['out'] = upper_1 + upper_2 if self.intermediate_features == 'conv' and self.auxnet: hourglass_dict['feature_{}'.format(self.n)] = hourglass_feature_map return hourglass_dict else: if self.intermediate_features == 'conv' and self.auxnet: return {'out': upper_1 + upper_2, 'feature_1': hourglass_feature_map} else: return {'out': upper_1 + upper_2}
4,697
31.625
114
py