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
pmb-nll
pmb-nll-main/src/probabilistic_inference/probabilistic_detr_predictor.py
import numpy as np import torch import torch.nn.functional as F # DETR imports from detr.util.box_ops import box_cxcywh_to_xyxy # Detectron Imports from detectron2.structures import Boxes # Project Imports from probabilistic_inference import inference_utils from probabilistic_inference.inference_core import ProbabilisticPredictor from probabilistic_modeling.modeling_utils import covariance_output_to_cholesky, clamp_log_variance class DetrProbabilisticPredictor(ProbabilisticPredictor): def __init__(self, cfg): super().__init__(cfg) # These are mock variables to be compatible with probabilistic detectron library. No NMS is performed for DETR. # Only needed for ensemble methods self.test_nms_thresh = 0.5 self.test_topk_per_image = self.model.detr.num_queries def detr_probabilistic_inference(self, input_im): outputs = self.model(input_im, return_raw_results=True, is_mc_dropout=self.mc_dropout_enabled) image_width = input_im[0]['image'].shape[2] image_height = input_im[0]['image'].shape[1] # Handle logits and classes predicted_logits = outputs['pred_logits'][0] if 'pred_logits_var' in outputs.keys(): predicted_logits_var = outputs['pred_logits_var'][0] box_cls_dists = torch.distributions.normal.Normal( predicted_logits, scale=torch.sqrt( torch.exp(predicted_logits_var))) predicted_logits = box_cls_dists.rsample( (self.model.cls_var_num_samples,)) predicted_prob_vectors = F.softmax(predicted_logits, dim=-1) predicted_prob_vectors = predicted_prob_vectors.mean(0) else: predicted_prob_vectors = F.softmax(predicted_logits, dim=-1) predicted_prob, classes_idxs = predicted_prob_vectors[:, :-1].max(-1) # Handle boxes and covariance matrices predicted_boxes = outputs['pred_boxes'][0] # Rescale boxes to inference image size (not COCO original size) pred_boxes = Boxes(box_cxcywh_to_xyxy(predicted_boxes)) pred_boxes.scale(scale_x=image_width, scale_y=image_height) predicted_boxes = pred_boxes.tensor # Rescale boxes to inference image size (not COCO original size) if 'pred_boxes_cov' in outputs.keys(): predicted_boxes_covariance = covariance_output_to_cholesky( outputs['pred_boxes_cov'][0]) predicted_boxes_covariance = torch.matmul( predicted_boxes_covariance, predicted_boxes_covariance.transpose( 1, 2)) transform_mat = torch.tensor([[[1.0, 0.0, -0.5, 0.0], [0.0, 1.0, 0.0, -0.5], [1.0, 0.0, 0.5, 0.0], [0.0, 1.0, 0.0, 0.5]]]).to(self.model.device) predicted_boxes_covariance = torch.matmul( torch.matmul( transform_mat, predicted_boxes_covariance), transform_mat.transpose( 1, 2)) scale_mat = torch.diag_embed( torch.as_tensor( (image_width, image_height, image_width, image_height), dtype=torch.float32)).to( self.model.device).unsqueeze(0) predicted_boxes_covariance = torch.matmul( torch.matmul( scale_mat, predicted_boxes_covariance), torch.transpose(scale_mat, 2, 1)) else: predicted_boxes_covariance = [] if 'ppp' in outputs: ppp = outputs['ppp'] else: ppp = [] return predicted_boxes, predicted_boxes_covariance, predicted_prob, classes_idxs, predicted_prob_vectors, ppp def post_processing_standard_nms(self, input_im): """ This function produces results using standard non-maximum suppression. The function takes into account any probabilistic modeling method when computing the results. Args: input_im (list): an input im list generated from dataset handler. Returns: result (instances): object instances """ outputs = self.detr_probabilistic_inference(input_im) return inference_utils.general_standard_nms_postprocessing( input_im, outputs) def post_processing_topk_detections(self, input_im): """ This function produces results using topk selection based on confidence scores. Args: input_im (list): an input im list generated from dataset handler. Returns: result (instances): object instances """ outputs = self.detr_probabilistic_inference(input_im) return inference_utils.general_topk_detection_postprocessing(input_im, outputs) def post_processing_output_statistics(self, input_im): """ Output statistics does not make much sense for DETR architecture. There is some redundancy due to forced 100 detections per image, but cluster sizes would be too small for meaningful estimates. Might implement it later on. """ raise NotImplementedError pass def post_processing_mc_dropout_ensembles(self, input_im): if self.cfg.PROBABILISTIC_INFERENCE.ENSEMBLES.BOX_MERGE_MODE == 'pre_nms': raise NotImplementedError else: # Merge results: results = [ inference_utils.general_standard_nms_postprocessing( input_im, self.detr_probabilistic_inference(input_im), self.test_nms_thresh, self.test_topk_per_image) for _ in range( self.num_mc_dropout_runs)] # Append per-ensemble outputs after NMS has been performed. ensemble_pred_box_list = [ result.pred_boxes.tensor for result in results] ensemble_pred_prob_vectors_list = [ result.pred_cls_probs for result in results] ensembles_class_idxs_list = [ result.pred_classes for result in results] ensembles_pred_box_covariance_list = [ result.pred_boxes_covariance for result in results] return inference_utils.general_black_box_ensembles_post_processing( input_im, ensemble_pred_box_list, ensembles_class_idxs_list, ensemble_pred_prob_vectors_list, ensembles_pred_box_covariance_list, self.test_nms_thresh, self.test_topk_per_image, self.cfg.PROBABILISTIC_INFERENCE.AFFINITY_THRESHOLD, is_generalized_rcnn=True, merging_method=self.cfg.PROBABILISTIC_INFERENCE.ENSEMBLES.BOX_FUSION_MODE) def post_processing_ensembles(self, input_im, model_dict): if self.cfg.PROBABILISTIC_INFERENCE.ENSEMBLES.BOX_MERGE_MODE == 'pre_nms': raise NotImplementedError else: outputs_list = [] for model in model_dict: self.model = model outputs_list.append( self.post_processing_standard_nms(input_im)) # Merge results: ensemble_pred_box_list = [] ensemble_pred_prob_vectors_list = [] ensembles_class_idxs_list = [] ensembles_pred_box_covariance_list = [] for results in outputs_list: # Append per-ensemble outputs after NMS has been performed. ensemble_pred_box_list.append(results.pred_boxes.tensor) ensemble_pred_prob_vectors_list.append(results.pred_cls_probs) ensembles_class_idxs_list.append(results.pred_classes) ensembles_pred_box_covariance_list.append( results.pred_boxes_covariance) return inference_utils.general_black_box_ensembles_post_processing( input_im, ensemble_pred_box_list, ensembles_class_idxs_list, ensemble_pred_prob_vectors_list, ensembles_pred_box_covariance_list, self.test_nms_thresh, self.test_topk_per_image, self.cfg.PROBABILISTIC_INFERENCE.AFFINITY_THRESHOLD, is_generalized_rcnn=True, merging_method=self.cfg.PROBABILISTIC_INFERENCE.ENSEMBLES.BOX_FUSION_MODE) def post_processing_bayes_od(self, input_im): """ Since there is no NMS step in DETR, bayesod is not implemented. Although possible to add NMS and implement it later on. """ raise NotImplementedError pass
9,040
40.095455
119
py
pmb-nll
pmb-nll-main/src/probabilistic_modeling/losses.py
from collections import defaultdict from math import comb from math import factorial from typing import Dict, List, Optional, Tuple, Union import numpy as np import torch from core.fastmurty.mhtdaClink import (allocateWorkvarsforDA, deallocateWorkvarsforDA, mhtda, sparse) from core.fastmurty.mhtdaClink import sparsifyByRow as sparsify from scipy.optimize import linear_sum_assignment from torch.distributions.multivariate_normal import MultivariateNormal from probabilistic_modeling.modeling_utils import ( clamp_log_variance, covariance_output_to_cholesky) def reshape_box_preds(preds, num_classes): """ Tiny helper function to reshape box predictions from [numpreds,classes*boxdim] to [numpreds,classes,boxdim] """ num_preds, *_ = preds.shape if num_preds == 0: return preds if len(preds.shape) == 2: preds = preds.unsqueeze(1) if preds.shape[-1] > num_classes: # if box predicted per class preds = preds.reshape(num_preds, num_classes, -1) else: preds = preds.repeat(1, num_classes, 1) return preds def run_murtys(cost_matrix: torch.tensor, nsolutions: int): """ Run fastmurtys given cost_matrix and number of assignments to search for. Returns associations and costs. Based on example_simplest.py in fastmurty. """ # make all costs negative for algo to work properly cost_matrix_max = cost_matrix.max() if cost_matrix_max >= 0: cost_matrix = cost_matrix - (cost_matrix_max + 1) cost_matrix = cost_matrix.detach().numpy() nrows, ncolumns = cost_matrix.shape # sparse cost matrices only include a certain number of elements # the rest are implicitly infinity # in this case, the sparse matrix includes all elements # The sparse and dense versions are compiled differently (see the Makefile). # The variable "sparse" in mhtdaClink needs to match the version compiled cost_matrix_to_use = sparsify(cost_matrix, ncolumns) if sparse else cost_matrix # mhtda is set up to potentially take multiple input hypotheses for both rows and columns # input hypotheses specify a subset of rows or columns. # In this case, we just want to use the whole matrix. row_priors = np.ones((1, nrows), dtype=np.bool8) col_priors = np.ones((1, ncolumns), dtype=np.bool8) # Each hypothesis has a relative weight too. # These values don't matter if there is only one hypothesis... row_prior_weights = np.zeros(1) col_prior_weights = np.zeros(1) # The mhtda function modifies preallocated outputs rather than # allocating new ones. This is slightly more efficient for repeated use # within a tracker. # The cost of each returned association: out_costs = np.zeros(nsolutions) # The row-column pairs in each association: # Generally there will be less than nrows+ncolumns pairs in an association. # The unused pairs are currently set to (-2, -2) out_associations = np.zeros((nsolutions, nrows + ncolumns, 2), dtype=np.int32) # variables needed within the algorithm (a C function sets this up): workvars = allocateWorkvarsforDA(nrows, ncolumns, nsolutions) # run! mhtda( cost_matrix_to_use, row_priors, row_prior_weights, col_priors, col_prior_weights, out_associations, out_costs, workvars, ) deallocateWorkvarsforDA(workvars) return out_associations, out_costs def compute_negative_log_likelihood( box_scores: torch.tensor, box_regs: torch.tensor, box_covars: torch.tensor, gt_box: torch.tensor, gt_class: torch.tensor, image_size: List[int], reg_distribution: torch.distributions.distribution.Distribution, associations: np.ndarray, device: torch.device, intensity_func=lambda x: 0.00000001, scores_have_bg_cls=False, target_delta=None, pred_delta=None, pred_delta_chol=None, ): """Compute NLL for given associations. Args: box_scores (torch.tensor): [description] box_regs (torch.tensor): [description] box_covars (torch.tensor): [description] gt_box (torch.tensor): [description] gt_class (torch.tensor): [description] image_size (List[int]): [description] reg_distribution (torch.distributions.distribution.Distribution): [description] associations (np.ndarray[np.int32]): [description] device (torch.device): [description] intensity_func ([type], optional): [description]. Defaults to lambdax:0.00000001. Returns: [type]: [description] """ if type(image_size) is not torch.tensor: image_size = torch.tensor(image_size) img_size = image_size.unsqueeze(0).to(device) existance_prob = 1 - box_scores[:, -1] num_preds, num_classes = box_scores.shape if scores_have_bg_cls: num_classes -= 1 # do not count background class num_gt, _ = gt_box.shape out_dict = defaultdict(list) out_dict.update( { "matched_bernoulli": [], "unmatched_bernoulli": [], "matched_ppp": [], "matched_bernoulli_reg": [], "matched_bernoulli_cls": [], "num_matched_bernoulli": [], "num_unmatched_bernoulli": [], "num_matched_ppp": [], "ppp_integral": None, } ) nll = torch.zeros(len(associations), dtype=torch.float64, device=device) for a, association in enumerate(associations): log_matched_bernoulli = torch.tensor(0, dtype=torch.float64, device=device) log_unmatched_bernoulli = torch.tensor(0, dtype=torch.float64, device=device) log_poisson = torch.tensor(0, dtype=torch.float64, device=device) log_matched_regression = torch.tensor(0, dtype=torch.float64, device=device) log_matched_classification = torch.tensor(0, dtype=torch.float64, device=device) num_matched_bernoulli = 0 num_unmatched_bernoulli = 0 num_matched_ppp = 0 log_matched_bernoulli_regs = [] log_matched_bernoulli_cls = [] log_unmatched_bernoullis = [] log_matched_ppps = [] for pair in association: pred = pair[0] gt = pair[1] if ( 0 <= pred < num_preds ) and gt >= 0: # if bernoulli was assigned to a GT element num_matched_bernoulli += 1 assigned_gt = gt k = pred gt_c = gt_class[assigned_gt] if scores_have_bg_cls: r = existance_prob[k] else: r = box_scores[k, gt_c] covar = box_covars[k, gt_c] if target_delta is None: covar = box_covars[k, gt_c] dist = reg_distribution(box_regs[k, gt_c, :], covar) regression = dist.log_prob(gt_box[assigned_gt, :]).sum() classification = torch.log(box_scores[k, gt_c]) else: covar = pred_delta_chol[k, gt_c] dist = reg_distribution(pred_delta[k, gt_c, :], covar) regression = dist.log_prob(target_delta[k, assigned_gt, :]).sum() classification = torch.log(box_scores[k, gt_c]) log_f = regression + classification # Save stats log_matched_bernoulli_regs.append(-regression.squeeze().item()) log_matched_bernoulli_cls.append(-classification.squeeze().item()) # Update total bernoulli component log_matched_bernoulli = log_matched_bernoulli + log_f.squeeze() log_matched_regression = log_matched_regression + regression.squeeze() log_matched_classification = ( log_matched_classification + classification.squeeze() ) elif ( 0 <= pred < num_preds ) and gt == -1: # if bernoulli was not assigned to a GT element num_unmatched_bernoulli += 1 k = pred if scores_have_bg_cls: log_f = torch.log(1 - existance_prob[k]) else: log_f = torch.log(1 - box_scores[k].max()) log_unmatched_bernoulli = log_unmatched_bernoulli + log_f.squeeze() # Save stats log_unmatched_bernoullis.append(-log_f.squeeze().item()) elif (pred >= num_preds) and ( gt >= 0 ): # if poisson was assigned to a GT element num_matched_ppp += 1 assigned_gt = gt gt_c = gt_class[assigned_gt].unsqueeze(0) gt_vec = torch.cat([gt_box[assigned_gt, :], gt_c]) log_f = intensity_func(gt_vec.unsqueeze(0), img_size).squeeze() log_poisson = log_poisson + log_f # Save stats log_matched_ppps.append(-log_f.item()) association_sum = log_matched_bernoulli + log_unmatched_bernoulli + log_poisson out_dict["matched_bernoulli"].append(-log_matched_bernoulli.item()) out_dict["matched_bernoulli_reg"].append(-log_matched_regression.item()) out_dict["matched_bernoulli_cls"].append(-log_matched_classification.item()) out_dict["num_matched_bernoulli"].append(num_matched_bernoulli) out_dict["unmatched_bernoulli"].append(-log_unmatched_bernoulli.item()) out_dict["num_unmatched_bernoulli"].append(num_unmatched_bernoulli) out_dict["matched_ppp"].append(-log_poisson.item()) out_dict["num_matched_ppp"].append(num_matched_ppp) out_dict["matched_bernoulli_regs"].append(log_matched_bernoulli_regs) out_dict["matched_bernoulli_clss"].append(log_matched_bernoulli_cls) out_dict["unmatched_bernoullis"].append(log_unmatched_bernoullis) out_dict["matched_ppps"].append(log_matched_ppps) nll[a] = association_sum nll = torch.logsumexp(nll, -1) n_class = torch.tensor(num_classes).unsqueeze(0).to(device) ppp_regularizer = intensity_func(None, img_size, n_class, integrate=True).squeeze() nll = ppp_regularizer - nll out_dict["ppp_integral"] = ppp_regularizer.item() out_dict["total"] = [ out_dict["matched_bernoulli"][i] + out_dict["unmatched_bernoulli"][i] + out_dict["matched_ppp"][i] + out_dict["ppp_integral"] for i in range(len(associations)) ] return nll, out_dict def negative_log_likelihood_matching( box_scores: torch.tensor, box_regs: torch.tensor, box_covars: torch.tensor, gt_box: torch.tensor, gt_class: torch.tensor, image_size: List[int], reg_distribution: torch.distributions.distribution.Distribution, device: torch.device, intensity_func=lambda x: 0.00000001, max_n_solutions: int = 5, scores_have_bg_cls=False, target_delta=None, distance_type="log_prob", covar_scaling = 1, use_target_delta_matching=True, pred_delta=None, pred_delta_chol=None, ): img_size = torch.tensor(image_size).unsqueeze(0).to(device) num_preds, num_classes = box_scores.shape if scores_have_bg_cls: num_classes -= 1 # do not count background class num_gt = gt_box.shape[0] existance_prob = 1 - box_scores[:, -1] # Init potential covar scaling for matching covar_scaling = torch.eye(box_covars.shape[-1]).to(box_covars.device)*covar_scaling # save indices of inf cost infinite_costs = [] with torch.no_grad(): if not(num_gt > 0 and num_preds > 0): associations = -np.ones((1, num_preds + num_gt, 2)) if num_gt > 0: associations[0, -num_gt:, 1] = np.arange(num_gt) associations[0, :, 0] = np.arange(num_preds + num_gt) associations = associations.astype(np.int32) return associations # Assemble and fill cost matrix cost_matrix = torch.zeros((num_preds + num_gt, num_gt), dtype=torch.float64) if scores_have_bg_cls: r = existance_prob.unsqueeze(-1).repeat(1, num_gt) else: r = box_scores[:, gt_class] # assume existance prob == class prob covar = box_covars[:, gt_class] if pred_delta_chol is None or not use_target_delta_matching else pred_delta_chol[:, gt_class] reg_means = box_regs if pred_delta is None or not use_target_delta_matching else pred_delta # Repeat gt to be [num_preds,num_gt,dim] if needed if len(gt_box.shape) < len(reg_means[:, gt_class].shape): gt_box = gt_box.unsqueeze(0).repeat(num_preds, 1, 1) if distance_type == "log_prob": # Covar is actually cholesky decomposed, hence only one multiplication with scaling scaled_covar = covar_scaling@covar dist = reg_distribution(reg_means[:, gt_class], scaled_covar) if target_delta is None or not use_target_delta_matching: log_p = dist.log_prob(gt_box) else: log_p = dist.log_prob(target_delta) elif distance_type == "euclidian_squared": # We use minus since its sign is reversed later (and cost should be minimized) if target_delta is None or not use_target_delta_matching: log_p = -(reg_means[:, gt_class] - gt_box).pow(2).sum(-1) else: log_p = -(reg_means[:, gt_class] - target_delta).pow(2).sum(-1) elif distance_type == "euclidian": # We use minus since its sign is reversed later (and cost should be minimized) if target_delta is None or not use_target_delta_matching: log_p = -(reg_means[:, gt_class] - gt_box).pow(2).sum(-1).sqrt() else: log_p = ( -(reg_means[:, gt_class] - target_delta).pow(2).sum(-1).sqrt() ) else: raise NotImplementedError( f'Distance type for PMB-NLL matching "{distance_type}" not implemented.' ) log_p = log_p.sum(-1) if len(log_p.shape) > 2 else log_p log_p = log_p + torch.log( box_scores[:, gt_class] ) # box regression + class scores conditioned on existance cost = -(log_p - torch.log(1 - r)) cost_matrix[:num_preds] = cost if not torch.isfinite(cost).all(): for k, l in torch.isfinite(cost).logical_not().nonzero(): infinite_costs.append((k, l)) cost_matrix[k, l] = 0 # Build GT vector with [box, class] if target_delta is None or not use_target_delta_matching: gt_vec = torch.cat([gt_box[0, :, :], gt_class.unsqueeze(-1)], -1) else: gt_vec = torch.cat([target_delta[0, :, :], gt_class.unsqueeze(-1)], -1) # PPP cost cost = -intensity_func(gt_vec, img_size, dist_type=distance_type) if torch.isfinite(cost).all(): cost_matrix[num_preds:] = torch.diag(cost) else: cost_matrix[num_preds:] = torch.diag(cost) for l in torch.isfinite(cost).logical_not().nonzero(): infinite_costs.append((num_preds + l, l)) cost_matrix[num_preds + l, l] = 0 # Fill in "inf" if cost_matrix.numel() > 0: largest_cost = cost_matrix.max() for k in range(num_preds, num_preds + num_gt): # loop over predictions for l in range(num_gt): # loop over ground truths if k != (l + num_preds): cost_matrix[k, l] = largest_cost * 3 for coord in infinite_costs: k, l = coord cost_matrix[k, l] = largest_cost * 2 # Find nsolutions best solutions nsolutions = 0 for i in range(num_gt+1): if i > num_preds or nsolutions > max_n_solutions: break nsolutions += (factorial(num_preds)//factorial(num_preds-i))*comb(num_gt, i) nsolutions = min( max_n_solutions, nsolutions ) # comb gives maximum number unique associations try: associations, _ = run_murtys(cost_matrix, nsolutions) except AssertionError: print( "[NLLOD] Murtys could not find solution! Using linear sum assignment." ) row_ind, col_ind = linear_sum_assignment(cost_matrix.cpu().numpy()) associations = -np.ones((1, num_preds + num_gt, 2)) associations[0, :, 0] = np.arange(num_preds + num_gt) associations[0, row_ind, 1] = col_ind associations = associations.astype(np.int32) return associations def negative_log_likelihood( pred_box_scores: List[torch.tensor], pred_box_regs: List[torch.tensor], pred_box_covars: List[torch.tensor], gt_boxes: List[torch.tensor], gt_classes: List[torch.tensor], image_sizes: List[List[int]], reg_distribution: torch.distributions.distribution.Distribution, intensity_func=lambda x: 0.00000001, max_n_solutions: int = 5, training: bool = True, scores_have_bg_cls: bool = True, target_deltas: torch.tensor = None, matching_distance: str = "log_prob", covar_scaling: float = 1.0, use_target_delta_matching=False, pred_deltas=None, pred_delta_chols=None, ): """ Calculate NLL for a PMB prediction. """ assert len(pred_box_scores) == len(pred_box_regs) == len(pred_box_covars) device = pred_box_scores[0].device nll_total_losses = torch.tensor( 0, dtype=torch.float64, device=device, requires_grad=training ) bs = len(pred_box_scores) total_associations = [] total_decompositions = [] for i in range(bs): # loop over images if type(intensity_func) == list: if type(intensity_func[i]) != dict: ppp = {"matching": intensity_func[i], "loss": intensity_func[i]} else: ppp = intensity_func[i] else: if type(intensity_func) != dict: ppp = {"matching": intensity_func, "loss": intensity_func} else: ppp = intensity_func # [N, num_classes] or [N, num_classes+1] box_scores = pred_box_scores[i] num_preds, num_classes = box_scores.shape if scores_have_bg_cls: num_classes -= 1 # do not count background class # [N, num_classes, boxdims] box_regs = pred_box_regs[i] # [N, num_classes, boxdims, boxdims] box_covars = pred_box_covars[i] # [M, boxdims] gt_box = gt_boxes[i] # [M, 1] gt_class = gt_classes[i] if target_deltas is None: target_delta = None else: # [N, M, boxdims] target_delta = target_deltas[i] if pred_deltas is None: pred_delta = None else: # [N, M, boxdims] pred_delta = pred_deltas[i] if pred_delta_chols is None: pred_delta_chol = None else: # [N, M, boxdims] pred_delta_chol = pred_delta_chols[i] image_size = image_sizes[i] associations = negative_log_likelihood_matching( box_scores, box_regs, box_covars, gt_box, gt_class, image_size, reg_distribution, device, ppp["matching"], max_n_solutions, scores_have_bg_cls, target_delta, matching_distance, covar_scaling, use_target_delta_matching, pred_delta, pred_delta_chol, ) nll, decomposition = compute_negative_log_likelihood( box_scores=box_scores, box_regs=box_regs, box_covars=box_covars, gt_box=gt_box, gt_class=gt_class, image_size=image_size, reg_distribution=reg_distribution, associations=associations, device=device, intensity_func=ppp["loss"], scores_have_bg_cls=scores_have_bg_cls, target_delta=target_delta, pred_delta=pred_delta, pred_delta_chol=pred_delta_chol, ) if torch.isfinite(nll): # Normalize by num predictions if training if training: number_preds = decomposition["num_matched_ppp"][0]+decomposition["num_matched_bernoulli"][0]+decomposition["num_unmatched_bernoulli"][0] regularizer = max(1, number_preds) nll_total_losses = nll_total_losses + nll / regularizer else: nll_total_losses = nll_total_losses + nll else: bs = max(1, bs - 1) print("WARNING: Infinite loss in NLL!") print(f"box scores: {box_scores}") print(f"box_regs: {box_regs}") print(f"box_covars: {box_covars}") print(f"gt_box: {gt_box}") print(f"gt_class: {gt_class}") print(f"associations: {associations}") total_associations.append(associations) total_decompositions.append(decomposition) return nll_total_losses / bs, total_associations, total_decompositions
21,481
37.846293
152
py
pmb-nll
pmb-nll-main/src/probabilistic_modeling/probabilistic_retinanet.py
import logging import math from typing import List, Tuple import numpy as np import torch from core.visualization_tools.probabilistic_visualizer import ProbabilisticVisualizer from detectron2.data.detection_utils import convert_image_to_rgb # Detectron Imports from detectron2.layers import ShapeSpec, batched_nms, cat, nonzero_tuple from detectron2.modeling.anchor_generator import build_anchor_generator from detectron2.modeling.meta_arch.build import META_ARCH_REGISTRY from detectron2.modeling.meta_arch.retinanet import ( RetinaNet, RetinaNetHead, permute_to_N_HWA_K, ) from detectron2.modeling.postprocessing import detector_postprocess from detectron2.structures import Boxes, Instances from detectron2.utils.events import get_event_storage from fvcore.nn import sigmoid_focal_loss_jit, smooth_l1_loss from matplotlib import cm from probabilistic_inference import inference_utils from torch import Tensor, distributions, nn from probabilistic_modeling.losses import ( negative_log_likelihood, negative_log_likelihood_matching, ) # Project Imports from probabilistic_modeling.modeling_utils import ( PoissonPointProcessIntensityFunction, clamp_log_variance, covariance_output_to_cholesky, get_probabilistic_loss_weight, unscented_transform, PoissonPointUnion, ) @META_ARCH_REGISTRY.register() class ProbabilisticRetinaNet(RetinaNet): """ Probabilistic retinanet class. """ def __init__(self, cfg): super().__init__(cfg) # Parse configs self.cls_var_loss = cfg.MODEL.PROBABILISTIC_MODELING.CLS_VAR_LOSS.NAME self.compute_cls_var = self.cls_var_loss != "none" self.cls_var_num_samples = ( cfg.MODEL.PROBABILISTIC_MODELING.CLS_VAR_LOSS.NUM_SAMPLES ) self.bbox_cov_loss = cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.NAME self.compute_bbox_cov = self.bbox_cov_loss != "none" self.bbox_cov_num_samples = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.NUM_SAMPLES ) self.bbox_cov_dist_type = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE ) self.bbox_cov_type = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.COVARIANCE_TYPE ) if self.bbox_cov_type == "diagonal": # Diagonal covariance matrix has N elements self.bbox_cov_dims = 4 else: # Number of elements required to describe an NxN covariance matrix is # computed as: (N * (N + 1)) / 2 self.bbox_cov_dims = 10 if self.bbox_cov_loss == "pmb_negative_log_likelihood": self.ppp_constructor = lambda x: PoissonPointProcessIntensityFunction( cfg, **x ) self.ppp_intensity_function = PoissonPointProcessIntensityFunction(cfg, device=self.device) self.nll_max_num_solutions = ( cfg.MODEL.PROBABILISTIC_MODELING.NLL_MAX_NUM_SOLUTIONS ) self.matching_distance = cfg.MODEL.PROBABILISTIC_MODELING.MATCHING_DISTANCE self.use_prediction_mixture = cfg.MODEL.PROBABILISTIC_MODELING.PPP.USE_PREDICTION_MIXTURE self.dropout_rate = cfg.MODEL.PROBABILISTIC_MODELING.DROPOUT_RATE self.use_dropout = self.dropout_rate != 0.0 self.current_step = 0 self.annealing_step = ( cfg.SOLVER.STEPS[1] if cfg.MODEL.PROBABILISTIC_MODELING.ANNEALING_STEP <= 0 else cfg.MODEL.PROBABILISTIC_MODELING.ANNEALING_STEP ) # Define custom probabilistic head backbone_shape = self.backbone.output_shape() feature_shapes = [backbone_shape[f] for f in self.head_in_features] self.head = ProbabilisticRetinaNetHead( cfg, self.use_dropout, self.dropout_rate, self.compute_cls_var, self.compute_bbox_cov, self.bbox_cov_dims, feature_shapes, ) # Send to device self.to(self.device) def get_ppp_intensity_function(self): return self.ppp_intensity_function def forward( self, batched_inputs, return_anchorwise_output=False, num_mc_dropout_runs=-1 ): """ Args: batched_inputs: a list, batched outputs of :class:`DatasetMapper` . Each item in the list contains the inputs for one image. For now, each item in the list is a dict that contains: * image: Tensor, image in (C, H, W) format. * instances: Instances Other information that's included in the original dicts, such as: * "height", "width" (int): the output resolution of the model, used in inference. See :meth:`postprocess` for details. return_anchorwise_output (bool): returns raw output for probabilistic inference num_mc_dropout_runs (int): perform efficient monte-carlo dropout runs by running only the head and not full neural network. Returns: dict[str: Tensor]: mapping from a named loss to a tensor storing the loss. Used during training only. """ # Update step try: self.current_step += get_event_storage().iter except: self.current_step += 1 # Preprocess image images = self.preprocess_image(batched_inputs) # Extract features and generate anchors features = self.backbone(images.tensor) features = [features[f] for f in self.head_in_features] anchors = self.anchor_generator(features) # MC_Dropout inference forward if num_mc_dropout_runs > 1: anchors = anchors * num_mc_dropout_runs features = features * num_mc_dropout_runs output_dict = self.produce_raw_output(anchors, features) return output_dict # Regular inference forward if return_anchorwise_output: return self.produce_raw_output(anchors, features) # Training and validation forward ( pred_logits, pred_anchor_deltas, pred_logits_vars, pred_anchor_deltas_vars, ) = self.head(features) # Transpose the Hi*Wi*A dimension to the middle: pred_logits = [permute_to_N_HWA_K(x, self.num_classes) for x in pred_logits] pred_anchor_deltas = [permute_to_N_HWA_K(x, 4) for x in pred_anchor_deltas] if pred_logits_vars is not None: pred_logits_vars = [ permute_to_N_HWA_K(x, self.num_classes) for x in pred_logits_vars ] if pred_anchor_deltas_vars is not None: pred_anchor_deltas_vars = [ permute_to_N_HWA_K(x, self.bbox_cov_dims) for x in pred_anchor_deltas_vars ] if self.training: assert ( "instances" in batched_inputs[0] ), "Instance annotations are missing in training!" gt_instances = [x["instances"].to(self.device) for x in batched_inputs] gt_classes, gt_boxes = self.label_anchors(anchors, gt_instances) self.anchors = torch.cat( [Boxes.cat(anchors).tensor for i in range(len(gt_instances))], 0 ) # Loss is computed based on what values are to be estimated by the neural # network losses = self.losses( anchors, gt_classes, gt_boxes, pred_logits, pred_anchor_deltas, pred_logits_vars, pred_anchor_deltas_vars, gt_instances, images.image_sizes, ) if self.vis_period > 0: storage = get_event_storage() if storage.iter % self.vis_period == 0: results = self.inference( anchors, pred_logits, pred_anchor_deltas, images.image_sizes ) self.visualize_training( batched_inputs, results, pred_logits, pred_anchor_deltas, pred_anchor_deltas_vars, anchors, ) return losses else: results = self.inference( anchors, pred_logits, pred_anchor_deltas, images.image_sizes ) processed_results = [] for results_per_image, input_per_image, image_size in zip( results, batched_inputs, images.image_sizes ): height = input_per_image.get("height", image_size[0]) width = input_per_image.get("width", image_size[1]) r = detector_postprocess(results_per_image[0], height, width) processed_results.append({"instances": r}) return processed_results def visualize_training( self, batched_inputs, results, pred_logits, pred_anchor_deltas, pred_anchor_deltas_vars, anchors, ): """ A function used to visualize ground truth images and final network predictions. It shows ground truth bounding boxes on the original image and up to 20 predicted object bounding boxes on the original image. Args: batched_inputs (list): a list that contains input to the model. results (List[Instances]): a list of #images elements. """ from detectron2.utils.visualizer import Visualizer pred_instaces, kept_idx = results assert len(batched_inputs) == len( pred_instaces ), "Cannot visualize inputs and results of different sizes" storage = get_event_storage() max_boxes = 20 image_index = 0 # only visualize a single image img = batched_inputs[image_index]["image"] img = convert_image_to_rgb(img.permute(1, 2, 0), self.input_format) # Extract NMS kept predictions box_scores = torch.cat([logits.squeeze() for logits in pred_logits])[ kept_idx ].sigmoid() box_scores = torch.cat( (box_scores, 1 - pred_instaces[image_index].scores.unsqueeze(-1)), dim=-1 ) anchor_deltas = torch.cat([delta.squeeze() for delta in pred_anchor_deltas])[ kept_idx ] anchor_delta_vars = torch.cat( [var.squeeze() for var in pred_anchor_deltas_vars] )[kept_idx] anchor_boxes = torch.cat([box.tensor.squeeze() for box in anchors])[kept_idx] cholesky_decomp = covariance_output_to_cholesky(anchor_delta_vars) ######## Get covariance for corner coordinates instead ######### multivariate_normal_samples = torch.distributions.MultivariateNormal( anchor_deltas, scale_tril=cholesky_decomp ) # Define monte-carlo samples distributions_samples = multivariate_normal_samples.rsample((1000,)) distributions_samples = torch.transpose( torch.transpose(distributions_samples, 0, 1), 1, 2 ) samples_proposals = torch.repeat_interleave( anchor_boxes.unsqueeze(2), 1000, dim=2 ) # Transform samples from deltas to boxes box_transform = inference_utils.SampleBox2BoxTransform( self.box2box_transform.weights ) t_dist_samples = box_transform.apply_samples_deltas( distributions_samples, samples_proposals ) # Compute samples mean and covariance matrices. _, boxes_covars = inference_utils.compute_mean_covariance_torch(t_dist_samples) # Scale if image has been reshaped during processing scale_x, scale_y = ( img.shape[1] / pred_instaces[image_index].image_size[1], img.shape[0] / pred_instaces[image_index].image_size[0], ) scaling = torch.tensor(np.stack([scale_x, scale_y, scale_x, scale_y]) ** 2).to( device=boxes_covars.device ) boxes_covars = (boxes_covars * scaling).float() processed_results = detector_postprocess( pred_instaces[image_index], img.shape[0], img.shape[1] ) predicted_boxes = processed_results.pred_boxes.tensor if self.bbox_cov_dist_type == "gaussian": reg_distribution = ( lambda x, y: distributions.multivariate_normal.MultivariateNormal(x, y) ) elif self.bbox_cov_dist_type == "laplacian": reg_distribution = lambda x, y: distributions.laplace.Laplace( loc=x, scale=(y.diagonal(dim1=-2, dim2=-1) / np.sqrt(2)) ) else: raise Exception( f"Bounding box uncertainty distribution {self.bbox_cov_dist_type} is not available." ) associations = negative_log_likelihood_matching( box_scores, box_regs=predicted_boxes.unsqueeze(1).repeat(1, 80, 1), box_covars=boxes_covars.unsqueeze(1).repeat(1, 80, 1, 1), gt_box=batched_inputs[image_index]["instances"].gt_boxes.tensor, gt_class=batched_inputs[image_index]["instances"].gt_classes, image_size=img.shape, reg_distribution=reg_distribution, device=boxes_covars.device, intensity_func=self.ppp_intensity_function, max_n_solutions=1, ) ################# Draw results #################### color_map = cm.get_cmap("tab20") num_gt = batched_inputs[image_index]["instances"].gt_boxes.tensor.shape[0] gt_colors = [color_map(i) for i in range(num_gt)] v_gt = Visualizer(img, None) v_gt = v_gt.overlay_instances( boxes=batched_inputs[image_index]["instances"].gt_boxes, assigned_colors=gt_colors, ) anno_img = v_gt.get_image() num_preds = len(boxes_covars) pred_colors = [(0.0, 0.0, 0.0, 1.0)] * num_preds for i in range(num_preds): matched_gt = associations[0, i, 1] if matched_gt >= 0: pred_colors[i] = color_map(matched_gt) pred_labels = [ f"{pred_class.item()}: {round(pred_score.item(),2)}" for pred_class, pred_score in zip( pred_instaces[image_index].pred_classes, pred_instaces[image_index].scores, ) ] v_pred = ProbabilisticVisualizer(img, None) v_pred = v_pred.overlay_covariance_instances( boxes=predicted_boxes[:max_boxes].detach().cpu().numpy(), covariance_matrices=boxes_covars[:max_boxes].detach().cpu().numpy(), assigned_colors=pred_colors, labels=pred_labels[:max_boxes], ) prop_img = v_pred.get_image() vis_img = np.vstack((anno_img, prop_img)) vis_img = vis_img.transpose(2, 0, 1) vis_name = ( f"Top: GT bounding boxes; Bottom: {max_boxes} Highest Scoring Results" ) storage.put_image(vis_name, vis_img) def losses( self, anchors, gt_classes, gt_boxes, pred_class_logits, pred_anchor_deltas, pred_class_logits_var=None, pred_bbox_cov=None, gt_instances=None, image_sizes: List[Tuple[int, int]] = [], ): """ Args: For `gt_classes` and `gt_anchors_deltas` parameters, see :meth:`RetinaNet.get_ground_truth`. Their shapes are (N, R) and (N, R, 4), respectively, where R is the total number of anchors across levels, i.e. sum(Hi x Wi x A) For `pred_class_logits`, `pred_anchor_deltas`, `pred_class_logits_var` and `pred_bbox_cov`, see :meth:`RetinaNetHead.forward`. Returns: dict[str: Tensor]: mapping from a named loss to a scalar tensor storing the loss. Used during training only. The dict keys are: "loss_cls" and "loss_box_reg" """ num_images = len(gt_classes) gt_labels = torch.stack(gt_classes) # (N, R) # Do NMS before reshaping stuff if self.bbox_cov_loss == "pmb_negative_log_likelihood": with torch.no_grad(): nms_results = self.inference( anchors, pred_class_logits, pred_anchor_deltas, image_sizes ) anchors = type(anchors[0]).cat(anchors).tensor # (R, 4) gt_anchor_deltas = [ self.box2box_transform.get_deltas(anchors, k) for k in gt_boxes ] gt_anchor_deltas = torch.stack(gt_anchor_deltas) # (N, R, 4) valid_mask = gt_labels >= 0 pos_mask = (gt_labels >= 0) & (gt_labels != self.num_classes) num_pos_anchors = pos_mask.sum().item() get_event_storage().put_scalar("num_pos_anchors", num_pos_anchors / num_images) self.loss_normalizer = self.loss_normalizer_momentum * self.loss_normalizer + ( 1 - self.loss_normalizer_momentum ) * max(num_pos_anchors, 1) # classification and regression loss # Shapes: # (N x R, K) for class_logits and class_logits_var. # (N x R, 4), (N x R x 10) for pred_anchor_deltas and pred_class_bbox_cov respectively. # Transform per-feature layer lists to a single tensor pred_class_logits = cat(pred_class_logits, dim=1) pred_anchor_deltas = cat(pred_anchor_deltas, dim=1) if pred_class_logits_var is not None: pred_class_logits_var = cat(pred_class_logits_var, dim=1) if pred_bbox_cov is not None: pred_bbox_cov = cat(pred_bbox_cov, dim=1) gt_classes_target = torch.nn.functional.one_hot( gt_labels[valid_mask], num_classes=self.num_classes + 1 )[:, :-1].to( pred_class_logits[0].dtype ) # no loss for the last (background) class # Classification losses if self.compute_cls_var: # Compute classification variance according to: # "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", NIPS 2017 if self.cls_var_loss == "loss_attenuation": num_samples = self.cls_var_num_samples # Compute standard deviation pred_class_logits_var = torch.sqrt( torch.exp(pred_class_logits_var[valid_mask]) ) pred_class_logits = pred_class_logits[valid_mask] # Produce normal samples using logits as the mean and the standard deviation computed above # Scales with GPU memory. 12 GB ---> 3 Samples per anchor for # COCO dataset. univariate_normal_dists = distributions.normal.Normal( pred_class_logits, scale=pred_class_logits_var ) pred_class_stochastic_logits = univariate_normal_dists.rsample( (num_samples,) ) pred_class_stochastic_logits = pred_class_stochastic_logits.view( ( pred_class_stochastic_logits.shape[1] * num_samples, pred_class_stochastic_logits.shape[2], -1, ) ) pred_class_stochastic_logits = pred_class_stochastic_logits.squeeze(2) # Produce copies of the target classes to match the number of # stochastic samples. gt_classes_target = torch.unsqueeze(gt_classes_target, 0) gt_classes_target = torch.repeat_interleave( gt_classes_target, num_samples, dim=0 ).view( ( gt_classes_target.shape[1] * num_samples, gt_classes_target.shape[2], -1, ) ) gt_classes_target = gt_classes_target.squeeze(2) # Produce copies of the target classes to form the stochastic # focal loss. loss_cls = ( sigmoid_focal_loss_jit( pred_class_stochastic_logits, gt_classes_target, alpha=self.focal_loss_alpha, gamma=self.focal_loss_gamma, reduction="sum", ) / (num_samples * max(1, self.loss_normalizer)) ) else: raise ValueError( "Invalid classification loss name {}.".format(self.bbox_cov_loss) ) else: # Standard loss computation in case one wants to use this code # without any probabilistic inference. loss_cls = ( sigmoid_focal_loss_jit( pred_class_logits[valid_mask], gt_classes_target, alpha=self.focal_loss_alpha, gamma=self.focal_loss_gamma, reduction="sum", ) / max(1, self.loss_normalizer) ) # Compute Regression Loss if self.bbox_cov_loss == "pmb_negative_log_likelihood": og_pred_anchor_deltas = pred_anchor_deltas pred_anchor_deltas = pred_anchor_deltas[pos_mask] gt_anchors_deltas = gt_anchor_deltas[pos_mask] if self.compute_bbox_cov: # We have to clamp the output variance else probabilistic metrics # go to infinity. if self.bbox_cov_loss == "pmb_negative_log_likelihood": og_pred_bbox_cov = pred_bbox_cov pred_bbox_cov = clamp_log_variance(pred_bbox_cov[pos_mask]) if self.bbox_cov_loss == "negative_log_likelihood": if self.bbox_cov_type == "diagonal": # Compute regression variance according to: # "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", NIPS 2017 # This implementation with smooth_l1_loss outperforms using # torch.distribution.multivariate_normal. Losses might have different numerical values # since we do not include constants in this implementation. loss_box_reg = ( 0.5 * torch.exp(-pred_bbox_cov) * smooth_l1_loss( pred_anchor_deltas, gt_anchors_deltas, beta=self.smooth_l1_beta, ) ) loss_covariance_regularize = 0.5 * pred_bbox_cov loss_box_reg += loss_covariance_regularize # Sum over all elements loss_box_reg = torch.sum(loss_box_reg) / max( 1, self.loss_normalizer ) else: # Multivariate negative log likelihood. Implemented with # pytorch multivariate_normal.log_prob function. Custom implementations fail to finish training # due to NAN loss. # This is the Cholesky decomposition of the covariance matrix. We reconstruct it from 10 estimated # parameters as a lower triangular matrix. forecaster_cholesky = covariance_output_to_cholesky(pred_bbox_cov) # Compute multivariate normal distribution using torch # distribution functions. multivariate_normal_dists = ( distributions.multivariate_normal.MultivariateNormal( pred_anchor_deltas, scale_tril=forecaster_cholesky ) ) loss_box_reg = -multivariate_normal_dists.log_prob( gt_anchors_deltas ) loss_box_reg = torch.sum(loss_box_reg) / max( 1, self.loss_normalizer ) elif self.bbox_cov_loss == "second_moment_matching": # Compute regression covariance using second moment matching. loss_box_reg = smooth_l1_loss( pred_anchor_deltas, gt_anchors_deltas, beta=self.smooth_l1_beta ) # Compute errors errors = pred_anchor_deltas - gt_anchors_deltas if self.bbox_cov_type == "diagonal": # Compute second moment matching term. second_moment_matching_term = smooth_l1_loss( torch.exp(pred_bbox_cov), errors ** 2, beta=self.smooth_l1_beta ) loss_box_reg += second_moment_matching_term loss_box_reg = torch.sum(loss_box_reg) / max( 1, self.loss_normalizer ) else: # Compute second moment matching term. errors = torch.unsqueeze(errors, 2) gt_error_covar = torch.matmul(errors, torch.transpose(errors, 2, 1)) # This is the cholesky decomposition of the covariance matrix. We reconstruct it from 10 estimated # parameters as a lower triangular matrix. forecaster_cholesky = covariance_output_to_cholesky(pred_bbox_cov) predicted_covar = torch.matmul( forecaster_cholesky, torch.transpose(forecaster_cholesky, 2, 1) ) second_moment_matching_term = smooth_l1_loss( predicted_covar, gt_error_covar, beta=self.smooth_l1_beta, reduction="sum", ) loss_box_reg = ( torch.sum(loss_box_reg) + second_moment_matching_term ) / max(1, self.loss_normalizer) elif self.bbox_cov_loss == "energy_loss": # Compute regression variance according to energy score loss. forecaster_means = pred_anchor_deltas # Compute forecaster cholesky. Takes care of diagonal case # automatically. forecaster_cholesky = covariance_output_to_cholesky(pred_bbox_cov) # Define normal distribution samples. To compute energy score, # we need i+1 samples. # Define per-anchor Distributions multivariate_normal_dists = ( distributions.multivariate_normal.MultivariateNormal( forecaster_means, scale_tril=forecaster_cholesky ) ) # Define Monte-Carlo Samples distributions_samples = multivariate_normal_dists.rsample( (self.bbox_cov_num_samples + 1,) ) distributions_samples_1 = distributions_samples[ 0 : self.bbox_cov_num_samples, :, : ] distributions_samples_2 = distributions_samples[ 1 : self.bbox_cov_num_samples + 1, :, : ] # Compute energy score gt_anchors_deltas_samples = torch.repeat_interleave( gt_anchors_deltas.unsqueeze(0), self.bbox_cov_num_samples, dim=0 ) energy_score_first_term = ( 2.0 * smooth_l1_loss( distributions_samples_1, gt_anchors_deltas_samples, beta=self.smooth_l1_beta, reduction="sum", ) / self.bbox_cov_num_samples ) # First term energy_score_second_term = ( -smooth_l1_loss( distributions_samples_1, distributions_samples_2, beta=self.smooth_l1_beta, reduction="sum", ) / self.bbox_cov_num_samples ) # Second term # Final Loss loss_box_reg = ( energy_score_first_term + energy_score_second_term ) / max(1, self.loss_normalizer) elif self.bbox_cov_loss == "pmb_negative_log_likelihood": pred_class_scores = pred_class_logits.sigmoid() losses = self.nll_od_loss_with_nms( nms_results, gt_instances, anchors, pred_class_scores, og_pred_anchor_deltas, og_pred_bbox_cov, image_sizes, ) loss_box_reg = losses["loss_box_reg"] use_nll_loss = True else: raise ValueError( "Invalid regression loss name {}.".format(self.bbox_cov_loss) ) # Perform loss annealing. Essential for reliably training variance estimates using NLL in RetinaNet. # For energy score and second moment matching, this is optional. standard_regression_loss = ( smooth_l1_loss( pred_anchor_deltas, gt_anchors_deltas, beta=self.smooth_l1_beta, reduction="sum", ) / max(1, self.loss_normalizer) ) probabilistic_loss_weight = get_probabilistic_loss_weight( self.current_step, self.annealing_step ) loss_box_reg = ( 1.0 - probabilistic_loss_weight ) * standard_regression_loss + probabilistic_loss_weight * loss_box_reg if self.bbox_cov_loss == "pmb_negative_log_likelihood": loss_cls = (1.0 - probabilistic_loss_weight) * loss_cls else: # Standard regression loss in case no variance is needed to be # estimated. loss_box_reg = ( smooth_l1_loss( pred_anchor_deltas, gt_anchors_deltas, beta=self.smooth_l1_beta, reduction="sum", ) / max(1, self.loss_normalizer) ) if use_nll_loss: losses["loss_cls"] = loss_cls losses["loss_box_reg"] = loss_box_reg else: losses = {"loss_cls": loss_cls, "loss_box_reg": loss_box_reg} return losses def nll_od_loss_with_nms( self, nms_results, gt_instances, anchors, scores, deltas, pred_covs, image_shapes, ): if "log_prob" in self.matching_distance and self.matching_distance != "log_prob": covar_scaling = float(self.matching_distance.split("_")[-1]) matching_distance = "log_prob" else: covar_scaling = 1 matching_distance = self.matching_distance self.ppp_intensity_function.update_distribution() instances, kept_idx = nms_results bs = len(instances) boxes = [ self.box2box_transform.apply_deltas(delta, anchors) for delta in deltas ] nll_pred_cov = [ pred_cov[kept].unsqueeze(1).repeat(1, self.num_classes, 1) for pred_cov, kept in zip(pred_covs, kept_idx) ] nll_pred_cov = [covariance_output_to_cholesky(cov) for cov in nll_pred_cov] nll_scores = [score[kept] for score, kept in zip(scores, kept_idx)] nll_pred_deltas = [ delta[kept].unsqueeze(1).repeat(1, self.num_classes, 1) for delta, kept in zip(deltas, kept_idx) ] gt_boxes = [instances.gt_boxes.tensor for instances in gt_instances] nll_gt_classes = [instances.gt_classes for instances in gt_instances] kept_proposals = [anchors[idx] for idx in kept_idx] trans_func = lambda x,y: self.box2box_transform.apply_deltas(x,y) box_means = [] box_chols = [] for i in range(bs): box_mean, box_chol = unscented_transform(nll_pred_deltas[i], nll_pred_cov[i], kept_proposals[i], trans_func) box_means.append(box_mean) box_chols.append(box_chol) if self.bbox_cov_dist_type == "gaussian": regression_dist = ( lambda x, y: distributions.multivariate_normal.MultivariateNormal( loc=x, scale_tril=y ) ) elif self.bbox_cov_dist_type == "laplacian": # Map cholesky decomp to laplacian scale regression_dist = lambda x, y: distributions.laplace.Laplace( loc=x, scale=y.diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) else: raise Exception( f"Bounding box uncertainty distribution {self.bbox_cov_dist_type} is not available." ) nll_scores = [ torch.cat( ( nll_scores[i], ( 1 - nll_scores[i][ torch.arange(len(kept_idx[i])), instances[i].pred_classes ] ).unsqueeze(-1), ), dim=-1, ) for i in range(bs) ] # Clamp for numerical stability nll_scores = [scores.clamp(1e-6, 1 - 1e-6) for scores in nll_scores] if self.use_prediction_mixture: ppps = [] src_boxes_tot = [] src_box_chol_tot = [] src_boxes_deltas_tot = [] src_boxes_deltas_chol_tot = [] src_scores_tot = [] gt_box_deltas = [] for i in range(bs): image_shape = image_shapes[i] h,w = image_shape scaling = torch.tensor([1/w,1/h],device=box_means[i].device).repeat(2) pred_box_means = box_means[i]*scaling pred_box_chols = torch.diag_embed(scaling)@box_chols[i] pred_box_deltas = nll_pred_deltas[i] pred_box_delta_chols = nll_pred_cov[i] pred_cls_probs = nll_scores[i] #max_conf = pred_cls_probs[..., :num_classes].max(dim=1)[0] max_conf = 1 - pred_cls_probs[..., -1] ppp_preds_idx = ( max_conf <= self.ppp_intensity_function.ppp_confidence_thres ) props = kept_proposals[i][ppp_preds_idx.logical_not()] # Get delta between each GT and proposal, batch-wise tmp = torch.stack( [ self.box2box_transform.get_deltas( props, gt_boxes[i][j].unsqueeze(0).repeat(len(props), 1), ) for j in range(len(gt_boxes[i])) ] ) gt_box_deltas.append( tmp.permute(1, 0, 2) ) # [gt,pred,boxdim] -> [pred, gt, boxdim] gt_boxes[i] = gt_boxes[i]*scaling mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] mixture_dict["covs"] = pred_box_chols[ppp_preds_idx, 0]@pred_box_chols[ppp_preds_idx, 0].transpose(-1,-2) mixture_dict["cls_probs"] = pred_cls_probs[ppp_preds_idx, :self.num_classes] mixture_dict["reg_dist_type"] = self.bbox_cov_dist_type if self.bbox_cov_dist_type == "gaussian": mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "scale_tril": pred_box_chols[ppp_preds_idx, 0] } elif self.bbox_cov_dist_type == "laplacian": mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": ( pred_box_chols[ppp_preds_idx, 0].diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) } loss_ppp = PoissonPointUnion() loss_ppp.add_ppp(self.ppp_constructor({"predictions": mixture_dict})) loss_ppp.add_ppp(self.ppp_intensity_function) mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] scale_mat = torch.eye(pred_box_chols.shape[-1]).to(pred_box_chols.device)*covar_scaling scaled_chol = scale_mat@pred_box_chols[ppp_preds_idx, 0] mixture_dict["covs"] = (scaled_chol)@(scaled_chol.transpose(-1,-2)) mixture_dict["cls_probs"] = pred_cls_probs[ppp_preds_idx, :self.num_classes] mixture_dict["reg_dist_type"] = self.bbox_cov_dist_type if self.bbox_cov_dist_type == "gaussian": mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "scale_tril": scaled_chol } elif self.bbox_cov_dist_type == "laplacian": mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": ( (scaled_chol).diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) } match_ppp = PoissonPointUnion() match_ppp.add_ppp(self.ppp_constructor({"predictions": mixture_dict})) match_ppp.add_ppp(self.ppp_intensity_function) ppps.append({"matching": match_ppp, "loss": loss_ppp}) src_boxes_tot.append(pred_box_means[ppp_preds_idx.logical_not()]) src_box_chol_tot.append(pred_box_chols[ppp_preds_idx.logical_not()]) src_scores_tot.append(pred_cls_probs[ppp_preds_idx.logical_not()]) src_boxes_deltas_tot.append(pred_box_deltas[ppp_preds_idx.logical_not()]) src_boxes_deltas_chol_tot.append(pred_box_delta_chols[ppp_preds_idx.logical_not()]) nll_pred_deltas = src_boxes_deltas_tot nll_pred_delta_chols = src_boxes_deltas_chol_tot nll_pred_boxes = src_boxes_tot nll_pred_cov = src_box_chol_tot nll_scores = src_scores_tot use_target_delta_matching = False elif self.ppp_intensity_function.ppp_intensity_type == "gaussian_mixture": ppps = [] src_boxes_tot = [] src_box_chol_tot = [] src_boxes_deltas_tot = [] src_boxes_deltas_chol_tot = [] src_scores_tot = [] gt_box_deltas = [] for i in range(bs): image_shape = image_shapes[i] h,w = image_shape scaling = torch.tensor([1/w,1/h],device=box_means[i].device).repeat(2) pred_box_means = box_means[i]*scaling pred_box_chols = torch.diag_embed(scaling)@box_chols[i] pred_box_deltas = nll_pred_deltas[i] pred_box_delta_chols = nll_pred_cov[i] pred_cls_probs = nll_scores[i] props = kept_proposals[i] # Get delta between each GT and proposal, batch-wise tmp = torch.stack( [ self.box2box_transform.get_deltas( props, gt_boxes[i][j].unsqueeze(0).repeat(len(props), 1), ) for j in range(len(gt_boxes[i])) ] ) gt_box_deltas.append( tmp.permute(1, 0, 2) ) # [gt,pred,boxdim] -> [pred, gt, boxdim] gt_boxes[i] = gt_boxes[i]*scaling src_boxes_tot.append(pred_box_means) src_box_chol_tot.append(pred_box_chols) src_scores_tot.append(pred_cls_probs) src_boxes_deltas_tot.append(pred_box_deltas) src_boxes_deltas_chol_tot.append(pred_box_delta_chols) nll_pred_deltas = src_boxes_deltas_tot nll_pred_delta_chols = src_boxes_deltas_chol_tot nll_pred_boxes = src_boxes_tot nll_pred_cov = src_box_chol_tot nll_scores = src_scores_tot use_target_delta_matching = False ppps = [{"loss": self.ppp_intensity_function, "matching": self.ppp_intensity_function}]*bs else: gt_box_deltas = [] for i in range(len(gt_boxes)): # Get delta between each GT and proposal, batch-wise tmp = torch.stack( [ self.box2box_transform.get_deltas( kept_proposals[i], gt_boxes[i][j].unsqueeze(0).repeat(len(kept_proposals[i]), 1), ) for j in range(len(gt_boxes[i])) ] ) gt_box_deltas.append( tmp.permute(1, 0, 2) ) # [gt,pred,boxdim] -> [pred, gt, boxdim] use_target_delta_matching = True ppps = [{"loss": self.ppp_intensity_function, "matching": self.ppp_intensity_function}]*bs nll_pred_delta_chols = nll_pred_cov nll_pred_deltas = nll_pred_deltas nll_pred_boxes = nll_pred_deltas nll_pred_cov = nll_pred_cov nll, associations, decompositions = negative_log_likelihood( nll_scores, nll_pred_boxes, nll_pred_cov, gt_boxes, nll_gt_classes, image_shapes, regression_dist, ppps, self.nll_max_num_solutions, target_deltas=gt_box_deltas, matching_distance=matching_distance, use_target_delta_matching=use_target_delta_matching, pred_deltas=nll_pred_deltas, pred_delta_chols=nll_pred_delta_chols, ) # Save some stats storage = get_event_storage() num_classes = self.num_classes mean_variance = np.mean( [ cov.diagonal(dim1=-2,dim2=-1) .pow(2) .mean() .item() for cov in nll_pred_cov if cov.shape[0] > 0 ] ) storage.put_scalar("nll/mean_covariance", mean_variance) ppp_intens = np.sum([ppp["loss"].integrate( torch.as_tensor(image_shapes).to(self.device), num_classes ) .mean() .item() for ppp in ppps ]) storage.put_scalar("nll/ppp_intensity", ppp_intens) reg_loss = np.mean( [ np.clip( decomp["matched_bernoulli_reg"][0] / (decomp["num_matched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) cls_loss_match = np.mean( [ np.clip( decomp["matched_bernoulli_cls"][0] / (decomp["num_matched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) cls_loss_no_match = np.mean( [ np.clip( decomp["unmatched_bernoulli"][0] / (decomp["num_unmatched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) # Collect all losses losses = dict() losses["loss_box_reg"] = nll # Add losses for logging, these do not propagate gradients losses["loss_regression"] = torch.tensor(reg_loss).to(nll.device) losses["loss_cls_matched"] = torch.tensor(cls_loss_match).to(nll.device) losses["loss_cls_unmatched"] = torch.tensor(cls_loss_no_match).to(nll.device) return losses def produce_raw_output(self, anchors, features): """ Given anchors and features, produces raw pre-nms output to be used for custom fusion operations. """ # Perform inference run ( pred_logits, pred_anchor_deltas, pred_logits_vars, pred_anchor_deltas_vars, ) = self.head(features) # Transpose the Hi*Wi*A dimension to the middle: pred_logits = [permute_to_N_HWA_K(x, self.num_classes) for x in pred_logits] pred_anchor_deltas = [permute_to_N_HWA_K(x, 4) for x in pred_anchor_deltas] if pred_logits_vars is not None: pred_logits_vars = [ permute_to_N_HWA_K(x, self.num_classes) for x in pred_logits_vars ] if pred_anchor_deltas_vars is not None: pred_anchor_deltas_vars = [ permute_to_N_HWA_K(x, self.bbox_cov_dims) for x in pred_anchor_deltas_vars ] # Create raw output dictionary raw_output = {"anchors": anchors} # Shapes: # (N x R, K) for class_logits and class_logits_var. # (N x R, 4), (N x R x 10) for pred_anchor_deltas and pred_class_bbox_cov respectively. raw_output.update( { "box_cls": pred_logits, "box_delta": pred_anchor_deltas, "box_cls_var": pred_logits_vars, "box_reg_var": pred_anchor_deltas_vars, } ) if ( self.compute_bbox_cov and self.bbox_cov_loss == "pmb_negative_log_likelihood" ): ppp_output = self.ppp_intensity_function.get_weights() raw_output.update({"ppp": ppp_output}) return raw_output def inference( self, anchors: List[Boxes], pred_logits: List[Tensor], pred_anchor_deltas: List[Tensor], image_sizes: List[Tuple[int, int]], ): """ Arguments: anchors (list[Boxes]): A list of #feature level Boxes. The Boxes contain anchors of this image on the specific feature level. pred_logits, pred_anchor_deltas: list[Tensor], one per level. Each has shape (N, Hi * Wi * Ai, K or 4) image_sizes (List[(h, w)]): the input image sizes Returns: results (List[Instances]): a list of #images elements. """ results: List[Instances] = [] for img_idx, image_size in enumerate(image_sizes): pred_logits_per_image = [x[img_idx] for x in pred_logits] deltas_per_image = [x[img_idx] for x in pred_anchor_deltas] results_per_image = self.inference_single_image( anchors, pred_logits_per_image, deltas_per_image, image_size ) results.append(results_per_image) return [x[0] for x in results], [x[1] for x in results] def inference_single_image( self, anchors: List[Boxes], box_cls: List[Tensor], box_delta: List[Tensor], image_size: Tuple[int, int], ): """ Single-image inference. Return bounding-box detection results by thresholding on scores and applying non-maximum suppression (NMS). Arguments: anchors (list[Boxes]): list of #feature levels. Each entry contains a Boxes object, which contains all the anchors in that feature level. box_cls (list[Tensor]): list of #feature levels. Each entry contains tensor of size (H x W x A, K) box_delta (list[Tensor]): Same shape as 'box_cls' except that K becomes 4. image_size (tuple(H, W)): a tuple of the image height and width. Returns: Same as `inference`, but for only one image. """ boxes_all = [] scores_all = [] class_idxs_all = [] anchor_idxs_all = [] # Iterate over every feature level for box_cls_i, box_reg_i, anchors_i in zip(box_cls, box_delta, anchors): # (HxWxAxK,) predicted_prob = box_cls_i.flatten().sigmoid() # Apply two filtering below to make NMS faster. # 1. Keep boxes with confidence score higher than threshold keep_idxs = predicted_prob > self.test_score_thresh predicted_prob = predicted_prob[keep_idxs] topk_idxs = nonzero_tuple(keep_idxs)[0] # 2. Keep top k top scoring boxes only num_topk = min(self.test_topk_candidates, topk_idxs.size(0)) # torch.sort is actually faster than .topk (at least on GPUs) predicted_prob, idxs = predicted_prob.sort(descending=True) predicted_prob = predicted_prob[:num_topk] topk_idxs = topk_idxs[idxs[:num_topk]] anchor_idxs = topk_idxs // self.num_classes classes_idxs = topk_idxs % self.num_classes box_reg_i = box_reg_i[anchor_idxs] anchors_i = anchors_i[anchor_idxs] # predict boxes predicted_boxes = self.box2box_transform.apply_deltas( box_reg_i, anchors_i.tensor ) boxes_all.append(predicted_boxes) scores_all.append(predicted_prob) class_idxs_all.append(classes_idxs) anchor_idxs_all.append(anchor_idxs) num_anchors_per_feat_lvl = [anchor.tensor.shape[0] for anchor in anchors] accum_anchor_nums = np.cumsum(num_anchors_per_feat_lvl).tolist() accum_anchor_nums = [0] + accum_anchor_nums anchor_idxs_all = [ anchor_idx + prev_num_feats for anchor_idx, prev_num_feats in zip(anchor_idxs_all, accum_anchor_nums) ] boxes_all, scores_all, class_idxs_all, anchor_idxs_all = [ cat(x) for x in [boxes_all, scores_all, class_idxs_all, anchor_idxs_all] ] keep = batched_nms(boxes_all, scores_all, class_idxs_all, self.test_nms_thresh) keep = keep[: self.max_detections_per_image] result = Instances(image_size) result.pred_boxes = Boxes(boxes_all[keep]) result.scores = scores_all[keep] result.pred_classes = class_idxs_all[keep] return result, anchor_idxs_all[keep] class ProbabilisticRetinaNetHead(RetinaNetHead): """ The head used in ProbabilisticRetinaNet for object class probability estimation, box regression, box covariance estimation. It has three subnets for the three tasks, with a common structure but separate parameters. """ def __init__( self, cfg, use_dropout, dropout_rate, compute_cls_var, compute_bbox_cov, bbox_cov_dims, input_shape: List[ShapeSpec], ): super().__init__(cfg, input_shape) # Extract config information # fmt: off in_channels = input_shape[0].channels num_classes = cfg.MODEL.RETINANET.NUM_CLASSES num_convs = cfg.MODEL.RETINANET.NUM_CONVS prior_prob = cfg.MODEL.RETINANET.PRIOR_PROB num_anchors = build_anchor_generator(cfg, input_shape).num_cell_anchors # fmt: on assert ( len(set(num_anchors)) == 1 ), "Using different number of anchors between levels is not currently supported!" num_anchors = num_anchors[0] self.compute_cls_var = compute_cls_var self.compute_bbox_cov = compute_bbox_cov self.bbox_cov_dims = bbox_cov_dims # For consistency all configs are grabbed from original RetinaNet self.use_dropout = use_dropout self.dropout_rate = dropout_rate cls_subnet = [] bbox_subnet = [] for _ in range(num_convs): cls_subnet.append( nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1) ) cls_subnet.append(nn.ReLU()) bbox_subnet.append( nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1) ) bbox_subnet.append(nn.ReLU()) if self.use_dropout: cls_subnet.append(nn.Dropout(p=self.dropout_rate)) bbox_subnet.append(nn.Dropout(p=self.dropout_rate)) self.cls_subnet = nn.Sequential(*cls_subnet) self.bbox_subnet = nn.Sequential(*bbox_subnet) self.cls_score = nn.Conv2d( in_channels, num_anchors * num_classes, kernel_size=3, stride=1, padding=1 ) self.bbox_pred = nn.Conv2d( in_channels, num_anchors * 4, kernel_size=3, stride=1, padding=1 ) for modules in [ self.cls_subnet, self.bbox_subnet, self.cls_score, self.bbox_pred, ]: for layer in modules.modules(): if isinstance(layer, nn.Conv2d): torch.nn.init.normal_(layer.weight, mean=0, std=0.01) torch.nn.init.constant_(layer.bias, 0) # Use prior in model initialization to improve stability bias_value = -math.log((1 - prior_prob) / prior_prob) torch.nn.init.constant_(self.cls_score.bias, bias_value) # Create subnet for classification variance estimation. if self.compute_cls_var: self.cls_var = nn.Conv2d( in_channels, num_anchors * num_classes, kernel_size=3, stride=1, padding=1, ) for layer in self.cls_var.modules(): if isinstance(layer, nn.Conv2d): torch.nn.init.normal_(layer.weight, mean=0, std=0.01) torch.nn.init.constant_(layer.bias, -10.0) # Create subnet for bounding box covariance estimation. if self.compute_bbox_cov: self.bbox_cov = nn.Conv2d( in_channels, num_anchors * self.bbox_cov_dims, kernel_size=3, stride=1, padding=1, ) for layer in self.bbox_cov.modules(): if isinstance(layer, nn.Conv2d): torch.nn.init.normal_(layer.weight, mean=0, std=0.0001) torch.nn.init.constant_(layer.bias, 0) def forward(self, features): """ Arguments: features (list[Tensor]): FPN feature map tensors in high to low resolution. Each tensor in the list correspond to different feature levels. Returns: logits (list[Tensor]): #lvl tensors, each has shape (N, AxK, Hi, Wi). The tensor predicts the classification probability at each spatial position for each of the A anchors and K object classes. logits_var (list[Tensor]): #lvl tensors, each has shape (N, AxK, Hi, Wi). The tensor predicts the variance of the logits modeled as a univariate Gaussian distribution at each spatial position for each of the A anchors and K object classes. bbox_reg (list[Tensor]): #lvl tensors, each has shape (N, Ax4, Hi, Wi). The tensor predicts 4-vector (dx,dy,dw,dh) box regression values for every anchor. These values are the relative offset between the anchor and the ground truth box. bbox_cov (list[Tensor]): #lvl tensors, each has shape (N, Ax4 or Ax10, Hi, Wi). The tensor predicts elements of the box covariance values for every anchor. The dimensions of the box covarianc depends on estimating a full covariance (10) or a diagonal covariance matrix (4). """ logits = [] bbox_reg = [] logits_var = [] bbox_cov = [] for feature in features: logits.append(self.cls_score(self.cls_subnet(feature))) bbox_reg.append(self.bbox_pred(self.bbox_subnet(feature))) if self.compute_cls_var: logits_var.append(self.cls_var(self.cls_subnet(feature))) if self.compute_bbox_cov: bbox_cov.append(self.bbox_cov(self.bbox_subnet(feature))) return_vector = [logits, bbox_reg] if self.compute_cls_var: return_vector.append(logits_var) else: return_vector.append(None) if self.compute_bbox_cov: return_vector.append(bbox_cov) else: return_vector.append(None) return return_vector
58,037
39.164706
127
py
pmb-nll
pmb-nll-main/src/probabilistic_modeling/modeling_utils.py
import copy import math import torch from sklearn.mixture._gaussian_mixture import _compute_precision_cholesky from torch import nn from torch.distributions import Distribution from torch.distributions.categorical import Categorical from torch.distributions.independent import Independent from torch.distributions.laplace import Laplace from torch.distributions.mixture_same_family import MixtureSameFamily from torch.distributions.multivariate_normal import MultivariateNormal class ClassRegDist(Distribution): def __init__( self, loc, reg_dist, reg_kwargs, probs=None, logits=None, independent_reg_dist=False, ): batch_shape = loc.shape[:-1] event_shape = torch.Size([1 + loc.shape[-1]]) self.reg_dist = reg_dist(loc, **reg_kwargs) if independent_reg_dist: self.reg_dist = Independent(self.reg_dist, 1) self.cls_dist = Categorical(probs=probs, logits=logits) self.dist_type = "log_prob" super().__init__(batch_shape, event_shape, validate_args=False) def log_prob(self, value): cls_log_prob = self.cls_dist.log_prob(value[..., -1]) if self.dist_type == "euclidian": reg_log_prob = -(self.reg_dist.mean - value[..., :-1]).pow(2).sum(-1).sqrt() elif self.dist_type == "euclidian_squared": reg_log_prob = -(self.reg_dist.mean - value[..., :-1]).pow(2).sum(-1) else: reg_log_prob = self.reg_dist.log_prob(value[..., :-1]) return cls_log_prob + reg_log_prob def set_dist_mode(self, dist_type): self.dist_type = dist_type def unscented_transform(means, chols, anchors, trans_func): """ Definition 1 in https://arxiv.org/abs/2104.01958 Args: means (_type_): _description_ chols (_type_): _description_ anchors (_type_): _description_ trans_func (_type_): _description_ Returns: _type_: _description_ """ n = means.shape[-1] kappa = n-3 if len(means.shape) > 2: old_means_shape = means.shape means = means.reshape(-1,n) if len(chols > 3): old_chol_shape = chols.shape chols = chols.reshape(-1,n,n) N = len(means) weights = torch.ones((1,2*n+1,1), device=means.device)/(2*(n+kappa)) weights[0,0,0] = kappa / (n+kappa) # means [N, n], chols [N, n, n] # [N, 1, n] sigma_points1 = means.unsqueeze(1) # [N, n, n] sigma_points2 = means.unsqueeze(1) + math.sqrt(n+kappa)*chols # [N, n, n] sigma_points3 = means.unsqueeze(1) - math.sqrt(n+kappa)*chols # [N, 2n+1, n] sigma_points = torch.cat((sigma_points1, sigma_points2, sigma_points3), dim=1) repeated_anchors = anchors.repeat_interleave(len(means)//len(anchors),dim=0).unsqueeze(1).repeat(1,2*n+1,1).reshape(-1,n) transformed_sigma_points = trans_func(sigma_points.reshape(-1, n), repeated_anchors) transformed_sigma_points = transformed_sigma_points.reshape(N, 2*n+1, n) transformed_means = (transformed_sigma_points*weights).sum(dim=1) residuals = transformed_sigma_points-transformed_means.unsqueeze(1) # [N, 2n+1, n, 1] residuals = residuals.unsqueeze(-1) # [N, n, n] transformed_covs = (weights.unsqueeze(-1)*residuals@residuals.transpose(-1,-2)).sum(dim=1) transformed_chols, info = torch.linalg.cholesky_ex(transformed_covs) if not (info==0).all(): # Clamp to avoid errors transformed_chols = torch.diag_embed(torch.diagonal(transformed_chols,dim1=-2,dim2=-1).clamp(math.exp(-7),math.exp(10)))+torch.tril(transformed_chols,-1) print("***************************") for cov,res,trans_mean,mean,anchor,chol in zip(transformed_covs[info!=0], residuals[info!=0].squeeze(-1), transformed_means[info!=0], means[info!=0], anchors.repeat_interleave(len(means)//len(anchors),dim=0)[info!=0], chols[info!=0]): print(cov) print(res) print(trans_mean) print(mean) print(anchor) print(chol) print("+++++++++++++++++++++++++++++++++++") print("***************************") return transformed_means.reshape(old_means_shape), transformed_chols.reshape(old_chol_shape) def covariance_output_to_cholesky(pred_bbox_cov): """ Transforms output to covariance cholesky decomposition. Args: pred_bbox_cov (kx4 or kx10): Output covariance matrix elements. Returns: predicted_cov_cholesky (kx4x4): cholesky factor matrix """ # Embed diagonal variance if pred_bbox_cov.shape[0] == 0: return pred_bbox_cov.reshape((0, 4, 4)) diag_vars = torch.sqrt(torch.exp(pred_bbox_cov[..., :4])) predicted_cov_cholesky = torch.diag_embed(diag_vars) if pred_bbox_cov.shape[-1] > 4: tril_indices = torch.tril_indices(row=4, col=4, offset=-1) predicted_cov_cholesky[..., tril_indices[0], tril_indices[1]] = pred_bbox_cov[ ..., 4: ] return predicted_cov_cholesky def clamp_log_variance(pred_bbox_cov, clamp_min=-7.0, clamp_max=10.0): """ Tiny function that clamps variance for consistency across all methods. """ pred_bbox_var_component = torch.clamp(pred_bbox_cov[..., 0:4], clamp_min, clamp_max) return torch.cat((pred_bbox_var_component, pred_bbox_cov[..., 4:]), dim=-1) def get_probabilistic_loss_weight(current_step, annealing_step): """ Tiny function to get adaptive probabilistic loss weight for consistency across all methods. """ probabilistic_loss_weight = min(1.0, current_step / annealing_step) probabilistic_loss_weight = (100 ** probabilistic_loss_weight - 1.0) / (100.0 - 1.0) return probabilistic_loss_weight def freeze_non_probabilistic_weights(cfg, model): """ Tiny function to only keep a small subset of weight non-frozen. """ if cfg.MODEL.TRAIN_ONLY_PPP: print("[NLLOD]: Freezing all non-PPP weights") for name, p in model.named_parameters(): if "ppp_intensity_function" in name: p.requires_grad = cfg.MODEL.TRAIN_PPP else: p.requires_grad = False print("[NLLOD]: Froze all non-PPP weights") elif cfg.MODEL.TRAIN_ONLY_UNCERTAINTY_PREDS: print("[NLLOD]: Freezing all non-probabilistic weights") for name, p in model.named_parameters(): if "ppp_intensity_function" in name: p.requires_grad = cfg.MODEL.TRAIN_PPP elif "bbox_cov" in name: p.requires_grad = True else: p.requires_grad = False print("[NLLOD]: Froze all non-probabilistic weights") else: for name, p in model.named_parameters(): if "ppp_intensity_function" in name: p.requires_grad = cfg.MODEL.TRAIN_PPP class PoissonPointProcessBase(nn.Module): def __init__(self): super().__init__() self.normalize_bboxes = False def set_normalization_of_bboxes(self, normalize_bboxes): self.normalize_bboxes = normalize_bboxes class PoissonPointUnion(PoissonPointProcessBase): def __init__(self): super().__init__() self.ppps = [] def add_ppp(self, ppp): self.ppps.append(ppp) def set_normalization_of_bboxes(self, normalize_bboxes): for ppp in self.ppps: ppp.normalize_bboxes = normalize_bboxes def integrate(self, image_sizes, num_classes): out = 0 for ppp in self.ppps: out = out + ppp.integrate(image_sizes, num_classes) return out def forward( self, src, image_sizes=[], num_classes=-1, integrate=False, src_is_features=False, dist_type="log_prob", ): if integrate: out = self.integrate(image_sizes, num_classes) return out outs = [] for ppp in self.ppps: outs.append( ppp(src, image_sizes, num_classes, integrate, src_is_features, dist_type)[:, None] ) outs = torch.cat(outs, 1) return torch.logsumexp(outs, 1) class PoissonPointProcessUniform(PoissonPointProcessBase): def __init__( self, class_dist_log, ppp_rate, uniform_center_pos, device=torch.device("cpu"), ): super().__init__() if not type(class_dist_log) == torch.Tensor: class_dist_log = torch.tensor(class_dist_log) self.class_dist_log = class_dist_log.to(device) self.ppp_rate = torch.tensor([ppp_rate]).to(device) self.uniform_center_pos = uniform_center_pos self.device = device def forward( self, src, image_sizes=[], num_classes=-1, integrate=False, src_is_features=False, ): if integrate: return self.integrate(image_sizes, num_classes) assert len(image_sizes) == 1 img_size = image_sizes[0].flip(0).repeat(2) # w,h,w,h cls_log_probs = self.class_dist_log[src[..., -1].long()] # log(1/(W^2/2) * 1/(H^2/2)) box_log_probs = (-image_sizes[0].log()*2+math.log(2)).sum() total_log_probs = cls_log_probs + box_log_probs + self.ppp_rate.log() return total_log_probs def integrate(self, image_sizes, num_classes): return self.ppp_rate class PoissonPointProcessGMM(PoissonPointProcessBase): def __init__( self, gmm, class_dist_log, ppp_rate, uniform_center_pos, device=torch.device("cpu"), ): super().__init__() if not type(class_dist_log) == torch.Tensor: class_dist_log = torch.tensor(class_dist_log) self.class_dist_log = class_dist_log.to(device) self.gmm = gmm self.ppp_rate = torch.tensor([ppp_rate]).to(device) self.uniform_center_pos = uniform_center_pos self.device = device def forward( self, src, image_sizes=[], num_classes=-1, integrate=False, src_is_features=False, ): if integrate: return self.integrate(image_sizes, num_classes) assert len(image_sizes) == 1 img_size = image_sizes[0].flip(0).repeat(2) # w,h,w,h scale = torch.diag_embed(img_size).cpu().numpy() gmm = copy.deepcopy(self.gmm) boxes = src[..., :-1] if self.uniform_center_pos: gmm.means_ = gmm.means_ * img_size.cpu().numpy()[:2] gmm.covariances_ = scale[:2, :2] @ gmm.covariances_ @ scale[:2, :2].T gmm.precisions_cholesky_ = _compute_precision_cholesky( gmm.covariances_, gmm.covariance_type ) img_area = img_size[0] * img_size[1] # N, 2 (w,h) box_sizes = torch.cat( ( (boxes[..., 2] - boxes[..., 0])[:, None], (boxes[..., 3] - boxes[..., 1])[:, None], ), 1, ) box_log_probs = torch.tensor(gmm.score_samples(box_sizes.cpu().numpy())).to( box_sizes.device ) box_log_probs = box_log_probs - img_area.log() else: gmm.means_ = gmm.means_ * img_size.cpu().numpy() gmm.covariances_ = scale @ gmm.covariances_ @ scale.T gmm.precisions_cholesky_ = _compute_precision_cholesky( gmm.covariances_, gmm.covariance_type ) box_log_probs = torch.tensor(gmm.score_samples(boxes.cpu().numpy())).to( boxes.device ) cls_log_probs = self.class_dist_log[src[..., -1].long()] total_log_probs = cls_log_probs + box_log_probs + self.ppp_rate.log() return total_log_probs def integrate(self, image_sizes, num_classes): return self.ppp_rate class ZeroDistribution(PoissonPointProcessBase): def __init__(self, device=torch.device("cuda"))-> None: super().__init__() self.device = device self.component_distribution = None def log_prob(self, src, *args, **kwargs): return torch.tensor(0.0).to(src.device).unsqueeze(0).repeat(len(src)).log() class PoissonPointProcessIntensityFunction(PoissonPointProcessBase): """ Class representing a Poisson Point Process RFS intensity function. Currently assuming DETR/RCNN/RetinaNet. """ def __init__( self, cfg, log_intensity=None, ppp_feature_net=None, predictions=None, device="cuda" ) -> None: super().__init__() self.device = device if cfg.PROBABILISTIC_INFERENCE.PPP_CONFIDENCE_THRES and predictions is not None: self.ppp_intensity_type = "prediction_mixture" elif log_intensity is not None: self.ppp_intensity_type = "uniform" self.num_classes = 1 else: self.ppp_intensity_type = ( cfg.MODEL.PROBABILISTIC_MODELING.PPP.INTENSITY_TYPE ) self.num_classes = cfg.MODEL.ROI_HEADS.NUM_CLASSES self.ppp_confidence_thres = cfg.PROBABILISTIC_INFERENCE.PPP_CONFIDENCE_THRES self.ppp_feature_net = ppp_feature_net if self.ppp_intensity_type == "uniform": self.ppp_intensity_per_coord = nn.Parameter( torch.tensor(1.0).to(self.device), requires_grad=True ) self.log_ppp_intensity_class = nn.Parameter( torch.tensor(1.0).to(self.device), requires_grad=True ) if log_intensity is None: nn.init.constant_( self.ppp_intensity_per_coord, cfg.MODEL.PROBABILISTIC_MODELING.PPP.UNIFORM_INTENSITY, ) nn.init.constant_( self.log_ppp_intensity_class, math.log(1 / cfg.MODEL.ROI_HEADS.NUM_CLASSES), ) else: nn.init.constant_(self.ppp_intensity_per_coord, log_intensity) nn.init.constant_(self.log_ppp_intensity_class, 0) self.log_ppp_intensity_class.requires_grad = False elif self.ppp_intensity_type == "gaussian_mixture": num_mixture_comps = cfg.MODEL.PROBABILISTIC_MODELING.PPP.NUM_GAUSS_MIXTURES cov_type = cfg.MODEL.PROBABILISTIC_MODELING.PPP.COV_TYPE if cov_type == "diagonal": cov_dims = 4 elif cov_type == "full": cov_dims = 10 else: cov_dims = 4 self.log_gmm_weights = nn.Parameter( (torch.ones(num_mixture_comps)*0.5).log().to(self.device), requires_grad=True, ) nn.init.normal_(self.log_gmm_weights, mean=0, std=0.1) means = torch.distributions.Normal(torch.tensor([0.5]).to(self.device), scale=torch.tensor([0.16]).to(self.device)).rsample((num_mixture_comps, 4,)).squeeze(-1) xywh_to_xyxy = torch.tensor([[1,0,-0.5,0],[0,1,0,-0.5],[1,0,0.5,0],[0,1,0,0.5]]).to(self.device) means = (xywh_to_xyxy@(means.unsqueeze(-1))).squeeze(-1) means = means.clamp(0,1) self.gmm_means = nn.Parameter( means, requires_grad=True ) self.gmm_chols = nn.Parameter( torch.zeros(num_mixture_comps, cov_dims).to(self.device), requires_grad=True ) nn.init.normal_(self.gmm_chols, std=1) cls_probs = torch.ones(num_mixture_comps, self.num_classes).to(self.device)/self.num_classes + torch.rand((num_mixture_comps, self.num_classes)).to(self.device)*0.1 cls_logits = (cls_probs/(1-cls_probs)).log() self.class_logits = nn.Parameter( cls_logits, requires_grad=True ) # these are softmaxed later #self.mvn = MultivariateNormal(self.gmm_means, scale_tril=self.gmm_chols) reg_kwargs = {"scale_tril": covariance_output_to_cholesky(self.gmm_chols)} mixture_dict = {} mixture_dict["means"] = self.gmm_means mixture_dict["weights"] = self.log_gmm_weights.exp() mixture_dict["reg_dist"] = torch.distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = reg_kwargs mixture_dict["cls_probs"] = self.class_logits.softmax(dim=-1) mixture_dict["reg_dist_type"] = "gaussian" mixture_dict["covs"] = None self.mixture_from_predictions(mixture_dict) elif self.ppp_intensity_type == "prediction_mixture": if predictions is not None: self.mixture_from_predictions(predictions) elif self.ppp_intensity_type == "zero": self.dist = ZeroDistribution(self.device) else: raise NotImplementedError( f"PPP intensity type {cfg.MODEL.PROBABILISTIC_MODELING.PPP_INTENSITY_TYPE} not implemented." ) def mixture_from_predictions(self, mixture_dict): reg_dist_str = mixture_dict["reg_dist_type"] means = mixture_dict["means"] covs = mixture_dict["covs"] weights = mixture_dict["weights"] cls_probs = mixture_dict["cls_probs"] reg_kwargs = mixture_dict["reg_kwargs"] independent_reg_dist = False reg_dist = mixture_dict["reg_dist"] if reg_dist_str == "laplacian": independent_reg_dist = True if not len(weights): self.mixture_dist = ZeroDistribution(means.device) self.ppp_rate = torch.tensor(0.0).to(means.device) else: self.mixture_dist = MixtureSameFamily( Categorical(weights), ClassRegDist( means, reg_dist, reg_kwargs, probs=cls_probs, independent_reg_dist=independent_reg_dist, ), validate_args=False, ) self.ppp_rate = weights.sum() def get_weights(self): weights = dict() if self.ppp_intensity_type == "uniform": weights["ppp_intensity_per_coord"] = self.ppp_intensity_per_coord weights["log_ppp_intensity_class"] = self.log_ppp_intensity_class elif self.ppp_intensity_type == "gaussian_mixture": return weights weights["log_gmm_weights"] = self.log_gmm_weights weights["gmm_means"] = self.gmm_means weights["gmm_covs"] = self.gmm_covs weights["class_weights"] = self.class_weights weights["log_class_scaling"] = self.log_class_scaling return weights def load_weights(self, weights): if self.ppp_intensity_type == "uniform": self.ppp_intensity_per_coord = nn.Parameter( torch.as_tensor(weights["ppp_intensity_per_coord"]) ) self.log_ppp_intensity_class = nn.Parameter( torch.as_tensor(weights["log_ppp_intensity_class"]) ) elif self.ppp_intensity_type == "gaussian_mixture": self.log_gmm_weights = nn.Parameter( torch.as_tensor(weights["log_gmm_weights"]) ) self.gmm_means = nn.Parameter(torch.as_tensor(weights["gmm_means"])) self.gmm_covs = nn.Parameter(torch.as_tensor(weights["gmm_covs"])) self.class_weights = nn.Parameter(torch.as_tensor(weights["class_weights"])) self.log_class_scaling = nn.Parameter( torch.as_tensor(weights["log_class_scaling"]) ) self.update_distribution() def update_distribution(self): if self.ppp_intensity_type == "gaussian_mixture": mixture_dict = {} mixture_dict["means"] = self.gmm_means mixture_dict["weights"] = self.log_gmm_weights.exp() mixture_dict["reg_dist"] = torch.distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = {"scale_tril": covariance_output_to_cholesky(self.gmm_chols)} mixture_dict["cls_probs"] = self.class_logits.softmax(dim=-1) mixture_dict["reg_dist_type"] = "gaussian" mixture_dict["covs"] = None self.mixture_from_predictions(mixture_dict) def forward_features(self, src): print("[NLLOD] Data dependent PPP not available yet") return out = self.ppp_feature_net(src) if self.ppp_intensity_type == "gaussian_mixture": pass # translate output to gmm params return def forward( self, src, image_sizes=[], num_classes=-1, integrate=False, src_is_features=False, dist_type="log_prob" ): """Calculate log PPP intensity for given input. If numclasses =! -1, returns integral over intensity Args: src ([type]): [description] image_sizes (list, optional): [description]. Defaults to []. num_classes (int, optional): [description]. Defaults to -1. Returns: [type]: [description] """ if src_is_features: return self.forward_features(src) if integrate: return self.integrate(image_sizes, num_classes) if self.ppp_intensity_type == "uniform": # Returns log intensity func value coord_log_prob = self.ppp_intensity_per_coord if src.shape[-1] > 4: src = src[..., :4] # keep gradients trough src, +1 to handle coodinates in zero out = (src + 1) / (src.detach() + 1) * coord_log_prob out = out.sum(-1) class_log_prob = self.log_ppp_intensity_class out = out + class_log_prob elif self.ppp_intensity_type == "gaussian_mixture": if self.normalize_bboxes: # H,W -> (flip) -> W,H -> (repeat) -> W,H,W,H box_scaling = 1/image_sizes.flip((-1)).repeat(1,2).float() class_scaling = torch.ones((len(image_sizes),1)).to(src.device) # [1, 5] scaling = torch.cat([box_scaling, class_scaling], dim=-1) # [num_gt, 5] scaling = scaling.repeat(src.shape[0],1) src = src*scaling else: scaling = torch.ones_like(src) if self.mixture_dist.component_distribution: self.mixture_dist.component_distribution.set_dist_mode(dist_type) out = self.mixture_dist.log_prob(src) out = out + self.ppp_rate.log() out = out + scaling.log().sum(dim=-1) elif self.ppp_intensity_type == "prediction_mixture": if self.mixture_dist.component_distribution: self.mixture_dist.component_distribution.set_dist_mode(dist_type) out = self.mixture_dist.log_prob(src) out = out + self.ppp_rate.log() elif self.ppp_intensity_type == "zero": out = self.dist.log_prob(src) return out def integrate(self, image_sizes, num_classes): if self.ppp_intensity_type == "uniform": # Evaluate the integral of the intensity funciton of all possible inputs coord_log_prob = self.ppp_intensity_per_coord class_log_prob = self.log_ppp_intensity_class # Divide by 2 because x1 < x2 and y1 < y2 image_part = torch.log( image_sizes[:, 0] ** 2 / 2 * image_sizes[:, 1] ** 2 / 2 ) + (4 * coord_log_prob) class_part = math.log(num_classes) + class_log_prob out = (image_part + class_part).exp() elif self.ppp_intensity_type == "gaussian_mixture": out = self.ppp_rate elif self.ppp_intensity_type == "prediction_mixture": out = self.ppp_rate elif self.ppp_intensity_type == "zero": out = torch.zeros(len(image_sizes)).to(image_sizes.device) else: out = torch.zeros(len(image_sizes)).to(image_sizes.device) return out
24,254
36.488408
242
py
pmb-nll
pmb-nll-main/src/probabilistic_modeling/probabilistic_generalized_rcnn.py
import logging from typing import Dict, List, Optional, Tuple, Union # Detectron imports import fvcore.nn.weight_init as weight_init import numpy as np import torch from detectron2.config import configurable from detectron2.data.detection_utils import convert_image_to_rgb from detectron2.layers import Conv2d, Linear, ShapeSpec, cat, get_norm from detectron2.modeling.box_regression import Box2BoxTransform from detectron2.modeling.meta_arch.build import META_ARCH_REGISTRY from detectron2.modeling.meta_arch.rcnn import GeneralizedRCNN from detectron2.modeling.roi_heads import ROI_HEADS_REGISTRY, StandardROIHeads from detectron2.modeling.roi_heads.box_head import ROI_BOX_HEAD_REGISTRY from detectron2.modeling.roi_heads.fast_rcnn import fast_rcnn_inference from detectron2.structures import Boxes, ImageList, Instances from detectron2.utils.events import get_event_storage from detectron2.utils.logger import log_first_n from fvcore.nn import smooth_l1_loss # Project imports from probabilistic_inference.inference_utils import get_dir_alphas from torch import distributions, nn from torch.nn import functional as F from probabilistic_modeling.losses import negative_log_likelihood, reshape_box_preds from probabilistic_modeling.modeling_utils import ( PoissonPointProcessIntensityFunction, clamp_log_variance, covariance_output_to_cholesky, get_probabilistic_loss_weight, unscented_transform, PoissonPointUnion, ) device = torch.device("cuda" if torch.cuda.is_available() else "cpu") @META_ARCH_REGISTRY.register() class ProbabilisticGeneralizedRCNN(GeneralizedRCNN): """ Probabilistic GeneralizedRCNN class. """ def __init__(self, cfg): super().__init__(cfg) # Parse configs self.cls_var_loss = cfg.MODEL.PROBABILISTIC_MODELING.CLS_VAR_LOSS.NAME self.compute_cls_var = self.cls_var_loss != "none" self.cls_var_num_samples = ( cfg.MODEL.PROBABILISTIC_MODELING.CLS_VAR_LOSS.NUM_SAMPLES ) self.bbox_cov_loss = cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.NAME self.compute_bbox_cov = self.bbox_cov_loss != "none" self.bbox_cov_num_samples = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.NUM_SAMPLES ) self.bbox_cov_dist_type = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE ) self.bbox_cov_type = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.COVARIANCE_TYPE ) if self.bbox_cov_type == "diagonal": # Diagonal covariance matrix has N elements self.bbox_cov_dims = 4 else: # Number of elements required to describe an NxN covariance matrix is # computed as: (N * (N + 1)) / 2 self.bbox_cov_dims = 10 self.dropout_rate = cfg.MODEL.PROBABILISTIC_MODELING.DROPOUT_RATE self.use_dropout = self.dropout_rate != 0.0 self.num_mc_dropout_runs = -1 if ( self.compute_bbox_cov and self.bbox_cov_loss == "pmb_negative_log_likelihood" ): ppp_constructor = lambda x: PoissonPointProcessIntensityFunction( cfg, **x ) self.nll_max_num_solutions = ( cfg.MODEL.PROBABILISTIC_MODELING.NLL_MAX_NUM_SOLUTIONS ) self.current_step = 0 # Define custom probabilistic head self.roi_heads.box_predictor = ProbabilisticFastRCNNOutputLayers( cfg, input_shape=self.roi_heads.box_head.output_shape, compute_cls_var=self.compute_cls_var, cls_var_loss=self.cls_var_loss, cls_var_num_samples=self.cls_var_num_samples, compute_bbox_cov=self.compute_bbox_cov, bbox_cov_loss=self.bbox_cov_loss, bbox_cov_type=self.bbox_cov_type, bbox_cov_dims=self.bbox_cov_dims, bbox_cov_num_samples=self.bbox_cov_num_samples, ppp_constructor=ppp_constructor, nll_max_num_solutions=self.nll_max_num_solutions, bbox_cov_dist_type=self.bbox_cov_dist_type, matching_distance=cfg.MODEL.PROBABILISTIC_MODELING.MATCHING_DISTANCE, use_prediction_mixture=cfg.MODEL.PROBABILISTIC_MODELING.PPP.USE_PREDICTION_MIXTURE, ) # Send to device self.to(self.device) def get_ppp_intensity_function(self): return self.roi_heads.box_predictor.ppp_intensity_function def forward( self, batched_inputs, return_anchorwise_output=False, num_mc_dropout_runs=-1 ): """ Args: batched_inputs: a list, batched outputs of :class:`DatasetMapper` . Each item in the list contains the inputs for one image. For now, each item in the list is a dict that contains: * image: Tensor, image in (C, H, W) format. * instances (optional): groundtruth :class:`Instances` * proposals (optional): :class:`Instances`, precomputed proposals. Other information that's included in the original dicts, such as: * "height", "width" (int): the output resolution of the model, used in inference. See :meth:`postprocess` for details. return_anchorwise_output (bool): returns raw output for probabilistic inference num_mc_dropout_runs (int): perform efficient monte-carlo dropout runs by running only the head and not full neural network. Returns: dict[str: Tensor]: mapping from a named loss to a tensor storing the loss. Used during training only. """ try: self.current_step += get_event_storage().iter except: self.current_step += 1 if not self.training and num_mc_dropout_runs == -1: if return_anchorwise_output: return self.produce_raw_output(batched_inputs) else: return self.inference(batched_inputs) elif self.training and num_mc_dropout_runs > 1: self.num_mc_dropout_runs = num_mc_dropout_runs output_list = [] for i in range(num_mc_dropout_runs): output_list.append(self.produce_raw_output(batched_inputs)) return output_list images = self.preprocess_image(batched_inputs) if "instances" in batched_inputs[0]: gt_instances = [x["instances"].to(self.device) for x in batched_inputs] elif "targets" in batched_inputs[0]: log_first_n( logging.WARN, "'targets' in the model inputs is now renamed to 'instances'!", n=10, ) gt_instances = [x["targets"].to(self.device) for x in batched_inputs] else: gt_instances = None features = self.backbone(images.tensor) if self.proposal_generator: proposals, proposal_losses = self.proposal_generator( images, features, gt_instances ) else: assert "proposals" in batched_inputs[0] proposals = [x["proposals"].to(self.device) for x in batched_inputs] proposal_losses = {} _, detector_losses = self.roi_heads( images, features, proposals, gt_instances, current_step=self.current_step ) if self.vis_period > 0: storage = get_event_storage() if storage.iter % self.vis_period == 0: # TODO: implement to visualize probabilistic outputs self.visualize_training(batched_inputs, proposals) losses = {} losses.update(detector_losses) losses.update(proposal_losses) return losses def produce_raw_output(self, batched_inputs, detected_instances=None): """ Run inference on the given inputs and return proposal-wise output for later postprocessing. Args: batched_inputs (list[dict]): same as in :meth:`forward` detected_instances (None or list[Instances]): if not None, it contains an `Instances` object per image. The `Instances` object contains "pred_boxes" and "pred_classes" which are known boxes in the image. The inference will then skip the detection of bounding boxes, and only predict other per-ROI outputs. Returns: same as in :meth:`forward`. """ raw_output = dict() images = self.preprocess_image(batched_inputs) features = self.backbone(images.tensor) if detected_instances is None: if self.proposal_generator: proposals, _ = self.proposal_generator(images, features, None) else: assert "proposals" in batched_inputs[0] proposals = [x["proposals"].to(self.device) for x in batched_inputs] # Create raw output dictionary raw_output.update({"proposals": proposals[0]}) results, _ = self.roi_heads( images, features, proposals, None, produce_raw_output=True, num_mc_dropout_runs=self.num_mc_dropout_runs, ) else: detected_instances = [x.to(self.device) for x in detected_instances] results = self.roi_heads.forward_with_given_boxes( features, detected_instances ) box_cls, box_delta, box_cls_var, box_reg_var = results raw_output.update( { "box_cls": box_cls, "box_delta": box_delta, "box_cls_var": box_cls_var, "box_reg_var": box_reg_var, } ) if ( self.compute_bbox_cov and self.bbox_cov_loss == "pmb_negative_log_likelihood" ): ppp_output = ( self.roi_heads.box_predictor.ppp_intensity_function.get_weights() ) raw_output.update({"ppp": ppp_output}) return raw_output def visualize_training(self, batched_inputs, proposals): """ A function used to visualize images and proposals. It shows ground truth bounding boxes on the original image and up to 20 top-scoring predicted object proposals on the original image. Users can implement different visualization functions for different models. Args: batched_inputs (list): a list that contains input to the model. proposals (list): a list that contains predicted proposals. Both batched_inputs and proposals should have the same length. """ from core.visualization_tools.probabilistic_visualizer import ( ProbabilisticVisualizer as Visualizer, ) storage = get_event_storage() max_vis_prop = 20 with torch.no_grad(): self.eval() predictions = self.produce_raw_output(batched_inputs) self.train() predictions = ( predictions["box_cls"], predictions["box_delta"], predictions["box_cls_var"], predictions["box_reg_var"], ) _, _, _, pred_covs = predictions boxes = self.roi_heads.box_predictor.predict_boxes(predictions, proposals) scores = self.roi_heads.box_predictor.predict_probs(predictions, proposals) image_shapes = [x.image_size for x in proposals] # Apply NMS without score threshold instances, kept_idx = fast_rcnn_inference( boxes, scores, image_shapes, 0.0, self.roi_heads.box_predictor.test_nms_thresh, self.roi_heads.box_predictor.test_topk_per_image, ) num_prop_per_image = [len(p) for p in proposals] pred_covs = pred_covs.split(num_prop_per_image) pred_covs = [pred_cov[kept] for pred_cov, kept in zip(pred_covs, kept_idx)] pred_scores = [score[kept] for score, kept in zip(scores, kept_idx)] pred_boxes = [box[kept] for box, kept in zip(boxes, kept_idx)] for i, (input, prop) in enumerate(zip(batched_inputs, proposals)): img = input["image"] img = convert_image_to_rgb(img.permute(1, 2, 0), self.input_format) v_gt = Visualizer(img, None) v_gt = v_gt.overlay_instances(boxes=input["instances"].gt_boxes) anno_img = v_gt.get_image() box_size = min(len(prop.proposal_boxes), max_vis_prop) v_pred = Visualizer(img, None) boxes = pred_boxes[i][0:box_size, :4].cpu().numpy() pred_cov_matrix = pred_covs[i][0:box_size, :4] pred_cov_matrix = clamp_log_variance(pred_cov_matrix) chol = covariance_output_to_cholesky(pred_cov_matrix) cov = ( torch.matmul(chol, torch.transpose(chol, -1, -2)).cpu().detach().numpy() ) v_pred = v_pred.overlay_covariance_instances( boxes=boxes, covariance_matrices=cov ) prop_img = v_pred.get_image() vis_img = np.concatenate((anno_img, prop_img), axis=1) vis_img = vis_img.transpose(2, 0, 1) vis_name = "Left: GT bounding boxes; Right: Predicted proposals" storage.put_image(vis_name, vis_img) break # only visualize one image in a batch @ROI_HEADS_REGISTRY.register() class ProbabilisticROIHeads(StandardROIHeads): """ Probabilistic ROI heads, inherit from standard ROI heads so can be used with mask RCNN in theory. """ def __init__(self, cfg, input_shape): super(ProbabilisticROIHeads, self).__init__(cfg, input_shape) self.is_mc_dropout_inference = False self.produce_raw_output = False self.current_step = 0 def forward( self, images: ImageList, features: Dict[str, torch.Tensor], proposals: List[Instances], targets: Optional[List[Instances]] = None, num_mc_dropout_runs=-1, produce_raw_output=False, current_step=0.0, ) -> Tuple[List[Instances], Dict[str, torch.Tensor]]: """ See :class:`ROIHeads.forward`. """ self.is_mc_dropout_inference = num_mc_dropout_runs > 1 self.produce_raw_output = produce_raw_output self.current_step = current_step del images if self.training and not self.is_mc_dropout_inference: assert targets proposals = self.label_and_sample_proposals(proposals, targets) # del targets if self.training and not self.is_mc_dropout_inference: losses = self._forward_box(features, proposals, targets) # Usually the original proposals used by the box head are used by the mask, keypoint # heads. But when `self.train_on_pred_boxes is True`, proposals will contain boxes # predicted by the box head. losses.update(self._forward_mask(features, proposals)) losses.update(self._forward_keypoint(features, proposals)) return proposals, losses else: pred_instances = self._forward_box(features, proposals, targets) if self.produce_raw_output: return pred_instances, {} # During inference cascaded prediction is used: the mask and keypoints heads are only # applied to the top scoring box detections. pred_instances = self.forward_with_given_boxes(features, pred_instances) return pred_instances, {} def _forward_box( self, features: Dict[str, torch.Tensor], proposals: List[Instances], gt_instances: List[Instances], ) -> Union[Dict[str, torch.Tensor], List[Instances]]: """ Forward logic of the box prediction branch. If `self.train_on_pred_boxes is True`, the function puts predicted boxes in the `proposal_boxes` field of `proposals` argument. Args: features (dict[str, Tensor]): mapping from feature map names to tensor. Same as in :meth:`ROIHeads.forward`. proposals (list[Instances]): the per-image object proposals with their matching ground truth. Each has fields "proposal_boxes", and "objectness_logits", "gt_classes", "gt_boxes". Returns: In training, a dict of losses. In inference, a list of `Instances`, the predicted instances. """ features = [features[f] for f in self.in_features] box_features = self.box_pooler(features, [x.proposal_boxes for x in proposals]) box_features = self.box_head(box_features) predictions = self.box_predictor(box_features) del box_features if self.produce_raw_output: return predictions if self.training: losses = self.box_predictor.losses( predictions, proposals, self.current_step, gt_instances ) # proposals is modified in-place below, so losses must be computed first. if self.train_on_pred_boxes: with torch.no_grad(): pred_boxes = self.box_predictor.predict_boxes_for_gt_classes( predictions, proposals ) for proposals_per_image, pred_boxes_per_image in zip( proposals, pred_boxes ): proposals_per_image.proposal_boxes = Boxes(pred_boxes_per_image) return losses else: pred_instances, _ = self.box_predictor.inference(predictions, proposals) return pred_instances class ProbabilisticFastRCNNOutputLayers(nn.Module): """ Four linear layers for predicting Fast R-CNN outputs: (1) proposal-to-detection box regression deltas (2) classification scores (3) box regression deltas covariance parameters (if needed) (4) classification logits variance (if needed) """ @configurable def __init__( self, input_shape, *, box2box_transform, num_classes, cls_agnostic_bbox_reg=False, smooth_l1_beta=0.0, test_score_thresh=0.0, test_nms_thresh=0.5, test_topk_per_image=100, compute_cls_var=False, compute_bbox_cov=False, bbox_cov_dims=4, cls_var_loss="none", cls_var_num_samples=10, bbox_cov_loss="none", bbox_cov_type="diagonal", dropout_rate=0.0, annealing_step=0, bbox_cov_num_samples=1000, ppp_constructor=None, nll_max_num_solutions=5, bbox_cov_dist_type=None, matching_distance="log_prob", use_prediction_mixture=False, ): """ NOTE: this interface is experimental. Args: input_shape (ShapeSpec): shape of the input feature to this module box2box_transform (Box2BoxTransform or Box2BoxTransformRotated): num_classes (int): number of foreground classes cls_agnostic_bbox_reg (bool): whether to use class agnostic for bbox regression smooth_l1_beta (float): transition point from L1 to L2 loss. test_score_thresh (float): threshold to filter predictions results. test_nms_thresh (float): NMS threshold for prediction results. test_topk_per_image (int): number of top predictions to produce per image. compute_cls_var (bool): compute classification variance compute_bbox_cov (bool): compute box covariance regression parameters. bbox_cov_dims (int): 4 for diagonal covariance, 10 for full covariance. cls_var_loss (str): name of classification variance loss. cls_var_num_samples (int): number of samples to be used for loss computation. Usually between 10-100. bbox_cov_loss (str): name of box covariance loss. bbox_cov_type (str): 'diagonal' or 'full'. This is used to train with loss functions that accept both types. dropout_rate (float): 0-1, probability of drop. annealing_step (int): step used for KL-divergence in evidential loss to fully be functional. ppp_intensity_function (func): function that returns PPP intensity given sample box nll_max_num_solutions (int): Maximum NLL solutions to consider when computing NLL-PMB loss """ super().__init__() if isinstance(input_shape, int): # some backward compatibility input_shape = ShapeSpec(channels=input_shape) self.num_classes = num_classes input_size = ( input_shape.channels * (input_shape.width or 1) * (input_shape.height or 1) ) self.compute_cls_var = compute_cls_var self.compute_bbox_cov = compute_bbox_cov self.bbox_cov_dims = bbox_cov_dims self.bbox_cov_num_samples = bbox_cov_num_samples self.dropout_rate = dropout_rate self.use_dropout = self.dropout_rate != 0.0 self.cls_var_loss = cls_var_loss self.cls_var_num_samples = cls_var_num_samples self.annealing_step = annealing_step self.bbox_cov_loss = bbox_cov_loss self.bbox_cov_type = bbox_cov_type self.bbox_cov_dist_type = bbox_cov_dist_type # The prediction layer for num_classes foreground classes and one background class # (hence + 1) self.cls_score = Linear(input_size, num_classes + 1) num_bbox_reg_classes = 1.0 if cls_agnostic_bbox_reg else num_classes box_dim = len(box2box_transform.weights) self.bbox_pred = Linear(input_size, num_bbox_reg_classes * box_dim) nn.init.normal_(self.cls_score.weight, std=0.01) nn.init.normal_(self.bbox_pred.weight, std=0.001) for l in [self.cls_score, self.bbox_pred]: nn.init.constant_(l.bias, 0) if self.compute_cls_var: self.cls_var = Linear(input_size, num_classes + 1) nn.init.normal_(self.cls_var.weight, std=0.0001) nn.init.constant_(self.cls_var.bias, 0) if self.compute_bbox_cov: self.bbox_cov = Linear(input_size, num_bbox_reg_classes * bbox_cov_dims) nn.init.normal_(self.bbox_cov.weight, std=0.0001) nn.init.constant_(self.bbox_cov.bias, 0.0) self.box2box_transform = box2box_transform self.smooth_l1_beta = smooth_l1_beta self.test_score_thresh = test_score_thresh self.test_nms_thresh = test_nms_thresh self.test_topk_per_image = test_topk_per_image self.ppp_intensity_function = ppp_constructor({"device": device}) if ppp_constructor is not None else None self.ppp_constructor = ppp_constructor self.nll_max_num_solutions = nll_max_num_solutions self.matching_distance = matching_distance self.use_prediction_mixture = use_prediction_mixture @classmethod def from_config( cls, cfg, input_shape, compute_cls_var, cls_var_loss, cls_var_num_samples, compute_bbox_cov, bbox_cov_loss, bbox_cov_type, bbox_cov_dims, bbox_cov_num_samples, ppp_constructor, nll_max_num_solutions, ): return { "input_shape": input_shape, "box2box_transform": Box2BoxTransform( weights=cfg.MODEL.ROI_BOX_HEAD.BBOX_REG_WEIGHTS ), # fmt: off "num_classes": cfg.MODEL.ROI_HEADS.NUM_CLASSES, "cls_agnostic_bbox_reg": cfg.MODEL.ROI_BOX_HEAD.CLS_AGNOSTIC_BBOX_REG, "smooth_l1_beta": cfg.MODEL.ROI_BOX_HEAD.SMOOTH_L1_BETA, "test_score_thresh": cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST, "test_nms_thresh": cfg.MODEL.ROI_HEADS.NMS_THRESH_TEST, "test_topk_per_image": cfg.TEST.DETECTIONS_PER_IMAGE, "compute_cls_var": compute_cls_var, "cls_var_loss": cls_var_loss, "cls_var_num_samples": cls_var_num_samples, "compute_bbox_cov": compute_bbox_cov, "bbox_cov_dims": bbox_cov_dims, "bbox_cov_loss": bbox_cov_loss, "bbox_cov_type": bbox_cov_type, "dropout_rate": cfg.MODEL.PROBABILISTIC_MODELING.DROPOUT_RATE, "annealing_step": cfg.SOLVER.STEPS[1] if cfg.MODEL.PROBABILISTIC_MODELING.ANNEALING_STEP <= 0 else cfg.MODEL.PROBABILISTIC_MODELING.ANNEALING_STEP, "bbox_cov_num_samples": bbox_cov_num_samples, "ppp_constructor": ppp_constructor, "nll_max_num_solutions" : nll_max_num_solutions, 'bbox_cov_dist_type': cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.COVARIANCE_TYPE, "use_prediction_mixture": cfg.MODEL.PROBABILISTIC_MODELING.PPP.USE_PREDICTION_MIXTURE # fmt: on } def forward(self, x): """ Args: x: per-region features of shape (N, ...) for N bounding boxes to predict. Returns: Tensor: Nx(K+1) logits for each box Tensor: Nx4 or Nx(Kx4) bounding box regression deltas. Tensor: Nx(K+1) logits variance for each box. Tensor: Nx4(10) or Nx(Kx4(10)) covariance matrix parameters. 4 if diagonal, 10 if full. """ if x.dim() > 2: x = torch.flatten(x, start_dim=1) scores = self.cls_score(x) proposal_deltas = self.bbox_pred(x) # Compute logits variance if needed if self.compute_cls_var: score_vars = self.cls_var(x) else: score_vars = None # Compute box covariance if needed if self.compute_bbox_cov: proposal_covs = self.bbox_cov(x) else: proposal_covs = None return scores, proposal_deltas, score_vars, proposal_covs def losses(self, predictions, proposals, current_step=0, gt_instances=None): """ Args: predictions: return values of :meth:`forward()`. proposals (list[Instances]): proposals that match the features that were used to compute predictions. current_step: current optimizer step. Used for losses with an annealing component. gt_instances: list of ground truth instances Returns: Dict[str, Tensor]: dict of losses """ global device # Overwrite later use_nll_loss = False ( pred_class_logits, pred_proposal_deltas, pred_class_logits_var, pred_proposal_covs, ) = predictions if len(proposals): box_type = type(proposals[0].proposal_boxes) # cat(..., dim=0) concatenates over all images in the batch proposals_boxes = box_type.cat([p.proposal_boxes for p in proposals]) assert ( not proposals_boxes.tensor.requires_grad ), "Proposals should not require gradients!" # The following fields should exist only when training. if proposals[0].has("gt_boxes"): gt_boxes = box_type.cat([p.gt_boxes for p in proposals]) assert proposals[0].has("gt_classes") gt_classes = cat([p.gt_classes for p in proposals], dim=0) else: proposals_boxes = Boxes( torch.zeros(0, 4, device=pred_proposal_deltas.device) ) no_instances = len(proposals) == 0 # no instances found # Compute Classification Loss if no_instances: # TODO 0.0 * pred.sum() is enough since PT1.6 loss_cls = 0.0 * F.cross_entropy( pred_class_logits, torch.zeros(0, dtype=torch.long, device=pred_class_logits.device), reduction="sum", ) else: if self.compute_cls_var: # Compute classification variance according to: # "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", NIPS 2017 if self.cls_var_loss == "loss_attenuation": num_samples = self.cls_var_num_samples # Compute standard deviation pred_class_logits_var = torch.sqrt(torch.exp(pred_class_logits_var)) # Produce normal samples using logits as the mean and the standard deviation computed above # Scales with GPU memory. 12 GB ---> 3 Samples per anchor for # COCO dataset. univariate_normal_dists = distributions.normal.Normal( pred_class_logits, scale=pred_class_logits_var ) pred_class_stochastic_logits = univariate_normal_dists.rsample( (num_samples,) ) pred_class_stochastic_logits = pred_class_stochastic_logits.view( ( pred_class_stochastic_logits.shape[1] * num_samples, pred_class_stochastic_logits.shape[2], -1, ) ) pred_class_logits = pred_class_stochastic_logits.squeeze(2) # Produce copies of the target classes to match the number of # stochastic samples. gt_classes_target = torch.unsqueeze(gt_classes, 0) gt_classes_target = torch.repeat_interleave( gt_classes_target, num_samples, dim=0 ).view((gt_classes_target.shape[1] * num_samples, -1)) gt_classes_target = gt_classes_target.squeeze(1) loss_cls = F.cross_entropy( pred_class_logits, gt_classes_target, reduction="mean" ) elif self.cls_var_loss == "evidential": # ToDo: Currently does not provide any reasonable mAP Results # (15% mAP) # Assume dirichlet parameters are output. alphas = get_dir_alphas(pred_class_logits) # Get sum of all alphas dirichlet_s = alphas.sum(1).unsqueeze(1) # Generate one hot vectors for ground truth one_hot_vectors = torch.nn.functional.one_hot( gt_classes, alphas.shape[1] ) # Compute loss. This loss attempts to put all evidence on the # correct location. per_instance_loss = one_hot_vectors * ( torch.digamma(dirichlet_s) - torch.digamma(alphas) ) # Compute KL divergence regularizer loss estimated_dirichlet = torch.distributions.dirichlet.Dirichlet( (alphas - 1.0) * (1.0 - one_hot_vectors) + 1.0 ) uniform_dirichlet = torch.distributions.dirichlet.Dirichlet( torch.ones_like(one_hot_vectors).type(torch.FloatTensor).to(device) ) kl_regularization_loss = torch.distributions.kl.kl_divergence( estimated_dirichlet, uniform_dirichlet ) # Compute final loss annealing_multiplier = torch.min( torch.as_tensor(current_step / self.annealing_step).to(device), torch.as_tensor(1.0).to(device), ) per_proposal_loss = ( per_instance_loss.sum(1) + annealing_multiplier * kl_regularization_loss ) # Compute evidence auxiliary loss evidence_maximization_loss = smooth_l1_loss( dirichlet_s, 100.0 * torch.ones_like(dirichlet_s).to(device), beta=self.smooth_l1_beta, reduction="mean", ) evidence_maximization_loss *= annealing_multiplier # Compute final loss foreground_loss = per_proposal_loss[ (gt_classes >= 0) & (gt_classes < pred_class_logits.shape[1] - 1) ] background_loss = per_proposal_loss[ gt_classes == pred_class_logits.shape[1] - 1 ] loss_cls = ( torch.mean(foreground_loss) + torch.mean(background_loss) ) / 2 + 0.01 * evidence_maximization_loss else: loss_cls = F.cross_entropy( pred_class_logits, gt_classes, reduction="mean" ) # Compute regression loss: if no_instances: # TODO 0.0 * pred.sum() is enough since PT1.6 loss_box_reg = 0.0 * smooth_l1_loss( pred_proposal_deltas, torch.zeros_like(pred_proposal_deltas), 0.0, reduction="sum", ) else: gt_proposal_deltas = self.box2box_transform.get_deltas( proposals_boxes.tensor, gt_boxes.tensor ) box_dim = gt_proposal_deltas.size(1) # 4 or 5 cls_agnostic_bbox_reg = pred_proposal_deltas.size(1) == box_dim device = pred_proposal_deltas.device bg_class_ind = pred_class_logits.shape[1] - 1 # Box delta loss is only computed between the prediction for the gt class k # (if 0 <= k < bg_class_ind) and the target; there is no loss defined on predictions # for non-gt classes and background. # Empty fg_inds produces a valid loss of zero as long as the size_average # arg to smooth_l1_loss is False (otherwise it uses torch.mean internally # and would produce a nan loss). fg_inds = torch.nonzero( (gt_classes >= 0) & (gt_classes < bg_class_ind), as_tuple=True )[0] if cls_agnostic_bbox_reg: # pred_proposal_deltas only corresponds to foreground class for # agnostic gt_class_cols = torch.arange(box_dim, device=device) else: fg_gt_classes = gt_classes[fg_inds] # pred_proposal_deltas for class k are located in columns [b * k : b * k + b], # where b is the dimension of box representation (4 or 5) # Note that compared to Detectron1, # we do not perform bounding box regression for background # classes. gt_class_cols = box_dim * fg_gt_classes[:, None] + torch.arange( box_dim, device=device ) gt_covar_class_cols = self.bbox_cov_dims * fg_gt_classes[ :, None ] + torch.arange(self.bbox_cov_dims, device=device) loss_reg_normalizer = gt_classes.numel() pred_proposal_deltas = pred_proposal_deltas[fg_inds[:, None], gt_class_cols] gt_proposals_delta = gt_proposal_deltas[fg_inds] if self.compute_bbox_cov: pred_proposal_covs = pred_proposal_covs[ fg_inds[:, None], gt_covar_class_cols ] pred_proposal_covs = clamp_log_variance(pred_proposal_covs) if self.bbox_cov_loss == "negative_log_likelihood": if self.bbox_cov_type == "diagonal": # Ger foreground proposals. _proposals_boxes = proposals_boxes.tensor[fg_inds] # Compute regression negative log likelihood loss according to: # "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?", NIPS 2017 loss_box_reg = ( 0.5 * torch.exp(-pred_proposal_covs) * smooth_l1_loss( pred_proposal_deltas, gt_proposals_delta, beta=self.smooth_l1_beta, ) ) loss_covariance_regularize = 0.5 * pred_proposal_covs loss_box_reg += loss_covariance_regularize loss_box_reg = torch.sum(loss_box_reg) / loss_reg_normalizer else: # Multivariate Gaussian Negative Log Likelihood loss using pytorch # distributions.multivariate_normal.log_prob() forecaster_cholesky = covariance_output_to_cholesky( pred_proposal_covs ) multivariate_normal_dists = ( distributions.multivariate_normal.MultivariateNormal( pred_proposal_deltas, scale_tril=forecaster_cholesky ) ) loss_box_reg = -multivariate_normal_dists.log_prob( gt_proposals_delta ) loss_box_reg = torch.sum(loss_box_reg) / loss_reg_normalizer elif self.bbox_cov_loss == "second_moment_matching": # Compute regression covariance using second moment # matching. loss_box_reg = smooth_l1_loss( pred_proposal_deltas, gt_proposals_delta, self.smooth_l1_beta ) errors = pred_proposal_deltas - gt_proposals_delta if self.bbox_cov_type == "diagonal": # Handel diagonal case second_moment_matching_term = smooth_l1_loss( torch.exp(pred_proposal_covs), errors ** 2, beta=self.smooth_l1_beta, ) loss_box_reg += second_moment_matching_term loss_box_reg = torch.sum(loss_box_reg) / loss_reg_normalizer else: # Handel full covariance case errors = torch.unsqueeze(errors, 2) gt_error_covar = torch.matmul( errors, torch.transpose(errors, 2, 1) ) # This is the cholesky decomposition of the covariance matrix. # We reconstruct it from 10 estimated parameters as a # lower triangular matrix. forecaster_cholesky = covariance_output_to_cholesky( pred_proposal_covs ) predicted_covar = torch.matmul( forecaster_cholesky, torch.transpose(forecaster_cholesky, 2, 1), ) second_moment_matching_term = smooth_l1_loss( predicted_covar, gt_error_covar, beta=self.smooth_l1_beta, reduction="sum", ) loss_box_reg = ( torch.sum(loss_box_reg) + second_moment_matching_term ) / loss_reg_normalizer elif self.bbox_cov_loss == "energy_loss": forecaster_cholesky = covariance_output_to_cholesky( pred_proposal_covs ) # Define per-anchor Distributions multivariate_normal_dists = ( distributions.multivariate_normal.MultivariateNormal( pred_proposal_deltas, scale_tril=forecaster_cholesky ) ) # Define Monte-Carlo Samples distributions_samples = multivariate_normal_dists.rsample( (self.bbox_cov_num_samples + 1,) ) distributions_samples_1 = distributions_samples[ 0 : self.bbox_cov_num_samples, :, : ] distributions_samples_2 = distributions_samples[ 1 : self.bbox_cov_num_samples + 1, :, : ] # Compute energy score loss_covariance_regularize = ( -smooth_l1_loss( distributions_samples_1, distributions_samples_2, beta=self.smooth_l1_beta, reduction="sum", ) / self.bbox_cov_num_samples ) # Second term gt_proposals_delta_samples = torch.repeat_interleave( gt_proposals_delta.unsqueeze(0), self.bbox_cov_num_samples, dim=0, ) loss_first_moment_match = ( 2.0 * smooth_l1_loss( distributions_samples_1, gt_proposals_delta_samples, beta=self.smooth_l1_beta, reduction="sum", ) / self.bbox_cov_num_samples ) # First term # Final Loss loss_box_reg = ( loss_first_moment_match + loss_covariance_regularize ) / loss_reg_normalizer elif self.bbox_cov_loss == "pmb_negative_log_likelihood": losses = self.nll_od_loss_with_nms( predictions, proposals, gt_instances ) loss_box_reg = losses["loss_box_reg"] use_nll_loss = True else: raise ValueError( "Invalid regression loss name {}.".format(self.bbox_cov_loss) ) # Perform loss annealing. Not really essential in Generalized-RCNN case, but good practice for more # elaborate regression variance losses. standard_regression_loss = smooth_l1_loss( pred_proposal_deltas, gt_proposals_delta, self.smooth_l1_beta, reduction="sum", ) standard_regression_loss = ( standard_regression_loss / loss_reg_normalizer ) probabilistic_loss_weight = get_probabilistic_loss_weight( current_step, self.annealing_step ) loss_box_reg = ( (1.0 - probabilistic_loss_weight) * standard_regression_loss + probabilistic_loss_weight * loss_box_reg ) if use_nll_loss: loss_cls = (1.0 - probabilistic_loss_weight) * loss_cls else: loss_box_reg = smooth_l1_loss( pred_proposal_deltas, gt_proposals_delta, self.smooth_l1_beta, reduction="sum", ) loss_box_reg = loss_box_reg / loss_reg_normalizer if use_nll_loss: losses["loss_cls"] = loss_cls losses["loss_box_reg"] = loss_box_reg else: losses = {"loss_cls": loss_cls, "loss_box_reg": loss_box_reg} return losses def nll_od_loss_with_nms( self, predictions: Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor], proposals: List[Instances], gt_instances, ): if "log_prob" in self.matching_distance and self.matching_distance != "log_prob": covar_scaling = float(self.matching_distance.split("_")[-1]) matching_distance = "log_prob" else: covar_scaling = 1 matching_distance = self.matching_distance self.ppp_intensity_function.update_distribution() _, pred_deltas, _, pred_covs = predictions boxes = self.predict_boxes(predictions, proposals) scores = self.predict_probs(predictions, proposals) scores = [score.clamp(1e-6, 1 - 1e-6) for score in scores] _, num_classes = scores[0].shape num_classes -= 1 # do not count background class image_shapes = [x.image_size for x in proposals] num_prop_per_image = [len(p) for p in proposals] # Apply NMS without score threshold instances, kept_idx = fast_rcnn_inference( boxes, scores, image_shapes, 0.0, self.test_nms_thresh, self.test_topk_per_image, ) kept_idx = [k.unique() for k in kept_idx] pred_covs = pred_covs.split(num_prop_per_image) pred_deltas = pred_deltas.split(num_prop_per_image) kept_proposals = [ prop.proposal_boxes.tensor[idx] for prop, idx in zip(proposals, kept_idx) ] pred_covs = [pred_cov[kept] for pred_cov, kept in zip(pred_covs, kept_idx)] nll_pred_cov = [ covariance_output_to_cholesky(clamp_log_variance(reshape_box_preds(cov, num_classes))) for cov in pred_covs ] nll_scores = [score[kept] for score, kept in zip(scores, kept_idx)] nll_pred_deltas = [ reshape_box_preds(delta[kept], num_classes) for delta, kept in zip(pred_deltas, kept_idx) ] trans_func = lambda x,y: self.box2box_transform.apply_deltas(x,y) box_means = [] box_chols = [] bs = len(nll_pred_deltas) for i in range(bs): box_mean, box_chol = unscented_transform(nll_pred_deltas[i], nll_pred_cov[i], kept_proposals[i], trans_func) box_means.append(box_mean) box_chols.append(box_chol) nll_gt_classes = [instances.gt_classes for instances in gt_instances] gt_boxes = [instances.gt_boxes.tensor for instances in gt_instances] if self.bbox_cov_dist_type == "gaussian": regression_dist = ( lambda x, y: distributions.multivariate_normal.MultivariateNormal( loc=x, scale_tril=y ) ) elif self.bbox_cov_dist_type == "laplacian": regression_dist = lambda x, y: distributions.laplace.Laplace( loc=x, scale=y.diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) else: raise Exception( f"Bounding box uncertainty distribution {self.bbox_cov_dist_type} is not available." ) if self.use_prediction_mixture: ppps = [] src_boxes_tot = [] src_box_chol_tot = [] src_boxes_deltas_tot = [] src_boxes_deltas_chol_tot = [] src_scores_tot = [] gt_box_deltas = [] for i in range(bs): image_shape = image_shapes[i] h,w = image_shape scaling = torch.tensor([1/w,1/h],device=box_means[i].device).repeat(2) pred_box_means = box_means[i]*scaling pred_box_chols = torch.diag_embed(scaling)@box_chols[i] pred_box_deltas = nll_pred_deltas[i] pred_box_delta_chols = nll_pred_cov[i] pred_cls_probs = nll_scores[i] #max_conf = pred_cls_probs[..., :num_classes].max(dim=1)[0] max_conf = 1 - pred_cls_probs[..., -1] ppp_preds_idx = ( max_conf <= self.ppp_intensity_function.ppp_confidence_thres ) props = kept_proposals[i][ppp_preds_idx.logical_not()] # Get delta between each GT and proposal, batch-wise tmp = torch.stack( [ self.box2box_transform.get_deltas( props, gt_boxes[i][j].unsqueeze(0).repeat(len(props), 1), ) for j in range(len(gt_boxes[i])) ] ) gt_box_deltas.append( tmp.permute(1, 0, 2) ) # [gt,pred,boxdim] -> [pred, gt, boxdim] gt_boxes[i] = gt_boxes[i]*scaling mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] selected_chols = pred_box_chols[ppp_preds_idx, 0] mixture_dict["covs"] = selected_chols@(selected_chols.transpose(-1,-2)) mixture_dict["cls_probs"] = pred_cls_probs[ppp_preds_idx, :self.num_classes] mixture_dict["reg_dist_type"] = self.bbox_cov_dist_type if self.bbox_cov_dist_type == "gaussian": mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "scale_tril": selected_chols } elif self.bbox_cov_dist_type == "laplacian": mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": ( selected_chols.diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) } loss_ppp = PoissonPointUnion() loss_ppp.add_ppp(self.ppp_constructor({"predictions": mixture_dict})) loss_ppp.add_ppp(self.ppp_intensity_function) mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] scale_mat = torch.eye(pred_box_chols.shape[-1]).to(pred_box_chols.device)*covar_scaling scaled_chol = scale_mat@pred_box_chols[ppp_preds_idx, 0] mixture_dict["covs"] = (scaled_chol)@(scaled_chol.transpose(-1,-2)) mixture_dict["cls_probs"] = pred_cls_probs[ppp_preds_idx, :self.num_classes] mixture_dict["reg_dist_type"] = self.bbox_cov_dist_type if self.bbox_cov_dist_type == "gaussian": mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "scale_tril": scaled_chol } elif self.bbox_cov_dist_type == "laplacian": mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": ( (scaled_chol).diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) } match_ppp = PoissonPointUnion() match_ppp.add_ppp(self.ppp_constructor({"predictions": mixture_dict})) match_ppp.add_ppp(self.ppp_intensity_function) ppps.append({"matching": match_ppp, "loss": loss_ppp}) src_boxes_tot.append(pred_box_means[ppp_preds_idx.logical_not()]) src_box_chol_tot.append(pred_box_chols[ppp_preds_idx.logical_not()]) src_scores_tot.append(pred_cls_probs[ppp_preds_idx.logical_not()]) src_boxes_deltas_tot.append(pred_box_deltas[ppp_preds_idx.logical_not()]) src_boxes_deltas_chol_tot.append(pred_box_delta_chols[ppp_preds_idx.logical_not()]) nll_pred_deltas = src_boxes_deltas_tot nll_pred_delta_chols = src_boxes_deltas_chol_tot nll_pred_boxes = src_boxes_tot nll_pred_cov = src_box_chol_tot nll_scores = src_scores_tot use_target_delta_matching = False elif self.ppp_intensity_function.ppp_intensity_type == "gaussian_mixture": ppps = [] src_boxes_tot = [] src_box_chol_tot = [] src_boxes_deltas_tot = [] src_boxes_deltas_chol_tot = [] src_scores_tot = [] gt_box_deltas = [] for i in range(bs): image_shape = image_shapes[i] h,w = image_shape scaling = torch.tensor([1/w,1/h],device=box_means[i].device).repeat(2) pred_box_means = box_means[i]*scaling pred_box_chols = torch.diag_embed(scaling)@box_chols[i] pred_box_deltas = nll_pred_deltas[i] pred_box_delta_chols = nll_pred_cov[i] pred_cls_probs = nll_scores[i] props = kept_proposals[i] # Get delta between each GT and proposal, batch-wise tmp = torch.stack( [ self.box2box_transform.get_deltas( props, gt_boxes[i][j].unsqueeze(0).repeat(len(props), 1), ) for j in range(len(gt_boxes[i])) ] ) gt_box_deltas.append( tmp.permute(1, 0, 2) ) # [gt,pred,boxdim] -> [pred, gt, boxdim] gt_boxes[i] = gt_boxes[i]*scaling src_boxes_tot.append(pred_box_means) src_box_chol_tot.append(pred_box_chols) src_scores_tot.append(pred_cls_probs) src_boxes_deltas_tot.append(pred_box_deltas) src_boxes_deltas_chol_tot.append(pred_box_delta_chols) nll_pred_deltas = src_boxes_deltas_tot nll_pred_delta_chols = src_boxes_deltas_chol_tot nll_pred_boxes = src_boxes_tot nll_pred_cov = src_box_chol_tot nll_scores = src_scores_tot use_target_delta_matching = False ppps = [{"loss": self.ppp_intensity_function, "matching": self.ppp_intensity_function}]*bs else: gt_box_deltas = [] for i in range(len(gt_boxes)): # Get delta between each GT and proposal, batch-wise tmp = torch.stack( [ self.box2box_transform.get_deltas( kept_proposals[i], gt_boxes[i][j].unsqueeze(0).repeat(len(kept_proposals[i]), 1), ) for j in range(len(gt_boxes[i])) ] ) gt_box_deltas.append( tmp.permute(1, 0, 2) ) # [gt,pred,boxdim] -> [pred, gt, boxdim] use_target_delta_matching = True ppps = [{"loss": self.ppp_intensity_function, "matching": self.ppp_intensity_function}]*bs nll_pred_delta_chols = nll_pred_cov nll_pred_deltas = nll_pred_deltas nll_pred_boxes = nll_pred_deltas nll_pred_cov = nll_pred_cov nll, associations, decompositions = negative_log_likelihood( nll_scores, nll_pred_boxes, nll_pred_cov, gt_boxes, nll_gt_classes, image_shapes, regression_dist, ppps, self.nll_max_num_solutions, scores_have_bg_cls=True, target_deltas=gt_box_deltas, matching_distance=matching_distance, use_target_delta_matching=use_target_delta_matching, pred_deltas=nll_pred_deltas, pred_delta_chols=nll_pred_delta_chols, ) # Save some stats storage = get_event_storage() num_classes = self.num_classes mean_variance = np.mean( [ cov.diagonal(dim1=-2,dim2=-1) .pow(2) .mean() .item() for cov in nll_pred_cov if cov.shape[0] > 0 ] ) storage.put_scalar("nll/mean_covariance", mean_variance) ppp_intens = np.sum([ppp["loss"].integrate( torch.as_tensor(image_shapes).to(device), num_classes ) .mean() .item() for ppp in ppps ]) storage.put_scalar("nll/ppp_intensity", ppp_intens) reg_loss = np.mean( [ np.clip( decomp["matched_bernoulli_reg"][0] / (decomp["num_matched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) cls_loss_match = np.mean( [ np.clip( decomp["matched_bernoulli_cls"][0] / (decomp["num_matched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) cls_loss_no_match = np.mean( [ np.clip( decomp["unmatched_bernoulli"][0] / (decomp["num_unmatched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) # Collect all losses losses = dict() losses["loss_box_reg"] = nll # Add losses for logging, these do not propagate gradients losses["loss_regression"] = torch.tensor(reg_loss).to(nll.device) losses["loss_cls_matched"] = torch.tensor(cls_loss_match).to(nll.device) losses["loss_cls_unmatched"] = torch.tensor(cls_loss_no_match).to(nll.device) return losses def inference(self, predictions, proposals): """ Returns: list[Instances]: same as `fast_rcnn_inference`. list[Tensor]: same as `fast_rcnn_inference`. """ boxes = self.predict_boxes(predictions, proposals) scores = self.predict_probs(predictions, proposals) image_shapes = [x.image_size for x in proposals] return fast_rcnn_inference( boxes, scores, image_shapes, self.test_score_thresh, self.test_nms_thresh, self.test_topk_per_image, ) def predict_boxes_for_gt_classes(self, predictions, proposals): """ Returns: list[Tensor]: A list of Tensors of predicted boxes for GT classes in case of class-specific box head. Element i of the list has shape (Ri, B), where Ri is the number of predicted objects for image i and B is the box dimension (4 or 5) """ if not len(proposals): return [] scores, proposal_deltas = predictions proposal_boxes = [p.proposal_boxes for p in proposals] proposal_boxes = proposal_boxes[0].cat(proposal_boxes).tensor N, B = proposal_boxes.shape predict_boxes = self.box2box_transform.apply_deltas( proposal_deltas, proposal_boxes ) # Nx(KxB) K = predict_boxes.shape[1] // B if K > 1: gt_classes = torch.cat([p.gt_classes for p in proposals], dim=0) # Some proposals are ignored or have a background class. Their gt_classes # cannot be used as index. gt_classes = gt_classes.clamp_(0, K - 1) predict_boxes = predict_boxes.view(N, K, B)[ torch.arange(N, dtype=torch.long, device=predict_boxes.device), gt_classes, ] num_prop_per_image = [len(p) for p in proposals] return predict_boxes.split(num_prop_per_image) def predict_boxes(self, predictions, proposals): """ Args: predictions: return values of :meth:`forward()`. proposals (list[Instances]): proposals that match the features that were used to compute predictions. The ``proposal_boxes`` field is expected. Returns: list[Tensor]: A list of Tensors of predicted class-specific or class-agnostic boxes for each image. Element i has shape (Ri, K * B) or (Ri, B), where Ri is the number of predicted objects for image i and B is the box dimension (4 or 5) """ if not len(proposals): return [] _, proposal_deltas, _, _ = predictions num_prop_per_image = [len(p) for p in proposals] proposal_boxes = [p.proposal_boxes for p in proposals] proposal_boxes = proposal_boxes[0].cat(proposal_boxes).tensor predict_boxes = self.box2box_transform.apply_deltas( proposal_deltas, proposal_boxes ) # Nx(KxB) return predict_boxes.split(num_prop_per_image) def predict_probs(self, predictions, proposals): """ Args: predictions: return values of :meth:`forward()`. proposals (list[Instances]): proposals that match the features that were used to compute predictions. Returns: list[Tensor]: A list of Tensors of predicted class probabilities for each image. Element i has shape (Ri, K + 1), where Ri is the number of predicted objects for image i. """ scores, _, _, _ = predictions num_inst_per_image = [len(p) for p in proposals] if self.cls_var_loss == "evidential": alphas = get_dir_alphas(scores) dirichlet_s = alphas.sum(1).unsqueeze(1) # Compute probabilities probs = alphas / dirichlet_s else: probs = F.softmax(scores, dim=-1) return probs.split(num_inst_per_image, dim=0) # Todo: new detectron interface required copying code. Check for better # way to inherit from FastRCNNConvFCHead. @ROI_BOX_HEAD_REGISTRY.register() class DropoutFastRCNNConvFCHead(nn.Module): """ A head with several 3x3 conv layers (each followed by norm & relu) and then several fc layers (each followed by relu) and dropout. """ @configurable def __init__( self, input_shape: ShapeSpec, *, conv_dims: List[int], fc_dims: List[int], conv_norm="", dropout_rate, ): """ NOTE: this interface is experimental. Args: input_shape (ShapeSpec): shape of the input feature. conv_dims (list[int]): the output dimensions of the conv layers fc_dims (list[int]): the output dimensions of the fc layers conv_norm (str or callable): normalization for the conv layers. See :func:`detectron2.layers.get_norm` for supported types. dropout_rate (float): p for dropout layer """ super().__init__() assert len(conv_dims) + len(fc_dims) > 0 self.dropout_rate = dropout_rate self.use_dropout = self.dropout_rate != 0.0 self._output_size = ( input_shape.channels, input_shape.height, input_shape.width, ) self.conv_norm_relus = [] for k, conv_dim in enumerate(conv_dims): conv = Conv2d( self._output_size[0], conv_dim, kernel_size=3, padding=1, bias=not conv_norm, norm=get_norm(conv_norm, conv_dim), activation=F.relu, ) self.add_module("conv{}".format(k + 1), conv) self.conv_norm_relus.append(conv) self._output_size = (conv_dim, self._output_size[1], self._output_size[2]) self.fcs = [] self.fcs_dropout = [] for k, fc_dim in enumerate(fc_dims): fc = Linear(np.prod(self._output_size), fc_dim) fc_dropout = nn.Dropout(p=self.dropout_rate) self.add_module("fc{}".format(k + 1), fc) self.add_module("fc_dropout{}".format(k + 1), fc_dropout) self.fcs.append(fc) self.fcs_dropout.append(fc_dropout) self._output_size = fc_dim for layer in self.conv_norm_relus: weight_init.c2_msra_fill(layer) for layer in self.fcs: weight_init.c2_xavier_fill(layer) @classmethod def from_config(cls, cfg, input_shape): num_conv = cfg.MODEL.ROI_BOX_HEAD.NUM_CONV conv_dim = cfg.MODEL.ROI_BOX_HEAD.CONV_DIM num_fc = cfg.MODEL.ROI_BOX_HEAD.NUM_FC fc_dim = cfg.MODEL.ROI_BOX_HEAD.FC_DIM return { "input_shape": input_shape, "conv_dims": [conv_dim] * num_conv, "fc_dims": [fc_dim] * num_fc, "conv_norm": cfg.MODEL.ROI_BOX_HEAD.NORM, "dropout_rate": cfg.MODEL.PROBABILISTIC_MODELING.DROPOUT_RATE, } def forward(self, x): for layer in self.conv_norm_relus: x = layer(x) if len(self.fcs): if x.dim() > 2: x = torch.flatten(x, start_dim=1) for layer, dropout in zip(self.fcs, self.fcs_dropout): x = F.relu(dropout(layer(x))) return x @property def output_shape(self): """ Returns: ShapeSpec: the output feature shape """ o = self._output_size if isinstance(o, int): return ShapeSpec(channels=o) else: return ShapeSpec(channels=o[0], height=o[1], width=o[2])
66,644
40.523364
159
py
pmb-nll
pmb-nll-main/src/probabilistic_modeling/probabilistic_detr.py
import numpy as np import torch import torch.nn.functional as F # Detectron imports from detectron2.modeling import META_ARCH_REGISTRY, detector_postprocess from detectron2.utils.events import get_event_storage # Detr imports from models.detr import DETR, MLP, SetCriterion from torch import distributions, nn from torch._C import device from util import box_ops from util.misc import NestedTensor, accuracy, nested_tensor_from_tensor_list from probabilistic_modeling.losses import negative_log_likelihood # Project imports from probabilistic_modeling.modeling_utils import ( PoissonPointProcessIntensityFunction, clamp_log_variance, covariance_output_to_cholesky, get_probabilistic_loss_weight, PoissonPointUnion) @META_ARCH_REGISTRY.register() class ProbabilisticDetr(META_ARCH_REGISTRY.get("Detr")): """ Implement Probabilistic Detr """ def __init__(self, cfg): super().__init__(cfg) # Parse configs self.cls_var_loss = cfg.MODEL.PROBABILISTIC_MODELING.CLS_VAR_LOSS.NAME self.compute_cls_var = self.cls_var_loss != "none" self.cls_var_num_samples = ( cfg.MODEL.PROBABILISTIC_MODELING.CLS_VAR_LOSS.NUM_SAMPLES ) self.bbox_cov_loss = cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.NAME self.compute_bbox_cov = self.bbox_cov_loss != "none" self.bbox_cov_num_samples = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.NUM_SAMPLES ) self.bbox_cov_dist_type = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE ) self.bbox_cov_type = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.COVARIANCE_TYPE ) if self.bbox_cov_type == "diagonal": # Diagonal covariance matrix has N elements self.bbox_cov_dims = 4 else: # Number of elements required to describe an NxN covariance matrix is # computed as: (N * (N + 1)) / 2 self.bbox_cov_dims = 10 self.dropout_rate = cfg.MODEL.PROBABILISTIC_MODELING.DROPOUT_RATE self.use_dropout = self.dropout_rate != 0.0 self.current_step = 0 self.annealing_step = ( cfg.SOLVER.STEPS[0] if cfg.MODEL.PROBABILISTIC_MODELING.ANNEALING_STEP <= 0 else cfg.MODEL.PROBABILISTIC_MODELING.ANNEALING_STEP ) if self.bbox_cov_loss == "pmb_negative_log_likelihood": ppp_intensity_function = lambda x: PoissonPointProcessIntensityFunction( cfg, device=self.device, **x ) self.nll_max_num_solutions = ( cfg.MODEL.PROBABILISTIC_MODELING.NLL_MAX_NUM_SOLUTIONS ) else: ppp_intensity_function = None self.nll_max_num_solutions = 0 # Create probabilistic output layers self.detr = CustomDetr( self.detr.backbone, self.detr.transformer, num_classes=self.num_classes, num_queries=self.detr.num_queries, aux_loss=self.detr.aux_loss, compute_cls_var=self.compute_cls_var, compute_bbox_cov=self.compute_bbox_cov, bbox_cov_dims=self.bbox_cov_dims, ) self.detr.to(self.device) losses = ["cardinality"] if self.compute_cls_var: losses.append("labels_" + self.cls_var_loss) elif not self.bbox_cov_loss == "pmb_negative_log_likelihood": losses.append("labels") if self.compute_bbox_cov: losses.append("boxes_" + self.bbox_cov_loss) else: losses.append("boxes") # Replace setcriterion with our own implementation self.criterion = ProbabilisticSetCriterion( self.num_classes, matcher=self.criterion.matcher, weight_dict=self.criterion.weight_dict, eos_coef=self.criterion.eos_coef, losses=losses, nll_max_num_solutions=self.nll_max_num_solutions, ppp=ppp_intensity_function, bbox_cov_dist_type=self.bbox_cov_dist_type, matching_distance=cfg.MODEL.PROBABILISTIC_MODELING.MATCHING_DISTANCE, use_prediction_mixture=cfg.MODEL.PROBABILISTIC_MODELING.PPP.USE_PREDICTION_MIXTURE, ) self.criterion.set_bbox_cov_num_samples(self.bbox_cov_num_samples) self.criterion.set_cls_var_num_samples(self.cls_var_num_samples) self.criterion.to(self.device) self.input_format = "RGB" def get_ppp_intensity_function(self): return self.criterion.ppp_intensity_function def forward(self, batched_inputs, return_raw_results=False, is_mc_dropout=False): """ Args: batched_inputs: a list, batched outputs of :class:`DatasetMapper` . Each item in the list contains the inputs for one image. For now, each item in the list is a dict that contains: * image: Tensor, image in (C, H, W) format. * instances: Instances Other information that's included in the original dicts, such as: * "height", "width" (int): the output resolution of the model, used in inference. See :meth:`postprocess` for details. return_raw_results (bool): if True return unprocessed results for probabilistic inference. is_mc_dropout (bool): if True, return unprocessed results even if self.is_training flag is on. Returns: dict[str: Tensor]: mapping from a named loss to a tensor storing the loss. Used during training only. """ try: self.current_step += get_event_storage().iter except: self.current_step += 1 images = self.preprocess_image(batched_inputs) output = self.detr(images) if self.training and not is_mc_dropout: gt_instances = [x["instances"].to(self.device) for x in batched_inputs] targets = self.prepare_targets(gt_instances) loss_dict = self.criterion(output, targets) weight_dict = self.criterion.weight_dict prob_weight = get_probabilistic_loss_weight( self.current_step, self.annealing_step ) for k in loss_dict.keys(): if k in weight_dict: loss_dict[k] *= weight_dict[k] if not "loss" in k: # some "losses" are here for logging purposes only probabilistic_loss_weight = 1 elif "nll" in k: probabilistic_loss_weight = prob_weight else: probabilistic_loss_weight = 1 - prob_weight # uncomment for weighted prob loss # loss_dict[k] *= probabilistic_loss_weight return loss_dict elif return_raw_results: if ( self.compute_bbox_cov and self.bbox_cov_loss == "pmb_negative_log_likelihood" ): output["ppp"] = self.criterion.ppp_intensity_function.get_weights() return output else: box_cls = output["pred_logits"] box_pred = output["pred_boxes"] mask_pred = output["pred_masks"] if self.mask_on else None results = self.inference(box_cls, box_pred, mask_pred, images.image_sizes) processed_results = [] for results_per_image, input_per_image, image_size in zip( results, batched_inputs, images.image_sizes ): height = input_per_image.get("height", image_size[0]) width = input_per_image.get("width", image_size[1]) r = detector_postprocess(results_per_image, height, width) processed_results.append({"instances": r}) return processed_results class CustomDetr(DETR): """This is the DETR module that performs PROBABILISTIC object detection""" def __init__( self, backbone, transformer, num_classes, num_queries, aux_loss=False, compute_cls_var=False, compute_bbox_cov=False, bbox_cov_dims=4, ): super().__init__(backbone, transformer, num_classes, num_queries, aux_loss) hidden_dim = self.transformer.d_model self.compute_cls_var = compute_cls_var if self.compute_cls_var: self.class_var_embed = nn.Linear(hidden_dim, num_classes + 1) nn.init.normal_(self.class_var_embed.weight, std=0.0001) nn.init.constant_(self.class_var_embed.bias, 2 * np.log(0.01)) self.compute_bbox_cov = compute_bbox_cov if self.compute_bbox_cov: self.bbox_covar_embed = MLP(hidden_dim, hidden_dim, bbox_cov_dims, 3) def forward(self, samples: NestedTensor): if isinstance(samples, (list, torch.Tensor)): samples = nested_tensor_from_tensor_list(samples) features, pos = self.backbone(samples) src, mask = features[-1].decompose() assert mask is not None hs = self.transformer( self.input_proj(src), mask, self.query_embed.weight, pos[-1] )[0] outputs_class = self.class_embed(hs) outputs_coord = self.bbox_embed(hs).sigmoid() # Only change to detr code happens here. We need to expose the features from # the transformer to compute variance parameters. out = {"pred_logits": outputs_class[-1], "pred_boxes": outputs_coord[-1]} if self.compute_cls_var: cls_var_out = self.class_var_embed(hs[-1]) out.update({"pred_logits_var": cls_var_out}) if self.compute_bbox_cov: bbox_cov_out = self.bbox_covar_embed(hs) out.update({"pred_boxes_cov": bbox_cov_out[-1]}) else: bbox_cov_out = None if self.aux_loss: out["aux_outputs"] = self._set_aux_loss( outputs_class, outputs_coord, bbox_cov_out ) return out def _set_aux_loss(self, outputs_class, outputs_coord, bbox_cov_out=None): # this is a workaround to make torchscript happy, as torchscript # doesn't support dictionary with non-homogeneous values, such # as a dict having both a Tensor and a list. if bbox_cov_out is None: return [ {"pred_logits": a, "pred_boxes": b} for a, b in zip(outputs_class[:-1], outputs_coord[:-1]) ] else: return [ {"pred_logits": a, "pred_boxes": b, "pred_boxes_cov": c} for a, b, c in zip( outputs_class[:-1], outputs_coord[:-1], bbox_cov_out[:-1] ) ] class ProbabilisticSetCriterion(SetCriterion): """ This is custom set criterion to allow probabilistic estimates """ def __init__( self, num_classes, matcher, weight_dict, eos_coef, losses, nll_max_num_solutions, ppp, bbox_cov_dist_type, matching_distance, use_prediction_mixture, ): super().__init__(num_classes, matcher, weight_dict, eos_coef, losses) self.probabilistic_loss_weight = 0.0 self.bbox_cov_num_samples = 1000 self.cls_var_num_samples = 1000 self.nll_max_num_solutions = nll_max_num_solutions self.ppp_intensity_function = ppp({}) self.ppp_constructor = ppp self.bbox_cov_dist_type = bbox_cov_dist_type self.matching_distance = matching_distance self.use_prediction_mixture = use_prediction_mixture def set_bbox_cov_num_samples(self, bbox_cov_num_samples): self.bbox_cov_num_samples = bbox_cov_num_samples def set_cls_var_num_samples(self, cls_var_num_samples): self.cls_var_num_samples = cls_var_num_samples def loss_labels_att(self, outputs, targets, indices, num_boxes, log=True): """Classification loss (NLL + Loss attenuation) targets dicts must contain the key "labels" containing a tensor of dim [nb_target_boxes] outputs must contain the mean pred_logits and the variance pred_logits_var """ if "pred_logits_var" not in outputs: return self.loss_labels(outputs, targets, indices, num_boxes, log) assert "pred_logits" in outputs src_logits = outputs["pred_logits"] src_logits_var = outputs["pred_logits_var"] src_logits_var = torch.sqrt(torch.exp(src_logits_var)) univariate_normal_dists = distributions.normal.Normal( src_logits, scale=src_logits_var ) pred_class_stochastic_logits = univariate_normal_dists.rsample( (self.cls_var_num_samples,) ) pred_class_stochastic_logits = pred_class_stochastic_logits.view( pred_class_stochastic_logits.shape[1], pred_class_stochastic_logits.shape[2] * pred_class_stochastic_logits.shape[0], -1, ) idx = self._get_src_permutation_idx(indices) target_classes_o = torch.cat( [t["labels"][J] for t, (_, J) in zip(targets, indices)] ) target_classes = torch.full( src_logits.shape[:2], self.num_classes, dtype=torch.int64, device=src_logits.device, ) target_classes[idx] = target_classes_o target_classes = torch.unsqueeze(target_classes, dim=0) target_classes = torch.repeat_interleave( target_classes, self.cls_var_num_samples, dim=0 ) target_classes = target_classes.view( target_classes.shape[1], target_classes.shape[2] * target_classes.shape[0] ) loss_ce = F.cross_entropy( pred_class_stochastic_logits.transpose(1, 2), target_classes, self.empty_weight, ) losses = {"loss_ce": loss_ce} if log: # TODO this should probably be a separate loss, not hacked in this # one here losses["class_error"] = 100 - accuracy(src_logits[idx], target_classes_o)[0] return losses def loss_boxes_var_nll(self, outputs, targets, indices, num_boxes): """Compute the losses related to the bounding boxes, the nll probabilistic regression loss and the GIoU loss targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4] The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size. """ if "pred_boxes_cov" not in outputs: return self.loss_boxes(outputs, targets, indices, num_boxes) assert "pred_boxes" in outputs idx = self._get_src_permutation_idx(indices) src_boxes = outputs["pred_boxes"][idx] src_vars = clamp_log_variance(outputs["pred_boxes_cov"][idx]) target_boxes = torch.cat( [t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0 ) loss_bbox = F.l1_loss(src_boxes, target_boxes, reduction="none") if src_vars.shape[1] == 4: loss_nll = 0.5 * torch.exp(-src_vars) * loss_bbox + 0.5 * src_vars else: forecaster_cholesky = covariance_output_to_cholesky(src_vars) if forecaster_cholesky.shape[0] != 0: multivariate_normal_dists = ( distributions.multivariate_normal.MultivariateNormal( src_boxes, scale_tril=forecaster_cholesky ) ) loss_nll = -multivariate_normal_dists.log_prob(target_boxes) else: loss_nll = loss_bbox loss_nll_final = loss_nll.sum() / num_boxes # Collect all losses losses = dict() losses["loss_bbox"] = loss_nll_final # Add iou loss losses = update_with_iou_loss(losses, src_boxes, target_boxes, num_boxes) return losses def loss_boxes_energy(self, outputs, targets, indices, num_boxes): """Compute the losses related to the bounding boxes, the energy distance loss and the GIoU loss targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4] The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size. """ if "pred_boxes_cov" not in outputs: return self.loss_boxes(outputs, targets, indices, num_boxes) assert "pred_boxes" in outputs idx = self._get_src_permutation_idx(indices) src_boxes = outputs["pred_boxes"][idx] target_boxes = torch.cat( [t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0 ) # Begin probabilistic loss computation src_vars = clamp_log_variance(outputs["pred_boxes_cov"][idx]) forecaster_cholesky = covariance_output_to_cholesky(src_vars) multivariate_normal_dists = ( distributions.multivariate_normal.MultivariateNormal( src_boxes, scale_tril=forecaster_cholesky ) ) # Define Monte-Carlo Samples distributions_samples = multivariate_normal_dists.rsample( (self.bbox_cov_num_samples + 1,) ) distributions_samples_1 = distributions_samples[ 0 : self.bbox_cov_num_samples, :, : ] distributions_samples_2 = distributions_samples[ 1 : self.bbox_cov_num_samples + 1, :, : ] # Compute energy score. Smooth L1 loss is preferred in this case to # maintain the proper scoring properties. loss_covariance_regularize = ( -F.l1_loss( distributions_samples_1, distributions_samples_2, reduction="sum" ) / self.bbox_cov_num_samples ) # Second term gt_proposals_delta_samples = torch.repeat_interleave( target_boxes.unsqueeze(0), self.bbox_cov_num_samples, dim=0 ) loss_first_moment_match = ( 2 * F.l1_loss( distributions_samples_1, gt_proposals_delta_samples, reduction="sum" ) / self.bbox_cov_num_samples ) # First term loss_energy = loss_first_moment_match + loss_covariance_regularize # Normalize and add losses loss_energy_final = loss_energy.sum() / num_boxes # Collect all losses losses = dict() losses["loss_bbox"] = loss_energy_final # Add iou loss losses = update_with_iou_loss(losses, src_boxes, target_boxes, num_boxes) return losses def loss_boxes_smm(self, outputs, targets, indices, num_boxes): """Compute the losses related to the bounding boxes, the L1 regression loss, SMM variance and Covariance loss and the GIoU loss targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4] The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size. """ if "pred_boxes_cov" not in outputs: return self.loss_boxes(outputs, targets, indices, num_boxes) assert "pred_boxes" in outputs idx = self._get_src_permutation_idx(indices) src_boxes = outputs["pred_boxes"][idx] target_boxes = torch.cat( [t["boxes"][i] for t, (_, i) in zip(targets, indices)], dim=0 ) loss_bbox = F.l1_loss(src_boxes, target_boxes, reduction="none") # Begin probabilistic loss computation src_vars = clamp_log_variance(outputs["pred_boxes_cov"][idx]) errors = src_boxes - target_boxes if src_vars.shape[1] == 4: second_moment_matching_term = F.l1_loss( torch.exp(src_vars), errors ** 2, reduction="none" ) else: errors = torch.unsqueeze(errors, 2) gt_error_covar = torch.matmul(errors, torch.transpose(errors, 2, 1)) # This is the cholesky decomposition of the covariance matrix. # We reconstruct it from 10 estimated parameters as a # lower triangular matrix. forecaster_cholesky = covariance_output_to_cholesky(src_vars) predicted_covar = torch.matmul( forecaster_cholesky, torch.transpose(forecaster_cholesky, 2, 1) ) second_moment_matching_term = F.l1_loss( predicted_covar, gt_error_covar, reduction="none" ) loss_smm = second_moment_matching_term.sum() / num_boxes # Normalize and add losses loss_bbox_final = loss_bbox.sum() / num_boxes loss_smm_final = loss_smm + loss_bbox_final # Collect all losses losses = dict() losses["loss_bbox"] = loss_smm_final # Add iou loss losses = update_with_iou_loss(losses, src_boxes, target_boxes, num_boxes) return losses def loss_pmb_nll(self, outputs, targets, indices, num_boxes): if "pred_boxes_cov" not in outputs: return self.loss_boxes(outputs, targets, indices, num_boxes) assert "pred_logits" in outputs src_logits = outputs["pred_logits"] src_scores = src_logits.softmax(-1).clamp(1e-6, 1 - 1e-6) num_classes = src_scores.shape[-1] - 1 assert "pred_boxes" in outputs src_boxes = outputs["pred_boxes"] src_boxes = src_boxes.unsqueeze(2).repeat(1, 1, num_classes, 1) assert "pred_boxes_cov" in outputs src_box_cov = outputs["pred_boxes_cov"] src_box_chol = covariance_output_to_cholesky(src_box_cov) src_box_chol = src_box_chol.unsqueeze(2).repeat(1, 1, num_classes, 1, 1) tgt_classes = [t["labels"] for t in targets] tgt_boxes = [t["boxes"] for t in targets] self.ppp_intensity_function.update_distribution() if self.bbox_cov_dist_type == "gaussian": regression_dist = ( lambda x, y: distributions.multivariate_normal.MultivariateNormal( loc=x, scale_tril=y ) ) elif self.bbox_cov_dist_type == "laplacian": regression_dist = lambda x, y: distributions.laplace.Laplace( loc=x, scale=(y.diagonal(dim1=-2, dim2=-1) / np.sqrt(2)) ) else: raise Exception( f"Bounding box uncertainty distribution {self.bbox_cov_dist_type} is not available." ) if "log_prob" in self.matching_distance and self.matching_distance != "log_prob": covar_scaling = float(self.matching_distance.split("_")[-1]) matching_distance = "log_prob" else: covar_scaling = 1 matching_distance = self.matching_distance bs = src_logits.shape[0] image_shapes = torch.as_tensor([[1, 1] for i in range(bs)]).to(src_boxes.device) if self.use_prediction_mixture: ppps = [] src_boxes_tot = [] src_box_chol_tot = [] src_scores_tot = [] for i in range(bs): pred_box_means = src_boxes[i] pred_box_chols = src_box_chol[i] pred_cls_probs = src_scores[i] #max_conf = pred_cls_probs[..., :num_classes].max(dim=1)[0] max_conf = 1 - pred_cls_probs[..., -1] ppp_preds_idx = ( max_conf <= self.ppp_intensity_function.ppp_confidence_thres ) mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] mixture_dict["covs"] = pred_box_chols[ppp_preds_idx, 0]@pred_box_chols[ppp_preds_idx, 0].transpose(-1,-2) mixture_dict["cls_probs"] = pred_cls_probs[ppp_preds_idx, :num_classes] mixture_dict["reg_dist_type"] = self.bbox_cov_dist_type if self.bbox_cov_dist_type == "gaussian": mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "scale_tril": pred_box_chols[ppp_preds_idx, 0] } elif self.bbox_cov_dist_type == "laplacian": mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": ( pred_box_chols[ppp_preds_idx, 0].diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) } loss_ppp = PoissonPointUnion() loss_ppp.add_ppp(self.ppp_constructor({"predictions": mixture_dict})) loss_ppp.add_ppp(self.ppp_intensity_function) mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] scale_mat = torch.eye(pred_box_chols.shape[-1]).to(pred_box_chols.device)*covar_scaling scaled_cov = scale_mat@pred_box_chols[ppp_preds_idx, 0] mixture_dict["covs"] = (scaled_cov)@(scaled_cov).transpose(-1,-2) mixture_dict["cls_probs"] = pred_cls_probs[ppp_preds_idx, :num_classes] mixture_dict["reg_dist_type"] = self.bbox_cov_dist_type if self.bbox_cov_dist_type == "gaussian": mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "scale_tril": scale_mat@pred_box_chols[ppp_preds_idx, 0] } elif self.bbox_cov_dist_type == "laplacian": mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": ( (scale_mat@pred_box_chols[ppp_preds_idx, 0]).diagonal(dim1=-2, dim2=-1) / np.sqrt(2) ) } match_ppp = PoissonPointUnion() match_ppp.add_ppp(self.ppp_constructor({"predictions": mixture_dict})) match_ppp.add_ppp(self.ppp_intensity_function) ppps.append({"matching": match_ppp, "loss": loss_ppp}) src_boxes_tot.append(pred_box_means[ppp_preds_idx.logical_not()]) src_box_chol_tot.append(pred_box_chols[ppp_preds_idx.logical_not()]) src_scores_tot.append(pred_cls_probs[ppp_preds_idx.logical_not()]) src_boxes = src_boxes_tot src_box_chol = src_box_chol_tot src_scores = src_scores_tot elif self.ppp_intensity_function.ppp_intensity_type == "gaussian_mixture": ppps = [{"loss": self.ppp_intensity_function, "matching": self.ppp_intensity_function}]*bs else: ppps = [{"loss": self.ppp_intensity_function, "matching": self.ppp_intensity_function}]*bs nll, associations, decompositions = negative_log_likelihood( src_scores, src_boxes, src_box_chol, tgt_boxes, tgt_classes, image_shapes, regression_dist, ppps, self.nll_max_num_solutions, scores_have_bg_cls=True, matching_distance=matching_distance, covar_scaling=covar_scaling ) # Save some stats storage = get_event_storage() num_classes = self.num_classes mean_variance = np.mean( [ cov.diagonal(dim1=-2,dim2=-1) .pow(2) .mean() .item() for cov in src_box_chol if cov.shape[0] > 0 ] ) storage.put_scalar("nll/mean_covariance", mean_variance) ppp_intens = np.sum([ppp["loss"].integrate( image_shapes, num_classes ) .mean() .item() for ppp in ppps ]) storage.put_scalar("nll/ppp_intensity", ppp_intens) reg_loss = np.mean( [ np.clip( decomp["matched_bernoulli_reg"][0] / (decomp["num_matched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) cls_loss_match = np.mean( [ np.clip( decomp["matched_bernoulli_cls"][0] / (decomp["num_matched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) cls_loss_no_match = np.mean( [ np.clip( decomp["unmatched_bernoulli"][0] / (decomp["num_unmatched_bernoulli"][0] + 1e-6), -1e25, 1e25, ) for decomp in decompositions ] ) # Collect all losses losses = dict() losses["loss_nll"] = nll # Add losses for logging, these do not propagate gradients losses["regression_matched_nll"] = torch.tensor(reg_loss).to(nll.device) losses["cls_matched_nll"] = torch.tensor(cls_loss_match).to(nll.device) losses["cls_unmatched_nll"] = torch.tensor(cls_loss_no_match).to(nll.device) # Extract matched boxes iou_src_boxes = [] iou_target_boxes = [] for i, association in enumerate(associations): association = torch.as_tensor(association).to(src_boxes[i].device).long() permutation_association = association[ 0, association[0, :, 1] >= 0 ] # select all predictions associated with GT permutation_association = permutation_association[ permutation_association[:, 0] < src_boxes[i].shape[0] ] iou_src_boxes.append(src_boxes[i][permutation_association[:, 0], 0]) iou_target_boxes.append(tgt_boxes[i][permutation_association[:, 1]]) # Add iou loss losses = update_with_iou_loss( losses, torch.cat(iou_src_boxes), torch.cat(iou_target_boxes), num_boxes ) return losses def get_loss(self, loss, outputs, targets, indices, num_boxes, **kwargs): loss_map = { "labels": self.loss_labels, "labels_loss_attenuation": self.loss_labels_att, "cardinality": self.loss_cardinality, "boxes": self.loss_boxes, "boxes_negative_log_likelihood": self.loss_boxes_var_nll, "boxes_energy_loss": self.loss_boxes_energy, "boxes_second_moment_matching": self.loss_boxes_smm, "boxes_pmb_negative_log_likelihood": self.loss_pmb_nll, "masks": self.loss_masks, } assert loss in loss_map, f"do you really want to compute {loss} loss?" return loss_map[loss](outputs, targets, indices, num_boxes, **kwargs) def update_with_iou_loss(losses, src_boxes, target_boxes, num_boxes): loss_giou = 1 - torch.diag( box_ops.generalized_box_iou( box_ops.box_cxcywh_to_xyxy(src_boxes), box_ops.box_cxcywh_to_xyxy(target_boxes), ) ) losses["loss_giou"] = loss_giou.sum() / num_boxes return losses
31,909
38.541512
135
py
pmb-nll
pmb-nll-main/src/detr/main.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import argparse import datetime import json import random import time from pathlib import Path import numpy as np import torch from torch.utils.data import DataLoader, DistributedSampler import datasets import util.misc as utils from datasets import build_dataset, get_coco_api_from_dataset from engine import evaluate, train_one_epoch from models import build_model def get_args_parser(): parser = argparse.ArgumentParser('Set transformer detector', add_help=False) parser.add_argument('--lr', default=1e-4, type=float) parser.add_argument('--lr_backbone', default=1e-5, type=float) parser.add_argument('--batch_size', default=2, type=int) parser.add_argument('--weight_decay', default=1e-4, type=float) parser.add_argument('--epochs', default=300, type=int) parser.add_argument('--lr_drop', default=200, type=int) parser.add_argument('--clip_max_norm', default=0.1, type=float, help='gradient clipping max norm') # Model parameters parser.add_argument('--frozen_weights', type=str, default=None, help="Path to the pretrained model. If set, only the mask head will be trained") # * Backbone parser.add_argument('--backbone', default='resnet50', type=str, help="Name of the convolutional backbone to use") parser.add_argument('--dilation', action='store_true', help="If true, we replace stride with dilation in the last convolutional block (DC5)") parser.add_argument('--position_embedding', default='sine', type=str, choices=('sine', 'learned'), help="Type of positional embedding to use on top of the image features") # * Transformer parser.add_argument('--enc_layers', default=6, type=int, help="Number of encoding layers in the transformer") parser.add_argument('--dec_layers', default=6, type=int, help="Number of decoding layers in the transformer") parser.add_argument('--dim_feedforward', default=2048, type=int, help="Intermediate size of the feedforward layers in the transformer blocks") parser.add_argument('--hidden_dim', default=256, type=int, help="Size of the embeddings (dimension of the transformer)") parser.add_argument('--dropout', default=0.1, type=float, help="Dropout applied in the transformer") parser.add_argument('--nheads', default=8, type=int, help="Number of attention heads inside the transformer's attentions") parser.add_argument('--num_queries', default=100, type=int, help="Number of query slots") parser.add_argument('--pre_norm', action='store_true') # * Segmentation parser.add_argument('--masks', action='store_true', help="Train segmentation head if the flag is provided") # Loss parser.add_argument('--no_aux_loss', dest='aux_loss', action='store_false', help="Disables auxiliary decoding losses (loss at each layer)") # * Matcher parser.add_argument('--set_cost_class', default=1, type=float, help="Class coefficient in the matching cost") parser.add_argument('--set_cost_bbox', default=5, type=float, help="L1 box coefficient in the matching cost") parser.add_argument('--set_cost_giou', default=2, type=float, help="giou box coefficient in the matching cost") # * Loss coefficients parser.add_argument('--mask_loss_coef', default=1, type=float) parser.add_argument('--dice_loss_coef', default=1, type=float) parser.add_argument('--bbox_loss_coef', default=5, type=float) parser.add_argument('--giou_loss_coef', default=2, type=float) parser.add_argument('--eos_coef', default=0.1, type=float, help="Relative classification weight of the no-object class") # dataset parameters parser.add_argument('--dataset_file', default='coco') parser.add_argument('--coco_path', type=str) parser.add_argument('--coco_panoptic_path', type=str) parser.add_argument('--remove_difficult', action='store_true') parser.add_argument('--output_dir', default='', help='path where to save, empty for no saving') parser.add_argument('--device', default='cuda', help='device to use for training / testing') parser.add_argument('--seed', default=42, type=int) parser.add_argument('--resume', default='', help='resume from checkpoint') parser.add_argument('--start_epoch', default=0, type=int, metavar='N', help='start epoch') parser.add_argument('--eval', action='store_true') parser.add_argument('--num_workers', default=2, type=int) # distributed training parameters parser.add_argument('--world_size', default=1, type=int, help='number of distributed processes') parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training') return parser def main(args): utils.init_distributed_mode(args) print("git:\n {}\n".format(utils.get_sha())) if args.frozen_weights is not None: assert args.masks, "Frozen training is meant for segmentation only" print(args) device = torch.device(args.device) # fix the seed for reproducibility seed = args.seed + utils.get_rank() torch.manual_seed(seed) np.random.seed(seed) random.seed(seed) model, criterion, postprocessors = build_model(args) model.to(device) model_without_ddp = model if args.distributed: model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu]) model_without_ddp = model.module n_parameters = sum(p.numel() for p in model.parameters() if p.requires_grad) print('number of params:', n_parameters) param_dicts = [ {"params": [p for n, p in model_without_ddp.named_parameters() if "backbone" not in n and p.requires_grad]}, { "params": [p for n, p in model_without_ddp.named_parameters() if "backbone" in n and p.requires_grad], "lr": args.lr_backbone, }, ] optimizer = torch.optim.AdamW(param_dicts, lr=args.lr, weight_decay=args.weight_decay) lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, args.lr_drop) dataset_train = build_dataset(image_set='train', args=args) dataset_val = build_dataset(image_set='val', args=args) if args.distributed: sampler_train = DistributedSampler(dataset_train) sampler_val = DistributedSampler(dataset_val, shuffle=False) else: sampler_train = torch.utils.data.RandomSampler(dataset_train) sampler_val = torch.utils.data.SequentialSampler(dataset_val) batch_sampler_train = torch.utils.data.BatchSampler( sampler_train, args.batch_size, drop_last=True) data_loader_train = DataLoader(dataset_train, batch_sampler=batch_sampler_train, collate_fn=utils.collate_fn, num_workers=args.num_workers) data_loader_val = DataLoader(dataset_val, args.batch_size, sampler=sampler_val, drop_last=False, collate_fn=utils.collate_fn, num_workers=args.num_workers) if args.dataset_file == "coco_panoptic": # We also evaluate AP during panoptic training, on original coco DS coco_val = datasets.coco.build("val", args) base_ds = get_coco_api_from_dataset(coco_val) else: base_ds = get_coco_api_from_dataset(dataset_val) if args.frozen_weights is not None: checkpoint = torch.load(args.frozen_weights, map_location='cpu') model_without_ddp.detr.load_state_dict(checkpoint['model']) output_dir = Path(args.output_dir) if args.resume: if args.resume.startswith('https'): checkpoint = torch.hub.load_state_dict_from_url( args.resume, map_location='cpu', check_hash=True) else: checkpoint = torch.load(args.resume, map_location='cpu') model_without_ddp.load_state_dict(checkpoint['model']) if not args.eval and 'optimizer' in checkpoint and 'lr_scheduler' in checkpoint and 'epoch' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) lr_scheduler.load_state_dict(checkpoint['lr_scheduler']) args.start_epoch = checkpoint['epoch'] + 1 if args.eval: test_stats, coco_evaluator = evaluate(model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir) if args.output_dir: utils.save_on_master(coco_evaluator.coco_eval["bbox"].eval, output_dir / "eval.pth") return print("Start training") start_time = time.time() for epoch in range(args.start_epoch, args.epochs): if args.distributed: sampler_train.set_epoch(epoch) train_stats = train_one_epoch( model, criterion, data_loader_train, optimizer, device, epoch, args.clip_max_norm) lr_scheduler.step() if args.output_dir: checkpoint_paths = [output_dir / 'checkpoint.pth'] # extra checkpoint before LR drop and every 100 epochs if (epoch + 1) % args.lr_drop == 0 or (epoch + 1) % 100 == 0: checkpoint_paths.append(output_dir / f'checkpoint{epoch:04}.pth') for checkpoint_path in checkpoint_paths: utils.save_on_master({ 'model': model_without_ddp.state_dict(), 'optimizer': optimizer.state_dict(), 'lr_scheduler': lr_scheduler.state_dict(), 'epoch': epoch, 'args': args, }, checkpoint_path) test_stats, coco_evaluator = evaluate( model, criterion, postprocessors, data_loader_val, base_ds, device, args.output_dir ) log_stats = {**{f'train_{k}': v for k, v in train_stats.items()}, **{f'test_{k}': v for k, v in test_stats.items()}, 'epoch': epoch, 'n_parameters': n_parameters} if args.output_dir and utils.is_main_process(): with (output_dir / "log.txt").open("a") as f: f.write(json.dumps(log_stats) + "\n") # for evaluation logs if coco_evaluator is not None: (output_dir / 'eval').mkdir(exist_ok=True) if "bbox" in coco_evaluator.coco_eval: filenames = ['latest.pth'] if epoch % 50 == 0: filenames.append(f'{epoch:03}.pth') for name in filenames: torch.save(coco_evaluator.coco_eval["bbox"].eval, output_dir / "eval" / name) total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('Training time {}'.format(total_time_str)) if __name__ == '__main__': parser = argparse.ArgumentParser('DETR training and evaluation script', parents=[get_args_parser()]) args = parser.parse_args() if args.output_dir: Path(args.output_dir).mkdir(parents=True, exist_ok=True) main(args)
11,532
45.317269
116
py
pmb-nll
pmb-nll-main/src/detr/engine.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Train and eval functions used in main.py """ import math import os import sys from typing import Iterable import torch import util.misc as utils from datasets.coco_eval import CocoEvaluator from datasets.panoptic_eval import PanopticEvaluator def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module, data_loader: Iterable, optimizer: torch.optim.Optimizer, device: torch.device, epoch: int, max_norm: float = 0): model.train() criterion.train() metric_logger = utils.MetricLogger(delimiter=" ") metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}')) metric_logger.add_meter('class_error', utils.SmoothedValue(window_size=1, fmt='{value:.2f}')) header = 'Epoch: [{}]'.format(epoch) print_freq = 10 for samples, targets in metric_logger.log_every(data_loader, print_freq, header): samples = samples.to(device) targets = [{k: v.to(device) for k, v in t.items()} for t in targets] outputs = model(samples) loss_dict = criterion(outputs, targets) weight_dict = criterion.weight_dict losses = sum(loss_dict[k] * weight_dict[k] for k in loss_dict.keys() if k in weight_dict) # reduce losses over all GPUs for logging purposes loss_dict_reduced = utils.reduce_dict(loss_dict) loss_dict_reduced_unscaled = {f'{k}_unscaled': v for k, v in loss_dict_reduced.items()} loss_dict_reduced_scaled = {k: v * weight_dict[k] for k, v in loss_dict_reduced.items() if k in weight_dict} losses_reduced_scaled = sum(loss_dict_reduced_scaled.values()) loss_value = losses_reduced_scaled.item() if not math.isfinite(loss_value): print("Loss is {}, stopping training".format(loss_value)) print(loss_dict_reduced) sys.exit(1) optimizer.zero_grad() losses.backward() if max_norm > 0: torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm) optimizer.step() metric_logger.update(loss=loss_value, **loss_dict_reduced_scaled, **loss_dict_reduced_unscaled) metric_logger.update(class_error=loss_dict_reduced['class_error']) metric_logger.update(lr=optimizer.param_groups[0]["lr"]) # gather the stats from all processes metric_logger.synchronize_between_processes() print("Averaged stats:", metric_logger) return {k: meter.global_avg for k, meter in metric_logger.meters.items()} @torch.no_grad() def evaluate(model, criterion, postprocessors, data_loader, base_ds, device, output_dir): model.eval() criterion.eval() metric_logger = utils.MetricLogger(delimiter=" ") metric_logger.add_meter('class_error', utils.SmoothedValue(window_size=1, fmt='{value:.2f}')) header = 'Test:' iou_types = tuple(k for k in ('segm', 'bbox') if k in postprocessors.keys()) coco_evaluator = CocoEvaluator(base_ds, iou_types) # coco_evaluator.coco_eval[iou_types[0]].params.iouThrs = [0, 0.1, 0.5, 0.75] panoptic_evaluator = None if 'panoptic' in postprocessors.keys(): panoptic_evaluator = PanopticEvaluator( data_loader.dataset.ann_file, data_loader.dataset.ann_folder, output_dir=os.path.join(output_dir, "panoptic_eval"), ) for samples, targets in metric_logger.log_every(data_loader, 10, header): samples = samples.to(device) targets = [{k: v.to(device) for k, v in t.items()} for t in targets] outputs = model(samples) loss_dict = criterion(outputs, targets) weight_dict = criterion.weight_dict # reduce losses over all GPUs for logging purposes loss_dict_reduced = utils.reduce_dict(loss_dict) loss_dict_reduced_scaled = {k: v * weight_dict[k] for k, v in loss_dict_reduced.items() if k in weight_dict} loss_dict_reduced_unscaled = {f'{k}_unscaled': v for k, v in loss_dict_reduced.items()} metric_logger.update(loss=sum(loss_dict_reduced_scaled.values()), **loss_dict_reduced_scaled, **loss_dict_reduced_unscaled) metric_logger.update(class_error=loss_dict_reduced['class_error']) orig_target_sizes = torch.stack([t["orig_size"] for t in targets], dim=0) results = postprocessors['bbox'](outputs, orig_target_sizes) if 'segm' in postprocessors.keys(): target_sizes = torch.stack([t["size"] for t in targets], dim=0) results = postprocessors['segm'](results, outputs, orig_target_sizes, target_sizes) res = {target['image_id'].item(): output for target, output in zip(targets, results)} if coco_evaluator is not None: coco_evaluator.update(res) if panoptic_evaluator is not None: res_pano = postprocessors["panoptic"](outputs, target_sizes, orig_target_sizes) for i, target in enumerate(targets): image_id = target["image_id"].item() file_name = f"{image_id:012d}.png" res_pano[i]["image_id"] = image_id res_pano[i]["file_name"] = file_name panoptic_evaluator.update(res_pano) # gather the stats from all processes metric_logger.synchronize_between_processes() print("Averaged stats:", metric_logger) if coco_evaluator is not None: coco_evaluator.synchronize_between_processes() if panoptic_evaluator is not None: panoptic_evaluator.synchronize_between_processes() # accumulate predictions from all images if coco_evaluator is not None: coco_evaluator.accumulate() coco_evaluator.summarize() panoptic_res = None if panoptic_evaluator is not None: panoptic_res = panoptic_evaluator.summarize() stats = {k: meter.global_avg for k, meter in metric_logger.meters.items()} if coco_evaluator is not None: if 'bbox' in postprocessors.keys(): stats['coco_eval_bbox'] = coco_evaluator.coco_eval['bbox'].stats.tolist() if 'segm' in postprocessors.keys(): stats['coco_eval_masks'] = coco_evaluator.coco_eval['segm'].stats.tolist() if panoptic_res is not None: stats['PQ_all'] = panoptic_res["All"] stats['PQ_th'] = panoptic_res["Things"] stats['PQ_st'] = panoptic_res["Stuff"] return stats, coco_evaluator
6,626
42.598684
103
py
pmb-nll
pmb-nll-main/src/detr/hubconf.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import torch from models.backbone import Backbone, Joiner from models.detr import DETR, PostProcess from models.position_encoding import PositionEmbeddingSine from models.segmentation import DETRsegm, PostProcessPanoptic from models.transformer import Transformer dependencies = ["torch", "torchvision"] def _make_detr(backbone_name: str, dilation=False, num_classes=91, mask=False): hidden_dim = 256 backbone = Backbone(backbone_name, train_backbone=True, return_interm_layers=mask, dilation=dilation) pos_enc = PositionEmbeddingSine(hidden_dim // 2, normalize=True) backbone_with_pos_enc = Joiner(backbone, pos_enc) backbone_with_pos_enc.num_channels = backbone.num_channels transformer = Transformer(d_model=hidden_dim, return_intermediate_dec=True) detr = DETR(backbone_with_pos_enc, transformer, num_classes=num_classes, num_queries=100) if mask: return DETRsegm(detr) return detr def detr_resnet50(pretrained=False, num_classes=91, return_postprocessor=False): """ DETR R50 with 6 encoder and 6 decoder layers. Achieves 42/62.4 AP/AP50 on COCO val5k. """ model = _make_detr("resnet50", dilation=False, num_classes=num_classes) if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth", map_location="cpu", check_hash=True ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcess() return model def detr_resnet50_dc5(pretrained=False, num_classes=91, return_postprocessor=False): """ DETR-DC5 R50 with 6 encoder and 6 decoder layers. The last block of ResNet-50 has dilation to increase output resolution. Achieves 43.3/63.1 AP/AP50 on COCO val5k. """ model = _make_detr("resnet50", dilation=True, num_classes=num_classes) if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r50-dc5-f0fb7ef5.pth", map_location="cpu", check_hash=True ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcess() return model def detr_resnet101(pretrained=False, num_classes=91, return_postprocessor=False): """ DETR-DC5 R101 with 6 encoder and 6 decoder layers. Achieves 43.5/63.8 AP/AP50 on COCO val5k. """ model = _make_detr("resnet101", dilation=False, num_classes=num_classes) if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r101-2c7b67e5.pth", map_location="cpu", check_hash=True ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcess() return model def detr_resnet101_dc5(pretrained=False, num_classes=91, return_postprocessor=False): """ DETR-DC5 R101 with 6 encoder and 6 decoder layers. The last block of ResNet-101 has dilation to increase output resolution. Achieves 44.9/64.7 AP/AP50 on COCO val5k. """ model = _make_detr("resnet101", dilation=True, num_classes=num_classes) if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r101-dc5-a2e86def.pth", map_location="cpu", check_hash=True ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcess() return model def detr_resnet50_panoptic( pretrained=False, num_classes=250, threshold=0.85, return_postprocessor=False ): """ DETR R50 with 6 encoder and 6 decoder layers. Achieves 43.4 PQ on COCO val5k. threshold is the minimum confidence required for keeping segments in the prediction """ model = _make_detr("resnet50", dilation=False, num_classes=num_classes, mask=True) is_thing_map = {i: i <= 90 for i in range(250)} if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r50-panoptic-00ce5173.pth", map_location="cpu", check_hash=True, ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcessPanoptic(is_thing_map, threshold=threshold) return model def detr_resnet50_dc5_panoptic( pretrained=False, num_classes=250, threshold=0.85, return_postprocessor=False ): """ DETR-DC5 R50 with 6 encoder and 6 decoder layers. The last block of ResNet-50 has dilation to increase output resolution. Achieves 44.6 on COCO val5k. threshold is the minimum confidence required for keeping segments in the prediction """ model = _make_detr("resnet50", dilation=True, num_classes=num_classes, mask=True) is_thing_map = {i: i <= 90 for i in range(250)} if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r50-dc5-panoptic-da08f1b1.pth", map_location="cpu", check_hash=True, ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcessPanoptic(is_thing_map, threshold=threshold) return model def detr_resnet101_panoptic( pretrained=False, num_classes=250, threshold=0.85, return_postprocessor=False ): """ DETR-DC5 R101 with 6 encoder and 6 decoder layers. Achieves 45.1 PQ on COCO val5k. threshold is the minimum confidence required for keeping segments in the prediction """ model = _make_detr("resnet101", dilation=False, num_classes=num_classes, mask=True) is_thing_map = {i: i <= 90 for i in range(250)} if pretrained: checkpoint = torch.hub.load_state_dict_from_url( url="https://dl.fbaipublicfiles.com/detr/detr-r101-panoptic-40021d53.pth", map_location="cpu", check_hash=True, ) model.load_state_dict(checkpoint["model"]) if return_postprocessor: return model, PostProcessPanoptic(is_thing_map, threshold=threshold) return model
6,265
36.076923
117
py
pmb-nll
pmb-nll-main/src/detr/test_all.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import io import unittest import torch from torch import nn, Tensor from typing import List from models.matcher import HungarianMatcher from models.position_encoding import PositionEmbeddingSine, PositionEmbeddingLearned from models.backbone import Backbone, Joiner, BackboneBase from util import box_ops from util.misc import nested_tensor_from_tensor_list from hubconf import detr_resnet50, detr_resnet50_panoptic # onnxruntime requires python 3.5 or above try: import onnxruntime except ImportError: onnxruntime = None class Tester(unittest.TestCase): def test_box_cxcywh_to_xyxy(self): t = torch.rand(10, 4) r = box_ops.box_xyxy_to_cxcywh(box_ops.box_cxcywh_to_xyxy(t)) self.assertLess((t - r).abs().max(), 1e-5) @staticmethod def indices_torch2python(indices): return [(i.tolist(), j.tolist()) for i, j in indices] def test_hungarian(self): n_queries, n_targets, n_classes = 100, 15, 91 logits = torch.rand(1, n_queries, n_classes + 1) boxes = torch.rand(1, n_queries, 4) tgt_labels = torch.randint(high=n_classes, size=(n_targets,)) tgt_boxes = torch.rand(n_targets, 4) matcher = HungarianMatcher() targets = [{'labels': tgt_labels, 'boxes': tgt_boxes}] indices_single = matcher({'pred_logits': logits, 'pred_boxes': boxes}, targets) indices_batched = matcher({'pred_logits': logits.repeat(2, 1, 1), 'pred_boxes': boxes.repeat(2, 1, 1)}, targets * 2) self.assertEqual(len(indices_single[0][0]), n_targets) self.assertEqual(len(indices_single[0][1]), n_targets) self.assertEqual(self.indices_torch2python(indices_single), self.indices_torch2python([indices_batched[0]])) self.assertEqual(self.indices_torch2python(indices_single), self.indices_torch2python([indices_batched[1]])) # test with empty targets tgt_labels_empty = torch.randint(high=n_classes, size=(0,)) tgt_boxes_empty = torch.rand(0, 4) targets_empty = [{'labels': tgt_labels_empty, 'boxes': tgt_boxes_empty}] indices = matcher({'pred_logits': logits.repeat(2, 1, 1), 'pred_boxes': boxes.repeat(2, 1, 1)}, targets + targets_empty) self.assertEqual(len(indices[1][0]), 0) indices = matcher({'pred_logits': logits.repeat(2, 1, 1), 'pred_boxes': boxes.repeat(2, 1, 1)}, targets_empty * 2) self.assertEqual(len(indices[0][0]), 0) def test_position_encoding_script(self): m1, m2 = PositionEmbeddingSine(), PositionEmbeddingLearned() mm1, mm2 = torch.jit.script(m1), torch.jit.script(m2) # noqa def test_backbone_script(self): backbone = Backbone('resnet50', True, False, False) torch.jit.script(backbone) # noqa def test_model_script_detection(self): model = detr_resnet50(pretrained=False).eval() scripted_model = torch.jit.script(model) x = nested_tensor_from_tensor_list([torch.rand(3, 200, 200), torch.rand(3, 200, 250)]) out = model(x) out_script = scripted_model(x) self.assertTrue(out["pred_logits"].equal(out_script["pred_logits"])) self.assertTrue(out["pred_boxes"].equal(out_script["pred_boxes"])) def test_model_script_panoptic(self): model = detr_resnet50_panoptic(pretrained=False).eval() scripted_model = torch.jit.script(model) x = nested_tensor_from_tensor_list([torch.rand(3, 200, 200), torch.rand(3, 200, 250)]) out = model(x) out_script = scripted_model(x) self.assertTrue(out["pred_logits"].equal(out_script["pred_logits"])) self.assertTrue(out["pred_boxes"].equal(out_script["pred_boxes"])) self.assertTrue(out["pred_masks"].equal(out_script["pred_masks"])) def test_model_detection_different_inputs(self): model = detr_resnet50(pretrained=False).eval() # support NestedTensor x = nested_tensor_from_tensor_list([torch.rand(3, 200, 200), torch.rand(3, 200, 250)]) out = model(x) self.assertIn('pred_logits', out) # and 4d Tensor x = torch.rand(1, 3, 200, 200) out = model(x) self.assertIn('pred_logits', out) # and List[Tensor[C, H, W]] x = torch.rand(3, 200, 200) out = model([x]) self.assertIn('pred_logits', out) def test_warpped_model_script_detection(self): class WrappedDETR(nn.Module): def __init__(self, model): super().__init__() self.model = model def forward(self, inputs: List[Tensor]): sample = nested_tensor_from_tensor_list(inputs) return self.model(sample) model = detr_resnet50(pretrained=False) wrapped_model = WrappedDETR(model) wrapped_model.eval() scripted_model = torch.jit.script(wrapped_model) x = [torch.rand(3, 200, 200), torch.rand(3, 200, 250)] out = wrapped_model(x) out_script = scripted_model(x) self.assertTrue(out["pred_logits"].equal(out_script["pred_logits"])) self.assertTrue(out["pred_boxes"].equal(out_script["pred_boxes"])) @unittest.skipIf(onnxruntime is None, 'ONNX Runtime unavailable') class ONNXExporterTester(unittest.TestCase): @classmethod def setUpClass(cls): torch.manual_seed(123) def run_model(self, model, inputs_list, tolerate_small_mismatch=False, do_constant_folding=True, dynamic_axes=None, output_names=None, input_names=None): model.eval() onnx_io = io.BytesIO() # export to onnx with the first input torch.onnx.export(model, inputs_list[0], onnx_io, do_constant_folding=do_constant_folding, opset_version=12, dynamic_axes=dynamic_axes, input_names=input_names, output_names=output_names) # validate the exported model with onnx runtime for test_inputs in inputs_list: with torch.no_grad(): if isinstance(test_inputs, torch.Tensor) or isinstance(test_inputs, list): test_inputs = (nested_tensor_from_tensor_list(test_inputs),) test_ouputs = model(*test_inputs) if isinstance(test_ouputs, torch.Tensor): test_ouputs = (test_ouputs,) self.ort_validate(onnx_io, test_inputs, test_ouputs, tolerate_small_mismatch) def ort_validate(self, onnx_io, inputs, outputs, tolerate_small_mismatch=False): inputs, _ = torch.jit._flatten(inputs) outputs, _ = torch.jit._flatten(outputs) def to_numpy(tensor): if tensor.requires_grad: return tensor.detach().cpu().numpy() else: return tensor.cpu().numpy() inputs = list(map(to_numpy, inputs)) outputs = list(map(to_numpy, outputs)) ort_session = onnxruntime.InferenceSession(onnx_io.getvalue()) # compute onnxruntime output prediction ort_inputs = dict((ort_session.get_inputs()[i].name, inpt) for i, inpt in enumerate(inputs)) ort_outs = ort_session.run(None, ort_inputs) for i in range(0, len(outputs)): try: torch.testing.assert_allclose(outputs[i], ort_outs[i], rtol=1e-03, atol=1e-05) except AssertionError as error: if tolerate_small_mismatch: self.assertIn("(0.00%)", str(error), str(error)) else: raise def test_model_onnx_detection(self): model = detr_resnet50(pretrained=False).eval() dummy_image = torch.ones(1, 3, 800, 800) * 0.3 model(dummy_image) # Test exported model on images of different size, or dummy input self.run_model( model, [(torch.rand(1, 3, 750, 800),)], input_names=["inputs"], output_names=["pred_logits", "pred_boxes"], tolerate_small_mismatch=True, ) @unittest.skip("CI doesn't have enough memory") def test_model_onnx_detection_panoptic(self): model = detr_resnet50_panoptic(pretrained=False).eval() dummy_image = torch.ones(1, 3, 800, 800) * 0.3 model(dummy_image) # Test exported model on images of different size, or dummy input self.run_model( model, [(torch.rand(1, 3, 750, 800),)], input_names=["inputs"], output_names=["pred_logits", "pred_boxes", "pred_masks"], tolerate_small_mismatch=True, ) if __name__ == '__main__': unittest.main()
8,804
40.928571
119
py
pmb-nll
pmb-nll-main/src/detr/models/detr.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ DETR model and criterion classes. """ import torch import torch.nn.functional as F from torch import nn from util import box_ops from util.misc import (NestedTensor, nested_tensor_from_tensor_list, accuracy, get_world_size, interpolate, is_dist_avail_and_initialized) from .backbone import build_backbone from .matcher import build_matcher from .segmentation import (DETRsegm, PostProcessPanoptic, PostProcessSegm, dice_loss, sigmoid_focal_loss) from .transformer import build_transformer class DETR(nn.Module): """ This is the DETR module that performs object detection """ def __init__(self, backbone, transformer, num_classes, num_queries, aux_loss=False): """ Initializes the model. Parameters: backbone: torch module of the backbone to be used. See backbone.py transformer: torch module of the transformer architecture. See transformer.py num_classes: number of object classes num_queries: number of object queries, ie detection slot. This is the maximal number of objects DETR can detect in a single image. For COCO, we recommend 100 queries. aux_loss: True if auxiliary decoding losses (loss at each decoder layer) are to be used. """ super().__init__() self.num_queries = num_queries self.transformer = transformer hidden_dim = transformer.d_model self.class_embed = nn.Linear(hidden_dim, num_classes + 1) self.bbox_embed = MLP(hidden_dim, hidden_dim, 4, 3) self.query_embed = nn.Embedding(num_queries, hidden_dim) self.input_proj = nn.Conv2d(backbone.num_channels, hidden_dim, kernel_size=1) self.backbone = backbone self.aux_loss = aux_loss def forward(self, samples: NestedTensor): """ The forward expects a NestedTensor, which consists of: - samples.tensor: batched images, of shape [batch_size x 3 x H x W] - samples.mask: a binary mask of shape [batch_size x H x W], containing 1 on padded pixels It returns a dict with the following elements: - "pred_logits": the classification logits (including no-object) for all queries. Shape= [batch_size x num_queries x (num_classes + 1)] - "pred_boxes": The normalized boxes coordinates for all queries, represented as (center_x, center_y, height, width). These values are normalized in [0, 1], relative to the size of each individual image (disregarding possible padding). See PostProcess for information on how to retrieve the unnormalized bounding box. - "aux_outputs": Optional, only returned when auxilary losses are activated. It is a list of dictionnaries containing the two above keys for each decoder layer. """ if isinstance(samples, (list, torch.Tensor)): samples = nested_tensor_from_tensor_list(samples) features, pos = self.backbone(samples) src, mask = features[-1].decompose() assert mask is not None hs = self.transformer(self.input_proj(src), mask, self.query_embed.weight, pos[-1])[0] outputs_class = self.class_embed(hs) outputs_coord = self.bbox_embed(hs).sigmoid() out = {'pred_logits': outputs_class[-1], 'pred_boxes': outputs_coord[-1]} if self.aux_loss: out['aux_outputs'] = self._set_aux_loss(outputs_class, outputs_coord) return out @torch.jit.unused def _set_aux_loss(self, outputs_class, outputs_coord): # this is a workaround to make torchscript happy, as torchscript # doesn't support dictionary with non-homogeneous values, such # as a dict having both a Tensor and a list. return [{'pred_logits': a, 'pred_boxes': b} for a, b in zip(outputs_class[:-1], outputs_coord[:-1])] class SetCriterion(nn.Module): """ This class computes the loss for DETR. The process happens in two steps: 1) we compute hungarian assignment between ground truth boxes and the outputs of the model 2) we supervise each pair of matched ground-truth / prediction (supervise class and box) """ def __init__(self, num_classes, matcher, weight_dict, eos_coef, losses): """ Create the criterion. Parameters: num_classes: number of object categories, omitting the special no-object category matcher: module able to compute a matching between targets and proposals weight_dict: dict containing as key the names of the losses and as values their relative weight. eos_coef: relative classification weight applied to the no-object category losses: list of all the losses to be applied. See get_loss for list of available losses. """ super().__init__() self.num_classes = num_classes self.matcher = matcher self.weight_dict = weight_dict self.eos_coef = eos_coef self.losses = losses empty_weight = torch.ones(self.num_classes + 1) empty_weight[-1] = self.eos_coef self.register_buffer('empty_weight', empty_weight) def loss_labels(self, outputs, targets, indices, num_boxes, log=True): """Classification loss (NLL) targets dicts must contain the key "labels" containing a tensor of dim [nb_target_boxes] """ assert 'pred_logits' in outputs src_logits = outputs['pred_logits'] idx = self._get_src_permutation_idx(indices) target_classes_o = torch.cat([t["labels"][J] for t, (_, J) in zip(targets, indices)]) target_classes = torch.full(src_logits.shape[:2], self.num_classes, dtype=torch.int64, device=src_logits.device) target_classes[idx] = target_classes_o loss_ce = F.cross_entropy(src_logits.transpose(1, 2), target_classes, self.empty_weight) losses = {'loss_ce': loss_ce} if log: # TODO this should probably be a separate loss, not hacked in this one here losses['class_error'] = 100 - accuracy(src_logits[idx], target_classes_o)[0] return losses @torch.no_grad() def loss_cardinality(self, outputs, targets, indices, num_boxes): """ Compute the cardinality error, ie the absolute error in the number of predicted non-empty boxes This is not really a loss, it is intended for logging purposes only. It doesn't propagate gradients """ pred_logits = outputs['pred_logits'] device = pred_logits.device tgt_lengths = torch.as_tensor([len(v["labels"]) for v in targets], device=device) # Count the number of predictions that are NOT "no-object" (which is the last class) card_pred = (pred_logits.argmax(-1) != pred_logits.shape[-1] - 1).sum(1) card_err = F.l1_loss(card_pred.float(), tgt_lengths.float()) losses = {'cardinality_error': card_err} return losses def loss_boxes(self, outputs, targets, indices, num_boxes): """Compute the losses related to the bounding boxes, the L1 regression loss and the GIoU loss targets dicts must contain the key "boxes" containing a tensor of dim [nb_target_boxes, 4] The target boxes are expected in format (center_x, center_y, w, h), normalized by the image size. """ assert 'pred_boxes' in outputs idx = self._get_src_permutation_idx(indices) src_boxes = outputs['pred_boxes'][idx] target_boxes = torch.cat([t['boxes'][i] for t, (_, i) in zip(targets, indices)], dim=0) loss_bbox = F.l1_loss(src_boxes, target_boxes, reduction='none') losses = {} losses['loss_bbox'] = loss_bbox.sum() / num_boxes loss_giou = 1 - torch.diag(box_ops.generalized_box_iou( box_ops.box_cxcywh_to_xyxy(src_boxes), box_ops.box_cxcywh_to_xyxy(target_boxes))) losses['loss_giou'] = loss_giou.sum() / num_boxes return losses def loss_masks(self, outputs, targets, indices, num_boxes): """Compute the losses related to the masks: the focal loss and the dice loss. targets dicts must contain the key "masks" containing a tensor of dim [nb_target_boxes, h, w] """ assert "pred_masks" in outputs src_idx = self._get_src_permutation_idx(indices) tgt_idx = self._get_tgt_permutation_idx(indices) src_masks = outputs["pred_masks"] src_masks = src_masks[src_idx] masks = [t["masks"] for t in targets] # TODO use valid to mask invalid areas due to padding in loss target_masks, valid = nested_tensor_from_tensor_list(masks).decompose() target_masks = target_masks.to(src_masks) target_masks = target_masks[tgt_idx] # upsample predictions to the target size src_masks = interpolate(src_masks[:, None], size=target_masks.shape[-2:], mode="bilinear", align_corners=False) src_masks = src_masks[:, 0].flatten(1) target_masks = target_masks.flatten(1) target_masks = target_masks.view(src_masks.shape) losses = { "loss_mask": sigmoid_focal_loss(src_masks, target_masks, num_boxes), "loss_dice": dice_loss(src_masks, target_masks, num_boxes), } return losses def _get_src_permutation_idx(self, indices): # permute predictions following indices batch_idx = torch.cat([torch.full_like(src, i) for i, (src, _) in enumerate(indices)]) src_idx = torch.cat([src for (src, _) in indices]) return batch_idx, src_idx def _get_tgt_permutation_idx(self, indices): # permute targets following indices batch_idx = torch.cat([torch.full_like(tgt, i) for i, (_, tgt) in enumerate(indices)]) tgt_idx = torch.cat([tgt for (_, tgt) in indices]) return batch_idx, tgt_idx def get_loss(self, loss, outputs, targets, indices, num_boxes, **kwargs): loss_map = { 'labels': self.loss_labels, 'cardinality': self.loss_cardinality, 'boxes': self.loss_boxes, 'masks': self.loss_masks } assert loss in loss_map, f'do you really want to compute {loss} loss?' return loss_map[loss](outputs, targets, indices, num_boxes, **kwargs) def forward(self, outputs, targets): """ This performs the loss computation. Parameters: outputs: dict of tensors, see the output specification of the model for the format targets: list of dicts, such that len(targets) == batch_size. The expected keys in each dict depends on the losses applied, see each loss' doc """ outputs_without_aux = {k: v for k, v in outputs.items() if k != 'aux_outputs'} # Retrieve the matching between the outputs of the last layer and the targets indices = self.matcher(outputs_without_aux, targets) # Compute the average number of target boxes accross all nodes, for normalization purposes num_boxes = sum(len(t["labels"]) for t in targets) num_boxes = torch.as_tensor([num_boxes], dtype=torch.float, device=next(iter(outputs.values())).device) if is_dist_avail_and_initialized(): torch.distributed.all_reduce(num_boxes) num_boxes = torch.clamp(num_boxes / get_world_size(), min=1).item() # Compute all the requested losses losses = {} for loss in self.losses: losses.update(self.get_loss(loss, outputs, targets, indices, num_boxes)) # In case of auxiliary losses, we repeat this process with the output of each intermediate layer. if 'aux_outputs' in outputs: for i, aux_outputs in enumerate(outputs['aux_outputs']): indices = self.matcher(aux_outputs, targets) for loss in self.losses: if loss == 'masks': # Intermediate masks losses are too costly to compute, we ignore them. continue kwargs = {} if loss == 'labels': # Logging is enabled only for the last layer kwargs = {'log': False} l_dict = self.get_loss(loss, aux_outputs, targets, indices, num_boxes, **kwargs) l_dict = {k + f'_{i}': v for k, v in l_dict.items()} losses.update(l_dict) return losses class PostProcess(nn.Module): """ This module converts the model's output into the format expected by the coco api""" @torch.no_grad() def forward(self, outputs, target_sizes): """ Perform the computation Parameters: outputs: raw outputs of the model target_sizes: tensor of dimension [batch_size x 2] containing the size of each images of the batch For evaluation, this must be the original image size (before any data augmentation) For visualization, this should be the image size after data augment, but before padding """ out_logits, out_bbox = outputs['pred_logits'], outputs['pred_boxes'] assert len(out_logits) == len(target_sizes) assert target_sizes.shape[1] == 2 prob = F.softmax(out_logits, -1) scores, labels = prob[..., :-1].max(-1) # convert to [x0, y0, x1, y1] format boxes = box_ops.box_cxcywh_to_xyxy(out_bbox) # and from relative [0, 1] to absolute [0, height] coordinates img_h, img_w = target_sizes.unbind(1) scale_fct = torch.stack([img_w, img_h, img_w, img_h], dim=1) boxes = boxes * scale_fct[:, None, :] results = [{'scores': s, 'labels': l, 'boxes': b} for s, l, b in zip(scores, labels, boxes)] return results class MLP(nn.Module): """ Very simple multi-layer perceptron (also called FFN)""" def __init__(self, input_dim, hidden_dim, output_dim, num_layers): super().__init__() self.num_layers = num_layers h = [hidden_dim] * (num_layers - 1) self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])) def forward(self, x): for i, layer in enumerate(self.layers): x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x) return x def build(args): # the `num_classes` naming here is somewhat misleading. # it indeed corresponds to `max_obj_id + 1`, where max_obj_id # is the maximum id for a class in your dataset. For example, # COCO has a max_obj_id of 90, so we pass `num_classes` to be 91. # As another example, for a dataset that has a single class with id 1, # you should pass `num_classes` to be 2 (max_obj_id + 1). # For more details on this, check the following discussion # https://github.com/facebookresearch/detr/issues/108#issuecomment-650269223 num_classes = 20 if args.dataset_file != 'coco' else 91 if args.dataset_file == "coco_panoptic": # for panoptic, we just add a num_classes that is large enough to hold # max_obj_id + 1, but the exact value doesn't really matter num_classes = 250 device = torch.device(args.device) backbone = build_backbone(args) transformer = build_transformer(args) model = DETR( backbone, transformer, num_classes=num_classes, num_queries=args.num_queries, aux_loss=args.aux_loss, ) if args.masks: model = DETRsegm(model, freeze_detr=(args.frozen_weights is not None)) matcher = build_matcher(args) weight_dict = {'loss_ce': 1, 'loss_bbox': args.bbox_loss_coef} weight_dict['loss_giou'] = args.giou_loss_coef if args.masks: weight_dict["loss_mask"] = args.mask_loss_coef weight_dict["loss_dice"] = args.dice_loss_coef # TODO this is a hack if args.aux_loss: aux_weight_dict = {} for i in range(args.dec_layers - 1): aux_weight_dict.update({k + f'_{i}': v for k, v in weight_dict.items()}) weight_dict.update(aux_weight_dict) losses = ['labels', 'boxes', 'cardinality'] if args.masks: losses += ["masks"] criterion = SetCriterion(num_classes, matcher=matcher, weight_dict=weight_dict, eos_coef=args.eos_coef, losses=losses) criterion.to(device) postprocessors = {'bbox': PostProcess()} if args.masks: postprocessors['segm'] = PostProcessSegm() if args.dataset_file == "coco_panoptic": is_thing_map = {i: i <= 90 for i in range(201)} postprocessors["panoptic"] = PostProcessPanoptic(is_thing_map, threshold=0.85) return model, criterion, postprocessors
17,088
46.469444
113
py
pmb-nll
pmb-nll-main/src/detr/models/matcher.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Modules to compute the matching cost and solve the corresponding LSAP. """ import torch from scipy.optimize import linear_sum_assignment from torch import nn from util.box_ops import box_cxcywh_to_xyxy, generalized_box_iou class HungarianMatcher(nn.Module): """This class computes an assignment between the targets and the predictions of the network For efficiency reasons, the targets don't include the no_object. Because of this, in general, there are more predictions than targets. In this case, we do a 1-to-1 matching of the best predictions, while the others are un-matched (and thus treated as non-objects). """ def __init__(self, cost_class: float = 1, cost_bbox: float = 1, cost_giou: float = 1): """Creates the matcher Params: cost_class: This is the relative weight of the classification error in the matching cost cost_bbox: This is the relative weight of the L1 error of the bounding box coordinates in the matching cost cost_giou: This is the relative weight of the giou loss of the bounding box in the matching cost """ super().__init__() self.cost_class = cost_class self.cost_bbox = cost_bbox self.cost_giou = cost_giou assert cost_class != 0 or cost_bbox != 0 or cost_giou != 0, "all costs cant be 0" @torch.no_grad() def forward(self, outputs, targets): """ Performs the matching Params: outputs: This is a dict that contains at least these entries: "pred_logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits "pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates targets: This is a list of targets (len(targets) = batch_size), where each target is a dict containing: "labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of ground-truth objects in the target) containing the class labels "boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates Returns: A list of size batch_size, containing tuples of (index_i, index_j) where: - index_i is the indices of the selected predictions (in order) - index_j is the indices of the corresponding selected targets (in order) For each batch element, it holds: len(index_i) = len(index_j) = min(num_queries, num_target_boxes) """ bs, num_queries = outputs["pred_logits"].shape[:2] # We flatten to compute the cost matrices in a batch out_prob = outputs["pred_logits"].flatten(0, 1).softmax(-1) # [batch_size * num_queries, num_classes] out_bbox = outputs["pred_boxes"].flatten(0, 1) # [batch_size * num_queries, 4] # Also concat the target labels and boxes tgt_ids = torch.cat([v["labels"] for v in targets]) tgt_bbox = torch.cat([v["boxes"] for v in targets]) # Compute the classification cost. Contrary to the loss, we don't use the NLL, # but approximate it in 1 - proba[target class]. # The 1 is a constant that doesn't change the matching, it can be ommitted. cost_class = -out_prob[:, tgt_ids] # Compute the L1 cost between boxes cost_bbox = torch.cdist(out_bbox, tgt_bbox, p=1) # Compute the giou cost betwen boxes cost_giou = -generalized_box_iou(box_cxcywh_to_xyxy(out_bbox), box_cxcywh_to_xyxy(tgt_bbox)) # Final cost matrix C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou C = C.view(bs, num_queries, -1).cpu() sizes = [len(v["boxes"]) for v in targets] indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(sizes, -1))] return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices] def build_matcher(args): return HungarianMatcher(cost_class=args.set_cost_class, cost_bbox=args.set_cost_bbox, cost_giou=args.set_cost_giou)
4,250
47.862069
119
py
pmb-nll
pmb-nll-main/src/detr/models/segmentation.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ This file provides the definition of the convolutional heads used to predict masks, as well as the losses """ import io from collections import defaultdict from typing import List, Optional import torch import torch.nn as nn import torch.nn.functional as F from torch import Tensor from PIL import Image import util.box_ops as box_ops from util.misc import NestedTensor, interpolate, nested_tensor_from_tensor_list try: from panopticapi.utils import id2rgb, rgb2id except ImportError: pass class DETRsegm(nn.Module): def __init__(self, detr, freeze_detr=False): super().__init__() self.detr = detr if freeze_detr: for p in self.parameters(): p.requires_grad_(False) hidden_dim, nheads = detr.transformer.d_model, detr.transformer.nhead self.bbox_attention = MHAttentionMap(hidden_dim, hidden_dim, nheads, dropout=0.0) self.mask_head = MaskHeadSmallConv(hidden_dim + nheads, [1024, 512, 256], hidden_dim) def forward(self, samples: NestedTensor): if isinstance(samples, (list, torch.Tensor)): samples = nested_tensor_from_tensor_list(samples) features, pos = self.detr.backbone(samples) bs = features[-1].tensors.shape[0] src, mask = features[-1].decompose() assert mask is not None src_proj = self.detr.input_proj(src) hs, memory = self.detr.transformer(src_proj, mask, self.detr.query_embed.weight, pos[-1]) outputs_class = self.detr.class_embed(hs) outputs_coord = self.detr.bbox_embed(hs).sigmoid() out = {"pred_logits": outputs_class[-1], "pred_boxes": outputs_coord[-1]} if self.detr.aux_loss: out['aux_outputs'] = self.detr._set_aux_loss(outputs_class, outputs_coord) # FIXME h_boxes takes the last one computed, keep this in mind bbox_mask = self.bbox_attention(hs[-1], memory, mask=mask) seg_masks = self.mask_head(src_proj, bbox_mask, [features[2].tensors, features[1].tensors, features[0].tensors]) outputs_seg_masks = seg_masks.view(bs, self.detr.num_queries, seg_masks.shape[-2], seg_masks.shape[-1]) out["pred_masks"] = outputs_seg_masks return out def _expand(tensor, length: int): return tensor.unsqueeze(1).repeat(1, int(length), 1, 1, 1).flatten(0, 1) class MaskHeadSmallConv(nn.Module): """ Simple convolutional head, using group norm. Upsampling is done using a FPN approach """ def __init__(self, dim, fpn_dims, context_dim): super().__init__() inter_dims = [dim, context_dim // 2, context_dim // 4, context_dim // 8, context_dim // 16, context_dim // 64] self.lay1 = torch.nn.Conv2d(dim, dim, 3, padding=1) self.gn1 = torch.nn.GroupNorm(8, dim) self.lay2 = torch.nn.Conv2d(dim, inter_dims[1], 3, padding=1) self.gn2 = torch.nn.GroupNorm(8, inter_dims[1]) self.lay3 = torch.nn.Conv2d(inter_dims[1], inter_dims[2], 3, padding=1) self.gn3 = torch.nn.GroupNorm(8, inter_dims[2]) self.lay4 = torch.nn.Conv2d(inter_dims[2], inter_dims[3], 3, padding=1) self.gn4 = torch.nn.GroupNorm(8, inter_dims[3]) self.lay5 = torch.nn.Conv2d(inter_dims[3], inter_dims[4], 3, padding=1) self.gn5 = torch.nn.GroupNorm(8, inter_dims[4]) self.out_lay = torch.nn.Conv2d(inter_dims[4], 1, 3, padding=1) self.dim = dim self.adapter1 = torch.nn.Conv2d(fpn_dims[0], inter_dims[1], 1) self.adapter2 = torch.nn.Conv2d(fpn_dims[1], inter_dims[2], 1) self.adapter3 = torch.nn.Conv2d(fpn_dims[2], inter_dims[3], 1) for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_uniform_(m.weight, a=1) nn.init.constant_(m.bias, 0) def forward(self, x: Tensor, bbox_mask: Tensor, fpns: List[Tensor]): x = torch.cat([_expand(x, bbox_mask.shape[1]), bbox_mask.flatten(0, 1)], 1) x = self.lay1(x) x = self.gn1(x) x = F.relu(x) x = self.lay2(x) x = self.gn2(x) x = F.relu(x) cur_fpn = self.adapter1(fpns[0]) if cur_fpn.size(0) != x.size(0): cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0)) x = cur_fpn + F.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest") x = self.lay3(x) x = self.gn3(x) x = F.relu(x) cur_fpn = self.adapter2(fpns[1]) if cur_fpn.size(0) != x.size(0): cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0)) x = cur_fpn + F.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest") x = self.lay4(x) x = self.gn4(x) x = F.relu(x) cur_fpn = self.adapter3(fpns[2]) if cur_fpn.size(0) != x.size(0): cur_fpn = _expand(cur_fpn, x.size(0) // cur_fpn.size(0)) x = cur_fpn + F.interpolate(x, size=cur_fpn.shape[-2:], mode="nearest") x = self.lay5(x) x = self.gn5(x) x = F.relu(x) x = self.out_lay(x) return x class MHAttentionMap(nn.Module): """This is a 2D attention module, which only returns the attention softmax (no multiplication by value)""" def __init__(self, query_dim, hidden_dim, num_heads, dropout=0.0, bias=True): super().__init__() self.num_heads = num_heads self.hidden_dim = hidden_dim self.dropout = nn.Dropout(dropout) self.q_linear = nn.Linear(query_dim, hidden_dim, bias=bias) self.k_linear = nn.Linear(query_dim, hidden_dim, bias=bias) nn.init.zeros_(self.k_linear.bias) nn.init.zeros_(self.q_linear.bias) nn.init.xavier_uniform_(self.k_linear.weight) nn.init.xavier_uniform_(self.q_linear.weight) self.normalize_fact = float(hidden_dim / self.num_heads) ** -0.5 def forward(self, q, k, mask: Optional[Tensor] = None): q = self.q_linear(q) k = F.conv2d(k, self.k_linear.weight.unsqueeze(-1).unsqueeze(-1), self.k_linear.bias) qh = q.view(q.shape[0], q.shape[1], self.num_heads, self.hidden_dim // self.num_heads) kh = k.view(k.shape[0], self.num_heads, self.hidden_dim // self.num_heads, k.shape[-2], k.shape[-1]) weights = torch.einsum("bqnc,bnchw->bqnhw", qh * self.normalize_fact, kh) if mask is not None: weights.masked_fill_(mask.unsqueeze(1).unsqueeze(1), float("-inf")) weights = F.softmax(weights.flatten(2), dim=-1).view(weights.size()) weights = self.dropout(weights) return weights def dice_loss(inputs, targets, num_boxes): """ Compute the DICE loss, similar to generalized IOU for masks Args: inputs: A float tensor of arbitrary shape. The predictions for each example. targets: A float tensor with the same shape as inputs. Stores the binary classification label for each element in inputs (0 for the negative class and 1 for the positive class). """ inputs = inputs.sigmoid() inputs = inputs.flatten(1) numerator = 2 * (inputs * targets).sum(1) denominator = inputs.sum(-1) + targets.sum(-1) loss = 1 - (numerator + 1) / (denominator + 1) return loss.sum() / num_boxes def sigmoid_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2): """ Loss used in RetinaNet for dense detection: https://arxiv.org/abs/1708.02002. Args: inputs: A float tensor of arbitrary shape. The predictions for each example. targets: A float tensor with the same shape as inputs. Stores the binary classification label for each element in inputs (0 for the negative class and 1 for the positive class). alpha: (optional) Weighting factor in range (0,1) to balance positive vs negative examples. Default = -1 (no weighting). gamma: Exponent of the modulating factor (1 - p_t) to balance easy vs hard examples. Returns: Loss tensor """ prob = inputs.sigmoid() ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none") p_t = prob * targets + (1 - prob) * (1 - targets) loss = ce_loss * ((1 - p_t) ** gamma) if alpha >= 0: alpha_t = alpha * targets + (1 - alpha) * (1 - targets) loss = alpha_t * loss return loss.mean(1).sum() / num_boxes class PostProcessSegm(nn.Module): def __init__(self, threshold=0.5): super().__init__() self.threshold = threshold @torch.no_grad() def forward(self, results, outputs, orig_target_sizes, max_target_sizes): assert len(orig_target_sizes) == len(max_target_sizes) max_h, max_w = max_target_sizes.max(0)[0].tolist() outputs_masks = outputs["pred_masks"].squeeze(2) outputs_masks = F.interpolate(outputs_masks, size=(max_h, max_w), mode="bilinear", align_corners=False) outputs_masks = (outputs_masks.sigmoid() > self.threshold).cpu() for i, (cur_mask, t, tt) in enumerate(zip(outputs_masks, max_target_sizes, orig_target_sizes)): img_h, img_w = t[0], t[1] results[i]["masks"] = cur_mask[:, :img_h, :img_w].unsqueeze(1) results[i]["masks"] = F.interpolate( results[i]["masks"].float(), size=tuple(tt.tolist()), mode="nearest" ).byte() return results class PostProcessPanoptic(nn.Module): """This class converts the output of the model to the final panoptic result, in the format expected by the coco panoptic API """ def __init__(self, is_thing_map, threshold=0.85): """ Parameters: is_thing_map: This is a whose keys are the class ids, and the values a boolean indicating whether the class is a thing (True) or a stuff (False) class threshold: confidence threshold: segments with confidence lower than this will be deleted """ super().__init__() self.threshold = threshold self.is_thing_map = is_thing_map def forward(self, outputs, processed_sizes, target_sizes=None): """ This function computes the panoptic prediction from the model's predictions. Parameters: outputs: This is a dict coming directly from the model. See the model doc for the content. processed_sizes: This is a list of tuples (or torch tensors) of sizes of the images that were passed to the model, ie the size after data augmentation but before batching. target_sizes: This is a list of tuples (or torch tensors) corresponding to the requested final size of each prediction. If left to None, it will default to the processed_sizes """ if target_sizes is None: target_sizes = processed_sizes assert len(processed_sizes) == len(target_sizes) out_logits, raw_masks, raw_boxes = outputs["pred_logits"], outputs["pred_masks"], outputs["pred_boxes"] assert len(out_logits) == len(raw_masks) == len(target_sizes) preds = [] def to_tuple(tup): if isinstance(tup, tuple): return tup return tuple(tup.cpu().tolist()) for cur_logits, cur_masks, cur_boxes, size, target_size in zip( out_logits, raw_masks, raw_boxes, processed_sizes, target_sizes ): # we filter empty queries and detection below threshold scores, labels = cur_logits.softmax(-1).max(-1) keep = labels.ne(outputs["pred_logits"].shape[-1] - 1) & (scores > self.threshold) cur_scores, cur_classes = cur_logits.softmax(-1).max(-1) cur_scores = cur_scores[keep] cur_classes = cur_classes[keep] cur_masks = cur_masks[keep] cur_masks = interpolate(cur_masks[:, None], to_tuple(size), mode="bilinear").squeeze(1) cur_boxes = box_ops.box_cxcywh_to_xyxy(cur_boxes[keep]) h, w = cur_masks.shape[-2:] assert len(cur_boxes) == len(cur_classes) # It may be that we have several predicted masks for the same stuff class. # In the following, we track the list of masks ids for each stuff class (they are merged later on) cur_masks = cur_masks.flatten(1) stuff_equiv_classes = defaultdict(lambda: []) for k, label in enumerate(cur_classes): if not self.is_thing_map[label.item()]: stuff_equiv_classes[label.item()].append(k) def get_ids_area(masks, scores, dedup=False): # This helper function creates the final panoptic segmentation image # It also returns the area of the masks that appears on the image m_id = masks.transpose(0, 1).softmax(-1) if m_id.shape[-1] == 0: # We didn't detect any mask :( m_id = torch.zeros((h, w), dtype=torch.long, device=m_id.device) else: m_id = m_id.argmax(-1).view(h, w) if dedup: # Merge the masks corresponding to the same stuff class for equiv in stuff_equiv_classes.values(): if len(equiv) > 1: for eq_id in equiv: m_id.masked_fill_(m_id.eq(eq_id), equiv[0]) final_h, final_w = to_tuple(target_size) seg_img = Image.fromarray(id2rgb(m_id.view(h, w).cpu().numpy())) seg_img = seg_img.resize(size=(final_w, final_h), resample=Image.NEAREST) np_seg_img = ( torch.ByteTensor(torch.ByteStorage.from_buffer(seg_img.tobytes())).view(final_h, final_w, 3).numpy() ) m_id = torch.from_numpy(rgb2id(np_seg_img)) area = [] for i in range(len(scores)): area.append(m_id.eq(i).sum().item()) return area, seg_img area, seg_img = get_ids_area(cur_masks, cur_scores, dedup=True) if cur_classes.numel() > 0: # We know filter empty masks as long as we find some while True: filtered_small = torch.as_tensor( [area[i] <= 4 for i, c in enumerate(cur_classes)], dtype=torch.bool, device=keep.device ) if filtered_small.any().item(): cur_scores = cur_scores[~filtered_small] cur_classes = cur_classes[~filtered_small] cur_masks = cur_masks[~filtered_small] area, seg_img = get_ids_area(cur_masks, cur_scores) else: break else: cur_classes = torch.ones(1, dtype=torch.long, device=cur_classes.device) segments_info = [] for i, a in enumerate(area): cat = cur_classes[i].item() segments_info.append({"id": i, "isthing": self.is_thing_map[cat], "category_id": cat, "area": a}) del cur_classes with io.BytesIO() as out: seg_img.save(out, format="PNG") predictions = {"png_string": out.getvalue(), "segments_info": segments_info} preds.append(predictions) return preds
15,573
41.785714
120
py
pmb-nll
pmb-nll-main/src/detr/models/position_encoding.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Various positional encodings for the transformer. """ import math import torch from torch import nn from util.misc import NestedTensor class PositionEmbeddingSine(nn.Module): """ This is a more standard version of the position embedding, very similar to the one used by the Attention is all you need paper, generalized to work on images. """ def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None): super().__init__() self.num_pos_feats = num_pos_feats self.temperature = temperature self.normalize = normalize if scale is not None and normalize is False: raise ValueError("normalize should be True if scale is passed") if scale is None: scale = 2 * math.pi self.scale = scale def forward(self, tensor_list: NestedTensor): x = tensor_list.tensors mask = tensor_list.mask assert mask is not None not_mask = ~mask y_embed = not_mask.cumsum(1, dtype=torch.float32) x_embed = not_mask.cumsum(2, dtype=torch.float32) if self.normalize: eps = 1e-6 y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device) dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats) pos_x = x_embed[:, :, :, None] / dim_t pos_y = y_embed[:, :, :, None] / dim_t pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) return pos class PositionEmbeddingLearned(nn.Module): """ Absolute pos embedding, learned. """ def __init__(self, num_pos_feats=256): super().__init__() self.row_embed = nn.Embedding(50, num_pos_feats) self.col_embed = nn.Embedding(50, num_pos_feats) self.reset_parameters() def reset_parameters(self): nn.init.uniform_(self.row_embed.weight) nn.init.uniform_(self.col_embed.weight) def forward(self, tensor_list: NestedTensor): x = tensor_list.tensors h, w = x.shape[-2:] i = torch.arange(w, device=x.device) j = torch.arange(h, device=x.device) x_emb = self.col_embed(i) y_emb = self.row_embed(j) pos = torch.cat([ x_emb.unsqueeze(0).repeat(h, 1, 1), y_emb.unsqueeze(1).repeat(1, w, 1), ], dim=-1).permute(2, 0, 1).unsqueeze(0).repeat(x.shape[0], 1, 1, 1) return pos def build_position_encoding(args): N_steps = args.hidden_dim // 2 if args.position_embedding in ('v2', 'sine'): # TODO find a better way of exposing other arguments position_embedding = PositionEmbeddingSine(N_steps, normalize=True) elif args.position_embedding in ('v3', 'learned'): position_embedding = PositionEmbeddingLearned(N_steps) else: raise ValueError(f"not supported {args.position_embedding}") return position_embedding
3,336
36.077778
103
py
pmb-nll
pmb-nll-main/src/detr/models/backbone.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Backbone modules. """ from collections import OrderedDict import torch import torch.nn.functional as F import torchvision from torch import nn from torchvision.models._utils import IntermediateLayerGetter from typing import Dict, List from util.misc import NestedTensor, is_main_process from .position_encoding import build_position_encoding class FrozenBatchNorm2d(torch.nn.Module): """ BatchNorm2d where the batch statistics and the affine parameters are fixed. Copy-paste from torchvision.misc.ops with added eps before rqsrt, without which any other models than torchvision.models.resnet[18,34,50,101] produce nans. """ def __init__(self, n): super(FrozenBatchNorm2d, self).__init__() self.register_buffer("weight", torch.ones(n)) self.register_buffer("bias", torch.zeros(n)) self.register_buffer("running_mean", torch.zeros(n)) self.register_buffer("running_var", torch.ones(n)) def _load_from_state_dict(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs): num_batches_tracked_key = prefix + 'num_batches_tracked' if num_batches_tracked_key in state_dict: del state_dict[num_batches_tracked_key] super(FrozenBatchNorm2d, self)._load_from_state_dict( state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs) def forward(self, x): # move reshapes to the beginning # to make it fuser-friendly w = self.weight.reshape(1, -1, 1, 1) b = self.bias.reshape(1, -1, 1, 1) rv = self.running_var.reshape(1, -1, 1, 1) rm = self.running_mean.reshape(1, -1, 1, 1) eps = 1e-5 scale = w * (rv + eps).rsqrt() bias = b - rm * scale return x * scale + bias class BackboneBase(nn.Module): def __init__(self, backbone: nn.Module, train_backbone: bool, num_channels: int, return_interm_layers: bool): super().__init__() for name, parameter in backbone.named_parameters(): if not train_backbone or 'layer2' not in name and 'layer3' not in name and 'layer4' not in name: parameter.requires_grad_(False) if return_interm_layers: return_layers = {"layer1": "0", "layer2": "1", "layer3": "2", "layer4": "3"} else: return_layers = {'layer4': "0"} self.body = IntermediateLayerGetter(backbone, return_layers=return_layers) self.num_channels = num_channels def forward(self, tensor_list: NestedTensor): xs = self.body(tensor_list.tensors) out: Dict[str, NestedTensor] = {} for name, x in xs.items(): m = tensor_list.mask assert m is not None mask = F.interpolate(m[None].float(), size=x.shape[-2:]).to(torch.bool)[0] out[name] = NestedTensor(x, mask) return out class Backbone(BackboneBase): """ResNet backbone with frozen BatchNorm.""" def __init__(self, name: str, train_backbone: bool, return_interm_layers: bool, dilation: bool): backbone = getattr(torchvision.models, name)( replace_stride_with_dilation=[False, False, dilation], pretrained=is_main_process(), norm_layer=FrozenBatchNorm2d) num_channels = 512 if name in ('resnet18', 'resnet34') else 2048 super().__init__(backbone, train_backbone, num_channels, return_interm_layers) class Joiner(nn.Sequential): def __init__(self, backbone, position_embedding): super().__init__(backbone, position_embedding) def forward(self, tensor_list: NestedTensor): xs = self[0](tensor_list) out: List[NestedTensor] = [] pos = [] for name, x in xs.items(): out.append(x) # position encoding pos.append(self[1](x).to(x.tensors.dtype)) return out, pos def build_backbone(args): position_embedding = build_position_encoding(args) train_backbone = args.lr_backbone > 0 return_interm_layers = args.masks backbone = Backbone(args.backbone, train_backbone, return_interm_layers, args.dilation) model = Joiner(backbone, position_embedding) model.num_channels = backbone.num_channels return model
4,437
35.983333
113
py
pmb-nll
pmb-nll-main/src/detr/models/transformer.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ DETR Transformer class. Copy-paste from torch.nn.Transformer with modifications: * positional encodings are passed in MHattention * extra LN at the end of encoder is removed * decoder returns a stack of activations from all decoding layers """ import copy from typing import Optional, List import torch import torch.nn.functional as F from torch import nn, Tensor class Transformer(nn.Module): def __init__(self, d_model=512, nhead=8, num_encoder_layers=6, num_decoder_layers=6, dim_feedforward=2048, dropout=0.1, activation="relu", normalize_before=False, return_intermediate_dec=False): super().__init__() encoder_layer = TransformerEncoderLayer(d_model, nhead, dim_feedforward, dropout, activation, normalize_before) encoder_norm = nn.LayerNorm(d_model) if normalize_before else None self.encoder = TransformerEncoder(encoder_layer, num_encoder_layers, encoder_norm) decoder_layer = TransformerDecoderLayer(d_model, nhead, dim_feedforward, dropout, activation, normalize_before) decoder_norm = nn.LayerNorm(d_model) self.decoder = TransformerDecoder(decoder_layer, num_decoder_layers, decoder_norm, return_intermediate=return_intermediate_dec) self._reset_parameters() self.d_model = d_model self.nhead = nhead def _reset_parameters(self): for p in self.parameters(): if p.dim() > 1: nn.init.xavier_uniform_(p) def forward(self, src, mask, query_embed, pos_embed): # flatten NxCxHxW to HWxNxC bs, c, h, w = src.shape src = src.flatten(2).permute(2, 0, 1) pos_embed = pos_embed.flatten(2).permute(2, 0, 1) query_embed = query_embed.unsqueeze(1).repeat(1, bs, 1) mask = mask.flatten(1) tgt = torch.zeros_like(query_embed) memory = self.encoder(src, src_key_padding_mask=mask, pos=pos_embed) hs = self.decoder(tgt, memory, memory_key_padding_mask=mask, pos=pos_embed, query_pos=query_embed) return hs.transpose(1, 2), memory.permute(1, 2, 0).view(bs, c, h, w) class TransformerEncoder(nn.Module): def __init__(self, encoder_layer, num_layers, norm=None): super().__init__() self.layers = _get_clones(encoder_layer, num_layers) self.num_layers = num_layers self.norm = norm def forward(self, src, mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None): output = src for layer in self.layers: output = layer(output, src_mask=mask, src_key_padding_mask=src_key_padding_mask, pos=pos) if self.norm is not None: output = self.norm(output) return output class TransformerDecoder(nn.Module): def __init__(self, decoder_layer, num_layers, norm=None, return_intermediate=False): super().__init__() self.layers = _get_clones(decoder_layer, num_layers) self.num_layers = num_layers self.norm = norm self.return_intermediate = return_intermediate def forward(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None): output = tgt intermediate = [] for layer in self.layers: output = layer(output, memory, tgt_mask=tgt_mask, memory_mask=memory_mask, tgt_key_padding_mask=tgt_key_padding_mask, memory_key_padding_mask=memory_key_padding_mask, pos=pos, query_pos=query_pos) if self.return_intermediate: intermediate.append(self.norm(output)) if self.norm is not None: output = self.norm(output) if self.return_intermediate: intermediate.pop() intermediate.append(output) if self.return_intermediate: return torch.stack(intermediate) return output.unsqueeze(0) class TransformerEncoderLayer(nn.Module): def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation="relu", normalize_before=False): super().__init__() self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) # Implementation of Feedforward model self.linear1 = nn.Linear(d_model, dim_feedforward) self.dropout = nn.Dropout(dropout) self.linear2 = nn.Linear(dim_feedforward, d_model) self.norm1 = nn.LayerNorm(d_model) self.norm2 = nn.LayerNorm(d_model) self.dropout1 = nn.Dropout(dropout) self.dropout2 = nn.Dropout(dropout) self.activation = _get_activation_fn(activation) self.normalize_before = normalize_before def with_pos_embed(self, tensor, pos: Optional[Tensor]): return tensor if pos is None else tensor + pos def forward_post(self, src, src_mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None): q = k = self.with_pos_embed(src, pos) src2 = self.self_attn(q, k, value=src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0] src = src + self.dropout1(src2) src = self.norm1(src) src2 = self.linear2(self.dropout(self.activation(self.linear1(src)))) src = src + self.dropout2(src2) src = self.norm2(src) return src def forward_pre(self, src, src_mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None): src2 = self.norm1(src) q = k = self.with_pos_embed(src2, pos) src2 = self.self_attn(q, k, value=src2, attn_mask=src_mask, key_padding_mask=src_key_padding_mask)[0] src = src + self.dropout1(src2) src2 = self.norm2(src) src2 = self.linear2(self.dropout(self.activation(self.linear1(src2)))) src = src + self.dropout2(src2) return src def forward(self, src, src_mask: Optional[Tensor] = None, src_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None): if self.normalize_before: return self.forward_pre(src, src_mask, src_key_padding_mask, pos) return self.forward_post(src, src_mask, src_key_padding_mask, pos) class TransformerDecoderLayer(nn.Module): def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation="relu", normalize_before=False): super().__init__() self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) self.multihead_attn = nn.MultiheadAttention(d_model, nhead, dropout=dropout) # Implementation of Feedforward model self.linear1 = nn.Linear(d_model, dim_feedforward) self.dropout = nn.Dropout(dropout) self.linear2 = nn.Linear(dim_feedforward, d_model) self.norm1 = nn.LayerNorm(d_model) self.norm2 = nn.LayerNorm(d_model) self.norm3 = nn.LayerNorm(d_model) self.dropout1 = nn.Dropout(dropout) self.dropout2 = nn.Dropout(dropout) self.dropout3 = nn.Dropout(dropout) self.activation = _get_activation_fn(activation) self.normalize_before = normalize_before def with_pos_embed(self, tensor, pos: Optional[Tensor]): return tensor if pos is None else tensor + pos def forward_post(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None): q = k = self.with_pos_embed(tgt, query_pos) tgt2 = self.self_attn(q, k, value=tgt, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask)[0] tgt = tgt + self.dropout1(tgt2) tgt = self.norm1(tgt) tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt, query_pos), key=self.with_pos_embed(memory, pos), value=memory, attn_mask=memory_mask, key_padding_mask=memory_key_padding_mask)[0] tgt = tgt + self.dropout2(tgt2) tgt = self.norm2(tgt) tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt)))) tgt = tgt + self.dropout3(tgt2) tgt = self.norm3(tgt) return tgt def forward_pre(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None): tgt2 = self.norm1(tgt) q = k = self.with_pos_embed(tgt2, query_pos) tgt2 = self.self_attn(q, k, value=tgt2, attn_mask=tgt_mask, key_padding_mask=tgt_key_padding_mask)[0] tgt = tgt + self.dropout1(tgt2) tgt2 = self.norm2(tgt) tgt2 = self.multihead_attn(query=self.with_pos_embed(tgt2, query_pos), key=self.with_pos_embed(memory, pos), value=memory, attn_mask=memory_mask, key_padding_mask=memory_key_padding_mask)[0] tgt = tgt + self.dropout2(tgt2) tgt2 = self.norm3(tgt) tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2)))) tgt = tgt + self.dropout3(tgt2) return tgt def forward(self, tgt, memory, tgt_mask: Optional[Tensor] = None, memory_mask: Optional[Tensor] = None, tgt_key_padding_mask: Optional[Tensor] = None, memory_key_padding_mask: Optional[Tensor] = None, pos: Optional[Tensor] = None, query_pos: Optional[Tensor] = None): if self.normalize_before: return self.forward_pre(tgt, memory, tgt_mask, memory_mask, tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos) return self.forward_post(tgt, memory, tgt_mask, memory_mask, tgt_key_padding_mask, memory_key_padding_mask, pos, query_pos) def _get_clones(module, N): return nn.ModuleList([copy.deepcopy(module) for i in range(N)]) def build_transformer(args): return Transformer( d_model=args.hidden_dim, dropout=args.dropout, nhead=args.nheads, dim_feedforward=args.dim_feedforward, num_encoder_layers=args.enc_layers, num_decoder_layers=args.dec_layers, normalize_before=args.pre_norm, return_intermediate_dec=True, ) def _get_activation_fn(activation): """Return an activation function given a string""" if activation == "relu": return F.relu if activation == "gelu": return F.gelu if activation == "glu": return F.glu raise RuntimeError(F"activation should be relu/gelu, not {activation}.")
12,162
39.815436
98
py
pmb-nll
pmb-nll-main/src/detr/d2/converter.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Helper script to convert models trained with the main version of DETR to be used with the Detectron2 version. """ import json import argparse import numpy as np import torch def parse_args(): parser = argparse.ArgumentParser("D2 model converter") parser.add_argument("--source_model", default="", type=str, help="Path or url to the DETR model to convert") parser.add_argument("--output_model", default="", type=str, help="Path where to save the converted model") return parser.parse_args() def main(): args = parse_args() # D2 expects contiguous classes, so we need to remap the 92 classes from DETR # fmt: off coco_idx = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90, 91] # fmt: on coco_idx = np.array(coco_idx) if args.source_model.startswith("https"): checkpoint = torch.hub.load_state_dict_from_url(args.source_model, map_location="cpu", check_hash=True) else: checkpoint = torch.load(args.source_model, map_location="cpu") model_to_convert = checkpoint["model"] model_converted = {} for k in model_to_convert.keys(): old_k = k if "backbone" in k: k = k.replace("backbone.0.body.", "") if "layer" not in k: k = "stem." + k for t in [1, 2, 3, 4]: k = k.replace(f"layer{t}", f"res{t + 1}") for t in [1, 2, 3]: k = k.replace(f"bn{t}", f"conv{t}.norm") k = k.replace("downsample.0", "shortcut") k = k.replace("downsample.1", "shortcut.norm") k = "backbone.0.backbone." + k k = "detr." + k print(old_k, "->", k) if "class_embed" in old_k: v = model_to_convert[old_k].detach() if v.shape[0] == 92: shape_old = v.shape model_converted[k] = v[coco_idx] print("Head conversion: changing shape from {} to {}".format(shape_old, model_converted[k].shape)) continue model_converted[k] = model_to_convert[old_k].detach() model_to_save = {"model": model_converted} torch.save(model_to_save, args.output_model) if __name__ == "__main__": main()
2,590
36.014286
114
py
pmb-nll
pmb-nll-main/src/detr/d2/train_net.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ DETR Training Script. This script is a simplified version of the training script in detectron2/tools. """ import os import sys import itertools # fmt: off sys.path.insert(1, os.path.join(sys.path[0], '..')) # fmt: on import time from typing import Any, Dict, List, Set import torch import detectron2.utils.comm as comm from d2.detr import DetrDatasetMapper, add_detr_config from detectron2.checkpoint import DetectionCheckpointer from detectron2.config import get_cfg from detectron2.data import MetadataCatalog, build_detection_train_loader from detectron2.engine import DefaultTrainer, default_argument_parser, default_setup, launch from detectron2.evaluation import COCOEvaluator, verify_results from detectron2.solver.build import maybe_add_gradient_clipping class Trainer(DefaultTrainer): """ Extension of the Trainer class adapted to DETR. """ @classmethod def build_evaluator(cls, cfg, dataset_name, output_folder=None): """ Create evaluator(s) for a given dataset. This uses the special metadata "evaluator_type" associated with each builtin dataset. For your own dataset, you can simply create an evaluator manually in your script and do not have to worry about the hacky if-else logic here. """ if output_folder is None: output_folder = os.path.join(cfg.OUTPUT_DIR, "inference") return COCOEvaluator(dataset_name, cfg, True, output_folder) @classmethod def build_train_loader(cls, cfg): if "Detr" == cfg.MODEL.META_ARCHITECTURE: mapper = DetrDatasetMapper(cfg, True) else: mapper = None return build_detection_train_loader(cfg, mapper=mapper) @classmethod def build_optimizer(cls, cfg, model): params: List[Dict[str, Any]] = [] memo: Set[torch.nn.parameter.Parameter] = set() for key, value in model.named_parameters(recurse=True): if not value.requires_grad: continue # Avoid duplicating parameters if value in memo: continue memo.add(value) lr = cfg.SOLVER.BASE_LR weight_decay = cfg.SOLVER.WEIGHT_DECAY if "backbone" in key: lr = lr * cfg.SOLVER.BACKBONE_MULTIPLIER params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}] def maybe_add_full_model_gradient_clipping(optim): # optim: the optimizer class # detectron2 doesn't have full model gradient clipping now clip_norm_val = cfg.SOLVER.CLIP_GRADIENTS.CLIP_VALUE enable = ( cfg.SOLVER.CLIP_GRADIENTS.ENABLED and cfg.SOLVER.CLIP_GRADIENTS.CLIP_TYPE == "full_model" and clip_norm_val > 0.0 ) class FullModelGradientClippingOptimizer(optim): def step(self, closure=None): all_params = itertools.chain(*[x["params"] for x in self.param_groups]) torch.nn.utils.clip_grad_norm_(all_params, clip_norm_val) super().step(closure=closure) return FullModelGradientClippingOptimizer if enable else optim optimizer_type = cfg.SOLVER.OPTIMIZER if optimizer_type == "SGD": optimizer = maybe_add_full_model_gradient_clipping(torch.optim.SGD)( params, cfg.SOLVER.BASE_LR, momentum=cfg.SOLVER.MOMENTUM ) elif optimizer_type == "ADAMW": optimizer = maybe_add_full_model_gradient_clipping(torch.optim.AdamW)( params, cfg.SOLVER.BASE_LR ) else: raise NotImplementedError(f"no optimizer type {optimizer_type}") if not cfg.SOLVER.CLIP_GRADIENTS.CLIP_TYPE == "full_model": optimizer = maybe_add_gradient_clipping(cfg, optimizer) return optimizer def setup(args): """ Create configs and perform basic setups. """ cfg = get_cfg() add_detr_config(cfg) cfg.merge_from_file(args.config_file) cfg.merge_from_list(args.opts) cfg.freeze() default_setup(cfg, args) return cfg def main(args): cfg = setup(args) if args.eval_only: model = Trainer.build_model(cfg) DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load(cfg.MODEL.WEIGHTS, resume=args.resume) res = Trainer.test(cfg, model) if comm.is_main_process(): verify_results(cfg, res) return res trainer = Trainer(cfg) trainer.resume_or_load(resume=args.resume) return trainer.train() if __name__ == "__main__": args = default_argument_parser().parse_args() print("Command Line Args:", args) launch( main, args.num_gpus, num_machines=args.num_machines, machine_rank=args.machine_rank, dist_url=args.dist_url, args=(args,), )
4,999
33.246575
115
py
pmb-nll
pmb-nll-main/src/detr/d2/detr/detr.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import logging import math from typing import List import numpy as np import torch import torch.distributed as dist import torch.nn.functional as F from scipy.optimize import linear_sum_assignment from torch import nn from detectron2.layers import ShapeSpec from detectron2.modeling import META_ARCH_REGISTRY, build_backbone, detector_postprocess from detectron2.structures import Boxes, ImageList, Instances, BitMasks, PolygonMasks from detectron2.utils.logger import log_first_n from fvcore.nn import giou_loss, smooth_l1_loss from models.backbone import Joiner from models.detr import DETR, SetCriterion from models.matcher import HungarianMatcher from models.position_encoding import PositionEmbeddingSine from models.transformer import Transformer from models.segmentation import DETRsegm, PostProcessPanoptic, PostProcessSegm from util.box_ops import box_cxcywh_to_xyxy, box_xyxy_to_cxcywh from util.misc import NestedTensor from datasets.coco import convert_coco_poly_to_mask __all__ = ["Detr"] class MaskedBackbone(nn.Module): """ This is a thin wrapper around D2's backbone to provide padding masking""" def __init__(self, cfg): super().__init__() self.backbone = build_backbone(cfg) backbone_shape = self.backbone.output_shape() self.feature_strides = [backbone_shape[f].stride for f in backbone_shape.keys()] self.num_channels = backbone_shape[list(backbone_shape.keys())[-1]].channels def forward(self, images): features = self.backbone(images.tensor) masks = self.mask_out_padding( [features_per_level.shape for features_per_level in features.values()], images.image_sizes, images.tensor.device, ) assert len(features) == len(masks) for i, k in enumerate(features.keys()): features[k] = NestedTensor(features[k], masks[i]) return features def mask_out_padding(self, feature_shapes, image_sizes, device): masks = [] assert len(feature_shapes) == len(self.feature_strides) for idx, shape in enumerate(feature_shapes): N, _, H, W = shape masks_per_feature_level = torch.ones((N, H, W), dtype=torch.bool, device=device) for img_idx, (h, w) in enumerate(image_sizes): masks_per_feature_level[ img_idx, : int(np.ceil(float(h) / self.feature_strides[idx])), : int(np.ceil(float(w) / self.feature_strides[idx])), ] = 0 masks.append(masks_per_feature_level) return masks @META_ARCH_REGISTRY.register() class Detr(nn.Module): """ Implement Detr """ def __init__(self, cfg): super().__init__() self.device = torch.device(cfg.MODEL.DEVICE) self.num_classes = cfg.MODEL.DETR.NUM_CLASSES self.mask_on = cfg.MODEL.MASK_ON hidden_dim = cfg.MODEL.DETR.HIDDEN_DIM num_queries = cfg.MODEL.DETR.NUM_OBJECT_QUERIES # Transformer parameters: nheads = cfg.MODEL.DETR.NHEADS dropout = cfg.MODEL.DETR.DROPOUT dim_feedforward = cfg.MODEL.DETR.DIM_FEEDFORWARD enc_layers = cfg.MODEL.DETR.ENC_LAYERS dec_layers = cfg.MODEL.DETR.DEC_LAYERS pre_norm = cfg.MODEL.DETR.PRE_NORM # Loss parameters: giou_weight = cfg.MODEL.DETR.GIOU_WEIGHT l1_weight = cfg.MODEL.DETR.L1_WEIGHT deep_supervision = cfg.MODEL.DETR.DEEP_SUPERVISION no_object_weight = cfg.MODEL.DETR.NO_OBJECT_WEIGHT N_steps = hidden_dim // 2 d2_backbone = MaskedBackbone(cfg) backbone = Joiner(d2_backbone, PositionEmbeddingSine(N_steps, normalize=True)) backbone.num_channels = d2_backbone.num_channels transformer = Transformer( d_model=hidden_dim, dropout=dropout, nhead=nheads, dim_feedforward=dim_feedforward, num_encoder_layers=enc_layers, num_decoder_layers=dec_layers, normalize_before=pre_norm, return_intermediate_dec=deep_supervision, ) self.detr = DETR( backbone, transformer, num_classes=self.num_classes, num_queries=num_queries, aux_loss=deep_supervision ) if self.mask_on: frozen_weights = cfg.MODEL.DETR.FROZEN_WEIGHTS if frozen_weights != '': print("LOAD pre-trained weights") weight = torch.load(frozen_weights, map_location=lambda storage, loc: storage)['model'] new_weight = {} for k, v in weight.items(): if 'detr.' in k: new_weight[k.replace('detr.', '')] = v else: print(f"Skipping loading weight {k} from frozen model") del weight self.detr.load_state_dict(new_weight) del new_weight self.detr = DETRsegm(self.detr, freeze_detr=(frozen_weights != '')) self.seg_postprocess = PostProcessSegm self.detr.to(self.device) # building criterion matcher = HungarianMatcher(cost_class=1, cost_bbox=l1_weight, cost_giou=giou_weight) weight_dict = {"loss_ce": 1, "loss_bbox": l1_weight} weight_dict["loss_giou"] = giou_weight if deep_supervision: aux_weight_dict = {} for i in range(dec_layers - 1): aux_weight_dict.update({k + f"_{i}": v for k, v in weight_dict.items()}) weight_dict.update(aux_weight_dict) losses = ["labels", "boxes", "cardinality"] if self.mask_on: losses += ["masks"] self.criterion = SetCriterion( self.num_classes, matcher=matcher, weight_dict=weight_dict, eos_coef=no_object_weight, losses=losses, ) self.criterion.to(self.device) pixel_mean = torch.Tensor(cfg.MODEL.PIXEL_MEAN).to(self.device).view(3, 1, 1) pixel_std = torch.Tensor(cfg.MODEL.PIXEL_STD).to(self.device).view(3, 1, 1) self.normalizer = lambda x: (x - pixel_mean) / pixel_std self.to(self.device) def forward(self, batched_inputs): """ Args: batched_inputs: a list, batched outputs of :class:`DatasetMapper` . Each item in the list contains the inputs for one image. For now, each item in the list is a dict that contains: * image: Tensor, image in (C, H, W) format. * instances: Instances Other information that's included in the original dicts, such as: * "height", "width" (int): the output resolution of the model, used in inference. See :meth:`postprocess` for details. Returns: dict[str: Tensor]: mapping from a named loss to a tensor storing the loss. Used during training only. """ images = self.preprocess_image(batched_inputs) output = self.detr(images) if self.training: gt_instances = [x["instances"].to(self.device) for x in batched_inputs] targets = self.prepare_targets(gt_instances) loss_dict = self.criterion(output, targets) weight_dict = self.criterion.weight_dict for k in loss_dict.keys(): if k in weight_dict: loss_dict[k] *= weight_dict[k] return loss_dict else: box_cls = output["pred_logits"] box_pred = output["pred_boxes"] mask_pred = output["pred_masks"] if self.mask_on else None results = self.inference(box_cls, box_pred, mask_pred, images.image_sizes) processed_results = [] for results_per_image, input_per_image, image_size in zip(results, batched_inputs, images.image_sizes): height = input_per_image.get("height", image_size[0]) width = input_per_image.get("width", image_size[1]) r = detector_postprocess(results_per_image, height, width) processed_results.append({"instances": r}) return processed_results def prepare_targets(self, targets): new_targets = [] for targets_per_image in targets: h, w = targets_per_image.image_size image_size_xyxy = torch.as_tensor([w, h, w, h], dtype=torch.float, device=self.device) gt_classes = targets_per_image.gt_classes gt_boxes = targets_per_image.gt_boxes.tensor / image_size_xyxy gt_boxes = box_xyxy_to_cxcywh(gt_boxes) new_targets.append({"labels": gt_classes, "boxes": gt_boxes}) if self.mask_on and hasattr(targets_per_image, 'gt_masks'): gt_masks = targets_per_image.gt_masks gt_masks = convert_coco_poly_to_mask(gt_masks.polygons, h, w) new_targets[-1].update({'masks': gt_masks}) return new_targets def inference(self, box_cls, box_pred, mask_pred, image_sizes): """ Arguments: box_cls (Tensor): tensor of shape (batch_size, num_queries, K). The tensor predicts the classification probability for each query. box_pred (Tensor): tensors of shape (batch_size, num_queries, 4). The tensor predicts 4-vector (x,y,w,h) box regression values for every queryx image_sizes (List[torch.Size]): the input image sizes Returns: results (List[Instances]): a list of #images elements. """ assert len(box_cls) == len(image_sizes) results = [] # For each box we assign the best class or the second best if the best on is `no_object`. scores, labels = F.softmax(box_cls, dim=-1)[:, :, :-1].max(-1) for i, (scores_per_image, labels_per_image, box_pred_per_image, image_size) in enumerate(zip( scores, labels, box_pred, image_sizes )): result = Instances(image_size) result.pred_boxes = Boxes(box_cxcywh_to_xyxy(box_pred_per_image)) result.pred_boxes.scale(scale_x=image_size[1], scale_y=image_size[0]) if self.mask_on: mask = F.interpolate(mask_pred[i].unsqueeze(0), size=image_size, mode='bilinear', align_corners=False) mask = mask[0].sigmoid() > 0.5 B, N, H, W = mask_pred.shape mask = BitMasks(mask.cpu()).crop_and_resize(result.pred_boxes.tensor.cpu(), 32) result.pred_masks = mask.unsqueeze(1).to(mask_pred[0].device) result.scores = scores_per_image result.pred_classes = labels_per_image results.append(result) return results def preprocess_image(self, batched_inputs): """ Normalize, pad and batch the input images. """ images = [self.normalizer(x["image"].to(self.device)) for x in batched_inputs] images = ImageList.from_tensors(images) return images
11,143
41.534351
118
py
pmb-nll
pmb-nll-main/src/detr/d2/detr/dataset_mapper.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import copy import logging import numpy as np import torch from detectron2.data import detection_utils as utils from detectron2.data import transforms as T from detectron2.data.transforms import TransformGen __all__ = ["DetrDatasetMapper"] def build_transform_gen(cfg, is_train): """ Create a list of :class:`TransformGen` from config. Returns: list[TransformGen] """ if is_train: min_size = cfg.INPUT.MIN_SIZE_TRAIN max_size = cfg.INPUT.MAX_SIZE_TRAIN sample_style = cfg.INPUT.MIN_SIZE_TRAIN_SAMPLING else: min_size = cfg.INPUT.MIN_SIZE_TEST max_size = cfg.INPUT.MAX_SIZE_TEST sample_style = "choice" if sample_style == "range": assert len(min_size) == 2, "more than 2 ({}) min_size(s) are provided for ranges".format(len(min_size)) logger = logging.getLogger(__name__) tfm_gens = [] if is_train: tfm_gens.append(T.RandomFlip()) tfm_gens.append(T.ResizeShortestEdge(min_size, max_size, sample_style)) if is_train: logger.info("TransformGens used in training: " + str(tfm_gens)) return tfm_gens class DetrDatasetMapper: """ A callable which takes a dataset dict in Detectron2 Dataset format, and map it into a format used by DETR. The callable currently does the following: 1. Read the image from "file_name" 2. Applies geometric transforms to the image and annotation 3. Find and applies suitable cropping to the image and annotation 4. Prepare image and annotation to Tensors """ def __init__(self, cfg, is_train=True): if cfg.INPUT.CROP.ENABLED and is_train: self.crop_gen = [ T.ResizeShortestEdge([400, 500, 600], sample_style="choice"), T.RandomCrop(cfg.INPUT.CROP.TYPE, cfg.INPUT.CROP.SIZE), ] else: self.crop_gen = None self.mask_on = cfg.MODEL.MASK_ON self.tfm_gens = build_transform_gen(cfg, is_train) logging.getLogger(__name__).info( "Full TransformGens used in training: {}, crop: {}".format(str(self.tfm_gens), str(self.crop_gen)) ) self.img_format = cfg.INPUT.FORMAT self.is_train = is_train def __call__(self, dataset_dict): """ Args: dataset_dict (dict): Metadata of one image, in Detectron2 Dataset format. Returns: dict: a format that builtin models in detectron2 accept """ dataset_dict = copy.deepcopy(dataset_dict) # it will be modified by code below image = utils.read_image(dataset_dict["file_name"], format=self.img_format) utils.check_image_size(dataset_dict, image) if self.crop_gen is None: image, transforms = T.apply_transform_gens(self.tfm_gens, image) else: if np.random.rand() > 0.5: image, transforms = T.apply_transform_gens(self.tfm_gens, image) else: image, transforms = T.apply_transform_gens( self.tfm_gens[:-1] + self.crop_gen + self.tfm_gens[-1:], image ) image_shape = image.shape[:2] # h, w # Pytorch's dataloader is efficient on torch.Tensor due to shared-memory, # but not efficient on large generic data structures due to the use of pickle & mp.Queue. # Therefore it's important to use torch.Tensor. dataset_dict["image"] = torch.as_tensor(np.ascontiguousarray(image.transpose(2, 0, 1))) if not self.is_train: # USER: Modify this if you want to keep them for some reason. dataset_dict.pop("annotations", None) return dataset_dict if "annotations" in dataset_dict: # USER: Modify this if you want to keep them for some reason. for anno in dataset_dict["annotations"]: if not self.mask_on: anno.pop("segmentation", None) anno.pop("keypoints", None) # USER: Implement additional transformations if you have other types of data annos = [ utils.transform_instance_annotations(obj, transforms, image_shape) for obj in dataset_dict.pop("annotations") if obj.get("iscrowd", 0) == 0 ] instances = utils.annotations_to_instances(annos, image_shape) dataset_dict["instances"] = utils.filter_empty_instances(instances) return dataset_dict
4,570
36.162602
111
py
pmb-nll
pmb-nll-main/src/detr/util/plot_utils.py
""" Plotting utilities to visualize training logs. """ import torch import pandas as pd import numpy as np import seaborn as sns import matplotlib.pyplot as plt from pathlib import Path, PurePath def plot_logs(logs, fields=('class_error', 'loss_bbox_unscaled', 'mAP'), ewm_col=0, log_name='log.txt'): ''' Function to plot specific fields from training log(s). Plots both training and test results. :: Inputs - logs = list containing Path objects, each pointing to individual dir with a log file - fields = which results to plot from each log file - plots both training and test for each field. - ewm_col = optional, which column to use as the exponential weighted smoothing of the plots - log_name = optional, name of log file if different than default 'log.txt'. :: Outputs - matplotlib plots of results in fields, color coded for each log file. - solid lines are training results, dashed lines are test results. ''' func_name = "plot_utils.py::plot_logs" # verify logs is a list of Paths (list[Paths]) or single Pathlib object Path, # convert single Path to list to avoid 'not iterable' error if not isinstance(logs, list): if isinstance(logs, PurePath): logs = [logs] print(f"{func_name} info: logs param expects a list argument, converted to list[Path].") else: raise ValueError(f"{func_name} - invalid argument for logs parameter.\n \ Expect list[Path] or single Path obj, received {type(logs)}") # Quality checks - verify valid dir(s), that every item in list is Path object, and that log_name exists in each dir for i, dir in enumerate(logs): if not isinstance(dir, PurePath): raise ValueError(f"{func_name} - non-Path object in logs argument of {type(dir)}: \n{dir}") if not dir.exists(): raise ValueError(f"{func_name} - invalid directory in logs argument:\n{dir}") # verify log_name exists fn = Path(dir / log_name) if not fn.exists(): print(f"-> missing {log_name}. Have you gotten to Epoch 1 in training?") print(f"--> full path of missing log file: {fn}") return # load log file(s) and plot dfs = [pd.read_json(Path(p) / log_name, lines=True) for p in logs] fig, axs = plt.subplots(ncols=len(fields), figsize=(16, 5)) for df, color in zip(dfs, sns.color_palette(n_colors=len(logs))): for j, field in enumerate(fields): if field == 'mAP': coco_eval = pd.DataFrame( np.stack(df.test_coco_eval_bbox.dropna().values)[:, 1] ).ewm(com=ewm_col).mean() axs[j].plot(coco_eval, c=color) else: df.interpolate().ewm(com=ewm_col).mean().plot( y=[f'train_{field}', f'test_{field}'], ax=axs[j], color=[color] * 2, style=['-', '--'] ) for ax, field in zip(axs, fields): ax.legend([Path(p).name for p in logs]) ax.set_title(field) def plot_precision_recall(files, naming_scheme='iter'): if naming_scheme == 'exp_id': # name becomes exp_id names = [f.parts[-3] for f in files] elif naming_scheme == 'iter': names = [f.stem for f in files] else: raise ValueError(f'not supported {naming_scheme}') fig, axs = plt.subplots(ncols=2, figsize=(16, 5)) for f, color, name in zip(files, sns.color_palette("Blues", n_colors=len(files)), names): data = torch.load(f) # precision is n_iou, n_points, n_cat, n_area, max_det precision = data['precision'] recall = data['params'].recThrs scores = data['scores'] # take precision for all classes, all areas and 100 detections precision = precision[0, :, :, 0, -1].mean(1) scores = scores[0, :, :, 0, -1].mean(1) prec = precision.mean() rec = data['recall'][0, :, 0, -1].mean() print(f'{naming_scheme} {name}: mAP@50={prec * 100: 05.1f}, ' + f'score={scores.mean():0.3f}, ' + f'f1={2 * prec * rec / (prec + rec + 1e-8):0.3f}' ) axs[0].plot(recall, precision, c=color) axs[1].plot(recall, scores, c=color) axs[0].set_title('Precision / Recall') axs[0].legend(names) axs[1].set_title('Scores / Recall') axs[1].legend(names) return fig, axs
4,514
40.805556
120
py
pmb-nll
pmb-nll-main/src/detr/util/misc.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Misc functions, including distributed helpers. Mostly copy-paste from torchvision references. """ import os import subprocess import time from collections import defaultdict, deque import datetime import pickle from typing import Optional, List import torch import torch.distributed as dist from torch import Tensor # needed due to empty tensor bug in pytorch and torchvision 0.5 import torchvision if float(torchvision.__version__.split(".")[1]) < 7.0: from torchvision.ops import _new_empty_tensor from torchvision.ops.misc import _output_size class SmoothedValue(object): """Track a series of values and provide access to smoothed values over a window or the global series average. """ def __init__(self, window_size=20, fmt=None): if fmt is None: fmt = "{median:.4f} ({global_avg:.4f})" self.deque = deque(maxlen=window_size) self.total = 0.0 self.count = 0 self.fmt = fmt def update(self, value, n=1): self.deque.append(value) self.count += n self.total += value * n def synchronize_between_processes(self): """ Warning: does not synchronize the deque! """ if not is_dist_avail_and_initialized(): return t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda') dist.barrier() dist.all_reduce(t) t = t.tolist() self.count = int(t[0]) self.total = t[1] @property def median(self): d = torch.tensor(list(self.deque)) return d.median().item() @property def avg(self): d = torch.tensor(list(self.deque), dtype=torch.float32) return d.mean().item() @property def global_avg(self): return self.total / self.count @property def max(self): return max(self.deque) @property def value(self): return self.deque[-1] def __str__(self): return self.fmt.format( median=self.median, avg=self.avg, global_avg=self.global_avg, max=self.max, value=self.value) def all_gather(data): """ Run all_gather on arbitrary picklable data (not necessarily tensors) Args: data: any picklable object Returns: list[data]: list of data gathered from each rank """ world_size = get_world_size() if world_size == 1: return [data] # serialized to a Tensor buffer = pickle.dumps(data) storage = torch.ByteStorage.from_buffer(buffer) tensor = torch.ByteTensor(storage).to("cuda") # obtain Tensor size of each rank local_size = torch.tensor([tensor.numel()], device="cuda") size_list = [torch.tensor([0], device="cuda") for _ in range(world_size)] dist.all_gather(size_list, local_size) size_list = [int(size.item()) for size in size_list] max_size = max(size_list) # receiving Tensor from all ranks # we pad the tensor because torch all_gather does not support # gathering tensors of different shapes tensor_list = [] for _ in size_list: tensor_list.append(torch.empty((max_size,), dtype=torch.uint8, device="cuda")) if local_size != max_size: padding = torch.empty(size=(max_size - local_size,), dtype=torch.uint8, device="cuda") tensor = torch.cat((tensor, padding), dim=0) dist.all_gather(tensor_list, tensor) data_list = [] for size, tensor in zip(size_list, tensor_list): buffer = tensor.cpu().numpy().tobytes()[:size] data_list.append(pickle.loads(buffer)) return data_list def reduce_dict(input_dict, average=True): """ Args: input_dict (dict): all the values will be reduced average (bool): whether to do average or sum Reduce the values in the dictionary from all processes so that all processes have the averaged results. Returns a dict with the same fields as input_dict, after reduction. """ world_size = get_world_size() if world_size < 2: return input_dict with torch.no_grad(): names = [] values = [] # sort the keys so that they are consistent across processes for k in sorted(input_dict.keys()): names.append(k) values.append(input_dict[k]) values = torch.stack(values, dim=0) dist.all_reduce(values) if average: values /= world_size reduced_dict = {k: v for k, v in zip(names, values)} return reduced_dict class MetricLogger(object): def __init__(self, delimiter="\t"): self.meters = defaultdict(SmoothedValue) self.delimiter = delimiter def update(self, **kwargs): for k, v in kwargs.items(): if isinstance(v, torch.Tensor): v = v.item() assert isinstance(v, (float, int)) self.meters[k].update(v) def __getattr__(self, attr): if attr in self.meters: return self.meters[attr] if attr in self.__dict__: return self.__dict__[attr] raise AttributeError("'{}' object has no attribute '{}'".format( type(self).__name__, attr)) def __str__(self): loss_str = [] for name, meter in self.meters.items(): loss_str.append( "{}: {}".format(name, str(meter)) ) return self.delimiter.join(loss_str) def synchronize_between_processes(self): for meter in self.meters.values(): meter.synchronize_between_processes() def add_meter(self, name, meter): self.meters[name] = meter def log_every(self, iterable, print_freq, header=None): i = 0 if not header: header = '' start_time = time.time() end = time.time() iter_time = SmoothedValue(fmt='{avg:.4f}') data_time = SmoothedValue(fmt='{avg:.4f}') space_fmt = ':' + str(len(str(len(iterable)))) + 'd' if torch.cuda.is_available(): log_msg = self.delimiter.join([ header, '[{0' + space_fmt + '}/{1}]', 'eta: {eta}', '{meters}', 'time: {time}', 'data: {data}', 'max mem: {memory:.0f}' ]) else: log_msg = self.delimiter.join([ header, '[{0' + space_fmt + '}/{1}]', 'eta: {eta}', '{meters}', 'time: {time}', 'data: {data}' ]) MB = 1024.0 * 1024.0 for obj in iterable: data_time.update(time.time() - end) yield obj iter_time.update(time.time() - end) if i % print_freq == 0 or i == len(iterable) - 1: eta_seconds = iter_time.global_avg * (len(iterable) - i) eta_string = str(datetime.timedelta(seconds=int(eta_seconds))) if torch.cuda.is_available(): print(log_msg.format( i, len(iterable), eta=eta_string, meters=str(self), time=str(iter_time), data=str(data_time), memory=torch.cuda.max_memory_allocated() / MB)) else: print(log_msg.format( i, len(iterable), eta=eta_string, meters=str(self), time=str(iter_time), data=str(data_time))) i += 1 end = time.time() total_time = time.time() - start_time total_time_str = str(datetime.timedelta(seconds=int(total_time))) print('{} Total time: {} ({:.4f} s / it)'.format( header, total_time_str, total_time / len(iterable))) def get_sha(): cwd = os.path.dirname(os.path.abspath(__file__)) def _run(command): return subprocess.check_output(command, cwd=cwd).decode('ascii').strip() sha = 'N/A' diff = "clean" branch = 'N/A' try: sha = _run(['git', 'rev-parse', 'HEAD']) subprocess.check_output(['git', 'diff'], cwd=cwd) diff = _run(['git', 'diff-index', 'HEAD']) diff = "has uncommited changes" if diff else "clean" branch = _run(['git', 'rev-parse', '--abbrev-ref', 'HEAD']) except Exception: pass message = f"sha: {sha}, status: {diff}, branch: {branch}" return message def collate_fn(batch): batch = list(zip(*batch)) batch[0] = nested_tensor_from_tensor_list(batch[0]) return tuple(batch) def _max_by_axis(the_list): # type: (List[List[int]]) -> List[int] maxes = the_list[0] for sublist in the_list[1:]: for index, item in enumerate(sublist): maxes[index] = max(maxes[index], item) return maxes class NestedTensor(object): def __init__(self, tensors, mask: Optional[Tensor]): self.tensors = tensors self.mask = mask def to(self, device): # type: (Device) -> NestedTensor # noqa cast_tensor = self.tensors.to(device) mask = self.mask if mask is not None: assert mask is not None cast_mask = mask.to(device) else: cast_mask = None return NestedTensor(cast_tensor, cast_mask) def decompose(self): return self.tensors, self.mask def __repr__(self): return str(self.tensors) def nested_tensor_from_tensor_list(tensor_list: List[Tensor]): # TODO make this more general if tensor_list[0].ndim == 3: if torchvision._is_tracing(): # nested_tensor_from_tensor_list() does not export well to ONNX # call _onnx_nested_tensor_from_tensor_list() instead return _onnx_nested_tensor_from_tensor_list(tensor_list) # TODO make it support different-sized images max_size = _max_by_axis([list(img.shape) for img in tensor_list]) # min_size = tuple(min(s) for s in zip(*[img.shape for img in tensor_list])) batch_shape = [len(tensor_list)] + max_size b, c, h, w = batch_shape dtype = tensor_list[0].dtype device = tensor_list[0].device tensor = torch.zeros(batch_shape, dtype=dtype, device=device) mask = torch.ones((b, h, w), dtype=torch.bool, device=device) for img, pad_img, m in zip(tensor_list, tensor, mask): pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img) m[: img.shape[1], :img.shape[2]] = False else: raise ValueError('not supported') return NestedTensor(tensor, mask) # _onnx_nested_tensor_from_tensor_list() is an implementation of # nested_tensor_from_tensor_list() that is supported by ONNX tracing. @torch.jit.unused def _onnx_nested_tensor_from_tensor_list(tensor_list: List[Tensor]) -> NestedTensor: max_size = [] for i in range(tensor_list[0].dim()): max_size_i = torch.max(torch.stack([img.shape[i] for img in tensor_list]).to(torch.float32)).to(torch.int64) max_size.append(max_size_i) max_size = tuple(max_size) # work around for # pad_img[: img.shape[0], : img.shape[1], : img.shape[2]].copy_(img) # m[: img.shape[1], :img.shape[2]] = False # which is not yet supported in onnx padded_imgs = [] padded_masks = [] for img in tensor_list: padding = [(s1 - s2) for s1, s2 in zip(max_size, tuple(img.shape))] padded_img = torch.nn.functional.pad(img, (0, padding[2], 0, padding[1], 0, padding[0])) padded_imgs.append(padded_img) m = torch.zeros_like(img[0], dtype=torch.int, device=img.device) padded_mask = torch.nn.functional.pad(m, (0, padding[2], 0, padding[1]), "constant", 1) padded_masks.append(padded_mask.to(torch.bool)) tensor = torch.stack(padded_imgs) mask = torch.stack(padded_masks) return NestedTensor(tensor, mask=mask) def setup_for_distributed(is_master): """ This function disables printing when not in master process """ import builtins as __builtin__ builtin_print = __builtin__.print def print(*args, **kwargs): force = kwargs.pop('force', False) if is_master or force: builtin_print(*args, **kwargs) __builtin__.print = print def is_dist_avail_and_initialized(): if not dist.is_available(): return False if not dist.is_initialized(): return False return True def get_world_size(): if not is_dist_avail_and_initialized(): return 1 return dist.get_world_size() def get_rank(): if not is_dist_avail_and_initialized(): return 0 return dist.get_rank() def is_main_process(): return get_rank() == 0 def save_on_master(*args, **kwargs): if is_main_process(): torch.save(*args, **kwargs) def init_distributed_mode(args): if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ: args.rank = int(os.environ["RANK"]) args.world_size = int(os.environ['WORLD_SIZE']) args.gpu = int(os.environ['LOCAL_RANK']) elif 'SLURM_PROCID' in os.environ: args.rank = int(os.environ['SLURM_PROCID']) args.gpu = args.rank % torch.cuda.device_count() else: print('Not using distributed mode') args.distributed = False return args.distributed = True torch.cuda.set_device(args.gpu) args.dist_backend = 'nccl' print('| distributed init (rank {}): {}'.format( args.rank, args.dist_url), flush=True) torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url, world_size=args.world_size, rank=args.rank) torch.distributed.barrier() setup_for_distributed(args.rank == 0) @torch.no_grad() def accuracy(output, target, topk=(1,)): """Computes the precision@k for the specified values of k""" if target.numel() == 0: return [torch.zeros([], device=output.device)] maxk = max(topk) batch_size = target.size(0) _, pred = output.topk(maxk, 1, True, True) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) res = [] for k in topk: correct_k = correct[:k].view(-1).float().sum(0) res.append(correct_k.mul_(100.0 / batch_size)) return res def interpolate(input, size=None, scale_factor=None, mode="nearest", align_corners=None): # type: (Tensor, Optional[List[int]], Optional[float], str, Optional[bool]) -> Tensor """ Equivalent to nn.functional.interpolate, but with support for empty batch sizes. This will eventually be supported natively by PyTorch, and this class can go away. """ if float(torchvision.__version__.split(".")[1]) < 7.0: if input.numel() > 0: return torch.nn.functional.interpolate( input, size, scale_factor, mode, align_corners ) output_shape = _output_size(2, input, size, scale_factor) output_shape = list(input.shape[:-2]) + list(output_shape) return _new_empty_tensor(input, output_shape) else: return torchvision.ops.misc.interpolate(input, size, scale_factor, mode, align_corners)
15,304
31.702991
116
py
pmb-nll
pmb-nll-main/src/detr/util/box_ops.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Utilities for bounding box manipulation and GIoU. """ import torch from torchvision.ops.boxes import box_area def box_cxcywh_to_xyxy(x): x_c, y_c, w, h = x.unbind(-1) b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)] return torch.stack(b, dim=-1) def box_xyxy_to_cxcywh(x): x0, y0, x1, y1 = x.unbind(-1) b = [(x0 + x1) / 2, (y0 + y1) / 2, (x1 - x0), (y1 - y0)] return torch.stack(b, dim=-1) # modified from torchvision to also return the union def box_iou(boxes1, boxes2): area1 = box_area(boxes1) area2 = box_area(boxes2) lt = torch.max(boxes1[:, None, :2], boxes2[:, :2]) # [N,M,2] rb = torch.min(boxes1[:, None, 2:], boxes2[:, 2:]) # [N,M,2] wh = (rb - lt).clamp(min=0) # [N,M,2] inter = wh[:, :, 0] * wh[:, :, 1] # [N,M] union = area1[:, None] + area2 - inter iou = inter / union return iou, union def generalized_box_iou(boxes1, boxes2): """ Generalized IoU from https://giou.stanford.edu/ The boxes should be in [x0, y0, x1, y1] format Returns a [N, M] pairwise matrix, where N = len(boxes1) and M = len(boxes2) """ # degenerate boxes gives inf / nan results # so do an early check assert (boxes1[:, 2:] >= boxes1[:, :2]).all() assert (boxes2[:, 2:] >= boxes2[:, :2]).all() iou, union = box_iou(boxes1, boxes2) lt = torch.min(boxes1[:, None, :2], boxes2[:, :2]) rb = torch.max(boxes1[:, None, 2:], boxes2[:, 2:]) wh = (rb - lt).clamp(min=0) # [N,M,2] area = wh[:, :, 0] * wh[:, :, 1] return iou - (area - union) / area def masks_to_boxes(masks): """Compute the bounding boxes around the provided masks The masks should be in format [N, H, W] where N is the number of masks, (H, W) are the spatial dimensions. Returns a [N, 4] tensors, with the boxes in xyxy format """ if masks.numel() == 0: return torch.zeros((0, 4), device=masks.device) h, w = masks.shape[-2:] y = torch.arange(0, h, dtype=torch.float) x = torch.arange(0, w, dtype=torch.float) y, x = torch.meshgrid(y, x) x_mask = (masks * x.unsqueeze(0)) x_max = x_mask.flatten(1).max(-1)[0] x_min = x_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0] y_mask = (masks * y.unsqueeze(0)) y_max = y_mask.flatten(1).max(-1)[0] y_min = y_mask.masked_fill(~(masks.bool()), 1e8).flatten(1).min(-1)[0] return torch.stack([x_min, y_min, x_max, y_max], 1)
2,561
27.786517
110
py
pmb-nll
pmb-nll-main/src/detr/datasets/__init__.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import torch.utils.data import torchvision from .coco import build as build_coco def get_coco_api_from_dataset(dataset): for _ in range(10): # if isinstance(dataset, torchvision.datasets.CocoDetection): # break if isinstance(dataset, torch.utils.data.Subset): dataset = dataset.dataset if isinstance(dataset, torchvision.datasets.CocoDetection): return dataset.coco def build_dataset(image_set, args): if args.dataset_file == 'coco': return build_coco(image_set, args) if args.dataset_file == 'coco_panoptic': # to avoid making panopticapi required for coco from .coco_panoptic import build as build_coco_panoptic return build_coco_panoptic(image_set, args) raise ValueError(f'dataset {args.dataset_file} not supported')
897
33.538462
70
py
pmb-nll
pmb-nll-main/src/detr/datasets/coco_eval.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ COCO evaluator that works in distributed mode. Mostly copy-paste from https://github.com/pytorch/vision/blob/edfd5a7/references/detection/coco_eval.py The difference is that there is less copy-pasting from pycocotools in the end of the file, as python3 can suppress prints with contextlib """ import os import contextlib import copy import numpy as np import torch from pycocotools.cocoeval import COCOeval from pycocotools.coco import COCO import pycocotools.mask as mask_util from util.misc import all_gather class CocoEvaluator(object): def __init__(self, coco_gt, iou_types): assert isinstance(iou_types, (list, tuple)) coco_gt = copy.deepcopy(coco_gt) self.coco_gt = coco_gt self.iou_types = iou_types self.coco_eval = {} for iou_type in iou_types: self.coco_eval[iou_type] = COCOeval(coco_gt, iouType=iou_type) self.img_ids = [] self.eval_imgs = {k: [] for k in iou_types} def update(self, predictions): img_ids = list(np.unique(list(predictions.keys()))) self.img_ids.extend(img_ids) for iou_type in self.iou_types: results = self.prepare(predictions, iou_type) # suppress pycocotools prints with open(os.devnull, 'w') as devnull: with contextlib.redirect_stdout(devnull): coco_dt = COCO.loadRes(self.coco_gt, results) if results else COCO() coco_eval = self.coco_eval[iou_type] coco_eval.cocoDt = coco_dt coco_eval.params.imgIds = list(img_ids) img_ids, eval_imgs = evaluate(coco_eval) self.eval_imgs[iou_type].append(eval_imgs) def synchronize_between_processes(self): for iou_type in self.iou_types: self.eval_imgs[iou_type] = np.concatenate(self.eval_imgs[iou_type], 2) create_common_coco_eval(self.coco_eval[iou_type], self.img_ids, self.eval_imgs[iou_type]) def accumulate(self): for coco_eval in self.coco_eval.values(): coco_eval.accumulate() def summarize(self): for iou_type, coco_eval in self.coco_eval.items(): print("IoU metric: {}".format(iou_type)) coco_eval.summarize() def prepare(self, predictions, iou_type): if iou_type == "bbox": return self.prepare_for_coco_detection(predictions) elif iou_type == "segm": return self.prepare_for_coco_segmentation(predictions) elif iou_type == "keypoints": return self.prepare_for_coco_keypoint(predictions) else: raise ValueError("Unknown iou type {}".format(iou_type)) def prepare_for_coco_detection(self, predictions): coco_results = [] for original_id, prediction in predictions.items(): if len(prediction) == 0: continue boxes = prediction["boxes"] boxes = convert_to_xywh(boxes).tolist() scores = prediction["scores"].tolist() labels = prediction["labels"].tolist() coco_results.extend( [ { "image_id": original_id, "category_id": labels[k], "bbox": box, "score": scores[k], } for k, box in enumerate(boxes) ] ) return coco_results def prepare_for_coco_segmentation(self, predictions): coco_results = [] for original_id, prediction in predictions.items(): if len(prediction) == 0: continue scores = prediction["scores"] labels = prediction["labels"] masks = prediction["masks"] masks = masks > 0.5 scores = prediction["scores"].tolist() labels = prediction["labels"].tolist() rles = [ mask_util.encode(np.array(mask[0, :, :, np.newaxis], dtype=np.uint8, order="F"))[0] for mask in masks ] for rle in rles: rle["counts"] = rle["counts"].decode("utf-8") coco_results.extend( [ { "image_id": original_id, "category_id": labels[k], "segmentation": rle, "score": scores[k], } for k, rle in enumerate(rles) ] ) return coco_results def prepare_for_coco_keypoint(self, predictions): coco_results = [] for original_id, prediction in predictions.items(): if len(prediction) == 0: continue boxes = prediction["boxes"] boxes = convert_to_xywh(boxes).tolist() scores = prediction["scores"].tolist() labels = prediction["labels"].tolist() keypoints = prediction["keypoints"] keypoints = keypoints.flatten(start_dim=1).tolist() coco_results.extend( [ { "image_id": original_id, "category_id": labels[k], 'keypoints': keypoint, "score": scores[k], } for k, keypoint in enumerate(keypoints) ] ) return coco_results def convert_to_xywh(boxes): xmin, ymin, xmax, ymax = boxes.unbind(1) return torch.stack((xmin, ymin, xmax - xmin, ymax - ymin), dim=1) def merge(img_ids, eval_imgs): all_img_ids = all_gather(img_ids) all_eval_imgs = all_gather(eval_imgs) merged_img_ids = [] for p in all_img_ids: merged_img_ids.extend(p) merged_eval_imgs = [] for p in all_eval_imgs: merged_eval_imgs.append(p) merged_img_ids = np.array(merged_img_ids) merged_eval_imgs = np.concatenate(merged_eval_imgs, 2) # keep only unique (and in sorted order) images merged_img_ids, idx = np.unique(merged_img_ids, return_index=True) merged_eval_imgs = merged_eval_imgs[..., idx] return merged_img_ids, merged_eval_imgs def create_common_coco_eval(coco_eval, img_ids, eval_imgs): img_ids, eval_imgs = merge(img_ids, eval_imgs) img_ids = list(img_ids) eval_imgs = list(eval_imgs.flatten()) coco_eval.evalImgs = eval_imgs coco_eval.params.imgIds = img_ids coco_eval._paramsEval = copy.deepcopy(coco_eval.params) ################################################################# # From pycocotools, just removed the prints and fixed # a Python3 bug about unicode not defined ################################################################# def evaluate(self): ''' Run per image evaluation on given images and store results (a list of dict) in self.evalImgs :return: None ''' # tic = time.time() # print('Running per image evaluation...') p = self.params # add backward compatibility if useSegm is specified in params if p.useSegm is not None: p.iouType = 'segm' if p.useSegm == 1 else 'bbox' print('useSegm (deprecated) is not None. Running {} evaluation'.format(p.iouType)) # print('Evaluate annotation type *{}*'.format(p.iouType)) p.imgIds = list(np.unique(p.imgIds)) if p.useCats: p.catIds = list(np.unique(p.catIds)) p.maxDets = sorted(p.maxDets) self.params = p self._prepare() # loop through images, area range, max detection number catIds = p.catIds if p.useCats else [-1] if p.iouType == 'segm' or p.iouType == 'bbox': computeIoU = self.computeIoU elif p.iouType == 'keypoints': computeIoU = self.computeOks self.ious = { (imgId, catId): computeIoU(imgId, catId) for imgId in p.imgIds for catId in catIds} evaluateImg = self.evaluateImg maxDet = p.maxDets[-1] evalImgs = [ evaluateImg(imgId, catId, areaRng, maxDet) for catId in catIds for areaRng in p.areaRng for imgId in p.imgIds ] # this is NOT in the pycocotools code, but could be done outside evalImgs = np.asarray(evalImgs).reshape(len(catIds), len(p.areaRng), len(p.imgIds)) self._paramsEval = copy.deepcopy(self.params) # toc = time.time() # print('DONE (t={:0.2f}s).'.format(toc-tic)) return p.imgIds, evalImgs ################################################################# # end of straight copy from pycocotools, just removing the prints #################################################################
8,735
32.860465
103
py
pmb-nll
pmb-nll-main/src/detr/datasets/coco_panoptic.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved import json from pathlib import Path import numpy as np import torch from PIL import Image from panopticapi.utils import rgb2id from util.box_ops import masks_to_boxes from .coco import make_coco_transforms class CocoPanoptic: def __init__(self, img_folder, ann_folder, ann_file, transforms=None, return_masks=True): with open(ann_file, 'r') as f: self.coco = json.load(f) # sort 'images' field so that they are aligned with 'annotations' # i.e., in alphabetical order self.coco['images'] = sorted(self.coco['images'], key=lambda x: x['id']) # sanity check if "annotations" in self.coco: for img, ann in zip(self.coco['images'], self.coco['annotations']): assert img['file_name'][:-4] == ann['file_name'][:-4] self.img_folder = img_folder self.ann_folder = ann_folder self.ann_file = ann_file self.transforms = transforms self.return_masks = return_masks def __getitem__(self, idx): ann_info = self.coco['annotations'][idx] if "annotations" in self.coco else self.coco['images'][idx] img_path = Path(self.img_folder) / ann_info['file_name'].replace('.png', '.jpg') ann_path = Path(self.ann_folder) / ann_info['file_name'] img = Image.open(img_path).convert('RGB') w, h = img.size if "segments_info" in ann_info: masks = np.asarray(Image.open(ann_path), dtype=np.uint32) masks = rgb2id(masks) ids = np.array([ann['id'] for ann in ann_info['segments_info']]) masks = masks == ids[:, None, None] masks = torch.as_tensor(masks, dtype=torch.uint8) labels = torch.tensor([ann['category_id'] for ann in ann_info['segments_info']], dtype=torch.int64) target = {} target['image_id'] = torch.tensor([ann_info['image_id'] if "image_id" in ann_info else ann_info["id"]]) if self.return_masks: target['masks'] = masks target['labels'] = labels target["boxes"] = masks_to_boxes(masks) target['size'] = torch.as_tensor([int(h), int(w)]) target['orig_size'] = torch.as_tensor([int(h), int(w)]) if "segments_info" in ann_info: for name in ['iscrowd', 'area']: target[name] = torch.tensor([ann[name] for ann in ann_info['segments_info']]) if self.transforms is not None: img, target = self.transforms(img, target) return img, target def __len__(self): return len(self.coco['images']) def get_height_and_width(self, idx): img_info = self.coco['images'][idx] height = img_info['height'] width = img_info['width'] return height, width def build(image_set, args): img_folder_root = Path(args.coco_path) ann_folder_root = Path(args.coco_panoptic_path) assert img_folder_root.exists(), f'provided COCO path {img_folder_root} does not exist' assert ann_folder_root.exists(), f'provided COCO path {ann_folder_root} does not exist' mode = 'panoptic' PATHS = { "train": ("train2017", Path("annotations") / f'{mode}_train2017.json'), "val": ("val2017", Path("annotations") / f'{mode}_val2017.json'), } img_folder, ann_file = PATHS[image_set] img_folder_path = img_folder_root / img_folder ann_folder = ann_folder_root / f'{mode}_{img_folder}' ann_file = ann_folder_root / ann_file dataset = CocoPanoptic(img_folder_path, ann_folder, ann_file, transforms=make_coco_transforms(image_set), return_masks=args.masks) return dataset
3,723
36.24
111
py
pmb-nll
pmb-nll-main/src/detr/datasets/coco.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ COCO dataset which returns image_id for evaluation. Mostly copy-paste from https://github.com/pytorch/vision/blob/13b35ff/references/detection/coco_utils.py """ from pathlib import Path import torch import torch.utils.data import torchvision from pycocotools import mask as coco_mask import datasets.transforms as T class CocoDetection(torchvision.datasets.CocoDetection): def __init__(self, img_folder, ann_file, transforms, return_masks): super(CocoDetection, self).__init__(img_folder, ann_file) self._transforms = transforms self.prepare = ConvertCocoPolysToMask(return_masks) def __getitem__(self, idx): img, target = super(CocoDetection, self).__getitem__(idx) image_id = self.ids[idx] target = {'image_id': image_id, 'annotations': target} img, target = self.prepare(img, target) if self._transforms is not None: img, target = self._transforms(img, target) return img, target def convert_coco_poly_to_mask(segmentations, height, width): masks = [] for polygons in segmentations: rles = coco_mask.frPyObjects(polygons, height, width) mask = coco_mask.decode(rles) if len(mask.shape) < 3: mask = mask[..., None] mask = torch.as_tensor(mask, dtype=torch.uint8) mask = mask.any(dim=2) masks.append(mask) if masks: masks = torch.stack(masks, dim=0) else: masks = torch.zeros((0, height, width), dtype=torch.uint8) return masks class ConvertCocoPolysToMask(object): def __init__(self, return_masks=False): self.return_masks = return_masks def __call__(self, image, target): w, h = image.size image_id = target["image_id"] image_id = torch.tensor([image_id]) anno = target["annotations"] anno = [obj for obj in anno if 'iscrowd' not in obj or obj['iscrowd'] == 0] boxes = [obj["bbox"] for obj in anno] # guard against no boxes via resizing boxes = torch.as_tensor(boxes, dtype=torch.float32).reshape(-1, 4) boxes[:, 2:] += boxes[:, :2] boxes[:, 0::2].clamp_(min=0, max=w) boxes[:, 1::2].clamp_(min=0, max=h) classes = [obj["category_id"] for obj in anno] classes = torch.tensor(classes, dtype=torch.int64) if self.return_masks: segmentations = [obj["segmentation"] for obj in anno] masks = convert_coco_poly_to_mask(segmentations, h, w) keypoints = None if anno and "keypoints" in anno[0]: keypoints = [obj["keypoints"] for obj in anno] keypoints = torch.as_tensor(keypoints, dtype=torch.float32) num_keypoints = keypoints.shape[0] if num_keypoints: keypoints = keypoints.view(num_keypoints, -1, 3) keep = (boxes[:, 3] > boxes[:, 1]) & (boxes[:, 2] > boxes[:, 0]) boxes = boxes[keep] classes = classes[keep] if self.return_masks: masks = masks[keep] if keypoints is not None: keypoints = keypoints[keep] target = {} target["boxes"] = boxes target["labels"] = classes if self.return_masks: target["masks"] = masks target["image_id"] = image_id if keypoints is not None: target["keypoints"] = keypoints # for conversion to coco api area = torch.tensor([obj["area"] for obj in anno]) iscrowd = torch.tensor([obj["iscrowd"] if "iscrowd" in obj else 0 for obj in anno]) target["area"] = area[keep] target["iscrowd"] = iscrowd[keep] target["orig_size"] = torch.as_tensor([int(h), int(w)]) target["size"] = torch.as_tensor([int(h), int(w)]) return image, target def make_coco_transforms(image_set): normalize = T.Compose([ T.ToTensor(), T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) scales = [480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800] if image_set == 'train': return T.Compose([ T.RandomHorizontalFlip(), T.RandomSelect( T.RandomResize(scales, max_size=1333), T.Compose([ T.RandomResize([400, 500, 600]), T.RandomSizeCrop(384, 600), T.RandomResize(scales, max_size=1333), ]) ), normalize, ]) if image_set == 'val': return T.Compose([ T.RandomResize([800], max_size=1333), normalize, ]) raise ValueError(f'unknown {image_set}') def build(image_set, args): root = Path(args.coco_path) assert root.exists(), f'provided COCO path {root} does not exist' mode = 'instances' PATHS = { "train": (root / "train2017", root / "annotations" / f'{mode}_train2017.json'), "val": (root / "val2017", root / "annotations" / f'{mode}_val2017.json'), } img_folder, ann_file = PATHS[image_set] dataset = CocoDetection(img_folder, ann_file, transforms=make_coco_transforms(image_set), return_masks=args.masks) return dataset
5,253
32.044025
118
py
pmb-nll
pmb-nll-main/src/detr/datasets/transforms.py
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved """ Transforms and data augmentation for both image + bbox. """ import random import PIL import torch import torchvision.transforms as T import torchvision.transforms.functional as F from util.box_ops import box_xyxy_to_cxcywh from util.misc import interpolate def crop(image, target, region): cropped_image = F.crop(image, *region) target = target.copy() i, j, h, w = region # should we do something wrt the original size? target["size"] = torch.tensor([h, w]) fields = ["labels", "area", "iscrowd"] if "boxes" in target: boxes = target["boxes"] max_size = torch.as_tensor([w, h], dtype=torch.float32) cropped_boxes = boxes - torch.as_tensor([j, i, j, i]) cropped_boxes = torch.min(cropped_boxes.reshape(-1, 2, 2), max_size) cropped_boxes = cropped_boxes.clamp(min=0) area = (cropped_boxes[:, 1, :] - cropped_boxes[:, 0, :]).prod(dim=1) target["boxes"] = cropped_boxes.reshape(-1, 4) target["area"] = area fields.append("boxes") if "masks" in target: # FIXME should we update the area here if there are no boxes? target['masks'] = target['masks'][:, i:i + h, j:j + w] fields.append("masks") # remove elements for which the boxes or masks that have zero area if "boxes" in target or "masks" in target: # favor boxes selection when defining which elements to keep # this is compatible with previous implementation if "boxes" in target: cropped_boxes = target['boxes'].reshape(-1, 2, 2) keep = torch.all(cropped_boxes[:, 1, :] > cropped_boxes[:, 0, :], dim=1) else: keep = target['masks'].flatten(1).any(1) for field in fields: target[field] = target[field][keep] return cropped_image, target def hflip(image, target): flipped_image = F.hflip(image) w, h = image.size target = target.copy() if "boxes" in target: boxes = target["boxes"] boxes = boxes[:, [2, 1, 0, 3]] * torch.as_tensor([-1, 1, -1, 1]) + torch.as_tensor([w, 0, w, 0]) target["boxes"] = boxes if "masks" in target: target['masks'] = target['masks'].flip(-1) return flipped_image, target def resize(image, target, size, max_size=None): # size can be min_size (scalar) or (w, h) tuple def get_size_with_aspect_ratio(image_size, size, max_size=None): w, h = image_size if max_size is not None: min_original_size = float(min((w, h))) max_original_size = float(max((w, h))) if max_original_size / min_original_size * size > max_size: size = int(round(max_size * min_original_size / max_original_size)) if (w <= h and w == size) or (h <= w and h == size): return (h, w) if w < h: ow = size oh = int(size * h / w) else: oh = size ow = int(size * w / h) return (oh, ow) def get_size(image_size, size, max_size=None): if isinstance(size, (list, tuple)): return size[::-1] else: return get_size_with_aspect_ratio(image_size, size, max_size) size = get_size(image.size, size, max_size) rescaled_image = F.resize(image, size) if target is None: return rescaled_image, None ratios = tuple(float(s) / float(s_orig) for s, s_orig in zip(rescaled_image.size, image.size)) ratio_width, ratio_height = ratios target = target.copy() if "boxes" in target: boxes = target["boxes"] scaled_boxes = boxes * torch.as_tensor([ratio_width, ratio_height, ratio_width, ratio_height]) target["boxes"] = scaled_boxes if "area" in target: area = target["area"] scaled_area = area * (ratio_width * ratio_height) target["area"] = scaled_area h, w = size target["size"] = torch.tensor([h, w]) if "masks" in target: target['masks'] = interpolate( target['masks'][:, None].float(), size, mode="nearest")[:, 0] > 0.5 return rescaled_image, target def pad(image, target, padding): # assumes that we only pad on the bottom right corners padded_image = F.pad(image, (0, 0, padding[0], padding[1])) if target is None: return padded_image, None target = target.copy() # should we do something wrt the original size? target["size"] = torch.tensor(padded_image.size[::-1]) if "masks" in target: target['masks'] = torch.nn.functional.pad(target['masks'], (0, padding[0], 0, padding[1])) return padded_image, target class RandomCrop(object): def __init__(self, size): self.size = size def __call__(self, img, target): region = T.RandomCrop.get_params(img, self.size) return crop(img, target, region) class RandomSizeCrop(object): def __init__(self, min_size: int, max_size: int): self.min_size = min_size self.max_size = max_size def __call__(self, img: PIL.Image.Image, target: dict): w = random.randint(self.min_size, min(img.width, self.max_size)) h = random.randint(self.min_size, min(img.height, self.max_size)) region = T.RandomCrop.get_params(img, [h, w]) return crop(img, target, region) class CenterCrop(object): def __init__(self, size): self.size = size def __call__(self, img, target): image_width, image_height = img.size crop_height, crop_width = self.size crop_top = int(round((image_height - crop_height) / 2.)) crop_left = int(round((image_width - crop_width) / 2.)) return crop(img, target, (crop_top, crop_left, crop_height, crop_width)) class RandomHorizontalFlip(object): def __init__(self, p=0.5): self.p = p def __call__(self, img, target): if random.random() < self.p: return hflip(img, target) return img, target class RandomResize(object): def __init__(self, sizes, max_size=None): assert isinstance(sizes, (list, tuple)) self.sizes = sizes self.max_size = max_size def __call__(self, img, target=None): size = random.choice(self.sizes) return resize(img, target, size, self.max_size) class RandomPad(object): def __init__(self, max_pad): self.max_pad = max_pad def __call__(self, img, target): pad_x = random.randint(0, self.max_pad) pad_y = random.randint(0, self.max_pad) return pad(img, target, (pad_x, pad_y)) class RandomSelect(object): """ Randomly selects between transforms1 and transforms2, with probability p for transforms1 and (1 - p) for transforms2 """ def __init__(self, transforms1, transforms2, p=0.5): self.transforms1 = transforms1 self.transforms2 = transforms2 self.p = p def __call__(self, img, target): if random.random() < self.p: return self.transforms1(img, target) return self.transforms2(img, target) class ToTensor(object): def __call__(self, img, target): return F.to_tensor(img), target class RandomErasing(object): def __init__(self, *args, **kwargs): self.eraser = T.RandomErasing(*args, **kwargs) def __call__(self, img, target): return self.eraser(img), target class Normalize(object): def __init__(self, mean, std): self.mean = mean self.std = std def __call__(self, image, target=None): image = F.normalize(image, mean=self.mean, std=self.std) if target is None: return image, None target = target.copy() h, w = image.shape[-2:] if "boxes" in target: boxes = target["boxes"] boxes = box_xyxy_to_cxcywh(boxes) boxes = boxes / torch.tensor([w, h, w, h], dtype=torch.float32) target["boxes"] = boxes return image, target class Compose(object): def __init__(self, transforms): self.transforms = transforms def __call__(self, image, target): for t in self.transforms: image, target = t(image, target) return image, target def __repr__(self): format_string = self.__class__.__name__ + "(" for t in self.transforms: format_string += "\n" format_string += " {0}".format(t) format_string += "\n)" return format_string
8,524
29.776173
104
py
pmb-nll
pmb-nll-main/src/offline_evaluation/compute_probabilistic_metrics.py
import json import os import pickle from collections import defaultdict import numpy as np import torch import torch.distributions as distributions import tqdm # Project imports from core.evaluation_tools import evaluation_utils, scoring_rules from core.evaluation_tools.evaluation_utils import ( calculate_iou, get_test_thing_dataset_id_to_train_contiguous_id_dict, ) from core.setup import setup_arg_parser, setup_config from detectron2.checkpoint import DetectionCheckpointer # Detectron imports from detectron2.data import MetadataCatalog from detectron2.engine import launch from detectron2.modeling import build_model from matplotlib import image from matplotlib import pyplot as plt from matplotlib.pyplot import hist from prettytable import PrettyTable from probabilistic_inference.inference_utils import get_inference_output_dir from probabilistic_modeling.losses import ( compute_negative_log_likelihood, negative_log_likelihood, ) from probabilistic_modeling.modeling_utils import ( PoissonPointProcessGMM, PoissonPointProcessIntensityFunction, PoissonPointProcessUniform, PoissonPointUnion, ) from scipy.spatial.distance import mahalanobis device = torch.device("cuda" if torch.cuda.is_available() else "cpu") AREA_LIMITS = {"small": [0, 1024], "medium": [1024, 9216], "large": [9216, np.inf]} def try_squeeze(to_squeeze, dim): return to_squeeze.squeeze(dim) if len(to_squeeze.shape) > dim else to_squeeze def print_nll_results_by_size( out, gt_boxes, inference_output_dir, area_limits=AREA_LIMITS, prefix="" ): title_dict = { "matched_bernoulli_clss": "Matched Bernoulli Classification", "matched_bernoulli_cls": "Matched Bernoulli Classification", "matched_bernoulli_reg": "Matched Bernoulli Regression", "matched_bernoulli_regs": "Matched Bernoulli Regression", "matched_bernoulli": "Matched Bernoulli", "matched_bernoullis": "Matched Bernoulli", "matched_ppp": "Matched PPP", "matched_ppps": "Matched PPP", } def plot_histogram( size_decomp, decomp_key, area_limits, filepath, max_limit=40, nbins=100 ): plt.clf() for size in size_decomp.keys(): hist( np.clip(size_decomp[size][decomp_key], 0, max_limit), nbins, alpha=0.33, label=size, ec=(0, 0, 0, 0), lw=0.0, ) plt.title(title_dict[decomp_key]) plt.legend() plt.xlim(0, max_limit) plt.savefig( os.path.join(filepath, f"{prefix}{decomp_key}.svg"), format="svg", transparent=True, ) size_decomp = {size: defaultdict(list) for size in area_limits.keys()} for img_id, out_dict in out.items(): boxes = gt_boxes[img_id].reshape(-1, 4) decomp = out_dict["decomposition"] # Remove unmatched detections and sort in gt-order instead association = np.array(out_dict["associations"][0]) if not len(association): continue association = association[association[:, 1] > -1] areas = (boxes[:, 2] - boxes[:, 0]) * (boxes[:, 3] - boxes[:, 1]) num_gts = len(areas) num_preds = ( decomp["num_unmatched_bernoulli"][0] + decomp["num_matched_bernoulli"][0] ) ppp_association = association[association[:, 0] >= num_preds] for size, limit in area_limits.items(): mask = torch.logical_and(limit[0] < areas, limit[1] > areas) gt_idx = mask.nonzero() matched_bernoulli_regs = [ comp for assoc, comp in zip(association, decomp["matched_bernoulli_regs"][0]) if assoc[1] in gt_idx ] size_decomp[size]["matched_bernoulli_regs"] += matched_bernoulli_regs size_decomp[size]["matched_bernoulli_reg"] += [sum(matched_bernoulli_regs)] matched_bernoulli_clss = [ comp for assoc, comp in zip(association, decomp["matched_bernoulli_clss"][0]) if assoc[1] in gt_idx ] size_decomp[size]["matched_bernoulli_clss"] += matched_bernoulli_clss size_decomp[size]["matched_bernoulli_cls"] += [sum(matched_bernoulli_clss)] size_decomp[size]["matched_bernoullis"] += [ cls_part + reg_part for cls_part, reg_part in zip( matched_bernoulli_clss, matched_bernoulli_regs ) ] size_decomp[size]["matched_bernoulli"] += [ sum(matched_bernoulli_regs) + sum(matched_bernoulli_clss) ] matched_ppps = [ comp for assoc, comp in zip(ppp_association, decomp["matched_ppps"][0]) if assoc[1] in gt_idx ] size_decomp[size]["matched_ppps"] += matched_ppps size_decomp[size]["matched_ppp"] += [sum(matched_ppps)] for size, limit in area_limits.items(): print(f"******** Size: {size} ********") print( f"Mean matched Bernoulli: {np.mean(size_decomp[size]['matched_bernoulli']):.2f}/", end="", ) print(f"{np.mean(size_decomp[size]['matched_bernoullis']):.2f}") print( f"Mean matched Bernoulli reg: {np.mean(size_decomp[size]['matched_bernoulli_reg']):.2f}/", end="", ) print(f"{np.mean(size_decomp[size]['matched_bernoulli_regs']):.2f}") print( f"Mean matched Bernoulli cls: {np.mean(size_decomp[size]['matched_bernoulli_cls']):.2f}/", end="", ) print(f"{np.mean(size_decomp[size]['matched_bernoulli_clss']):.2f}") print( f"Mean matched PPP: {np.mean(size_decomp[size]['matched_ppp']):.2f}/", end="", ) print(f"{np.mean(size_decomp[size]['matched_ppps']):.2f}") print(f"**************************") for decomp_key in size_decomp[list(area_limits.keys())[0]]: plot_histogram(size_decomp, decomp_key, area_limits, inference_output_dir) def print_nll_results(out): nlls = torch.tensor([el["nll"] for el in out.values() if el["nll"] > 0]) print("*" * 40) print("*" * 12 + "PMB NLL results" + "*" * 13) print("*" * 40) print(f"Min NLL: {nlls.min().item()}") print(f"Mean NLL: {nlls.mean().item()}") print(f"Median NLL: {nlls.median().item()}") print(f"Max NLL: {nlls.max().item()}") print(f"Binned NLL: {torch.histc(nlls, bins=20).tolist()}") print("*" * 40) matched_bernoulli = [] matched_bernoulli_reg = [] matched_bernoulli_cls = [] num_matched_bernoulli = [] unmatched_bernoulli = [] num_unmatched_bernoulli = [] matched_ppp = [] num_matched_ppp = [] ppp_integral = [] for img_id, out_dict in out.items(): decomp = out_dict["decomposition"] matched_bernoulli.append(decomp["matched_bernoulli"][0]) matched_bernoulli_reg.append(decomp["matched_bernoulli_reg"][0]) matched_bernoulli_cls.append(decomp["matched_bernoulli_cls"][0]) num_matched_bernoulli.append(decomp["num_matched_bernoulli"][0]) unmatched_bernoulli.append(decomp["unmatched_bernoulli"][0]) num_unmatched_bernoulli.append(decomp["num_unmatched_bernoulli"][0]) matched_ppp.append(decomp["matched_ppp"][0]) num_matched_ppp.append(decomp["num_matched_ppp"][0]) ppp_integral.append(decomp["ppp_integral"]) matched_bernoulli = np.array(matched_bernoulli) matched_bernoulli_reg = np.array(matched_bernoulli_reg) matched_bernoulli_cls = np.array(matched_bernoulli_cls) num_matched_bernoulli = np.array(num_matched_bernoulli) unmatched_bernoulli = np.array(unmatched_bernoulli) num_unmatched_bernoulli = np.array(num_unmatched_bernoulli) matched_ppp = np.array(matched_ppp) num_matched_ppp = np.array(num_matched_ppp) num_matched_ppp = num_matched_ppp[matched_ppp < np.inf] matched_ppp = matched_ppp[matched_ppp < np.inf] matched_bernoulli_norm = matched_bernoulli.sum() / (num_matched_bernoulli.sum()) matched_bernoulli_reg_norm = matched_bernoulli_reg.sum() / ( num_matched_bernoulli.sum() ) matched_bernoulli_cls_norm = matched_bernoulli_cls.sum() / ( num_matched_bernoulli.sum() ) print(f"Mean matched Bernoulli: {np.mean(matched_bernoulli):.2f}/", end="") print(f"{matched_bernoulli_norm:.2f}") print(f"Mean matched Bernoulli reg: {np.mean(matched_bernoulli_reg):.2f}/", end="") print(f"{matched_bernoulli_reg_norm:.2f}") print(f"Mean matched Bernoulli cls: {np.mean(matched_bernoulli_cls):.2f}/", end="") print(f"{matched_bernoulli_cls_norm:.2f}") unmatched_bernoulli_norm = unmatched_bernoulli.sum() / ( num_unmatched_bernoulli.sum() ) print(f"Mean unmatched Bernoulli: {np.mean(unmatched_bernoulli):.2f}/", end="") print(f"{unmatched_bernoulli_norm:.2f}") matched_ppp_norm = matched_ppp.sum() / num_matched_ppp.sum() print(f"Mean matched PPP: {np.mean(matched_ppp):.2f}/", end="") print(f"{matched_ppp_norm:.2f}") print(f"Mean PPP integral: {np.mean(ppp_integral):.2f}") print("*" * 40) def plot_nll_results(out, inference_output_dir, prefix=""): matched_bernoulli = [] matched_bernoulli_reg = [] matched_bernoulli_cls = [] num_matched_bernoulli = [] unmatched_bernoulli = [] num_unmatched_bernoulli = [] matched_ppp = [] num_matched_ppp = [] ppp_integral = [] for img_id, out_dict in out.items(): decomp = out_dict["decomposition"] matched_bernoulli += [ reg + classification for reg, classification in zip( decomp["matched_bernoulli_regs"][0], decomp["matched_bernoulli_clss"][0], ) ] matched_bernoulli_reg += decomp["matched_bernoulli_regs"][0] matched_bernoulli_cls += decomp["matched_bernoulli_clss"][0] num_matched_bernoulli.append(decomp["num_matched_bernoulli"][0]) unmatched_bernoulli += decomp["unmatched_bernoullis"][0] num_unmatched_bernoulli.append(decomp["num_unmatched_bernoulli"][0]) matched_ppp += decomp["matched_ppps"][0] num_matched_ppp.append(decomp["num_matched_ppp"][0]) ppp_integral.append(decomp["ppp_integral"]) plt.figure() plt.hist(np.clip(matched_bernoulli, 0, 40), 100, ec=(0, 0, 0, 0), lw=0.0) plt.xlim(0, 40) plt.title("Matched Bernoulli") plt.savefig( os.path.join(inference_output_dir, f"{prefix}matched_bernoulli_histogram.svg"), format="svg", transparent=True, ) plt.clf() plt.hist(np.clip(matched_bernoulli_reg, 0, 40), 100, ec=(0, 0, 0, 0), lw=0.0) plt.xlim(0, 40) plt.title("Matched Bernoulli regression") plt.savefig( os.path.join( inference_output_dir, f"{prefix}matched_bernoulli_reg_histogram.svg" ), format="svg", transparent=True, ) plt.clf() plt.hist(np.clip(matched_bernoulli_cls, 0, 5), 100, ec=(0, 0, 0, 0), lw=0.0) plt.xlim(0, 5) plt.title("Matched Bernoulli Classification") plt.savefig( os.path.join( inference_output_dir, f"{prefix}matched_bernoulli_cls_histogram.svg" ), format="svg", transparent=True, ) plt.clf() plt.hist(np.clip(unmatched_bernoulli, 0, 10), 100, ec=(0, 0, 0, 0), lw=0.0) plt.xlim(0, 10) plt.title("Unmatched Bernoulli") plt.savefig( os.path.join( inference_output_dir, f"{prefix}unmatched_bernoulli_histogram.svg" ), format="svg", transparent=True, ) plt.clf() plt.hist(np.clip(matched_ppp, 0, 40), 100, ec=(0, 0, 0, 0), lw=0.0) plt.xlim(0, 40) plt.title("Matched PPP") plt.savefig( os.path.join(inference_output_dir, f"{prefix}matched_ppp_histogram.svg"), format="svg", transparent=True, ) def compute_pmb_nll( cfg, inference_output_dir, cat_mapping_dict, min_allowed_score=0.0, print_results=True, plot_results=True, print_by_size=True, load_nll_results=True, ): results_file = os.path.join( inference_output_dir, f"nll_results_minallowedscore_{min_allowed_score}.pkl" ) if load_nll_results and os.path.isfile(results_file): with open(results_file, "rb") as f: out = pickle.load(f) if print_results: print_nll_results(out) if plot_results: plot_nll_results(out, inference_output_dir) if print_by_size: ( preprocessed_predicted_instances, preprocessed_gt_instances, ) = evaluation_utils.get_per_frame_preprocessed_instances( cfg, inference_output_dir, min_allowed_score ) gt_boxes = preprocessed_gt_instances["gt_boxes"] print_nll_results_by_size(out, gt_boxes, inference_output_dir) return out with torch.no_grad(): # Load predictions and GT ( preprocessed_predicted_instances, preprocessed_gt_instances, ) = evaluation_utils.get_per_frame_preprocessed_instances( cfg, inference_output_dir, min_allowed_score ) predicted_box_means = preprocessed_predicted_instances["predicted_boxes"] predicted_cls_probs = preprocessed_predicted_instances["predicted_cls_probs"] predicted_box_covariances = preprocessed_predicted_instances[ "predicted_covar_mats" ] if "ppp_weights" in preprocessed_predicted_instances: predicted_ppp = preprocessed_predicted_instances["ppp_weights"] elif "log_ppp_intensity" in preprocessed_predicted_instances: predicted_ppp = preprocessed_predicted_instances["log_ppp_intensity"] else: predicted_ppp = defaultdict(list) if cfg.PROBABILISTIC_INFERENCE.LOAD_PPP_FROM_MODEL: model = build_model(cfg) DetectionCheckpointer(model, save_dir=cfg.OUTPUT_DIR).resume_or_load( cfg.MODEL.WEIGHTS, resume=True ) ppp = model.get_ppp_intensity_function() ppp.set_normalization_of_bboxes(True) ppp.update_distribution() predicted_ppp = defaultdict(int) image_sizes = preprocessed_predicted_instances["image_size"] gt_box_means = preprocessed_gt_instances["gt_boxes"] gt_cat_idxs = preprocessed_gt_instances["gt_cat_idxs"] # Initialize results out = defaultdict(dict) print("[NLLOD] Started evaluating NLL for dataset.") with tqdm.tqdm(total=len(predicted_box_means)) as pbar: for image_id in predicted_box_means: ppp_mix = PoissonPointUnion() pbar.update(1) image_size = image_sizes[image_id] ################ GT STUFF ########################### gt_boxes = gt_box_means[image_id] if len(gt_boxes.shape) < 2: gt_boxes = gt_boxes.view(-1, 4) gt_classes = ( torch.as_tensor( [ cat_mapping_dict[cat_id.item()] for cat_id in gt_cat_idxs[image_id].long().view(-1, 1) ] ) .long() .to(device) ) ################# PREDICTION STUFF #################### pred_cls_probs = predicted_cls_probs[image_id].clamp(1e-6, 1 - 1e-6) if cfg.MODEL.META_ARCHITECTURE == "ProbabilisticRetinaNet": num_classes = pred_cls_probs.shape[-1] scores_have_bg_cls = False else: num_classes = pred_cls_probs.shape[-1] - 1 scores_have_bg_cls = True pred_box_means = ( predicted_box_means[image_id].unsqueeze(1).repeat(1, num_classes, 1) ) pred_box_covs = predicted_box_covariances[image_id] pred_box_covs = pred_box_covs.unsqueeze(1).repeat(1, num_classes, 1, 1) pred_ppp_weights = predicted_ppp[image_id] if not cfg.PROBABILISTIC_INFERENCE.TREAT_AS_MB: if cfg.PROBABILISTIC_INFERENCE.PPP_CONFIDENCE_THRES > 0: if scores_have_bg_cls: max_conf = 1 - pred_cls_probs[..., -1] else: max_conf = pred_cls_probs[..., :num_classes].max(dim=1)[0] ppp_preds_idx = ( max_conf <= cfg.PROBABILISTIC_INFERENCE.PPP_CONFIDENCE_THRES ) if not ppp_preds_idx.any(): ppp_preds = PoissonPointProcessIntensityFunction( cfg, log_intensity=-np.inf, device=gt_boxes.device ) else: mixture_dict = {} mixture_dict["weights"] = max_conf[ppp_preds_idx] mixture_dict["means"] = pred_box_means[ppp_preds_idx, 0] mixture_dict["covs"] = pred_box_covs[ppp_preds_idx, 0] mixture_dict["cls_probs"] = pred_cls_probs[ ppp_preds_idx, :num_classes ] mixture_dict[ "reg_dist_type" ] = ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE ) if ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE == "gaussian" ): mixture_dict[ "reg_dist" ] = distributions.multivariate_normal.MultivariateNormal mixture_dict["reg_kwargs"] = { "covariance_matrix": mixture_dict["covs"] } elif ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE == "laplacian" ): mixture_dict["reg_dist"] = distributions.laplace.Laplace mixture_dict["reg_kwargs"] = { "scale": torch.sqrt( mixture_dict["covs"].diagonal(dim1=-2, dim2=-1) / 2 ) } ppp_preds = PoissonPointProcessIntensityFunction( cfg, predictions=mixture_dict ) pred_box_means = pred_box_means[ppp_preds_idx.logical_not()] pred_box_covs = pred_box_covs[ppp_preds_idx.logical_not()] pred_cls_probs = pred_cls_probs[ppp_preds_idx.logical_not()] ppp_mix.add_ppp(ppp_preds) if cfg.PROBABILISTIC_INFERENCE.LOAD_PPP_FROM_MODEL: ppp = ppp elif isinstance(pred_ppp_weights, dict): ppp = PoissonPointProcessIntensityFunction( cfg, device=gt_boxes.device ) ppp.load_weights(pred_ppp_weights) elif isinstance(pred_ppp_weights, torch.Tensor): ppp = PoissonPointProcessIntensityFunction( cfg, log_intensity=pred_ppp_weights, device=gt_boxes.device ) else: print( "[NLLOD] PPP intensity function not found in annotations, using config" ) pred_ppp_weights = -np.inf ppp = PoissonPointProcessIntensityFunction( cfg, log_intensity=pred_ppp_weights, device=gt_boxes.device ) else: pred_ppp_weights = -np.inf ppp = PoissonPointProcessIntensityFunction( cfg, log_intensity=pred_ppp_weights ) ppp_mix.add_ppp(ppp) if ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE == "gaussian" ): reg_distribution = lambda x, y: distributions.multivariate_normal.MultivariateNormal( x, y ) elif ( cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE == "laplacian" ): reg_distribution = lambda x, y: distributions.laplace.Laplace( loc=x, scale=torch.sqrt(y.diagonal(dim1=-2, dim2=-1) / 2) ) else: raise Exception( f"Bounding box uncertainty distribution {cfg.MODEL.PROBABILISTIC_MODELING.BBOX_COV_LOSS.DISTRIBUTION_TYPE} is not available." ) try: nll, associations, decompositions = negative_log_likelihood( pred_box_scores=[pred_cls_probs], pred_box_regs=[pred_box_means], pred_box_covars=[pred_box_covs], gt_boxes=[gt_boxes], gt_classes=[gt_classes], image_sizes=[image_size], reg_distribution=reg_distribution, intensity_func=ppp_mix, max_n_solutions=cfg.MODEL.PROBABILISTIC_MODELING.NLL_MAX_NUM_SOLUTIONS, training=False, scores_have_bg_cls=scores_have_bg_cls, ) out[image_id] = { "nll": nll.item(), "associations": associations[0].tolist(), "decomposition": decompositions[0], } except Exception as e: print( f"Image {image_id} raised error. Will not be used to calculate NLL." ) print(e) with open( os.path.join( inference_output_dir, f"nll_results_minallowedscore_{min_allowed_score}.pkl", ), "wb", ) as f: pickle.dump(out, f) if print_results: print_nll_results(out) if plot_results: plot_nll_results(out, inference_output_dir) if print_by_size: gt_boxes = preprocessed_gt_instances["gt_boxes"] print_nll_results_by_size(out, gt_boxes, inference_output_dir) return out def main( args, cfg=None, iou_min=None, iou_correct=None, min_allowed_score=None, print_results=True, inference_output_dir="", image_ids=[], ): # Setup config if cfg is None: cfg = setup_config(args, random_seed=args.random_seed, is_testing=True) cfg.defrost() cfg.ACTUAL_TEST_DATASET = args.test_dataset # Setup torch device and num_threads torch.set_num_threads(cfg.DATALOADER.NUM_WORKERS) # Build path to gt instances and inference output if inference_output_dir == "": inference_output_dir = get_inference_output_dir( cfg["OUTPUT_DIR"], args.test_dataset, args.inference_config, args.image_corruption_level, ) # Get thresholds to perform evaluation on if iou_min is None: iou_min = args.iou_min if iou_correct is None: iou_correct = args.iou_correct if min_allowed_score is None or min_allowed_score < 0: # Check if F-1 Score has been previously computed ON THE ORIGINAL # DATASET such as COCO even when evaluating on OpenImages. try: with open(os.path.join(inference_output_dir, "mAP_res.txt"), "r") as f: min_allowed_score = f.read().strip("][\n").split(", ")[-1] min_allowed_score = round(float(min_allowed_score), 4) except FileNotFoundError: # If not, process all detections. Not recommended as the results might be influenced by very low scoring # detections that would normally be removed in robotics/vision # applications. min_allowed_score = 0.0 # Get category mapping dictionary: train_thing_dataset_id_to_contiguous_id = MetadataCatalog.get( cfg.DATASETS.TRAIN[0] ).thing_dataset_id_to_contiguous_id test_thing_dataset_id_to_contiguous_id = MetadataCatalog.get( args.test_dataset ).thing_dataset_id_to_contiguous_id cat_mapping_dict = get_test_thing_dataset_id_to_train_contiguous_id_dict( cfg, args, train_thing_dataset_id_to_contiguous_id, test_thing_dataset_id_to_contiguous_id, ) # Compute NLL results load_nll_results = len(image_ids) == 0 nll_results = compute_pmb_nll( cfg, inference_output_dir, cat_mapping_dict, min_allowed_score, print_results, load_nll_results=load_nll_results ) # Get matched results by either generating them or loading from file. with torch.no_grad(): matched_results = evaluation_utils.get_matched_results( cfg, inference_output_dir, iou_min=iou_min, iou_correct=iou_correct, min_allowed_score=min_allowed_score, ) # Build preliminary dicts required for computing classification scores. for matched_results_key in matched_results.keys(): if "gt_cat_idxs" in matched_results[matched_results_key].keys(): # First we convert the written things indices to contiguous # indices. gt_converted_cat_idxs = matched_results[matched_results_key][ "gt_cat_idxs" ] gt_converted_cat_idxs = try_squeeze(gt_converted_cat_idxs, 1) gt_converted_cat_idxs = torch.as_tensor( [ cat_mapping_dict[class_idx.cpu().tolist()] for class_idx in gt_converted_cat_idxs ] ).to(device) matched_results[matched_results_key][ "gt_converted_cat_idxs" ] = gt_converted_cat_idxs.to(device) if "predicted_cls_probs" in matched_results[matched_results_key].keys(): predicted_cls_probs = matched_results[matched_results_key][ "predicted_cls_probs" ] # This is required for evaluation of retinanet based # detections. matched_results[matched_results_key][ "predicted_score_of_gt_category" ] = torch.gather( predicted_cls_probs, 1, gt_converted_cat_idxs.unsqueeze(1) ).squeeze( 1 ) matched_results[matched_results_key][ "gt_cat_idxs" ] = gt_converted_cat_idxs else: if cfg.MODEL.META_ARCHITECTURE == "ProbabilisticRetinaNet": # For false positives, the correct category is background. For retinanet, since no explicit # background category is available, this value is computed as 1.0 - score of the predicted # category. predicted_class_probs, predicted_class_idx = matched_results[ matched_results_key ]["predicted_cls_probs"].max(1) matched_results[matched_results_key][ "predicted_score_of_gt_category" ] = (1.0 - predicted_class_probs) matched_results[matched_results_key][ "predicted_cat_idxs" ] = predicted_class_idx else: # For RCNN/DETR based networks, a background category is # explicitly available. matched_results[matched_results_key][ "predicted_score_of_gt_category" ] = matched_results[matched_results_key]["predicted_cls_probs"][ :, -1 ] _, predicted_class_idx = matched_results[matched_results_key][ "predicted_cls_probs" ][:, :-1].max(1) matched_results[matched_results_key][ "predicted_cat_idxs" ] = predicted_class_idx # Load the different detection partitions true_positives = matched_results["true_positives"] duplicates = matched_results["duplicates"] localization_errors = matched_results["localization_errors"] false_negatives = matched_results["false_negatives"] false_positives = matched_results["false_positives"] # Get the number of elements in each partition num_true_positives = true_positives["predicted_box_means"].shape[0] num_duplicates = duplicates["predicted_box_means"].shape[0] num_localization_errors = localization_errors["predicted_box_means"].shape[0] num_false_negatives = false_negatives["gt_box_means"].shape[0] num_false_positives = false_positives["predicted_box_means"].shape[0] per_class_output_list = [] for class_idx in cat_mapping_dict.values(): true_positives_valid_idxs = ( true_positives["gt_converted_cat_idxs"] == class_idx ) localization_errors_valid_idxs = ( localization_errors["gt_converted_cat_idxs"] == class_idx ) duplicates_valid_idxs = duplicates["gt_converted_cat_idxs"] == class_idx false_positives_valid_idxs = ( false_positives["predicted_cat_idxs"] == class_idx ) if cfg.MODEL.META_ARCHITECTURE == "ProbabilisticRetinaNet": # Compute classification metrics for every partition true_positives_cls_analysis = scoring_rules.sigmoid_compute_cls_scores( true_positives, true_positives_valid_idxs ) localization_errors_cls_analysis = ( scoring_rules.sigmoid_compute_cls_scores( localization_errors, localization_errors_valid_idxs ) ) duplicates_cls_analysis = scoring_rules.sigmoid_compute_cls_scores( duplicates, duplicates_valid_idxs ) false_positives_cls_analysis = scoring_rules.sigmoid_compute_cls_scores( false_positives, false_positives_valid_idxs ) else: # Compute classification metrics for every partition true_positives_cls_analysis = scoring_rules.softmax_compute_cls_scores( true_positives, true_positives_valid_idxs ) localization_errors_cls_analysis = ( scoring_rules.softmax_compute_cls_scores( localization_errors, localization_errors_valid_idxs ) ) duplicates_cls_analysis = scoring_rules.softmax_compute_cls_scores( duplicates, duplicates_valid_idxs ) false_positives_cls_analysis = scoring_rules.softmax_compute_cls_scores( false_positives, false_positives_valid_idxs ) # Compute regression metrics for every partition true_positives_reg_analysis = scoring_rules.compute_reg_scores( true_positives, true_positives_valid_idxs ) localization_errors_reg_analysis = scoring_rules.compute_reg_scores( localization_errors, localization_errors_valid_idxs ) duplicates_reg_analysis = scoring_rules.compute_reg_scores( duplicates, duplicates_valid_idxs ) false_positives_reg_analysis = scoring_rules.compute_reg_scores_fn( false_positives, false_positives_valid_idxs ) per_class_output_list.append( { "true_positives_cls_analysis": true_positives_cls_analysis, "true_positives_reg_analysis": true_positives_reg_analysis, "localization_errors_cls_analysis": localization_errors_cls_analysis, "localization_errors_reg_analysis": localization_errors_reg_analysis, "duplicates_cls_analysis": duplicates_cls_analysis, "duplicates_reg_analysis": duplicates_reg_analysis, "false_positives_cls_analysis": false_positives_cls_analysis, "false_positives_reg_analysis": false_positives_reg_analysis, } ) final_accumulated_output_dict = dict() final_average_output_dict = dict() for key in per_class_output_list[0].keys(): average_output_dict = dict() for inner_key in per_class_output_list[0][key].keys(): collected_values = [ per_class_output[key][inner_key] if per_class_output[key][inner_key] is not None else np.NaN for per_class_output in per_class_output_list ] collected_values = np.array(collected_values) if key in average_output_dict.keys(): # Use nan mean since some classes do not have duplicates for # instance or has one duplicate for instance. torch.std returns nan in that case # so we handle those here. This should not have any effect on the final results, as # it only affects inter-class variance which we do not # report anyways. average_output_dict[key].update( { inner_key: np.nanmean(collected_values), inner_key + "_std": np.nanstd(collected_values, ddof=1), } ) final_accumulated_output_dict[key].update( {inner_key: collected_values} ) else: average_output_dict.update( { key: { inner_key: np.nanmean(collected_values), inner_key + "_std": np.nanstd(collected_values, ddof=1), } } ) final_accumulated_output_dict.update( {key: {inner_key: collected_values}} ) final_average_output_dict.update(average_output_dict) final_accumulated_output_dict.update( { "num_instances": { "num_true_positives": num_true_positives, "num_duplicates": num_duplicates, "num_localization_errors": num_localization_errors, "num_false_positives": num_false_positives, "num_false_negatives": num_false_negatives, } } ) if print_results: # Summarize and print all table = PrettyTable() table.field_names = [ "Output Type", "Number of Instances", "Cls Negative Log Likelihood", "Cls Brier Score", "Reg TP Negative Log Likelihood / FP Entropy", "Reg Energy Score", ] table.add_row( [ "True Positives:", num_true_positives, "{:.4f} ± {:.4f}".format( final_average_output_dict["true_positives_cls_analysis"][ "ignorance_score_mean" ], final_average_output_dict["true_positives_cls_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["true_positives_cls_analysis"][ "brier_score_mean" ], final_average_output_dict["true_positives_cls_analysis"][ "brier_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["true_positives_reg_analysis"][ "ignorance_score_mean" ], final_average_output_dict["true_positives_reg_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["true_positives_reg_analysis"][ "energy_score_mean" ], final_average_output_dict["true_positives_reg_analysis"][ "energy_score_mean_std" ], ), ] ) table.add_row( [ "Duplicates:", num_duplicates, "{:.4f} ± {:.4f}".format( final_average_output_dict["duplicates_cls_analysis"][ "ignorance_score_mean" ], final_average_output_dict["duplicates_cls_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["duplicates_cls_analysis"][ "brier_score_mean" ], final_average_output_dict["duplicates_cls_analysis"][ "brier_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["duplicates_reg_analysis"][ "ignorance_score_mean" ], final_average_output_dict["duplicates_reg_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["duplicates_reg_analysis"][ "energy_score_mean" ], final_average_output_dict["duplicates_reg_analysis"][ "energy_score_mean_std" ], ), ] ) table.add_row( [ "Localization Errors:", num_localization_errors, "{:.4f} ± {:.4f}".format( final_average_output_dict["localization_errors_cls_analysis"][ "ignorance_score_mean" ], final_average_output_dict["localization_errors_cls_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["localization_errors_cls_analysis"][ "brier_score_mean" ], final_average_output_dict["localization_errors_cls_analysis"][ "brier_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["localization_errors_reg_analysis"][ "ignorance_score_mean" ], final_average_output_dict["localization_errors_reg_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["localization_errors_reg_analysis"][ "energy_score_mean" ], final_average_output_dict["localization_errors_reg_analysis"][ "energy_score_mean_std" ], ), ] ) table.add_row( [ "False Positives:", num_false_positives, "{:.4f} ± {:.4f}".format( final_average_output_dict["false_positives_cls_analysis"][ "ignorance_score_mean" ], final_average_output_dict["false_positives_cls_analysis"][ "ignorance_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["false_positives_cls_analysis"][ "brier_score_mean" ], final_average_output_dict["false_positives_cls_analysis"][ "brier_score_mean_std" ], ), "{:.4f} ± {:.4f}".format( final_average_output_dict["false_positives_reg_analysis"][ "total_entropy_mean" ], final_average_output_dict["false_positives_reg_analysis"][ "total_entropy_mean_std" ], ), "-", ] ) table.add_row(["False Negatives:", num_false_negatives, "-", "-", "-", "-"]) print(table) text_file_name = os.path.join( inference_output_dir, "probabilistic_scoring_res_{}_{}_{}.txt".format( iou_min, iou_correct, min_allowed_score ), ) with open(text_file_name, "w") as text_file: print(table, file=text_file) dictionary_file_name = os.path.join( inference_output_dir, "probabilistic_scoring_res_{}_{}_{}.pkl".format( iou_min, iou_correct, min_allowed_score ), ) with open(dictionary_file_name, "wb") as pickle_file: pickle.dump(final_accumulated_output_dict, pickle_file) if __name__ == "__main__": # Create arg parser arg_parser = setup_arg_parser() args = arg_parser.parse_args() print("Command Line Args:", args) launch( main, args.num_gpus, num_machines=args.num_machines, machine_rank=args.machine_rank, dist_url=args.dist_url, args=(args,), )
44,542
40.785178
149
py
pmb-nll
pmb-nll-main/src/offline_evaluation/compute_ood_probabilistic_metrics.py
import itertools import os import torch import ujson as json import pickle from prettytable import PrettyTable # Detectron imports from detectron2.engine import launch # Project imports from core.evaluation_tools import scoring_rules from core.evaluation_tools.evaluation_utils import eval_predictions_preprocess from core.setup import setup_config, setup_arg_parser from probabilistic_inference.inference_utils import get_inference_output_dir device = torch.device("cuda" if torch.cuda.is_available() else "cpu") def main( args, cfg=None, min_allowed_score=None): # Setup config if cfg is None: cfg = setup_config(args, random_seed=args.random_seed, is_testing=True) cfg.defrost() cfg.ACTUAL_TEST_DATASET = args.test_dataset # Setup torch device and num_threads torch.set_num_threads(cfg.DATALOADER.NUM_WORKERS) # Build path to gt instances and inference output inference_output_dir = get_inference_output_dir( cfg['OUTPUT_DIR'], args.test_dataset, args.inference_config, args.image_corruption_level) if min_allowed_score is None: # Check if F-1 Score has been previously computed ON THE ORIGINAL # DATASET, and not on VOC. try: train_set_inference_output_dir = get_inference_output_dir( cfg['OUTPUT_DIR'], cfg.DATASETS.TEST[0], args.inference_config, 0) with open(os.path.join(train_set_inference_output_dir, "mAP_res.txt"), "r") as f: min_allowed_score = f.read().strip('][\n').split(', ')[-1] min_allowed_score = round(float(min_allowed_score), 4) except FileNotFoundError: # If not, process all detections. Not recommended as the results might be influenced by very low scoring # detections that would normally be removed in robotics/vision # applications. min_allowed_score = 0.0 # Get matched results by either generating them or loading from file. with torch.no_grad(): try: preprocessed_predicted_instances = torch.load( os.path.join( inference_output_dir, "preprocessed_predicted_instances_odd_{}.pth".format(min_allowed_score)), map_location=device) # Process predictions except FileNotFoundError: prediction_file_name = os.path.join( inference_output_dir, 'coco_instances_results.json') predicted_instances = json.load(open(prediction_file_name, 'r')) preprocessed_predicted_instances = eval_predictions_preprocess( predicted_instances, min_allowed_score=min_allowed_score, is_odd=True) torch.save( preprocessed_predicted_instances, os.path.join( inference_output_dir, "preprocessed_predicted_instances_odd_{}.pth".format(min_allowed_score))) predicted_boxes = preprocessed_predicted_instances['predicted_boxes'] predicted_cov_mats = preprocessed_predicted_instances['predicted_covar_mats'] predicted_cls_probs = preprocessed_predicted_instances['predicted_cls_probs'] predicted_boxes = list(itertools.chain.from_iterable( [predicted_boxes[key] for key in predicted_boxes.keys()])) predicted_cov_mats = list(itertools.chain.from_iterable( [predicted_cov_mats[key] for key in predicted_cov_mats.keys()])) predicted_cls_probs = list(itertools.chain.from_iterable( [predicted_cls_probs[key] for key in predicted_cls_probs.keys()])) num_false_positives = len(predicted_boxes) valid_idxs = torch.as_tensor( [i for i in range(num_false_positives)]).to(device) predicted_boxes = torch.stack(predicted_boxes, 1).transpose(0, 1) predicted_cov_mats = torch.stack(predicted_cov_mats, 1).transpose(0, 1) predicted_cls_probs = torch.stack( predicted_cls_probs, 1).transpose( 0, 1) false_positives_dict = { 'predicted_box_means': predicted_boxes, 'predicted_box_covariances': predicted_cov_mats, 'predicted_cls_probs': predicted_cls_probs} false_positives_reg_analysis = scoring_rules.compute_reg_scores_fn( false_positives_dict, valid_idxs) if cfg.MODEL.META_ARCHITECTURE == 'ProbabilisticRetinaNet': predicted_class_probs, predicted_class_idx = predicted_cls_probs.max( 1) false_positives_dict['predicted_score_of_gt_category'] = 1.0 - \ predicted_class_probs false_positives_cls_analysis = scoring_rules.sigmoid_compute_cls_scores( false_positives_dict, valid_idxs) else: false_positives_dict['predicted_score_of_gt_category'] = predicted_cls_probs[:, -1] _, predicted_class_idx = predicted_cls_probs[:, :-1].max( 1) false_positives_cls_analysis = scoring_rules.softmax_compute_cls_scores( false_positives_dict, valid_idxs) # Summarize and print all table = PrettyTable() table.field_names = (['Output Type', 'Number of Instances', 'Cls Ignorance Score', 'Cls Brier/Probability Score', 'Reg Ignorance Score', 'Reg Energy Score']) table.add_row( [ "False Positives:", num_false_positives, '{:.4f}'.format( false_positives_cls_analysis['ignorance_score_mean'],), '{:.4f}'.format( false_positives_cls_analysis['brier_score_mean']), '{:.4f}'.format( false_positives_reg_analysis['total_entropy_mean']), '{:.4f}'.format( false_positives_reg_analysis['fp_energy_score_mean'])]) print(table) text_file_name = os.path.join( inference_output_dir, 'probabilistic_scoring_res_odd_{}.txt'.format(min_allowed_score)) with open(text_file_name, "w") as text_file: print(table, file=text_file) dictionary_file_name = os.path.join( inference_output_dir, 'probabilistic_scoring_res_odd_{}.pkl'.format(min_allowed_score)) false_positives_reg_analysis.update(false_positives_cls_analysis) with open(dictionary_file_name, "wb") as pickle_file: pickle.dump(false_positives_reg_analysis, pickle_file) if __name__ == "__main__": # Create arg parser arg_parser = setup_arg_parser() args = arg_parser.parse_args() print("Command Line Args:", args) launch( main, args.num_gpus, num_machines=args.num_machines, machine_rank=args.machine_rank, dist_url=args.dist_url, args=(args,), )
7,146
38.486188
116
py
pmb-nll
pmb-nll-main/src/offline_evaluation/compute_calibration_errors.py
import calibration as cal import os import pickle import torch from prettytable import PrettyTable # Detectron imports from detectron2.data import MetadataCatalog from detectron2.engine import launch # Project imports from core.evaluation_tools import evaluation_utils from core.evaluation_tools.evaluation_utils import get_test_thing_dataset_id_to_train_contiguous_id_dict from core.setup import setup_config, setup_arg_parser from probabilistic_inference.inference_utils import get_inference_output_dir device = torch.device("cuda" if torch.cuda.is_available() else "cpu") def main( args, cfg=None, iou_min=None, iou_correct=None, min_allowed_score=None, print_results=True, inference_output_dir=""): # Setup config if cfg is None: cfg = setup_config(args, random_seed=args.random_seed, is_testing=True) cfg.defrost() cfg.ACTUAL_TEST_DATASET = args.test_dataset # Setup torch device and num_threads torch.set_num_threads(cfg.DATALOADER.NUM_WORKERS) # Build path to gt instances and inference output if inference_output_dir == "": inference_output_dir = get_inference_output_dir( cfg['OUTPUT_DIR'], args.test_dataset, args.inference_config, args.image_corruption_level) # Get thresholds to perform evaluation on if iou_min is None: iou_min = args.iou_min if iou_correct is None: iou_correct = args.iou_correct if min_allowed_score is None: # Check if F-1 Score has been previously computed ON THE ORIGINAL # DATASET such as COCO even when evaluating on OpenImages. try: train_set_inference_output_dir = get_inference_output_dir( cfg['OUTPUT_DIR'], cfg.DATASETS.TEST[0], args.inference_config, 0) with open(os.path.join(train_set_inference_output_dir, "mAP_res.txt"), "r") as f: min_allowed_score = f.read().strip('][\n').split(', ')[-1] min_allowed_score = round(float(min_allowed_score), 4) except FileNotFoundError: # If not, process all detections. Not recommended as the results might be influenced by very low scoring # detections that would normally be removed in robotics/vision # applications. min_allowed_score = 0.0 # Get category mapping dictionary: train_thing_dataset_id_to_contiguous_id = MetadataCatalog.get( cfg.DATASETS.TRAIN[0]).thing_dataset_id_to_contiguous_id test_thing_dataset_id_to_contiguous_id = MetadataCatalog.get( args.test_dataset).thing_dataset_id_to_contiguous_id cat_mapping_dict = get_test_thing_dataset_id_to_train_contiguous_id_dict( cfg, args, train_thing_dataset_id_to_contiguous_id, test_thing_dataset_id_to_contiguous_id) # Get matched results by either generating them or loading from file. with torch.no_grad(): matched_results = evaluation_utils.get_matched_results( cfg, inference_output_dir, iou_min=iou_min, iou_correct=iou_correct, min_allowed_score=min_allowed_score) # Build preliminary dicts required for computing classification scores. for matched_results_key in matched_results.keys(): if 'gt_cat_idxs' in matched_results[matched_results_key].keys(): # First we convert the written things indices to contiguous # indices. gt_converted_cat_idxs = matched_results[matched_results_key]['gt_cat_idxs'].squeeze( 1) gt_converted_cat_idxs = torch.as_tensor([cat_mapping_dict[class_idx.cpu( ).tolist()] for class_idx in gt_converted_cat_idxs]).to(device) matched_results[matched_results_key]['gt_converted_cat_idxs'] = gt_converted_cat_idxs.to( device) matched_results[matched_results_key]['gt_cat_idxs'] = gt_converted_cat_idxs if 'predicted_cls_probs' in matched_results[matched_results_key].keys( ): if cfg.MODEL.META_ARCHITECTURE == 'ProbabilisticRetinaNet': # For false positives, the correct category is background. For retinanet, since no explicit # background category is available, this value is computed as 1.0 - score of the predicted # category. predicted_class_probs, predicted_cat_idxs = matched_results[matched_results_key][ 'predicted_cls_probs'].max( 1) matched_results[matched_results_key]['output_logits'] = predicted_class_probs else: predicted_class_probs, predicted_cat_idxs = matched_results[ matched_results_key]['predicted_cls_probs'][:, :-1].max(1) matched_results[matched_results_key]['predicted_cat_idxs'] = predicted_cat_idxs # Load the different detection partitions true_positives = matched_results['true_positives'] duplicates = matched_results['duplicates'] localization_errors = matched_results['localization_errors'] false_positives = matched_results['false_positives'] reg_maximum_calibration_error_list = [] reg_expected_calibration_error_list = [] if cfg.MODEL.META_ARCHITECTURE == 'ProbabilisticRetinaNet': all_predicted_scores = torch.cat( (true_positives['predicted_cls_probs'].flatten(), duplicates['predicted_cls_probs'].flatten(), localization_errors['predicted_cls_probs'].flatten(), false_positives['predicted_cls_probs'].flatten()), 0) all_gt_scores = torch.cat( (torch.nn.functional.one_hot( true_positives['gt_cat_idxs'], true_positives['predicted_cls_probs'].shape[1]).flatten().to(device), torch.nn.functional.one_hot( duplicates['gt_cat_idxs'], duplicates['predicted_cls_probs'].shape[1]).flatten().to(device), torch.zeros_like( localization_errors['predicted_cls_probs'].type( torch.LongTensor).flatten()).to(device), torch.zeros_like( false_positives['predicted_cls_probs'].type( torch.LongTensor).flatten()).to(device)), 0) else: # For RCNN based networks, a background category is # explicitly available. all_predicted_scores = torch.cat( (true_positives['predicted_cls_probs'], duplicates['predicted_cls_probs'], localization_errors['predicted_cls_probs'], false_positives['predicted_cls_probs']), 0) all_gt_scores = torch.cat( (true_positives['gt_cat_idxs'], duplicates['gt_cat_idxs'], torch.ones_like( localization_errors['predicted_cls_probs'][:, 0]).fill_(80.0).type( torch.LongTensor).to(device), torch.ones_like( false_positives['predicted_cls_probs'][:, 0]).fill_(80.0).type( torch.LongTensor).to(device)), 0) # Compute classification calibration error using calibration # library cls_marginal_calibration_error = cal.get_calibration_error( all_predicted_scores.cpu().numpy(), all_gt_scores.cpu().numpy()) for class_idx in cat_mapping_dict.values(): true_positives_valid_idxs = true_positives['gt_converted_cat_idxs'] == class_idx localization_errors_valid_idxs = localization_errors['gt_converted_cat_idxs'] == class_idx duplicates_valid_idxs = duplicates['gt_converted_cat_idxs'] == class_idx # Compute regression calibration errors. False negatives cant be evaluated since # those do not have ground truth. all_predicted_means = torch.cat( (true_positives['predicted_box_means'][true_positives_valid_idxs], duplicates['predicted_box_means'][duplicates_valid_idxs], localization_errors['predicted_box_means'][localization_errors_valid_idxs]), 0) all_predicted_covariances = torch.cat( (true_positives['predicted_box_covariances'][true_positives_valid_idxs], duplicates['predicted_box_covariances'][duplicates_valid_idxs], localization_errors['predicted_box_covariances'][localization_errors_valid_idxs]), 0) all_predicted_gt = torch.cat( (true_positives['gt_box_means'][true_positives_valid_idxs], duplicates['gt_box_means'][duplicates_valid_idxs], localization_errors['gt_box_means'][localization_errors_valid_idxs]), 0) all_predicted_covariances = torch.diagonal( all_predicted_covariances, dim1=1, dim2=2) # The assumption of uncorrelated components is not accurate, especially when estimating full # covariance matrices. However, using scipy to compute multivariate cdfs is very very # time consuming for such large amounts of data. reg_maximum_calibration_error = [] reg_expected_calibration_error = [] # Regression calibration is computed for every box dimension # separately, and averaged after. for box_dim in range(all_predicted_gt.shape[1]): all_predicted_means_current_dim = all_predicted_means[:, box_dim] all_predicted_gt_current_dim = all_predicted_gt[:, box_dim] all_predicted_covariances_current_dim = all_predicted_covariances[:, box_dim] normal_dists = torch.distributions.Normal( all_predicted_means_current_dim, scale=torch.sqrt(all_predicted_covariances_current_dim)) all_predicted_scores = normal_dists.cdf( all_predicted_gt_current_dim) reg_calibration_error = [] histogram_bin_step_size = 1 / 15.0 for i in torch.arange( 0.0, 1.0 - histogram_bin_step_size, histogram_bin_step_size): # Get number of elements in bin elements_in_bin = ( all_predicted_scores < (i + histogram_bin_step_size)) num_elems_in_bin_i = elements_in_bin.type( torch.FloatTensor).to(device).sum() # Compute calibration error from "Accurate uncertainties for deep # learning using calibrated regression" paper. reg_calibration_error.append( (num_elems_in_bin_i / all_predicted_scores.shape[0] - (i + histogram_bin_step_size)) ** 2) calibration_error = torch.stack( reg_calibration_error).to(device) reg_maximum_calibration_error.append(calibration_error.max()) reg_expected_calibration_error.append(calibration_error.mean()) reg_maximum_calibration_error_list.append( reg_maximum_calibration_error) reg_expected_calibration_error_list.append( reg_expected_calibration_error) # Summarize and print all reg_expected_calibration_error = torch.stack([torch.stack( reg, 0) for reg in reg_expected_calibration_error_list], 0) reg_expected_calibration_error = reg_expected_calibration_error[ ~torch.isnan(reg_expected_calibration_error)].mean() reg_maximum_calibration_error = torch.stack([torch.stack( reg, 0) for reg in reg_maximum_calibration_error_list], 0) reg_maximum_calibration_error = reg_maximum_calibration_error[ ~torch.isnan(reg_maximum_calibration_error)].mean() if print_results: table = PrettyTable() table.field_names = (['Cls Marginal Calibration Error', 'Reg Expected Calibration Error', 'Reg Maximum Calibration Error']) table.add_row([cls_marginal_calibration_error, reg_expected_calibration_error.cpu().numpy().tolist(), reg_maximum_calibration_error.cpu().numpy().tolist()]) print(table) text_file_name = os.path.join( inference_output_dir, 'calibration_errors_{}_{}_{}.txt'.format( iou_min, iou_correct, min_allowed_score)) with open(text_file_name, "w") as text_file: print([ cls_marginal_calibration_error, reg_expected_calibration_error.cpu().numpy().tolist(), reg_maximum_calibration_error.cpu().numpy().tolist()], file=text_file) dictionary_file_name = os.path.join( inference_output_dir, 'calibration_errors_res_{}_{}_{}.pkl'.format( iou_min, iou_correct, min_allowed_score)) final_accumulated_output_dict = { 'cls_marginal_calibration_error': cls_marginal_calibration_error, 'reg_expected_calibration_error': reg_expected_calibration_error.cpu().numpy(), 'reg_maximum_calibration_error': reg_maximum_calibration_error.cpu().numpy()} with open(dictionary_file_name, "wb") as pickle_file: pickle.dump(final_accumulated_output_dict, pickle_file) if __name__ == "__main__": # Create arg parser arg_parser = setup_arg_parser() args = arg_parser.parse_args() print("Command Line Args:", args) launch( main, args.num_gpus, num_machines=args.num_machines, machine_rank=args.machine_rank, dist_url=args.dist_url, args=(args,), )
14,295
45.718954
116
py
FDS
FDS-main/main.py
""" This is the base code to learn the learning rate, momentum and weight decay non-greedily with forward mode differentiation, over long horizons (e.g. CIFAR10) """ import os import time import shutil import torch import torch.optim as optim import pickle from utils.logger import * from utils.helpers import * from utils.datasets import * from models.selector import * class MetaLearner(object): def __init__(self, args): self.args = args ## Optimization self.hypers_init() self.cross_entropy = nn.CrossEntropyLoss() ## Experiment Set Up self.best_outer_step = 0 self.best_validation_acc = 0 ns, learnables = (self.args.n_lrs, self.args.n_moms, self.args.n_wds), (self.args.learn_lr, self.args.learn_mom, self.args.learn_wd) self.all_lr_schedules, self.all_mom_schedules, self.all_wd_schedules = [torch.zeros((self.args.n_outer_steps+1, n)) for n in ns] #+1 since save init schedules and last schedule self.all_lr_raw_grads, self.all_mom_raw_grads, self.all_wd_raw_grads = [torch.zeros((self.args.n_outer_steps, n)) if l else None for (n,l) in zip(ns, learnables)] self.all_lr_smooth_grads, self.all_mom_smooth_grads, self.all_wd_smooth_grads = [torch.zeros((self.args.n_outer_steps, n)) if l else None for (n,l) in zip(ns, learnables)] self.experiment_path = os.path.join(self.args.log_directory_path, self.args.experiment_name) self.checkpoint_path = os.path.join(self.experiment_path, 'checkpoint.pth.tar') if os.path.exists(self.experiment_path): if self.args.use_gpu and os.path.isfile(self.checkpoint_path): raise NotImplementedError(f"Experiment folder {self.experiment_path} already exists") #TODO: restore code from ckpt else: shutil.rmtree(self.experiment_path) # clear debug logs on cpu os.makedirs(self.experiment_path) else: os.makedirs(self.experiment_path) copy_file(os.path.realpath(__file__), self.experiment_path) # save this python file in logs folder self.logger = Logger(self.experiment_path, 'run_results.csv') ## Save and Print Args print('\n---------') with open(os.path.join(self.experiment_path, 'args.txt'), 'w+') as f: for k, v in self.args.__dict__.items(): print(k, v) f.write("{} \t {}\n".format(k, v)) print('---------\n') print('\nLogging every {} outer_steps and every {} epochs per outer_step\n'.format(self.args.outer_step_log_freq, self.args.epoch_log_freq)) def hypers_init(self): """ initialize hyperparameters """ self.inner_lrs = self.args.inner_lr_init*torch.ones(self.args.n_lrs, device=self.args.device) self.inner_lrs_grad = torch.zeros_like(self.inner_lrs) # lr hypergradient self.lr_hypersigns = torch.zeros(self.args.n_lrs, device=self.args.device) self.lr_step_sizes = self.args.lr_init_step_size*torch.ones(self.args.n_lrs, device=self.args.device) self.inner_moms = self.args.inner_mom_init*torch.ones(self.args.n_moms, device=self.args.device) self.inner_moms_grad = torch.zeros_like(self.inner_moms) self.mom_hypersigns = torch.zeros(self.args.n_moms, device=self.args.device) self.mom_step_sizes = self.args.mom_init_step_size*torch.ones(self.args.n_moms, device=self.args.device) self.inner_wds = self.args.inner_wd_init*torch.ones(self.args.n_wds, device=self.args.device) self.inner_wds_grad = torch.zeros_like(self.inner_wds) self.wd_hypersigns = torch.zeros(self.args.n_wds, device=self.args.device) self.wd_step_sizes = self.args.wd_init_step_size*torch.ones(self.args.n_wds, device=self.args.device) def get_hypers(self, epoch, batch_idx): """return hyperparameters to be used for given batch""" lr_index = int(self.args.n_lrs * (epoch*self.n_batches_per_epoch + batch_idx)/self.n_total_batches_for_this_outer_step) lr = float(self.inner_lrs[lr_index]) mom_index = int(self.args.n_moms * (epoch*self.n_batches_per_epoch + batch_idx)/self.n_total_batches_for_this_outer_step) mom = float(self.inner_moms[mom_index]) wd_index = int(self.args.n_wds * (epoch*self.n_batches_per_epoch + batch_idx)/self.n_total_batches_for_this_outer_step) wd = float(self.inner_wds[wd_index]) return lr, mom, wd, lr_index, mom_index, wd_index def to_prune(self, epoch, batch_idx, n_hypers): """ Do we skip calculation of Z for this batch?""" if self.args.pruning_ratio==0: to_prune=False else: n_batches_per_hyper = int(self.n_total_batches_for_this_outer_step/n_hypers) current_global_batch_idx = epoch*self.n_batches_per_epoch + batch_idx current_global_batch_idx_per_hyper = current_global_batch_idx % n_batches_per_hyper if self.args.pruning_mode=='alternate': #rounded to nearest integer, so r=0.25 -> prune 1 in 4 but r=0.21 -> 1 in 4 also if self.args.pruning_ratio>=0.5: #at least 1 in 2 pruned keep_freq = int(1/(1-self.args.pruning_ratio)) to_prune = (current_global_batch_idx_per_hyper % keep_freq != 0) else: prune_freq = int(1/(self.args.pruning_ratio)) to_prune = (current_global_batch_idx_per_hyper % prune_freq == 0) elif self.args.pruning_mode=='truncate': to_prune = current_global_batch_idx_per_hyper < self.args.pruning_ratio*n_batches_per_hyper return to_prune def inner_loop(self): """ Compute Z for each hyperparameter to learn over all epochs in the run """ ## Network self.classifier = select_model(True, self.args.dataset, self.args.architecture, self.args.init_type, self.args.init_param, self.args.device).to(self.args.device) self.classifier.train() self.weights = self.classifier.get_param() velocity = torch.zeros(self.weights.numel(), requires_grad=False, device=self.args.device) ## Forward Mode Init if self.args.learn_lr: self.n_batches_per_lr = 0 Z_lr = torch.zeros((self.weights.numel(), self.args.n_lrs), device=self.args.device) C_lr = torch.zeros((self.weights.numel(), self.args.n_lrs), device=self.args.device) else: Z_lr = None if self.args.learn_mom: self.n_batches_per_mom = 0 Z_mom = torch.zeros((self.weights.numel(), self.args.n_moms), device=self.args.device) C_mom = torch.zeros((self.weights.numel(), self.args.n_moms), device=self.args.device) else: Z_mom = None if self.args.learn_wd: self.n_batches_per_wd = 0 Z_wd = torch.zeros((self.weights.numel(), self.args.n_wds), device=self.args.device) C_wd = torch.zeros((self.weights.numel(), self.args.n_wds), device=self.args.device) else: Z_wd = None ## Inner Loop Over All Epochs for epoch in range(self.n_inner_epochs_for_this_outer_step): t0_epoch = time.time() for batch_idx, (x_train, y_train) in enumerate(self.train_loader): lr, mom, wd, lr_index, mom_index, wd_index = self.get_hypers(epoch, batch_idx) #print(f'epoch {epoch} batch {batch_idx} -- lr idx {lr_index} -- mom idx {mom_index} -- wd index {wd_index}') x_train, y_train = x_train.to(device=self.args.device), y_train.to(device=self.args.device) train_logits = self.classifier.forward_with_param(x_train, self.weights) train_loss = self.cross_entropy(train_logits, y_train) grads = torch.autograd.grad(train_loss, self.weights, create_graph=True)[0] if self.args.clamp_grads: grads.clamp_(-self.args.clamp_grads_range, self.args.clamp_grads_range) if self.args.learn_lr and not self.to_prune(epoch, batch_idx, self.args.n_lrs): #print('update lr') self.n_batches_per_lr += 1 H_times_Z = torch.zeros((self.weights.numel(), self.args.n_lrs),device=self.args.device) for j in range(lr_index + 1): retain = (j != lr_index) or self.args.learn_mom or self.args.learn_wd H_times_Z[:, j] = torch.autograd.grad(grads @ Z_lr[:, j], self.weights, retain_graph=retain)[0] if self.args.clamp_HZ: H_times_Z.clamp_(-self.args.clamp_HZ_range, self.args.clamp_HZ_range) A_times_Z = Z_lr*(1 - lr*wd) - lr*H_times_Z B = - mom*lr*C_lr B[:,lr_index] -= grads.detach() + wd*self.weights.detach() + mom*velocity C_lr = mom*C_lr + H_times_Z + wd*Z_lr Z_lr = A_times_Z + B if self.args.learn_mom and not self.to_prune(epoch, batch_idx, self.args.n_moms): #print('update mom') self.n_batches_per_mom += 1 H_times_Z = torch.zeros((self.weights.numel(), self.args.n_moms),device=self.args.device) for j in range(mom_index + 1): retain = (j != mom_index) or self.args.learn_wd H_times_Z[:, j] = torch.autograd.grad(grads @ Z_mom[:, j], self.weights, retain_graph=retain)[0] if self.args.clamp_HZ: H_times_Z.clamp_(-self.args.clamp_HZ_range, self.args.clamp_HZ_range) A_times_Z = (1 - lr*wd)*Z_mom - lr*H_times_Z B = -lr*mom*C_mom B[:, mom_index] -= lr*velocity C_mom = mom*C_mom + H_times_Z + wd * Z_mom C_mom[:, mom_index] += velocity Z_mom = A_times_Z + B if self.args.learn_wd and not self.to_prune(epoch, batch_idx, self.args.n_wds): #print('update wd') self.n_batches_per_wd += 1 H_times_Z = torch.zeros((self.weights.numel(), self.args.n_wds),device=self.args.device) for j in range(wd_index + 1): retain = (j != wd_index) H_times_Z[:, j] = torch.autograd.grad(grads @ Z_wd[:, j], self.weights, retain_graph=retain)[0] if self.args.clamp_HZ: H_times_Z.clamp_(-self.args.clamp_HZ_range, self.args.clamp_HZ_range) A_times_Z = (1 - lr*wd)*Z_wd - lr*H_times_Z B = - lr*mom*C_wd B[:, wd_index] -= lr*self.weights.detach() C_wd = mom*C_wd + H_times_Z + wd*Z_wd C_wd[:, wd_index] += self.weights.detach() Z_wd = A_times_Z + B ## SGD inner update self.weights.detach_(), grads.detach_() velocity = velocity*mom + (grads + wd*self.weights) self.weights = self.weights - lr*velocity self.weights.requires_grad_() print(f'--- Ran epoch {epoch+1} in {format_time(time.time()-t0_epoch)} ---') if self.args.learn_lr: self.n_batches_per_lr /= self.args.n_lrs # each hyper gets same # of updates regardless of pruning mode if self.args.learn_mom: self.n_batches_per_mom /= self.args.n_moms if self.args.learn_wd: self.n_batches_per_wd /= self.args.n_wds return Z_lr, Z_mom, Z_wd def outer_step(self, outer_step_idx, Z_lr_final, Z_mom_final, Z_wd_final): """ Calculate hypergradients and update hyperparameters accordingly. """ ## Calculate validation gradients with final weights of inner loop self.running_val_grad = AggregateTensor() for batch_idx, (x_val, y_val) in enumerate(self.val_loader): #need as big batches as train mode for BN train mode x_val, y_val = x_val.to(device=self.args.device), y_val.to(device=self.args.device) val_logits = self.classifier.forward_with_param(x_val, self.weights) val_loss = self.cross_entropy(val_logits, y_val) dLval_dw = torch.autograd.grad(val_loss, self.weights)[0] self.running_val_grad.update(dLval_dw) ## Update hyperparams print('') if self.args.learn_lr: self.inner_lrs_grad = self.running_val_grad.avg() @ Z_lr_final / self.n_batches_per_lr self.all_lr_raw_grads[outer_step_idx] = self.inner_lrs_grad.detach() print('RAW LR GRADS: ', ["{:.2E}".format(float(i)) for i in self.inner_lrs_grad]) new_hypersigns = torch.sign(self.inner_lrs_grad) #Nans and zero have sign 0 flipped_signs = self.lr_hypersigns*new_hypersigns # 1, -1 or 0 multipliers = torch.tensor([self.args.lr_step_decay if f==-1.0 else 1.0 for f in flipped_signs], device=self.args.device) self.lr_step_sizes = multipliers*self.lr_step_sizes self.lr_hypersigns = new_hypersigns deltas = new_hypersigns*self.lr_step_sizes # how much to change hyperparameter by self.lr_converged = ((self.lr_step_sizes/self.inner_lrs) < self.args.converged_frac).all() self.inner_lrs = self.inner_lrs - deltas self.all_lr_smooth_grads[outer_step_idx] = deltas print('SMOOTH LR DELTAS: ', ["{:02.2f}".format(float(i)) for i in deltas]) if self.args.learn_mom: self.inner_moms_grad = self.running_val_grad.avg() @ Z_mom_final / self.n_batches_per_mom self.all_mom_raw_grads[outer_step_idx] = self.inner_moms_grad.detach() print('RAW MOM GRADS: ', ["{:.2E}".format(float(i)) for i in self.inner_moms_grad]) new_hypersigns = torch.sign(self.inner_moms_grad) #Nans and zero have sign 0 flipped_signs = self.mom_hypersigns*new_hypersigns # 1, -1 or 0 multipliers = torch.tensor([self.args.mom_step_decay if f==-1.0 else 1.0 for f in flipped_signs], device=self.args.device) self.mom_step_sizes = multipliers*self.mom_step_sizes self.mom_hypersigns = new_hypersigns deltas = new_hypersigns*self.mom_step_sizes # how much to change hyperparameter by self.mom_converged = ((self.mom_step_sizes/self.inner_moms) < self.args.converged_frac).all() self.inner_moms = self.inner_moms - deltas self.all_mom_smooth_grads[outer_step_idx] = deltas print('SMOOTH MOM DELTAS: ', ["{:02.2f}".format(float(i)) for i in deltas]) if self.args.learn_wd: self.inner_wds_grad = self.running_val_grad.avg() @ Z_wd_final / self.n_batches_per_wd self.all_wd_raw_grads[outer_step_idx] = self.inner_wds_grad.detach() print('RAW WD GRADS: ', ["{:.2E}".format(float(i)) for i in self.inner_wds_grad]) new_hypersigns = torch.sign(self.inner_wds_grad) #Nans and zero have sign 0 flipped_signs = self.wd_hypersigns*new_hypersigns # 1, -1 or 0 multipliers = torch.tensor([self.args.wd_step_decay if f==-1.0 else 1.0 for f in flipped_signs], device=self.args.device) self.wd_step_sizes = multipliers*self.wd_step_sizes self.wd_hypersigns = new_hypersigns deltas = new_hypersigns*self.wd_step_sizes # how much to change hyperparameter by self.wd_converged = ((self.wd_step_sizes/self.inner_wds) < self.args.converged_frac).all() self.inner_wds = self.inner_wds - deltas self.all_wd_smooth_grads[outer_step_idx] = deltas print('SMOOTH WD DELTAS: ', ["{:02.2f}".format(float(i)) for i in deltas]) self.converged = (self.lr_converged if self.args.learn_lr else True) and (self.mom_converged if self.args.learn_mom else True) and (self.wd_converged if self.args.learn_wd else True) def run(self): """ Run meta learning experiment """ t0 = time.time() for outer_step_idx in range(self.args.n_outer_steps): # number of outer steps ## Set up self.n_inner_epochs_for_this_outer_step = self.args.n_inner_epochs_per_outer_steps[outer_step_idx] print(f'\nOuter step {outer_step_idx+1}/{self.args.n_outer_steps} --- current budget of {self.n_inner_epochs_for_this_outer_step} epochs --- using:') print('lrs = ', [float('{:02.2e}'.format(el)) for el in self.inner_lrs], 'moms = ', [float('{:02.2e}'.format(el)) for el in self.inner_moms], 'wds = ', [float('{:02.2e}'.format(el)) for el in self.inner_wds]) self.all_lr_schedules[outer_step_idx], self.all_mom_schedules[outer_step_idx], self.all_wd_schedules[outer_step_idx] = self.inner_lrs.detach(), self.inner_moms.detach(), self.inner_wds.detach() self.save_state(outer_step_idx) # state and lrs saved correspond to those set at the beginning of the outer_step ## New data split for each outer_step self.train_loader, self.val_loader, self.test_loader = get_loaders(datasets_path=self.args.datasets_path, dataset=self.args.dataset, train_batch_size=self.args.train_batch_size, val_batch_size=self.args.val_batch_size, val_source='train', val_train_fraction=self.args.val_train_fraction, val_train_overlap=self.args.val_train_overlap, workers=self.args.workers, train_infinite=False, val_infinite=False, cutout=self.args.cutout, cutout_length=self.args.cutout_length, cutout_prob=self.args.cutout_prob) self.n_batches_per_epoch = len(self.train_loader) self.n_total_batches_for_this_outer_step = self.n_inner_epochs_for_this_outer_step * self.n_batches_per_epoch ## Update Hypers Z_lr_final, Z_mom_final, Z_wd_final = self.inner_loop() self.outer_step(outer_step_idx, Z_lr_final, Z_mom_final, Z_wd_final) ## See if schedule used for this outer_step led to best validation _, val_acc = self.validate(self.weights) _, test_acc = self.test(self.weights) if val_acc > self.best_validation_acc: self.best_validation_acc = val_acc self.best_outer_step = outer_step_idx #print(f'Best validation acc at outer_step idx {outer_step_idx}') ## Break if all hyperparameters have converged if self.converged: print('STOP HYPERTRAINING BECAUSE ALL HYPERPARAMETERS HAVE CONVERGED') break ## Time time_so_far = time.time() - t0 self.logger.write({'budget': self.n_inner_epochs_for_this_outer_step, 'time': time_so_far, 'val_acc': val_acc, 'test_acc': test_acc}) print(f'final val acc {100*val_acc:.2g} -- final test_acc: {100*test_acc:.2g}') ## Logging Final Metrics self.all_lr_schedules[outer_step_idx+1], self.all_mom_schedules[outer_step_idx+1], self.all_wd_schedules[outer_step_idx+1] = self.inner_lrs.detach(), self.inner_moms.detach(), self.inner_wds.detach() #last schedule was never trained on self.save_state(outer_step_idx+1) avg_test_loss, avg_test_acc = self.test(self.weights) return avg_test_acc def validate(self, weights, fraction=1.0): """ Fraction allows trading accuracy for speed when logging many times""" self.classifier.eval() running_acc, running_loss = AggregateTensor(), AggregateTensor() with torch.no_grad(): for batch_idx, (x, y) in enumerate(self.val_loader): x, y = x.to(device=self.args.device), y.to(device=self.args.device) logits = self.classifier.forward_with_param(x, weights) running_loss.update(self.cross_entropy(logits, y), x.shape[0]) running_acc.update(accuracy(logits, y, topk=(1,))[0], x.shape[0]) if fraction < 1 and (batch_idx + 1) >= fraction*len(self.val_loader): break self.classifier.train() return float(running_loss.avg()), float(running_acc.avg()) def test(self, weights, fraction=1.0): """ Fraction allows trading accuracy for speed when logging many times""" self.classifier.eval() running_acc, running_loss = AggregateTensor(), AggregateTensor() with torch.no_grad(): for batch_idx, (x, y) in enumerate(self.test_loader): x, y = x.to(device=self.args.device), y.to(device=self.args.device) logits = self.classifier.forward_with_param(x, weights) running_loss.update(self.cross_entropy(logits, y), x.shape[0]) running_acc.update(accuracy(logits, y, topk=(1,))[0], x.shape[0]) if fraction < 1 and (batch_idx + 1) >= fraction*len(self.test_loader): break self.classifier.train() return float(running_loss.avg()), float(running_acc.avg()) def save_state(self, outer_step_idx): torch.save({'args': self.args, 'outer_step_idx': outer_step_idx, 'best_outer_step': self.best_outer_step, 'best_validation_acc': self.best_validation_acc, 'all_lr_schedules': self.all_lr_schedules, 'all_lr_raw_grads': self.all_lr_raw_grads, 'all_lr_smooth_grads': self.all_lr_smooth_grads, 'all_mom_schedules': self.all_mom_schedules, 'all_mom_raw_grads': self.all_mom_raw_grads, 'all_mom_smooth_grads': self.all_mom_smooth_grads, 'all_wd_schedules': self.all_wd_schedules, 'all_wd_raw_grads': self.all_wd_raw_grads, 'all_wd_smooth_grads': self.all_wd_smooth_grads}, self.checkpoint_path) class BaseLearner(object): """ Retrain from scratch using learned schedule and whole training set """ def __init__(self, args, lr_schedule, mom_schedule, wd_schedule, log_name): self.args = args self.inner_lrs = lr_schedule self.inner_moms = mom_schedule self.inner_wds = wd_schedule ## Loaders self.args.val_source = 'test' # retrain on full train set from scratch self.train_loader, _, self.test_loader = get_loaders(datasets_path=self.args.datasets_path, dataset=self.args.dataset, train_batch_size=self.args.train_batch_size, val_batch_size=self.args.val_batch_size, val_source=self.args.val_source, val_train_fraction=self.args.val_train_fraction, val_train_overlap=self.args.val_train_overlap, workers=self.args.workers, train_infinite=False, val_infinite=False, cutout=self.args.cutout, cutout_length=self.args.cutout_length, cutout_prob=self.args.cutout_prob) self.n_batches_per_epoch = len(self.train_loader) self.n_total_batches = self.args.retrain_n_epochs * self.n_batches_per_epoch ## Optimizer self.classifier = select_model(False, self.args.dataset, self.args.architecture, self.args.init_type, self.args.init_param, self.args.device).to(self.args.device) self.optimizer = optim.SGD(self.classifier.parameters(), lr=0.0, momentum=0.0, weight_decay=0.0) #set hypers manually later self.cross_entropy = nn.CrossEntropyLoss() ### Set up self.experiment_path = os.path.join(args.log_directory_path, args.experiment_name) self.logger = Logger(self.experiment_path, log_name) def log_init(self): self.running_train_loss, self.running_train_acc = AggregateTensor(), AggregateTensor() def log(self, epoch, avg_train_loss, avg_train_acc): avg_test_loss, avg_test_acc = self.test(fraction=0.1 if epoch!=self.args.retrain_n_epochs-1 else 1) print('Retrain epoch {}/{} --- Train Acc: {:02.2f}% -- Test Acc: {:02.2f}%'.format(epoch+1, self.args.retrain_n_epochs, avg_train_acc * 100, avg_test_acc * 100)) self.logger.write({'train_loss': avg_train_loss, 'train_acc': avg_train_acc, 'test_loss': avg_test_loss, 'test_acc': avg_test_acc}) self.log_init() def get_hypers(self, epoch, batch_idx): """return hyperparameters to be used for given batch""" lr_index = int(self.args.n_lrs * (epoch*self.n_batches_per_epoch + batch_idx)/self.n_total_batches) lr = float(self.inner_lrs[lr_index]) mom_index = int(self.args.n_moms * (epoch*self.n_batches_per_epoch + batch_idx)/self.n_total_batches) mom = float(self.inner_moms[mom_index]) wd_index = int(self.args.n_wds * (epoch*self.n_batches_per_epoch + batch_idx)/self.n_total_batches) wd = float(self.inner_wds[wd_index]) return lr, mom, wd, lr_index, mom_index, wd_index def set_hypers(self, epoch, batch_idx): lr, mom, wd, lr_index, mom_index, wd_index = self.get_hypers(epoch, batch_idx) for param_group in self.optimizer.param_groups: param_group['lr'] = lr param_group['momentum'] = mom param_group['weight_decay'] = wd #print(f'Setting: lr={lr}, mom={mom}, wd={wd}') def run(self): for epoch in range(self.args.retrain_n_epochs): avg_train_loss, avg_train_acc = self.train(epoch) self.log(epoch, avg_train_loss, avg_train_acc) test_loss, test_acc = self.test() return test_acc def train(self, epoch): self.classifier.train() running_acc, running_loss = AggregateTensor(), AggregateTensor() for batch_idx, (x,y) in enumerate(self.train_loader): self.set_hypers(epoch, batch_idx) x, y = x.to(device=self.args.device), y.to(device=self.args.device) logits = self.classifier(x) loss = self.cross_entropy(input=logits, target=y) acc1 = accuracy(logits.data, y, topk=(1,))[0] running_loss.update(loss, x.shape[0]) running_acc.update(acc1, x.shape[0]) self.optimizer.zero_grad() loss.backward() self.optimizer.step() return float(running_loss.avg()), float(running_acc.avg()) def test(self, fraction=1.0): """ fraction allows trading accuracy for speed when logging many times""" self.classifier.eval() running_acc, running_loss = AggregateTensor(), AggregateTensor() with torch.no_grad(): for batch_idx, (x, y) in enumerate(self.test_loader): x, y = x.to(device=self.args.device), y.to(device=self.args.device) logits = self.classifier(x) running_loss.update(self.cross_entropy(logits, y), x.shape[0]) running_acc.update(accuracy(logits, y, topk=(1,))[0], x.shape[0]) if fraction < 1 and (batch_idx + 1) >= fraction*len(self.test_loader): break self.classifier.train() return float(running_loss.avg()), float(running_acc.avg()) # ________________________________________________________________________________ # ________________________________________________________________________________ # ________________________________________________________________________________ def make_experiment_name(args): """ Warning: Windows can have a weird behaviour for long filenames. Protip: switch to Ubuntu ;) """ ## Main nepr = ''.join([str(i) for i in args.n_inner_epochs_per_outer_steps]) experiment_name = f'FSL_{args.dataset}_{args.architecture}_nepr{nepr}' experiment_name += f'_init{args.init_type}-{args.init_param}' experiment_name += f'_tbs{args.train_batch_size}' if args.cutout: experiment_name += f'_cutout-p{args.cutout_prob}' if args.clamp_HZ: experiment_name += f'_HZclamp{args.clamp_HZ_range}' experiment_name += f'_S{args.seed}' return experiment_name def main(args): set_torch_seeds(args.seed) torch.backends.cudnn.enabled = True torch.backends.cudnn.deterministic = False torch.backends.cudnn.benchmark = True t0 = time.time() meta_learner = MetaLearner(args) meta_test_acc = meta_learner.run() total_time = time.time() - t0 to_print = '\n\nMETA TEST ACC: {:02.2f}%'.format(meta_test_acc*100) file_name = "final_meta_test_acc_{:02.2f}_total_time_{}".format(meta_test_acc*100, format_time(total_time)) create_empty_file(os.path.join(args.log_directory_path, args.experiment_name, file_name)) if args.retrain_from_scratch: ## Fetch schedules # best_idx = meta_learner.best_outer_step final_lr_schedule, final_mom_schedule, final_wd_schedule = meta_learner.all_lr_schedules[-1], meta_learner.all_mom_schedules[-1], meta_learner.all_wd_schedules[-1] # best_lr_schedule, best_mom_schedule, best_wd_schedule = meta_learner.all_lr_schedules[best_idx], meta_learner.all_mom_schedules[best_idx], meta_learner.all_wd_schedules[best_idx] del meta_learner ## Retrain Last print(f'\n\n\n---------- RETRAINING FROM SCRATCH WITH LAST SCHEDULE (idx {args.n_outer_steps}) ----------') print(f'lrs = {final_lr_schedule.tolist()}') print(f'moms = {final_mom_schedule.tolist()}') print(f'wds = {final_wd_schedule.tolist()}') log_name = f'Rerun_last_outer_step.csv' base_learner = BaseLearner(args, final_lr_schedule, final_mom_schedule, final_wd_schedule, log_name) if args.use_gpu: torch.cuda.empty_cache() base_test_acc = base_learner.run() to_print += '\nRE-RUN LAST SCHEDULE TEST ACC: {:02.2f}%'.format(base_test_acc*100) file_name = "Rerun_last_test_acc_{:02.2f}".format(base_test_acc*100) create_empty_file(os.path.join(args.log_directory_path, args.experiment_name, file_name)) # ## Retrain Best Val # print(f'\n\n\n---------- RETRAINING FROM SCRATCH WITH BEST VAL SCHEDULE (idx {best_idx}) ----------') # print(f'lrs = {best_lr_schedule.tolist()}') # print(f'moms = {best_mom_schedule.tolist()}') # print(f'wds = {best_wd_schedule.tolist()}') # # log_name = f'Rerun_best_outer_step_idx_{best_idx}.csv' # base_learner = BaseLearner(args, best_lr_schedule, best_mom_schedule, best_wd_schedule, log_name) # if args.use_gpu: torch.cuda.empty_cache() # base_test_acc = base_learner.run() # to_print += '\nRE-RUN BEST SCHEDULE TEST ACC: {:02.2f}%'.format(base_test_acc*100) # file_name = "Rerun_best_test_acc_{:02.2f}".format(base_test_acc*100) # create_empty_file(os.path.join(args.log_directory_path, args.experiment_name, file_name)) print(to_print) if __name__ == "__main__": import argparse print('Running...') parser = argparse.ArgumentParser(description='Welcome to GreedyGrad') ## Main parser.add_argument('--learn_lr', type=str2bool, default=True) parser.add_argument('--learn_mom', type=str2bool, default=True) parser.add_argument('--learn_wd', type=str2bool, default=True) parser.add_argument('--n_lrs', type=int, default=7) parser.add_argument('--n_moms', type=int, default=1) parser.add_argument('--n_wds', type=int, default=1) parser.add_argument('--dataset', type=str, default='CIFAR10') parser.add_argument('--n_inner_epochs_per_outer_steps', nargs='*', type=int, default=[1, 10, 10, 10, 10, 10, 10, 10, 10, 10], help='number of epochs to run for each outer step') parser.add_argument('--pruning_mode', type=str, choices=['alternate', 'truncate'], default='alternate') parser.add_argument('--pruning_ratio', type=float, default=0.0, help='<1, how many inner steps to skip Z calculation for expressed as a fraction of total inner steps per hyper') ## Architecture parser.add_argument('--architecture', type=str, default='WRN-16-1') parser.add_argument('--init_type', type=str, default='xavier', choices=['normal', 'xavier', 'kaiming', 'orthogonal', 'zero', 'default'], help='network initialization scheme') parser.add_argument('--init_param', type=float, default=1, help='network initialization param: gain, std, etc.') parser.add_argument('--init_norm_weights', type=float, default=1, help='init gammas of BN') ## Inner Loop parser.add_argument('--inner_lr_init', type=float, default=0, help='SGD inner learning rate init') parser.add_argument('--inner_mom_init', type=float, default=0, help='SGD inner momentum init') parser.add_argument('--inner_wd_init', type=float, default=0, help='SGD inner weight decay init') parser.add_argument('--train_batch_size', type=int, default=256) parser.add_argument('--clamp_grads', type=str2bool, default=True) parser.add_argument('--clamp_grads_range', type=float, default=3, help='clamp inner grads for each batch to +/- that') parser.add_argument('--cutout', type=str2bool, default=False) parser.add_argument('--cutout_length', type=int, default=16) parser.add_argument('--cutout_prob', type=float, default=1, help='clamp inner grads for each batch to +/- that') ## Outer Loop parser.add_argument('--val_batch_size', type=int, default=500) parser.add_argument('--val_train_fraction', type=float, default=0.05) parser.add_argument('--val_train_overlap', type=str2bool, default=False, help='if True and val_source=train, val images are also in train set') parser.add_argument('--lr_init_step_size', type=float, default=0.1, help='at each iteration grads changed so that each hyper can only change by this fraction (ignoring outer momentum)') parser.add_argument('--mom_init_step_size', type=float, default=0.1) parser.add_argument('--wd_init_step_size', type=float, default=3e-4) parser.add_argument('--lr_step_decay', type=float, default=0.5, help='step size multiplied by this much if hypergrad sign changes') parser.add_argument('--mom_step_decay', type=float, default=0.5, help='step size multiplied by this much if hypergrad sign changes') parser.add_argument('--wd_step_decay', type=float, default=0.5, help='step size multiplied by this much if hypergrad sign changes') parser.add_argument('--clamp_HZ', type=str2bool, default=True) parser.add_argument('--clamp_HZ_range', type=float, default=1, help='clamp to +/- that') parser.add_argument('--converged_frac', type=float, default=0.05, help='if steps are smaller than this percentage of hypers, stop experiment') ## Other parser.add_argument('--retrain_from_scratch', type=str2bool, default=True, help='retrain from scratch with learned lr schedule') parser.add_argument('--retrain_n_epochs', type=int, default=50, help='interpolates from learned schedule, -1 for same as n_inner_epochs_per_outer_steps[-1]') parser.add_argument('--datasets_path', type=str, default="~/Datasets/Pytorch/") parser.add_argument('--log_directory_path', type=str, default="./logs/") parser.add_argument('--epoch_log_freq', type=int, default=1, help='every how many epochs to save summaries') parser.add_argument('--outer_step_log_freq', type=int, default=1, help='every how many outer_steps to save the whole run') parser.add_argument('--seed', type=int, default=0) parser.add_argument('--workers', type=int, default=0) parser.add_argument('--use_gpu', type=str2bool, default=True) args = parser.parse_args() args.dataset_path = os.path.join(args.datasets_path, args.dataset) args.use_gpu = args.use_gpu and torch.cuda.is_available() args.device = torch.device('cuda') if args.use_gpu else torch.device('cpu') assert args.lr_step_decay < 1 assert args.mom_step_decay < 1 assert args.wd_step_decay < 1 assert args.converged_frac < 1 if args.retrain_n_epochs < 0: args.retrain_n_epochs = args.n_inner_epochs_per_outer_steps[-1] assert args.pruning_ratio <= 1 args.n_outer_steps = len(args.n_inner_epochs_per_outer_steps) args.experiment_name = make_experiment_name(args) print('\nRunning on device: {}'.format(args.device)) if args.use_gpu: print(torch.cuda.get_device_name(0)) main(args)
37,988
54.866176
243
py
FDS
FDS-main/figure2_hypergradients_fluctuation.py
""" Here we measure hypergradients for several runs when perturbing the training data and weight initialization. This must be done on toy datasets where reverse-mode differentiation is tractable. This corresponds to figure 2 in the paper. """ import torch.optim as optim import pickle import os import warnings import sys import shutil import torch import torch.nn.functional as F import torch.optim as optimw from utils.helpers import * from utils.datasets import * from models.selector import * class HyperGradFluctuation(object): def __init__(self, args): self.args = args self.hypergrads_all = torch.zeros((self.args.n_runs, self.args.T)) self.cross_entropy = nn.CrossEntropyLoss() self.init_lr_schedule() ## Loaders self.infinite_train_loader, self.val_loader, _ = get_loaders(datasets_path=self.args.datasets_path, dataset=self.args.dataset, train_batch_size=self.args.train_batch_size, val_batch_size=self.args.n_val_images, val_source='test', workers=self.args.workers, train_infinite=True, val_infinite=False) for x,y in self.val_loader: self.X_val, self.Y_val = x.to(device=self.args.device), y.to(device=self.args.device) ## Set up experiment folder self.experiment_path = os.path.join(self.args.log_directory_path, self.args.experiment_name) if os.path.isfile(os.path.join(self.experiment_path, 'hypergrads.pth.tar')): if args.use_gpu: raise FileExistsError(f'Experiment already ran and exists at {self.experiment_path}. \nStopping now') else: if os.path.exists(self.experiment_path): shutil.rmtree(self.experiment_path) os.makedirs(self.experiment_path) ## Save and Print Args copy_file(os.path.realpath(__file__), self.experiment_path) # save this python file in logs folder print('\n---------') with open(os.path.join(self.experiment_path, 'args.txt'), 'w+') as f: for k, v in self.args.__dict__.items(): print(k, v) f.write("{} \t {}\n".format(k, v)) print('---------\n') def init_lr_schedule(self): if self.args.inner_lr_cosine_anneal: dummy_opt = optim.SGD([torch.ones([1], requires_grad=True)], lr=self.args.inner_lr_init) dummy_scheduler = optim.lr_scheduler.CosineAnnealingLR(dummy_opt, T_max=self.args.T) lrs = [] for i in range(self.args.T): lrs.append(dummy_scheduler.get_last_lr()[0]) dummy_opt.step() dummy_scheduler.step() self.inner_lrs = torch.tensor(lrs, requires_grad=True, device=self.args.device) else: self.inner_lrs = torch.full((self.args.T,), self.args.inner_lr_init, requires_grad=True, device=self.args.device) def inner_and_outer_loop(self): for self.inner_step_idx, (x_train, y_train) in enumerate(self.infinite_train_loader): x_train, y_train = x_train.to(self.args.device, self.args.dtype), y_train.to(self.args.device) train_logits = self.classifier.forward_with_param(x_train, self.weights) train_loss = self.cross_entropy(train_logits, y_train) grads = torch.autograd.grad(train_loss, self.weights, create_graph=True)[0] if self.args.clamp_inner_grads: grads.clamp_(-self.args.clamp_inner_grads_range, self.args.clamp_inner_grads_range) self.velocity = self.args.inner_momentum * self.velocity + (grads + self.args.inner_weight_decay * self.weights) self.weights = self.weights - self.inner_lrs[self.inner_step_idx] * self.velocity if self.args.greedy: self.compute_hypergradients() #only populates .grad of one item in self.inner_lrs self.weights.detach_().requires_grad_() self.velocity.detach_().requires_grad_() if self.inner_step_idx+1 == self.args.T: break if not self.args.greedy: self.compute_hypergradients() #populates .grad of all items in self.inner_lrs def compute_hypergradients(self): val_logits = self.classifier.forward_with_param(self.X_val, self.weights) val_loss = self.cross_entropy(val_logits, self.Y_val) val_loss.backward() def run(self): for self.run_idx in range(self.args.n_runs): self.classifier = select_model(True, self.args.dataset, self.args.architecture, self.args.init_type, self.args.init_param, self.args.device).to(self.args.device) self.weights = self.classifier.get_param() self.velocity = torch.zeros(self.weights.numel(), device=self.args.device) self.inner_and_outer_loop() self.hypergrads_all[self.run_idx] = self.inner_lrs.grad.detach() self.inner_lrs.grad.data.zero_() self.save_final() def save_final(self): torch.save({'args': self.args, 'hypergrads_all': self.hypergrads_all}, os.path.join(self.experiment_path, 'hypergrads.pth.tar')) print(f"Saved hypergrads to {os.path.join(self.experiment_path, 'hypergrads.pth.tar')}") # ________________________________________________________________________________ # ________________________________________________________________________________ # ________________________________________________________________________________ def make_experiment_name(args): experiment_name = f'Hg_{args.dataset}_{args.init_type}_T{args.T}_tbs{args.train_batch_size}_mom{args.inner_momentum}_wd{args.inner_weight_decay}_ilr{args.inner_lr_init}' if args.inner_lr_cosine_anneal: experiment_name += f'cosine' if args.greedy: experiment_name += f'_GREEDY' if args.dtype == torch.float64: experiment_name += '_FL64' experiment_name += f'_S{args.seed}' return experiment_name def main(args): set_torch_seeds(args.seed) torch.backends.cudnn.enabled = True torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True t0 = time.time() hypervariance_learner = HyperGradFluctuation(args) hypervariance_learner.run() total_time = time.time() - t0 with open(os.path.join(args.log_directory_path, args.experiment_name, 'TOTAL_TIME_' + format_time(total_time)), 'w+') as f: f.write("NA") if __name__ == "__main__": import argparse print('Running...') parser = argparse.ArgumentParser(description='Welcome to GreedyGrad') ## Main parser.add_argument('--T', type=int, default=250, help='number of batches for the task and to learn a schedule over') parser.add_argument('--n_runs', type=int, default=100, help='how many times to compute hypergrads, with different train-val-split each time') parser.add_argument('--dataset', type=str, default='SVHN') parser.add_argument('--greedy', type=str2bool, default=False) parser.add_argument('--architecture', type=str, default='LeNet') parser.add_argument('--init_type', type=str, default='xavier', choices=['normal', 'xavier', 'kaiming', 'orthogonal', 'zero', 'default'], help='network initialization scheme') parser.add_argument('--init_param', type=float, default=1, help='network initialization param: gain, std, etc.') parser.add_argument('--n_val_images', type=int, default=2000, help='ignored unless val_source=train') #20% of 60k=12000 ## Inner Loop parser.add_argument('--inner_lr_init', type=float, default=0.01, help='Used to initialize inner learning rate(s).') parser.add_argument('--inner_lr_cosine_anneal', type=str2bool, default=True, help='Initial schedule is cosine annealing') parser.add_argument('--inner_momentum', type=float, default=0.9, help='SGD inner momentum') parser.add_argument('--inner_weight_decay', type=float, default=0.0, help='SGD + ADAM inner weight decay') parser.add_argument('--train_batch_size', type=int, default=128) parser.add_argument('--clamp_inner_grads', type=str2bool, default=True) parser.add_argument('--clamp_inner_grads_range', type=float, default=1, help='clamp inner grads for each batch to +/- that') ## Misc parser.add_argument('--datasets_path', type=str, default="~/Datasets/Pytorch/") parser.add_argument('--log_directory_path', type=str, default="./logs/") parser.add_argument('--dtype', type=str, default='float32', choices=['float32', 'float64']) parser.add_argument('--seed', type=int, default=0) parser.add_argument('--workers', type=int, default=0) parser.add_argument('--use_gpu', type=str2bool, default=True) args = parser.parse_args() args.dataset_path = os.path.join(args.datasets_path, args.dataset) args.use_gpu = args.use_gpu and torch.cuda.is_available() args.device = torch.device('cuda') if args.use_gpu else torch.device('cpu') if args.dtype == 'float64': torch.set_default_tensor_type(torch.DoubleTensor) # changes weights and tensors but not loaders args.dtype = torch.float64 if args.dtype == 'float64' else torch.float32 print('\nRunning on device: {}'.format(args.device)) args.experiment_name = make_experiment_name(args) main(args)
9,738
49.201031
178
py
FDS
FDS-main/models/wresnet.py
""" Base architecture taken from https://github.com/xternalz/WideResNet-pytorch """ import torch import torch.nn as nn import torch.nn.functional as F from models.meta_factory import ReparamModule class BasicBlock(nn.Module): def __init__(self, in_planes, out_planes, stride, dropRate): super(BasicBlock, self).__init__() self.dropRate = dropRate self.bn1 = nn.BatchNorm2d(in_planes) self.relu1 = nn.ReLU(inplace=True) self.conv1 = nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(out_planes) self.relu2 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(out_planes, out_planes, kernel_size=3, stride=1, padding=1, bias=False) self.equalInOut = (in_planes == out_planes) self.convShortcut = (not self.equalInOut) and nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False) or None def forward(self, x): if self.equalInOut: out = self.relu1(self.bn1(x)) out = self.relu2(self.bn2(self.conv1(out))) else: #keep x var so can add it in skip connection x = self.relu1(self.bn1(x)) out = self.relu2(self.bn2(self.conv1(x))) if self.dropRate > 0: out = F.dropout(out, p=self.dropRate, training=self.training) out = self.conv2(out) return torch.add(x if self.equalInOut else self.convShortcut(x), out) class NetworkBlock(nn.Module): def __init__(self, nb_layers, in_planes, out_planes, block, stride, dropRate): super(NetworkBlock, self).__init__() self.layer = self._make_layer(block, in_planes, out_planes, nb_layers, stride, dropRate) def _make_layer(self, block, in_planes, out_planes, nb_layers, stride, dropRate): layers = [] for i in range(int(nb_layers)): layers.append(block(i == 0 and in_planes or out_planes, out_planes, i == 0 and stride or 1, dropRate)) return nn.Sequential(*layers) def forward(self, x): return self.layer(x) class WideResNet(nn.Module): def __init__(self, depth, n_classes, n_channels, widen_factor=1, dropRate=0.0): super(WideResNet, self).__init__() nChannels = [16, 16*widen_factor, 32*widen_factor, 64*widen_factor] assert((depth - 4) % 6 == 0) n = (depth - 4) / 6 block = BasicBlock # 1st conv before any network block self.conv1 = nn.Conv2d(n_channels, nChannels[0], kernel_size=3, stride=1, padding=1, bias=False) # 1st block self.block1 = NetworkBlock(n, nChannels[0], nChannels[1], block, 1, dropRate) # 2nd block self.block2 = NetworkBlock(n, nChannels[1], nChannels[2], block, 2, dropRate) # 3rd block self.block3 = NetworkBlock(n, nChannels[2], nChannels[3], block, 2, dropRate) # global average pooling and classifier self.final_bn = nn.BatchNorm2d(nChannels[3], affine=True) self.final_relu = nn.ReLU(inplace=True) self.fc = nn.Linear(nChannels[3], n_classes) self.nChannels = nChannels[3] def forward(self, x): out = self.conv1(x) out = self.block1(out) out = self.block2(out) out = self.block3(out) out = self.final_relu(self.final_bn(out)) out = F.avg_pool2d(out, 8) out = out.view(-1, self.nChannels) return self.fc(out) class MetaWideResNet(ReparamModule): def __init__(self, depth, n_classes, n_channels, widen_factor=1, dropRate=0.0, device='cpu'): super(MetaWideResNet, self).__init__() self.device = device nChannels = [16, 16*widen_factor, 32*widen_factor, 64*widen_factor] assert((depth - 4) % 6 == 0) n = (depth - 4) / 6 block = BasicBlock # 1st conv before any network block self.conv1 = nn.Conv2d(n_channels, nChannels[0], kernel_size=3, stride=1, padding=1, bias=False) # 1st block self.block1 = NetworkBlock(n, nChannels[0], nChannels[1], block, 1, dropRate) # 2nd block self.block2 = NetworkBlock(n, nChannels[1], nChannels[2], block, 2, dropRate) # 3rd block self.block3 = NetworkBlock(n, nChannels[2], nChannels[3], block, 2, dropRate) # global average pooling and classifier self.final_bn = nn.BatchNorm2d(nChannels[3], affine=True) self.final_relu = nn.ReLU(inplace=True) self.fc = nn.Linear(nChannels[3], n_classes) self.nChannels = nChannels[3] def forward(self, x): out = self.conv1(x) out = self.block1(out) out = self.block2(out) out = self.block3(out) out = self.final_relu(self.final_bn(out)) out = F.avg_pool2d(out, 8) out = out.view(-1, self.nChannels) return self.fc(out) if __name__ == '__main__': import time from torchsummary import summary from utils.helpers import * set_torch_seeds(0) x = torch.FloatTensor(2, 3, 32, 32).uniform_(0, 1) ## Test normal WRN model = WideResNet(depth=40, widen_factor=2, n_channels=3, n_classes=10, dropRate=0.0) t0 = time.time() out = model(x) print(f'time for normal fw pass: {time.time() - t0}s') summary(model, (3, 32, 32)) ## Test meta WRN model = MetaWideResNet(depth=40, widen_factor=2, n_channels=3, n_classes=10, device='cpu') weights = model.get_param() t0 = time.time() out = model.forward_with_param(x, weights) print(f'time for meta fw pass: {time.time() - t0}s') summary(model, (3, 32, 32))
5,745
38.627586
116
py
FDS
FDS-main/models/lenet.py
import torch import torch.nn as nn import torch.nn.functional as F from models.meta_factory import ReparamModule from models.helpers import * class Flatten(nn.Module): """ NN module version of torch.nn.functional.flatten """ def __init__(self): super().__init__() def forward(self, input): return torch.flatten(input, start_dim=1, end_dim=-1) class LeNet(nn.Module): def __init__(self, n_classes, n_channels, im_size): super(LeNet, self).__init__() assert im_size in [28, 32] h = 16*5*5 if im_size==32 else 16*4*4 self.n_classes = n_classes self.n_channels = n_channels self.im_size = im_size self.layers = nn.Sequential( nn.Conv2d(n_channels, 6, kernel_size=5, stride=1, padding=0), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2, padding=0), nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2, padding=0), Flatten(), nn.Linear(h, 120), nn.ReLU(inplace=True), nn.Linear(120, 84), nn.ReLU(inplace=True), nn.Linear(84, n_classes)) def forward(self, x): return self.layers(x) class MetaLeNet(ReparamModule): def __init__(self, n_classes, n_channels, im_size, device='cpu'): super(MetaLeNet, self).__init__() assert im_size in [28, 32] h = 16*5*5 if im_size==32 else 16*4*4 self.n_classes = n_classes self.n_channels = n_channels self.im_size = im_size self.device = device # must be defined for parent class self.layers = nn.Sequential( nn.Conv2d(n_channels, 6, kernel_size=5, stride=1, padding=0), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2, padding=0), nn.Conv2d(6, 16, kernel_size=5, stride=1, padding=0), nn.ReLU(inplace=True), nn.MaxPool2d(kernel_size=2, stride=2, padding=0), Flatten(), nn.Linear(h, 120), nn.ReLU(inplace=True), nn.Linear(120, 84), nn.ReLU(inplace=True), nn.Linear(84, n_classes)) def forward(self, x): return self.layers(x) if __name__ == '__main__': import time from torchsummary import summary from utils.helpers import * set_torch_seeds(0) x = torch.FloatTensor(256, 3, 32, 32).uniform_(0, 1) ## Test meta LeNet model = MetaLeNet(n_classes=10, n_channels=3, im_size=32, device='cpu') weights = model.get_param() t0 = time.time() out = model.forward_with_param(x, weights) print(f'time for meta fw pass: {time.time() - t0}s') summary(model, (3, 32, 32))
2,829
25.203704
75
py
FDS
FDS-main/models/meta_factory.py
""" This is a slim version of the code from https://github.com/SsnL/dataset-distillation """ import torch import torchvision import logging import torch.nn as nn import torch.nn.functional as F import functools import math import types from contextlib import contextmanager from torch.optim import lr_scheduler from six import add_metaclass from itertools import chain from copy import deepcopy from models.helpers import * class MetaFactory(type): def __call__(cls, *args, **kwargs): r"""Called when you call ReparamModule(...) """ net = type.__call__(cls, *args, **kwargs) # collect weight (module, name) pairs # flatten weights w_modules_names = [] for m in net.modules(): for n, p in m.named_parameters(recurse=False): if p is not None: w_modules_names.append((m, n)) for n, b in m.named_buffers(recurse=False): if b is not None: pass # logging.warning(( # '{} contains buffer {}. The buffer will be treated as ' # 'a constant and assumed not to change during gradient ' # 'steps. If this assumption is violated (e.g., ' # 'BatcHNorm*d\' running_mean/var), the computation will ' # 'be incorrect.').format(m.__class__.__name__, n)) net._weights_module_names = tuple(w_modules_names) # Put to correct device before we do stuff on parameters #net = net.to(device) ws = tuple(m._parameters[n].detach() for m, n in w_modules_names) assert len(set(w.dtype for w in ws)) == 1 # reparam to a single flat parameter net._weights_numels = tuple(w.numel() for w in ws) net._weights_shapes = tuple(w.shape for w in ws) with torch.no_grad(): flat_w = torch.cat([w.reshape(-1) for w in ws], 0) # remove old parameters, assign the names as buffers for m, n in net._weights_module_names: delattr(m, n) m.register_buffer(n, None) # register the flat one net.register_parameter('flat_w', nn.Parameter(flat_w, requires_grad=True)) return net @add_metaclass(MetaFactory) class ReparamModule(nn.Module): """ Make an architecture inherit this class instead of nn.Module to allow .forward_with_params() This changes state_dict() to a one value dict containing 'flat_w' This requires self.device to be defined in the module """ def _apply(self, *args, **kwargs): rv = super(ReparamModule, self)._apply(*args, **kwargs) return rv def get_param(self, clone=False): if clone: return self.flat_w.detach().clone().requires_grad_(self.flat_w.requires_grad).to(device=self.device) return self.flat_w.to(device=self.device) @contextmanager def unflatten_weight(self, flat_w): """ This changes self.state_dict() from --> odict_keys(['flat_w']) to --> odict_keys(['flat_w', 'layers.0.weight', 'layers.0.bias', ... ] Somehow removes 'bias=False' in self._weights_module_names conv names, and replaces 'bias=False' by 'bias=True' in linear layers. type(self.state_dict()) = <class 'collections.OrderedDict'> before and after """ ws = (t.view(s) for (t, s) in zip(flat_w.split(self._weights_numels), self._weights_shapes)) for (m, n), w in zip(self._weights_module_names, ws): setattr(m, n, w) yield for m, n in self._weights_module_names: setattr(m, n, None) def forward_with_param(self, inp, new_w): #print(type(self.state_dict())) with self.unflatten_weight(new_w): # print('FLATTENED') # print('state_dict: ', type(self.state_dict()), [(k, v.shape) for k,v in self.state_dict().items()]) # print('self._weights_module_names: ', self._weights_module_names) return nn.Module.__call__(self, inp) def __call__(self, inp): return self.forward_with_param(inp, self.flat_w) def load_state_dict(self, state_dict, *args, **kwargs): """ Make load_state_dict work on both singleton dicts containing a flattened weight tensor and full dicts containing unflattened weight tensors. Useful when loading weights from non-meta architectures """ if len(state_dict) == 1 and 'flat_w' in state_dict: return super(ReparamModule, self).load_state_dict(state_dict, *args, **kwargs) with self.unflatten_weight(self.flat_w): flat_w = self.flat_w del self.flat_w super(ReparamModule, self).load_state_dict(state_dict, *args, **kwargs) self.register_parameter('flat_w', flat_w) def unflattened_weights(self): #print(float(torch.sum(self.state_dict()['flat_w']))) with self.unflatten_weight(self.flat_w): state_dict = deepcopy(self.state_dict()) del state_dict['flat_w'] return state_dict def layer_names(self): layer_names = [] layer_count = 0 prev_layer = None for (name, n) in zip(self._weights_module_names, self._weights_numels): if name[0] != prev_layer: layer_count += 1 prev_layer = name[0] if isinstance(name[0], torch.nn.Conv2d) and name[1]=='weight': layer_names.append('L{}_conv_W_s{}'.format(layer_count, n)) elif isinstance(name[0], torch.nn.Conv2d) and name[1]=='bias': layer_names.append('L{}_conv_b_s{}'.format(layer_count, n)) elif isinstance(name[0], torch.nn.BatchNorm2d) and name[1]=='weight': layer_names.append('L{}_bn_W_s{}'.format(layer_count, n)) elif isinstance(name[0], torch.nn.BatchNorm2d) and name[1]=='bias': layer_names.append('L{}_bn_b_s{}'.format(layer_count, n)) elif isinstance(name[0], torch.nn.Linear) and name[1]=='weight': layer_names.append('L{}_fc_W_s{}'.format(layer_count, n)) elif isinstance(name[0], torch.nn.Linear) and name[1]=='bias': layer_names.append('L{}_fc_b_s{}'.format(layer_count, n)) else: raise ValueError('Unknown layer type {}'.format(name)) return layer_names def get_bn_masks(self): """ Returns 2 boolean masks of size n_weights, where ones correspond to batchnorm gammas in first mask, and batchnorm betas in second mask """ gammas_mask = torch.zeros(self.flat_w.shape[0], dtype=torch.bool) betas_mask = torch.zeros(self.flat_w.shape[0], dtype=torch.bool) i = 0 for (name, n) in zip(self._weights_module_names, self._weights_numels): is_BN = isinstance(name[0], torch.nn.BatchNorm2d) or isinstance(name[0], torch.nn.BatchNorm1d) if is_BN and name[1]=='weight': gammas_mask[i:i+n] = 1 elif is_BN and name[1]=='bias': betas_mask[i:i+n] = 1 i += n return gammas_mask, betas_mask def flattened_unflattened_weights(self): """ somehow unflattening weights changes the value of their sum. This looks like it's because permutation matters in float 32 sum operation and so different data structures give different results to the same operations even though they contain the same values. Here unflattening and reflattening recovers the sum value of the original self.get_param() method. """ with self.unflatten_weight(self.flat_w): state_dict = deepcopy(self.state_dict()) del state_dict['flat_w'] flat_w = torch.cat([w.reshape(-1) for w in state_dict.values()], 0) #.type(torch.DoubleTensor) doesn't change behaviour return flat_w def initialize(self, init_type='xavier', init_param=1, init_norm_weights=1, inplace=True): if inplace: flat_w = self.flat_w else: flat_w = torch.empty_like(self.flat_w).requires_grad_() with torch.no_grad(): with self.unflatten_weight(flat_w): initialize(self, init_type=init_type, init_param=init_param, init_norm_weights=init_norm_weights) return flat_w
8,482
35.722944
127
py
FDS
FDS-main/models/helpers.py
import torch.nn as nn from torch.nn import init def initialize(net, init_type, init_param, init_norm_weights=1): """ various initialization schemes """ def init_func(m): classname = m.__class__.__name__ if classname.startswith('Conv') or classname == 'Linear': if getattr(m, 'bias', None) is not None: init.constant_(m.bias, 0.0) #if init_type = default bias isn't kept to zero if getattr(m, 'weight', None) is not None: if init_type == 'normal': init.normal_(m.weight, 0.0, init_param) elif init_type == 'xavier': init.xavier_normal_(m.weight, gain=init_param) elif init_type == 'xavier_unif': init.xavier_uniform_(m.weight, gain=init_param) elif init_type == 'kaiming': init.kaiming_normal_(m.weight, a=init_param, mode='fan_in') elif init_type == 'kaiming_out': init.kaiming_normal_(m.weight, a=init_param, mode='fan_out') elif init_type == 'orthogonal': init.orthogonal_(m.weight, gain=init_param) elif init_type == 'default': if hasattr(m, 'reset_parameters'): m.reset_parameters() else: raise NotImplementedError('initialization method [%s] is not implemented' % init_type) elif 'Norm' in classname: #different Pytorch versions differ in BN init so do it manually if getattr(m, 'weight', None) is not None: m.weight.data.fill_(init_norm_weights) if getattr(m, 'bias', None) is not None: m.bias.data.zero_() net.apply(init_func) return net
1,798
43.975
106
py
FDS
FDS-main/models/selector.py
from models.lenet import * from models.wresnet import * def select_model(meta, dataset, architecture, init_type='xavier', init_param=1, device='cpu'): """ Meta models require device to be provided during init. """ if dataset in ['MNIST', 'FashionMNIST']: n_classes, n_channels, im_size = 10, 1, 28 kwargs0 = {'n_classes':n_classes, 'n_channels':n_channels, 'im_size':im_size} if architecture == 'LeNet': model = MetaLeNet(**kwargs0, device=device) if meta else LeNet(**kwargs0) elif architecture == 'LeNet-BN': #debug neg learning rates model = MetaLeNetBN(**kwargs0, device=device) if meta else LeNetBN(**kwargs0) else: raise NotImplementedError elif dataset in ['SVHN', 'CIFAR10', 'CIFAR100']: n_channels, im_size = 3, 32 n_classes = 100 if dataset == 'CIFAR100' else 10 kwargs0 = {'n_classes':n_classes, 'n_channels':n_channels} if architecture == 'LeNet': kwargs1 = {'im_size':im_size} model = MetaLeNet(**kwargs0, **kwargs1, device=device) if meta else LeNet(**kwargs0, **kwargs1) elif architecture == 'LeNetBN': kwargs1 = {'im_size':im_size} model = MetaLeNetBN(**kwargs0, **kwargs1, device=device) if meta else LeNetBN(**kwargs0, **kwargs1) elif architecture == 'WRN-10-1': kwargs1 = {'depth':10, 'widen_factor':1, 'dropRate':0.0} model = MetaWideResNet(**kwargs0, **kwargs1, device=device) if meta else WideResNet(**kwargs0, **kwargs1) elif architecture == 'WRN-16-1': kwargs1 = {'depth':16, 'widen_factor':1, 'dropRate':0.0} model = MetaWideResNet(**kwargs0, **kwargs1, device=device) if meta else WideResNet(**kwargs0, **kwargs1) elif architecture == 'WRN-40-2': kwargs1 = {'depth':40, 'widen_factor':2, 'dropRate':0.0} model = MetaWideResNet(**kwargs0, **kwargs1, device=device) if meta else WideResNet(**kwargs0, **kwargs1) else: raise NotImplementedError else: raise NotImplementedError ## Initialization schemes if meta: model.initialize(init_type=init_type, init_param=init_param, init_norm_weights=1, inplace=True) else: initialize(model, init_type=init_type, init_param=init_param, init_norm_weights=1) return model if __name__ == '__main__': from torchsummary import summary from utils.helpers import * ## Check meta and normal models do the same calculations # x1 = torch.FloatTensor(64, 3, 32, 32).uniform_(0, 1) # x2 = torch.FloatTensor(64, 3, 32, 32).uniform_(0, 1) # set_torch_seeds(0) # model = select_model(False, dataset='CIFAR10', architecture='WRN-10-1', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False) # set_torch_seeds(0) # meta_model = select_model(True, dataset='CIFAR10', architecture='WRN-10-1', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False) # meta_weights = meta_model.get_param() # # model.train(), meta_model.train() #x1 before and after x2 if comment out eval mode below. # # model_output = model(x1) # meta_model_output = meta_model.forward_with_param(x1, meta_weights) # print(float(torch.sum(model_output)), float(torch.sum(meta_model_output))) # # model_output = model(x2) # meta_model_output = meta_model.forward_with_param(x2, meta_weights) # print(float(torch.sum(model_output)), float(torch.sum(meta_model_output))) # # model.eval(), meta_model.eval() #x1 output changes in eval now because running stats were calculated # model_output = model(x1) # meta_model_output = meta_model.forward_with_param(x1, meta_weights) # print(float(torch.sum(model_output)), float(torch.sum(meta_model_output))) # x = torch.FloatTensor(64, 3, 32, 32).uniform_(0, 1) # # t0 = time.time() # model = select_model(True, dataset='CIFAR10', architecture='WRN-16-1', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False) # output = model(x) # print("Time taken for forward pass: {} s".format(time.time.time() - t0)) # print("\nOUTPUT SHAPE: ", output.shape) # summary(model, (3, 32, 32), max_depth=5) ## Weights init for normal model # model = select_model(False, dataset='CIFAR10', architecture='WRN-16-1', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False) # def weights_to_gaussian(m): # if isinstance(m, nn.Conv2d) or isinstance(m, nn.Linear): #TODO separate init for linear layer? # torch.nn.init.normal_(m.weight, mean=0, std=0.1) # if m.bias is not None: # torch.nn.init.zeros_(m.bias) # model.apply(weights_to_gaussian) ## Weights init for meta model # model = select_model(True, dataset='CIFAR10', architecture='WRN-16-1', activation='ReLU', norm_type='BN', norm_affine=False, noRes=False) # weights = model.get_param() # print(len(weights)) # print(torch.sum(weights)) # torch.nn.init.normal_(weights, mean=0, std=0.1) # print(torch.sum(weights)) ## Change BN init for meta model #model = select_model(False, dataset='CIFAR10', architecture='LeNet', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False) #model = select_model(True, dataset='CIFAR10', architecture='LeNet', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False) #summary(model, (3, 32, 32), max_depth=10) # weights = model.get_param() # #weights_numels = model._weights_numels # #layer_names = model.layer_names() # #print(weights.shape[0], sum(weights_numels)) #same # #print(len(weights_numels), len(layer_names)) #same # gammas_mask, betas_mask = model.get_bn_masks() # print(len(weights[gammas_mask]), len(weights[betas_mask])) # #print(weights[gammas_mask]) # print(weights[betas_mask]) ## Check init set_torch_seeds(0) model = select_model(False, dataset='CIFAR10', architecture='ShuffleNetv2-s05', activation='ReLU', norm_type='BN', norm_affine=True, noRes=False, init_type='normal', init_param=1, init_norm_weights=1) for n,p in model.named_parameters(): print(n, float(torch.sum(p)))
6,359
45.764706
150
py
FDS
FDS-main/utils/datasets.py
import torch import torch.nn as nn import torch.nn.functional as F import torchvision.transforms as transforms import torchvision.datasets as datasets from torch.utils.data import Dataset, DataLoader from torch.utils.data.sampler import SubsetRandomSampler import os import math import numpy as np import matplotlib.pyplot as plt import warnings from utils.helpers import * def unormalize_CIFAR10_image(image): return image*torch.tensor([0.2023, 0.1994, 0.2010]).view(3,1,1) + torch.tensor([0.4914, 0.4822, 0.4465]).view(3,1,1) def plot_image(input, unormalize=True): if len(input.shape) > 3: print("Use plot_images function instead!") raise NotImplementedError npimg = input.numpy() if unormalize: npimg = npimg * np.array([0.2023, 0.1994, 0.2010]).reshape(3,1,1) + np.array([0.4914, 0.4822, 0.4465]).reshape(3,1,1) npimg = np.transpose(npimg, (1, 2, 0)) if npimg.shape[-1] != 3: npimg = npimg[:, :, 0] #print(npimg.shape) fig = plt.figure(figsize=(20, 20)) ax = fig.add_subplot(111) ax.axis('off') ax.set_xticklabels([]) ax.set_yticklabels([]) ax.imshow(npimg, cmap='gray') plt.show() return fig def plot_images(batch, padding=2, unormalize=True): if len(batch.shape) == 3: plot_image(batch, unormalize=unormalize) elif len(batch.shape) == 4: n_images = batch.shape[0] if n_images == 1: plot_image(batch[0], unormalize=unormalize) else: grid_img = torchvision.utils.make_grid(batch, nrow=int(np.ceil(np.sqrt(n_images))), padding=padding) plot_image(grid_img, unormalize=unormalize) class Cutout(object): def __init__(self, length, prob=1.0): self.length = length self.prob = prob assert prob<=1, f"Cutout prob given ({prob}) must be <=1" def __call__(self, img): if np.random.binomial(1, self.prob): h, w = img.size(1), img.size(2) mask = np.ones((h, w), np.float32) y = np.random.randint(h) x = np.random.randint(w) y1 = np.clip(y - self.length // 2, 0, h) y2 = np.clip(y + self.length // 2, 0, h) x1 = np.clip(x - self.length // 2, 0, w) x2 = np.clip(x + self.length // 2, 0, w) mask[y1: y2, x1: x2] = 0. mask = torch.from_numpy(mask) mask = mask.expand_as(img) img *= mask return img class InfiniteDataLoader(DataLoader): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) self.dataset_iterator = super().__iter__() def __iter__(self): return self def __next__(self): try: batch = next(self.dataset_iterator) except StopIteration: self.dataset_iterator = super().__iter__() batch = next(self.dataset_iterator) return batch def get_loaders(datasets_path, dataset, train_batch_size=128, val_batch_size=128, val_source='train', val_train_fraction=0.1, val_train_overlap=False, workers=0, train_infinite=False, val_infinite=False, cutout=False, cutout_length=16, cutout_prob=1): """ NB: val_train_fraction and val_train_overlap only used if val_source='train' Note that infinite=True changes the seed/order of the batches Validation is never augmented since validation stochasticity comes from sampling different validation images anyways """ assert val_source in ['test', 'train'] TrainLoader = InfiniteDataLoader if train_infinite else DataLoader ValLoader = InfiniteDataLoader if val_infinite else DataLoader ## Select relevant dataset if dataset in ['MNIST', 'FashionMNIST']: mean, std = (0.1307,), (0.3081,) transform_train = transforms.Compose([transforms.ToTensor(), transforms.Normalize(mean, std)]) if cutout: transform_train.transforms.append(Cutout(length=cutout_length, prob=cutout_prob)) transform_test = transforms.Compose([transforms.ToTensor(), transforms.Normalize(mean, std)]) if dataset == 'MNIST': train_dataset = datasets.MNIST(datasets_path, train=True, download=True, transform=transform_train) test_dataset = datasets.MNIST(datasets_path, train=False, download=True, transform=transform_test) val_dataset = test_dataset if val_source=='test' else datasets.MNIST(datasets_path, train=True, download=True, transform=transform_test) elif dataset == 'FashionMNIST': train_dataset = datasets.FashionMNIST(datasets_path, train=True, download=True, transform=transform_train) test_dataset = datasets.FashionMNIST(datasets_path, train=False, download=True, transform=transform_test) val_dataset = test_dataset if val_source=='test' else datasets.FashionMNIST(datasets_path, train=True, download=True, transform=transform_test) elif dataset == 'SVHN': mean = (0.4377, 0.4438, 0.4728) std = (0.1980, 0.2010, 0.1970) dataset_path = os.path.join(datasets_path, 'SVHN') #Pytorch is inconsistent in folder structure transform_train = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean, std)]) if cutout: transform_train.transforms.append(Cutout(length=cutout_length, prob=cutout_prob)) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean, std)]) train_dataset = datasets.SVHN(dataset_path, split='train', download=True, transform=transform_train) test_dataset = datasets.SVHN(dataset_path, split='test', download=True, transform=transform_test) val_dataset = test_dataset if val_source=='test' else datasets.SVHN(dataset_path, split='train', download=True, transform=transform_test) #print(len(train_dataset)) elif dataset in ['CIFAR10', 'CIFAR100']: # official CIFAR10 std seems to be wrong (actual is [0.2470, 0.2435, 0.2616]) mean = (0.4914, 0.4822, 0.4465) if dataset == 'CIFAR10' else (0.5071, 0.4867, 0.4408) std = (0.2023, 0.1994, 0.2010) if dataset == 'CIFAR10' else (0.2675, 0.2565, 0.2761) transform_train = transforms.Compose([ transforms.RandomCrop(32, padding=4), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean, std)]) if cutout: transform_train.transforms.append(Cutout(length=cutout_length, prob=cutout_prob)) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize(mean, std)]) if dataset == 'CIFAR10': dataset_path = os.path.join(datasets_path, 'CIFAR10') #Pytorch is inconsistent in folder structure train_dataset = datasets.CIFAR10(dataset_path, train=True, download=True, transform=transform_train) test_dataset = datasets.CIFAR10(dataset_path, train=False, download=True, transform=transform_test) val_dataset = test_dataset if val_source=='test' else datasets.CIFAR10(datasets_path, train=True, download=True, transform=transform_test) elif dataset == 'CIFAR100': dataset_path = os.path.join(datasets_path, 'CIFAR100') train_dataset = datasets.CIFAR100(dataset_path, train=True, download=True, transform=transform_train) test_dataset = datasets.CIFAR100(dataset_path, train=False, download=True, transform=transform_test) val_dataset = test_dataset if val_source=='test' else datasets.CIFAR10(datasets_path, train=True, download=True, transform=transform_test) else: print(f'{dataset} is not implemented') raise NotImplementedError ## Create dataloaders n_train_images = len(train_dataset) #print(train_dataset) pin_memory = True if dataset == 'ImageNet' else False if val_source == 'test': train_loader = TrainLoader( dataset=train_dataset, batch_size=train_batch_size, shuffle=True, drop_last=True, num_workers=workers, pin_memory=pin_memory) val_loader = ValLoader( dataset=val_dataset, batch_size=val_batch_size, shuffle=True, drop_last=True, num_workers=workers, pin_memory=pin_memory) elif val_source == 'train': all_indices = list(range(n_train_images)) val_indices = np.random.choice(all_indices, size=int(val_train_fraction * n_train_images), replace=False) val_loader = ValLoader( dataset=val_dataset, batch_size=val_batch_size, sampler=SubsetRandomSampler(val_indices), drop_last=True, num_workers=workers, pin_memory=pin_memory) if val_train_overlap: train_loader = TrainLoader( dataset=train_dataset, batch_size=train_batch_size, shuffle=True, drop_last=True, num_workers=workers, pin_memory=pin_memory) else: train_indices = list(set(all_indices) - set(val_indices)) train_loader = TrainLoader( dataset=train_dataset, batch_size=train_batch_size, sampler=SubsetRandomSampler(train_indices), drop_last=True, num_workers=workers, pin_memory=pin_memory) test_loader = DataLoader( dataset=test_dataset, batch_size=val_batch_size, shuffle=True, drop_last=True, num_workers=workers, pin_memory=pin_memory) # test loader never infinite return train_loader, val_loader, test_loader if __name__ == '__main__': train_loader, val_loader, test_loader = get_loaders('~/Datasets/Pytorch/', 'MNIST', train_batch_size=500, val_batch_size=500, val_source='train', val_train_fraction=0.05, val_train_overlap=False, workers=0, train_infinite=False, val_infinite=False, cutout=True, cutout_length=16, cutout_prob=1) print(len(train_loader)*500) print(len(val_loader)*500) for x_val, y_val in val_loader: print(x_val.shape) for x_train, y_train in train_loader: break #plot_images(x_val[:100]) plot_images(x_train[:100])
10,895
41.232558
155
py
FDS
FDS-main/utils/helpers.py
import csv import torch import torch.nn.functional as F from torchvision import datasets, transforms from torch.utils.data import Dataset, DataLoader import shutil import datetime import json import os import argparse import gc import numpy as np import torchvision import functools import time import warnings #warnings.simplefilter("ignore", UserWarning) ### Metrics class AggregateTensor(object): """ Computes and stores the average of stream. Mostly used to average losses and accuracies. Works for both scalars and vectors but input needs to be a pytorch tensor. """ def __init__(self): self.reset() def reset(self): self.count = 0.0001 # DIV/0! self.sum = 0 #self.sum2 = 0 def update(self, val, w=1): """ :param val: new running value :param w: weight, e.g batch size Turn everything into floats so that we don't keep bits of the graph """ self.sum += w * val.detach() self.count += w def avg(self): return self.sum / self.count # def std(self): # return np.sqrt(self.sum2/self.count - self.avg()**2) def accuracy(output, target, topk=(1,)): """Computes the precision@k for the specified values of k""" maxk = max(topk) batch_size = target.size(0) _, pred = output.topk(maxk, 1, True, True) pred = pred.t() correct = pred.eq(target.view(1, -1).expand_as(pred)) res = [] for k in topk: correct_k = correct[:k].view(-1).float().sum(0) res.append(correct_k.mul_(1/batch_size)) return res def avg_entropy(pmf): """ :param pmf: pytorch tensor pmf of shape [batch_size, n_classes] :return: average entropy of pmf across entire batch """ #assert assert ((pmf>=0)*(pmf<=1.00001)).all(), "All inputs must be in range [0,1] but min/max is {}/{}".format(float(torch.min(pmf)), float(torch.max(pmf))) p_log_p = torch.log2(torch.clamp(pmf, min=0.0001, max=1.0))*pmf #log(0) causes error return torch.mean(-p_log_p.sum(1)) def avg_max(pmf): """ :param pmf: pytorch tensor pmf of shape [batch_size, n_classes] when learned the pmf doesn't have to be within [0,1] :return: average of max predictions of pmf across entire batch """ assert ((pmf >= 0) * (pmf <= 1)).all(), "All inputs must be in range [0,1]" return torch.mean(torch.max(pmf, 1)[0]) def onehot(targets, n_classes): """ Convert labels of form [[2], [7], ...] to [0,0,1,0,0,0,0,0,0,0], [0,0,0,0,0,0,0,0,1,0,0], ...] :param targets: :param n_classes: :param device: :return: """ return torch.zeros((targets.shape[0], n_classes), device=targets.device).scatter(1, targets.unsqueeze(-1), 1) def gc_tensor_view(verbose=True): """ Doesn't catch intermediate variables stored by Pytorch graph if they are not in the Python scope. assumes all tensors are torch.float() i.e. 32 bit (4MB) """ total_MB_size = 0 object_counts = {} object_MBs = {} if verbose: print('\n------- TENSORS SEEN BY GARBAGE COLLECTOR -------') for obj in gc.get_objects(): try: if torch.is_tensor(obj) or (hasattr(obj, 'data') and torch.is_tensor(obj.data)): MB_size = np.prod(obj.size()) * 4 / 1024**2 #assume float32 total_MB_size += MB_size #str(type(obj)) key = str(obj.size())[10:] object_counts[key] = object_counts.get(key, 0) + 1 object_MBs[key] = MB_size except: pass if verbose: object_totals = {k:object_counts[k] * object_MBs[k] for k in object_MBs.keys()} for key, value in sorted(object_totals.items(), key=lambda item: item[1], reverse=True): print("{} x {} ({:.0f}MB) = {:.0f}MB".format(object_counts[key], key, object_MBs[key], object_counts[key]*object_MBs[key])) print("TOTAL MEMORY USED BY PYTORCH TENSORS: {:.0f} MB".format(total_MB_size)) def set_torch_seeds(seed): import random import numpy as np import torch random.seed(seed) np.random.seed(seed) torch.manual_seed(seed) torch.cuda.manual_seed(seed) def timer(func): """Print the runtime of the decorated function""" @functools.wraps(func) def wrapper_timer(*args, **kwargs): start_time = time.perf_counter() # 1 value = func(*args, **kwargs) end_time = time.perf_counter() # 2 run_time = end_time - start_time # 3 #print('\n------------------------------') print(f"--- Ran func {func.__name__!r} in {format_time(run_time)} ---") #print('------------------------------\n') return value return wrapper_timer ### Data view and read def unormalize_CIFAR10_image(image): return image*torch.tensor([0.2023, 0.1994, 0.2010]).view(3,1,1) + torch.tensor([0.4914, 0.4822, 0.4465]).view(3,1,1) # def plot_image(input, unormalize=False): # if len(input.shape) > 3: # print("Use plot_images function instead!") # raise NotImplementedError # npimg = input.numpy() # if unormalize: # npimg = npimg * np.array([0.2023, 0.1994, 0.2010]).reshape(3,1,1) + np.array([0.4914, 0.4822, 0.4465]).reshape(3,1,1) # npimg = np.transpose(npimg, (1, 2, 0)) # if npimg.shape[-1] != 3: # npimg = npimg[:, :, 0] # #print(npimg.shape) # # fig = plt.figure(figsize=(20, 20)) # ax = fig.add_subplot(111) # ax.axis('off') # ax.set_xticklabels([]) # ax.set_yticklabels([]) # # ax.imshow(npimg, cmap='gray') # plt.show() # return fig # def plot_images(batch, padding=2, unormalize=False): # if len(batch.shape) == 3: # plot_image(batch, unormalize=unormalize) # elif len(batch.shape) == 4: # n_images = batch.shape[0] # if n_images == 1: # plot_image(batch[0], unormalize=unormalize) # else: # grid_img = torchvision.utils.make_grid(batch, nrow=int(np.ceil(np.sqrt(n_images))), padding=padding) # plot_image(grid_img, unormalize=unormalize) def str2bool(v): # codes from : https://stackoverflow.com/questions/15008758/parsing-boolean-values-with-argparse if v.lower() in ('yes', 'true', 't', 'y', '1'): return True elif v.lower() in ('no', 'false', 'f', 'n', '0'): return False else: raise argparse.ArgumentTypeError('Boolean value expected.') def delete_files_from_name(folder_path, file_name, type='contains'): assert type in ['is', 'contains'] for f in os.listdir(folder_path): if (type=='is' and file_name==f) or (type=='contains' and file_name in f): os.remove(os.path.join(folder_path, f)) def copy_file(file_path, folder_path): destination_path = os.path.join(folder_path, os.path.basename(file_path)) shutil.copyfile(file_path, destination_path) def format_time(seconds): minutes, seconds = divmod(seconds, 60) hours, minutes = divmod(minutes, 60) return "%dh%02dm%02ds" % (hours, minutes, seconds) def create_empty_file(path): """Easy way to log final test accuracy in some experiment folder""" with open(path, 'w+') as f: f.write("NA") if __name__ == '__main__': from time import time import torch ## Test AggregateTensor # x = np.random.rand(1000)*50 # w = np.random.rand(1000)*5 # true_mu = w@x/np.sum(w) # true_std = np.sqrt(np.sum(w*(x-true_mu)**2)/((len(x)-1)*np.sum(w)/len(x))) # # t0 = time.time() # a = "yolo" # print("Init of string takes: {} us".format(1e6*(time()-t0))) # # t0 = time.time() # meter = AggregateTensor() # print("Init of AggregateTensor takes: {} us".format(1e6*(time()-t0))) # # t0 = time.time() # a = 1000*5 # print("Multiplication takes: {} us".format(1e6 * (time.time() - t0))) # # t = 0 # for val,weight in zip(x,w): # t0 = time.time() # meter.update(val, weight) # t += time.time() - t0 # print("Avg update time: {} us".format(1e6*t/len(x))) # # print(true_mu, meter.avg()) # #print(np.std(x), true_std, meter.std()) # # ### Test AggregateDict # keys = ['loss', 'acc', 'yolo'] # meter = AggregateDict() # # values = [[1,2,3], [3,4,5], [1,1,1]] # true_mus = [np.mean(el) for el in values] # true_stds = [np.std(el) for el in values] # # for i in range(3): # dict = {k: v[i] for k,v in zip(keys, values)} # print(dict) # meter.update(val=dict, w=1) # # # print(true_mus, meter.avg()) #print(true_stds, meter.std()) ### Test cutout # Data loader tests # from time import time # import torchvision.datasets as datasets # import torchvision.transforms as transforms # # device = torch.device('cpu') # dataset_path = "~/Datasets/Pytorch/" # # # transform = transforms.Compose([ # transforms.RandomCrop(32, padding=4), # transforms.RandomHorizontalFlip(), # transforms.ToTensor(), # #transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), # Cutout(n_holes=1, length=16, cutout_proba=0.5)]) # # dataset = datasets.CIFAR10(dataset_path, train=False, download=True, transform=transform) # loader = torch.utils.data.DataLoader( # dataset=dataset, # batch_size=5, # shuffle=False, drop_last=False, num_workers=4) # # for x,y in loader: # print(x.shape, y.shape) # image = x[4]#*torch.Tensor[0.2023, 0.1994, 0.2010])-torch.Tensor([0.4914, 0.4822, 0.4465] # plot_image(image) # break ### Test entropy # output = torch.Tensor([[0.1, 0.5, 0.4], # [0.3,0.3,0.4], # [0.99, 0.005, 0.005], # [0.5, 0.5, 0.000001]]) # # print(avg_entropy(output)) ## Test Dataloader indices # transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]) # dataset = datasets.MNIST("~/Datasets/Pytorch/", train=True, download=True, transform=transform) # dataset = DatasetWithIndices(dataset) # loader = torch.utils.data.DataLoader(dataset=dataset,batch_size=5,shuffle=True,drop_last=True,num_workers=1) # # cnt = 0 # for x, y, indices in loader: # print(x.shape, y.shape, indices.shape) # print(indices) # if cnt>5: # break # cnt+=1 #print(len(dataset), len(loader)) ## Test Dataloader coefficient # transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))]) # dataset = datasets.MNIST("~/Datasets/Pytorch/", train=True, download=True, transform=transform) # dataset = DatasetWithLearnableCoefficients(dataset) # loader = torch.utils.data.DataLoader(dataset=dataset,batch_size=5,shuffle=True,drop_last=True,num_workers=1) # # cnt = 0 # for x, y, indices in loader: # print(x.shape, y.shape, indices.shape) # print(indices) # if cnt>5: # break # cnt+=1 # # print(len(dataset), len(loader)) ## Test Corrupter # corrupter = Corrupter(n_images=50000, fraction_to_corrupt=0.1, n_classes=10) # indices = torch.arange(10000, 10000+260, dtype=torch.long) # targets = torch.arange(10, dtype=torch.long).repeat(26) # # t0 = time.time() # corrupted = corrupter(indices, targets) # print(time.time() - t0) # # print(corrupted) # print(len(corrupted)) ## Test Aggregate for vector #a = torch.tensor([1,2,3]) #b = torch.tensor([3,4,5]) a = torch.FloatTensor([1]) b = torch.FloatTensor([2]) c = AggregateVector() c.update(a) c.update(b) print(c.avg()) ### pass
11,853
30.442971
153
py
corrupted_data_classification
corrupted_data_classification-main/main.py
# -*- coding: utf-8 -*- ''' The following libraries are used: [1] NIFTy – Numerical Information Field Theory, https://gitlab.mpcdf.mpg.de/ift/nifty [2] NumPy - Numerical Python, https://numpy.org/ [3] Tensorflow - Tensorflow, https://www.tensorflow.org/ [4] Keras - Keras, https://keras.io/ [5] Matplotlib - Matplotlib, https://matplotlib.org/ [6] SciPy - Scientific Python, https://www.scipy.org/ [7] random - random, https://docs.python.org/3/library/random.html [8] sklearn - https://scikit-learn.org/ Within helper_functions.py, Conv.py and Mask.py, the following libraries are used (these may be obsolete and omittable for the core task): [9] PIL - Pillow (only Image-function), https://pillow.readthedocs.io/en/stable/ [10] warnings - warnings, https://docs.python.org/3/library/warnings.html [11] random - random, https://docs.python.org/3/library/random.html [12] skimage - scikit-image (only resize-function), https://scikit-image.org All Neural Networks were built with Keras and saved as tensorflow-objects. Neural Netowrks are optimized for MNIST, good performance is observed for F-MNIST. ''' # Commented out IPython magic to ensure Python compatibility. # Colab and system related import os import sys import nifty6 as ift ### # Necessary to convert tensorflow-object (e.g. Neural Network) to Nifty-Operator sys.path.append('corrupted_data_classification/helper_functions/') from operators.tensorflow_operator import TensorFlowOperator ### import tensorflow as tf # Include path to access helper functions and Mask / Conv Operator sys.path.append('corrupted_data_classification/helper_functions/') from helper_functions import clear_axis, gaussian, get_cmap, info_text, get_noise, rotation, split_validation_set import Mask # Masking Operator import Conv # Convolution Operator sys.path.remove # Tensorflow # Plotting import matplotlib as mpl import matplotlib.pyplot as plt import matplotlib.cm as cm import matplotlib.pyplot as plt # %matplotlib inline plt.rcParams['figure.dpi'] = 200 # 200 e.g. is really fine, but slower # Numerics import random import numpy as np from sklearn.neighbors import KernelDensity from scipy.stats import multivariate_normal import sklearn as sk from sklearn import decomposition # Choose dataset dataset = 'mnist' #'mnist, 'fashion_mnist' datasource = getattr(tf.keras.datasets, dataset) (XTrain, YTrain), (XTest, YTest) = datasource.load_data() XTrain, XTest = XTrain / 255.0, XTest / 255.0 x_shape = XTrain[1].shape[0] y_shape = XTrain[1].shape[1] try: z_shape = XTrain[1].shape[2] img_shape = [x_shape, y_shape, z_shape] except: img_shape = [x_shape, y_shape] xy_shape = x_shape * y_shape flattened_shape = np.prod(img_shape) # Reshape Xtrain and XTest to flattened Vectors instead of square arrays if dataset == 'mnist' or dataset== 'fashion_mnist': XTrain = XTrain.reshape((len(XTrain), np.prod(XTrain.shape[1:]))) XTest = XTest.reshape((len(XTest), np.prod(XTest.shape[1:]))) n_classes = len(np.unique(YTrain)) # Session for tensorflow v1 compatibility sess = tf.compat.v1.InteractiveSession() graph = tf.compat.v1.get_default_graph() ### # [4] ### # Split Training-Dataset into additional validation set. XTrain, YTrain, XVal, YVal = split_validation_set(XTrain, YTrain, val_perc=0.2) # Read in model# if dataset=='mnist': Decoder_tf = tf.keras.models.load_model('./corrupted_data_classification/NNs/MNIST/pretrained_supervised_ae10/Decoder', compile=False) Encoder_tf = tf.keras.models.load_model('./corrupted_data_classification/NNs/MNIST/pretrained_supervised_ae10/Encoder', compile=False) if dataset=='fashion_mnist': Decoder_tf = tf.keras.models.load_model('./corrupted_data_classification/NNs/Fashion-MNIST/pretrained_supervised_ae10/Decoder', compile=False) Encoder_tf = tf.keras.models.load_model('./corrupted_data_classification/NNs/Fashion-MNIST/pretrained_supervised_ae10/Encoder', compile=False) # Define ift-space # position_space: Also data-space. Equal to the vectorized image dimension. For MNIST-Images, the position-space's # dimensions are 784x1 position_space = ift.UnstructuredDomain(Decoder_tf.get_layer(index=-1).output_shape[1:]) # n_latent: number of latent space activations n_latent = Encoder_tf.get_layer(index=-1).output_shape[-1] # latent_space: Domain with dimensions of the latent space latent_space = ift.UnstructuredDomain([n_latent]) # Initialize Parameters # Pre-Defined parameters by Max-Planck-Institute comm, _, _, master = ift.utilities.get_MPI_params() # Convert Encoder and Decoder to nifty-operators (``TensorFlowOperator``) Decoder = TensorFlowOperator(Decoder_tf.layers[-1].output, Decoder_tf.layers[0].output, latent_space, position_space) Encoder = TensorFlowOperator(Encoder_tf.layers[-1].output, Encoder_tf.layers[0].output, position_space, latent_space) # Choose how to classify data, once it has been reconstructed (any classifier of MNIST data my be chosen here). Classifier = Encoder #Classifier = TensorFlowOperator(Classifier_tf.layers[-2].output, Classifier_tf.layers[0].output, position_space,ift.UnstructuredDomain(n_classes)) # Get all activations in the latent space from Encoder with Validation Dataset -> latent_values latent_values = np.zeros((len(XVal), n_latent)) for i, pic in enumerate(XVal): pic = np.reshape(pic, position_space.shape) latent_values[i, :] = Encoder(ift.Field.from_raw(position_space, pic)).val # Fill means-array with mean activation of every picture means = np.zeros([n_latent, n_classes]) for pic in range(n_classes): for weight in range(n_latent): means[weight, pic] = np.mean(latent_values[np.where(YVal == pic), weight]) # Define overall mean of all activations in latent-space mean = ift.Field.from_raw(latent_space, np.mean(latent_values, axis=0)) #mean of all activations in latent Mean = ift.Adder(mean) # Fill cov_all_variables with covariances of activation of every digit; # Get cov_supervised_variables with covariances of only supervised activations cov_all_variables = [[np.zeros([n_latent, n_latent])] for y in range(n_classes)] cov_supervised_variables = [[np.zeros([n_classes, n_classes])] for y in range(n_classes)] for i in range(n_classes): cov_all_variables[i] = np.cov(latent_values[np.where(YVal==i)[0]][:,:], rowvar=False) cov_supervised_variables[i] = np.cov(latent_values[np.where(YVal==i)[0]][:,:10], rowvar=False) # Fill overall covariance of all activations in latent space cov = np.zeros([n_latent, n_latent]) cov = np.cov(latent_values, rowvar=False) # Transform covariance matrix into standardized space by Cholesky factorization # cov = AA^T A = ift.MatrixProductOperator(ift.UnstructuredDomain([n_latent]), np.linalg.cholesky(cov)) ''' Generate Ground Truth either --> from Sampling from latent distribution OR --> from drawing a sample from independent partition of dataset ''' ## Sampling from latent distribution #xi = ift.from_random(latent_space, 'normal') #s = A.apply(xi, 1) + mean #ground_truth = Decoder(s) ## Drawing sample from dataset p=3 #p = 10 ground_truth = ift.Field.from_raw(position_space, np.reshape(XTest[p], position_space.shape)) ''' Data Corruption: 1. Mask --> Operator: M (no_mask, half_mask, corner_mask, checkerboard_mask, random_mask) 2. Noise --> Operator: N 3. Convolution --> Operator: C (sobel, gaussian_blur, edge_detection, own) Data Modification (not included in modeling-process; thus the Model "does not know" these modifications): 4. Rotation (angle) X. Response --> Operator: R (Concatenated Mask, Noise and Convolution) ''' p = 10 # Specify element of XTest that is to be corrupted and to be evaluated; can be arbitrary integer within length of XTest ground_truth = ift.Field.from_raw(position_space, np.reshape(XTest[p], position_space.shape)) # 1. Mask M = Mask.no_mask(position_space=position_space) #M = Mask.half_mask(position_space=position_space, mask_range=0.5) #M = Mask.random_mask(position_space=position_space, seed=10, n_blobs=25) # 2. Noise N, n = get_noise(noise_level=1, position_space=position_space, seed=10) # 3. Convolution #C = Conv.gaussian_blur(7, 1, position_space=position_space) # sobel, edge_detection, # 4. Rotation (not included in data-model, reconstruction may be poor!) # Specify angle in degrees (clockwise rotation) ground_truth_rot = rotation(ground_truth, img_shape, angle=0) # Apply Data Corruption to Ground Truth and creeate Response operator GR = ift.GeometryRemover(position_space) R = GR(M) # Without Convolution #R = GR(M @ C) # With Convolution data = R((ground_truth_rot))+n # Apply Response R on (rotated) ground truth --> Noise is applicated after masking plt.imshow(np.reshape(data.val, [28,28])) # Define Hyperparameters for minimizer via Iteration-Controllers # These Hyperparameters are not fully optimized! ic_sampling = ift.AbsDeltaEnergyController(name='Sampling', deltaE=1e-2, iteration_limit=150) ic_newton = ift.AbsDeltaEnergyController(name='Newton', deltaE=5e-2, iteration_limit=150) minimizer = ift.NewtonCG(ic_newton) ''' Define Likelihood as Gaussian Energy mean: data (corruped image with R applied) inverse_covariance: Inverse of Noise-Matrix N R: Response Operator Decoder: Generator mapping data from latent space to image space Mean: Adder Operator; Mean of all latent Space activations A: Product Operator; Transformed Covariance of all latent space activations Mean and A originate from the following transformation: s = A*xi+Mean ''' likelihood = ift.GaussianEnergy(mean=data, inverse_covariance=N.inverse) @ R @ Decoder @ Mean @ A H = ift.StandardHamiltonian(likelihood, ic_sampling) # Run MGVI (Metric Gaussian Variational Inference) n_samples = 50 # Define number of samples with which posterior distribution is approximated; more samples => higher runtime, higher accuracy def MGVI(n_samples, H): initial_mean = ift.Field.full(latent_space, 0.) # Define initial activation; random initialization works as well mu = initial_mean for i in range(5): # Draw new samples and minimize KL KL = ift.MetricGaussianKL(mu, H, n_samples, mirror_samples=False) # Set up KL with current mu KL, convergence = minimizer(KL) # Minimize KL and check for convergence mu = KL.position # Set minimized KL as new mu KL = ift.MetricGaussianKL(mu, H, n_samples, mirror_samples=False) KL, convergence = minimizer(KL) return KL iters=1 # Define number of iterations of posterior approximation. This might be helpful to check "how certain" the approximation is and if only an unstable local minimum is found KL_iterations = [] for i in range(iters): KL_iterations.append(MGVI(n_samples, H)) # Draw inferred signal from posterior samples and transform to original space sc = ift.StatCalculator() for i in range(iters): KL = KL_iterations[i] for sample in KL.samples: sc.add(A.apply(sample + KL.position, 1) + mean) # Retransform signal s = A*xi+mu posterior_mean = sc.mean # Get mean of all samples posterior_std = ift.sqrt(sc.var) # Get standard deviation of all samples # Classify posteriors via mahalanobis-distance and by classifying all posterior samples # with seperatly trained network ('Classifier') mahalanobis_distance_supervised = np.zeros([iters*n_samples, n_classes]) mahalanobis_distance = np.zeros([iters*n_samples, n_classes]) classified_posteriors = np.zeros([iters*n_samples, n_latent]) latent_posteriors = np.zeros([iters*n_samples, n_latent]) for k in range(iters): KL = KL_iterations[k] for j, sample in enumerate(KL.samples): s_posterior = A.apply(sample + KL.position, 1) + mean latent_posteriors[j+k*n_samples, :] = s_posterior.val classified_posteriors[j+k*n_samples, :] = Classifier(Decoder(s_posterior)).val for i in range(n_classes): mahalanobis_distance_supervised[j+k*n_samples, i] = np.sqrt((s_posterior.val[:n_classes] - means[:n_classes,i]).T @ np.linalg.inv(cov_supervised_variables[i]) @ (s_posterior.val[:n_classes] - means[:n_classes,i])) mahalanobis_distance[j+k*n_samples, i] = np.sqrt((s_posterior.val - means[:,i]).T @ np.linalg.inv(cov_all_variables[i]) @ (s_posterior.val - means[:,i])) #mahalanobis_distance[j+k*n_samples, i] = np.sqrt((s_posterior.val - means[:,i]).T @ (s_posterior.val - means[:,i])) # Euclidian Distance mahalanobis_mean = np.mean(mahalanobis_distance, axis=0) mahalanobis_std = np.sqrt(np.var(mahalanobis_distance, axis=0)) mahalanobis_mean_supervised = np.mean(mahalanobis_distance_supervised, axis=0) mahalanobis_std_supervised = np.sqrt(np.var(mahalanobis_distance_supervised, axis=0)) classified_mean = np.mean(classified_posteriors, axis=0) classified_std = np.std(classified_posteriors, axis=0) # Get all classifications of posterior samples for pie-plot visualization classified_posteriors_nn = np.sort(np.argmax(classified_posteriors, axis=1)) classified_posteriors_dm = np.sort(np.argmin(mahalanobis_distance, axis=1)) for i in range(n_classes): unique_digit_nn, count_nn = np.unique(classified_posteriors_nn, return_counts=True) unique_digit_dm, count_dm = np.unique(classified_posteriors_dm, return_counts=True) counts_nn = dict(zip(unique_digit_nn, count_nn)) counts_dm = dict(zip(unique_digit_dm, count_dm)) viridis = cm.get_cmap('viridis', n_classes) pie_colors = viridis(np.linspace(0, 1, n_classes)) # Create dictionary with important information: # Top scores of respective classification method (M-Dist, NN) # True or false classification (only valid if Labels given) # Overlapping standard-deviations n_scores = 3 # Number of top scoring elements to be displayed (max: n_classes) top_scores_nn = list(reversed(np.argsort(classified_mean)[-n_scores:])) top_scores_dm = list(np.argsort(mahalanobis_mean)[:n_scores]) overlap_bottom_nn = np.zeros(n_scores-1) overlap_bottom_dm = np.zeros(n_scores-1) for i in range(n_scores-1): overlap_bottom_nn[i] = (classified_mean[top_scores_nn[0]] - classified_std[top_scores_nn[0]]) - (classified_mean[top_scores_nn[i+1]] + classified_std[top_scores_nn[i+1]]) overlap_bottom_dm[i] = (mahalanobis_mean[top_scores_dm[i+1]] - mahalanobis_std[top_scores_dm[i+1]]) - (mahalanobis_mean[top_scores_dm[0]] + mahalanobis_std[top_scores_dm[0]]) keys_nn = ['Measure','Top Scores:', 'Classification:', 'ID:', 'N Samples:'] keys_dm = ['Measure','Top Scores:', 'Classification:', 'ID:', 'N Samples:', 'M-Dist of {}:'.format(top_scores_dm[0])] if top_scores_nn[0] == YTrain[-p]: values_nn = ['Neural Net Classifier','{}'.format(tuple(top_scores_nn)), 'True', 'YTrain[-{}]'.format(p), '{}'.format(n_samples)] if top_scores_dm[0] == YTrain[-p]: values_dm = ['Mahalanobis Distance','{}'.format(tuple(top_scores_dm)), 'True', 'YTrain[-{}]'.format(p), '{}'.format(n_samples), '{}'.format(mahalanobis_mean[top_scores_dm[0]])] if top_scores_nn[0] != YTrain[-p]: values_nn = ['Neural Net Classifier','{}'.format(tuple(top_scores_nn)), 'False', 'YTrain[-{}]'.format(p), '{}'.format(n_samples)] if top_scores_dm[0] != YTrain[-p]: values_dm = ['Mahalanobis Distance','{}'.format(tuple(top_scores_dm)), 'False', 'YTrain[-{}]'.format(p), '{}'.format(n_samples), '{}'.format(mahalanobis_mean[top_scores_dm[0]])] # Store Overlapping in Dictionary, expressed in terms of sigmas/STD of top Scoring digit for i in range(n_scores - 1): keys_nn.append('Overlap [sigmas] {} --> {}'.format(top_scores_nn[0], top_scores_nn[i+1])) values_nn.append(overlap_bottom_nn[i] / classified_std[top_scores_nn[0]]) keys_dm.append('Overlap [sigmas] {} --> {}'.format(top_scores_dm[0], top_scores_dm[i+1])) values_dm.append(overlap_bottom_dm[i] / mahalanobis_std[top_scores_dm[0]]) overlapping_nn = dict(zip(keys_nn, values_nn)) overlapping_dm = dict(zip(keys_dm, values_dm)) min = np.min([posterior_mean.val]) max = np.max([posterior_mean.val]) plt.subplot(3, 4, 1) barplot = plt.bar(range(n_classes), posterior_mean.val[0:n_classes], alpha=1, width=0.8, yerr=posterior_std.val[0:n_classes], label='MGVI with STD') barplot[np.where(posterior_mean.val == np.max(posterior_mean.val[:10]))[0][0]].set_color('r') plt.legend(fontsize=3) plt.title('$h\pm\delta_r$', fontsize=8) plt.xticks(range(n_classes), fontsize=6) plt.yticks(fontsize=6) plt.subplot(3, 4, 2) barplot = plt.bar(range(n_classes), classified_mean[:10], yerr=classified_std[:10]) plt.xticks(np.arange(n_classes), fontsize=6) plt.yticks(fontsize=6) barplot[np.where(classified_mean == np.max(classified_mean))[0][0]].set_color('r') plt.title('$f(g(h))\pm \delta_r$', fontsize=8) plt.subplot(3, 4, 3) m_mean = mahalanobis_mean_supervised m_std = mahalanobis_std_supervised barplot = plt.bar(range(n_classes), m_mean, yerr=m_std) barplot[np.where(m_mean == np.min(m_mean))[0][0]].set_color('r') for bar in barplot: yval = bar.get_height() yval = np.round(yval, decimals=2) plt.annotate('{}'.format(yval), xy=(bar.get_x() + bar.get_width() / 2, bar.get_height()), xytext=(0, 3), # 3 points vertical offset textcoords="offset points", ha='center', va='bottom', fontsize=5, rotation=45) plt.title('$\delta_m\pm \delta_r$', fontsize=8) plt.ylim(0, 1.3*np.max(m_mean)) plt.xticks(np.arange(n_classes), fontsize=6) plt.yticks(fontsize=6) plt.subplot(3, 4, 5) plt.imshow(np.reshape(data.val, img_shape)) plt.xlabel('Mock Signal') clear_axis() plt.xticks(fontsize=6) plt.yticks(fontsize=6) plt.subplot(3, 4, 6) plt.imshow(np.reshape(ground_truth.val, img_shape)) plt.xlabel('Ground Truth: {}'.format(YTest[p]), fontsize=8) clear_axis() plt.subplot(3, 4, 7) plt.imshow(np.reshape(Decoder(posterior_mean).val, img_shape)) plt.xlabel('Reconstruction', fontsize=8) clear_axis() plt.subplot(3, 4, 4) plt.pie([float(v) for v in counts_nn.values()], labels=[float(k) for k in counts_nn.keys()],autopct='%1.1f%%', colors=pie_colors[list(counts_nn.keys())], textprops={'fontsize': 4} ) plt.xlabel('Class. Post. NN', fontsize=8) plt.subplot(3, 4, 8) plt.pie([float(v) for v in counts_dm.values()], labels=[float(k) for k in counts_dm.keys()],autopct='%1.1f%%', colors=pie_colors[list(counts_dm.keys())], textprops={'fontsize': 4}) plt.xlabel('Class. Post. $d_M$', fontsize=8) plt.savefig('./corrupted_data_classification/{}'.format('example_results')) # Visualize reconstructions of all posterior samples. Output dependent on n_samples. grid = plt.GridSpec(np.int(np.floor(np.sqrt(len(latent_posteriors)))), np.int(np.ceil(np.sqrt(len(latent_posteriors)))), wspace=0.1, hspace=0.1) k=0 latent_posteriors=latent_posteriors[latent_posteriors[:,5].argsort()] for i in range(np.int(np.floor(np.sqrt(len(latent_posteriors))))): for j in range(np.int(np.ceil(np.sqrt(len(latent_posteriors))))): if k < iters*n_samples: plt.subplot(grid[i, j]) plt.imshow(np.reshape(Decoder(ift.Field.from_raw(latent_space, latent_posteriors[k, :])).val, img_shape), 'gray') clear_axis() k += 1 else: break fig = plt.gcf() plt.savefig('./corrupted_data_classification/{}'.format('example_samples')) print('Done. Results saved.')
18,972
42.71659
219
py
corrupted_data_classification
corrupted_data_classification-main/helper_functions/helper_functions.py
import pandas as pd import numpy as np import math import torch.optim as optim from torch.autograd import Variable import matplotlib.pyplot as plt from torchvision import datasets, transforms import matplotlib.pyplot as plt import numpy as np import io import cv2 import numpy as np import matplotlib.pyplot as plt import random from skimage.transform import resize from scipy.special import binom import warnings try: import nifty6 as ift except: warnings.warn("Failed importing nifty6") from PIL import Image def clear_axis(): ax = plt.gca() ax.axes.yaxis.set_ticks([]) ax.axes.xaxis.set_ticks([]) def convolution(colatitude): angle = colatitude * (180 / np.pi) return angle def gaussian(x, mu, sig): return np.exp(-np.power(x - mu, 2.) / (2 * np.power(sig, 2.))) n= 14 x_values = np.linspace(0, 1, n) kernel = np.ones(n) kernel = gaussian(x_values, 1, 3) kernels = np.zeros(784) for i in range(784//n): kernels[i*n:(i+1)*n] = kernel def conv(colatitude): #plt.imshow(np.reshape(colatitude, [28, 28])) #GT = convolve(GT, kernel=[0, 0.5, 1, 2, 3.5, 5, 3.5, 2, 1, 0.5, 0], boundary='extend') return convolve(colatitude, kernel=[0.1, 0.5, 1, 2, 3.5, 5, 3.5, 2, 1, 0.5, 0.1], boundary='extend') def get_cmap(n, name='hsv'): '''Returns a function that maps each index in 0, 1, ..., n-1 to a distinct RGB color; the keyword argument name must be a standard mpl colormap name.''' return plt.cm.get_cmap(name, n) def info_text(overlapping_nn, overlapping_dm): text = [] text.append('----------------------------------------------------') text.append('{:<40} {}'.format('Key','Label')) for k, v in overlapping_nn.items(): text.append("{:<40} {}".format(k, v)) text.append('----------------------------------------------------') text.append('{:<40} {}'.format('Key','Label')) for k, v in overlapping_dm.items(): text.append("{:<40} {}".format(k, v)) return text def get_noise(noise_level, position_space, seed): N_ift = ift.ScalingOperator(position_space, noise_level) with ift.random.Context(seed): n = N_ift.draw_sample_with_dtype(dtype=np.float64) return N_ift, n # N respresents the noise operator (diagnonal covariance), n represents acutal sampled noise values def rotation(image, img_shape, angle): im = np.reshape(image.val, img_shape) im = Image.fromarray(np.uint8(im*255)) im = im.rotate(angle) im = np.asarray(im)/255 im = np.reshape(im, image.shape) return ift.Field.from_raw(image.domain, im) def split_validation_set(XTrain, YTrain, val_perc): ''' Permutation of Training Dataset is inspired by an article pusblished on Medium: https://medium.com/@mjbhobe/mnist-digits-classification-with-keras-ed6c2374bd0e Author: Bhobeé, Manish Date of Publication: 29.09.2018 Relevant Code Section: Permutation of Data and Cut-Out of Validation Set Visit: 23.10.2020 Minor modifications were made on val_percent and names of variables (adjusted to my given variable names) and dimensionality of Datasets (mine is reshaped to vectors, the author used 2D Arrays.) ''' # shuffle the training dataset (5 times!) for i in range(5): np.random.seed(i) indexes = np.random.permutation(len(XTrain)) XTrain = XTrain[indexes] YTrain = YTrain[indexes] # now set-aside 20% of the train_data/labels as the # cross-validation sets val_perc = 0.2 val_count = int(val_perc * len(XTrain)) # first pick validation set from train_data/labels XVal = XTrain[:val_count] YVal = YTrain[:val_count] # leave rest in training set XTrain = XTrain[val_count:] YTrain = YTrain[val_count:] return XTrain, YTrain, XVal, YVal
3,772
30.705882
120
py
corrupted_data_classification
corrupted_data_classification-main/NNs/Fashion-MNIST/pretrained_supervised_ae10/autoencoder_fmnist.py
# -*- coding: utf-8 -*- # Commented out IPython magic to ensure Python compatibility. # %matplotlib inline # Commented out IPython magic to ensure Python compatibility. # Colab and system related import os import sys ### # Necessary to convert tensorflow-object (e.g. Neural Network) to Nifty-Operator sys.path.append('corrupted_data_classification/helper_functions/') ### import tensorflow as tf # Include path to access helper functions and Mask / Conv Operator sys.path.append('corrupted_data_classification/helper_functions/') from helper_functions import clear_axis, gaussian, get_cmap, info_text, get_noise, rotation, split_validation_set import Mask # Masking Operator import Conv # Convolution Operator sys.path.remove # Tensorflow # Plotting import matplotlib as mpl import matplotlib.pyplot as plt import matplotlib.cm as cm import matplotlib.pyplot as plt # %matplotlib inline plt.rcParams['figure.dpi'] = 200 # 200 e.g. is really fine, but slower # Numerics import random import numpy as np from sklearn.neighbors import KernelDensity from scipy.stats import multivariate_normal import sklearn as sk from sklearn import decomposition # Load MNIST Dataset mnist = tf.keras.datasets.fashion_mnist (XTrain, YTrain), (XTest, YTest) = mnist.load_data() XTrain, XTest = XTrain / 255.0, XTest / 255.0 # Cut out last 100 Training images for comparison XTrain = XTrain[0:-100] YTrain = YTrain[0:-100] # Reshape Xtrain and XTest to 1x784 Vectors instead of 28x28 arrays XTrain = XTrain.reshape((len(XTrain), np.prod(XTrain.shape[1:]))) XTest = XTest.reshape((len(XTest), np.prod(XTest.shape[1:]))) XTrain, YTrain, XVal, YVal = split_validation_set(XTrain, YTrain, val_perc=0.2) def autoencoder_deep(latent_space_size): Input = tf.keras.layers.Input(shape=784) h1 = tf.keras.layers.Dense(512, activation='selu', kernel_initializer='lecun_normal')(Input) h2 = tf.keras.layers.Dense(256, activation='selu', kernel_initializer='lecun_normal')(h1) h3 = tf.keras.layers.Dense(128, activation='selu', kernel_initializer='lecun_normal')(h2) encoded = tf.keras.layers.Dense(latent_space_size, activation='linear', activity_regularizer=tf.keras.regularizers.L2(0.1))(h3) # Decoder Decoder_Input = tf.keras.layers.Input(shape=latent_space_size) # Input for Decoder h5 = tf.keras.layers.Dense(128, activation='selu', kernel_initializer='lecun_normal')(Decoder_Input) h6 = tf.keras.layers.Dense(256, activation='selu', kernel_initializer='lecun_normal')(h5) h7 = tf.keras.layers.Dense(512, activation='selu', kernel_initializer='lecun_normal')(h6) decoded = tf.keras.layers.Dense(784, activation='sigmoid')(h7) # Decouple Encoder and Decoder from overall model Encoder = tf.keras.Model(Input, encoded) Decoder = tf.keras.Model(Decoder_Input, decoded) decoded = Decoder(encoded) model = tf.keras.Model(Input, [decoded, encoded]) return Encoder, Decoder, model Encoder, Decoder, model = autoencoder_deep(10) # Loss Function for Reconstruction of images (i.e. overall Autoencoder) def loss_fn_AE(y_true, y_pred): # y_pred = tf.nn.elu(y_pred) * tf.nn.softplus(y_pred) # return tf.losses.categorical_crossentropy(y_true, y_pred) # y_pred = tf.nn.softmax(y_pred) return tf.losses.binary_crossentropy(y_true,y_pred) #return tf.keras.losses.MeanSquaredError(y_true, y_pred) # Loss Function for Classification of Images in latent space def loss_fn_Encoder(y_true, y_pred): y_pred = tf.nn.softmax(y_pred) return tf.losses.sparse_categorical_crossentropy(y_true, y_pred) # Training Options model.compile(optimizer='adam', #loss=[loss_fn_AE, loss_fn_Encoder], loss=[loss_fn_AE, loss_fn_Encoder], metrics=['accuracy']) # Training and Testing # Training and Testing with tf.device('/device:GPU:0'): results = model.fit(XTrain, [XTrain, YTrain], epochs=25) model.evaluate(XTest, [XTest, YTest], verbose=2) # Save trained Decoder and trained Encoder Decoder.save('./corrupted_data_classification/NNs/Fashion-MNIST/pretrained_supervised_ae10/Decoder/', save_format='tf') Encoder.save('./corrupted_data_classification/NNs/Fashion-MNIST/pretrained_supervised_ae10/Encoder/', save_format='tf')
4,227
40.048544
119
py
corrupted_data_classification
corrupted_data_classification-main/NNs/MNIST/pretrained_supervised_ae10/autoencoder.py
# -*- coding: utf-8 -*- # Commented out IPython magic to ensure Python compatibility. # %matplotlib inline # Commented out IPython magic to ensure Python compatibility. # Colab and system related import os import sys ### # Necessary to convert tensorflow-object (e.g. Neural Network) to Nifty-Operator sys.path.append('corrupted_data_classification/helper_functions/') ### import tensorflow as tf # Include path to access helper functions and Mask / Conv Operator sys.path.append('corrupted_data_classification/helper_functions/') from helper_functions import clear_axis, gaussian, get_cmap, info_text, get_noise, rotation, split_validation_set import Mask # Masking Operator import Conv # Convolution Operator sys.path.remove # Tensorflow # Plotting import matplotlib as mpl import matplotlib.pyplot as plt import matplotlib.cm as cm import matplotlib.pyplot as plt # %matplotlib inline plt.rcParams['figure.dpi'] = 200 # 200 e.g. is really fine, but slower # Numerics import random import numpy as np from sklearn.neighbors import KernelDensity from scipy.stats import multivariate_normal import sklearn as sk from sklearn import decomposition # Load MNIST Dataset mnist = tf.keras.datasets.mnist (XTrain, YTrain), (XTest, YTest) = mnist.load_data() XTrain, XTest = XTrain / 255.0, XTest / 255.0 # Cut out last 100 Training images for comparison XTrain = XTrain[0:-100] YTrain = YTrain[0:-100] # Reshape Xtrain and XTest to 1x784 Vectors instead of 28x28 arrays XTrain = XTrain.reshape((len(XTrain), np.prod(XTrain.shape[1:]))) XTest = XTest.reshape((len(XTest), np.prod(XTest.shape[1:]))) XTrain, YTrain, XVal, YVal = split_validation_set(XTrain, YTrain, val_perc=0.2) def autoencoder_deep(latent_space_size): Input = tf.keras.layers.Input(shape=784) h1 = tf.keras.layers.Dense(512, activation='selu', kernel_initializer='lecun_normal')(Input) h2 = tf.keras.layers.Dense(256, activation='selu', kernel_initializer='lecun_normal')(h1) h3 = tf.keras.layers.Dense(128, activation='selu', kernel_initializer='lecun_normal')(h2) encoded = tf.keras.layers.Dense(latent_space_size, activation='linear', activity_regularizer=tf.keras.regularizers.L2(0.001))(h3) # Decoder Decoder_Input = tf.keras.layers.Input(shape=latent_space_size) # Input for Decoder h5 = tf.keras.layers.Dense(128, activation='selu', kernel_initializer='lecun_normal')(Decoder_Input) h6 = tf.keras.layers.Dense(256, activation='selu', kernel_initializer='lecun_normal')(h5) h7 = tf.keras.layers.Dense(512, activation='selu', kernel_initializer='lecun_normal')(h6) decoded = tf.keras.layers.Dense(784, activation='sigmoid')(h7) # Decouple Encoder and Decoder from overall model Encoder = tf.keras.Model(Input, encoded) Decoder = tf.keras.Model(Decoder_Input, decoded) decoded = Decoder(encoded) model = tf.keras.Model(Input, [decoded, encoded]) return Encoder, Decoder, model Encoder, Decoder, model = autoencoder_deep(10) # Loss Function for Reconstruction of images (i.e. overall Autoencoder) def loss_fn_AE(y_true, y_pred): # y_pred = tf.nn.elu(y_pred) * tf.nn.softplus(y_pred) # return tf.losses.categorical_crossentropy(y_true, y_pred) # y_pred = tf.nn.softmax(y_pred) return tf.losses.binary_crossentropy(y_true,y_pred) #return tf.keras.losses.MeanSquaredError(y_true, y_pred) # Loss Function for Classification of Images in latent space def loss_fn_Encoder(y_true, y_pred): y_pred = tf.nn.softmax(y_pred) return tf.losses.sparse_categorical_crossentropy(y_true, y_pred) # Training Options model.compile(optimizer='adam', #loss=[loss_fn_AE, loss_fn_Encoder], loss=[loss_fn_AE, loss_fn_Encoder], metrics=['accuracy']) # Training and Testing results = model.fit(XTrain, [XTrain, YTrain], epochs=25) model.evaluate(XTest, [XTest, YTest], verbose=2) # Save trained Decoder and trained Encoder Decoder.save('./corrupted_data_classification/NNs/MNIST/pretrained_supervised_ae10/Decoder/', save_format='tf') Encoder.save('./corrupted_data_classification/NNs/MNIST/pretrained_supervised_ae10/Encoder/', save_format='tf') plt.plot(results.history['dense_3_accuracy'])
4,195
39.346154
113
py
mmyolo
mmyolo-main/setup.py
#!/usr/bin/env python # Copyright (c) OpenMMLab. All rights reserved. import os import os.path as osp import platform import shutil import sys import warnings from setuptools import find_packages, setup from torch.utils.cpp_extension import BuildExtension def readme(): with open('README.md', encoding='utf-8') as f: content = f.read() return content version_file = 'mmyolo/version.py' def get_version(): with open(version_file) as f: exec(compile(f.read(), version_file, 'exec')) return locals()['__version__'] def parse_requirements(fname='requirements.txt', with_version=True): """Parse the package dependencies listed in a requirements file but strips specific versioning information. Args: fname (str): path to requirements file with_version (bool, default=False): if True include version specs Returns: List[str]: list of requirements items CommandLine: python -c "import setup; print(setup.parse_requirements())" """ import re import sys from os.path import exists require_fpath = fname def parse_line(line): """Parse information from a line in a requirements text file.""" if line.startswith('-r '): # Allow specifying requirements in other files target = line.split(' ')[1] for info in parse_require_file(target): yield info else: info = {'line': line} if line.startswith('-e '): info['package'] = line.split('#egg=')[1] elif '@git+' in line: info['package'] = line else: # Remove versioning from the package pat = '(' + '|'.join(['>=', '==', '>']) + ')' parts = re.split(pat, line, maxsplit=1) parts = [p.strip() for p in parts] info['package'] = parts[0] if len(parts) > 1: op, rest = parts[1:] if ';' in rest: # Handle platform specific dependencies # http://setuptools.readthedocs.io/en/latest/setuptools.html#declaring-platform-specific-dependencies version, platform_deps = map(str.strip, rest.split(';')) info['platform_deps'] = platform_deps else: version = rest # NOQA info['version'] = (op, version) yield info def parse_require_file(fpath): with open(fpath) as f: for line in f.readlines(): line = line.strip() if line and not line.startswith('#'): yield from parse_line(line) def gen_packages_items(): if exists(require_fpath): for info in parse_require_file(require_fpath): parts = [info['package']] if with_version and 'version' in info: parts.extend(info['version']) if not sys.version.startswith('3.4'): # apparently package_deps are broken in 3.4 platform_deps = info.get('platform_deps') if platform_deps is not None: parts.append(';' + platform_deps) item = ''.join(parts) yield item packages = list(gen_packages_items()) return packages def add_mim_extension(): """Add extra files that are required to support MIM into the package. These files will be added by creating a symlink to the originals if the package is installed in `editable` mode (e.g. pip install -e .), or by copying from the originals otherwise. """ # parse installment mode if 'develop' in sys.argv: # installed by `pip install -e .` if platform.system() == 'Windows': # set `copy` mode here since symlink fails on Windows. mode = 'copy' else: mode = 'symlink' elif 'sdist' in sys.argv or 'bdist_wheel' in sys.argv: # installed by `pip install .` # or create source distribution by `python setup.py sdist` mode = 'copy' else: return filenames = ['tools', 'configs', 'demo', 'model-index.yml'] repo_path = osp.dirname(__file__) mim_path = osp.join(repo_path, 'mmyolo', '.mim') os.makedirs(mim_path, exist_ok=True) for filename in filenames: if osp.exists(filename): src_path = osp.join(repo_path, filename) tar_path = osp.join(mim_path, filename) if osp.isfile(tar_path) or osp.islink(tar_path): os.remove(tar_path) elif osp.isdir(tar_path): shutil.rmtree(tar_path) if mode == 'symlink': src_relpath = osp.relpath(src_path, osp.dirname(tar_path)) os.symlink(src_relpath, tar_path) elif mode == 'copy': if osp.isfile(src_path): shutil.copyfile(src_path, tar_path) elif osp.isdir(src_path): shutil.copytree(src_path, tar_path) else: warnings.warn(f'Cannot copy file {src_path}.') else: raise ValueError(f'Invalid mode {mode}') if __name__ == '__main__': add_mim_extension() setup( name='mmyolo', version=get_version(), description='OpenMMLab Toolbox of YOLO', long_description=readme(), long_description_content_type='text/markdown', author='MMYOLO Contributors', author_email='openmmlab@gmail.com', keywords='computer vision, object detection', url='https://github.com/open-mmlab/mmyolo', packages=find_packages(exclude=('configs', 'tools', 'demo')), include_package_data=True, classifiers=[ 'Development Status :: 5 - Production/Stable', 'License :: OSI Approved :: Apache Software License', 'Operating System :: OS Independent', 'Programming Language :: Python :: 3', 'Programming Language :: Python :: 3.7', 'Programming Language :: Python :: 3.8', 'Programming Language :: Python :: 3.9', ], license='GPL License 3.0', install_requires=parse_requirements('requirements/runtime.txt'), extras_require={ 'all': parse_requirements('requirements.txt'), 'tests': parse_requirements('requirements/tests.txt'), 'build': parse_requirements('requirements/build.txt'), 'mim': parse_requirements('requirements/mminstall.txt'), }, ext_modules=[], cmdclass={'build_ext': BuildExtension}, zip_safe=False)
6,862
34.744792
125
py
mmyolo
mmyolo-main/tools/test.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import os import os.path as osp from mmdet.engine.hooks.utils import trigger_visualization_hook from mmengine.config import Config, ConfigDict, DictAction from mmengine.evaluator import DumpResults from mmengine.runner import Runner from mmyolo.registry import RUNNERS from mmyolo.utils import is_metainfo_lower # TODO: support fuse_conv_bn def parse_args(): parser = argparse.ArgumentParser( description='MMYOLO test (and eval) a model') parser.add_argument('config', help='test config file path') parser.add_argument('checkpoint', help='checkpoint file') parser.add_argument( '--work-dir', help='the directory to save the file containing evaluation metrics') parser.add_argument( '--out', type=str, help='output result file (must be a .pkl file) in pickle format') parser.add_argument( '--json-prefix', type=str, help='the prefix of the output json file without perform evaluation, ' 'which is useful when you want to format the result to a specific ' 'format and submit it to the test server') parser.add_argument( '--tta', action='store_true', help='Whether to use test time augmentation') parser.add_argument( '--show', action='store_true', help='show prediction results') parser.add_argument( '--deploy', action='store_true', help='Switch model to deployment mode') parser.add_argument( '--show-dir', help='directory where painted images will be saved. ' 'If specified, it will be automatically saved ' 'to the work_dir/timestamp/show_dir') parser.add_argument( '--wait-time', type=float, default=2, help='the interval of show (s)') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) return args def main(): args = parse_args() # load config cfg = Config.fromfile(args.config) # replace the ${key} with the value of cfg.key # cfg = replace_cfg_vals(cfg) cfg.launcher = args.launcher if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) # work_dir is determined in this priority: CLI > segment in file > filename if args.work_dir is not None: # update configs according to CLI args if args.work_dir is not None cfg.work_dir = args.work_dir elif cfg.get('work_dir', None) is None: # use config filename as default work_dir if cfg.work_dir is None cfg.work_dir = osp.join('./work_dirs', osp.splitext(osp.basename(args.config))[0]) cfg.load_from = args.checkpoint if args.show or args.show_dir: cfg = trigger_visualization_hook(cfg, args) if args.deploy: cfg.custom_hooks.append(dict(type='SwitchToDeployHook')) # add `format_only` and `outfile_prefix` into cfg if args.json_prefix is not None: cfg_json = { 'test_evaluator.format_only': True, 'test_evaluator.outfile_prefix': args.json_prefix } cfg.merge_from_dict(cfg_json) # Determine whether the custom metainfo fields are all lowercase is_metainfo_lower(cfg) if args.tta: assert 'tta_model' in cfg, 'Cannot find ``tta_model`` in config.' \ " Can't use tta !" assert 'tta_pipeline' in cfg, 'Cannot find ``tta_pipeline`` ' \ "in config. Can't use tta !" cfg.model = ConfigDict(**cfg.tta_model, module=cfg.model) test_data_cfg = cfg.test_dataloader.dataset while 'dataset' in test_data_cfg: test_data_cfg = test_data_cfg['dataset'] # batch_shapes_cfg will force control the size of the output image, # it is not compatible with tta. if 'batch_shapes_cfg' in test_data_cfg: test_data_cfg.batch_shapes_cfg = None test_data_cfg.pipeline = cfg.tta_pipeline # build the runner from config if 'runner_type' not in cfg: # build the default runner runner = Runner.from_cfg(cfg) else: # build customized runner from the registry # if 'runner_type' is set in the cfg runner = RUNNERS.build(cfg) # add `DumpResults` dummy metric if args.out is not None: assert args.out.endswith(('.pkl', '.pickle')), \ 'The dump file must be a pkl file.' runner.test_evaluator.metrics.append( DumpResults(out_file_path=args.out)) # start testing runner.test() if __name__ == '__main__': main()
5,443
35.05298
79
py
mmyolo
mmyolo-main/tools/train.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import logging import os import os.path as osp from mmengine.config import Config, DictAction from mmengine.logging import print_log from mmengine.runner import Runner from mmyolo.registry import RUNNERS from mmyolo.utils import is_metainfo_lower def parse_args(): parser = argparse.ArgumentParser(description='Train a detector') parser.add_argument('config', help='train config file path') parser.add_argument('--work-dir', help='the dir to save logs and models') parser.add_argument( '--amp', action='store_true', default=False, help='enable automatic-mixed-precision training') parser.add_argument( '--resume', nargs='?', type=str, const='auto', help='If specify checkpoint path, resume from it, while if not ' 'specify, try to auto resume from the latest checkpoint ' 'in the work directory.') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) return args def main(): args = parse_args() # load config cfg = Config.fromfile(args.config) # replace the ${key} with the value of cfg.key # cfg = replace_cfg_vals(cfg) cfg.launcher = args.launcher if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) # work_dir is determined in this priority: CLI > segment in file > filename if args.work_dir is not None: # update configs according to CLI args if args.work_dir is not None cfg.work_dir = args.work_dir elif cfg.get('work_dir', None) is None: # use config filename as default work_dir if cfg.work_dir is None cfg.work_dir = osp.join('./work_dirs', osp.splitext(osp.basename(args.config))[0]) # enable automatic-mixed-precision training if args.amp is True: optim_wrapper = cfg.optim_wrapper.type if optim_wrapper == 'AmpOptimWrapper': print_log( 'AMP training is already enabled in your config.', logger='current', level=logging.WARNING) else: assert optim_wrapper == 'OptimWrapper', ( '`--amp` is only supported when the optimizer wrapper type is ' f'`OptimWrapper` but got {optim_wrapper}.') cfg.optim_wrapper.type = 'AmpOptimWrapper' cfg.optim_wrapper.loss_scale = 'dynamic' # resume is determined in this priority: resume from > auto_resume if args.resume == 'auto': cfg.resume = True cfg.load_from = None elif args.resume is not None: cfg.resume = True cfg.load_from = args.resume # Determine whether the custom metainfo fields are all lowercase is_metainfo_lower(cfg) # build the runner from config if 'runner_type' not in cfg: # build the default runner runner = Runner.from_cfg(cfg) else: # build customized runner from the registry # if 'runner_type' is set in the cfg runner = RUNNERS.build(cfg) # start training runner.train() if __name__ == '__main__': main()
3,969
33.224138
79
py
mmyolo
mmyolo-main/tools/misc/download_dataset.py
import argparse from itertools import repeat from multiprocessing.pool import ThreadPool from pathlib import Path from tarfile import TarFile from zipfile import ZipFile import torch def parse_args(): parser = argparse.ArgumentParser( description='Download datasets for training') parser.add_argument( '--dataset-name', type=str, help='dataset name', default='coco2017') parser.add_argument( '--save-dir', type=str, help='the dir to save dataset', default='data/coco') parser.add_argument( '--unzip', action='store_true', help='whether unzip dataset or not, zipped files will be saved') parser.add_argument( '--delete', action='store_true', help='delete the download zipped files') parser.add_argument( '--threads', type=int, help='number of threading', default=4) args = parser.parse_args() return args def download(url, dir, unzip=True, delete=False, threads=1): def download_one(url, dir): f = dir / Path(url).name if Path(url).is_file(): Path(url).rename(f) elif not f.exists(): print(f'Downloading {url} to {f}') torch.hub.download_url_to_file(url, f, progress=True) if unzip and f.suffix in ('.zip', '.tar'): print(f'Unzipping {f.name}') if f.suffix == '.zip': ZipFile(f).extractall(path=dir) elif f.suffix == '.tar': TarFile(f).extractall(path=dir) if delete: f.unlink() print(f'Delete {f}') dir = Path(dir) if threads > 1: pool = ThreadPool(threads) pool.imap(lambda x: download_one(*x), zip(url, repeat(dir))) pool.close() pool.join() else: for u in [url] if isinstance(url, (str, Path)) else url: download_one(u, dir) def main(): args = parse_args() path = Path(args.save_dir) if not path.exists(): path.mkdir(parents=True, exist_ok=True) data2url = dict( # TODO: Support for downloading Panoptic Segmentation of COCO coco2017=[ 'http://images.cocodataset.org/zips/train2017.zip', 'http://images.cocodataset.org/zips/val2017.zip', 'http://images.cocodataset.org/zips/test2017.zip', 'http://images.cocodataset.org/annotations/' + 'annotations_trainval2017.zip' ], lvis=[ 'https://s3-us-west-2.amazonaws.com/dl.fbaipublicfiles.com/LVIS/lvis_v1_train.json.zip', # noqa 'https://s3-us-west-2.amazonaws.com/dl.fbaipublicfiles.com/LVIS/lvis_v1_train.json.zip', # noqa ], voc2007=[ 'http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar', # noqa 'http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar', # noqa 'http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCdevkit_08-Jun-2007.tar', # noqa ], voc2012=[ 'http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar', # noqa ], balloon=[ # src link: https://github.com/matterport/Mask_RCNN/releases/download/v2.1/balloon_dataset.zip # noqa 'https://download.openmmlab.com/mmyolo/data/balloon_dataset.zip' ], cat=[ 'https://download.openmmlab.com/mmyolo/data/cat_dataset.zip' # noqa ], ) url = data2url.get(args.dataset_name, None) if url is None: print('Only support COCO, VOC, balloon, cat and LVIS now!') return download( url, dir=path, unzip=args.unzip, delete=args.delete, threads=args.threads) if __name__ == '__main__': main()
3,814
32.761062
113
py
mmyolo
mmyolo-main/tools/misc/publish_model.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import subprocess import torch def parse_args(): parser = argparse.ArgumentParser( description='Process a checkpoint to be published') parser.add_argument('in_file', help='input checkpoint filename') parser.add_argument('out_file', help='output checkpoint filename') args = parser.parse_args() return args def process_checkpoint(in_file, out_file): checkpoint = torch.load(in_file, map_location='cpu') # remove optimizer for smaller file size if 'optimizer' in checkpoint: del checkpoint['optimizer'] if 'message_hub' in checkpoint: del checkpoint['message_hub'] if 'ema_state_dict' in checkpoint: del checkpoint['ema_state_dict'] for key in list(checkpoint['state_dict']): if key.startswith('data_preprocessor'): checkpoint['state_dict'].pop(key) elif 'priors_base_sizes' in key: checkpoint['state_dict'].pop(key) elif 'grid_offset' in key: checkpoint['state_dict'].pop(key) elif 'prior_inds' in key: checkpoint['state_dict'].pop(key) if torch.__version__ >= '1.6': torch.save(checkpoint, out_file, _use_new_zipfile_serialization=False) else: torch.save(checkpoint, out_file) sha = subprocess.check_output(['sha256sum', out_file]).decode() if out_file.endswith('.pth'): out_file_name = out_file[:-4] else: out_file_name = out_file final_file = out_file_name + f'-{sha[:8]}.pth' subprocess.Popen(['mv', out_file, final_file]) def main(): args = parse_args() process_checkpoint(args.in_file, args.out_file) if __name__ == '__main__': main()
1,744
29.086207
78
py
mmyolo
mmyolo-main/tools/model_converters/yolov6_to_mmyolo.py
import argparse from collections import OrderedDict import torch def convert(src, dst): import sys sys.path.append('yolov6') try: ckpt = torch.load(src, map_location=torch.device('cpu')) except ModuleNotFoundError: raise RuntimeError( 'This script must be placed under the meituan/YOLOv6 repo,' ' because loading the official pretrained model need' ' some python files to build model.') # The saved model is the model before reparameterization model = ckpt['ema' if ckpt.get('ema') else 'model'].float() new_state_dict = OrderedDict() for k, v in model.state_dict().items(): name = k if 'detect' in k: if 'proj' in k: continue name = k.replace('detect', 'bbox_head.head_module') if k.find('anchors') >= 0 or k.find('anchor_grid') >= 0: continue if 'ERBlock_2' in k: name = k.replace('ERBlock_2', 'stage1.0') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'ERBlock_3' in k: name = k.replace('ERBlock_3', 'stage2.0') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'ERBlock_4' in k: name = k.replace('ERBlock_4', 'stage3.0') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'ERBlock_5' in k: name = k.replace('ERBlock_5', 'stage4.0') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') if 'stage4.0.2' in name: name = name.replace('stage4.0.2', 'stage4.1') name = name.replace('cv', 'conv') elif 'reduce_layer0' in k: name = k.replace('reduce_layer0', 'reduce_layers.2') elif 'Rep_p4' in k: name = k.replace('Rep_p4', 'top_down_layers.0.0') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'reduce_layer1' in k: name = k.replace('reduce_layer1', 'top_down_layers.0.1') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'Rep_p3' in k: name = k.replace('Rep_p3', 'top_down_layers.1') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'upsample0' in k: name = k.replace('upsample0.upsample_transpose', 'upsample_layers.0') elif 'upsample1' in k: name = k.replace('upsample1.upsample_transpose', 'upsample_layers.1') elif 'Rep_n3' in k: name = k.replace('Rep_n3', 'bottom_up_layers.0') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'Rep_n4' in k: name = k.replace('Rep_n4', 'bottom_up_layers.1') if '.cv' in k: name = name.replace('.cv', '.conv') if '.m.' in k: name = name.replace('.m.', '.block.') elif 'downsample2' in k: name = k.replace('downsample2', 'downsample_layers.0') elif 'downsample1' in k: name = k.replace('downsample1', 'downsample_layers.1') new_state_dict[name] = v data = {'state_dict': new_state_dict} torch.save(data, dst) # Note: This script must be placed under the yolov6 repo to run. def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument( '--src', default='yolov6s.pt', help='src yolov6 model path') parser.add_argument('--dst', default='mmyolov6.pt', help='save path') args = parser.parse_args() convert(args.src, args.dst) if __name__ == '__main__': main()
4,403
36.965517
73
py
mmyolo
mmyolo-main/tools/model_converters/yolox_to_mmyolo.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse from collections import OrderedDict import torch neck_dict = { 'backbone.lateral_conv0': 'neck.reduce_layers.2', 'backbone.C3_p4.conv': 'neck.top_down_layers.0.0.cv', 'backbone.C3_p4.m.0.': 'neck.top_down_layers.0.0.m.0.', 'backbone.reduce_conv1': 'neck.top_down_layers.0.1', 'backbone.C3_p3.conv': 'neck.top_down_layers.1.cv', 'backbone.C3_p3.m.0.': 'neck.top_down_layers.1.m.0.', 'backbone.bu_conv2': 'neck.downsample_layers.0', 'backbone.C3_n3.conv': 'neck.bottom_up_layers.0.cv', 'backbone.C3_n3.m.0.': 'neck.bottom_up_layers.0.m.0.', 'backbone.bu_conv1': 'neck.downsample_layers.1', 'backbone.C3_n4.conv': 'neck.bottom_up_layers.1.cv', 'backbone.C3_n4.m.0.': 'neck.bottom_up_layers.1.m.0.', } def convert_stem(model_key, model_weight, state_dict, converted_names): new_key = model_key[9:] state_dict[new_key] = model_weight converted_names.add(model_key) print(f'Convert {model_key} to {new_key}') def convert_backbone(model_key, model_weight, state_dict, converted_names): new_key = model_key.replace('backbone.dark', 'stage') num = int(new_key[14]) - 1 new_key = new_key[:14] + str(num) + new_key[15:] if '.m.' in model_key: new_key = new_key.replace('.m.', '.blocks.') elif not new_key[16] == '0' and 'stage4.1' not in new_key: new_key = new_key.replace('conv1', 'main_conv') new_key = new_key.replace('conv2', 'short_conv') new_key = new_key.replace('conv3', 'final_conv') state_dict[new_key] = model_weight converted_names.add(model_key) print(f'Convert {model_key} to {new_key}') def convert_neck(model_key, model_weight, state_dict, converted_names): for old, new in neck_dict.items(): if old in model_key: new_key = model_key.replace(old, new) if '.m.' in model_key: new_key = new_key.replace('.m.', '.blocks.') elif '.C' in model_key: new_key = new_key.replace('cv1', 'main_conv') new_key = new_key.replace('cv2', 'short_conv') new_key = new_key.replace('cv3', 'final_conv') state_dict[new_key] = model_weight converted_names.add(model_key) print(f'Convert {model_key} to {new_key}') def convert_head(model_key, model_weight, state_dict, converted_names): if 'stem' in model_key: new_key = model_key.replace('head.stem', 'neck.out_layer') elif 'cls_convs' in model_key: new_key = model_key.replace( 'head.cls_convs', 'bbox_head.head_module.multi_level_cls_convs') elif 'reg_convs' in model_key: new_key = model_key.replace( 'head.reg_convs', 'bbox_head.head_module.multi_level_reg_convs') elif 'preds' in model_key: new_key = model_key.replace('head.', 'bbox_head.head_module.multi_level_conv_') new_key = new_key.replace('_preds', '') state_dict[new_key] = model_weight converted_names.add(model_key) print(f'Convert {model_key} to {new_key}') def convert(src, dst): """Convert keys in detectron pretrained YOLOX models to mmyolo style.""" blobs = torch.load(src)['model'] state_dict = OrderedDict() converted_names = set() for key, weight in blobs.items(): if 'backbone.stem' in key: convert_stem(key, weight, state_dict, converted_names) elif 'backbone.backbone' in key: convert_backbone(key, weight, state_dict, converted_names) elif 'backbone.neck' not in key and 'head' not in key: convert_neck(key, weight, state_dict, converted_names) elif 'head' in key: convert_head(key, weight, state_dict, converted_names) # save checkpoint checkpoint = dict() checkpoint['state_dict'] = state_dict torch.save(checkpoint, dst) def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument( '--src', default='yolox_s.pth', help='src yolox model path') parser.add_argument('--dst', default='mmyoloxs.pt', help='save path') args = parser.parse_args() convert(args.src, args.dst) if __name__ == '__main__': main()
4,218
37.009009
78
py
mmyolo
mmyolo-main/tools/model_converters/yolov8_to_mmyolo.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse from collections import OrderedDict import torch convert_dict_s = { # backbone 'model.0': 'backbone.stem', 'model.1': 'backbone.stage1.0', 'model.2': 'backbone.stage1.1', 'model.3': 'backbone.stage2.0', 'model.4': 'backbone.stage2.1', 'model.5': 'backbone.stage3.0', 'model.6': 'backbone.stage3.1', 'model.7': 'backbone.stage4.0', 'model.8': 'backbone.stage4.1', 'model.9': 'backbone.stage4.2', # neck 'model.12': 'neck.top_down_layers.0', 'model.15': 'neck.top_down_layers.1', 'model.16': 'neck.downsample_layers.0', 'model.18': 'neck.bottom_up_layers.0', 'model.19': 'neck.downsample_layers.1', 'model.21': 'neck.bottom_up_layers.1', # Detector 'model.22': 'bbox_head.head_module', } def convert(src, dst): """Convert keys in pretrained YOLOv8 models to mmyolo style.""" convert_dict = convert_dict_s try: yolov8_model = torch.load(src)['model'] blobs = yolov8_model.state_dict() except ModuleNotFoundError: raise RuntimeError( 'This script must be placed under the ultralytics repo,' ' because loading the official pretrained model need' ' `model.py` to build model.' 'Also need to install hydra-core>=1.2.0 and thop>=0.1.1') state_dict = OrderedDict() for key, weight in blobs.items(): num, module = key.split('.')[1:3] prefix = f'model.{num}' new_key = key.replace(prefix, convert_dict[prefix]) if '.m.' in new_key: new_key = new_key.replace('.m.', '.blocks.') new_key = new_key.replace('.cv', '.conv') elif 'bbox_head.head_module' in new_key: new_key = new_key.replace('.cv2', '.reg_preds') new_key = new_key.replace('.cv3', '.cls_preds') elif 'backbone.stage4.2' in new_key: new_key = new_key.replace('.cv', '.conv') else: new_key = new_key.replace('.cv1', '.main_conv') new_key = new_key.replace('.cv2', '.final_conv') if 'bbox_head.head_module.dfl.conv.weight' == new_key: print('Drop "bbox_head.head_module.dfl.conv.weight", ' 'because it is useless') continue state_dict[new_key] = weight print(f'Convert {key} to {new_key}') # save checkpoint checkpoint = dict() checkpoint['state_dict'] = state_dict torch.save(checkpoint, dst) # Note: This script must be placed under the YOLOv8 repo to run. def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument( '--src', default='yolov8s.pt', help='src YOLOv8 model path') parser.add_argument('--dst', default='mmyolov8s.pth', help='save path') args = parser.parse_args() convert(args.src, args.dst) if __name__ == '__main__': main()
2,937
31.644444
75
py
mmyolo
mmyolo-main/tools/model_converters/rtmdet_to_mmyolo.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse from collections import OrderedDict import torch def convert(src, dst): """Convert keys in pretrained RTMDet models to MMYOLO style.""" blobs = torch.load(src)['state_dict'] state_dict = OrderedDict() for key, weight in blobs.items(): if 'neck.reduce_layers.0' in key: new_key = key.replace('.0', '.2') state_dict[new_key] = weight elif 'neck.reduce_layers.1' in key: new_key = key.replace('reduce_layers.1', 'top_down_layers.0.1') state_dict[new_key] = weight elif 'neck.top_down_blocks.0' in key: new_key = key.replace('down_blocks', 'down_layers.0') state_dict[new_key] = weight elif 'neck.top_down_blocks.1' in key: new_key = key.replace('down_blocks', 'down_layers') state_dict[new_key] = weight elif 'downsamples' in key: new_key = key.replace('downsamples', 'downsample_layers') state_dict[new_key] = weight elif 'bottom_up_blocks' in key: new_key = key.replace('bottom_up_blocks', 'bottom_up_layers') state_dict[new_key] = weight elif 'out_convs' in key: new_key = key.replace('out_convs', 'out_layers') state_dict[new_key] = weight elif 'bbox_head' in key: new_key = key.replace('bbox_head', 'bbox_head.head_module') state_dict[new_key] = weight elif 'data_preprocessor' in key: continue else: new_key = key state_dict[new_key] = weight print(f'Convert {key} to {new_key}') # save checkpoint checkpoint = dict() checkpoint['state_dict'] = state_dict checkpoint['meta'] = blobs.get('meta') torch.save(checkpoint, dst) def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument('src', help='src rtm model path') parser.add_argument('dst', help='save path') args = parser.parse_args() convert(args.src, args.dst) if __name__ == '__main__': main()
2,142
33.564516
75
py
mmyolo
mmyolo-main/tools/model_converters/ppyoloe_to_mmyolo.py
import argparse import pickle from collections import OrderedDict import torch def convert_bn(k: str): name = k.replace('._mean', '.running_mean').replace('._variance', '.running_var') return name def convert_repvgg(k: str): if '.conv2.conv1.' in k: name = k.replace('.conv2.conv1.', '.conv2.rbr_dense.') return name elif '.conv2.conv2.' in k: name = k.replace('.conv2.conv2.', '.conv2.rbr_1x1.') return name else: return k def convert(src: str, dst: str, imagenet_pretrain: bool = False): with open(src, 'rb') as f: model = pickle.load(f) new_state_dict = OrderedDict() if imagenet_pretrain: for k, v in model.items(): if '@@' in k: continue if 'stem.' in k: # backbone.stem.conv1.conv.weight # -> backbone.stem.0.conv.weight org_ind = k.split('.')[1][-1] new_ind = str(int(org_ind) - 1) name = k.replace('stem.conv%s.' % org_ind, 'stem.%s.' % new_ind) else: # backbone.stages.1.conv2.bn._variance # -> backbone.stage2.0.conv2.bn.running_var org_stage_ind = k.split('.')[1] new_stage_ind = str(int(org_stage_ind) + 1) name = k.replace('stages.%s.' % org_stage_ind, 'stage%s.0.' % new_stage_ind) name = convert_repvgg(name) if '.attn.' in k: name = name.replace('.attn.fc.', '.attn.fc.conv.') name = convert_bn(name) name = 'backbone.' + name new_state_dict[name] = torch.from_numpy(v) else: for k, v in model.items(): name = k if k.startswith('backbone.'): if '.stem.' in k: # backbone.stem.conv1.conv.weight # -> backbone.stem.0.conv.weight org_ind = k.split('.')[2][-1] new_ind = str(int(org_ind) - 1) name = k.replace('.stem.conv%s.' % org_ind, '.stem.%s.' % new_ind) else: # backbone.stages.1.conv2.bn._variance # -> backbone.stage2.0.conv2.bn.running_var org_stage_ind = k.split('.')[2] new_stage_ind = str(int(org_stage_ind) + 1) name = k.replace('.stages.%s.' % org_stage_ind, '.stage%s.0.' % new_stage_ind) name = convert_repvgg(name) if '.attn.' in k: name = name.replace('.attn.fc.', '.attn.fc.conv.') name = convert_bn(name) elif k.startswith('neck.'): # fpn_stages if k.startswith('neck.fpn_stages.'): # neck.fpn_stages.0.0.conv1.conv.weight # -> neck.reduce_layers.2.0.conv1.conv.weight if k.startswith('neck.fpn_stages.0.0.'): name = k.replace('neck.fpn_stages.0.0.', 'neck.reduce_layers.2.0.') if '.spp.' in name: name = name.replace('.spp.conv.', '.spp.conv2.') # neck.fpn_stages.1.0.conv1.conv.weight # -> neck.top_down_layers.0.0.conv1.conv.weight elif k.startswith('neck.fpn_stages.1.0.'): name = k.replace('neck.fpn_stages.1.0.', 'neck.top_down_layers.0.0.') elif k.startswith('neck.fpn_stages.2.0.'): name = k.replace('neck.fpn_stages.2.0.', 'neck.top_down_layers.1.0.') else: raise NotImplementedError('Not implemented.') name = name.replace('.0.convs.', '.0.blocks.') elif k.startswith('neck.fpn_routes.'): # neck.fpn_routes.0.conv.weight # -> neck.upsample_layers.0.0.conv.weight index = k.split('.')[2] name = 'neck.upsample_layers.' + index + '.0.' + '.'.join( k.split('.')[-2:]) name = name.replace('.0.convs.', '.0.blocks.') elif k.startswith('neck.pan_stages.'): # neck.pan_stages.0.0.conv1.conv.weight # -> neck.bottom_up_layers.1.0.conv1.conv.weight ind = k.split('.')[2] name = k.replace( 'neck.pan_stages.' + ind, 'neck.bottom_up_layers.' + ('0' if ind == '1' else '1')) name = name.replace('.0.convs.', '.0.blocks.') elif k.startswith('neck.pan_routes.'): # neck.pan_routes.0.conv.weight # -> neck.downsample_layers.0.conv.weight ind = k.split('.')[2] name = k.replace( 'neck.pan_routes.' + ind, 'neck.downsample_layers.' + ('0' if ind == '1' else '1')) name = name.replace('.0.convs.', '.0.blocks.') else: raise NotImplementedError('Not implement.') name = convert_repvgg(name) name = convert_bn(name) elif k.startswith('yolo_head.'): if ('anchor_points' in k) or ('stride_tensor' in k): continue if 'proj_conv' in k: name = k.replace('yolo_head.proj_conv.', 'bbox_head.head_module.proj_conv.') else: for org_key, rep_key in [ [ 'yolo_head.stem_cls.', 'bbox_head.head_module.cls_stems.' ], [ 'yolo_head.stem_reg.', 'bbox_head.head_module.reg_stems.' ], [ 'yolo_head.pred_cls.', 'bbox_head.head_module.cls_preds.' ], [ 'yolo_head.pred_reg.', 'bbox_head.head_module.reg_preds.' ] ]: name = name.replace(org_key, rep_key) name = name.split('.') ind = name[3] name[3] = str(2 - int(ind)) name = '.'.join(name) name = convert_bn(name) else: continue new_state_dict[name] = torch.from_numpy(v) data = {'state_dict': new_state_dict} torch.save(data, dst) def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument( '--src', default='ppyoloe_plus_crn_s_80e_coco.pdparams', help='src ppyoloe model path') parser.add_argument( '--dst', default='mmppyoloe_plus_s.pt', help='save path') parser.add_argument( '--imagenet-pretrain', action='store_true', default=False, help='Load model pretrained on imagenet dataset which only ' 'have weight for backbone.') args = parser.parse_args() convert(args.src, args.dst, args.imagenet_pretrain) if __name__ == '__main__': main()
7,738
40.832432
78
py
mmyolo
mmyolo-main/tools/model_converters/convert_kd_ckpt_to_student.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse from pathlib import Path import torch def parse_args(): parser = argparse.ArgumentParser( description='Convert KD checkpoint to student-only checkpoint') parser.add_argument('checkpoint', help='input checkpoint filename') parser.add_argument('--out-path', help='save checkpoint path') parser.add_argument( '--inplace', action='store_true', help='replace origin ckpt') args = parser.parse_args() return args def main(): args = parse_args() checkpoint = torch.load(args.checkpoint, map_location='cpu') new_state_dict = dict() new_meta = checkpoint['meta'] for key, value in checkpoint['state_dict'].items(): if key.startswith('architecture.'): new_key = key.replace('architecture.', '') new_state_dict[new_key] = value checkpoint = dict() checkpoint['meta'] = new_meta checkpoint['state_dict'] = new_state_dict if args.inplace: torch.save(checkpoint, args.checkpoint) else: ckpt_path = Path(args.checkpoint) ckpt_name = ckpt_path.stem if args.out_path: ckpt_dir = Path(args.out_path) else: ckpt_dir = ckpt_path.parent new_ckpt_path = ckpt_dir / f'{ckpt_name}_student.pth' torch.save(checkpoint, new_ckpt_path) if __name__ == '__main__': main()
1,412
27.836735
71
py
mmyolo
mmyolo-main/tools/model_converters/yolov7_to_mmyolo.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import os.path as osp from collections import OrderedDict import torch convert_dict_tiny = { # stem 'model.0': 'backbone.stem.0', 'model.1': 'backbone.stem.1', # stage1 TinyDownSampleBlock 'model.2': 'backbone.stage1.0.short_conv', 'model.3': 'backbone.stage1.0.main_convs.0', 'model.4': 'backbone.stage1.0.main_convs.1', 'model.5': 'backbone.stage1.0.main_convs.2', 'model.7': 'backbone.stage1.0.final_conv', # stage2 TinyDownSampleBlock 'model.9': 'backbone.stage2.1.short_conv', 'model.10': 'backbone.stage2.1.main_convs.0', 'model.11': 'backbone.stage2.1.main_convs.1', 'model.12': 'backbone.stage2.1.main_convs.2', 'model.14': 'backbone.stage2.1.final_conv', # stage3 TinyDownSampleBlock 'model.16': 'backbone.stage3.1.short_conv', 'model.17': 'backbone.stage3.1.main_convs.0', 'model.18': 'backbone.stage3.1.main_convs.1', 'model.19': 'backbone.stage3.1.main_convs.2', 'model.21': 'backbone.stage3.1.final_conv', # stage4 TinyDownSampleBlock 'model.23': 'backbone.stage4.1.short_conv', 'model.24': 'backbone.stage4.1.main_convs.0', 'model.25': 'backbone.stage4.1.main_convs.1', 'model.26': 'backbone.stage4.1.main_convs.2', 'model.28': 'backbone.stage4.1.final_conv', # neck SPPCSPBlock 'model.29': 'neck.reduce_layers.2.short_layer', 'model.30': 'neck.reduce_layers.2.main_layers', 'model.35': 'neck.reduce_layers.2.fuse_layers', 'model.37': 'neck.reduce_layers.2.final_conv', 'model.38': 'neck.upsample_layers.0.0', 'model.40': 'neck.reduce_layers.1', 'model.42': 'neck.top_down_layers.0.short_conv', 'model.43': 'neck.top_down_layers.0.main_convs.0', 'model.44': 'neck.top_down_layers.0.main_convs.1', 'model.45': 'neck.top_down_layers.0.main_convs.2', 'model.47': 'neck.top_down_layers.0.final_conv', 'model.48': 'neck.upsample_layers.1.0', 'model.50': 'neck.reduce_layers.0', 'model.52': 'neck.top_down_layers.1.short_conv', 'model.53': 'neck.top_down_layers.1.main_convs.0', 'model.54': 'neck.top_down_layers.1.main_convs.1', 'model.55': 'neck.top_down_layers.1.main_convs.2', 'model.57': 'neck.top_down_layers.1.final_conv', 'model.58': 'neck.downsample_layers.0', 'model.60': 'neck.bottom_up_layers.0.short_conv', 'model.61': 'neck.bottom_up_layers.0.main_convs.0', 'model.62': 'neck.bottom_up_layers.0.main_convs.1', 'model.63': 'neck.bottom_up_layers.0.main_convs.2', 'model.65': 'neck.bottom_up_layers.0.final_conv', 'model.66': 'neck.downsample_layers.1', 'model.68': 'neck.bottom_up_layers.1.short_conv', 'model.69': 'neck.bottom_up_layers.1.main_convs.0', 'model.70': 'neck.bottom_up_layers.1.main_convs.1', 'model.71': 'neck.bottom_up_layers.1.main_convs.2', 'model.73': 'neck.bottom_up_layers.1.final_conv', 'model.74': 'neck.out_layers.0', 'model.75': 'neck.out_layers.1', 'model.76': 'neck.out_layers.2', # head 'model.77.m.0': 'bbox_head.head_module.convs_pred.0.1', 'model.77.m.1': 'bbox_head.head_module.convs_pred.1.1', 'model.77.m.2': 'bbox_head.head_module.convs_pred.2.1' } convert_dict_l = { # stem 'model.0': 'backbone.stem.0', 'model.1': 'backbone.stem.1', 'model.2': 'backbone.stem.2', # stage1 # ConvModule 'model.3': 'backbone.stage1.0', # ELANBlock expand_channel_2x 'model.4': 'backbone.stage1.1.short_conv', 'model.5': 'backbone.stage1.1.main_conv', 'model.6': 'backbone.stage1.1.blocks.0.0', 'model.7': 'backbone.stage1.1.blocks.0.1', 'model.8': 'backbone.stage1.1.blocks.1.0', 'model.9': 'backbone.stage1.1.blocks.1.1', 'model.11': 'backbone.stage1.1.final_conv', # stage2 # MaxPoolBlock reduce_channel_2x 'model.13': 'backbone.stage2.0.maxpool_branches.1', 'model.14': 'backbone.stage2.0.stride_conv_branches.0', 'model.15': 'backbone.stage2.0.stride_conv_branches.1', # ELANBlock expand_channel_2x 'model.17': 'backbone.stage2.1.short_conv', 'model.18': 'backbone.stage2.1.main_conv', 'model.19': 'backbone.stage2.1.blocks.0.0', 'model.20': 'backbone.stage2.1.blocks.0.1', 'model.21': 'backbone.stage2.1.blocks.1.0', 'model.22': 'backbone.stage2.1.blocks.1.1', 'model.24': 'backbone.stage2.1.final_conv', # stage3 # MaxPoolBlock reduce_channel_2x 'model.26': 'backbone.stage3.0.maxpool_branches.1', 'model.27': 'backbone.stage3.0.stride_conv_branches.0', 'model.28': 'backbone.stage3.0.stride_conv_branches.1', # ELANBlock expand_channel_2x 'model.30': 'backbone.stage3.1.short_conv', 'model.31': 'backbone.stage3.1.main_conv', 'model.32': 'backbone.stage3.1.blocks.0.0', 'model.33': 'backbone.stage3.1.blocks.0.1', 'model.34': 'backbone.stage3.1.blocks.1.0', 'model.35': 'backbone.stage3.1.blocks.1.1', 'model.37': 'backbone.stage3.1.final_conv', # stage4 # MaxPoolBlock reduce_channel_2x 'model.39': 'backbone.stage4.0.maxpool_branches.1', 'model.40': 'backbone.stage4.0.stride_conv_branches.0', 'model.41': 'backbone.stage4.0.stride_conv_branches.1', # ELANBlock no_change_channel 'model.43': 'backbone.stage4.1.short_conv', 'model.44': 'backbone.stage4.1.main_conv', 'model.45': 'backbone.stage4.1.blocks.0.0', 'model.46': 'backbone.stage4.1.blocks.0.1', 'model.47': 'backbone.stage4.1.blocks.1.0', 'model.48': 'backbone.stage4.1.blocks.1.1', 'model.50': 'backbone.stage4.1.final_conv', # neck SPPCSPBlock 'model.51.cv1': 'neck.reduce_layers.2.main_layers.0', 'model.51.cv3': 'neck.reduce_layers.2.main_layers.1', 'model.51.cv4': 'neck.reduce_layers.2.main_layers.2', 'model.51.cv5': 'neck.reduce_layers.2.fuse_layers.0', 'model.51.cv6': 'neck.reduce_layers.2.fuse_layers.1', 'model.51.cv2': 'neck.reduce_layers.2.short_layer', 'model.51.cv7': 'neck.reduce_layers.2.final_conv', # neck 'model.52': 'neck.upsample_layers.0.0', 'model.54': 'neck.reduce_layers.1', # neck ELANBlock reduce_channel_2x 'model.56': 'neck.top_down_layers.0.short_conv', 'model.57': 'neck.top_down_layers.0.main_conv', 'model.58': 'neck.top_down_layers.0.blocks.0', 'model.59': 'neck.top_down_layers.0.blocks.1', 'model.60': 'neck.top_down_layers.0.blocks.2', 'model.61': 'neck.top_down_layers.0.blocks.3', 'model.63': 'neck.top_down_layers.0.final_conv', 'model.64': 'neck.upsample_layers.1.0', 'model.66': 'neck.reduce_layers.0', # neck ELANBlock reduce_channel_2x 'model.68': 'neck.top_down_layers.1.short_conv', 'model.69': 'neck.top_down_layers.1.main_conv', 'model.70': 'neck.top_down_layers.1.blocks.0', 'model.71': 'neck.top_down_layers.1.blocks.1', 'model.72': 'neck.top_down_layers.1.blocks.2', 'model.73': 'neck.top_down_layers.1.blocks.3', 'model.75': 'neck.top_down_layers.1.final_conv', # neck MaxPoolBlock no_change_channel 'model.77': 'neck.downsample_layers.0.maxpool_branches.1', 'model.78': 'neck.downsample_layers.0.stride_conv_branches.0', 'model.79': 'neck.downsample_layers.0.stride_conv_branches.1', # neck ELANBlock reduce_channel_2x 'model.81': 'neck.bottom_up_layers.0.short_conv', 'model.82': 'neck.bottom_up_layers.0.main_conv', 'model.83': 'neck.bottom_up_layers.0.blocks.0', 'model.84': 'neck.bottom_up_layers.0.blocks.1', 'model.85': 'neck.bottom_up_layers.0.blocks.2', 'model.86': 'neck.bottom_up_layers.0.blocks.3', 'model.88': 'neck.bottom_up_layers.0.final_conv', # neck MaxPoolBlock no_change_channel 'model.90': 'neck.downsample_layers.1.maxpool_branches.1', 'model.91': 'neck.downsample_layers.1.stride_conv_branches.0', 'model.92': 'neck.downsample_layers.1.stride_conv_branches.1', # neck ELANBlock reduce_channel_2x 'model.94': 'neck.bottom_up_layers.1.short_conv', 'model.95': 'neck.bottom_up_layers.1.main_conv', 'model.96': 'neck.bottom_up_layers.1.blocks.0', 'model.97': 'neck.bottom_up_layers.1.blocks.1', 'model.98': 'neck.bottom_up_layers.1.blocks.2', 'model.99': 'neck.bottom_up_layers.1.blocks.3', 'model.101': 'neck.bottom_up_layers.1.final_conv', # RepVGGBlock 'model.102.rbr_dense.0': 'neck.out_layers.0.rbr_dense.conv', 'model.102.rbr_dense.1': 'neck.out_layers.0.rbr_dense.bn', 'model.102.rbr_1x1.0': 'neck.out_layers.0.rbr_1x1.conv', 'model.102.rbr_1x1.1': 'neck.out_layers.0.rbr_1x1.bn', 'model.103.rbr_dense.0': 'neck.out_layers.1.rbr_dense.conv', 'model.103.rbr_dense.1': 'neck.out_layers.1.rbr_dense.bn', 'model.103.rbr_1x1.0': 'neck.out_layers.1.rbr_1x1.conv', 'model.103.rbr_1x1.1': 'neck.out_layers.1.rbr_1x1.bn', 'model.104.rbr_dense.0': 'neck.out_layers.2.rbr_dense.conv', 'model.104.rbr_dense.1': 'neck.out_layers.2.rbr_dense.bn', 'model.104.rbr_1x1.0': 'neck.out_layers.2.rbr_1x1.conv', 'model.104.rbr_1x1.1': 'neck.out_layers.2.rbr_1x1.bn', # head 'model.105.m.0': 'bbox_head.head_module.convs_pred.0.1', 'model.105.m.1': 'bbox_head.head_module.convs_pred.1.1', 'model.105.m.2': 'bbox_head.head_module.convs_pred.2.1' } convert_dict_x = { # stem 'model.0': 'backbone.stem.0', 'model.1': 'backbone.stem.1', 'model.2': 'backbone.stem.2', # stage1 # ConvModule 'model.3': 'backbone.stage1.0', # ELANBlock expand_channel_2x 'model.4': 'backbone.stage1.1.short_conv', 'model.5': 'backbone.stage1.1.main_conv', 'model.6': 'backbone.stage1.1.blocks.0.0', 'model.7': 'backbone.stage1.1.blocks.0.1', 'model.8': 'backbone.stage1.1.blocks.1.0', 'model.9': 'backbone.stage1.1.blocks.1.1', 'model.10': 'backbone.stage1.1.blocks.2.0', 'model.11': 'backbone.stage1.1.blocks.2.1', 'model.13': 'backbone.stage1.1.final_conv', # stage2 # MaxPoolBlock reduce_channel_2x 'model.15': 'backbone.stage2.0.maxpool_branches.1', 'model.16': 'backbone.stage2.0.stride_conv_branches.0', 'model.17': 'backbone.stage2.0.stride_conv_branches.1', # ELANBlock expand_channel_2x 'model.19': 'backbone.stage2.1.short_conv', 'model.20': 'backbone.stage2.1.main_conv', 'model.21': 'backbone.stage2.1.blocks.0.0', 'model.22': 'backbone.stage2.1.blocks.0.1', 'model.23': 'backbone.stage2.1.blocks.1.0', 'model.24': 'backbone.stage2.1.blocks.1.1', 'model.25': 'backbone.stage2.1.blocks.2.0', 'model.26': 'backbone.stage2.1.blocks.2.1', 'model.28': 'backbone.stage2.1.final_conv', # stage3 # MaxPoolBlock reduce_channel_2x 'model.30': 'backbone.stage3.0.maxpool_branches.1', 'model.31': 'backbone.stage3.0.stride_conv_branches.0', 'model.32': 'backbone.stage3.0.stride_conv_branches.1', # ELANBlock expand_channel_2x 'model.34': 'backbone.stage3.1.short_conv', 'model.35': 'backbone.stage3.1.main_conv', 'model.36': 'backbone.stage3.1.blocks.0.0', 'model.37': 'backbone.stage3.1.blocks.0.1', 'model.38': 'backbone.stage3.1.blocks.1.0', 'model.39': 'backbone.stage3.1.blocks.1.1', 'model.40': 'backbone.stage3.1.blocks.2.0', 'model.41': 'backbone.stage3.1.blocks.2.1', 'model.43': 'backbone.stage3.1.final_conv', # stage4 # MaxPoolBlock reduce_channel_2x 'model.45': 'backbone.stage4.0.maxpool_branches.1', 'model.46': 'backbone.stage4.0.stride_conv_branches.0', 'model.47': 'backbone.stage4.0.stride_conv_branches.1', # ELANBlock no_change_channel 'model.49': 'backbone.stage4.1.short_conv', 'model.50': 'backbone.stage4.1.main_conv', 'model.51': 'backbone.stage4.1.blocks.0.0', 'model.52': 'backbone.stage4.1.blocks.0.1', 'model.53': 'backbone.stage4.1.blocks.1.0', 'model.54': 'backbone.stage4.1.blocks.1.1', 'model.55': 'backbone.stage4.1.blocks.2.0', 'model.56': 'backbone.stage4.1.blocks.2.1', 'model.58': 'backbone.stage4.1.final_conv', # neck SPPCSPBlock 'model.59.cv1': 'neck.reduce_layers.2.main_layers.0', 'model.59.cv3': 'neck.reduce_layers.2.main_layers.1', 'model.59.cv4': 'neck.reduce_layers.2.main_layers.2', 'model.59.cv5': 'neck.reduce_layers.2.fuse_layers.0', 'model.59.cv6': 'neck.reduce_layers.2.fuse_layers.1', 'model.59.cv2': 'neck.reduce_layers.2.short_layer', 'model.59.cv7': 'neck.reduce_layers.2.final_conv', # neck 'model.60': 'neck.upsample_layers.0.0', 'model.62': 'neck.reduce_layers.1', # neck ELANBlock reduce_channel_2x 'model.64': 'neck.top_down_layers.0.short_conv', 'model.65': 'neck.top_down_layers.0.main_conv', 'model.66': 'neck.top_down_layers.0.blocks.0.0', 'model.67': 'neck.top_down_layers.0.blocks.0.1', 'model.68': 'neck.top_down_layers.0.blocks.1.0', 'model.69': 'neck.top_down_layers.0.blocks.1.1', 'model.70': 'neck.top_down_layers.0.blocks.2.0', 'model.71': 'neck.top_down_layers.0.blocks.2.1', 'model.73': 'neck.top_down_layers.0.final_conv', 'model.74': 'neck.upsample_layers.1.0', 'model.76': 'neck.reduce_layers.0', # neck ELANBlock reduce_channel_2x 'model.78': 'neck.top_down_layers.1.short_conv', 'model.79': 'neck.top_down_layers.1.main_conv', 'model.80': 'neck.top_down_layers.1.blocks.0.0', 'model.81': 'neck.top_down_layers.1.blocks.0.1', 'model.82': 'neck.top_down_layers.1.blocks.1.0', 'model.83': 'neck.top_down_layers.1.blocks.1.1', 'model.84': 'neck.top_down_layers.1.blocks.2.0', 'model.85': 'neck.top_down_layers.1.blocks.2.1', 'model.87': 'neck.top_down_layers.1.final_conv', # neck MaxPoolBlock no_change_channel 'model.89': 'neck.downsample_layers.0.maxpool_branches.1', 'model.90': 'neck.downsample_layers.0.stride_conv_branches.0', 'model.91': 'neck.downsample_layers.0.stride_conv_branches.1', # neck ELANBlock reduce_channel_2x 'model.93': 'neck.bottom_up_layers.0.short_conv', 'model.94': 'neck.bottom_up_layers.0.main_conv', 'model.95': 'neck.bottom_up_layers.0.blocks.0.0', 'model.96': 'neck.bottom_up_layers.0.blocks.0.1', 'model.97': 'neck.bottom_up_layers.0.blocks.1.0', 'model.98': 'neck.bottom_up_layers.0.blocks.1.1', 'model.99': 'neck.bottom_up_layers.0.blocks.2.0', 'model.100': 'neck.bottom_up_layers.0.blocks.2.1', 'model.102': 'neck.bottom_up_layers.0.final_conv', # neck MaxPoolBlock no_change_channel 'model.104': 'neck.downsample_layers.1.maxpool_branches.1', 'model.105': 'neck.downsample_layers.1.stride_conv_branches.0', 'model.106': 'neck.downsample_layers.1.stride_conv_branches.1', # neck ELANBlock reduce_channel_2x 'model.108': 'neck.bottom_up_layers.1.short_conv', 'model.109': 'neck.bottom_up_layers.1.main_conv', 'model.110': 'neck.bottom_up_layers.1.blocks.0.0', 'model.111': 'neck.bottom_up_layers.1.blocks.0.1', 'model.112': 'neck.bottom_up_layers.1.blocks.1.0', 'model.113': 'neck.bottom_up_layers.1.blocks.1.1', 'model.114': 'neck.bottom_up_layers.1.blocks.2.0', 'model.115': 'neck.bottom_up_layers.1.blocks.2.1', 'model.117': 'neck.bottom_up_layers.1.final_conv', # Conv 'model.118': 'neck.out_layers.0', 'model.119': 'neck.out_layers.1', 'model.120': 'neck.out_layers.2', # head 'model.121.m.0': 'bbox_head.head_module.convs_pred.0.1', 'model.121.m.1': 'bbox_head.head_module.convs_pred.1.1', 'model.121.m.2': 'bbox_head.head_module.convs_pred.2.1' } convert_dict_w = { # stem 'model.1': 'backbone.stem.conv', # stage1 # ConvModule 'model.2': 'backbone.stage1.0', # ELANBlock 'model.3': 'backbone.stage1.1.short_conv', 'model.4': 'backbone.stage1.1.main_conv', 'model.5': 'backbone.stage1.1.blocks.0.0', 'model.6': 'backbone.stage1.1.blocks.0.1', 'model.7': 'backbone.stage1.1.blocks.1.0', 'model.8': 'backbone.stage1.1.blocks.1.1', 'model.10': 'backbone.stage1.1.final_conv', # stage2 'model.11': 'backbone.stage2.0', # ELANBlock 'model.12': 'backbone.stage2.1.short_conv', 'model.13': 'backbone.stage2.1.main_conv', 'model.14': 'backbone.stage2.1.blocks.0.0', 'model.15': 'backbone.stage2.1.blocks.0.1', 'model.16': 'backbone.stage2.1.blocks.1.0', 'model.17': 'backbone.stage2.1.blocks.1.1', 'model.19': 'backbone.stage2.1.final_conv', # stage3 'model.20': 'backbone.stage3.0', # ELANBlock 'model.21': 'backbone.stage3.1.short_conv', 'model.22': 'backbone.stage3.1.main_conv', 'model.23': 'backbone.stage3.1.blocks.0.0', 'model.24': 'backbone.stage3.1.blocks.0.1', 'model.25': 'backbone.stage3.1.blocks.1.0', 'model.26': 'backbone.stage3.1.blocks.1.1', 'model.28': 'backbone.stage3.1.final_conv', # stage4 'model.29': 'backbone.stage4.0', # ELANBlock 'model.30': 'backbone.stage4.1.short_conv', 'model.31': 'backbone.stage4.1.main_conv', 'model.32': 'backbone.stage4.1.blocks.0.0', 'model.33': 'backbone.stage4.1.blocks.0.1', 'model.34': 'backbone.stage4.1.blocks.1.0', 'model.35': 'backbone.stage4.1.blocks.1.1', 'model.37': 'backbone.stage4.1.final_conv', # stage5 'model.38': 'backbone.stage5.0', # ELANBlock 'model.39': 'backbone.stage5.1.short_conv', 'model.40': 'backbone.stage5.1.main_conv', 'model.41': 'backbone.stage5.1.blocks.0.0', 'model.42': 'backbone.stage5.1.blocks.0.1', 'model.43': 'backbone.stage5.1.blocks.1.0', 'model.44': 'backbone.stage5.1.blocks.1.1', 'model.46': 'backbone.stage5.1.final_conv', # neck SPPCSPBlock 'model.47.cv1': 'neck.reduce_layers.3.main_layers.0', 'model.47.cv3': 'neck.reduce_layers.3.main_layers.1', 'model.47.cv4': 'neck.reduce_layers.3.main_layers.2', 'model.47.cv5': 'neck.reduce_layers.3.fuse_layers.0', 'model.47.cv6': 'neck.reduce_layers.3.fuse_layers.1', 'model.47.cv2': 'neck.reduce_layers.3.short_layer', 'model.47.cv7': 'neck.reduce_layers.3.final_conv', # neck 'model.48': 'neck.upsample_layers.0.0', 'model.50': 'neck.reduce_layers.2', # neck ELANBlock 'model.52': 'neck.top_down_layers.0.short_conv', 'model.53': 'neck.top_down_layers.0.main_conv', 'model.54': 'neck.top_down_layers.0.blocks.0', 'model.55': 'neck.top_down_layers.0.blocks.1', 'model.56': 'neck.top_down_layers.0.blocks.2', 'model.57': 'neck.top_down_layers.0.blocks.3', 'model.59': 'neck.top_down_layers.0.final_conv', 'model.60': 'neck.upsample_layers.1.0', 'model.62': 'neck.reduce_layers.1', # neck ELANBlock reduce_channel_2x 'model.64': 'neck.top_down_layers.1.short_conv', 'model.65': 'neck.top_down_layers.1.main_conv', 'model.66': 'neck.top_down_layers.1.blocks.0', 'model.67': 'neck.top_down_layers.1.blocks.1', 'model.68': 'neck.top_down_layers.1.blocks.2', 'model.69': 'neck.top_down_layers.1.blocks.3', 'model.71': 'neck.top_down_layers.1.final_conv', 'model.72': 'neck.upsample_layers.2.0', 'model.74': 'neck.reduce_layers.0', 'model.76': 'neck.top_down_layers.2.short_conv', 'model.77': 'neck.top_down_layers.2.main_conv', 'model.78': 'neck.top_down_layers.2.blocks.0', 'model.79': 'neck.top_down_layers.2.blocks.1', 'model.80': 'neck.top_down_layers.2.blocks.2', 'model.81': 'neck.top_down_layers.2.blocks.3', 'model.83': 'neck.top_down_layers.2.final_conv', 'model.84': 'neck.downsample_layers.0', # neck ELANBlock 'model.86': 'neck.bottom_up_layers.0.short_conv', 'model.87': 'neck.bottom_up_layers.0.main_conv', 'model.88': 'neck.bottom_up_layers.0.blocks.0', 'model.89': 'neck.bottom_up_layers.0.blocks.1', 'model.90': 'neck.bottom_up_layers.0.blocks.2', 'model.91': 'neck.bottom_up_layers.0.blocks.3', 'model.93': 'neck.bottom_up_layers.0.final_conv', 'model.94': 'neck.downsample_layers.1', # neck ELANBlock reduce_channel_2x 'model.96': 'neck.bottom_up_layers.1.short_conv', 'model.97': 'neck.bottom_up_layers.1.main_conv', 'model.98': 'neck.bottom_up_layers.1.blocks.0', 'model.99': 'neck.bottom_up_layers.1.blocks.1', 'model.100': 'neck.bottom_up_layers.1.blocks.2', 'model.101': 'neck.bottom_up_layers.1.blocks.3', 'model.103': 'neck.bottom_up_layers.1.final_conv', 'model.104': 'neck.downsample_layers.2', # neck ELANBlock reduce_channel_2x 'model.106': 'neck.bottom_up_layers.2.short_conv', 'model.107': 'neck.bottom_up_layers.2.main_conv', 'model.108': 'neck.bottom_up_layers.2.blocks.0', 'model.109': 'neck.bottom_up_layers.2.blocks.1', 'model.110': 'neck.bottom_up_layers.2.blocks.2', 'model.111': 'neck.bottom_up_layers.2.blocks.3', 'model.113': 'neck.bottom_up_layers.2.final_conv', 'model.114': 'bbox_head.head_module.main_convs_pred.0.0', 'model.115': 'bbox_head.head_module.main_convs_pred.1.0', 'model.116': 'bbox_head.head_module.main_convs_pred.2.0', 'model.117': 'bbox_head.head_module.main_convs_pred.3.0', # head 'model.118.m.0': 'bbox_head.head_module.main_convs_pred.0.2', 'model.118.m.1': 'bbox_head.head_module.main_convs_pred.1.2', 'model.118.m.2': 'bbox_head.head_module.main_convs_pred.2.2', 'model.118.m.3': 'bbox_head.head_module.main_convs_pred.3.2' } convert_dict_e = { # stem 'model.1': 'backbone.stem.conv', # stage1 'model.2.cv1': 'backbone.stage1.0.stride_conv_branches.0', 'model.2.cv2': 'backbone.stage1.0.stride_conv_branches.1', 'model.2.cv3': 'backbone.stage1.0.maxpool_branches.1', # ELANBlock 'model.3': 'backbone.stage1.1.short_conv', 'model.4': 'backbone.stage1.1.main_conv', 'model.5': 'backbone.stage1.1.blocks.0.0', 'model.6': 'backbone.stage1.1.blocks.0.1', 'model.7': 'backbone.stage1.1.blocks.1.0', 'model.8': 'backbone.stage1.1.blocks.1.1', 'model.9': 'backbone.stage1.1.blocks.2.0', 'model.10': 'backbone.stage1.1.blocks.2.1', 'model.12': 'backbone.stage1.1.final_conv', # stage2 'model.13.cv1': 'backbone.stage2.0.stride_conv_branches.0', 'model.13.cv2': 'backbone.stage2.0.stride_conv_branches.1', 'model.13.cv3': 'backbone.stage2.0.maxpool_branches.1', # ELANBlock 'model.14': 'backbone.stage2.1.short_conv', 'model.15': 'backbone.stage2.1.main_conv', 'model.16': 'backbone.stage2.1.blocks.0.0', 'model.17': 'backbone.stage2.1.blocks.0.1', 'model.18': 'backbone.stage2.1.blocks.1.0', 'model.19': 'backbone.stage2.1.blocks.1.1', 'model.20': 'backbone.stage2.1.blocks.2.0', 'model.21': 'backbone.stage2.1.blocks.2.1', 'model.23': 'backbone.stage2.1.final_conv', # stage3 'model.24.cv1': 'backbone.stage3.0.stride_conv_branches.0', 'model.24.cv2': 'backbone.stage3.0.stride_conv_branches.1', 'model.24.cv3': 'backbone.stage3.0.maxpool_branches.1', # ELANBlock 'model.25': 'backbone.stage3.1.short_conv', 'model.26': 'backbone.stage3.1.main_conv', 'model.27': 'backbone.stage3.1.blocks.0.0', 'model.28': 'backbone.stage3.1.blocks.0.1', 'model.29': 'backbone.stage3.1.blocks.1.0', 'model.30': 'backbone.stage3.1.blocks.1.1', 'model.31': 'backbone.stage3.1.blocks.2.0', 'model.32': 'backbone.stage3.1.blocks.2.1', 'model.34': 'backbone.stage3.1.final_conv', # stage4 'model.35.cv1': 'backbone.stage4.0.stride_conv_branches.0', 'model.35.cv2': 'backbone.stage4.0.stride_conv_branches.1', 'model.35.cv3': 'backbone.stage4.0.maxpool_branches.1', # ELANBlock 'model.36': 'backbone.stage4.1.short_conv', 'model.37': 'backbone.stage4.1.main_conv', 'model.38': 'backbone.stage4.1.blocks.0.0', 'model.39': 'backbone.stage4.1.blocks.0.1', 'model.40': 'backbone.stage4.1.blocks.1.0', 'model.41': 'backbone.stage4.1.blocks.1.1', 'model.42': 'backbone.stage4.1.blocks.2.0', 'model.43': 'backbone.stage4.1.blocks.2.1', 'model.45': 'backbone.stage4.1.final_conv', # stage5 'model.46.cv1': 'backbone.stage5.0.stride_conv_branches.0', 'model.46.cv2': 'backbone.stage5.0.stride_conv_branches.1', 'model.46.cv3': 'backbone.stage5.0.maxpool_branches.1', # ELANBlock 'model.47': 'backbone.stage5.1.short_conv', 'model.48': 'backbone.stage5.1.main_conv', 'model.49': 'backbone.stage5.1.blocks.0.0', 'model.50': 'backbone.stage5.1.blocks.0.1', 'model.51': 'backbone.stage5.1.blocks.1.0', 'model.52': 'backbone.stage5.1.blocks.1.1', 'model.53': 'backbone.stage5.1.blocks.2.0', 'model.54': 'backbone.stage5.1.blocks.2.1', 'model.56': 'backbone.stage5.1.final_conv', # neck SPPCSPBlock 'model.57.cv1': 'neck.reduce_layers.3.main_layers.0', 'model.57.cv3': 'neck.reduce_layers.3.main_layers.1', 'model.57.cv4': 'neck.reduce_layers.3.main_layers.2', 'model.57.cv5': 'neck.reduce_layers.3.fuse_layers.0', 'model.57.cv6': 'neck.reduce_layers.3.fuse_layers.1', 'model.57.cv2': 'neck.reduce_layers.3.short_layer', 'model.57.cv7': 'neck.reduce_layers.3.final_conv', # neck 'model.58': 'neck.upsample_layers.0.0', 'model.60': 'neck.reduce_layers.2', # neck ELANBlock 'model.62': 'neck.top_down_layers.0.short_conv', 'model.63': 'neck.top_down_layers.0.main_conv', 'model.64': 'neck.top_down_layers.0.blocks.0', 'model.65': 'neck.top_down_layers.0.blocks.1', 'model.66': 'neck.top_down_layers.0.blocks.2', 'model.67': 'neck.top_down_layers.0.blocks.3', 'model.68': 'neck.top_down_layers.0.blocks.4', 'model.69': 'neck.top_down_layers.0.blocks.5', 'model.71': 'neck.top_down_layers.0.final_conv', 'model.72': 'neck.upsample_layers.1.0', 'model.74': 'neck.reduce_layers.1', # neck ELANBlock 'model.76': 'neck.top_down_layers.1.short_conv', 'model.77': 'neck.top_down_layers.1.main_conv', 'model.78': 'neck.top_down_layers.1.blocks.0', 'model.79': 'neck.top_down_layers.1.blocks.1', 'model.80': 'neck.top_down_layers.1.blocks.2', 'model.81': 'neck.top_down_layers.1.blocks.3', 'model.82': 'neck.top_down_layers.1.blocks.4', 'model.83': 'neck.top_down_layers.1.blocks.5', 'model.85': 'neck.top_down_layers.1.final_conv', 'model.86': 'neck.upsample_layers.2.0', 'model.88': 'neck.reduce_layers.0', 'model.90': 'neck.top_down_layers.2.short_conv', 'model.91': 'neck.top_down_layers.2.main_conv', 'model.92': 'neck.top_down_layers.2.blocks.0', 'model.93': 'neck.top_down_layers.2.blocks.1', 'model.94': 'neck.top_down_layers.2.blocks.2', 'model.95': 'neck.top_down_layers.2.blocks.3', 'model.96': 'neck.top_down_layers.2.blocks.4', 'model.97': 'neck.top_down_layers.2.blocks.5', 'model.99': 'neck.top_down_layers.2.final_conv', 'model.100.cv1': 'neck.downsample_layers.0.stride_conv_branches.0', 'model.100.cv2': 'neck.downsample_layers.0.stride_conv_branches.1', 'model.100.cv3': 'neck.downsample_layers.0.maxpool_branches.1', # neck ELANBlock 'model.102': 'neck.bottom_up_layers.0.short_conv', 'model.103': 'neck.bottom_up_layers.0.main_conv', 'model.104': 'neck.bottom_up_layers.0.blocks.0', 'model.105': 'neck.bottom_up_layers.0.blocks.1', 'model.106': 'neck.bottom_up_layers.0.blocks.2', 'model.107': 'neck.bottom_up_layers.0.blocks.3', 'model.108': 'neck.bottom_up_layers.0.blocks.4', 'model.109': 'neck.bottom_up_layers.0.blocks.5', 'model.111': 'neck.bottom_up_layers.0.final_conv', 'model.112.cv1': 'neck.downsample_layers.1.stride_conv_branches.0', 'model.112.cv2': 'neck.downsample_layers.1.stride_conv_branches.1', 'model.112.cv3': 'neck.downsample_layers.1.maxpool_branches.1', # neck ELANBlock 'model.114': 'neck.bottom_up_layers.1.short_conv', 'model.115': 'neck.bottom_up_layers.1.main_conv', 'model.116': 'neck.bottom_up_layers.1.blocks.0', 'model.117': 'neck.bottom_up_layers.1.blocks.1', 'model.118': 'neck.bottom_up_layers.1.blocks.2', 'model.119': 'neck.bottom_up_layers.1.blocks.3', 'model.120': 'neck.bottom_up_layers.1.blocks.4', 'model.121': 'neck.bottom_up_layers.1.blocks.5', 'model.123': 'neck.bottom_up_layers.1.final_conv', 'model.124.cv1': 'neck.downsample_layers.2.stride_conv_branches.0', 'model.124.cv2': 'neck.downsample_layers.2.stride_conv_branches.1', 'model.124.cv3': 'neck.downsample_layers.2.maxpool_branches.1', # neck ELANBlock 'model.126': 'neck.bottom_up_layers.2.short_conv', 'model.127': 'neck.bottom_up_layers.2.main_conv', 'model.128': 'neck.bottom_up_layers.2.blocks.0', 'model.129': 'neck.bottom_up_layers.2.blocks.1', 'model.130': 'neck.bottom_up_layers.2.blocks.2', 'model.131': 'neck.bottom_up_layers.2.blocks.3', 'model.132': 'neck.bottom_up_layers.2.blocks.4', 'model.133': 'neck.bottom_up_layers.2.blocks.5', 'model.135': 'neck.bottom_up_layers.2.final_conv', 'model.136': 'bbox_head.head_module.main_convs_pred.0.0', 'model.137': 'bbox_head.head_module.main_convs_pred.1.0', 'model.138': 'bbox_head.head_module.main_convs_pred.2.0', 'model.139': 'bbox_head.head_module.main_convs_pred.3.0', # head 'model.140.m.0': 'bbox_head.head_module.main_convs_pred.0.2', 'model.140.m.1': 'bbox_head.head_module.main_convs_pred.1.2', 'model.140.m.2': 'bbox_head.head_module.main_convs_pred.2.2', 'model.140.m.3': 'bbox_head.head_module.main_convs_pred.3.2' } convert_dict_e2e = { # stem 'model.1': 'backbone.stem.conv', # stage1 'model.2.cv1': 'backbone.stage1.0.stride_conv_branches.0', 'model.2.cv2': 'backbone.stage1.0.stride_conv_branches.1', 'model.2.cv3': 'backbone.stage1.0.maxpool_branches.1', # E-ELANBlock 'model.3': 'backbone.stage1.1.e_elan_blocks.0.short_conv', 'model.4': 'backbone.stage1.1.e_elan_blocks.0.main_conv', 'model.5': 'backbone.stage1.1.e_elan_blocks.0.blocks.0.0', 'model.6': 'backbone.stage1.1.e_elan_blocks.0.blocks.0.1', 'model.7': 'backbone.stage1.1.e_elan_blocks.0.blocks.1.0', 'model.8': 'backbone.stage1.1.e_elan_blocks.0.blocks.1.1', 'model.9': 'backbone.stage1.1.e_elan_blocks.0.blocks.2.0', 'model.10': 'backbone.stage1.1.e_elan_blocks.0.blocks.2.1', 'model.12': 'backbone.stage1.1.e_elan_blocks.0.final_conv', 'model.13': 'backbone.stage1.1.e_elan_blocks.1.short_conv', 'model.14': 'backbone.stage1.1.e_elan_blocks.1.main_conv', 'model.15': 'backbone.stage1.1.e_elan_blocks.1.blocks.0.0', 'model.16': 'backbone.stage1.1.e_elan_blocks.1.blocks.0.1', 'model.17': 'backbone.stage1.1.e_elan_blocks.1.blocks.1.0', 'model.18': 'backbone.stage1.1.e_elan_blocks.1.blocks.1.1', 'model.19': 'backbone.stage1.1.e_elan_blocks.1.blocks.2.0', 'model.20': 'backbone.stage1.1.e_elan_blocks.1.blocks.2.1', 'model.22': 'backbone.stage1.1.e_elan_blocks.1.final_conv', # stage2 'model.24.cv1': 'backbone.stage2.0.stride_conv_branches.0', 'model.24.cv2': 'backbone.stage2.0.stride_conv_branches.1', 'model.24.cv3': 'backbone.stage2.0.maxpool_branches.1', # E-ELANBlock 'model.25': 'backbone.stage2.1.e_elan_blocks.0.short_conv', 'model.26': 'backbone.stage2.1.e_elan_blocks.0.main_conv', 'model.27': 'backbone.stage2.1.e_elan_blocks.0.blocks.0.0', 'model.28': 'backbone.stage2.1.e_elan_blocks.0.blocks.0.1', 'model.29': 'backbone.stage2.1.e_elan_blocks.0.blocks.1.0', 'model.30': 'backbone.stage2.1.e_elan_blocks.0.blocks.1.1', 'model.31': 'backbone.stage2.1.e_elan_blocks.0.blocks.2.0', 'model.32': 'backbone.stage2.1.e_elan_blocks.0.blocks.2.1', 'model.34': 'backbone.stage2.1.e_elan_blocks.0.final_conv', 'model.35': 'backbone.stage2.1.e_elan_blocks.1.short_conv', 'model.36': 'backbone.stage2.1.e_elan_blocks.1.main_conv', 'model.37': 'backbone.stage2.1.e_elan_blocks.1.blocks.0.0', 'model.38': 'backbone.stage2.1.e_elan_blocks.1.blocks.0.1', 'model.39': 'backbone.stage2.1.e_elan_blocks.1.blocks.1.0', 'model.40': 'backbone.stage2.1.e_elan_blocks.1.blocks.1.1', 'model.41': 'backbone.stage2.1.e_elan_blocks.1.blocks.2.0', 'model.42': 'backbone.stage2.1.e_elan_blocks.1.blocks.2.1', 'model.44': 'backbone.stage2.1.e_elan_blocks.1.final_conv', # stage3 'model.46.cv1': 'backbone.stage3.0.stride_conv_branches.0', 'model.46.cv2': 'backbone.stage3.0.stride_conv_branches.1', 'model.46.cv3': 'backbone.stage3.0.maxpool_branches.1', # E-ELANBlock 'model.47': 'backbone.stage3.1.e_elan_blocks.0.short_conv', 'model.48': 'backbone.stage3.1.e_elan_blocks.0.main_conv', 'model.49': 'backbone.stage3.1.e_elan_blocks.0.blocks.0.0', 'model.50': 'backbone.stage3.1.e_elan_blocks.0.blocks.0.1', 'model.51': 'backbone.stage3.1.e_elan_blocks.0.blocks.1.0', 'model.52': 'backbone.stage3.1.e_elan_blocks.0.blocks.1.1', 'model.53': 'backbone.stage3.1.e_elan_blocks.0.blocks.2.0', 'model.54': 'backbone.stage3.1.e_elan_blocks.0.blocks.2.1', 'model.56': 'backbone.stage3.1.e_elan_blocks.0.final_conv', 'model.57': 'backbone.stage3.1.e_elan_blocks.1.short_conv', 'model.58': 'backbone.stage3.1.e_elan_blocks.1.main_conv', 'model.59': 'backbone.stage3.1.e_elan_blocks.1.blocks.0.0', 'model.60': 'backbone.stage3.1.e_elan_blocks.1.blocks.0.1', 'model.61': 'backbone.stage3.1.e_elan_blocks.1.blocks.1.0', 'model.62': 'backbone.stage3.1.e_elan_blocks.1.blocks.1.1', 'model.63': 'backbone.stage3.1.e_elan_blocks.1.blocks.2.0', 'model.64': 'backbone.stage3.1.e_elan_blocks.1.blocks.2.1', 'model.66': 'backbone.stage3.1.e_elan_blocks.1.final_conv', # stage4 'model.68.cv1': 'backbone.stage4.0.stride_conv_branches.0', 'model.68.cv2': 'backbone.stage4.0.stride_conv_branches.1', 'model.68.cv3': 'backbone.stage4.0.maxpool_branches.1', # E-ELANBlock 'model.69': 'backbone.stage4.1.e_elan_blocks.0.short_conv', 'model.70': 'backbone.stage4.1.e_elan_blocks.0.main_conv', 'model.71': 'backbone.stage4.1.e_elan_blocks.0.blocks.0.0', 'model.72': 'backbone.stage4.1.e_elan_blocks.0.blocks.0.1', 'model.73': 'backbone.stage4.1.e_elan_blocks.0.blocks.1.0', 'model.74': 'backbone.stage4.1.e_elan_blocks.0.blocks.1.1', 'model.75': 'backbone.stage4.1.e_elan_blocks.0.blocks.2.0', 'model.76': 'backbone.stage4.1.e_elan_blocks.0.blocks.2.1', 'model.78': 'backbone.stage4.1.e_elan_blocks.0.final_conv', 'model.79': 'backbone.stage4.1.e_elan_blocks.1.short_conv', 'model.80': 'backbone.stage4.1.e_elan_blocks.1.main_conv', 'model.81': 'backbone.stage4.1.e_elan_blocks.1.blocks.0.0', 'model.82': 'backbone.stage4.1.e_elan_blocks.1.blocks.0.1', 'model.83': 'backbone.stage4.1.e_elan_blocks.1.blocks.1.0', 'model.84': 'backbone.stage4.1.e_elan_blocks.1.blocks.1.1', 'model.85': 'backbone.stage4.1.e_elan_blocks.1.blocks.2.0', 'model.86': 'backbone.stage4.1.e_elan_blocks.1.blocks.2.1', 'model.88': 'backbone.stage4.1.e_elan_blocks.1.final_conv', # stage5 'model.90.cv1': 'backbone.stage5.0.stride_conv_branches.0', 'model.90.cv2': 'backbone.stage5.0.stride_conv_branches.1', 'model.90.cv3': 'backbone.stage5.0.maxpool_branches.1', # E-ELANBlock 'model.91': 'backbone.stage5.1.e_elan_blocks.0.short_conv', 'model.92': 'backbone.stage5.1.e_elan_blocks.0.main_conv', 'model.93': 'backbone.stage5.1.e_elan_blocks.0.blocks.0.0', 'model.94': 'backbone.stage5.1.e_elan_blocks.0.blocks.0.1', 'model.95': 'backbone.stage5.1.e_elan_blocks.0.blocks.1.0', 'model.96': 'backbone.stage5.1.e_elan_blocks.0.blocks.1.1', 'model.97': 'backbone.stage5.1.e_elan_blocks.0.blocks.2.0', 'model.98': 'backbone.stage5.1.e_elan_blocks.0.blocks.2.1', 'model.100': 'backbone.stage5.1.e_elan_blocks.0.final_conv', 'model.101': 'backbone.stage5.1.e_elan_blocks.1.short_conv', 'model.102': 'backbone.stage5.1.e_elan_blocks.1.main_conv', 'model.103': 'backbone.stage5.1.e_elan_blocks.1.blocks.0.0', 'model.104': 'backbone.stage5.1.e_elan_blocks.1.blocks.0.1', 'model.105': 'backbone.stage5.1.e_elan_blocks.1.blocks.1.0', 'model.106': 'backbone.stage5.1.e_elan_blocks.1.blocks.1.1', 'model.107': 'backbone.stage5.1.e_elan_blocks.1.blocks.2.0', 'model.108': 'backbone.stage5.1.e_elan_blocks.1.blocks.2.1', 'model.110': 'backbone.stage5.1.e_elan_blocks.1.final_conv', # neck SPPCSPBlock 'model.112.cv1': 'neck.reduce_layers.3.main_layers.0', 'model.112.cv3': 'neck.reduce_layers.3.main_layers.1', 'model.112.cv4': 'neck.reduce_layers.3.main_layers.2', 'model.112.cv5': 'neck.reduce_layers.3.fuse_layers.0', 'model.112.cv6': 'neck.reduce_layers.3.fuse_layers.1', 'model.112.cv2': 'neck.reduce_layers.3.short_layer', 'model.112.cv7': 'neck.reduce_layers.3.final_conv', # neck 'model.113': 'neck.upsample_layers.0.0', 'model.115': 'neck.reduce_layers.2', # neck E-ELANBlock 'model.117': 'neck.top_down_layers.0.e_elan_blocks.0.short_conv', 'model.118': 'neck.top_down_layers.0.e_elan_blocks.0.main_conv', 'model.119': 'neck.top_down_layers.0.e_elan_blocks.0.blocks.0', 'model.120': 'neck.top_down_layers.0.e_elan_blocks.0.blocks.1', 'model.121': 'neck.top_down_layers.0.e_elan_blocks.0.blocks.2', 'model.122': 'neck.top_down_layers.0.e_elan_blocks.0.blocks.3', 'model.123': 'neck.top_down_layers.0.e_elan_blocks.0.blocks.4', 'model.124': 'neck.top_down_layers.0.e_elan_blocks.0.blocks.5', 'model.126': 'neck.top_down_layers.0.e_elan_blocks.0.final_conv', 'model.127': 'neck.top_down_layers.0.e_elan_blocks.1.short_conv', 'model.128': 'neck.top_down_layers.0.e_elan_blocks.1.main_conv', 'model.129': 'neck.top_down_layers.0.e_elan_blocks.1.blocks.0', 'model.130': 'neck.top_down_layers.0.e_elan_blocks.1.blocks.1', 'model.131': 'neck.top_down_layers.0.e_elan_blocks.1.blocks.2', 'model.132': 'neck.top_down_layers.0.e_elan_blocks.1.blocks.3', 'model.133': 'neck.top_down_layers.0.e_elan_blocks.1.blocks.4', 'model.134': 'neck.top_down_layers.0.e_elan_blocks.1.blocks.5', 'model.136': 'neck.top_down_layers.0.e_elan_blocks.1.final_conv', 'model.138': 'neck.upsample_layers.1.0', 'model.140': 'neck.reduce_layers.1', # neck E-ELANBlock 'model.142': 'neck.top_down_layers.1.e_elan_blocks.0.short_conv', 'model.143': 'neck.top_down_layers.1.e_elan_blocks.0.main_conv', 'model.144': 'neck.top_down_layers.1.e_elan_blocks.0.blocks.0', 'model.145': 'neck.top_down_layers.1.e_elan_blocks.0.blocks.1', 'model.146': 'neck.top_down_layers.1.e_elan_blocks.0.blocks.2', 'model.147': 'neck.top_down_layers.1.e_elan_blocks.0.blocks.3', 'model.148': 'neck.top_down_layers.1.e_elan_blocks.0.blocks.4', 'model.149': 'neck.top_down_layers.1.e_elan_blocks.0.blocks.5', 'model.151': 'neck.top_down_layers.1.e_elan_blocks.0.final_conv', 'model.152': 'neck.top_down_layers.1.e_elan_blocks.1.short_conv', 'model.153': 'neck.top_down_layers.1.e_elan_blocks.1.main_conv', 'model.154': 'neck.top_down_layers.1.e_elan_blocks.1.blocks.0', 'model.155': 'neck.top_down_layers.1.e_elan_blocks.1.blocks.1', 'model.156': 'neck.top_down_layers.1.e_elan_blocks.1.blocks.2', 'model.157': 'neck.top_down_layers.1.e_elan_blocks.1.blocks.3', 'model.158': 'neck.top_down_layers.1.e_elan_blocks.1.blocks.4', 'model.159': 'neck.top_down_layers.1.e_elan_blocks.1.blocks.5', 'model.161': 'neck.top_down_layers.1.e_elan_blocks.1.final_conv', 'model.163': 'neck.upsample_layers.2.0', 'model.165': 'neck.reduce_layers.0', 'model.167': 'neck.top_down_layers.2.e_elan_blocks.0.short_conv', 'model.168': 'neck.top_down_layers.2.e_elan_blocks.0.main_conv', 'model.169': 'neck.top_down_layers.2.e_elan_blocks.0.blocks.0', 'model.170': 'neck.top_down_layers.2.e_elan_blocks.0.blocks.1', 'model.171': 'neck.top_down_layers.2.e_elan_blocks.0.blocks.2', 'model.172': 'neck.top_down_layers.2.e_elan_blocks.0.blocks.3', 'model.173': 'neck.top_down_layers.2.e_elan_blocks.0.blocks.4', 'model.174': 'neck.top_down_layers.2.e_elan_blocks.0.blocks.5', 'model.176': 'neck.top_down_layers.2.e_elan_blocks.0.final_conv', 'model.177': 'neck.top_down_layers.2.e_elan_blocks.1.short_conv', 'model.178': 'neck.top_down_layers.2.e_elan_blocks.1.main_conv', 'model.179': 'neck.top_down_layers.2.e_elan_blocks.1.blocks.0', 'model.180': 'neck.top_down_layers.2.e_elan_blocks.1.blocks.1', 'model.181': 'neck.top_down_layers.2.e_elan_blocks.1.blocks.2', 'model.182': 'neck.top_down_layers.2.e_elan_blocks.1.blocks.3', 'model.183': 'neck.top_down_layers.2.e_elan_blocks.1.blocks.4', 'model.184': 'neck.top_down_layers.2.e_elan_blocks.1.blocks.5', 'model.186': 'neck.top_down_layers.2.e_elan_blocks.1.final_conv', 'model.188.cv1': 'neck.downsample_layers.0.stride_conv_branches.0', 'model.188.cv2': 'neck.downsample_layers.0.stride_conv_branches.1', 'model.188.cv3': 'neck.downsample_layers.0.maxpool_branches.1', # neck E-ELANBlock 'model.190': 'neck.bottom_up_layers.0.e_elan_blocks.0.short_conv', 'model.191': 'neck.bottom_up_layers.0.e_elan_blocks.0.main_conv', 'model.192': 'neck.bottom_up_layers.0.e_elan_blocks.0.blocks.0', 'model.193': 'neck.bottom_up_layers.0.e_elan_blocks.0.blocks.1', 'model.194': 'neck.bottom_up_layers.0.e_elan_blocks.0.blocks.2', 'model.195': 'neck.bottom_up_layers.0.e_elan_blocks.0.blocks.3', 'model.196': 'neck.bottom_up_layers.0.e_elan_blocks.0.blocks.4', 'model.197': 'neck.bottom_up_layers.0.e_elan_blocks.0.blocks.5', 'model.199': 'neck.bottom_up_layers.0.e_elan_blocks.0.final_conv', 'model.200': 'neck.bottom_up_layers.0.e_elan_blocks.1.short_conv', 'model.201': 'neck.bottom_up_layers.0.e_elan_blocks.1.main_conv', 'model.202': 'neck.bottom_up_layers.0.e_elan_blocks.1.blocks.0', 'model.203': 'neck.bottom_up_layers.0.e_elan_blocks.1.blocks.1', 'model.204': 'neck.bottom_up_layers.0.e_elan_blocks.1.blocks.2', 'model.205': 'neck.bottom_up_layers.0.e_elan_blocks.1.blocks.3', 'model.206': 'neck.bottom_up_layers.0.e_elan_blocks.1.blocks.4', 'model.207': 'neck.bottom_up_layers.0.e_elan_blocks.1.blocks.5', 'model.209': 'neck.bottom_up_layers.0.e_elan_blocks.1.final_conv', 'model.211.cv1': 'neck.downsample_layers.1.stride_conv_branches.0', 'model.211.cv2': 'neck.downsample_layers.1.stride_conv_branches.1', 'model.211.cv3': 'neck.downsample_layers.1.maxpool_branches.1', 'model.213': 'neck.bottom_up_layers.1.e_elan_blocks.0.short_conv', 'model.214': 'neck.bottom_up_layers.1.e_elan_blocks.0.main_conv', 'model.215': 'neck.bottom_up_layers.1.e_elan_blocks.0.blocks.0', 'model.216': 'neck.bottom_up_layers.1.e_elan_blocks.0.blocks.1', 'model.217': 'neck.bottom_up_layers.1.e_elan_blocks.0.blocks.2', 'model.218': 'neck.bottom_up_layers.1.e_elan_blocks.0.blocks.3', 'model.219': 'neck.bottom_up_layers.1.e_elan_blocks.0.blocks.4', 'model.220': 'neck.bottom_up_layers.1.e_elan_blocks.0.blocks.5', 'model.222': 'neck.bottom_up_layers.1.e_elan_blocks.0.final_conv', 'model.223': 'neck.bottom_up_layers.1.e_elan_blocks.1.short_conv', 'model.224': 'neck.bottom_up_layers.1.e_elan_blocks.1.main_conv', 'model.225': 'neck.bottom_up_layers.1.e_elan_blocks.1.blocks.0', 'model.226': 'neck.bottom_up_layers.1.e_elan_blocks.1.blocks.1', 'model.227': 'neck.bottom_up_layers.1.e_elan_blocks.1.blocks.2', 'model.228': 'neck.bottom_up_layers.1.e_elan_blocks.1.blocks.3', 'model.229': 'neck.bottom_up_layers.1.e_elan_blocks.1.blocks.4', 'model.230': 'neck.bottom_up_layers.1.e_elan_blocks.1.blocks.5', 'model.232': 'neck.bottom_up_layers.1.e_elan_blocks.1.final_conv', 'model.234.cv1': 'neck.downsample_layers.2.stride_conv_branches.0', 'model.234.cv2': 'neck.downsample_layers.2.stride_conv_branches.1', 'model.234.cv3': 'neck.downsample_layers.2.maxpool_branches.1', # neck E-ELANBlock 'model.236': 'neck.bottom_up_layers.2.e_elan_blocks.0.short_conv', 'model.237': 'neck.bottom_up_layers.2.e_elan_blocks.0.main_conv', 'model.238': 'neck.bottom_up_layers.2.e_elan_blocks.0.blocks.0', 'model.239': 'neck.bottom_up_layers.2.e_elan_blocks.0.blocks.1', 'model.240': 'neck.bottom_up_layers.2.e_elan_blocks.0.blocks.2', 'model.241': 'neck.bottom_up_layers.2.e_elan_blocks.0.blocks.3', 'model.242': 'neck.bottom_up_layers.2.e_elan_blocks.0.blocks.4', 'model.243': 'neck.bottom_up_layers.2.e_elan_blocks.0.blocks.5', 'model.245': 'neck.bottom_up_layers.2.e_elan_blocks.0.final_conv', 'model.246': 'neck.bottom_up_layers.2.e_elan_blocks.1.short_conv', 'model.247': 'neck.bottom_up_layers.2.e_elan_blocks.1.main_conv', 'model.248': 'neck.bottom_up_layers.2.e_elan_blocks.1.blocks.0', 'model.249': 'neck.bottom_up_layers.2.e_elan_blocks.1.blocks.1', 'model.250': 'neck.bottom_up_layers.2.e_elan_blocks.1.blocks.2', 'model.251': 'neck.bottom_up_layers.2.e_elan_blocks.1.blocks.3', 'model.252': 'neck.bottom_up_layers.2.e_elan_blocks.1.blocks.4', 'model.253': 'neck.bottom_up_layers.2.e_elan_blocks.1.blocks.5', 'model.255': 'neck.bottom_up_layers.2.e_elan_blocks.1.final_conv', 'model.257': 'bbox_head.head_module.main_convs_pred.0.0', 'model.258': 'bbox_head.head_module.main_convs_pred.1.0', 'model.259': 'bbox_head.head_module.main_convs_pred.2.0', 'model.260': 'bbox_head.head_module.main_convs_pred.3.0', # head 'model.261.m.0': 'bbox_head.head_module.main_convs_pred.0.2', 'model.261.m.1': 'bbox_head.head_module.main_convs_pred.1.2', 'model.261.m.2': 'bbox_head.head_module.main_convs_pred.2.2', 'model.261.m.3': 'bbox_head.head_module.main_convs_pred.3.2' } convert_dicts = { 'yolov7-tiny.pt': convert_dict_tiny, 'yolov7-w6.pt': convert_dict_w, 'yolov7-e6.pt': convert_dict_e, 'yolov7-e6e.pt': convert_dict_e2e, 'yolov7.pt': convert_dict_l, 'yolov7x.pt': convert_dict_x } def convert(src, dst): src_key = osp.basename(src) convert_dict = convert_dicts[osp.basename(src)] num_levels = 3 if src_key == 'yolov7.pt': indexes = [102, 51] in_channels = [256, 512, 1024] elif src_key == 'yolov7x.pt': indexes = [121, 59] in_channels = [320, 640, 1280] elif src_key == 'yolov7-tiny.pt': indexes = [77, 1000] in_channels = [128, 256, 512] elif src_key == 'yolov7-w6.pt': indexes = [118, 47] in_channels = [256, 512, 768, 1024] num_levels = 4 elif src_key == 'yolov7-e6.pt': indexes = [140, [2, 13, 24, 35, 46, 57, 100, 112, 124]] in_channels = 320, 640, 960, 1280 num_levels = 4 elif src_key == 'yolov7-e6e.pt': indexes = [261, [2, 24, 46, 68, 90, 112, 188, 211, 234]] in_channels = 320, 640, 960, 1280 num_levels = 4 if isinstance(indexes[1], int): indexes[1] = [indexes[1]] """Convert keys in detectron pretrained YOLOv7 models to mmyolo style.""" try: yolov7_model = torch.load(src)['model'].float() blobs = yolov7_model.state_dict() except ModuleNotFoundError: raise RuntimeError( 'This script must be placed under the WongKinYiu/yolov7 repo,' ' because loading the official pretrained model need' ' `model.py` to build model.') state_dict = OrderedDict() for key, weight in blobs.items(): if key.find('anchors') >= 0 or key.find('anchor_grid') >= 0: continue num, module = key.split('.')[1:3] if int(num) < indexes[0] and int(num) not in indexes[1]: prefix = f'model.{num}' new_key = key.replace(prefix, convert_dict[prefix]) state_dict[new_key] = weight print(f'Convert {key} to {new_key}') elif int(num) in indexes[1]: strs_key = key.split('.')[:3] new_key = key.replace('.'.join(strs_key), convert_dict['.'.join(strs_key)]) state_dict[new_key] = weight print(f'Convert {key} to {new_key}') else: strs_key = key.split('.')[:4] new_key = key.replace('.'.join(strs_key), convert_dict['.'.join(strs_key)]) state_dict[new_key] = weight print(f'Convert {key} to {new_key}') # Add ImplicitA and ImplicitM for i in range(num_levels): if num_levels == 3: implicit_a = f'bbox_head.head_module.' \ f'convs_pred.{i}.0.implicit' state_dict[implicit_a] = torch.zeros((1, in_channels[i], 1, 1)) implicit_m = f'bbox_head.head_module.' \ f'convs_pred.{i}.2.implicit' state_dict[implicit_m] = torch.ones((1, 3 * 85, 1, 1)) else: implicit_a = f'bbox_head.head_module.' \ f'main_convs_pred.{i}.1.implicit' state_dict[implicit_a] = torch.zeros((1, in_channels[i], 1, 1)) implicit_m = f'bbox_head.head_module.' \ f'main_convs_pred.{i}.3.implicit' state_dict[implicit_m] = torch.ones((1, 3 * 85, 1, 1)) # save checkpoint checkpoint = dict() checkpoint['state_dict'] = state_dict torch.save(checkpoint, dst) # Note: This script must be placed under the yolov7 repo to run. def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument( 'src', default='yolov7.pt', help='src yolov7 model path') parser.add_argument('dst', default='mm_yolov7l.pt', help='save path') args = parser.parse_args() convert(args.src, args.dst) print('If your model weights are from P6 models, such as W6, E6, D6, \ E6E, the auxiliary training module is not required to be loaded, \ so it is normal for the weights of the auxiliary module \ to be missing.') if __name__ == '__main__': main()
50,022
44.724863
78
py
mmyolo
mmyolo-main/tools/model_converters/yolov5_to_mmyolo.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse from collections import OrderedDict import torch convert_dict_p5 = { 'model.0': 'backbone.stem', 'model.1': 'backbone.stage1.0', 'model.2': 'backbone.stage1.1', 'model.3': 'backbone.stage2.0', 'model.4': 'backbone.stage2.1', 'model.5': 'backbone.stage3.0', 'model.6': 'backbone.stage3.1', 'model.7': 'backbone.stage4.0', 'model.8': 'backbone.stage4.1', 'model.9.cv1': 'backbone.stage4.2.conv1', 'model.9.cv2': 'backbone.stage4.2.conv2', 'model.10': 'neck.reduce_layers.2', 'model.13': 'neck.top_down_layers.0.0', 'model.14': 'neck.top_down_layers.0.1', 'model.17': 'neck.top_down_layers.1', 'model.18': 'neck.downsample_layers.0', 'model.20': 'neck.bottom_up_layers.0', 'model.21': 'neck.downsample_layers.1', 'model.23': 'neck.bottom_up_layers.1', 'model.24.m': 'bbox_head.head_module.convs_pred', } convert_dict_p6 = { 'model.0': 'backbone.stem', 'model.1': 'backbone.stage1.0', 'model.2': 'backbone.stage1.1', 'model.3': 'backbone.stage2.0', 'model.4': 'backbone.stage2.1', 'model.5': 'backbone.stage3.0', 'model.6': 'backbone.stage3.1', 'model.7': 'backbone.stage4.0', 'model.8': 'backbone.stage4.1', 'model.9': 'backbone.stage5.0', 'model.10': 'backbone.stage5.1', 'model.11.cv1': 'backbone.stage5.2.conv1', 'model.11.cv2': 'backbone.stage5.2.conv2', 'model.12': 'neck.reduce_layers.3', 'model.15': 'neck.top_down_layers.0.0', 'model.16': 'neck.top_down_layers.0.1', 'model.19': 'neck.top_down_layers.1.0', 'model.20': 'neck.top_down_layers.1.1', 'model.23': 'neck.top_down_layers.2', 'model.24': 'neck.downsample_layers.0', 'model.26': 'neck.bottom_up_layers.0', 'model.27': 'neck.downsample_layers.1', 'model.29': 'neck.bottom_up_layers.1', 'model.30': 'neck.downsample_layers.2', 'model.32': 'neck.bottom_up_layers.2', 'model.33.m': 'bbox_head.head_module.convs_pred', } def convert(src, dst): """Convert keys in pretrained YOLOv5 models to mmyolo style.""" if src.endswith('6.pt'): convert_dict = convert_dict_p6 is_p6_model = True print('Converting P6 model') else: convert_dict = convert_dict_p5 is_p6_model = False print('Converting P5 model') try: yolov5_model = torch.load(src)['model'] blobs = yolov5_model.state_dict() except ModuleNotFoundError: raise RuntimeError( 'This script must be placed under the ultralytics/yolov5 repo,' ' because loading the official pretrained model need' ' `model.py` to build model.') state_dict = OrderedDict() for key, weight in blobs.items(): num, module = key.split('.')[1:3] if (is_p6_model and (num == '11' or num == '33')) or (not is_p6_model and (num == '9' or num == '24')): if module == 'anchors': continue prefix = f'model.{num}.{module}' else: prefix = f'model.{num}' new_key = key.replace(prefix, convert_dict[prefix]) if '.m.' in new_key: new_key = new_key.replace('.m.', '.blocks.') new_key = new_key.replace('.cv', '.conv') else: new_key = new_key.replace('.cv1', '.main_conv') new_key = new_key.replace('.cv2', '.short_conv') new_key = new_key.replace('.cv3', '.final_conv') state_dict[new_key] = weight print(f'Convert {key} to {new_key}') # save checkpoint checkpoint = dict() checkpoint['state_dict'] = state_dict torch.save(checkpoint, dst) # Note: This script must be placed under the yolov5 repo to run. def main(): parser = argparse.ArgumentParser(description='Convert model keys') parser.add_argument( '--src', default='yolov5s.pt', help='src yolov5 model path') parser.add_argument('--dst', default='mmyolov5s.pt', help='save path') args = parser.parse_args() convert(args.src, args.dst) if __name__ == '__main__': main()
4,176
32.95935
75
py
mmyolo
mmyolo-main/tools/analysis_tools/benchmark.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import copy import os import time import torch from mmengine import Config, DictAction from mmengine.dist import get_world_size, init_dist from mmengine.logging import MMLogger, print_log from mmengine.registry import init_default_scope from mmengine.runner import Runner, load_checkpoint from mmengine.utils import mkdir_or_exist from mmengine.utils.dl_utils import set_multi_processing from mmyolo.registry import MODELS # TODO: Refactoring and improving def parse_args(): parser = argparse.ArgumentParser(description='MMYOLO benchmark a model') parser.add_argument('config', help='test config file path') parser.add_argument('checkpoint', help='checkpoint file') parser.add_argument( '--repeat-num', type=int, default=1, help='number of repeat times of measurement for averaging the results') parser.add_argument( '--max-iter', type=int, default=2000, help='num of max iter') parser.add_argument( '--log-interval', type=int, default=50, help='interval of logging') parser.add_argument( '--work-dir', help='the directory to save the file containing ' 'benchmark metrics') parser.add_argument( '--fuse-conv-bn', action='store_true', help='Whether to fuse conv and bn, this will slightly increase' 'the inference speed') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') parser.add_argument( '--launcher', choices=['none', 'pytorch', 'slurm', 'mpi'], default='none', help='job launcher') parser.add_argument('--local_rank', type=int, default=0) args = parser.parse_args() if 'LOCAL_RANK' not in os.environ: os.environ['LOCAL_RANK'] = str(args.local_rank) return args def measure_inference_speed(cfg, checkpoint, max_iter, log_interval, is_fuse_conv_bn): env_cfg = cfg.get('env_cfg') if env_cfg.get('cudnn_benchmark'): torch.backends.cudnn.benchmark = True mp_cfg: dict = env_cfg.get('mp_cfg', {}) set_multi_processing(**mp_cfg, distributed=cfg.distributed) # Because multiple processes will occupy additional CPU resources, # FPS statistics will be more unstable when num_workers is not 0. # It is reasonable to set num_workers to 0. dataloader_cfg = cfg.test_dataloader dataloader_cfg['num_workers'] = 0 dataloader_cfg['batch_size'] = 1 dataloader_cfg['persistent_workers'] = False data_loader = Runner.build_dataloader(dataloader_cfg) # build the model and load checkpoint model = MODELS.build(cfg.model) load_checkpoint(model, checkpoint, map_location='cpu') model = model.cuda() model.eval() # the first several iterations may be very slow so skip them num_warmup = 5 pure_inf_time = 0 fps = 0 # benchmark with 2000 image and take the average for i, data in enumerate(data_loader): torch.cuda.synchronize() start_time = time.perf_counter() with torch.no_grad(): model.test_step(data) torch.cuda.synchronize() elapsed = time.perf_counter() - start_time if i >= num_warmup: pure_inf_time += elapsed if (i + 1) % log_interval == 0: fps = (i + 1 - num_warmup) / pure_inf_time print_log( f'Done image [{i + 1:<3}/ {max_iter}], ' f'fps: {fps:.1f} img / s, ' f'times per image: {1000 / fps:.1f} ms / img', 'current') if (i + 1) == max_iter: fps = (i + 1 - num_warmup) / pure_inf_time print_log( f'Overall fps: {fps:.1f} img / s, ' f'times per image: {1000 / fps:.1f} ms / img', 'current') break return fps def repeat_measure_inference_speed(cfg, checkpoint, max_iter, log_interval, is_fuse_conv_bn, repeat_num=1): assert repeat_num >= 1 fps_list = [] for _ in range(repeat_num): cp_cfg = copy.deepcopy(cfg) fps_list.append( measure_inference_speed(cp_cfg, checkpoint, max_iter, log_interval, is_fuse_conv_bn)) if repeat_num > 1: fps_list_ = [round(fps, 1) for fps in fps_list] times_pre_image_list_ = [round(1000 / fps, 1) for fps in fps_list] mean_fps_ = sum(fps_list_) / len(fps_list_) mean_times_pre_image_ = sum(times_pre_image_list_) / len( times_pre_image_list_) print_log( f'Overall fps: {fps_list_}[{mean_fps_:.1f}] img / s, ' f'times per image: ' f'{times_pre_image_list_}[{mean_times_pre_image_:.1f}] ms / img', 'current') return fps_list return fps_list[0] # TODO: refactoring def main(): args = parse_args() cfg = Config.fromfile(args.config) if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) init_default_scope(cfg.get('default_scope', 'mmyolo')) distributed = False if args.launcher != 'none': init_dist(args.launcher, **cfg.get('env_cfg', {}).get('dist_cfg', {})) distributed = True assert get_world_size( ) == 1, 'Inference benchmark does not allow distributed multi-GPU' cfg.distributed = distributed log_file = None if args.work_dir: log_file = os.path.join(args.work_dir, 'benchmark.log') mkdir_or_exist(args.work_dir) MMLogger.get_instance('mmyolo', log_file=log_file, log_level='INFO') repeat_measure_inference_speed(cfg, args.checkpoint, args.max_iter, args.log_interval, args.fuse_conv_bn, args.repeat_num) if __name__ == '__main__': main()
6,460
33.185185
79
py
mmyolo
mmyolo-main/tools/analysis_tools/optimize_anchors.py
# Copyright (c) OpenMMLab. All rights reserved. """Optimize anchor settings on a specific dataset. This script provides three methods to optimize YOLO anchors including k-means anchor cluster, differential evolution and v5-k-means. You can use ``--algorithm k-means``, ``--algorithm differential_evolution`` and ``--algorithm v5-k-means`` to switch those methods. Example: Use k-means anchor cluster:: python tools/analysis_tools/optimize_anchors.py ${CONFIG} \ --algorithm k-means --input-shape ${INPUT_SHAPE [WIDTH HEIGHT]} \ --out-dir ${OUT_DIR} Use differential evolution to optimize anchors:: python tools/analysis_tools/optimize_anchors.py ${CONFIG} \ --algorithm differential_evolution \ --input-shape ${INPUT_SHAPE [WIDTH HEIGHT]} \ --out-dir ${OUT_DIR} Use v5-k-means to optimize anchors:: python tools/analysis_tools/optimize_anchors.py ${CONFIG} \ --algorithm v5-k-means \ --input-shape ${INPUT_SHAPE [WIDTH HEIGHT]} \ --prior_match_thr ${PRIOR_MATCH_THR} \ --out-dir ${OUT_DIR} """ import argparse import os.path as osp import random from typing import Tuple import numpy as np import torch from mmdet.structures.bbox import (bbox_cxcywh_to_xyxy, bbox_overlaps, bbox_xyxy_to_cxcywh) from mmdet.utils import replace_cfg_vals, update_data_root from mmengine.config import Config from mmengine.fileio import dump from mmengine.logging import MMLogger from mmengine.registry import init_default_scope from mmengine.utils import ProgressBar from scipy.optimize import differential_evolution from torch import Tensor from mmyolo.registry import DATASETS try: from scipy.cluster.vq import kmeans except ImportError: kmeans = None def parse_args(): parser = argparse.ArgumentParser(description='Optimize anchor parameters.') parser.add_argument('config', help='Train config file path.') parser.add_argument( '--input-shape', type=int, nargs='+', default=[640, 640], help='input image size, represent [width, height]') parser.add_argument( '--algorithm', default='DE', help='Algorithm used for anchor optimizing.' 'Support k-means and differential_evolution for YOLO,' 'and v5-k-means is special for YOLOV5.') parser.add_argument( '--iters', default=1000, type=int, help='Maximum iterations for optimizer.') parser.add_argument( '--prior-match-thr', default=4.0, type=float, help='anchor-label `gt_filter_sizes` ratio threshold ' 'hyperparameter used for training, default=4.0, this ' 'parameter is unique to v5-k-means') parser.add_argument( '--mutation-args', type=float, nargs='+', default=[0.9, 0.1], help='paramter of anchor optimize method genetic algorithm, ' 'represent [prob, sigma], this parameter is unique to v5-k-means') parser.add_argument( '--augment-args', type=float, nargs='+', default=[0.9, 1.1], help='scale factor of box size augment when metric box and anchor, ' 'represent [min, max], this parameter is unique to v5-k-means') parser.add_argument( '--device', default='cuda:0', help='Device used for calculating.') parser.add_argument( '--out-dir', default=None, type=str, help='Path to save anchor optimize result.') args = parser.parse_args() return args class BaseAnchorOptimizer: """Base class for anchor optimizer. Args: dataset (obj:`Dataset`): Dataset object. input_shape (list[int]): Input image shape of the model. Format in [width, height]. num_anchor_per_level (list[int]) : Number of anchors for each level. logger (obj:`logging.Logger`): The logger for logging. device (str, optional): Device used for calculating. Default: 'cuda:0' out_dir (str, optional): Path to save anchor optimize result. Default: None """ def __init__(self, dataset, input_shape, num_anchor_per_level, logger, device='cuda:0', out_dir=None): self.dataset = dataset self.input_shape = input_shape self.num_anchor_per_level = num_anchor_per_level self.num_anchors = sum(num_anchor_per_level) self.logger = logger self.device = device self.out_dir = out_dir bbox_whs, img_shapes = self.get_whs_and_shapes() ratios = img_shapes.max(1, keepdims=True) / np.array([input_shape]) # resize to input shape self.bbox_whs = bbox_whs / ratios def get_whs_and_shapes(self): """Get widths and heights of bboxes and shapes of images. Returns: tuple[np.ndarray]: Array of bbox shapes and array of image shapes with shape (num_bboxes, 2) in [width, height] format. """ self.logger.info('Collecting bboxes from annotation...') bbox_whs = [] img_shapes = [] prog_bar = ProgressBar(len(self.dataset)) for idx in range(len(self.dataset)): data_info = self.dataset.get_data_info(idx) img_shape = np.array([data_info['width'], data_info['height']]) gt_instances = data_info['instances'] for instance in gt_instances: bbox = np.array(instance['bbox']) gt_filter_sizes = bbox[2:4] - bbox[0:2] img_shapes.append(img_shape) bbox_whs.append(gt_filter_sizes) prog_bar.update() print('\n') bbox_whs = np.array(bbox_whs) img_shapes = np.array(img_shapes) self.logger.info(f'Collected {bbox_whs.shape[0]} bboxes.') return bbox_whs, img_shapes def get_zero_center_bbox_tensor(self): """Get a tensor of bboxes centered at (0, 0). Returns: Tensor: Tensor of bboxes with shape (num_bboxes, 4) in [xmin, ymin, xmax, ymax] format. """ whs = torch.from_numpy(self.bbox_whs).to( self.device, dtype=torch.float32) bboxes = bbox_cxcywh_to_xyxy( torch.cat([torch.zeros_like(whs), whs], dim=1)) return bboxes def optimize(self): raise NotImplementedError def save_result(self, anchors, path=None): anchor_results = [] start = 0 for num in self.num_anchor_per_level: end = num + start anchor_results.append([(round(w), round(h)) for w, h in anchors[start:end]]) start = end self.logger.info(f'Anchor optimize result:{anchor_results}') if path: json_path = osp.join(path, 'anchor_optimize_result.json') dump(anchor_results, json_path) self.logger.info(f'Result saved in {json_path}') class YOLOKMeansAnchorOptimizer(BaseAnchorOptimizer): r"""YOLO anchor optimizer using k-means. Code refer to `AlexeyAB/darknet. <https://github.com/AlexeyAB/darknet/blob/master/src/detector.c>`_. Args: iters (int): Maximum iterations for k-means. """ def __init__(self, iters, **kwargs): super().__init__(**kwargs) self.iters = iters def optimize(self): anchors = self.kmeans_anchors() self.save_result(anchors, self.out_dir) def kmeans_anchors(self): self.logger.info( f'Start cluster {self.num_anchors} YOLO anchors with K-means...') bboxes = self.get_zero_center_bbox_tensor() cluster_center_idx = torch.randint( 0, bboxes.shape[0], (self.num_anchors, )).to(self.device) assignments = torch.zeros((bboxes.shape[0], )).to(self.device) cluster_centers = bboxes[cluster_center_idx] if self.num_anchors == 1: cluster_centers = self.kmeans_maximization(bboxes, assignments, cluster_centers) anchors = bbox_xyxy_to_cxcywh(cluster_centers)[:, 2:].cpu().numpy() anchors = sorted(anchors, key=lambda x: x[0] * x[1]) return anchors prog_bar = ProgressBar(self.iters) for i in range(self.iters): converged, assignments = self.kmeans_expectation( bboxes, assignments, cluster_centers) if converged: self.logger.info(f'K-means process has converged at iter {i}.') break cluster_centers = self.kmeans_maximization(bboxes, assignments, cluster_centers) prog_bar.update() print('\n') avg_iou = bbox_overlaps(bboxes, cluster_centers).max(1)[0].mean().item() anchors = bbox_xyxy_to_cxcywh(cluster_centers)[:, 2:].cpu().numpy() anchors = sorted(anchors, key=lambda x: x[0] * x[1]) self.logger.info(f'Anchor cluster finish. Average IOU: {avg_iou}') return anchors def kmeans_maximization(self, bboxes, assignments, centers): """Maximization part of EM algorithm(Expectation-Maximization)""" new_centers = torch.zeros_like(centers) for i in range(centers.shape[0]): mask = (assignments == i) if mask.sum(): new_centers[i, :] = bboxes[mask].mean(0) return new_centers def kmeans_expectation(self, bboxes, assignments, centers): """Expectation part of EM algorithm(Expectation-Maximization)""" ious = bbox_overlaps(bboxes, centers) closest = ious.argmax(1) converged = (closest == assignments).all() return converged, closest class YOLOV5KMeansAnchorOptimizer(BaseAnchorOptimizer): r"""YOLOv5 anchor optimizer using shape k-means. Code refer to `ultralytics/yolov5. <https://github.com/ultralytics/yolov5/blob/master/utils/autoanchor.py>`_. Args: iters (int): Maximum iterations for k-means. prior_match_thr (float): anchor-label width height ratio threshold hyperparameter. """ def __init__(self, iters, prior_match_thr=4.0, mutation_args=[0.9, 0.1], augment_args=[0.9, 1.1], **kwargs): super().__init__(**kwargs) self.iters = iters self.prior_match_thr = prior_match_thr [self.mutation_prob, self.mutation_sigma] = mutation_args [self.augment_min, self.augment_max] = augment_args def optimize(self): self.logger.info( f'Start cluster {self.num_anchors} YOLOv5 anchors with K-means...') bbox_whs = torch.from_numpy(self.bbox_whs).to( self.device, dtype=torch.float32) anchors = self.anchor_generate( bbox_whs, num=self.num_anchors, img_size=self.input_shape[0], prior_match_thr=self.prior_match_thr, iters=self.iters) best_ratio, mean_matched = self.anchor_metric(bbox_whs, anchors) self.logger.info(f'{mean_matched:.2f} anchors/target {best_ratio:.3f} ' 'Best Possible Recall (BPR). ') self.save_result(anchors.tolist(), self.out_dir) def anchor_generate(self, box_size: Tensor, num: int = 9, img_size: int = 640, prior_match_thr: float = 4.0, iters: int = 1000) -> Tensor: """cluster boxes metric with anchors. Args: box_size (Tensor): The size of the bxes, which shape is (box_num, 2),the number 2 means width and height. num (int): number of anchors. img_size (int): image size used for training prior_match_thr (float): width/height ratio threshold used for training iters (int): iterations to evolve anchors using genetic algorithm Returns: anchors (Tensor): kmeans evolved anchors """ thr = 1 / prior_match_thr # step1: filter small bbox box_size = self._filter_box(box_size) assert num <= len(box_size) # step2: init anchors if kmeans: try: self.logger.info( 'beginning init anchors with scipy kmeans method') # sigmas for whitening sigmas = box_size.std(0).cpu().numpy() anchors = kmeans( box_size.cpu().numpy() / sigmas, num, iter=30)[0] * sigmas # kmeans may return fewer points than requested # if width/height is insufficient or too similar assert num == len(anchors) except Exception: self.logger.warning( 'scipy kmeans method cannot get enough points ' 'because of width/height is insufficient or too similar, ' 'now switching strategies from kmeans to random init.') anchors = np.sort(np.random.rand(num * 2)).reshape( num, 2) * img_size else: self.logger.info( 'cannot found scipy package, switching strategies from kmeans ' 'to random init, you can install scipy package to ' 'get better anchor init') anchors = np.sort(np.random.rand(num * 2)).reshape(num, 2) * img_size self.logger.info('init done, beginning evolve anchors...') # sort small to large anchors = torch.tensor(anchors[np.argsort(anchors.prod(1))]).to( box_size.device, dtype=torch.float32) # step3: evolve anchors use Genetic Algorithm prog_bar = ProgressBar(iters) fitness = self._anchor_fitness(box_size, anchors, thr) cluster_shape = anchors.shape for _ in range(iters): mutate_result = np.ones(cluster_shape) # mutate until a change occurs (prevent duplicates) while (mutate_result == 1).all(): # mutate_result is scale factor of anchors, between 0.3 and 3 mutate_result = ( (np.random.random(cluster_shape) < self.mutation_prob) * random.random() * np.random.randn(*cluster_shape) * self.mutation_sigma + 1).clip(0.3, 3.0) mutate_result = torch.from_numpy(mutate_result).to(box_size.device) new_anchors = (anchors.clone() * mutate_result).clip(min=2.0) new_fitness = self._anchor_fitness(box_size, new_anchors, thr) if new_fitness > fitness: fitness = new_fitness anchors = new_anchors.clone() prog_bar.update() print('\n') # sort small to large anchors = anchors[torch.argsort(anchors.prod(1))] self.logger.info(f'Anchor cluster finish. fitness = {fitness:.4f}') return anchors def anchor_metric(self, box_size: Tensor, anchors: Tensor, threshold: float = 4.0) -> Tuple: """compute boxes metric with anchors. Args: box_size (Tensor): The size of the bxes, which shape is (box_num, 2), the number 2 means width and height. anchors (Tensor): The size of the bxes, which shape is (anchor_num, 2), the number 2 means width and height. threshold (float): the compare threshold of ratio Returns: Tuple: a tuple of metric result, best_ratio_mean and mean_matched """ # step1: augment scale # According to the uniform distribution,the scaling scale between # augment_min and augment_max is randomly generated scale = np.random.uniform( self.augment_min, self.augment_max, size=(box_size.shape[0], 1)) box_size = torch.tensor( np.array( [l[:, ] * s for s, l in zip(scale, box_size.cpu().numpy())])).to( box_size.device, dtype=torch.float32) # step2: calculate ratio min_ratio, best_ratio = self._metric(box_size, anchors) mean_matched = (min_ratio > 1 / threshold).float().sum(1).mean() best_ratio_mean = (best_ratio > 1 / threshold).float().mean() return best_ratio_mean, mean_matched def _filter_box(self, box_size: Tensor) -> Tensor: small_cnt = (box_size < 3.0).any(1).sum() if small_cnt: self.logger.warning( f'Extremely small objects found: {small_cnt} ' f'of {len(box_size)} labels are <3 pixels in size') # filter > 2 pixels filter_sizes = box_size[(box_size >= 2.0).any(1)] return filter_sizes def _anchor_fitness(self, box_size: Tensor, anchors: Tensor, thr: float): """mutation fitness.""" _, best = self._metric(box_size, anchors) return (best * (best > thr).float()).mean() def _metric(self, box_size: Tensor, anchors: Tensor) -> Tuple: """compute boxes metric with anchors. Args: box_size (Tensor): The size of the bxes, which shape is (box_num, 2), the number 2 means width and height. anchors (Tensor): The size of the bxes, which shape is (anchor_num, 2), the number 2 means width and height. Returns: Tuple: a tuple of metric result, min_ratio and best_ratio """ # ratio means the (width_1/width_2 and height_1/height_2) ratio of each # box and anchor, the ratio shape is torch.Size([box_num,anchor_num,2]) ratio = box_size[:, None] / anchors[None] # min_ratio records the min ratio of each box with all anchor, # min_ratio.shape is torch.Size([box_num,anchor_num]) # notice: # smaller ratio means worse shape-match between boxes and anchors min_ratio = torch.min(ratio, 1 / ratio).min(2)[0] # find the best shape-match ratio for each box # box_best_ratio.shape is torch.Size([box_num]) best_ratio = min_ratio.max(1)[0] return min_ratio, best_ratio class YOLODEAnchorOptimizer(BaseAnchorOptimizer): """YOLO anchor optimizer using differential evolution algorithm. Args: iters (int): Maximum iterations for k-means. strategy (str): The differential evolution strategy to use. Should be one of: - 'best1bin' - 'best1exp' - 'rand1exp' - 'randtobest1exp' - 'currenttobest1exp' - 'best2exp' - 'rand2exp' - 'randtobest1bin' - 'currenttobest1bin' - 'best2bin' - 'rand2bin' - 'rand1bin' Default: 'best1bin'. population_size (int): Total population size of evolution algorithm. Default: 15. convergence_thr (float): Tolerance for convergence, the optimizing stops when ``np.std(pop) <= abs(convergence_thr) + convergence_thr * np.abs(np.mean(population_energies))``, respectively. Default: 0.0001. mutation (tuple[float]): Range of dithering randomly changes the mutation constant. Default: (0.5, 1). recombination (float): Recombination constant of crossover probability. Default: 0.7. """ def __init__(self, iters, strategy='best1bin', population_size=15, convergence_thr=0.0001, mutation=(0.5, 1), recombination=0.7, **kwargs): super().__init__(**kwargs) self.iters = iters self.strategy = strategy self.population_size = population_size self.convergence_thr = convergence_thr self.mutation = mutation self.recombination = recombination def optimize(self): anchors = self.differential_evolution() self.save_result(anchors, self.out_dir) def differential_evolution(self): bboxes = self.get_zero_center_bbox_tensor() bounds = [] for i in range(self.num_anchors): bounds.extend([(0, self.input_shape[0]), (0, self.input_shape[1])]) result = differential_evolution( func=self.avg_iou_cost, bounds=bounds, args=(bboxes, ), strategy=self.strategy, maxiter=self.iters, popsize=self.population_size, tol=self.convergence_thr, mutation=self.mutation, recombination=self.recombination, updating='immediate', disp=True) self.logger.info( f'Anchor evolution finish. Average IOU: {1 - result.fun}') anchors = [(w, h) for w, h in zip(result.x[::2], result.x[1::2])] anchors = sorted(anchors, key=lambda x: x[0] * x[1]) return anchors @staticmethod def avg_iou_cost(anchor_params, bboxes): assert len(anchor_params) % 2 == 0 anchor_whs = torch.tensor( [[w, h] for w, h in zip(anchor_params[::2], anchor_params[1::2])]).to( bboxes.device, dtype=bboxes.dtype) anchor_boxes = bbox_cxcywh_to_xyxy( torch.cat([torch.zeros_like(anchor_whs), anchor_whs], dim=1)) ious = bbox_overlaps(bboxes, anchor_boxes) max_ious, _ = ious.max(1) cost = 1 - max_ious.mean().item() return cost def main(): logger = MMLogger.get_current_instance() args = parse_args() cfg = args.config cfg = Config.fromfile(cfg) # replace the ${key} with the value of cfg.key cfg = replace_cfg_vals(cfg) # update data root according to MMDET_DATASETS update_data_root(cfg) init_default_scope(cfg.get('default_scope', 'mmyolo')) input_shape = args.input_shape assert len(input_shape) == 2 anchor_type = cfg.model.bbox_head.prior_generator.type assert anchor_type == 'mmdet.YOLOAnchorGenerator', \ f'Only support optimize YOLOAnchor, but get {anchor_type}.' base_sizes = cfg.model.bbox_head.prior_generator.base_sizes num_anchor_per_level = [len(sizes) for sizes in base_sizes] train_data_cfg = cfg.train_dataloader while 'dataset' in train_data_cfg: train_data_cfg = train_data_cfg['dataset'] dataset = DATASETS.build(train_data_cfg) if args.algorithm == 'k-means': optimizer = YOLOKMeansAnchorOptimizer( dataset=dataset, input_shape=input_shape, device=args.device, num_anchor_per_level=num_anchor_per_level, iters=args.iters, logger=logger, out_dir=args.out_dir) elif args.algorithm == 'DE': optimizer = YOLODEAnchorOptimizer( dataset=dataset, input_shape=input_shape, device=args.device, num_anchor_per_level=num_anchor_per_level, iters=args.iters, logger=logger, out_dir=args.out_dir) elif args.algorithm == 'v5-k-means': optimizer = YOLOV5KMeansAnchorOptimizer( dataset=dataset, input_shape=input_shape, device=args.device, num_anchor_per_level=num_anchor_per_level, iters=args.iters, prior_match_thr=args.prior_match_thr, mutation_args=args.mutation_args, augment_args=args.augment_args, logger=logger, out_dir=args.out_dir) else: raise NotImplementedError( f'Only support k-means and differential_evolution, ' f'but get {args.algorithm}') optimizer.optimize() if __name__ == '__main__': main()
24,296
36.49537
79
py
mmyolo
mmyolo-main/tools/analysis_tools/get_flops.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import tempfile from pathlib import Path import torch from mmdet.registry import MODELS from mmengine.analysis import get_model_complexity_info from mmengine.config import Config, DictAction from mmengine.logging import MMLogger from mmengine.model import revert_sync_batchnorm from mmengine.registry import init_default_scope from mmyolo.utils import switch_to_deploy def parse_args(): parser = argparse.ArgumentParser(description='Get a detector flops') parser.add_argument('config', help='train config file path') parser.add_argument( '--shape', type=int, nargs='+', default=[640, 640], help='input image size') parser.add_argument( '--show-arch', action='store_true', help='whether return the statistics in the form of network layers') parser.add_argument( '--not-show-table', action='store_true', help='whether return the statistics in the form of table'), parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') return parser.parse_args() def inference(args, logger): config_name = Path(args.config) if not config_name.exists(): logger.error(f'{config_name} not found.') cfg = Config.fromfile(args.config) cfg.work_dir = tempfile.TemporaryDirectory().name cfg.log_level = 'WARN' if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) init_default_scope(cfg.get('default_scope', 'mmyolo')) if len(args.shape) == 1: h = w = args.shape[0] elif len(args.shape) == 2: h, w = args.shape else: raise ValueError('invalid input shape') # model model = MODELS.build(cfg.model) if torch.cuda.is_available(): model.cuda() model = revert_sync_batchnorm(model) model.eval() switch_to_deploy(model) # input tensor # automatically generate a input tensor with the given input_shape. data_batch = {'inputs': [torch.rand(3, h, w)], 'batch_samples': [None]} data = model.data_preprocessor(data_batch) result = {'ori_shape': (h, w), 'pad_shape': data['inputs'].shape[-2:]} outputs = get_model_complexity_info( model, input_shape=None, inputs=data['inputs'], # the input tensor of the model show_table=not args.not_show_table, # show the complexity table show_arch=args.show_arch) # show the complexity arch result['flops'] = outputs['flops_str'] result['params'] = outputs['params_str'] result['out_table'] = outputs['out_table'] result['out_arch'] = outputs['out_arch'] return result def main(): args = parse_args() logger = MMLogger.get_instance(name='MMLogger') result = inference(args, logger) split_line = '=' * 30 ori_shape = result['ori_shape'] pad_shape = result['pad_shape'] flops = result['flops'] params = result['params'] print(result['out_table']) # print related information by table print(result['out_arch']) # print related information by network layers if pad_shape != ori_shape: print(f'{split_line}\nUse size divisor set input shape ' f'from {ori_shape} to {pad_shape}') print(f'{split_line}\n' f'Input shape: {pad_shape}\nModel Flops: {flops}\n' f'Model Parameters: {params}\n{split_line}') print('!!!Please be cautious if you use the results in papers. ' 'You may need to check if all ops are supported and verify ' 'that the flops computation is correct.') if __name__ == '__main__': main()
4,085
31.951613
78
py
mmyolo
mmyolo-main/tools/analysis_tools/vis_scheduler.py
# Copyright (c) OpenMMLab. All rights reserved. """Hyper-parameter Scheduler Visualization. This tool aims to help the user to check the hyper-parameter scheduler of the optimizer(without training), which support the "learning rate", "momentum", and "weight_decay". Example: ```shell python tools/analysis_tools/vis_scheduler.py \ configs/rtmdet/rtmdet_s_syncbn_fast_8xb32-300e_coco.py \ --dataset-size 118287 \ --ngpus 8 \ --out-dir ./output ``` Modified from: https://github.com/open-mmlab/mmclassification/blob/1.x/tools/visualizations/vis_scheduler.py # noqa """ import argparse import json import os.path as osp import re from pathlib import Path from unittest.mock import MagicMock import matplotlib.pyplot as plt import rich import torch.nn as nn from mmengine.config import Config, DictAction from mmengine.hooks import Hook from mmengine.model import BaseModel from mmengine.registry import init_default_scope from mmengine.runner import Runner from mmengine.utils.path import mkdir_or_exist from mmengine.visualization import Visualizer from rich.progress import BarColumn, MofNCompleteColumn, Progress, TextColumn def parse_args(): parser = argparse.ArgumentParser( description='Visualize a hyper-parameter scheduler') parser.add_argument('config', help='config file path') parser.add_argument( '-p', '--parameter', type=str, default='lr', choices=['lr', 'momentum', 'wd'], help='The parameter to visualize its change curve, choose from' '"lr", "wd" and "momentum". Defaults to "lr".') parser.add_argument( '-d', '--dataset-size', type=int, help='The size of the dataset. If specify, `DATASETS.build` will ' 'be skipped and use this size as the dataset size.') parser.add_argument( '-n', '--ngpus', type=int, default=1, help='The number of GPUs used in training.') parser.add_argument( '-o', '--out-dir', type=Path, help='Path to output file') parser.add_argument( '--log-level', default='WARNING', help='The log level of the handler and logger. Defaults to ' 'WARNING.') parser.add_argument('--title', type=str, help='title of figure') parser.add_argument( '--style', type=str, default='whitegrid', help='style of plt') parser.add_argument('--not-show', default=False, action='store_true') parser.add_argument( '--window-size', default='12*7', help='Size of the window to display images, in format of "$W*$H".') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') args = parser.parse_args() if args.window_size != '': assert re.match(r'\d+\*\d+', args.window_size), \ "'window-size' must be in format 'W*H'." return args class SimpleModel(BaseModel): """simple model that do nothing in train_step.""" def __init__(self): super().__init__() self.data_preprocessor = nn.Identity() self.conv = nn.Conv2d(1, 1, 1) def forward(self, inputs, data_samples, mode='tensor'): pass def train_step(self, data, optim_wrapper): pass class ParamRecordHook(Hook): def __init__(self, by_epoch): super().__init__() self.by_epoch = by_epoch self.lr_list = [] self.momentum_list = [] self.wd_list = [] self.task_id = 0 self.progress = Progress(BarColumn(), MofNCompleteColumn(), TextColumn('{task.description}')) def before_train(self, runner): if self.by_epoch: total = runner.train_loop.max_epochs self.task_id = self.progress.add_task( 'epochs', start=True, total=total) else: total = runner.train_loop.max_iters self.task_id = self.progress.add_task( 'iters', start=True, total=total) self.progress.start() def after_train_epoch(self, runner): if self.by_epoch: self.progress.update(self.task_id, advance=1) # TODO: Support multiple schedulers def after_train_iter(self, runner, batch_idx, data_batch, outputs): if not self.by_epoch: self.progress.update(self.task_id, advance=1) self.lr_list.append(runner.optim_wrapper.get_lr()['lr'][0]) self.momentum_list.append( runner.optim_wrapper.get_momentum()['momentum'][0]) self.wd_list.append( runner.optim_wrapper.param_groups[0]['weight_decay']) def after_train(self, runner): self.progress.stop() def plot_curve(lr_list, args, param_name, iters_per_epoch, by_epoch=True): """Plot learning rate vs iter graph.""" try: import seaborn as sns sns.set_style(args.style) except ImportError: pass wind_w, wind_h = args.window_size.split('*') wind_w, wind_h = int(wind_w), int(wind_h) plt.figure(figsize=(wind_w, wind_h)) ax: plt.Axes = plt.subplot() ax.plot(lr_list, linewidth=1) if by_epoch: ax.xaxis.tick_top() ax.set_xlabel('Iters') ax.xaxis.set_label_position('top') sec_ax = ax.secondary_xaxis( 'bottom', functions=(lambda x: x / iters_per_epoch, lambda y: y * iters_per_epoch)) sec_ax.set_xlabel('Epochs') else: plt.xlabel('Iters') plt.ylabel(param_name) if args.title is None: plt.title(f'{osp.basename(args.config)} {param_name} curve') else: plt.title(args.title) def simulate_train(data_loader, cfg, by_epoch): model = SimpleModel() param_record_hook = ParamRecordHook(by_epoch=by_epoch) default_hooks = dict( param_scheduler=cfg.default_hooks['param_scheduler'], runtime_info=None, timer=None, logger=None, checkpoint=None, sampler_seed=None, param_record=param_record_hook) runner = Runner( model=model, work_dir=cfg.work_dir, train_dataloader=data_loader, train_cfg=cfg.train_cfg, log_level=cfg.log_level, optim_wrapper=cfg.optim_wrapper, param_scheduler=cfg.param_scheduler, default_scope=cfg.default_scope, default_hooks=default_hooks, visualizer=MagicMock(spec=Visualizer), custom_hooks=cfg.get('custom_hooks', None)) runner.train() param_dict = dict( lr=param_record_hook.lr_list, momentum=param_record_hook.momentum_list, wd=param_record_hook.wd_list) return param_dict def main(): args = parse_args() cfg = Config.fromfile(args.config) if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) if cfg.get('work_dir', None) is None: # use config filename as default work_dir if cfg.work_dir is None cfg.work_dir = osp.join('./work_dirs', osp.splitext(osp.basename(args.config))[0]) cfg.log_level = args.log_level init_default_scope(cfg.get('default_scope', 'mmyolo')) # init logger print('Param_scheduler :') rich.print_json(json.dumps(cfg.param_scheduler)) # prepare data loader batch_size = cfg.train_dataloader.batch_size * args.ngpus if 'by_epoch' in cfg.train_cfg: by_epoch = cfg.train_cfg.get('by_epoch') elif 'type' in cfg.train_cfg: by_epoch = cfg.train_cfg.get('type') == 'EpochBasedTrainLoop' else: raise ValueError('please set `train_cfg`.') if args.dataset_size is None and by_epoch: from mmyolo.registry import DATASETS dataset_size = len(DATASETS.build(cfg.train_dataloader.dataset)) else: dataset_size = args.dataset_size or batch_size class FakeDataloader(list): dataset = MagicMock(metainfo=None) data_loader = FakeDataloader(range(dataset_size // batch_size)) dataset_info = ( f'\nDataset infos:' f'\n - Dataset size: {dataset_size}' f'\n - Batch size per GPU: {cfg.train_dataloader.batch_size}' f'\n - Number of GPUs: {args.ngpus}' f'\n - Total batch size: {batch_size}') if by_epoch: dataset_info += f'\n - Iterations per epoch: {len(data_loader)}' rich.print(dataset_info + '\n') # simulation training process param_dict = simulate_train(data_loader, cfg, by_epoch) param_list = param_dict[args.parameter] if args.parameter == 'lr': param_name = 'Learning Rate' elif args.parameter == 'momentum': param_name = 'Momentum' else: param_name = 'Weight Decay' plot_curve(param_list, args, param_name, len(data_loader), by_epoch) if args.out_dir: # make dir for output mkdir_or_exist(args.out_dir) # save the graph out_file = osp.join( args.out_dir, f'{osp.basename(args.config)}-{args.parameter}.jpg') plt.savefig(out_file) print(f'\nThe {param_name} graph is saved at {out_file}') if not args.not_show: plt.show() if __name__ == '__main__': main()
9,574
31.347973
115
py
mmyolo
mmyolo-main/projects/assigner_visualization/assigner_visualization.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import os import os.path as osp import sys import warnings import mmcv import numpy as np import torch from mmengine import ProgressBar from mmengine.config import Config, DictAction from mmengine.dataset import COLLATE_FUNCTIONS from mmengine.runner.checkpoint import load_checkpoint from numpy import random from mmyolo.registry import DATASETS, MODELS from mmyolo.utils import register_all_modules from projects.assigner_visualization.dense_heads import (RTMHeadAssigner, YOLOv5HeadAssigner, YOLOv7HeadAssigner, YOLOv8HeadAssigner) from projects.assigner_visualization.visualization import \ YOLOAssignerVisualizer def parse_args(): parser = argparse.ArgumentParser( description='MMYOLO show the positive sample assigning' ' results.') parser.add_argument('config', help='config file path') parser.add_argument('--checkpoint', '-c', type=str, help='checkpoint file') parser.add_argument( '--show-number', '-n', type=int, default=sys.maxsize, help='number of images selected to save, ' 'must bigger than 0. if the number is bigger than length ' 'of dataset, show all the images in dataset; ' 'default "sys.maxsize", show all images in dataset') parser.add_argument( '--output-dir', default='assigned_results', type=str, help='The name of the folder where the image is saved.') parser.add_argument( '--device', default='cuda:0', help='Device used for inference.') parser.add_argument( '--show-prior', default=False, action='store_true', help='Whether to show prior on image.') parser.add_argument( '--not-show-label', default=False, action='store_true', help='Whether to show label on image.') parser.add_argument('--seed', default=-1, type=int, help='random seed') parser.add_argument( '--cfg-options', nargs='+', action=DictAction, help='override some settings in the used config, the key-value pair ' 'in xxx=yyy format will be merged into config file. If the value to ' 'be overwritten is a list, it should be like key="[a,b]" or key=a,b ' 'It also allows nested list/tuple values, e.g. key="[(a,b),(c,d)]" ' 'Note that the quotation marks are necessary and that no white space ' 'is allowed.') args = parser.parse_args() return args def main(): args = parse_args() register_all_modules() # set random seed seed = int(args.seed) if seed != -1: print(f'Set the global seed: {seed}') random.seed(int(args.seed)) cfg = Config.fromfile(args.config) if args.cfg_options is not None: cfg.merge_from_dict(args.cfg_options) # build model model = MODELS.build(cfg.model) if args.checkpoint is not None: load_checkpoint(model, args.checkpoint) elif isinstance(model.bbox_head, (YOLOv7HeadAssigner, RTMHeadAssigner)): warnings.warn( 'if you use dynamic_assignment methods such as YOLOv7 or ' 'YOLOv8 or RTMDet assigner, please load the checkpoint.') assert isinstance(model.bbox_head, (YOLOv5HeadAssigner, YOLOv7HeadAssigner, YOLOv8HeadAssigner, RTMHeadAssigner)), \ 'Now, this script only support YOLOv5, YOLOv7, YOLOv8 and RTMdet, ' \ 'and bbox_head must use ' \ '`YOLOv5HeadAssigner or YOLOv7HeadAssigne or YOLOv8HeadAssigner ' \ 'or RTMHeadAssigner`. Please use `' \ 'yolov5_s-v61_syncbn_fast_8xb16-300e_coco_assignervisualization.py' \ 'or yolov7_tiny_syncbn_fast_8x16b-300e_coco_assignervisualization.py' \ 'or yolov8_s_syncbn_fast_8xb16-500e_coco_assignervisualization.py' \ 'or rtmdet_s_syncbn_fast_8xb32-300e_coco_assignervisualization.py' \ """` as config file.""" model.eval() model.to(args.device) # build dataset dataset_cfg = cfg.get('train_dataloader').get('dataset') dataset = DATASETS.build(dataset_cfg) # get collate_fn collate_fn_cfg = cfg.get('train_dataloader').pop( 'collate_fn', dict(type='pseudo_collate')) collate_fn_type = collate_fn_cfg.pop('type') collate_fn = COLLATE_FUNCTIONS.get(collate_fn_type) # init visualizer visualizer = YOLOAssignerVisualizer( vis_backends=[{ 'type': 'LocalVisBackend' }], name='visualizer') visualizer.dataset_meta = dataset.metainfo # need priors size to draw priors if hasattr(model.bbox_head.prior_generator, 'base_anchors'): visualizer.priors_size = model.bbox_head.prior_generator.base_anchors # make output dir os.makedirs(args.output_dir, exist_ok=True) print('Results will save to ', args.output_dir) # init visualization image number assert args.show_number > 0 display_number = min(args.show_number, len(dataset)) progress_bar = ProgressBar(display_number) for ind_img in range(display_number): data = dataset.prepare_data(ind_img) if data is None: print('Unable to visualize {} due to strong data augmentations'. format(dataset[ind_img]['data_samples'].img_path)) continue # convert data to batch format batch_data = collate_fn([data]) with torch.no_grad(): assign_results = model.assign(batch_data) img = data['inputs'].cpu().numpy().astype(np.uint8).transpose( (1, 2, 0)) # bgr2rgb img = mmcv.bgr2rgb(img) gt_instances = data['data_samples'].gt_instances img_show = visualizer.draw_assign(img, assign_results, gt_instances, args.show_prior, args.not_show_label) if hasattr(data['data_samples'], 'img_path'): filename = osp.basename(data['data_samples'].img_path) else: # some dataset have not image path filename = f'{ind_img}.jpg' out_file = osp.join(args.output_dir, filename) # convert rgb 2 bgr and save img mmcv.imwrite(mmcv.rgb2bgr(img_show), out_file) progress_bar.update() if __name__ == '__main__': main()
6,558
35.848315
79
py
mmyolo
mmyolo-main/projects/assigner_visualization/visualization/assigner_visualizer.py
# Copyright (c) OpenMMLab. All rights reserved. import math from typing import List, Union import mmcv import numpy as np import torch from mmdet.structures.bbox import HorizontalBoxes from mmdet.visualization import DetLocalVisualizer from mmdet.visualization.palette import _get_adaptive_scales, get_palette from mmengine.structures import InstanceData from torch import Tensor from mmyolo.registry import VISUALIZERS @VISUALIZERS.register_module() class YOLOAssignerVisualizer(DetLocalVisualizer): """MMYOLO Detection Assigner Visualizer. This class is provided to the `assigner_visualization.py` script. Args: name (str): Name of the instance. Defaults to 'visualizer'. """ def __init__(self, name: str = 'visualizer', *args, **kwargs): super().__init__(name=name, *args, **kwargs) # need priors_size from config self.priors_size = None def draw_grid(self, stride: int = 8, line_styles: Union[str, List[str]] = ':', colors: Union[str, tuple, List[str], List[tuple]] = (180, 180, 180), line_widths: Union[Union[int, float], List[Union[int, float]]] = 1): """Draw grids on image. Args: stride (int): Downsample factor of feature map. line_styles (Union[str, List[str]]): The linestyle of lines. ``line_styles`` can have the same length with texts or just single value. If ``line_styles`` is single value, all the lines will have the same linestyle. Reference to https://matplotlib.org/stable/api/collections_api.html?highlight=collection#matplotlib.collections.AsteriskPolygonCollection.set_linestyle for more details. Defaults to ':'. colors (Union[str, tuple, List[str], List[tuple]]): The colors of lines. ``colors`` can have the same length with lines or just single value. If ``colors`` is single value, all the lines will have the same colors. Reference to https://matplotlib.org/stable/gallery/color/named_colors.html for more details. Defaults to (180, 180, 180). line_widths (Union[Union[int, float], List[Union[int, float]]]): The linewidth of lines. ``line_widths`` can have the same length with lines or just single value. If ``line_widths`` is single value, all the lines will have the same linewidth. Defaults to 1. """ assert self._image is not None, 'Please set image using `set_image`' # draw vertical lines x_datas_vertical = ((np.arange(self.width // stride - 1) + 1) * stride).reshape((-1, 1)).repeat( 2, axis=1) y_datas_vertical = np.array([[0, self.height - 1]]).repeat( self.width // stride - 1, axis=0) self.draw_lines( x_datas_vertical, y_datas_vertical, colors=colors, line_styles=line_styles, line_widths=line_widths) # draw horizontal lines x_datas_horizontal = np.array([[0, self.width - 1]]).repeat( self.height // stride - 1, axis=0) y_datas_horizontal = ((np.arange(self.height // stride - 1) + 1) * stride).reshape((-1, 1)).repeat( 2, axis=1) self.draw_lines( x_datas_horizontal, y_datas_horizontal, colors=colors, line_styles=line_styles, line_widths=line_widths) def draw_instances_assign(self, instances: InstanceData, retained_gt_inds: Tensor, not_show_label: bool = False): """Draw instances of GT. Args: instances (:obj:`InstanceData`): gt_instance. It usually includes ``bboxes`` and ``labels`` attributes. retained_gt_inds (Tensor): The gt indexes assigned as the positive sample in the current prior. not_show_label (bool): Whether to show gt labels on images. """ assert self.dataset_meta is not None classes = self.dataset_meta['classes'] palette = self.dataset_meta['palette'] if len(retained_gt_inds) == 0: return self.get_image() draw_gt_inds = torch.from_numpy( np.array( list(set(retained_gt_inds.cpu().numpy())), dtype=np.int64)) bboxes = instances.bboxes[draw_gt_inds] labels = instances.labels[draw_gt_inds] if not isinstance(bboxes, Tensor): bboxes = bboxes.tensor edge_colors = [palette[i] for i in labels] max_label = int(max(labels) if len(labels) > 0 else 0) text_palette = get_palette(self.text_color, max_label + 1) text_colors = [text_palette[label] for label in labels] self.draw_bboxes( bboxes, edge_colors=edge_colors, alpha=self.alpha, line_widths=self.line_width) if not not_show_label: positions = bboxes[:, :2] + self.line_width areas = (bboxes[:, 3] - bboxes[:, 1]) * ( bboxes[:, 2] - bboxes[:, 0]) scales = _get_adaptive_scales(areas) for i, (pos, label) in enumerate(zip(positions, labels)): label_text = classes[ label] if classes is not None else f'class {label}' self.draw_texts( label_text, pos, colors=text_colors[i], font_sizes=int(13 * scales[i]), bboxes=[{ 'facecolor': 'black', 'alpha': 0.8, 'pad': 0.7, 'edgecolor': 'none' }]) def draw_positive_assign(self, grid_x_inds: Tensor, grid_y_inds: Tensor, class_inds: Tensor, stride: int, bboxes: Union[Tensor, HorizontalBoxes], retained_gt_inds: Tensor, offset: float = 0.5): """ Args: grid_x_inds (Tensor): The X-axis indexes of the positive sample in current prior. grid_y_inds (Tensor): The Y-axis indexes of the positive sample in current prior. class_inds (Tensor): The classes indexes of the positive sample in current prior. stride (int): Downsample factor of feature map. bboxes (Union[Tensor, HorizontalBoxes]): Bounding boxes of GT. retained_gt_inds (Tensor): The gt indexes assigned as the positive sample in the current prior. offset (float): The offset of points, the value is normalized with corresponding stride. Defaults to 0.5. """ if not isinstance(bboxes, Tensor): # Convert HorizontalBoxes to Tensor bboxes = bboxes.tensor # The PALETTE in the dataset_meta is required assert self.dataset_meta is not None palette = self.dataset_meta['palette'] x = ((grid_x_inds + offset) * stride).long() y = ((grid_y_inds + offset) * stride).long() center = torch.stack((x, y), dim=-1) retained_bboxes = bboxes[retained_gt_inds] bbox_wh = retained_bboxes[:, 2:] - retained_bboxes[:, :2] bbox_area = bbox_wh[:, 0] * bbox_wh[:, 1] radius = _get_adaptive_scales(bbox_area) * 4 colors = [palette[i] for i in class_inds] self.draw_circles( center, radius, colors, line_widths=0, face_colors=colors, alpha=1.0) def draw_prior(self, grid_x_inds: Tensor, grid_y_inds: Tensor, class_inds: Tensor, stride: int, feat_ind: int, prior_ind: int, offset: float = 0.5): """Draw priors on image. Args: grid_x_inds (Tensor): The X-axis indexes of the positive sample in current prior. grid_y_inds (Tensor): The Y-axis indexes of the positive sample in current prior. class_inds (Tensor): The classes indexes of the positive sample in current prior. stride (int): Downsample factor of feature map. feat_ind (int): Index of featmap. prior_ind (int): Index of prior in current featmap. offset (float): The offset of points, the value is normalized with corresponding stride. Defaults to 0.5. """ palette = self.dataset_meta['palette'] center_x = ((grid_x_inds + offset) * stride) center_y = ((grid_y_inds + offset) * stride) xyxy = torch.stack((center_x, center_y, center_x, center_y), dim=1) device = xyxy.device if self.priors_size is not None: xyxy += self.priors_size[feat_ind][prior_ind].to(device) else: xyxy += torch.tensor( [[-stride / 2, -stride / 2, stride / 2, stride / 2]], device=device) colors = [palette[i] for i in class_inds] self.draw_bboxes( xyxy, edge_colors=colors, alpha=self.alpha, line_styles='--', line_widths=math.ceil(self.line_width * 0.3)) def draw_assign(self, image: np.ndarray, assign_results: List[List[dict]], gt_instances: InstanceData, show_prior: bool = False, not_show_label: bool = False) -> np.ndarray: """Draw assigning results. Args: image (np.ndarray): The image to draw. assign_results (list): The assigning results. gt_instances (:obj:`InstanceData`): Data structure for instance-level annotations or predictions. show_prior (bool): Whether to show prior on image. not_show_label (bool): Whether to show gt labels on images. Returns: np.ndarray: the drawn image which channel is RGB. """ img_show_list = [] for feat_ind, assign_results_feat in enumerate(assign_results): img_show_list_feat = [] for prior_ind, assign_results_prior in enumerate( assign_results_feat): self.set_image(image) h, w = image.shape[:2] # draw grid stride = assign_results_prior['stride'] self.draw_grid(stride) # draw prior on matched gt grid_x_inds = assign_results_prior['grid_x_inds'] grid_y_inds = assign_results_prior['grid_y_inds'] class_inds = assign_results_prior['class_inds'] prior_ind = assign_results_prior['prior_ind'] offset = assign_results_prior.get('offset', 0.5) if show_prior: self.draw_prior(grid_x_inds, grid_y_inds, class_inds, stride, feat_ind, prior_ind, offset) # draw matched gt retained_gt_inds = assign_results_prior['retained_gt_inds'] self.draw_instances_assign(gt_instances, retained_gt_inds, not_show_label) # draw positive self.draw_positive_assign(grid_x_inds, grid_y_inds, class_inds, stride, gt_instances.bboxes, retained_gt_inds, offset) # draw title if self.priors_size is not None: base_prior = self.priors_size[feat_ind][prior_ind] else: base_prior = [stride, stride, stride * 2, stride * 2] prior_size = (base_prior[2] - base_prior[0], base_prior[3] - base_prior[1]) pos = np.array((20, 20)) text = f'feat_ind: {feat_ind} ' \ f'prior_ind: {prior_ind} ' \ f'prior_size: ({prior_size[0]}, {prior_size[1]})' scales = _get_adaptive_scales(np.array([h * w / 16])) font_sizes = int(13 * scales) self.draw_texts( text, pos, colors=self.text_color, font_sizes=font_sizes, bboxes=[{ 'facecolor': 'black', 'alpha': 0.8, 'pad': 0.7, 'edgecolor': 'none' }]) img_show = self.get_image() img_show = mmcv.impad(img_show, padding=(5, 5, 5, 5)) img_show_list_feat.append(img_show) img_show_list.append(np.concatenate(img_show_list_feat, axis=1)) # Merge all images into one image # setting axis is to beautify the merged image axis = 0 if len(assign_results[0]) > 1 else 1 return np.concatenate(img_show_list, axis=axis)
13,691
40.87156
154
py
mmyolo
mmyolo-main/projects/assigner_visualization/dense_heads/yolov5_head_assigner.py
# Copyright (c) OpenMMLab. All rights reserved. from typing import Sequence, Union import torch from mmdet.models.utils import unpack_gt_instances from mmengine.structures import InstanceData from torch import Tensor from mmyolo.models import YOLOv5Head from mmyolo.registry import MODELS @MODELS.register_module() class YOLOv5HeadAssigner(YOLOv5Head): def assign_by_gt_and_feat( self, batch_gt_instances: Sequence[InstanceData], batch_img_metas: Sequence[dict], inputs_hw: Union[Tensor, tuple] = (640, 640) ) -> dict: """Calculate the assigning results based on the gt and features extracted by the detection head. Args: batch_gt_instances (Sequence[InstanceData]): Batch of gt_instance. It usually includes ``bboxes`` and ``labels`` attributes. batch_img_metas (Sequence[dict]): Meta information of each image, e.g., image size, scaling factor, etc. batch_gt_instances_ignore (list[:obj:`InstanceData`], optional): Batch of gt_instances_ignore. It includes ``bboxes`` attribute data that is ignored during training and testing. Defaults to None. inputs_hw (Union[Tensor, tuple]): Height and width of inputs size. Returns: dict[str, Tensor]: A dictionary of assigning results. """ # 1. Convert gt to norm format batch_targets_normed = self._convert_gt_to_norm_format( batch_gt_instances, batch_img_metas) device = batch_targets_normed.device scaled_factor = torch.ones(7, device=device) gt_inds = torch.arange( batch_targets_normed.shape[1], dtype=torch.long, device=device, requires_grad=False).unsqueeze(0).repeat((self.num_base_priors, 1)) assign_results = [] for i in range(self.num_levels): assign_results_feat = [] h = inputs_hw[0] // self.featmap_strides[i] w = inputs_hw[1] // self.featmap_strides[i] # empty gt bboxes if batch_targets_normed.shape[1] == 0: for k in range(self.num_base_priors): assign_results_feat.append({ 'stride': self.featmap_strides[i], 'grid_x_inds': torch.zeros([0], dtype=torch.int64).to(device), 'grid_y_inds': torch.zeros([0], dtype=torch.int64).to(device), 'img_inds': torch.zeros([0], dtype=torch.int64).to(device), 'class_inds': torch.zeros([0], dtype=torch.int64).to(device), 'retained_gt_inds': torch.zeros([0], dtype=torch.int64).to(device), 'prior_ind': k }) assign_results.append(assign_results_feat) continue priors_base_sizes_i = self.priors_base_sizes[i] # feature map scale whwh scaled_factor[2:6] = torch.tensor([w, h, w, h]) # Scale batch_targets from range 0-1 to range 0-features_maps size. # (num_base_priors, num_bboxes, 7) batch_targets_scaled = batch_targets_normed * scaled_factor # 2. Shape match wh_ratio = batch_targets_scaled[..., 4:6] / priors_base_sizes_i[:, None] match_inds = torch.max( wh_ratio, 1 / wh_ratio).max(2)[0] < self.prior_match_thr batch_targets_scaled = batch_targets_scaled[match_inds] match_gt_inds = gt_inds[match_inds] # no gt bbox matches anchor if batch_targets_scaled.shape[0] == 0: for k in range(self.num_base_priors): assign_results_feat.append({ 'stride': self.featmap_strides[i], 'grid_x_inds': torch.zeros([0], dtype=torch.int64).to(device), 'grid_y_inds': torch.zeros([0], dtype=torch.int64).to(device), 'img_inds': torch.zeros([0], dtype=torch.int64).to(device), 'class_inds': torch.zeros([0], dtype=torch.int64).to(device), 'retained_gt_inds': torch.zeros([0], dtype=torch.int64).to(device), 'prior_ind': k }) assign_results.append(assign_results_feat) continue # 3. Positive samples with additional neighbors # check the left, up, right, bottom sides of the # targets grid, and determine whether assigned # them as positive samples as well. batch_targets_cxcy = batch_targets_scaled[:, 2:4] grid_xy = scaled_factor[[2, 3]] - batch_targets_cxcy left, up = ((batch_targets_cxcy % 1 < self.near_neighbor_thr) & (batch_targets_cxcy > 1)).T right, bottom = ((grid_xy % 1 < self.near_neighbor_thr) & (grid_xy > 1)).T offset_inds = torch.stack( (torch.ones_like(left), left, up, right, bottom)) batch_targets_scaled = batch_targets_scaled.repeat( (5, 1, 1))[offset_inds] retained_gt_inds = match_gt_inds.repeat((5, 1))[offset_inds] retained_offsets = self.grid_offset.repeat(1, offset_inds.shape[1], 1)[offset_inds] # prepare pred results and positive sample indexes to # calculate class loss and bbox lo _chunk_targets = batch_targets_scaled.chunk(4, 1) img_class_inds, grid_xy, grid_wh, priors_inds = _chunk_targets priors_inds, (img_inds, class_inds) = priors_inds.long().view( -1), img_class_inds.long().T grid_xy_long = (grid_xy - retained_offsets * self.near_neighbor_thr).long() grid_x_inds, grid_y_inds = grid_xy_long.T for k in range(self.num_base_priors): retained_inds = priors_inds == k assign_results_prior = { 'stride': self.featmap_strides[i], 'grid_x_inds': grid_x_inds[retained_inds], 'grid_y_inds': grid_y_inds[retained_inds], 'img_inds': img_inds[retained_inds], 'class_inds': class_inds[retained_inds], 'retained_gt_inds': retained_gt_inds[retained_inds], 'prior_ind': k } assign_results_feat.append(assign_results_prior) assign_results.append(assign_results_feat) return assign_results def assign(self, batch_data_samples: Union[list, dict], inputs_hw: Union[tuple, torch.Size]) -> dict: """Calculate assigning results. This function is provided to the `assigner_visualization.py` script. Args: batch_data_samples (List[:obj:`DetDataSample`], dict): The Data Samples. It usually includes information such as `gt_instance`, `gt_panoptic_seg` and `gt_sem_seg`. inputs_hw: Height and width of inputs size Returns: dict: A dictionary of assigning components. """ if isinstance(batch_data_samples, list): outputs = unpack_gt_instances(batch_data_samples) (batch_gt_instances, batch_gt_instances_ignore, batch_img_metas) = outputs assign_inputs = (batch_gt_instances, batch_img_metas, batch_gt_instances_ignore, inputs_hw) else: # Fast version assign_inputs = (batch_data_samples['bboxes_labels'], batch_data_samples['img_metas'], inputs_hw) assign_results = self.assign_by_gt_and_feat(*assign_inputs) return assign_results
8,364
43.259259
79
py
mmyolo
mmyolo-main/projects/assigner_visualization/dense_heads/yolov8_head_assigner.py
# Copyright (c) OpenMMLab. All rights reserved. from typing import List, Union import torch from mmdet.utils import InstanceList from torch import Tensor from mmyolo.models import YOLOv8Head from mmyolo.models.utils import gt_instances_preprocess from mmyolo.registry import MODELS @MODELS.register_module() class YOLOv8HeadAssigner(YOLOv8Head): def assign_by_gt_and_feat( self, cls_scores: List[Tensor], bbox_preds: List[Tensor], batch_gt_instances: InstanceList, batch_img_metas: List[dict], inputs_hw: Union[Tensor, tuple] = (640, 640) ) -> dict: """Calculate the assigning results based on the gt and features extracted by the detection head. Args: cls_scores (Sequence[Tensor]): Box scores for each scale level, each is a 4D-tensor, the channel number is num_priors * num_classes. bbox_preds (Sequence[Tensor]): Box energies / deltas for each scale level, each is a 4D-tensor, the channel number is num_priors * 4. bbox_dist_preds (Sequence[Tensor]): Box distribution logits for each scale level with shape (bs, reg_max + 1, H*W, 4). batch_gt_instances (list[:obj:`InstanceData`]): Batch of gt_instance. It usually includes ``bboxes`` and ``labels`` attributes. batch_img_metas (list[dict]): Meta information of each image, e.g., image size, scaling factor, etc. inputs_hw (Union[Tensor, tuple]): Height and width of inputs size. Returns: dict[str, Tensor]: A dictionary of assigning results. """ num_imgs = len(batch_img_metas) device = cls_scores[0].device current_featmap_sizes = [ cls_score.shape[2:] for cls_score in cls_scores ] # If the shape does not equal, generate new one if current_featmap_sizes != self.featmap_sizes_train: self.featmap_sizes_train = current_featmap_sizes mlvl_priors_with_stride = self.prior_generator.grid_priors( self.featmap_sizes_train, dtype=cls_scores[0].dtype, device=device, with_stride=True) self.num_level_priors = [len(n) for n in mlvl_priors_with_stride] self.flatten_priors_train = torch.cat( mlvl_priors_with_stride, dim=0) self.stride_tensor = self.flatten_priors_train[..., [2]] # gt info gt_info = gt_instances_preprocess(batch_gt_instances, num_imgs) gt_labels = gt_info[:, :, :1] gt_bboxes = gt_info[:, :, 1:] # xyxy pad_bbox_flag = (gt_bboxes.sum(-1, keepdim=True) > 0).float() # pred info flatten_cls_preds = [ cls_pred.permute(0, 2, 3, 1).reshape(num_imgs, -1, self.num_classes) for cls_pred in cls_scores ] flatten_pred_bboxes = [ bbox_pred.permute(0, 2, 3, 1).reshape(num_imgs, -1, 4) for bbox_pred in bbox_preds ] # (bs, n, 4 * reg_max) flatten_cls_preds = torch.cat(flatten_cls_preds, dim=1) flatten_pred_bboxes = torch.cat(flatten_pred_bboxes, dim=1) flatten_pred_bboxes = self.bbox_coder.decode( self.flatten_priors_train[..., :2], flatten_pred_bboxes, self.stride_tensor[..., 0]) assigned_result = self.assigner( (flatten_pred_bboxes.detach()).type(gt_bboxes.dtype), flatten_cls_preds.detach().sigmoid(), self.flatten_priors_train, gt_labels, gt_bboxes, pad_bbox_flag) labels = assigned_result['assigned_labels'].reshape(-1) bbox_targets = assigned_result['assigned_bboxes'].reshape(-1, 4) fg_mask_pre_prior = assigned_result['fg_mask_pre_prior'].squeeze(0) pos_inds = fg_mask_pre_prior.nonzero().squeeze(1) targets = bbox_targets[pos_inds] gt_bboxes = gt_bboxes.squeeze(0) matched_gt_inds = torch.tensor( [((t == gt_bboxes).sum(dim=1) == t.shape[0]).nonzero()[0] for t in targets], device=device) level_inds = torch.zeros_like(labels) img_inds = torch.zeros_like(labels) level_nums = [0] + self.num_level_priors for i in range(len(level_nums) - 1): level_nums[i + 1] = level_nums[i] + level_nums[i + 1] level_inds[level_nums[i]:level_nums[i + 1]] = i level_inds_pos = level_inds[pos_inds] img_inds = img_inds[pos_inds] labels = labels[pos_inds] assign_results = [] for i in range(self.num_levels): retained_inds = level_inds_pos == i if not retained_inds.any(): assign_results_prior = { 'stride': self.featmap_strides[i], 'grid_x_inds': torch.zeros([0], dtype=torch.int64).to(device), 'grid_y_inds': torch.zeros([0], dtype=torch.int64).to(device), 'img_inds': torch.zeros([0], dtype=torch.int64).to(device), 'class_inds': torch.zeros([0], dtype=torch.int64).to(device), 'retained_gt_inds': torch.zeros([0], dtype=torch.int64).to(device), 'prior_ind': 0 } else: w = inputs_hw[1] // self.featmap_strides[i] retained_pos_inds = pos_inds[retained_inds] - level_nums[i] grid_y_inds = retained_pos_inds // w grid_x_inds = retained_pos_inds - retained_pos_inds // w * w assign_results_prior = { 'stride': self.featmap_strides[i], 'grid_x_inds': grid_x_inds, 'grid_y_inds': grid_y_inds, 'img_inds': img_inds[retained_inds], 'class_inds': labels[retained_inds], 'retained_gt_inds': matched_gt_inds[retained_inds], 'prior_ind': 0 } assign_results.append([assign_results_prior]) return assign_results def assign(self, batch_data_samples: Union[list, dict], inputs_hw: Union[tuple, torch.Size]) -> dict: """Calculate assigning results. This function is provided to the `assigner_visualization.py` script. Args: batch_data_samples (List[:obj:`DetDataSample`], dict): The Data Samples. It usually includes information such as `gt_instance`, `gt_panoptic_seg` and `gt_sem_seg`. inputs_hw: Height and width of inputs size Returns: dict: A dictionary of assigning components. """ if isinstance(batch_data_samples, list): raise NotImplementedError( 'assigning results_list is not implemented') else: # Fast version cls_scores, bbox_preds = self(batch_data_samples['feats']) assign_inputs = (cls_scores, bbox_preds, batch_data_samples['bboxes_labels'], batch_data_samples['img_metas'], inputs_hw) assign_results = self.assign_by_gt_and_feat(*assign_inputs) return assign_results
7,515
40.524862
79
py
mmyolo
mmyolo-main/projects/assigner_visualization/dense_heads/yolov7_head_assigner.py
# Copyright (c) OpenMMLab. All rights reserved. from typing import List, Union import torch from mmdet.utils import InstanceList from torch import Tensor from mmyolo.models import YOLOv7Head from mmyolo.registry import MODELS @MODELS.register_module() class YOLOv7HeadAssigner(YOLOv7Head): def assign_by_gt_and_feat( self, cls_scores: List[Tensor], bbox_preds: List[Tensor], objectnesses: List[Tensor], batch_gt_instances: InstanceList, batch_img_metas: List[dict], inputs_hw: Union[Tensor, tuple], ) -> dict: """Calculate the assigning results based on the gt and features extracted by the detection head. Args: cls_scores (Sequence[Tensor]): Box scores for each scale level, each is a 4D-tensor, the channel number is num_priors * num_classes. bbox_preds (Sequence[Tensor]): Box energies / deltas for each scale level, each is a 4D-tensor, the channel number is num_priors * 4. objectnesses (Sequence[Tensor]): Score factor for all scale level, each is a 4D-tensor, has shape (batch_size, 1, H, W) batch_gt_instances (list[:obj:`InstanceData`]): Batch of gt_instance. It usually includes ``bboxes`` and ``labels`` attributes. batch_img_metas (list[dict]): Meta information of each image, e.g., image size, scaling factor, etc. inputs_hw (Union[Tensor, tuple]): Height and width of inputs size. Returns: dict[str, Tensor]: A dictionary of assigning results. """ device = cls_scores[0][0].device head_preds = self._merge_predict_results(bbox_preds, objectnesses, cls_scores) batch_targets_normed = self._convert_gt_to_norm_format( batch_gt_instances, batch_img_metas) # yolov5_assign and simota_assign assigner_results = self.assigner( head_preds, batch_targets_normed, batch_img_metas[0]['batch_input_shape'], self.priors_base_sizes, self.grid_offset, near_neighbor_thr=self.near_neighbor_thr) # multi-level positive sample position. mlvl_positive_infos = assigner_results['mlvl_positive_infos'] # assigned results with label and bboxes information. mlvl_targets_normed = assigner_results['mlvl_targets_normed'] assign_results = [] for i in range(self.num_levels): assign_results_feat = [] # no gt bbox matches anchor if mlvl_positive_infos[i].shape[0] == 0: for k in range(self.num_base_priors): assign_results_feat.append({ 'stride': self.featmap_strides[i], 'grid_x_inds': torch.zeros([0], dtype=torch.int64).to(device), 'grid_y_inds': torch.zeros([0], dtype=torch.int64).to(device), 'img_inds': torch.zeros([0], dtype=torch.int64).to(device), 'class_inds': torch.zeros([0], dtype=torch.int64).to(device), 'retained_gt_inds': torch.zeros([0], dtype=torch.int64).to(device), 'prior_ind': k }) assign_results.append(assign_results_feat) continue # (batch_idx, prior_idx, x_scaled, y_scaled) positive_info = mlvl_positive_infos[i] targets_normed = mlvl_targets_normed[i] priors_inds = positive_info[:, 1] grid_x_inds = positive_info[:, 2] grid_y_inds = positive_info[:, 3] img_inds = targets_normed[:, 0] class_inds = targets_normed[:, 1].long() retained_gt_inds = self.get_gt_inds( targets_normed, batch_targets_normed[0]).long() for k in range(self.num_base_priors): retained_inds = priors_inds == k assign_results_prior = { 'stride': self.featmap_strides[i], 'grid_x_inds': grid_x_inds[retained_inds], 'grid_y_inds': grid_y_inds[retained_inds], 'img_inds': img_inds[retained_inds], 'class_inds': class_inds[retained_inds], 'retained_gt_inds': retained_gt_inds[retained_inds], 'prior_ind': k } assign_results_feat.append(assign_results_prior) assign_results.append(assign_results_feat) return assign_results def get_gt_inds(self, assigned_target, gt_instance): """Judging which one gt_ind is assigned by comparing assign_target and origin target. Args: assigned_target (Tensor(assign_nums,7)): YOLOv7 assigning results. gt_instance (Tensor(gt_nums,7)): Normalized gt_instance, It usually includes ``bboxes`` and ``labels`` attributes. Returns: gt_inds (Tensor): the index which one gt is assigned. """ gt_inds = torch.zeros(assigned_target.shape[0]) for i in range(assigned_target.shape[0]): gt_inds[i] = ((assigned_target[i] == gt_instance).sum( dim=1) == 7).nonzero().squeeze() return gt_inds def assign(self, batch_data_samples: Union[list, dict], inputs_hw: Union[tuple, torch.Size]) -> dict: """Calculate assigning results. This function is provided to the `assigner_visualization.py` script. Args: batch_data_samples (List[:obj:`DetDataSample`], dict): The Data Samples. It usually includes information such as `gt_instance`, `gt_panoptic_seg` and `gt_sem_seg`. inputs_hw: Height and width of inputs size Returns: dict: A dictionary of assigning components. """ if isinstance(batch_data_samples, list): raise NotImplementedError( 'assigning results_list is not implemented') else: # Fast version cls_scores, bbox_preds, objectnesses = self( batch_data_samples['feats']) assign_inputs = (cls_scores, bbox_preds, objectnesses, batch_data_samples['bboxes_labels'], batch_data_samples['img_metas'], inputs_hw) assign_results = self.assign_by_gt_and_feat(*assign_inputs) return assign_results
6,836
41.73125
79
py
mmyolo
mmyolo-main/projects/assigner_visualization/dense_heads/rtmdet_head_assigner.py
# Copyright (c) OpenMMLab. All rights reserved. from typing import List, Union import torch from mmdet.structures.bbox import distance2bbox from mmdet.utils import InstanceList from torch import Tensor from mmyolo.models import RTMDetHead from mmyolo.models.utils import gt_instances_preprocess from mmyolo.registry import MODELS @MODELS.register_module() class RTMHeadAssigner(RTMDetHead): def assign_by_gt_and_feat( self, cls_scores: List[Tensor], bbox_preds: List[Tensor], batch_gt_instances: InstanceList, batch_img_metas: List[dict], inputs_hw: Union[Tensor, tuple] = (640, 640) ) -> dict: """Calculate the assigning results based on the gt and features extracted by the detection head. Args: cls_scores (list[Tensor]): Box scores for each scale level Has shape (N, num_anchors * num_classes, H, W) bbox_preds (list[Tensor]): Decoded box for each scale level with shape (N, num_anchors * 4, H, W) in [tl_x, tl_y, br_x, br_y] format. batch_gt_instances (list[:obj:`InstanceData`]): Batch of gt_instance. It usually includes ``bboxes`` and ``labels`` attributes. batch_img_metas (list[dict]): Meta information of each image, e.g., image size, scaling factor, etc. inputs_hw (Union[Tensor, tuple]): Height and width of inputs size. Returns: dict[str, Tensor]: A dictionary of assigning results. """ num_imgs = len(batch_img_metas) featmap_sizes = [featmap.size()[-2:] for featmap in cls_scores] assert len(featmap_sizes) == self.prior_generator.num_levels # rtmdet's prior offset differs from others prior_offset = self.prior_generator.offset gt_info = gt_instances_preprocess(batch_gt_instances, num_imgs) gt_labels = gt_info[:, :, :1] gt_bboxes = gt_info[:, :, 1:] # xyxy pad_bbox_flag = (gt_bboxes.sum(-1, keepdim=True) > 0).float() device = cls_scores[0].device # If the shape does not equal, generate new one if featmap_sizes != self.featmap_sizes_train: self.featmap_sizes_train = featmap_sizes mlvl_priors_with_stride = self.prior_generator.grid_priors( featmap_sizes, device=device, with_stride=True) self.flatten_priors_train = torch.cat( mlvl_priors_with_stride, dim=0) flatten_cls_scores = torch.cat([ cls_score.permute(0, 2, 3, 1).reshape(num_imgs, -1, self.cls_out_channels) for cls_score in cls_scores ], 1).contiguous() flatten_bboxes = torch.cat([ bbox_pred.permute(0, 2, 3, 1).reshape(num_imgs, -1, 4) for bbox_pred in bbox_preds ], 1) flatten_bboxes = flatten_bboxes * self.flatten_priors_train[..., -1, None] flatten_bboxes = distance2bbox(self.flatten_priors_train[..., :2], flatten_bboxes) assigned_result = self.assigner(flatten_bboxes.detach(), flatten_cls_scores.detach(), self.flatten_priors_train, gt_labels, gt_bboxes, pad_bbox_flag) labels = assigned_result['assigned_labels'].reshape(-1) bbox_targets = assigned_result['assigned_bboxes'].reshape(-1, 4) # FG cat_id: [0, num_classes -1], BG cat_id: num_classes bg_class_ind = self.num_classes pos_inds = ((labels >= 0) & (labels < bg_class_ind)).nonzero().squeeze(1) targets = bbox_targets[pos_inds] gt_bboxes = gt_bboxes.squeeze(0) matched_gt_inds = torch.tensor( [((t == gt_bboxes).sum(dim=1) == t.shape[0]).nonzero()[0] for t in targets], device=device) level_inds = torch.zeros_like(labels) img_inds = torch.zeros_like(labels) level_nums = [0] + [f[0] * f[1] for f in featmap_sizes] for i in range(len(level_nums) - 1): level_nums[i + 1] = level_nums[i] + level_nums[i + 1] level_inds[level_nums[i]:level_nums[i + 1]] = i level_inds_pos = level_inds[pos_inds] img_inds = img_inds[pos_inds] labels = labels[pos_inds] inputs_hw = batch_img_metas[0]['batch_input_shape'] assign_results = [] for i in range(self.num_levels): retained_inds = level_inds_pos == i if not retained_inds.any(): assign_results_prior = { 'stride': self.featmap_strides[i], 'grid_x_inds': torch.zeros([0], dtype=torch.int64).to(device), 'grid_y_inds': torch.zeros([0], dtype=torch.int64).to(device), 'img_inds': torch.zeros([0], dtype=torch.int64).to(device), 'class_inds': torch.zeros([0], dtype=torch.int64).to(device), 'retained_gt_inds': torch.zeros([0], dtype=torch.int64).to(device), 'prior_ind': 0, 'offset': prior_offset } else: w = inputs_hw[1] // self.featmap_strides[i] retained_pos_inds = pos_inds[retained_inds] - level_nums[i] grid_y_inds = retained_pos_inds // w grid_x_inds = retained_pos_inds - retained_pos_inds // w * w assign_results_prior = { 'stride': self.featmap_strides[i], 'grid_x_inds': grid_x_inds, 'grid_y_inds': grid_y_inds, 'img_inds': img_inds[retained_inds], 'class_inds': labels[retained_inds], 'retained_gt_inds': matched_gt_inds[retained_inds], 'prior_ind': 0, 'offset': prior_offset } assign_results.append([assign_results_prior]) return assign_results def assign(self, batch_data_samples: Union[list, dict], inputs_hw: Union[tuple, torch.Size]) -> dict: """Calculate assigning results. This function is provided to the `assigner_visualization.py` script. Args: batch_data_samples (List[:obj:`DetDataSample`], dict): The Data Samples. It usually includes information such as `gt_instance`, `gt_panoptic_seg` and `gt_sem_seg`. inputs_hw: Height and width of inputs size Returns: dict: A dictionary of assigning components. """ if isinstance(batch_data_samples, list): raise NotImplementedError( 'assigning results_list is not implemented') else: # Fast version cls_scores, bbox_preds = self(batch_data_samples['feats']) assign_inputs = (cls_scores, bbox_preds, batch_data_samples['bboxes_labels'], batch_data_samples['img_metas'], inputs_hw) assign_results = self.assign_by_gt_and_feat(*assign_inputs) return assign_results
7,517
41.715909
79
py
mmyolo
mmyolo-main/projects/easydeploy/backbone/common.py
import torch import torch.nn as nn from torch import Tensor class DeployC2f(nn.Module): def __init__(self, *args, **kwargs): super().__init__() def forward(self, x: Tensor) -> Tensor: x_main = self.main_conv(x) x_main = [x_main, x_main[:, self.mid_channels:, ...]] x_main.extend(blocks(x_main[-1]) for blocks in self.blocks) x_main.pop(1) return self.final_conv(torch.cat(x_main, 1))
444
25.176471
67
py
mmyolo
mmyolo-main/projects/easydeploy/backbone/focus.py
# Copyright (c) OpenMMLab. All rights reserved. import torch import torch.nn as nn import torch.nn.functional as F from torch import Tensor class DeployFocus(nn.Module): def __init__(self, orin_Focus: nn.Module): super().__init__() self.__dict__.update(orin_Focus.__dict__) def forward(self, x: Tensor) -> Tensor: batch_size, channel, height, width = x.shape x = x.reshape(batch_size, channel, -1, 2, width) x = x.reshape(batch_size, channel, x.shape[2], 2, -1, 2) half_h = x.shape[2] half_w = x.shape[4] x = x.permute(0, 5, 3, 1, 2, 4) x = x.reshape(batch_size, channel * 4, half_h, half_w) return self.conv(x) class NcnnFocus(nn.Module): def __init__(self, orin_Focus: nn.Module): super().__init__() self.__dict__.update(orin_Focus.__dict__) def forward(self, x: Tensor) -> Tensor: batch_size, c, h, w = x.shape assert h % 2 == 0 and w % 2 == 0, f'focus for yolox needs even feature\ height and width, got {(h, w)}.' x = x.reshape(batch_size, c * h, 1, w) _b, _c, _h, _w = x.shape g = _c // 2 # fuse to ncnn's shufflechannel x = x.view(_b, g, 2, _h, _w) x = torch.transpose(x, 1, 2).contiguous() x = x.view(_b, -1, _h, _w) x = x.reshape(_b, c * h * w, 1, 1) _b, _c, _h, _w = x.shape g = _c // 2 # fuse to ncnn's shufflechannel x = x.view(_b, g, 2, _h, _w) x = torch.transpose(x, 1, 2).contiguous() x = x.view(_b, -1, _h, _w) x = x.reshape(_b, c * 4, h // 2, w // 2) return self.conv(x) class GConvFocus(nn.Module): def __init__(self, orin_Focus: nn.Module): super().__init__() device = next(orin_Focus.parameters()).device self.weight1 = torch.tensor([[1., 0], [0, 0]]).expand(3, 1, 2, 2).to(device) self.weight2 = torch.tensor([[0, 0], [1., 0]]).expand(3, 1, 2, 2).to(device) self.weight3 = torch.tensor([[0, 1.], [0, 0]]).expand(3, 1, 2, 2).to(device) self.weight4 = torch.tensor([[0, 0], [0, 1.]]).expand(3, 1, 2, 2).to(device) self.__dict__.update(orin_Focus.__dict__) def forward(self, x: Tensor) -> Tensor: conv1 = F.conv2d(x, self.weight1, stride=2, groups=3) conv2 = F.conv2d(x, self.weight2, stride=2, groups=3) conv3 = F.conv2d(x, self.weight3, stride=2, groups=3) conv4 = F.conv2d(x, self.weight4, stride=2, groups=3) return self.conv(torch.cat([conv1, conv2, conv3, conv4], dim=1))
2,834
34.4375
79
py
mmyolo
mmyolo-main/projects/easydeploy/tools/build_engine.py
import argparse from pathlib import Path from typing import List, Optional, Tuple, Union try: import tensorrt as trt except Exception: trt = None import warnings import numpy as np import torch warnings.filterwarnings(action='ignore', category=DeprecationWarning) class EngineBuilder: def __init__( self, checkpoint: Union[str, Path], opt_shape: Union[Tuple, List] = (1, 3, 640, 640), device: Optional[Union[str, int, torch.device]] = None) -> None: checkpoint = Path(checkpoint) if isinstance(checkpoint, str) else checkpoint assert checkpoint.exists() and checkpoint.suffix == '.onnx' if isinstance(device, str): device = torch.device(device) elif isinstance(device, int): device = torch.device(f'cuda:{device}') self.checkpoint = checkpoint self.opt_shape = np.array(opt_shape, dtype=np.float32) self.device = device def __build_engine(self, scale: Optional[List[List]] = None, fp16: bool = True, with_profiling: bool = True) -> None: logger = trt.Logger(trt.Logger.WARNING) trt.init_libnvinfer_plugins(logger, namespace='') builder = trt.Builder(logger) config = builder.create_builder_config() config.max_workspace_size = torch.cuda.get_device_properties( self.device).total_memory flag = (1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)) network = builder.create_network(flag) parser = trt.OnnxParser(network, logger) if not parser.parse_from_file(str(self.checkpoint)): raise RuntimeError( f'failed to load ONNX file: {str(self.checkpoint)}') inputs = [network.get_input(i) for i in range(network.num_inputs)] outputs = [network.get_output(i) for i in range(network.num_outputs)] profile = None dshape = -1 in network.get_input(0).shape if dshape: profile = builder.create_optimization_profile() if scale is None: scale = np.array( [[1, 1, 0.5, 0.5], [1, 1, 1, 1], [4, 1, 1.5, 1.5]], dtype=np.float32) scale = (self.opt_shape * scale).astype(np.int32) elif isinstance(scale, List): scale = np.array(scale, dtype=np.int32) assert scale.shape[0] == 3, 'Input a wrong scale list' else: raise NotImplementedError for inp in inputs: logger.log( trt.Logger.WARNING, f'input "{inp.name}" with shape{inp.shape} {inp.dtype}') if dshape: profile.set_shape(inp.name, *scale) for out in outputs: logger.log( trt.Logger.WARNING, f'output "{out.name}" with shape{out.shape} {out.dtype}') if fp16 and builder.platform_has_fast_fp16: config.set_flag(trt.BuilderFlag.FP16) self.weight = self.checkpoint.with_suffix('.engine') if dshape: config.add_optimization_profile(profile) if with_profiling: config.profiling_verbosity = trt.ProfilingVerbosity.DETAILED with builder.build_engine(network, config) as engine: self.weight.write_bytes(engine.serialize()) logger.log( trt.Logger.WARNING, f'Build tensorrt engine finish.\n' f'Save in {str(self.weight.absolute())}') def build(self, scale: Optional[List[List]] = None, fp16: bool = True, with_profiling=True): self.__build_engine(scale, fp16, with_profiling) def parse_args(): parser = argparse.ArgumentParser() parser.add_argument('checkpoint', help='Checkpoint file') parser.add_argument( '--img-size', nargs='+', type=int, default=[640, 640], help='Image size of height and width') parser.add_argument( '--device', type=str, default='cuda:0', help='TensorRT builder device') parser.add_argument( '--scales', type=str, default='[[1,3,640,640],[1,3,640,640],[1,3,640,640]]', help='Input scales for build dynamic input shape engine') parser.add_argument( '--fp16', action='store_true', help='Build model with fp16 mode') args = parser.parse_args() args.img_size *= 2 if len(args.img_size) == 1 else 1 return args def main(args): img_size = (1, 3, *args.img_size) try: scales = eval(args.scales) except Exception: print('Input scales is not a python variable') print('Set scales default None') scales = None builder = EngineBuilder(args.checkpoint, img_size, args.device) builder.build(scales, fp16=args.fp16) if __name__ == '__main__': args = parse_args() main(args)
5,007
35.554745
79
py
mmyolo
mmyolo-main/projects/easydeploy/tools/export.py
import argparse import os import warnings from io import BytesIO import onnx import torch from mmdet.apis import init_detector from mmengine.config import ConfigDict from mmengine.utils.path import mkdir_or_exist from mmyolo.utils import register_all_modules from projects.easydeploy.model import DeployModel warnings.filterwarnings(action='ignore', category=torch.jit.TracerWarning) warnings.filterwarnings(action='ignore', category=torch.jit.ScriptWarning) warnings.filterwarnings(action='ignore', category=UserWarning) warnings.filterwarnings(action='ignore', category=FutureWarning) warnings.filterwarnings(action='ignore', category=ResourceWarning) def parse_args(): parser = argparse.ArgumentParser() parser.add_argument('config', help='Config file') parser.add_argument('checkpoint', help='Checkpoint file') parser.add_argument( '--model-only', action='store_true', help='Export model only') parser.add_argument( '--work-dir', default='./work_dir', help='Path to save export model') parser.add_argument( '--img-size', nargs='+', type=int, default=[640, 640], help='Image size of height and width') parser.add_argument('--batch-size', type=int, default=1, help='Batch size') parser.add_argument( '--device', default='cuda:0', help='Device used for inference') parser.add_argument( '--simplify', action='store_true', help='Simplify onnx model by onnx-sim') parser.add_argument( '--opset', type=int, default=11, help='ONNX opset version') parser.add_argument( '--backend', type=int, default=1, help='Backend for export onnx') parser.add_argument( '--pre-topk', type=int, default=1000, help='Postprocess pre topk bboxes feed into NMS') parser.add_argument( '--keep-topk', type=int, default=100, help='Postprocess keep topk bboxes out of NMS') parser.add_argument( '--iou-threshold', type=float, default=0.65, help='IoU threshold for NMS') parser.add_argument( '--score-threshold', type=float, default=0.25, help='Score threshold for NMS') args = parser.parse_args() args.img_size *= 2 if len(args.img_size) == 1 else 1 return args def build_model_from_cfg(config_path, checkpoint_path, device): model = init_detector(config_path, checkpoint_path, device=device) model.eval() return model def main(): args = parse_args() register_all_modules() mkdir_or_exist(args.work_dir) if args.model_only: postprocess_cfg = None output_names = None else: postprocess_cfg = ConfigDict( pre_top_k=args.pre_topk, keep_top_k=args.keep_topk, iou_threshold=args.iou_threshold, score_threshold=args.score_threshold, backend=args.backend) output_names = ['num_dets', 'boxes', 'scores', 'labels'] baseModel = build_model_from_cfg(args.config, args.checkpoint, args.device) deploy_model = DeployModel( baseModel=baseModel, postprocess_cfg=postprocess_cfg) deploy_model.eval() fake_input = torch.randn(args.batch_size, 3, *args.img_size).to(args.device) # dry run deploy_model(fake_input) save_onnx_path = os.path.join(args.work_dir, 'end2end.onnx') # export onnx with BytesIO() as f: torch.onnx.export( deploy_model, fake_input, f, input_names=['images'], output_names=output_names, opset_version=args.opset) f.seek(0) onnx_model = onnx.load(f) onnx.checker.check_model(onnx_model) # Fix tensorrt onnx output shape, just for view if args.backend in (2, 3): shapes = [ args.batch_size, 1, args.batch_size, args.keep_topk, 4, args.batch_size, args.keep_topk, args.batch_size, args.keep_topk ] for i in onnx_model.graph.output: for j in i.type.tensor_type.shape.dim: j.dim_param = str(shapes.pop(0)) if args.simplify: try: import onnxsim onnx_model, check = onnxsim.simplify(onnx_model) assert check, 'assert check failed' except Exception as e: print(f'Simplify failure: {e}') onnx.save(onnx_model, save_onnx_path) print(f'ONNX export success, save into {save_onnx_path}') if __name__ == '__main__': main()
4,623
31.335664
79
py
mmyolo
mmyolo-main/projects/easydeploy/tools/image-demo.py
# Copyright (c) OpenMMLab. All rights reserved. from projects.easydeploy.model import ORTWrapper, TRTWrapper # isort:skip import os import random from argparse import ArgumentParser import cv2 import mmcv import numpy as np import torch from mmcv.transforms import Compose from mmdet.utils import get_test_pipeline_cfg from mmengine.config import Config, ConfigDict from mmengine.utils import ProgressBar, path from mmyolo.utils import register_all_modules from mmyolo.utils.misc import get_file_list def parse_args(): parser = ArgumentParser() parser.add_argument( 'img', help='Image path, include image file, dir and URL.') parser.add_argument('config', help='Config file') parser.add_argument('checkpoint', help='Checkpoint file') parser.add_argument( '--out-dir', default='./output', help='Path to output file') parser.add_argument( '--device', default='cuda:0', help='Device used for inference') parser.add_argument( '--show', action='store_true', help='Show the detection results') args = parser.parse_args() return args def preprocess(config): data_preprocess = config.get('model', {}).get('data_preprocessor', {}) mean = data_preprocess.get('mean', [0., 0., 0.]) std = data_preprocess.get('std', [1., 1., 1.]) mean = torch.tensor(mean, dtype=torch.float32).reshape(1, 3, 1, 1) std = torch.tensor(std, dtype=torch.float32).reshape(1, 3, 1, 1) class PreProcess(torch.nn.Module): def __init__(self): super().__init__() def forward(self, x): x = x[None].float() x -= mean.to(x.device) x /= std.to(x.device) return x return PreProcess().eval() def main(): args = parse_args() # register all modules in mmdet into the registries register_all_modules() colors = [[random.randint(0, 255) for _ in range(3)] for _ in range(1000)] # build the model from a config file and a checkpoint file if args.checkpoint.endswith('.onnx'): model = ORTWrapper(args.checkpoint, args.device) elif args.checkpoint.endswith('.engine') or args.checkpoint.endswith( '.plan'): model = TRTWrapper(args.checkpoint, args.device) else: raise NotImplementedError model.to(args.device) cfg = Config.fromfile(args.config) class_names = cfg.get('class_name') test_pipeline = get_test_pipeline_cfg(cfg) test_pipeline[0] = ConfigDict({'type': 'mmdet.LoadImageFromNDArray'}) test_pipeline = Compose(test_pipeline) pre_pipeline = preprocess(cfg) if not args.show: path.mkdir_or_exist(args.out_dir) # get file list files, source_type = get_file_list(args.img) # start detector inference progress_bar = ProgressBar(len(files)) for i, file in enumerate(files): bgr = mmcv.imread(file) rgb = mmcv.imconvert(bgr, 'bgr', 'rgb') data, samples = test_pipeline(dict(img=rgb, img_id=i)).values() pad_param = samples.get('pad_param', np.array([0, 0, 0, 0], dtype=np.float32)) h, w = samples.get('ori_shape', rgb.shape[:2]) pad_param = torch.asarray( [pad_param[2], pad_param[0], pad_param[2], pad_param[0]], device=args.device) scale_factor = samples.get('scale_factor', [1., 1]) scale_factor = torch.asarray(scale_factor * 2, device=args.device) data = pre_pipeline(data).to(args.device) result = model(data) if source_type['is_dir']: filename = os.path.relpath(file, args.img).replace('/', '_') else: filename = os.path.basename(file) out_file = None if args.show else os.path.join(args.out_dir, filename) # Get candidate predict info by num_dets num_dets, bboxes, scores, labels = result scores = scores[0, :num_dets] bboxes = bboxes[0, :num_dets] labels = labels[0, :num_dets] bboxes -= pad_param bboxes /= scale_factor bboxes[:, 0::2].clamp_(0, w) bboxes[:, 1::2].clamp_(0, h) bboxes = bboxes.round().int() for (bbox, score, label) in zip(bboxes, scores, labels): bbox = bbox.tolist() color = colors[label] if class_names is not None: label_name = class_names[label] name = f'cls:{label_name}_score:{score:0.4f}' else: name = f'cls:{label}_score:{score:0.4f}' cv2.rectangle(bgr, bbox[:2], bbox[2:], color, 2) cv2.putText( bgr, name, (bbox[0], bbox[1] - 2), cv2.FONT_HERSHEY_SIMPLEX, 2.0, [225, 255, 255], thickness=3) if args.show: mmcv.imshow(bgr, 'result', 0) else: mmcv.imwrite(bgr, out_file) progress_bar.update() if __name__ == '__main__': main()
4,968
31.477124
78
py
mmyolo
mmyolo-main/projects/easydeploy/bbox_code/bbox_coder.py
# Copyright (c) OpenMMLab. All rights reserved. from typing import Optional import torch from torch import Tensor def yolov5_bbox_decoder(priors: Tensor, bbox_preds: Tensor, stride: Tensor) -> Tensor: bbox_preds = bbox_preds.sigmoid() x_center = (priors[..., 0] + priors[..., 2]) * 0.5 y_center = (priors[..., 1] + priors[..., 3]) * 0.5 w = priors[..., 2] - priors[..., 0] h = priors[..., 3] - priors[..., 1] x_center_pred = (bbox_preds[..., 0] - 0.5) * 2 * stride + x_center y_center_pred = (bbox_preds[..., 1] - 0.5) * 2 * stride + y_center w_pred = (bbox_preds[..., 2] * 2)**2 * w h_pred = (bbox_preds[..., 3] * 2)**2 * h decoded_bboxes = torch.stack( [x_center_pred, y_center_pred, w_pred, h_pred], dim=-1) return decoded_bboxes def rtmdet_bbox_decoder(priors: Tensor, bbox_preds: Tensor, stride: Optional[Tensor]) -> Tensor: stride = stride[None, :, None] bbox_preds *= stride tl_x = (priors[..., 0] - bbox_preds[..., 0]) tl_y = (priors[..., 1] - bbox_preds[..., 1]) br_x = (priors[..., 0] + bbox_preds[..., 2]) br_y = (priors[..., 1] + bbox_preds[..., 3]) decoded_bboxes = torch.stack([tl_x, tl_y, br_x, br_y], -1) return decoded_bboxes def yolox_bbox_decoder(priors: Tensor, bbox_preds: Tensor, stride: Optional[Tensor]) -> Tensor: stride = stride[None, :, None] xys = (bbox_preds[..., :2] * stride) + priors whs = bbox_preds[..., 2:].exp() * stride decoded_bboxes = torch.cat([xys, whs], -1) return decoded_bboxes
1,608
33.234043
70
py
mmyolo
mmyolo-main/projects/easydeploy/model/backendwrapper.py
import warnings from collections import namedtuple from functools import partial from pathlib import Path from typing import List, Optional, Union import numpy as np import onnxruntime try: import tensorrt as trt except Exception: trt = None import torch warnings.filterwarnings(action='ignore', category=DeprecationWarning) class TRTWrapper(torch.nn.Module): dtype_mapping = {} def __init__(self, weight: Union[str, Path], device: Optional[torch.device]): super().__init__() weight = Path(weight) if isinstance(weight, str) else weight assert weight.exists() and weight.suffix in ('.engine', '.plan') if isinstance(device, str): device = torch.device(device) elif isinstance(device, int): device = torch.device(f'cuda:{device}') self.weight = weight self.device = device self.stream = torch.cuda.Stream(device=device) self.__update_mapping() self.__init_engine() self.__init_bindings() def __update_mapping(self): self.dtype_mapping.update({ trt.bool: torch.bool, trt.int8: torch.int8, trt.int32: torch.int32, trt.float16: torch.float16, trt.float32: torch.float32 }) def __init_engine(self): logger = trt.Logger(trt.Logger.ERROR) self.log = partial(logger.log, trt.Logger.ERROR) trt.init_libnvinfer_plugins(logger, namespace='') self.logger = logger with trt.Runtime(logger) as runtime: model = runtime.deserialize_cuda_engine(self.weight.read_bytes()) context = model.create_execution_context() names = [model.get_binding_name(i) for i in range(model.num_bindings)] num_inputs, num_outputs = 0, 0 for i in range(model.num_bindings): if model.binding_is_input(i): num_inputs += 1 else: num_outputs += 1 self.is_dynamic = -1 in model.get_binding_shape(0) self.model = model self.context = context self.input_names = names[:num_inputs] self.output_names = names[num_inputs:] self.num_inputs = num_inputs self.num_outputs = num_outputs self.num_bindings = num_inputs + num_outputs self.bindings: List[int] = [0] * self.num_bindings def __init_bindings(self): Binding = namedtuple('Binding', ('name', 'dtype', 'shape')) inputs_info = [] outputs_info = [] for i, name in enumerate(self.input_names): assert self.model.get_binding_name(i) == name dtype = self.dtype_mapping[self.model.get_binding_dtype(i)] shape = tuple(self.model.get_binding_shape(i)) inputs_info.append(Binding(name, dtype, shape)) for i, name in enumerate(self.output_names): i += self.num_inputs assert self.model.get_binding_name(i) == name dtype = self.dtype_mapping[self.model.get_binding_dtype(i)] shape = tuple(self.model.get_binding_shape(i)) outputs_info.append(Binding(name, dtype, shape)) self.inputs_info = inputs_info self.outputs_info = outputs_info if not self.is_dynamic: self.output_tensor = [ torch.empty(o.shape, dtype=o.dtype, device=self.device) for o in outputs_info ] def forward(self, *inputs): assert len(inputs) == self.num_inputs contiguous_inputs: List[torch.Tensor] = [ i.contiguous() for i in inputs ] for i in range(self.num_inputs): self.bindings[i] = contiguous_inputs[i].data_ptr() if self.is_dynamic: self.context.set_binding_shape( i, tuple(contiguous_inputs[i].shape)) # create output tensors outputs: List[torch.Tensor] = [] for i in range(self.num_outputs): j = i + self.num_inputs if self.is_dynamic: shape = tuple(self.context.get_binding_shape(j)) output = torch.empty( size=shape, dtype=self.output_dtypes[i], device=self.device) else: output = self.output_tensor[i] outputs.append(output) self.bindings[j] = output.data_ptr() self.context.execute_async_v2(self.bindings, self.stream.cuda_stream) self.stream.synchronize() return tuple(outputs) class ORTWrapper(torch.nn.Module): def __init__(self, weight: Union[str, Path], device: Optional[torch.device]): super().__init__() weight = Path(weight) if isinstance(weight, str) else weight assert weight.exists() and weight.suffix == '.onnx' if isinstance(device, str): device = torch.device(device) elif isinstance(device, int): device = torch.device(f'cuda:{device}') self.weight = weight self.device = device self.__init_session() self.__init_bindings() def __init_session(self): providers = ['CPUExecutionProvider'] if 'cuda' in self.device.type: providers.insert(0, 'CUDAExecutionProvider') session = onnxruntime.InferenceSession( str(self.weight), providers=providers) self.session = session def __init_bindings(self): Binding = namedtuple('Binding', ('name', 'dtype', 'shape')) inputs_info = [] outputs_info = [] self.is_dynamic = False for i, tensor in enumerate(self.session.get_inputs()): if any(not isinstance(i, int) for i in tensor.shape): self.is_dynamic = True inputs_info.append( Binding(tensor.name, tensor.type, tuple(tensor.shape))) for i, tensor in enumerate(self.session.get_outputs()): outputs_info.append( Binding(tensor.name, tensor.type, tuple(tensor.shape))) self.inputs_info = inputs_info self.outputs_info = outputs_info self.num_inputs = len(inputs_info) def forward(self, *inputs): assert len(inputs) == self.num_inputs contiguous_inputs: List[np.ndarray] = [ i.contiguous().cpu().numpy() for i in inputs ] if not self.is_dynamic: # make sure input shape is right for static input shape for i in range(self.num_inputs): assert contiguous_inputs[i].shape == self.inputs_info[i].shape outputs = self.session.run([o.name for o in self.outputs_info], { j.name: contiguous_inputs[i] for i, j in enumerate(self.inputs_info) }) return tuple(torch.from_numpy(o).to(self.device) for o in outputs)
6,885
32.921182
78
py
mmyolo
mmyolo-main/projects/easydeploy/model/model.py
# Copyright (c) OpenMMLab. All rights reserved. from functools import partial from typing import List, Optional import torch import torch.nn as nn from mmdet.models.backbones.csp_darknet import Focus from mmengine.config import ConfigDict from torch import Tensor from mmyolo.models import RepVGGBlock from mmyolo.models.dense_heads import (RTMDetHead, YOLOv5Head, YOLOv7Head, YOLOXHead) from mmyolo.models.layers import CSPLayerWithTwoConv from ..backbone import DeployC2f, DeployFocus, GConvFocus, NcnnFocus from ..bbox_code import (rtmdet_bbox_decoder, yolov5_bbox_decoder, yolox_bbox_decoder) from ..nms import batched_nms, efficient_nms, onnx_nms class DeployModel(nn.Module): def __init__(self, baseModel: nn.Module, postprocess_cfg: Optional[ConfigDict] = None): super().__init__() self.baseModel = baseModel if postprocess_cfg is None: self.with_postprocess = False else: self.with_postprocess = True self.baseHead = baseModel.bbox_head self.__init_sub_attributes() self.detector_type = type(self.baseHead) self.pre_top_k = postprocess_cfg.get('pre_top_k', 1000) self.keep_top_k = postprocess_cfg.get('keep_top_k', 100) self.iou_threshold = postprocess_cfg.get('iou_threshold', 0.65) self.score_threshold = postprocess_cfg.get('score_threshold', 0.25) self.backend = postprocess_cfg.get('backend', 1) self.__switch_deploy() def __init_sub_attributes(self): self.bbox_decoder = self.baseHead.bbox_coder.decode self.prior_generate = self.baseHead.prior_generator.grid_priors self.num_base_priors = self.baseHead.num_base_priors self.featmap_strides = self.baseHead.featmap_strides self.num_classes = self.baseHead.num_classes def __switch_deploy(self): for layer in self.baseModel.modules(): if isinstance(layer, RepVGGBlock): layer.switch_to_deploy() elif isinstance(layer, Focus): # onnxruntime tensorrt8 tensorrt7 if self.backend in (1, 2, 3): self.baseModel.backbone.stem = DeployFocus(layer) # ncnn elif self.backend == 4: self.baseModel.backbone.stem = NcnnFocus(layer) # switch focus to group conv else: self.baseModel.backbone.stem = GConvFocus(layer) elif isinstance(layer, CSPLayerWithTwoConv): setattr(layer, '__class__', DeployC2f) def pred_by_feat(self, cls_scores: List[Tensor], bbox_preds: List[Tensor], objectnesses: Optional[List[Tensor]] = None, **kwargs): assert len(cls_scores) == len(bbox_preds) dtype = cls_scores[0].dtype device = cls_scores[0].device nms_func = self.select_nms() if self.detector_type in (YOLOv5Head, YOLOv7Head): bbox_decoder = yolov5_bbox_decoder elif self.detector_type is RTMDetHead: bbox_decoder = rtmdet_bbox_decoder elif self.detector_type is YOLOXHead: bbox_decoder = yolox_bbox_decoder else: bbox_decoder = self.bbox_decoder num_imgs = cls_scores[0].shape[0] featmap_sizes = [cls_score.shape[2:] for cls_score in cls_scores] mlvl_priors = self.prior_generate( featmap_sizes, dtype=dtype, device=device) flatten_priors = torch.cat(mlvl_priors) mlvl_strides = [ flatten_priors.new_full( (featmap_size[0] * featmap_size[1] * self.num_base_priors, ), stride) for featmap_size, stride in zip( featmap_sizes, self.featmap_strides) ] flatten_stride = torch.cat(mlvl_strides) # flatten cls_scores, bbox_preds and objectness flatten_cls_scores = [ cls_score.permute(0, 2, 3, 1).reshape(num_imgs, -1, self.num_classes) for cls_score in cls_scores ] cls_scores = torch.cat(flatten_cls_scores, dim=1).sigmoid() flatten_bbox_preds = [ bbox_pred.permute(0, 2, 3, 1).reshape(num_imgs, -1, 4) for bbox_pred in bbox_preds ] flatten_bbox_preds = torch.cat(flatten_bbox_preds, dim=1) if objectnesses is not None: flatten_objectness = [ objectness.permute(0, 2, 3, 1).reshape(num_imgs, -1) for objectness in objectnesses ] flatten_objectness = torch.cat(flatten_objectness, dim=1).sigmoid() cls_scores = cls_scores * (flatten_objectness.unsqueeze(-1)) scores = cls_scores bboxes = bbox_decoder(flatten_priors[None], flatten_bbox_preds, flatten_stride) return nms_func(bboxes, scores, self.keep_top_k, self.iou_threshold, self.score_threshold, self.pre_top_k, self.keep_top_k) def select_nms(self): if self.backend == 1: nms_func = onnx_nms elif self.backend == 2: nms_func = efficient_nms elif self.backend == 3: nms_func = batched_nms else: raise NotImplementedError if type(self.baseHead) in (YOLOv5Head, YOLOv7Head, YOLOXHead): nms_func = partial(nms_func, box_coding=1) return nms_func def forward(self, inputs: Tensor): neck_outputs = self.baseModel(inputs) if self.with_postprocess: return self.pred_by_feat(*neck_outputs) else: return neck_outputs
5,871
37.887417
79
py
mmyolo
mmyolo-main/projects/easydeploy/nms/ort_nms.py
# Copyright (c) OpenMMLab. All rights reserved. import torch from torch import Tensor _XYWH2XYXY = torch.tensor([[1.0, 0.0, 1.0, 0.0], [0.0, 1.0, 0.0, 1.0], [-0.5, 0.0, 0.5, 0.0], [0.0, -0.5, 0.0, 0.5]], dtype=torch.float32) def select_nms_index(scores: Tensor, boxes: Tensor, nms_index: Tensor, batch_size: int, keep_top_k: int = -1): batch_inds, cls_inds = nms_index[:, 0], nms_index[:, 1] box_inds = nms_index[:, 2] scores = scores[batch_inds, cls_inds, box_inds].unsqueeze(1) boxes = boxes[batch_inds, box_inds, ...] dets = torch.cat([boxes, scores], dim=1) batched_dets = dets.unsqueeze(0).repeat(batch_size, 1, 1) batch_template = torch.arange( 0, batch_size, dtype=batch_inds.dtype, device=batch_inds.device) batched_dets = batched_dets.where( (batch_inds == batch_template.unsqueeze(1)).unsqueeze(-1), batched_dets.new_zeros(1)) batched_labels = cls_inds.unsqueeze(0).repeat(batch_size, 1) batched_labels = batched_labels.where( (batch_inds == batch_template.unsqueeze(1)), batched_labels.new_ones(1) * -1) N = batched_dets.shape[0] batched_dets = torch.cat((batched_dets, batched_dets.new_zeros((N, 1, 5))), 1) batched_labels = torch.cat((batched_labels, -batched_labels.new_ones( (N, 1))), 1) _, topk_inds = batched_dets[:, :, -1].sort(dim=1, descending=True) topk_batch_inds = torch.arange( batch_size, dtype=topk_inds.dtype, device=topk_inds.device).view(-1, 1) batched_dets = batched_dets[topk_batch_inds, topk_inds, ...] batched_labels = batched_labels[topk_batch_inds, topk_inds, ...] batched_dets, batched_scores = batched_dets.split([4, 1], 2) batched_scores = batched_scores.squeeze(-1) num_dets = (batched_scores > 0).sum(1, keepdim=True) return num_dets, batched_dets, batched_scores, batched_labels class ONNXNMSop(torch.autograd.Function): @staticmethod def forward( ctx, boxes: Tensor, scores: Tensor, max_output_boxes_per_class: Tensor = torch.tensor([100]), iou_threshold: Tensor = torch.tensor([0.5]), score_threshold: Tensor = torch.tensor([0.05]) ) -> Tensor: device = boxes.device batch = scores.shape[0] num_det = 20 batches = torch.randint(0, batch, (num_det, )).sort()[0].to(device) idxs = torch.arange(100, 100 + num_det).to(device) zeros = torch.zeros((num_det, ), dtype=torch.int64).to(device) selected_indices = torch.cat([batches[None], zeros[None], idxs[None]], 0).T.contiguous() selected_indices = selected_indices.to(torch.int64) return selected_indices @staticmethod def symbolic( g, boxes: Tensor, scores: Tensor, max_output_boxes_per_class: Tensor = torch.tensor([100]), iou_threshold: Tensor = torch.tensor([0.5]), score_threshold: Tensor = torch.tensor([0.05]), ): return g.op( 'NonMaxSuppression', boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold, outputs=1) def onnx_nms( boxes: torch.Tensor, scores: torch.Tensor, max_output_boxes_per_class: int = 100, iou_threshold: float = 0.5, score_threshold: float = 0.05, pre_top_k: int = -1, keep_top_k: int = 100, box_coding: int = 0, ): max_output_boxes_per_class = torch.tensor([max_output_boxes_per_class]) iou_threshold = torch.tensor([iou_threshold]) score_threshold = torch.tensor([score_threshold]) batch_size, _, _ = scores.shape if box_coding == 1: boxes = boxes @ (_XYWH2XYXY.to(boxes.device)) scores = scores.transpose(1, 2).contiguous() selected_indices = ONNXNMSop.apply(boxes, scores, max_output_boxes_per_class, iou_threshold, score_threshold) num_dets, batched_dets, batched_scores, batched_labels = select_nms_index( scores, boxes, selected_indices, batch_size, keep_top_k=keep_top_k) return num_dets, batched_dets, batched_scores, batched_labels.to( torch.int32)
4,445
35.146341
79
py
mmyolo
mmyolo-main/projects/easydeploy/nms/trt_nms.py
# Copyright (c) OpenMMLab. All rights reserved. import torch from torch import Tensor _XYWH2XYXY = torch.tensor([[1.0, 0.0, 1.0, 0.0], [0.0, 1.0, 0.0, 1.0], [-0.5, 0.0, 0.5, 0.0], [0.0, -0.5, 0.0, 0.5]], dtype=torch.float32) class TRTEfficientNMSop(torch.autograd.Function): @staticmethod def forward( ctx, boxes: Tensor, scores: Tensor, background_class: int = -1, box_coding: int = 0, iou_threshold: float = 0.45, max_output_boxes: int = 100, plugin_version: str = '1', score_activation: int = 0, score_threshold: float = 0.25, ): batch_size, _, num_classes = scores.shape num_det = torch.randint( 0, max_output_boxes, (batch_size, 1), dtype=torch.int32) det_boxes = torch.randn(batch_size, max_output_boxes, 4) det_scores = torch.randn(batch_size, max_output_boxes) det_classes = torch.randint( 0, num_classes, (batch_size, max_output_boxes), dtype=torch.int32) return num_det, det_boxes, det_scores, det_classes @staticmethod def symbolic(g, boxes: Tensor, scores: Tensor, background_class: int = -1, box_coding: int = 0, iou_threshold: float = 0.45, max_output_boxes: int = 100, plugin_version: str = '1', score_activation: int = 0, score_threshold: float = 0.25): out = g.op( 'TRT::EfficientNMS_TRT', boxes, scores, background_class_i=background_class, box_coding_i=box_coding, iou_threshold_f=iou_threshold, max_output_boxes_i=max_output_boxes, plugin_version_s=plugin_version, score_activation_i=score_activation, score_threshold_f=score_threshold, outputs=4) num_det, det_boxes, det_scores, det_classes = out return num_det, det_boxes, det_scores, det_classes class TRTbatchedNMSop(torch.autograd.Function): """TensorRT NMS operation.""" @staticmethod def forward( ctx, boxes: Tensor, scores: Tensor, plugin_version: str = '1', shareLocation: int = 1, backgroundLabelId: int = -1, numClasses: int = 80, topK: int = 1000, keepTopK: int = 100, scoreThreshold: float = 0.25, iouThreshold: float = 0.45, isNormalized: int = 0, clipBoxes: int = 0, scoreBits: int = 16, caffeSemantics: int = 1, ): batch_size, _, numClasses = scores.shape num_det = torch.randint( 0, keepTopK, (batch_size, 1), dtype=torch.int32) det_boxes = torch.randn(batch_size, keepTopK, 4) det_scores = torch.randn(batch_size, keepTopK) det_classes = torch.randint(0, numClasses, (batch_size, keepTopK)).float() return num_det, det_boxes, det_scores, det_classes @staticmethod def symbolic( g, boxes: Tensor, scores: Tensor, plugin_version: str = '1', shareLocation: int = 1, backgroundLabelId: int = -1, numClasses: int = 80, topK: int = 1000, keepTopK: int = 100, scoreThreshold: float = 0.25, iouThreshold: float = 0.45, isNormalized: int = 0, clipBoxes: int = 0, scoreBits: int = 16, caffeSemantics: int = 1, ): out = g.op( 'TRT::BatchedNMSDynamic_TRT', boxes, scores, shareLocation_i=shareLocation, plugin_version_s=plugin_version, backgroundLabelId_i=backgroundLabelId, numClasses_i=numClasses, topK_i=topK, keepTopK_i=keepTopK, scoreThreshold_f=scoreThreshold, iouThreshold_f=iouThreshold, isNormalized_i=isNormalized, clipBoxes_i=clipBoxes, scoreBits_i=scoreBits, caffeSemantics_i=caffeSemantics, outputs=4) num_det, det_boxes, det_scores, det_classes = out return num_det, det_boxes, det_scores, det_classes def _efficient_nms( boxes: Tensor, scores: Tensor, max_output_boxes_per_class: int = 1000, iou_threshold: float = 0.5, score_threshold: float = 0.05, pre_top_k: int = -1, keep_top_k: int = 100, box_coding: int = 0, ): """Wrapper for `efficient_nms` with TensorRT. Args: boxes (Tensor): The bounding boxes of shape [N, num_boxes, 4]. scores (Tensor): The detection scores of shape [N, num_boxes, num_classes]. max_output_boxes_per_class (int): Maximum number of output boxes per class of nms. Defaults to 1000. iou_threshold (float): IOU threshold of nms. Defaults to 0.5. score_threshold (float): score threshold of nms. Defaults to 0.05. pre_top_k (int): Number of top K boxes to keep before nms. Defaults to -1. keep_top_k (int): Number of top K boxes to keep after nms. Defaults to -1. box_coding (int): Bounding boxes format for nms. Defaults to 0 means [x1, y1 ,x2, y2]. Set to 1 means [x, y, w, h]. Returns: tuple[Tensor, Tensor, Tensor, Tensor]: (num_det, det_boxes, det_scores, det_classes), `num_det` of shape [N, 1] `det_boxes` of shape [N, num_det, 4] `det_scores` of shape [N, num_det] `det_classes` of shape [N, num_det] """ num_det, det_boxes, det_scores, det_classes = TRTEfficientNMSop.apply( boxes, scores, -1, box_coding, iou_threshold, keep_top_k, '1', 0, score_threshold) return num_det, det_boxes, det_scores, det_classes def _batched_nms( boxes: Tensor, scores: Tensor, max_output_boxes_per_class: int = 1000, iou_threshold: float = 0.5, score_threshold: float = 0.05, pre_top_k: int = -1, keep_top_k: int = 100, box_coding: int = 0, ): """Wrapper for `efficient_nms` with TensorRT. Args: boxes (Tensor): The bounding boxes of shape [N, num_boxes, 4]. scores (Tensor): The detection scores of shape [N, num_boxes, num_classes]. max_output_boxes_per_class (int): Maximum number of output boxes per class of nms. Defaults to 1000. iou_threshold (float): IOU threshold of nms. Defaults to 0.5. score_threshold (float): score threshold of nms. Defaults to 0.05. pre_top_k (int): Number of top K boxes to keep before nms. Defaults to -1. keep_top_k (int): Number of top K boxes to keep after nms. Defaults to -1. box_coding (int): Bounding boxes format for nms. Defaults to 0 means [x1, y1 ,x2, y2]. Set to 1 means [x, y, w, h]. Returns: tuple[Tensor, Tensor, Tensor, Tensor]: (num_det, det_boxes, det_scores, det_classes), `num_det` of shape [N, 1] `det_boxes` of shape [N, num_det, 4] `det_scores` of shape [N, num_det] `det_classes` of shape [N, num_det] """ if box_coding == 1: boxes = boxes @ (_XYWH2XYXY.to(boxes.device)) boxes = boxes if boxes.dim() == 4 else boxes.unsqueeze(2) _, _, numClasses = scores.shape num_det, det_boxes, det_scores, det_classes = TRTbatchedNMSop.apply( boxes, scores, '1', 1, -1, int(numClasses), min(pre_top_k, 4096), keep_top_k, score_threshold, iou_threshold, 0, 0, 16, 1) det_classes = det_classes.int() return num_det, det_boxes, det_scores, det_classes def efficient_nms(*args, **kwargs): """Wrapper function for `_efficient_nms`.""" return _efficient_nms(*args, **kwargs) def batched_nms(*args, **kwargs): """Wrapper function for `_batched_nms`.""" return _batched_nms(*args, **kwargs)
8,045
34.444934
78
py
mmyolo
mmyolo-main/.dev_scripts/gather_models.py
# Copyright (c) OpenMMLab. All rights reserved. import argparse import glob import os import os.path as osp import shutil import subprocess import time from collections import OrderedDict import torch import yaml from mmengine.config import Config from mmengine.fileio import dump from mmengine.utils import mkdir_or_exist, scandir def ordered_yaml_dump(data, stream=None, Dumper=yaml.SafeDumper, **kwds): class OrderedDumper(Dumper): pass def _dict_representer(dumper, data): return dumper.represent_mapping( yaml.resolver.BaseResolver.DEFAULT_MAPPING_TAG, data.items()) OrderedDumper.add_representer(OrderedDict, _dict_representer) return yaml.dump(data, stream, OrderedDumper, **kwds) def process_checkpoint(in_file, out_file): checkpoint = torch.load(in_file, map_location='cpu') # remove optimizer for smaller file size if 'optimizer' in checkpoint: del checkpoint['optimizer'] if 'message_hub' in checkpoint: del checkpoint['message_hub'] if 'ema_state_dict' in checkpoint: del checkpoint['ema_state_dict'] for key in list(checkpoint['state_dict']): if key.startswith('data_preprocessor'): checkpoint['state_dict'].pop(key) elif 'priors_base_sizes' in key: checkpoint['state_dict'].pop(key) elif 'grid_offset' in key: checkpoint['state_dict'].pop(key) elif 'prior_inds' in key: checkpoint['state_dict'].pop(key) # if it is necessary to remove some sensitive data in checkpoint['meta'], # add the code here. if torch.__version__ >= '1.6': torch.save(checkpoint, out_file, _use_new_zipfile_serialization=False) else: torch.save(checkpoint, out_file) sha = subprocess.check_output(['sha256sum', out_file]).decode() final_file = out_file.rstrip('.pth') + f'-{sha[:8]}.pth' subprocess.Popen(['mv', out_file, final_file]) return final_file def is_by_epoch(config): cfg = Config.fromfile('./configs/' + config) return cfg.train_cfg.type == 'EpochBasedTrainLoop' def get_final_epoch_or_iter(config): cfg = Config.fromfile('./configs/' + config) if cfg.train_cfg.type == 'EpochBasedTrainLoop': return cfg.train_cfg.max_epochs else: return cfg.train_cfg.max_iters def get_best_epoch_or_iter(exp_dir): best_epoch_iter_full_path = list( sorted(glob.glob(osp.join(exp_dir, 'best_*.pth'))))[-1] best_epoch_or_iter_model_path = best_epoch_iter_full_path.split('/')[-1] best_epoch_or_iter = best_epoch_or_iter_model_path. \ split('_')[-1].split('.')[0] return best_epoch_or_iter_model_path, int(best_epoch_or_iter) def get_real_epoch_or_iter(config): cfg = Config.fromfile('./configs/' + config) if cfg.train_cfg.type == 'EpochBasedTrainLoop': epoch = cfg.train_cfg.max_epochs return epoch else: return cfg.runner.max_iters def get_final_results(log_json_path, epoch_or_iter, results_lut='coco/bbox_mAP', by_epoch=True): result_dict = dict() with open(log_json_path) as f: r = f.readlines()[-1] last_metric = r.split(',')[0].split(': ')[-1].strip() result_dict[results_lut] = last_metric return result_dict def get_dataset_name(config): # If there are more dataset, add here. name_map = dict( CityscapesDataset='Cityscapes', CocoDataset='COCO', YOLOv5CocoDataset='COCO', CocoPanopticDataset='COCO', YOLOv5DOTADataset='DOTA 1.0', DeepFashionDataset='Deep Fashion', LVISV05Dataset='LVIS v0.5', LVISV1Dataset='LVIS v1', VOCDataset='Pascal VOC', YOLOv5VOCDataset='Pascal VOC', WIDERFaceDataset='WIDER Face', OpenImagesDataset='OpenImagesDataset', OpenImagesChallengeDataset='OpenImagesChallengeDataset') cfg = Config.fromfile('./configs/' + config) return name_map[cfg.dataset_type] def find_last_dir(model_dir): dst_times = [] for time_stamp in os.scandir(model_dir): if osp.isdir(time_stamp): dst_time = time.mktime( time.strptime(time_stamp.name, '%Y%m%d_%H%M%S')) dst_times.append([dst_time, time_stamp.name]) return max(dst_times, key=lambda x: x[0])[1] def convert_model_info_to_pwc(model_infos): pwc_files = {} for model in model_infos: cfg_folder_name = osp.split(model['config'])[-2] pwc_model_info = OrderedDict() pwc_model_info['Name'] = osp.split(model['config'])[-1].split('.')[0] pwc_model_info['In Collection'] = 'Please fill in Collection name' pwc_model_info['Config'] = osp.join('configs', model['config']) # get metadata meta_data = OrderedDict() if 'epochs' in model: meta_data['Epochs'] = get_real_epoch_or_iter(model['config']) else: meta_data['Iterations'] = get_real_epoch_or_iter(model['config']) pwc_model_info['Metadata'] = meta_data # get dataset name dataset_name = get_dataset_name(model['config']) # get results results = [] # if there are more metrics, add here. if 'bbox_mAP' in model['results']: metric = round(model['results']['bbox_mAP'] * 100, 1) results.append( OrderedDict( Task='Object Detection', Dataset=dataset_name, Metrics={'box AP': metric})) if 'segm_mAP' in model['results']: metric = round(model['results']['segm_mAP'] * 100, 1) results.append( OrderedDict( Task='Instance Segmentation', Dataset=dataset_name, Metrics={'mask AP': metric})) if 'PQ' in model['results']: metric = round(model['results']['PQ'], 1) results.append( OrderedDict( Task='Panoptic Segmentation', Dataset=dataset_name, Metrics={'PQ': metric})) pwc_model_info['Results'] = results link_string = 'https://download.openmmlab.com/mmyolo/v0/' link_string += '{}/{}'.format(model['config'].rstrip('.py'), osp.split(model['model_path'])[-1]) pwc_model_info['Weights'] = link_string if cfg_folder_name in pwc_files: pwc_files[cfg_folder_name].append(pwc_model_info) else: pwc_files[cfg_folder_name] = [pwc_model_info] return pwc_files def parse_args(): parser = argparse.ArgumentParser(description='Gather benchmarked models') parser.add_argument( 'root', type=str, help='root path of benchmarked models to be gathered') parser.add_argument( 'out', type=str, help='output path of gathered models to be stored') parser.add_argument( '--best', action='store_true', help='whether to gather the best model.') args = parser.parse_args() return args # TODO: Refine def main(): args = parse_args() models_root = args.root models_out = args.out mkdir_or_exist(models_out) # find all models in the root directory to be gathered raw_configs = list(scandir('./configs', '.py', recursive=True)) # filter configs that is not trained in the experiments dir used_configs = [] for raw_config in raw_configs: if osp.exists(osp.join(models_root, raw_config)): used_configs.append(raw_config) print(f'Find {len(used_configs)} models to be gathered') # find final_ckpt and log file for trained each config # and parse the best performance model_infos = [] for used_config in used_configs: exp_dir = osp.join(models_root, used_config) by_epoch = is_by_epoch(used_config) # check whether the exps is finished if args.best is True: final_model, final_epoch_or_iter = get_best_epoch_or_iter(exp_dir) else: final_epoch_or_iter = get_final_epoch_or_iter(used_config) final_model = '{}_{}.pth'.format('epoch' if by_epoch else 'iter', final_epoch_or_iter) model_path = osp.join(exp_dir, final_model) # skip if the model is still training if not osp.exists(model_path): continue # get the latest logs latest_exp_name = find_last_dir(exp_dir) latest_exp_json = osp.join(exp_dir, latest_exp_name, 'vis_data', latest_exp_name + '.json') model_performance = get_final_results( latest_exp_json, final_epoch_or_iter, by_epoch=by_epoch) if model_performance is None: continue model_info = dict( config=used_config, results=model_performance, final_model=final_model, latest_exp_json=latest_exp_json, latest_exp_name=latest_exp_name) model_info['epochs' if by_epoch else 'iterations'] = \ final_epoch_or_iter model_infos.append(model_info) # publish model for each checkpoint publish_model_infos = [] for model in model_infos: model_publish_dir = osp.join(models_out, model['config'].rstrip('.py')) mkdir_or_exist(model_publish_dir) model_name = osp.split(model['config'])[-1].split('.')[0] model_name += '_' + model['latest_exp_name'] publish_model_path = osp.join(model_publish_dir, model_name) trained_model_path = osp.join(models_root, model['config'], model['final_model']) # convert model final_model_path = process_checkpoint(trained_model_path, publish_model_path) # copy log shutil.copy(model['latest_exp_json'], osp.join(model_publish_dir, f'{model_name}.log.json')) # copy config to guarantee reproducibility config_path = model['config'] config_path = osp.join( 'configs', config_path) if 'configs' not in config_path else config_path target_config_path = osp.split(config_path)[-1] shutil.copy(config_path, osp.join(model_publish_dir, target_config_path)) model['model_path'] = final_model_path publish_model_infos.append(model) models = dict(models=publish_model_infos) print(f'Totally gathered {len(publish_model_infos)} models') dump(models, osp.join(models_out, 'model_info.json')) pwc_files = convert_model_info_to_pwc(publish_model_infos) for name in pwc_files: with open(osp.join(models_out, name + '_metafile.yml'), 'w') as f: ordered_yaml_dump(pwc_files[name], f, encoding='utf-8') if __name__ == '__main__': main()
11,017
34.314103
79
py
mmyolo
mmyolo-main/tests/test_engine/test_hooks/test_yolox_mode_switch_hook.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase from unittest.mock import Mock import torch from mmengine.config import Config from mmengine.runner import Runner from torch.utils.data import Dataset from mmyolo.engine.hooks import YOLOXModeSwitchHook from mmyolo.utils import register_all_modules class DummyDataset(Dataset): METAINFO = dict() # type: ignore data = torch.randn(12, 2) label = torch.ones(12) @property def metainfo(self): return self.METAINFO def __len__(self): return self.data.size(0) def __getitem__(self, index): return dict(inputs=self.data[index], data_sample=self.label[index]) pipeline1 = [ dict(type='mmdet.Resize'), ] pipeline2 = [ dict(type='mmdet.RandomFlip'), ] register_all_modules() class TestYOLOXModeSwitchHook(TestCase): def test(self): train_dataloader = dict( dataset=DummyDataset(), sampler=dict(type='DefaultSampler', shuffle=True), batch_size=3, num_workers=0) runner = Mock() runner.model = Mock() runner.model.module = Mock() runner.model.bbox_head.use_bbox_aux = False runner.cfg.train_dataloader = Config(train_dataloader) runner.train_dataloader = Runner.build_dataloader(train_dataloader) runner.train_dataloader.dataset.pipeline = pipeline1 hook = YOLOXModeSwitchHook( num_last_epochs=15, new_train_pipeline=pipeline2) # test after change mode runner.epoch = 284 runner.max_epochs = 300 hook.before_train_epoch(runner) self.assertTrue(runner.model.bbox_head.use_bbox_aux) self.assertEqual(runner.train_loop.dataloader.dataset.pipeline, pipeline2)
1,810
25.632353
75
py
mmyolo
mmyolo-main/tests/test_engine/test_hooks/test_yolov5_param_scheduler_hook.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase from unittest.mock import Mock import torch from mmengine.config import Config from mmengine.optim import build_optim_wrapper from mmengine.runner import Runner from torch import nn from torch.utils.data import Dataset from mmyolo.engine.hooks import YOLOv5ParamSchedulerHook from mmyolo.utils import register_all_modules class ToyModel(nn.Module): def __init__(self): super().__init__() self.linear = nn.Linear(2, 1) def forward(self, inputs, data_samples, mode='tensor'): labels = torch.stack(data_samples) inputs = torch.stack(inputs) outputs = self.linear(inputs) if mode == 'tensor': return outputs elif mode == 'loss': loss = (labels - outputs).sum() outputs = dict(loss=loss) return outputs else: return outputs class DummyDataset(Dataset): METAINFO = dict() # type: ignore data = torch.randn(12, 2) label = torch.ones(12) @property def metainfo(self): return self.METAINFO def __len__(self): return self.data.size(0) def __getitem__(self, index): return dict(inputs=self.data[index], data_sample=self.label[index]) optim_wrapper = dict( type='OptimWrapper', optimizer=dict( type='SGD', lr=0.01, momentum=0.937, weight_decay=0.0005, nesterov=True, batch_size_per_gpu=1), constructor='YOLOv5OptimizerConstructor') register_all_modules() class TestYOLOv5ParamSchelerHook(TestCase): def test(self): model = ToyModel() train_dataloader = dict( dataset=DummyDataset(), sampler=dict(type='DefaultSampler', shuffle=True), batch_size=3, num_workers=0) runner = Mock() runner.model = model runner.optim_wrapper = build_optim_wrapper(model, optim_wrapper) runner.cfg.train_dataloader = Config(train_dataloader) runner.train_dataloader = Runner.build_dataloader(train_dataloader) hook = YOLOv5ParamSchedulerHook( scheduler_type='linear', lr_factor=0.01, max_epochs=300) # test before train runner.epoch = 0 runner.iter = 0 hook.before_train(runner) for group in runner.optim_wrapper.param_groups: self.assertEqual(group['lr'], 0.01) self.assertEqual(group['momentum'], 0.937) self.assertFalse(hook._warmup_end) # test after training 10 steps for i in range(10): runner.iter += 1 hook.before_train_iter(runner, 0) for group_idx, group in enumerate(runner.optim_wrapper.param_groups): if group_idx == 2: self.assertEqual(round(group['lr'], 5), 0.0991) self.assertEqual(group['momentum'], 0.80137) self.assertFalse(hook._warmup_end) # test after warm up runner.iter = 1000 hook.before_train_iter(runner, 0) self.assertFalse(hook._warmup_end) for group in runner.optim_wrapper.param_groups: self.assertEqual(group['lr'], 0.01) self.assertEqual(group['momentum'], 0.937) runner.iter = 1001 hook.before_train_iter(runner, 0) self.assertTrue(hook._warmup_end) # test after train_epoch hook.after_train_epoch(runner) for group in runner.optim_wrapper.param_groups: self.assertEqual(group['lr'], 0.01) self.assertEqual(group['momentum'], 0.937)
3,619
27.96
77
py
mmyolo
mmyolo-main/tests/test_engine/test_optimizers/test_yolov5_optim_constructor.py
# Copyright (c) OpenMMLab. All rights reserved. import copy from unittest import TestCase import torch import torch.nn as nn from mmengine.optim import build_optim_wrapper from mmyolo.engine import YOLOv5OptimizerConstructor from mmyolo.utils import register_all_modules register_all_modules() class ExampleModel(nn.Module): def __init__(self): super().__init__() self.param1 = nn.Parameter(torch.ones(1)) self.conv1 = nn.Conv2d(3, 4, kernel_size=1, bias=False) self.conv2 = nn.Conv2d(4, 2, kernel_size=1) self.bn = nn.BatchNorm2d(2) class TestYOLOv5OptimizerConstructor(TestCase): def setUp(self): self.model = ExampleModel() self.base_lr = 0.01 self.weight_decay = 0.0001 self.optim_wrapper_cfg = dict( type='OptimWrapper', optimizer=dict( type='SGD', lr=self.base_lr, momentum=0.9, weight_decay=self.weight_decay, batch_size_per_gpu=16)) def test_init(self): YOLOv5OptimizerConstructor(copy.deepcopy(self.optim_wrapper_cfg)) YOLOv5OptimizerConstructor( copy.deepcopy(self.optim_wrapper_cfg), paramwise_cfg={'base_total_batch_size': 64}) # `paramwise_cfg` must include `base_total_batch_size` if not None. with self.assertRaises(AssertionError): YOLOv5OptimizerConstructor( copy.deepcopy(self.optim_wrapper_cfg), paramwise_cfg={'a': 64}) def test_build(self): optim_wrapper = YOLOv5OptimizerConstructor( copy.deepcopy(self.optim_wrapper_cfg))( self.model) # test param_groups assert len(optim_wrapper.optimizer.param_groups) == 3 for i in range(3): param_groups_i = optim_wrapper.optimizer.param_groups[i] assert param_groups_i['lr'] == self.base_lr if i == 0: assert param_groups_i['weight_decay'] == self.weight_decay else: assert param_groups_i['weight_decay'] == 0 # test weight_decay linear scaling optim_wrapper_cfg = copy.deepcopy(self.optim_wrapper_cfg) optim_wrapper_cfg['optimizer']['batch_size_per_gpu'] = 128 optim_wrapper = YOLOv5OptimizerConstructor(optim_wrapper_cfg)( self.model) assert optim_wrapper.optimizer.param_groups[0][ 'weight_decay'] == self.weight_decay * 2 # test without batch_size_per_gpu optim_wrapper_cfg = copy.deepcopy(self.optim_wrapper_cfg) optim_wrapper_cfg['optimizer'].pop('batch_size_per_gpu') optim_wrapper = dict( optim_wrapper_cfg, constructor='YOLOv5OptimizerConstructor') optim_wrapper = build_optim_wrapper(self.model, optim_wrapper) assert optim_wrapper.optimizer.param_groups[0][ 'weight_decay'] == self.weight_decay
2,934
34.792683
79
py
mmyolo
mmyolo-main/tests/test_engine/test_optimizers/test_yolov7_optim_wrapper_constructor.py
# Copyright (c) OpenMMLab. All rights reserved. import copy from unittest import TestCase import torch import torch.nn as nn from mmengine.optim import build_optim_wrapper from mmyolo.engine import YOLOv7OptimWrapperConstructor from mmyolo.utils import register_all_modules register_all_modules() class ExampleModel(nn.Module): def __init__(self): super().__init__() self.param1 = nn.Parameter(torch.ones(1)) self.conv1 = nn.Conv2d(3, 4, kernel_size=1, bias=False) self.conv2 = nn.Conv2d(4, 2, kernel_size=1) self.bn = nn.BatchNorm2d(2) class TestYOLOv7OptimWrapperConstructor(TestCase): def setUp(self): self.model = ExampleModel() self.base_lr = 0.01 self.weight_decay = 0.0001 self.optim_wrapper_cfg = dict( type='OptimWrapper', optimizer=dict( type='SGD', lr=self.base_lr, momentum=0.9, weight_decay=self.weight_decay, batch_size_per_gpu=16)) def test_init(self): YOLOv7OptimWrapperConstructor(copy.deepcopy(self.optim_wrapper_cfg)) YOLOv7OptimWrapperConstructor( copy.deepcopy(self.optim_wrapper_cfg), paramwise_cfg={'base_total_batch_size': 64}) # `paramwise_cfg` must include `base_total_batch_size` if not None. with self.assertRaises(AssertionError): YOLOv7OptimWrapperConstructor( copy.deepcopy(self.optim_wrapper_cfg), paramwise_cfg={'a': 64}) def test_build(self): optim_wrapper = YOLOv7OptimWrapperConstructor( copy.deepcopy(self.optim_wrapper_cfg))( self.model) # test param_groups assert len(optim_wrapper.optimizer.param_groups) == 3 for i in range(3): param_groups_i = optim_wrapper.optimizer.param_groups[i] assert param_groups_i['lr'] == self.base_lr if i == 0: assert param_groups_i['weight_decay'] == self.weight_decay else: assert param_groups_i['weight_decay'] == 0 # test weight_decay linear scaling optim_wrapper_cfg = copy.deepcopy(self.optim_wrapper_cfg) optim_wrapper_cfg['optimizer']['batch_size_per_gpu'] = 128 optim_wrapper = YOLOv7OptimWrapperConstructor(optim_wrapper_cfg)( self.model) assert optim_wrapper.optimizer.param_groups[0][ 'weight_decay'] == self.weight_decay * 2 # test without batch_size_per_gpu optim_wrapper_cfg = copy.deepcopy(self.optim_wrapper_cfg) optim_wrapper_cfg['optimizer'].pop('batch_size_per_gpu') optim_wrapper = dict( optim_wrapper_cfg, constructor='YOLOv7OptimWrapperConstructor') optim_wrapper = build_optim_wrapper(self.model, optim_wrapper) assert optim_wrapper.optimizer.param_groups[0][ 'weight_decay'] == self.weight_decay
2,958
35.085366
79
py
mmyolo
mmyolo-main/tests/test_deploy/test_object_detection.py
# Copyright (c) OpenMMLab. All rights reserved. import os from tempfile import NamedTemporaryFile, TemporaryDirectory import numpy as np import pytest import torch from mmengine import Config try: import importlib importlib.import_module('mmdeploy') except ImportError: pytest.skip('mmdeploy is not installed.', allow_module_level=True) import mmdeploy.backend.onnxruntime as ort_apis from mmdeploy.apis import build_task_processor from mmdeploy.codebase import import_codebase from mmdeploy.utils import load_config from mmdeploy.utils.config_utils import register_codebase from mmdeploy.utils.test import SwitchBackendWrapper try: codebase = register_codebase('mmyolo') import_codebase(codebase, ['mmyolo.deploy']) except ImportError: pytest.skip('mmyolo is not installed.', allow_module_level=True) model_cfg_path = 'tests/test_deploy/data/model.py' model_cfg = load_config(model_cfg_path)[0] model_cfg.test_dataloader.dataset.data_root = \ 'tests/data' model_cfg.test_dataloader.dataset.ann_file = 'coco_sample.json' model_cfg.test_evaluator.ann_file = \ 'tests/coco_sample.json' deploy_cfg = Config( dict( backend_config=dict(type='onnxruntime'), codebase_config=dict( type='mmyolo', task='ObjectDetection', post_processing=dict( score_threshold=0.05, confidence_threshold=0.005, # for YOLOv3 iou_threshold=0.5, max_output_boxes_per_class=200, pre_top_k=5000, keep_top_k=100, background_label_id=-1, ), module=['mmyolo.deploy']), onnx_config=dict( type='onnx', export_params=True, keep_initializers_as_inputs=False, opset_version=11, input_shape=None, input_names=['input'], output_names=['dets', 'labels']))) onnx_file = NamedTemporaryFile(suffix='.onnx').name task_processor = None img_shape = (32, 32) img = np.random.rand(*img_shape, 3) @pytest.fixture(autouse=True) def init_task_processor(): global task_processor task_processor = build_task_processor(model_cfg, deploy_cfg, 'cpu') @pytest.fixture def backend_model(): from mmdeploy.backend.onnxruntime import ORTWrapper ort_apis.__dict__.update({'ORTWrapper': ORTWrapper}) wrapper = SwitchBackendWrapper(ORTWrapper) wrapper.set( outputs={ 'dets': torch.rand(1, 10, 5).sort(2).values, 'labels': torch.randint(0, 10, (1, 10)) }) yield task_processor.build_backend_model(['']) wrapper.recover() def test_visualize(backend_model): img_path = 'tests/data/color.jpg' input_dict, _ = task_processor.create_input( img_path, input_shape=img_shape) results = backend_model.test_step(input_dict)[0] with TemporaryDirectory() as dir: filename = dir + 'tmp.jpg' task_processor.visualize(img, results, filename, 'window') assert os.path.exists(filename)
3,052
30.474227
71
py
mmyolo
mmyolo-main/tests/test_deploy/test_mmyolo_models.py
# Copyright (c) OpenMMLab. All rights reserved. import os import random import numpy as np import pytest import torch from mmengine import Config try: import importlib importlib.import_module('mmdeploy') except ImportError: pytest.skip('mmdeploy is not installed.', allow_module_level=True) from mmdeploy.codebase import import_codebase from mmdeploy.utils import Backend from mmdeploy.utils.config_utils import register_codebase from mmdeploy.utils.test import (WrapModel, check_backend, get_model_outputs, get_rewrite_outputs) try: codebase = register_codebase('mmyolo') import_codebase(codebase, ['mmyolo.deploy']) except ImportError: pytest.skip('mmyolo is not installed.', allow_module_level=True) def seed_everything(seed=1029): random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) np.random.seed(seed) torch.manual_seed(seed) if torch.cuda.is_available(): torch.cuda.manual_seed(seed) torch.cuda.manual_seed_all(seed) # if you are using multi-GPU. torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True torch.backends.cudnn.enabled = False def get_yolov5_head_model(): """YOLOv5 Head Config.""" test_cfg = Config( dict( multi_label=True, nms_pre=30000, score_thr=0.001, nms=dict(type='nms', iou_threshold=0.65), max_per_img=300)) from mmyolo.models.dense_heads import YOLOv5Head head_module = dict( type='YOLOv5HeadModule', num_classes=4, in_channels=[2, 4, 8], featmap_strides=[8, 16, 32], num_base_priors=1) model = YOLOv5Head(head_module, test_cfg=test_cfg) model.requires_grad_(False) return model @pytest.mark.parametrize('backend_type', [Backend.ONNXRUNTIME]) def test_yolov5_head_predict_by_feat(backend_type: Backend): """Test predict_by_feat rewrite of YOLOXHead.""" check_backend(backend_type) yolov5_head = get_yolov5_head_model() yolov5_head.cpu().eval() s = 256 batch_img_metas = [{ 'scale_factor': (1.0, 1.0), 'pad_shape': (s, s, 3), 'img_shape': (s, s, 3), 'ori_shape': (s, s, 3) }] output_names = ['dets', 'labels'] deploy_cfg = Config( dict( backend_config=dict(type=backend_type.value), onnx_config=dict(output_names=output_names, input_shape=None), codebase_config=dict( type='mmyolo', task='ObjectDetection', post_processing=dict( score_threshold=0.05, iou_threshold=0.5, max_output_boxes_per_class=20, pre_top_k=-1, keep_top_k=10, background_label_id=-1, ), module=['mmyolo.deploy']))) seed_everything(1234) cls_scores = [ torch.rand(1, yolov5_head.num_classes * yolov5_head.num_base_priors, 4 * pow(2, i), 4 * pow(2, i)) for i in range(3, 0, -1) ] seed_everything(5678) bbox_preds = [ torch.rand(1, 4 * yolov5_head.num_base_priors, 4 * pow(2, i), 4 * pow(2, i)) for i in range(3, 0, -1) ] seed_everything(9101) objectnesses = [ torch.rand(1, 1 * yolov5_head.num_base_priors, 4 * pow(2, i), 4 * pow(2, i)) for i in range(3, 0, -1) ] # to get outputs of pytorch model model_inputs = { 'cls_scores': cls_scores, 'bbox_preds': bbox_preds, 'objectnesses': objectnesses, 'batch_img_metas': batch_img_metas, 'with_nms': True } model_outputs = get_model_outputs(yolov5_head, 'predict_by_feat', model_inputs) # to get outputs of onnx model after rewrite wrapped_model = WrapModel( yolov5_head, 'predict_by_feat', batch_img_metas=batch_img_metas, with_nms=True) rewrite_inputs = { 'cls_scores': cls_scores, 'bbox_preds': bbox_preds, 'objectnesses': objectnesses, } rewrite_outputs, is_backend_output = get_rewrite_outputs( wrapped_model=wrapped_model, model_inputs=rewrite_inputs, deploy_cfg=deploy_cfg) if is_backend_output: # hard code to make two tensors with the same shape # rewrite and original codes applied different nms strategy min_shape = min(model_outputs[0].bboxes.shape[0], rewrite_outputs[0].shape[1], 5) for i in range(len(model_outputs)): rewrite_outputs[0][i, :min_shape, 0::2] = \ rewrite_outputs[0][i, :min_shape, 0::2].clamp_(0, s) rewrite_outputs[0][i, :min_shape, 1::2] = \ rewrite_outputs[0][i, :min_shape, 1::2].clamp_(0, s) assert np.allclose( model_outputs[i].bboxes[:min_shape], rewrite_outputs[0][i, :min_shape, :4], rtol=1e-03, atol=1e-05) assert np.allclose( model_outputs[i].scores[:min_shape], rewrite_outputs[0][i, :min_shape, 4], rtol=1e-03, atol=1e-05) assert np.allclose( model_outputs[i].labels[:min_shape], rewrite_outputs[1][i, :min_shape], rtol=1e-03, atol=1e-05) else: assert rewrite_outputs is not None
5,546
32.415663
77
py
mmyolo
mmyolo-main/tests/test_models/test_layers/test_ema.py
# Copyright (c) OpenMMLab. All rights reserved. import itertools import math from unittest import TestCase import torch import torch.nn as nn from mmengine.testing import assert_allclose from mmyolo.models.layers import ExpMomentumEMA class TestEMA(TestCase): def test_exp_momentum_ema(self): model = nn.Sequential(nn.Conv2d(1, 5, kernel_size=3), nn.Linear(5, 10)) # Test invalid gamma with self.assertRaisesRegex(AssertionError, 'gamma must be greater than 0'): ExpMomentumEMA(model, gamma=-1) # Test EMA model = torch.nn.Sequential( torch.nn.Conv2d(1, 5, kernel_size=3), torch.nn.Linear(5, 10)) momentum = 0.1 gamma = 4 ema_model = ExpMomentumEMA(model, momentum=momentum, gamma=gamma) averaged_params = [ torch.zeros_like(param) for param in model.parameters() ] n_updates = 10 for i in range(n_updates): updated_averaged_params = [] for p, p_avg in zip(model.parameters(), averaged_params): p.detach().add_(torch.randn_like(p)) if i == 0: updated_averaged_params.append(p.clone()) else: m = (1 - momentum) * math.exp(-(1 + i) / gamma) + momentum updated_averaged_params.append( (p_avg * (1 - m) + p * m).clone()) ema_model.update_parameters(model) averaged_params = updated_averaged_params for p_target, p_ema in zip(averaged_params, ema_model.parameters()): assert_allclose(p_target, p_ema) def test_exp_momentum_ema_update_buffer(self): model = nn.Sequential( nn.Conv2d(1, 5, kernel_size=3), nn.BatchNorm2d(5, momentum=0.3), nn.Linear(5, 10)) # Test invalid gamma with self.assertRaisesRegex(AssertionError, 'gamma must be greater than 0'): ExpMomentumEMA(model, gamma=-1) # Test EMA with momentum annealing. momentum = 0.1 gamma = 4 ema_model = ExpMomentumEMA( model, gamma=gamma, momentum=momentum, update_buffers=True) averaged_params = [ torch.zeros_like(param) for param in itertools.chain(model.parameters(), model.buffers()) if param.size() != torch.Size([]) ] n_updates = 10 for i in range(n_updates): updated_averaged_params = [] params = [ param for param in itertools.chain(model.parameters(), model.buffers()) if param.size() != torch.Size([]) ] for p, p_avg in zip(params, averaged_params): p.detach().add_(torch.randn_like(p)) if i == 0: updated_averaged_params.append(p.clone()) else: m = (1 - momentum) * math.exp(-(1 + i) / gamma) + momentum updated_averaged_params.append( (p_avg * (1 - m) + p * m).clone()) ema_model.update_parameters(model) averaged_params = updated_averaged_params ema_params = [ param for param in itertools.chain(ema_model.module.parameters(), ema_model.module.buffers()) if param.size() != torch.Size([]) ] for p_target, p_ema in zip(averaged_params, ema_params): assert_allclose(p_target, p_ema)
3,634
37.263158
79
py
mmyolo
mmyolo-main/tests/test_models/test_layers/test_yolo_bricks.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models.layers import SPPFBottleneck from mmyolo.utils import register_all_modules register_all_modules() class TestSPPFBottleneck(TestCase): def test_forward(self): input_tensor = torch.randn((1, 3, 20, 20)) bottleneck = SPPFBottleneck(3, 16) out_tensor = bottleneck(input_tensor) self.assertEqual(out_tensor.shape, (1, 16, 20, 20)) bottleneck = SPPFBottleneck(3, 16, kernel_sizes=[3, 5, 7]) out_tensor = bottleneck(input_tensor) self.assertEqual(out_tensor.shape, (1, 16, 20, 20)) # set len(kernel_sizes)=4 bottleneck = SPPFBottleneck(3, 16, kernel_sizes=[3, 5, 7, 9]) out_tensor = bottleneck(input_tensor) self.assertEqual(out_tensor.shape, (1, 16, 20, 20)) # set use_conv_first=False bottleneck = SPPFBottleneck( 3, 16, use_conv_first=False, kernel_sizes=[3, 5, 7, 9]) out_tensor = bottleneck(input_tensor) self.assertEqual(out_tensor.shape, (1, 16, 20, 20))
1,111
30.771429
69
py
mmyolo
mmyolo-main/tests/test_models/test_backbone/test_csp_resnet.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import pytest import torch from torch.nn.modules.batchnorm import _BatchNorm from mmyolo.models import PPYOLOECSPResNet from mmyolo.utils import register_all_modules from .utils import check_norm_state, is_norm register_all_modules() class TestPPYOLOECSPResNet(TestCase): def test_init(self): # out_indices in range(len(arch_setting) + 1) with pytest.raises(AssertionError): PPYOLOECSPResNet(out_indices=(6, )) with pytest.raises(ValueError): # frozen_stages must in range(-1, len(arch_setting) + 1) PPYOLOECSPResNet(frozen_stages=6) def test_forward(self): # Test PPYOLOECSPResNet with first stage frozen frozen_stages = 1 model = PPYOLOECSPResNet(frozen_stages=frozen_stages) model.init_weights() model.train() for mod in model.stem.modules(): for param in mod.parameters(): assert param.requires_grad is False for i in range(1, frozen_stages + 1): layer = getattr(model, f'stage{i}') for mod in layer.modules(): if isinstance(mod, _BatchNorm): assert mod.training is False for param in layer.parameters(): assert param.requires_grad is False # Test PPYOLOECSPResNet with norm_eval=True model = PPYOLOECSPResNet(norm_eval=True) model.train() assert check_norm_state(model.modules(), False) # Test PPYOLOECSPResNet-P5 forward with widen_factor=0.25 model = PPYOLOECSPResNet( arch='P5', widen_factor=0.25, out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 16, 32, 32)) assert feat[1].shape == torch.Size((1, 32, 16, 16)) assert feat[2].shape == torch.Size((1, 64, 8, 8)) assert feat[3].shape == torch.Size((1, 128, 4, 4)) assert feat[4].shape == torch.Size((1, 256, 2, 2)) # Test PPYOLOECSPResNet forward with dict(type='ReLU') model = PPYOLOECSPResNet( widen_factor=0.125, act_cfg=dict(type='ReLU'), out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test PPYOLOECSPResNet with BatchNorm forward model = PPYOLOECSPResNet(widen_factor=0.125, out_indices=range(0, 5)) for m in model.modules(): if is_norm(m): assert isinstance(m, _BatchNorm) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test PPYOLOECSPResNet with BatchNorm forward model = PPYOLOECSPResNet(plugins=[ dict( cfg=dict(type='mmdet.DropBlock', drop_prob=0.1, block_size=3), stages=(False, False, True, True)), ]) assert len(model.stage1) == 1 assert len(model.stage2) == 1 assert len(model.stage3) == 2 # +DropBlock assert len(model.stage4) == 2 # +DropBlock model.train() imgs = torch.randn(1, 3, 256, 256) feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 256, 32, 32)) assert feat[1].shape == torch.Size((1, 512, 16, 16)) assert feat[2].shape == torch.Size((1, 1024, 8, 8))
4,162
35.517544
78
py
mmyolo
mmyolo-main/tests/test_models/test_backbone/test_efficient_rep.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import pytest import torch from torch.nn.modules.batchnorm import _BatchNorm from mmyolo.models.backbones import YOLOv6CSPBep, YOLOv6EfficientRep from mmyolo.utils import register_all_modules from .utils import check_norm_state, is_norm register_all_modules() class TestYOLOv6EfficientRep(TestCase): def test_init(self): # out_indices in range(len(arch_setting) + 1) with pytest.raises(AssertionError): YOLOv6EfficientRep(out_indices=(6, )) with pytest.raises(ValueError): # frozen_stages must in range(-1, len(arch_setting) + 1) YOLOv6EfficientRep(frozen_stages=6) def test_YOLOv6EfficientRep_forward(self): # Test YOLOv6EfficientRep with first stage frozen frozen_stages = 1 model = YOLOv6EfficientRep(frozen_stages=frozen_stages) model.init_weights() model.train() for mod in model.stem.modules(): for param in mod.parameters(): assert param.requires_grad is False for i in range(1, frozen_stages + 1): layer = getattr(model, f'stage{i}') for mod in layer.modules(): if isinstance(mod, _BatchNorm): assert mod.training is False for param in layer.parameters(): assert param.requires_grad is False # Test YOLOv6EfficientRep with norm_eval=True model = YOLOv6EfficientRep(norm_eval=True) model.train() assert check_norm_state(model.modules(), False) # Test YOLOv6EfficientRep-P5 forward with widen_factor=0.25 model = YOLOv6EfficientRep( arch='P5', widen_factor=0.25, out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 16, 32, 32)) assert feat[1].shape == torch.Size((1, 32, 16, 16)) assert feat[2].shape == torch.Size((1, 64, 8, 8)) assert feat[3].shape == torch.Size((1, 128, 4, 4)) assert feat[4].shape == torch.Size((1, 256, 2, 2)) # Test YOLOv6EfficientRep forward with dict(type='ReLU') model = YOLOv6EfficientRep( widen_factor=0.125, act_cfg=dict(type='ReLU'), out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test YOLOv6EfficientRep with BatchNorm forward model = YOLOv6EfficientRep(widen_factor=0.125, out_indices=range(0, 5)) for m in model.modules(): if is_norm(m): assert isinstance(m, _BatchNorm) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test YOLOv6EfficientRep with BatchNorm forward model = YOLOv6EfficientRep(plugins=[ dict( cfg=dict(type='mmdet.DropBlock', drop_prob=0.1, block_size=3), stages=(False, False, True, True)), ]) assert len(model.stage1) == 1 assert len(model.stage2) == 1 assert len(model.stage3) == 2 # +DropBlock assert len(model.stage4) == 3 # +SPPF+DropBlock model.train() imgs = torch.randn(1, 3, 256, 256) feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 256, 32, 32)) assert feat[1].shape == torch.Size((1, 512, 16, 16)) assert feat[2].shape == torch.Size((1, 1024, 8, 8)) def test_YOLOv6CSPBep_forward(self): # Test YOLOv6CSPBep with first stage frozen frozen_stages = 1 model = YOLOv6CSPBep(frozen_stages=frozen_stages) model.init_weights() model.train() for mod in model.stem.modules(): for param in mod.parameters(): assert param.requires_grad is False for i in range(1, frozen_stages + 1): layer = getattr(model, f'stage{i}') for mod in layer.modules(): if isinstance(mod, _BatchNorm): assert mod.training is False for param in layer.parameters(): assert param.requires_grad is False # Test YOLOv6CSPBep with norm_eval=True model = YOLOv6CSPBep(norm_eval=True) model.train() assert check_norm_state(model.modules(), False) # Test YOLOv6CSPBep forward with widen_factor=0.25 model = YOLOv6CSPBep( arch='P5', widen_factor=0.25, out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 16, 32, 32)) assert feat[1].shape == torch.Size((1, 32, 16, 16)) assert feat[2].shape == torch.Size((1, 64, 8, 8)) assert feat[3].shape == torch.Size((1, 128, 4, 4)) assert feat[4].shape == torch.Size((1, 256, 2, 2)) # Test YOLOv6CSPBep forward with dict(type='ReLU') model = YOLOv6CSPBep( widen_factor=0.125, act_cfg=dict(type='ReLU'), out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test YOLOv6CSPBep with BatchNorm forward model = YOLOv6CSPBep(widen_factor=0.125, out_indices=range(0, 5)) for m in model.modules(): if is_norm(m): assert isinstance(m, _BatchNorm) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test YOLOv6CSPBep with BatchNorm forward model = YOLOv6CSPBep(plugins=[ dict( cfg=dict(type='mmdet.DropBlock', drop_prob=0.1, block_size=3), stages=(False, False, True, True)), ]) assert len(model.stage1) == 1 assert len(model.stage2) == 1 assert len(model.stage3) == 2 # +DropBlock assert len(model.stage4) == 3 # +SPPF+DropBlock model.train() imgs = torch.randn(1, 3, 256, 256) feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 256, 32, 32)) assert feat[1].shape == torch.Size((1, 512, 16, 16)) assert feat[2].shape == torch.Size((1, 1024, 8, 8))
7,688
36.876847
79
py
mmyolo
mmyolo-main/tests/test_models/test_backbone/utils.py
# Copyright (c) OpenMMLab. All rights reserved. from mmdet.models.backbones.res2net import Bottle2neck from mmdet.models.backbones.resnet import BasicBlock, Bottleneck from mmdet.models.backbones.resnext import Bottleneck as BottleneckX from mmdet.models.layers import SimplifiedBasicBlock from torch.nn.modules import GroupNorm from torch.nn.modules.batchnorm import _BatchNorm def is_block(modules): """Check if is ResNet building block.""" if isinstance(modules, (BasicBlock, Bottleneck, BottleneckX, Bottle2neck, SimplifiedBasicBlock)): return True return False def is_norm(modules): """Check if is one of the norms.""" if isinstance(modules, (GroupNorm, _BatchNorm)): return True return False def check_norm_state(modules, train_state): """Check if norm layer is in correct train state.""" for mod in modules: if isinstance(mod, _BatchNorm): if mod.training != train_state: return False return True
1,026
31.09375
77
py
mmyolo
mmyolo-main/tests/test_models/test_backbone/test_yolov7_backbone.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import pytest import torch from torch.nn.modules.batchnorm import _BatchNorm from mmyolo.models.backbones import YOLOv7Backbone from mmyolo.utils import register_all_modules from .utils import check_norm_state register_all_modules() class TestYOLOv7Backbone(TestCase): def test_init(self): # out_indices in range(len(arch_setting) + 1) with pytest.raises(AssertionError): YOLOv7Backbone(out_indices=(6, )) with pytest.raises(ValueError): # frozen_stages must in range(-1, len(arch_setting) + 1) YOLOv7Backbone(frozen_stages=6) def test_forward(self): # Test YOLOv7Backbone-L with first stage frozen frozen_stages = 1 model = YOLOv7Backbone(frozen_stages=frozen_stages) model.init_weights() model.train() for mod in model.stem.modules(): for param in mod.parameters(): assert param.requires_grad is False for i in range(1, frozen_stages + 1): layer = getattr(model, f'stage{i}') for mod in layer.modules(): if isinstance(mod, _BatchNorm): assert mod.training is False for param in layer.parameters(): assert param.requires_grad is False # Test YOLOv7Backbone-L with norm_eval=True model = YOLOv7Backbone(norm_eval=True) model.train() assert check_norm_state(model.modules(), False) # Test YOLOv7Backbone-L forward with widen_factor=0.25 model = YOLOv7Backbone( widen_factor=0.25, out_indices=tuple(range(0, 5))) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 16, 32, 32)) assert feat[1].shape == torch.Size((1, 64, 16, 16)) assert feat[2].shape == torch.Size((1, 128, 8, 8)) assert feat[3].shape == torch.Size((1, 256, 4, 4)) assert feat[4].shape == torch.Size((1, 256, 2, 2)) # Test YOLOv7Backbone-L with plugins model = YOLOv7Backbone( widen_factor=0.25, plugins=[ dict( cfg=dict( type='mmdet.DropBlock', drop_prob=0.1, block_size=3), stages=(False, False, True, True)), ]) assert len(model.stage1) == 2 assert len(model.stage2) == 2 assert len(model.stage3) == 3 # +DropBlock assert len(model.stage4) == 3 # +DropBlock model.train() imgs = torch.randn(1, 3, 128, 128) feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 128, 16, 16)) assert feat[1].shape == torch.Size((1, 256, 8, 8)) assert feat[2].shape == torch.Size((1, 256, 4, 4)) # Test YOLOv7Backbone-X forward with widen_factor=0.25 model = YOLOv7Backbone(arch='X', widen_factor=0.25) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 160, 8, 8)) assert feat[1].shape == torch.Size((1, 320, 4, 4)) assert feat[2].shape == torch.Size((1, 320, 2, 2)) # Test YOLOv7Backbone-tiny forward with widen_factor=0.25 model = YOLOv7Backbone(arch='Tiny', widen_factor=0.25) model.train() feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 32, 8, 8)) assert feat[1].shape == torch.Size((1, 64, 4, 4)) assert feat[2].shape == torch.Size((1, 128, 2, 2)) # Test YOLOv7Backbone-w forward with widen_factor=0.25 model = YOLOv7Backbone( arch='W', widen_factor=0.25, out_indices=(2, 3, 4, 5)) model.train() imgs = torch.randn(1, 3, 128, 128) feat = model(imgs) assert len(feat) == 4 assert feat[0].shape == torch.Size((1, 64, 16, 16)) assert feat[1].shape == torch.Size((1, 128, 8, 8)) assert feat[2].shape == torch.Size((1, 192, 4, 4)) assert feat[3].shape == torch.Size((1, 256, 2, 2)) # Test YOLOv7Backbone-w forward with widen_factor=0.25 model = YOLOv7Backbone( arch='D', widen_factor=0.25, out_indices=(2, 3, 4, 5)) model.train() feat = model(imgs) assert len(feat) == 4 assert feat[0].shape == torch.Size((1, 96, 16, 16)) assert feat[1].shape == torch.Size((1, 192, 8, 8)) assert feat[2].shape == torch.Size((1, 288, 4, 4)) assert feat[3].shape == torch.Size((1, 384, 2, 2)) # Test YOLOv7Backbone-w forward with widen_factor=0.25 model = YOLOv7Backbone( arch='E', widen_factor=0.25, out_indices=(2, 3, 4, 5)) model.train() feat = model(imgs) assert len(feat) == 4 assert feat[0].shape == torch.Size((1, 80, 16, 16)) assert feat[1].shape == torch.Size((1, 160, 8, 8)) assert feat[2].shape == torch.Size((1, 240, 4, 4)) assert feat[3].shape == torch.Size((1, 320, 2, 2)) # Test YOLOv7Backbone-w forward with widen_factor=0.25 model = YOLOv7Backbone( arch='E2E', widen_factor=0.25, out_indices=(2, 3, 4, 5)) model.train() feat = model(imgs) assert len(feat) == 4 assert feat[0].shape == torch.Size((1, 80, 16, 16)) assert feat[1].shape == torch.Size((1, 160, 8, 8)) assert feat[2].shape == torch.Size((1, 240, 4, 4)) assert feat[3].shape == torch.Size((1, 320, 2, 2))
5,705
35.812903
77
py
mmyolo
mmyolo-main/tests/test_models/test_backbone/test_csp_darknet.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import pytest import torch from parameterized import parameterized from torch.nn.modules.batchnorm import _BatchNorm from mmyolo.models.backbones import (YOLOv5CSPDarknet, YOLOv8CSPDarknet, YOLOXCSPDarknet) from mmyolo.utils import register_all_modules from .utils import check_norm_state, is_norm register_all_modules() class TestCSPDarknet(TestCase): @parameterized.expand([(YOLOv5CSPDarknet, ), (YOLOXCSPDarknet, ), (YOLOv8CSPDarknet, )]) def test_init(self, module_class): # out_indices in range(len(arch_setting) + 1) with pytest.raises(AssertionError): module_class(out_indices=(6, )) with pytest.raises(ValueError): # frozen_stages must in range(-1, len(arch_setting) + 1) module_class(frozen_stages=6) @parameterized.expand([(YOLOv5CSPDarknet, ), (YOLOXCSPDarknet, ), (YOLOv8CSPDarknet, )]) def test_forward(self, module_class): # Test CSPDarknet with first stage frozen frozen_stages = 1 model = module_class(frozen_stages=frozen_stages) model.init_weights() model.train() for mod in model.stem.modules(): for param in mod.parameters(): assert param.requires_grad is False for i in range(1, frozen_stages + 1): layer = getattr(model, f'stage{i}') for mod in layer.modules(): if isinstance(mod, _BatchNorm): assert mod.training is False for param in layer.parameters(): assert param.requires_grad is False # Test CSPDarknet with norm_eval=True model = module_class(norm_eval=True) model.train() assert check_norm_state(model.modules(), False) # Test CSPDarknet-P5 forward with widen_factor=0.25 model = module_class( arch='P5', widen_factor=0.25, out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 16, 32, 32)) assert feat[1].shape == torch.Size((1, 32, 16, 16)) assert feat[2].shape == torch.Size((1, 64, 8, 8)) assert feat[3].shape == torch.Size((1, 128, 4, 4)) assert feat[4].shape == torch.Size((1, 256, 2, 2)) # Test CSPDarknet forward with dict(type='ReLU') model = module_class( widen_factor=0.125, act_cfg=dict(type='ReLU'), out_indices=range(0, 5)) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test CSPDarknet with BatchNorm forward model = module_class(widen_factor=0.125, out_indices=range(0, 5)) for m in model.modules(): if is_norm(m): assert isinstance(m, _BatchNorm) model.train() imgs = torch.randn(1, 3, 64, 64) feat = model(imgs) assert len(feat) == 5 assert feat[0].shape == torch.Size((1, 8, 32, 32)) assert feat[1].shape == torch.Size((1, 16, 16, 16)) assert feat[2].shape == torch.Size((1, 32, 8, 8)) assert feat[3].shape == torch.Size((1, 64, 4, 4)) assert feat[4].shape == torch.Size((1, 128, 2, 2)) # Test CSPDarknet with Dropout Block model = module_class(plugins=[ dict( cfg=dict(type='mmdet.DropBlock', drop_prob=0.1, block_size=3), stages=(False, False, True, True)), ]) assert len(model.stage1) == 2 assert len(model.stage2) == 2 assert len(model.stage3) == 3 # +DropBlock assert len(model.stage4) == 4 # +SPPF+DropBlock model.train() imgs = torch.randn(1, 3, 256, 256) feat = model(imgs) assert len(feat) == 3 assert feat[0].shape == torch.Size((1, 256, 32, 32)) assert feat[1].shape == torch.Size((1, 512, 16, 16)) assert feat[2].shape == torch.Size((1, 1024, 8, 8))
4,481
36.35
78
py
mmyolo
mmyolo-main/tests/test_models/test_data_preprocessor/test_data_preprocessor.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmdet.structures import DetDataSample from mmengine import MessageHub from mmyolo.models import PPYOLOEBatchRandomResize, PPYOLOEDetDataPreprocessor from mmyolo.models.data_preprocessors import (YOLOv5DetDataPreprocessor, YOLOXBatchSyncRandomResize) from mmyolo.utils import register_all_modules register_all_modules() class TestYOLOv5DetDataPreprocessor(TestCase): def test_forward(self): processor = YOLOv5DetDataPreprocessor(mean=[0, 0, 0], std=[1, 1, 1]) data = { 'inputs': [torch.randint(0, 256, (3, 11, 10))], 'data_samples': [DetDataSample()] } out_data = processor(data, training=False) batch_inputs, batch_data_samples = out_data['inputs'], out_data[ 'data_samples'] self.assertEqual(batch_inputs.shape, (1, 3, 11, 10)) self.assertEqual(len(batch_data_samples), 1) # test channel_conversion processor = YOLOv5DetDataPreprocessor( mean=[0., 0., 0.], std=[1., 1., 1.], bgr_to_rgb=True) out_data = processor(data, training=False) batch_inputs, batch_data_samples = out_data['inputs'], out_data[ 'data_samples'] self.assertEqual(batch_inputs.shape, (1, 3, 11, 10)) self.assertEqual(len(batch_data_samples), 1) # test padding, training=False data = { 'inputs': [ torch.randint(0, 256, (3, 10, 11)), torch.randint(0, 256, (3, 9, 14)) ] } processor = YOLOv5DetDataPreprocessor( mean=[0., 0., 0.], std=[1., 1., 1.], bgr_to_rgb=True) out_data = processor(data, training=False) batch_inputs, batch_data_samples = out_data['inputs'], out_data[ 'data_samples'] self.assertEqual(batch_inputs.shape, (2, 3, 10, 14)) self.assertIsNone(batch_data_samples) # test training data = { 'inputs': torch.randint(0, 256, (2, 3, 10, 11)), 'data_samples': { 'bboxes_labels': torch.randint(0, 11, (18, 6)) }, } out_data = processor(data, training=True) batch_inputs, batch_data_samples = out_data['inputs'], out_data[ 'data_samples'] self.assertIn('img_metas', batch_data_samples) self.assertIn('bboxes_labels', batch_data_samples) self.assertEqual(batch_inputs.shape, (2, 3, 10, 11)) self.assertIsInstance(batch_data_samples['bboxes_labels'], torch.Tensor) self.assertIsInstance(batch_data_samples['img_metas'], list) data = { 'inputs': [torch.randint(0, 256, (3, 11, 10))], 'data_samples': [DetDataSample()] } # data_samples must be dict with self.assertRaises(AssertionError): processor(data, training=True) class TestPPYOLOEDetDataPreprocessor(TestCase): def test_batch_random_resize(self): processor = PPYOLOEDetDataPreprocessor( pad_size_divisor=32, batch_augments=[ dict( type='PPYOLOEBatchRandomResize', random_size_range=(320, 480), interval=1, size_divisor=32, random_interp=True, keep_ratio=False) ], mean=[0., 0., 0.], std=[255., 255., 255.], bgr_to_rgb=True) self.assertTrue( isinstance(processor.batch_augments[0], PPYOLOEBatchRandomResize)) message_hub = MessageHub.get_instance('test_batch_random_resize') message_hub.update_info('iter', 0) # test training data = { 'inputs': [ torch.randint(0, 256, (3, 10, 11)), torch.randint(0, 256, (3, 10, 11)) ], 'data_samples': { 'bboxes_labels': torch.randint(0, 11, (18, 6)).float() }, } out_data = processor(data, training=True) batch_data_samples = out_data['data_samples'] self.assertIn('img_metas', batch_data_samples) self.assertIn('bboxes_labels', batch_data_samples) self.assertIsInstance(batch_data_samples['bboxes_labels'], torch.Tensor) self.assertIsInstance(batch_data_samples['img_metas'], list) data = { 'inputs': [torch.randint(0, 256, (3, 11, 10))], 'data_samples': DetDataSample() } # data_samples must be list with self.assertRaises(AssertionError): processor(data, training=True) class TestYOLOXDetDataPreprocessor(TestCase): def test_batch_sync_random_size(self): processor = YOLOXBatchSyncRandomResize( random_size_range=(480, 800), size_divisor=32, interval=1) self.assertTrue(isinstance(processor, YOLOXBatchSyncRandomResize)) message_hub = MessageHub.get_instance( 'test_yolox_batch_sync_random_resize') message_hub.update_info('iter', 0) # test training inputs = torch.randint(0, 256, (4, 3, 10, 11)) data_samples = {'bboxes_labels': torch.randint(0, 11, (18, 6)).float()} inputs, data_samples = processor(inputs, data_samples) self.assertIn('bboxes_labels', data_samples) self.assertIsInstance(data_samples['bboxes_labels'], torch.Tensor) self.assertIsInstance(inputs, torch.Tensor) inputs = torch.randint(0, 256, (4, 3, 10, 11)) data_samples = DetDataSample() # data_samples must be dict with self.assertRaises(AssertionError): processor(inputs, data_samples)
5,829
36.133758
79
py
mmyolo
mmyolo-main/tests/test_models/test_utils/test_misc.py
# Copyright (c) OpenMMLab. All rights reserved. import numpy as np import pytest import torch from mmengine.structures import InstanceData from torch import Tensor from mmyolo.models.utils import gt_instances_preprocess from mmyolo.utils import register_all_modules register_all_modules() class TestGtInstancesPreprocess: @pytest.mark.parametrize('box_dim', [4, 5]) def test(self, box_dim): gt_instances = InstanceData( bboxes=torch.empty((0, box_dim)), labels=torch.LongTensor([])) batch_size = 1 batch_instance = gt_instances_preprocess([gt_instances], batch_size) assert isinstance(batch_instance, Tensor) assert len(batch_instance.shape) == 3, 'the len of result must be 3.' assert batch_instance.size(-1) == box_dim + 1 @pytest.mark.parametrize('box_dim', [4, 5]) def test_fast_version(self, box_dim: int): gt_instances = torch.from_numpy( np.array([[0., 1., *(0., ) * box_dim]], dtype=np.float32)) batch_size = 1 batch_instance = gt_instances_preprocess(gt_instances, batch_size) assert isinstance(batch_instance, Tensor) assert len(batch_instance.shape) == 3, 'the len of result must be 3.' assert batch_instance.shape[1] == 1 assert batch_instance.shape[2] == box_dim + 1
1,330
35.972222
77
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_ppyoloe_csppan.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models import PPYOLOECSPPAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestPPYOLOECSPPAFPN(TestCase): def test_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = PPYOLOECSPPAFPN( in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i) def test_drop_block(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = PPYOLOECSPPAFPN( in_channels=in_channels, out_channels=out_channels, drop_block_cfg=dict( type='mmdet.DropBlock', drop_prob=0.1, block_size=3, warm_iters=0)) neck.train() outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
1,741
31.259259
71
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_cspnext_pafpn.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models.necks import CSPNeXtPAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestCSPNeXtPAFPN(TestCase): def test_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = 24 feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = CSPNeXtPAFPN(in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i) # test depth-wise neck = CSPNeXtPAFPN( in_channels=in_channels, out_channels=out_channels, use_depthwise=True) from mmcv.cnn.bricks import DepthwiseSeparableConvModule self.assertTrue(neck.conv, DepthwiseSeparableConvModule)
1,159
29.526316
79
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_yolov8_pafpn.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models import YOLOv8PAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestYOLOv8PAFPN(TestCase): def test_YOLOv8PAFPN_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOv8PAFPN(in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
879
29.344828
78
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_yolox_pafpn.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models.necks import YOLOXPAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestYOLOXPAFPN(TestCase): def test_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = 24 feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOXPAFPN(in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
858
28.62069
77
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_yolov7_pafpn.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmcv.cnn import ConvModule from mmyolo.models.necks import YOLOv7PAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestYOLOv7PAFPN(TestCase): def test_forward(self): # test P5 s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOv7PAFPN(in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] * 2 assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i) # test is_tiny_version neck = YOLOv7PAFPN( in_channels=in_channels, out_channels=out_channels, is_tiny_version=True) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] * 2 assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i) # test use_in_channels_in_downsample neck = YOLOv7PAFPN( in_channels=in_channels, out_channels=out_channels, use_in_channels_in_downsample=True) for f in feats: print(f.shape) outs = neck(feats) for f in outs: print(f.shape) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] * 2 assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i) # test use_repconv_outs is False neck = YOLOv7PAFPN( in_channels=in_channels, out_channels=out_channels, use_repconv_outs=False) self.assertIsInstance(neck.out_layers[0], ConvModule) # test P6 s = 64 in_channels = [8, 16, 32, 64] feat_sizes = [s // 2**i for i in range(4)] out_channels = [8, 16, 32, 64] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOv7PAFPN(in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
2,718
32.9875
78
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_yolov6_pafpn.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models.necks import YOLOv6CSPRepPAFPN, YOLOv6RepPAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestYOLOv6PAFPN(TestCase): def test_YOLOv6RepPAFP_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOv6RepPAFPN( in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i) def test_YOLOv6CSPRepPAFPN_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOv6CSPRepPAFPN( in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
1,593
32.914894
71
py
mmyolo
mmyolo-main/tests/test_models/test_necks/test_yolov5_pafpn.py
# Copyright (c) OpenMMLab. All rights reserved. from unittest import TestCase import torch from mmyolo.models.necks import YOLOv5PAFPN from mmyolo.utils import register_all_modules register_all_modules() class TestYOLOv5PAFPN(TestCase): def test_forward(self): s = 64 in_channels = [8, 16, 32] feat_sizes = [s // 2**i for i in range(4)] # [32, 16, 8] out_channels = [8, 16, 32] feats = [ torch.rand(1, in_channels[i], feat_sizes[i], feat_sizes[i]) for i in range(len(in_channels)) ] neck = YOLOv5PAFPN(in_channels=in_channels, out_channels=out_channels) outs = neck(feats) assert len(outs) == len(feats) for i in range(len(feats)): assert outs[i].shape[1] == out_channels[i] assert outs[i].shape[2] == outs[i].shape[3] == s // (2**i)
873
29.137931
78
py