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
NBFNet
NBFNet-master/nbfnet/model.py
from collections.abc import Sequence import torch from torch import nn from torch import autograd from torch_scatter import scatter_add from torchdrug import core, layers, utils from torchdrug.layers import functional from torchdrug.core import Registry as R from . import layer @R.register("model.NBFNet") class NeuralBellmanFordNetwork(nn.Module, core.Configurable): def __init__(self, input_dim, hidden_dims, num_relation=None, symmetric=False, message_func="distmult", aggregate_func="pna", short_cut=False, layer_norm=False, activation="relu", concat_hidden=False, num_mlp_layer=2, dependent=True, remove_one_hop=False, num_beam=10, path_topk=10): super(NeuralBellmanFordNetwork, self).__init__() if not isinstance(hidden_dims, Sequence): hidden_dims = [hidden_dims] if num_relation is None: double_relation = 1 else: num_relation = int(num_relation) double_relation = num_relation * 2 self.dims = [input_dim] + list(hidden_dims) self.num_relation = num_relation self.symmetric = symmetric self.short_cut = short_cut self.concat_hidden = concat_hidden self.remove_one_hop = remove_one_hop self.num_beam = num_beam self.path_topk = path_topk self.layers = nn.ModuleList() for i in range(len(self.dims) - 1): self.layers.append(layer.GeneralizedRelationalConv(self.dims[i], self.dims[i + 1], double_relation, self.dims[0], message_func, aggregate_func, layer_norm, activation, dependent)) feature_dim = hidden_dims[-1] * (len(hidden_dims) if concat_hidden else 1) + input_dim self.query = nn.Embedding(double_relation, input_dim) self.mlp = layers.MLP(feature_dim, [feature_dim] * (num_mlp_layer - 1) + [1]) def remove_easy_edges(self, graph, h_index, t_index, r_index=None): if self.remove_one_hop: h_index_ext = torch.cat([h_index, t_index], dim=-1) t_index_ext = torch.cat([t_index, h_index], dim=-1) if r_index is not None: any = -torch.ones_like(h_index_ext) pattern = torch.stack([h_index_ext, t_index_ext, any], dim=-1) else: pattern = torch.stack([h_index_ext, t_index_ext], dim=-1) else: if r_index is not None: pattern = torch.stack([h_index, t_index, r_index], dim=-1) else: pattern = torch.stack([h_index, t_index], dim=-1) pattern = pattern.flatten(0, -2) edge_index = graph.match(pattern)[0] edge_mask = ~functional.as_mask(edge_index, graph.num_edge) return graph.edge_mask(edge_mask) def negative_sample_to_tail(self, h_index, t_index, r_index): # convert p(h | t, r) to p(t' | h', r') # h' = t, r' = r^{-1}, t' = h is_t_neg = (h_index == h_index[:, [0]]).all(dim=-1, keepdim=True) new_h_index = torch.where(is_t_neg, h_index, t_index) new_t_index = torch.where(is_t_neg, t_index, h_index) new_r_index = torch.where(is_t_neg, r_index, r_index + self.num_relation) return new_h_index, new_t_index, new_r_index def as_relational_graph(self, graph, self_loop=True): # add self loop # convert homogeneous graphs to knowledge graphs with 1 relation edge_list = graph.edge_list edge_weight = graph.edge_weight if self_loop: node_in = node_out = torch.arange(graph.num_node, device=self.device) loop = torch.stack([node_in, node_out], dim=-1) edge_list = torch.cat([edge_list, loop]) edge_weight = torch.cat([edge_weight, torch.ones(graph.num_node, device=self.device)]) relation = torch.zeros(len(edge_list), 1, dtype=torch.long, device=self.device) edge_list = torch.cat([edge_list, relation], dim=-1) graph = type(graph)(edge_list, edge_weight=edge_weight, num_node=graph.num_node, num_relation=1, meta_dict=graph.meta_dict, **graph.data_dict) return graph @utils.cached def bellmanford(self, graph, h_index, r_index, separate_grad=False): query = self.query(r_index) index = h_index.unsqueeze(-1).expand_as(query) boundary = torch.zeros(graph.num_node, *query.shape, device=self.device) boundary.scatter_add_(0, index.unsqueeze(0), query.unsqueeze(0)) with graph.graph(): graph.query = query with graph.node(): graph.boundary = boundary hiddens = [] step_graphs = [] layer_input = boundary for layer in self.layers: if separate_grad: step_graph = graph.clone().requires_grad_() else: step_graph = graph hidden = layer(step_graph, layer_input) if self.short_cut and hidden.shape == layer_input.shape: hidden = hidden + layer_input hiddens.append(hidden) step_graphs.append(step_graph) layer_input = hidden node_query = query.expand(graph.num_node, -1, -1) if self.concat_hidden: output = torch.cat(hiddens + [node_query], dim=-1) else: output = torch.cat([hiddens[-1], node_query], dim=-1) return { "node_feature": output, "step_graphs": step_graphs, } def forward(self, graph, h_index, t_index, r_index=None, all_loss=None, metric=None): if all_loss is not None: graph = self.remove_easy_edges(graph, h_index, t_index, r_index) shape = h_index.shape if graph.num_relation: graph = graph.undirected(add_inverse=True) h_index, t_index, r_index = self.negative_sample_to_tail(h_index, t_index, r_index) else: graph = self.as_relational_graph(graph) h_index = h_index.view(-1, 1) t_index = t_index.view(-1, 1) r_index = torch.zeros_like(h_index) assert (h_index[:, [0]] == h_index).all() assert (r_index[:, [0]] == r_index).all() output = self.bellmanford(graph, h_index[:, 0], r_index[:, 0]) feature = output["node_feature"].transpose(0, 1) index = t_index.unsqueeze(-1).expand(-1, -1, feature.shape[-1]) feature = feature.gather(1, index) if self.symmetric: assert (t_index[:, [0]] == t_index).all() output = self.bellmanford(graph, t_index[:, 0], r_index[:, 0]) inv_feature = output["node_feature"].transpose(0, 1) index = h_index.unsqueeze(-1).expand(-1, -1, inv_feature.shape[-1]) inv_feature = inv_feature.gather(1, index) feature = (feature + inv_feature) / 2 score = self.mlp(feature).squeeze(-1) return score.view(shape) def visualize(self, graph, h_index, t_index, r_index): assert h_index.numel() == 1 and h_index.ndim == 1 graph = graph.undirected(add_inverse=True) output = self.bellmanford(graph, h_index, r_index, separate_grad=True) feature = output["node_feature"] step_graphs = output["step_graphs"] index = t_index.unsqueeze(0).unsqueeze(-1).expand(-1, -1, feature.shape[-1]) feature = feature.gather(0, index).squeeze(0) score = self.mlp(feature).squeeze(-1) edge_weights = [graph.edge_weight for graph in step_graphs] edge_grads = autograd.grad(score, edge_weights) for graph, edge_grad in zip(step_graphs, edge_grads): with graph.edge(): graph.edge_grad = edge_grad distances, back_edges = self.beam_search_distance(step_graphs, h_index, t_index, self.num_beam) paths, weights = self.topk_average_length(distances, back_edges, t_index, self.path_topk) return paths, weights @torch.no_grad() def beam_search_distance(self, graphs, h_index, t_index, num_beam=10): num_node = graphs[0].num_node input = torch.full((num_node, num_beam), float("-inf"), device=self.device) input[h_index, 0] = 0 distances = [] back_edges = [] for graph in graphs: graph = graph.edge_mask(graph.edge_list[:, 0] != t_index) node_in, node_out = graph.edge_list.t()[:2] message = input[node_in] + graph.edge_grad.unsqueeze(-1) msg_source = graph.edge_list.unsqueeze(1).expand(-1, num_beam, -1) is_duplicate = torch.isclose(message.unsqueeze(-1), message.unsqueeze(-2)) & \ (msg_source.unsqueeze(-2) == msg_source.unsqueeze(-3)).all(dim=-1) is_duplicate = is_duplicate.float() - \ torch.arange(num_beam, dtype=torch.float, device=self.device) / (num_beam + 1) # pick the first occurrence as the previous state prev_rank = is_duplicate.argmax(dim=-1, keepdim=True) msg_source = torch.cat([msg_source, prev_rank], dim=-1) node_out, order = node_out.sort() node_out_set = torch.unique(node_out) # sort message w.r.t. node_out message = message[order].flatten() msg_source = msg_source[order].flatten(0, -2) size = scatter_add(torch.ones_like(node_out), node_out, dim_size=num_node) msg2out = functional._size_to_index(size[node_out_set] * num_beam) # deduplicate is_duplicate = (msg_source[1:] == msg_source[:-1]).all(dim=-1) is_duplicate = torch.cat([torch.zeros(1, dtype=torch.bool, device=self.device), is_duplicate]) message = message[~is_duplicate] msg_source = msg_source[~is_duplicate] msg2out = msg2out[~is_duplicate] size = scatter_add(torch.ones_like(msg2out), msg2out, dim_size=len(node_out_set)) if not torch.isinf(message).all(): distance, rel_index = functional.variadic_topk(message, size, k=num_beam) abs_index = rel_index + (size.cumsum(0) - size).unsqueeze(-1) back_edge = msg_source[abs_index] distance = distance.view(len(node_out_set), num_beam) back_edge = back_edge.view(len(node_out_set), num_beam, 4) distance = scatter_add(distance, node_out_set, dim=0, dim_size=num_node) back_edge = scatter_add(back_edge, node_out_set, dim=0, dim_size=num_node) else: distance = torch.full((num_node, num_beam), float("-inf"), device=self.device) back_edge = torch.zeros(num_node, num_beam, 4, dtype=torch.long, device=self.device) distances.append(distance) back_edges.append(back_edge) input = distance return distances, back_edges def topk_average_length(self, distances, back_edges, t_index, k=10): paths = [] average_lengths = [] for i in range(len(distances)): distance, order = distances[i][t_index].flatten(0, -1).sort(descending=True) back_edge = back_edges[i][t_index].flatten(0, -2)[order] for d, (h, t, r, prev_rank) in zip(distance[:k].tolist(), back_edge[:k].tolist()): if d == float("-inf"): break path = [(h, t, r)] for j in range(i - 1, -1, -1): h, t, r, prev_rank = back_edges[j][h, prev_rank].tolist() path.append((h, t, r)) paths.append(path[::-1]) average_lengths.append(d / len(path)) if paths: average_lengths, paths = zip(*sorted(zip(average_lengths, paths), reverse=True)[:k]) return paths, average_lengths
11,950
44.441065
118
py
NBFNet
NBFNet-master/nbfnet/dataset.py
import os import csv import glob from tqdm import tqdm from ogb import linkproppred import torch from torch.utils import data as torch_data from torchdrug import data, datasets, utils from torchdrug.core import Registry as R class InductiveKnowledgeGraphDataset(data.KnowledgeGraphDataset): def load_inductive_tsvs(self, train_files, test_files, verbose=0): assert len(train_files) == len(test_files) == 2 inv_train_entity_vocab = {} inv_test_entity_vocab = {} inv_relation_vocab = {} triplets = [] num_samples = [] for txt_file in train_files: with open(txt_file, "r") as fin: reader = csv.reader(fin, delimiter="\t") if verbose: reader = tqdm(reader, "Loading %s" % txt_file, utils.get_line_count(txt_file)) num_sample = 0 for tokens in reader: h_token, r_token, t_token = tokens if h_token not in inv_train_entity_vocab: inv_train_entity_vocab[h_token] = len(inv_train_entity_vocab) h = inv_train_entity_vocab[h_token] if r_token not in inv_relation_vocab: inv_relation_vocab[r_token] = len(inv_relation_vocab) r = inv_relation_vocab[r_token] if t_token not in inv_train_entity_vocab: inv_train_entity_vocab[t_token] = len(inv_train_entity_vocab) t = inv_train_entity_vocab[t_token] triplets.append((h, t, r)) num_sample += 1 num_samples.append(num_sample) for txt_file in test_files: with open(txt_file, "r") as fin: reader = csv.reader(fin, delimiter="\t") if verbose: reader = tqdm(reader, "Loading %s" % txt_file, utils.get_line_count(txt_file)) num_sample = 0 for tokens in reader: h_token, r_token, t_token = tokens if h_token not in inv_test_entity_vocab: inv_test_entity_vocab[h_token] = len(inv_test_entity_vocab) h = inv_test_entity_vocab[h_token] assert r_token in inv_relation_vocab r = inv_relation_vocab[r_token] if t_token not in inv_test_entity_vocab: inv_test_entity_vocab[t_token] = len(inv_test_entity_vocab) t = inv_test_entity_vocab[t_token] triplets.append((h, t, r)) num_sample += 1 num_samples.append(num_sample) train_entity_vocab, inv_train_entity_vocab = self._standarize_vocab(None, inv_train_entity_vocab) test_entity_vocab, inv_test_entity_vocab = self._standarize_vocab(None, inv_test_entity_vocab) relation_vocab, inv_relation_vocab = self._standarize_vocab(None, inv_relation_vocab) self.train_graph = data.Graph(triplets[:num_samples[0]], num_node=len(train_entity_vocab), num_relation=len(relation_vocab)) self.valid_graph = self.train_graph self.test_graph = data.Graph(triplets[sum(num_samples[:2]): sum(num_samples[:3])], num_node=len(test_entity_vocab), num_relation=len(relation_vocab)) self.graph = self.train_graph self.triplets = torch.tensor(triplets[:sum(num_samples[:2])] + triplets[sum(num_samples[:3]):]) self.num_samples = num_samples[:2] + num_samples[3:] self.train_entity_vocab = train_entity_vocab self.test_entity_vocab = test_entity_vocab self.relation_vocab = relation_vocab self.inv_train_entity_vocab = inv_train_entity_vocab self.inv_test_entity_vocab = inv_test_entity_vocab self.inv_relation_vocab = inv_relation_vocab def __getitem__(self, index): return self.triplets[index] def split(self): offset = 0 splits = [] for num_sample in self.num_samples: split = torch_data.Subset(self, range(offset, offset + num_sample)) splits.append(split) offset += num_sample return splits @R.register("datasets.CoraLinkPrediction") class CoraLinkPrediction(datasets.Cora): def __init__(self, **kwargs): super(CoraLinkPrediction, self).__init__(**kwargs) self.transform = None def __getitem__(self, index): return self.graph.edge_list[index] def __len__(self): return self.graph.num_edge def split(self, ratios=(85, 5, 10)): length = self.graph.num_edge norm = sum(ratios) lengths = [int(r / norm * length) for r in ratios] lengths[-1] = length - sum(lengths[:-1]) g = torch.Generator() g.manual_seed(0) return torch_data.random_split(self, lengths, generator=g) @R.register("datasets.CiteSeerLinkPrediction") class CiteSeerLinkPrediction(datasets.CiteSeer): def __init__(self, **kwargs): super(CiteSeerLinkPrediction, self).__init__(**kwargs) self.transform = None def __getitem__(self, index): return self.graph.edge_list[index] def __len__(self): return self.graph.num_edge def split(self, ratios=(85, 5, 10)): length = self.graph.num_edge norm = sum(ratios) lengths = [int(r / norm * length) for r in ratios] lengths[-1] = length - sum(lengths[:-1]) g = torch.Generator() g.manual_seed(0) return torch_data.random_split(self, lengths, generator=g) @R.register("datasets.PubMedLinkPrediction") class PubMedLinkPrediction(datasets.PubMed): def __init__(self, **kwargs): super(PubMedLinkPrediction, self).__init__(**kwargs) self.transform = None def __getitem__(self, index): return self.graph.edge_list[index] def __len__(self): return self.graph.num_edge def split(self, ratios=(85, 5, 10)): length = self.graph.num_edge norm = sum(ratios) lengths = [int(r / norm * length) for r in ratios] lengths[-1] = length - sum(lengths[:-1]) g = torch.Generator() g.manual_seed(0) return torch_data.random_split(self, lengths, generator=g) @R.register("datasets.FB15k237Inductive") class FB15k237Inductive(InductiveKnowledgeGraphDataset): train_urls = [ "https://raw.githubusercontent.com/kkteru/grail/master/data/fb237_%s/train.txt", "https://raw.githubusercontent.com/kkteru/grail/master/data/fb237_%s/valid.txt", ] test_urls = [ "https://raw.githubusercontent.com/kkteru/grail/master/data/fb237_%s_ind/train.txt", "https://raw.githubusercontent.com/kkteru/grail/master/data/fb237_%s_ind/test.txt", ] def __init__(self, path, version="v1", verbose=1): path = os.path.expanduser(path) if not os.path.exists(path): os.makedirs(path) self.path = path train_files = [] for url in self.train_urls: url = url % version save_file = "fb15k237_%s_%s" % (version, os.path.basename(url)) txt_file = os.path.join(path, save_file) if not os.path.exists(txt_file): txt_file = utils.download(url, self.path, save_file=save_file) train_files.append(txt_file) test_files = [] for url in self.test_urls: url = url % version save_file = "fb15k237_%s_ind_%s" % (version, os.path.basename(url)) txt_file = os.path.join(path, save_file) if not os.path.exists(txt_file): txt_file = utils.download(url, self.path, save_file=save_file) test_files.append(txt_file) self.load_inductive_tsvs(train_files, test_files, verbose=verbose) @R.register("datasets.WN18RRInductive") class WN18RRInductive(InductiveKnowledgeGraphDataset): train_urls = [ "https://raw.githubusercontent.com/kkteru/grail/master/data/WN18RR_%s/train.txt", "https://raw.githubusercontent.com/kkteru/grail/master/data/WN18RR_%s/valid.txt", ] test_urls = [ "https://raw.githubusercontent.com/kkteru/grail/master/data/WN18RR_%s_ind/train.txt", "https://raw.githubusercontent.com/kkteru/grail/master/data/WN18RR_%s_ind/test.txt", ] def __init__(self, path, version="v1", verbose=1): path = os.path.expanduser(path) if not os.path.exists(path): os.makedirs(path) self.path = path train_files = [] for url in self.train_urls: url = url % version save_file = "wn18rr_%s_%s" % (version, os.path.basename(url)) txt_file = os.path.join(path, save_file) if not os.path.exists(txt_file): txt_file = utils.download(url, self.path, save_file=save_file) train_files.append(txt_file) test_files = [] for url in self.test_urls: url = url % version save_file = "wn18rr_%s_ind_%s" % (version, os.path.basename(url)) txt_file = os.path.join(path, save_file) if not os.path.exists(txt_file): txt_file = utils.download(url, self.path, save_file=save_file) test_files.append(txt_file) self.load_inductive_tsvs(train_files, test_files, verbose=verbose) @R.register("datasets.OGBLBioKG") class OGBLBioKG(data.KnowledgeGraphDataset): def __init__(self, path, verbose=1): path = os.path.expanduser(path) self.path = path dataset = linkproppred.LinkPropPredDataset("ogbl-biokg", path) self.load_ogb(dataset, verbose=verbose) def load_ogb(self, dataset, verbose=1): entity_vocab = [] relation_vocab = [] entity_type_vocab = [] inv_entity_type_offset = {} entity_type2num = [] zip_files = glob.glob(os.path.join(dataset.root, "mapping/*.gz")) for zip_file in zip_files: csv_file = utils.extract(zip_file) type = os.path.basename(csv_file).split("_")[0] with open(csv_file, "r") as fin: reader = csv.reader(fin) if verbose: reader = iter(tqdm(reader, "Loading %s" % csv_file, utils.get_line_count(csv_file))) fields = next(reader) if "relidx" in csv_file: for index, token in reader: relation_vocab.append(token) else: entity_type_vocab.append(type) inv_entity_type_offset[type] = len(entity_vocab) num_entity = 0 for index, token in reader: entity_vocab.append("%s (%s)" % (type, token)) num_entity += 1 entity_type2num.append(num_entity) edge_split = dataset.get_edge_split() triplets = [] num_samples = [] num_samples_with_neg = [] negative_heads = [] negative_tails = [] for key in ["train", "valid", "test"]: split_dict = edge_split[key] h = torch.as_tensor(split_dict["head"]) t = torch.as_tensor(split_dict["tail"]) r = torch.as_tensor(split_dict["relation"]) h_type = torch.tensor([inv_entity_type_offset[h] for h in split_dict["head_type"]]) t_type = torch.tensor([inv_entity_type_offset[t] for t in split_dict["tail_type"]]) h = h + h_type t = t + t_type triplet = torch.stack([h, t, r], dim=-1) triplets.append(triplet) num_samples.append(len(h)) if "head_neg" in split_dict: neg_h = torch.as_tensor(split_dict["head_neg"]) neg_t = torch.as_tensor(split_dict["tail_neg"]) neg_h = neg_h + h_type.unsqueeze(-1) neg_t = neg_t + t_type.unsqueeze(-1) negative_heads.append(neg_h) negative_tails.append(neg_t) num_samples_with_neg.append(len(h)) else: num_samples_with_neg.append(0) triplets = torch.cat(triplets) self.load_triplet(triplets, entity_vocab=entity_vocab, relation_vocab=relation_vocab) entity_type_vocab, inv_entity_type_vocab = self._standarize_vocab(entity_type_vocab, None) self.entity_type_vocab = entity_type_vocab self.inv_entity_type_vocab = inv_entity_type_vocab self.num_samples = num_samples self.num_samples_with_neg = num_samples_with_neg self.negative_heads = torch.cat(negative_heads) self.negative_tails = torch.cat(negative_tails) node_type = [] for i, num_entity in enumerate(entity_type2num): node_type += [i] * num_entity with self.graph.node(): self.graph.node_type = torch.tensor(node_type) def split(self, test_negative=True): offset = 0 neg_offset = 0 splits = [] for num_sample, num_sample_with_neg in zip(self.num_samples, self.num_samples_with_neg): if test_negative and num_sample_with_neg: pos_h, pos_t, pos_r = self[offset: offset + num_sample].t() neg_h = self.negative_heads[neg_offset: neg_offset + num_sample_with_neg] neg_t = self.negative_tails[neg_offset: neg_offset + num_sample_with_neg] num_negative = neg_h.shape[-1] h = pos_h.unsqueeze(-1).repeat(2, num_negative + 1) t = pos_t.unsqueeze(-1).repeat(2, num_negative + 1) r = pos_r.unsqueeze(-1).repeat(2, num_negative + 1) t[:num_sample_with_neg, 1:] = neg_t h[num_sample_with_neg:, 1:] = neg_h split = torch.stack([h, t, r], dim=-1) else: split = torch_data.Subset(self, range(offset, offset + num_sample)) splits.append(split) offset += num_sample neg_offset += num_sample_with_neg return splits
14,201
39.005634
105
py
NBFNet
NBFNet-master/nbfnet/util.py
import os import time import logging import argparse import yaml import jinja2 from jinja2 import meta import easydict import torch from torch.utils import data as torch_data from torch import distributed as dist from torchdrug import core, utils from torchdrug.utils import comm logger = logging.getLogger(__file__) def get_root_logger(file=True): logger = logging.getLogger("") logger.setLevel(logging.INFO) format = logging.Formatter("%(asctime)-10s %(message)s", "%H:%M:%S") if file: handler = logging.FileHandler("log.txt") handler.setFormatter(format) logger.addHandler(handler) return logger def create_working_directory(cfg): file_name = "working_dir.tmp" world_size = comm.get_world_size() if world_size > 1 and not dist.is_initialized(): comm.init_process_group("nccl", init_method="env://") working_dir = os.path.join(os.path.expanduser(cfg.output_dir), cfg.task["class"], cfg.dataset["class"], cfg.task.model["class"], time.strftime("%Y-%m-%d-%H-%M-%S")) # synchronize working directory if comm.get_rank() == 0: with open(file_name, "w") as fout: fout.write(working_dir) os.makedirs(working_dir) comm.synchronize() if comm.get_rank() != 0: with open(file_name, "r") as fin: working_dir = fin.read() comm.synchronize() if comm.get_rank() == 0: os.remove(file_name) os.chdir(working_dir) return working_dir def detect_variables(cfg_file): with open(cfg_file, "r") as fin: raw = fin.read() env = jinja2.Environment() ast = env.parse(raw) vars = meta.find_undeclared_variables(ast) return vars def load_config(cfg_file, context=None): with open(cfg_file, "r") as fin: raw = fin.read() template = jinja2.Template(raw) instance = template.render(context) cfg = yaml.safe_load(instance) cfg = easydict.EasyDict(cfg) return cfg def parse_args(): parser = argparse.ArgumentParser() parser.add_argument("-c", "--config", help="yaml configuration file", required=True) parser.add_argument("-s", "--seed", help="random seed for PyTorch", type=int, default=1024) args, unparsed = parser.parse_known_args() # get dynamic arguments defined in the config file vars = detect_variables(args.config) parser = argparse.ArgumentParser() for var in vars: parser.add_argument("--%s" % var, required=True) vars = parser.parse_known_args(unparsed)[0] vars = {k: utils.literal_eval(v) for k, v in vars._get_kwargs()} return args, vars def build_solver(cfg, dataset): train_set, valid_set, test_set = dataset.split() if comm.get_rank() == 0: logger.warning(dataset) logger.warning("#train: %d, #valid: %d, #test: %d" % (len(train_set), len(valid_set), len(test_set))) if "fast_test" in cfg: if comm.get_rank() == 0: logger.warning("Quick test mode on. Only evaluate on %d samples for valid / test." % cfg.fast_test) g = torch.Generator() g.manual_seed(1024) valid_set = torch_data.random_split(valid_set, [cfg.fast_test, len(valid_set) - cfg.fast_test], generator=g)[0] test_set = torch_data.random_split(test_set, [cfg.fast_test, len(test_set) - cfg.fast_test], generator=g)[0] if hasattr(dataset, "num_relation"): cfg.task.model.num_relation = dataset.num_relation task = core.Configurable.load_config_dict(cfg.task) cfg.optimizer.params = task.parameters() optimizer = core.Configurable.load_config_dict(cfg.optimizer) solver = core.Engine(task, train_set, valid_set, test_set, optimizer, **cfg.engine) if "checkpoint" in cfg: solver.load(cfg.checkpoint) return solver
3,838
30.467213
119
py
gca-rom
gca-rom-main/main.py
import sys sys.path.append('../') import torch from gca_rom import network, pde, loader, plotting, preprocessing, training, initialization, testing, error import numpy as np if __name__ == "__main__": problem_name, variable, mu1_range, mu2_range = pde.problem(int(sys.argv[1])) print("PROBLEM: ", problem_name, "for variable ", variable, "\n") AE_Params = network.AE_Params device = initialization.set_device() initialization.set_reproducibility(AE_Params) initialization.set_path(AE_Params) dataset_dir = '../dataset/'+problem_name+'_unstructured.mat' dataset = loader.LoadDataset(dataset_dir, variable) dataset_graph, graph_loader, train_loader, test_loader, \ val_loader, scaler_all, scaler_test, xx, yy, var, VAR_all, VAR_test, \ train_trajectories, test_trajectories = preprocessing.graphs_dataset(dataset, AE_Params) mu1, mu2 = np.meshgrid(mu1_range, mu2_range) params = torch.tensor(np.vstack((mu1.T, mu2.T)).reshape(2, -1).T) params = params.to(device) print('Shape of parameter space:', params.shape, '\n') model = network.Net() model = model.to(device) optimizer = torch.optim.Adam(model.parameters(), lr=AE_Params.learning_rate, weight_decay=AE_Params.weight_decay) scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=AE_Params.miles, gamma=AE_Params.gamma) history = dict(train=[], l1=[], l2=[]) history_test = dict(test=[], l1=[], l2=[]) min_test_loss = np.Inf try: model.load_state_dict(torch.load(AE_Params.net_dir+AE_Params.net_name+AE_Params.net_run+'.pt', map_location=torch.device('cpu'))) print('Loading saved network') except FileNotFoundError: print('Training net') # with torch.autograd.profiler.profile() as prof: for epoch in range(AE_Params.max_epochs): train_rmse = training.train(model, optimizer, device, scheduler, params, train_loader, train_trajectories, AE_Params, history) if AE_Params.cross_validation: test_rmse = training.val(model, device, params, test_loader, test_trajectories, AE_Params, history_test) print("Epoch[{}/{}, train_mse loss:{}, test_mse loss:{}".format(epoch + 1, AE_Params.max_epochs, history['train'][-1], history_test['test'][-1])) else: test_rmse = train_rmse print("Epoch[{}/{}, train_mse loss:{}".format(epoch + 1, AE_Params.max_epochs, history['train'][-1])) if test_rmse < min_test_loss: min_test_loss = test_rmse best_epoch = epoch torch.save(model.state_dict(), AE_Params.net_dir+AE_Params.net_name+AE_Params.net_run+'.pt') if AE_Params.tolerance >= train_rmse: print('Early stopping!') break np.save(AE_Params.net_dir+'history'+AE_Params.net_run+'.npy', history) np.save(AE_Params.net_dir+'history_test'+AE_Params.net_run+'.npy', history_test) # print(prof.key_averages().table(sort_by="self_cpu_time_total")) print("\nLoading best network for epoch: ", best_epoch) model.load_state_dict(torch.load(AE_Params.net_dir+AE_Params.net_name+AE_Params.net_run+'.pt', map_location=torch.device('cpu'))) model.to("cpu") params = params.to("cpu") vars = "GCA-ROM" results, latents_map, latents_gca = testing.evaluate(VAR_all, model, graph_loader, params, AE_Params, range(params.shape[0])) plotting.plot_loss(AE_Params) plotting.plot_latent(AE_Params, latents_map, latents_gca) plotting.plot_error(results, VAR_all, scaler_all, AE_Params, mu1_range, mu2_range, params, train_trajectories, vars) N = 5 snapshots = np.arange(params.shape[0]).tolist() np.random.shuffle(snapshots) for SNAP in snapshots[0:N]: plotting.plot_fields(SNAP, results, scaler_all, AE_Params, dataset, xx, yy, params) plotting.plot_error_fields(SNAP, results, VAR_all, scaler_all, AE_Params, dataset, xx, yy, params) results_test, _, _ = testing.evaluate(VAR_test, model, val_loader, params, AE_Params, test_trajectories) error_abs, norm = error.compute_error(results_test, VAR_test, scaler_test, AE_Params) error.print_error(error_abs, norm, vars) error.save_error(error_abs, norm, AE_Params, vars)
4,087
45.454545
157
py
gca-rom
gca-rom-main/gca_rom/preprocessing.py
import numpy as np import torch from torch_geometric.data import Data from torch_geometric.loader import DataLoader from gca_rom import scaling def graphs_dataset(dataset, AE_Params): """ graphs_dataset: function to process and scale the input dataset for graph autoencoder model. Inputs: dataset: an object containing the dataset to be processed. AE_Params: an object containing the hyperparameters of the graph autoencoder model. Outputs: dataset_graph: an object containing the processed and scaled dataset. loader: a DataLoader object of the processed and scaled dataset. train_loader: a DataLoader object of the training set. test_loader: a DataLoader object of the test set. val_loader: a DataLoader object of the validation set. scaler_all: a scaler object to scale the entire dataset. scaler_test: a scaler object to scale the test set. xx: an array of the x-coordinate of the nodes. yy: an array of the y-coordinate of the nodes. var: an array of the node features. VAR_all: an array of the scaled node features of the entire dataset. VAR_test: an array of the scaled node features of the test set. train_snapshots: a list of indices of the training set. test_snapshots: a list of indices of the test set. """ xx = dataset.xx yy = dataset.yy var = dataset.U # PROCESSING DATASET num_nodes = var.shape[0] num_graphs = var.shape[1] print("Number of nodes processed: ", num_nodes) print("Number of graphs processed: ", num_graphs) total_sims = int(num_graphs) rate = AE_Params.rate/100 train_sims = int(rate * total_sims) test_sims = total_sims - train_sims main_loop = np.arange(total_sims).tolist() np.random.shuffle(main_loop) train_snapshots = main_loop[0:train_sims] train_snapshots.sort() test_snapshots = main_loop[train_sims:total_sims] test_snapshots.sort() ## FEATURE SCALING var_test = dataset.U[:, test_snapshots] scaling_type = AE_Params.scaling_type scaler_all, VAR_all = scaling.tensor_scaling(var, scaling_type) scaler_test, VAR_test = scaling.tensor_scaling(var_test, scaling_type) print("Shape of scaled dataset: ", VAR_all.shape) print("Shape of scaled test set: ", VAR_test.shape) graphs = [] edge_index = torch.t(dataset.E) - 1 for graph in range(num_graphs): pos = torch.cat((xx[:, graph].unsqueeze(1), yy[:, graph].unsqueeze(1)), 1) ei = torch.index_select(pos, 0, edge_index[0, :]) ej = torch.index_select(pos, 0, edge_index[1, :]) edge_attr = ej - ei edge_attr = torch.sqrt(torch.pow(edge_attr[:, 0], 2) + torch.pow(edge_attr[:, 1], 2)) node_features = VAR_all[graph, :] dataset_graph = Data(x=node_features, edge_index=edge_index, edge_attr=edge_attr, pos=pos) graphs.append(dataset_graph) AE_Params.num_nodes = dataset_graph.num_nodes train_dataset = [graphs[i] for i in train_snapshots] test_dataset = [graphs[i] for i in test_snapshots] loader = DataLoader(graphs, batch_size=1) train_loader = DataLoader(train_dataset, batch_size=train_sims, shuffle=False) test_loader = DataLoader(test_dataset, batch_size=test_sims, shuffle=False) val_loader = DataLoader(test_dataset, batch_size=1, shuffle=False) return dataset_graph, loader, train_loader, test_loader, \ val_loader, scaler_all, scaler_test, xx, yy, var, VAR_all, VAR_test, train_snapshots, test_snapshots
3,515
38.954545
112
py
gca-rom
gca-rom-main/gca_rom/network.py
import sys import torch from torch import nn from gca_rom import gca, scaling, pde problem_name, variable, mu1_range, mu2_range = pde.problem(int(sys.argv[1])) class AE_Params(): """Class that holds the hyperparameters for the autoencoder model. Args: sparse_method (str): The method to use for sparsity constraint. rate (int): Amount of data used in training. seed (int): Seed for the random number generator. bottleneck_dim (int): The dimension of the bottleneck layer. tolerance (float): The tolerance value for stopping the training. lambda_map (float): The weight for the map loss. learning_rate (float): The learning rate for the optimizer. ffn (int): The number of feed-forward layers. in_channels (int): The number of input channels. hidden_channels (list): The number of hidden channels for each layer. act (function): The activation function to use. nodes (int): The number of nodes in each hidden layer. skip (int): The number of skipped connections. layer_vec (list): The structure of the network. net_name (str): The name of the network. scaler_name (str): The name of the scaler used for preprocessing. weight_decay (float): The weight decay for the optimizer. max_epochs (int): The maximum number of epochs to run training for. miles (list): The miles for learning rate update in scheduler. gamma (float): The gamma value for the optimizer. num_nodes (int): The number of nodes in the network. scaling_type (int): The type of scaling to use for preprocessing. net_dir (str): The directory to save the network in. cross_validation (bool): Whether to perform cross-validation. """ def __init__(self): self.scaling_type = int(sys.argv[2]) _, self.scaler_name = scaling.scaler_functions(int(sys.argv[3])) self.skip = int(sys.argv[4]) self.rate = int(sys.argv[5]) self.sparse_method = 'L1_mean' self.ffn = int(sys.argv[6]) self.nodes = int(sys.argv[7]) self.bottleneck_dim = int(sys.argv[8]) self.lambda_map = float(sys.argv[9]) self.in_channels = int(sys.argv[10]) self.seed = 10 self.tolerance = 1e-6 self.learning_rate = 0.001 self.hidden_channels = [1]*self.in_channels self.act = torch.tanh self.layer_vec=[2, self.nodes, self.nodes, self.nodes, self.nodes, self.bottleneck_dim] self.net_name = problem_name self.net_run = '_' + self.scaler_name self.weight_decay = 0.00001 self.max_epochs = 5000 self.miles = [] self.gamma = 0.0001 self.num_nodes = 0 self.net_dir = './' + problem_name + '/' + self.net_run+ '/' + variable + '_' + self.net_name + '_lmap' + str(self.lambda_map) + '_btt' + str(self.bottleneck_dim) \ + '_seed' + str(self.seed) + '_lv' + str(len(self.layer_vec)-2) + '_hc' + str(len(self.hidden_channels)) + '_nd' + str(self.nodes) \ + '_ffn' + str(self.ffn) + '_skip' + str(self.skip) + '_lr' + str(self.learning_rate) + '_sc' + str(self.scaling_type) + '_rate' + str(self.rate) + '/' self.cross_validation = True AE_Params = AE_Params() class Net(torch.nn.Module): """ Class Net --------- A PyTorch neural network class which consists of encoder, decoder and mapping modules. Attributes ---------- encoder : gca.Encoder An encoder module from the gca module. decoder : gca.Decoder A decoder module from the gca module. act_map : AE_Params.act The activation map specified in the AE_Params. layer_vec : AE_Params.layer_vec The layer vector specified in the AE_Params. Shape of the layers for the parameter space mapping. steps : int Number of layers for the parameter mapping MLP. maptovec : nn.ModuleList A list of linear layers for mapping. Methods ------- solo_encoder(data) Encodes the input data using the encoder module. Returns the encoded representation. solo_decoder(x, data) Decodes the encoded representation and the input data using the decoder module. Returns the decoded output. mapping(x) Maps the input using the linear modules in maptovec. Returns the mapped output. forward(data, parameters) Runs a forward pass through the network using the input data and parameters. Returns the decoded output, encoded representation, and estimated encoded representation. """ def __init__(self): super().__init__() self.encoder = gca.Encoder(AE_Params.hidden_channels, AE_Params.bottleneck_dim, AE_Params.num_nodes, ffn=AE_Params.ffn, skip=AE_Params.skip) self.decoder = gca.Decoder(AE_Params.hidden_channels, AE_Params.bottleneck_dim, AE_Params.num_nodes, ffn=AE_Params.ffn, skip=AE_Params.skip) self.act_map = AE_Params.act self.layer_vec = AE_Params.layer_vec self.steps = len(self.layer_vec) - 1 self.maptovec = nn.ModuleList() for k in range(self.steps): self.maptovec.append(nn.Linear(self.layer_vec[k], self.layer_vec[k+1])) def solo_encoder(self,data): x = self.encoder(data) return x def solo_decoder(self,x, data): x = self.decoder(x, data) return x def mapping(self, x): idx = 0 for layer in self.maptovec: if(idx==self.steps): x = layer(x) else: x =self.act_map(layer(x)) idx += 1 return x def forward(self, data, parameters): z = self.solo_encoder(data) z_estimation = self.mapping(parameters) x = self.solo_decoder(z, data) # x = self.solo_decoder(z_estimation, data) return x, z, z_estimation
5,946
39.732877
179
py
gca-rom
gca-rom-main/gca_rom/training.py
import torch import torch.nn.functional as F def train(model, optimizer, device, scheduler, params, train_loader, train_trajectories, AE_Params, history): """Trains the autoencoder model. This function trains the autoencoder model using mean squared error (MSE) loss and a map loss, where the map loss is the MSE between the estimated z (latent space) and the actual z. The final loss is the sum of the MSE loss and the map loss multiplied by the weight `AE_Params.lambda_map`. The model is trained on the data from `train_loader` and the optimization process is performed using the `optimizer`. The learning rate is updated after every iteration using `scheduler`. Use of mini-batching for reducing the computational cost. Args: model (torch.nn.Module): The autoencoder model to be trained. optimizer (torch.optim.Optimizer): The optimizer to update the model parameters. device (str): The device to run the model on ('cuda' or 'cpu'). scheduler (torch.optim.lr_scheduler._LRScheduler): The learning rate scheduler to update the learning rate. params (torch.Tensor): Tensor containing the parameters of the model. train_loader (torch.utils.data.DataLoader): The data loader to provide the training data. train_trajectories (int): The number of training trajectories. AE_Params (dict): The dictionary containing the hyperparameters for the autoencoder model. history (dict): The dictionary to store the loss history. Returns: float: The average loss over all training examples. """ model.train() total_loss_train = total_examples = sum_loss = 0 total_loss_train_1 = total_loss_train_2 = 0 sum_loss_1 = sum_loss_2 = 0 for data in train_loader: optimizer.zero_grad() data = data.to(device) out, z, z_estimation = model(data, params[train_trajectories, :]) loss_train_mse = F.mse_loss(out, data.x, reduction='mean') loss_train_map = F.mse_loss(z_estimation, z, reduction='mean') loss_train = loss_train_mse + AE_Params.lambda_map * loss_train_map loss_train.backward() optimizer.step() sum_loss += loss_train.item() sum_loss_1 += loss_train_mse.item() sum_loss_2 += loss_train_map.item() total_examples += 1 scheduler.step() total_loss_train = sum_loss / total_examples total_loss_train_1 = sum_loss_1 / total_examples total_loss_train_2 = sum_loss_2 / total_examples history['train'].append(total_loss_train) history['l1'].append(total_loss_train_1) history['l2'].append(total_loss_train_2) return total_loss_train def val(model, device, params, test_loader, test_trajectories, AE_Params, history_test): """ Evaluate the performance of a model on a test set. This function calculates the mean of the total loss, the mean of loss_test_mse and the mean of loss_test_map for all test examples. The losses are computed as the mean squared error (MSE) between the model's predictions and the true target variables, and between the estimated latent code and the true latent code. The lambda_map weight balances the contribution of each loss term to the total loss. The function adds the computed loss values to the history_test dictionary. Parameters: model (torch.nn.Module): The model to be evaluated. device (torch.device): The device to use for computations (e.g. 'cpu' or 'cuda'). params (torch.Tensor): Tensor containing the parameters of the model. test_loader (torch.utils.data.DataLoader): The test data to use for evaluation. test_trajectories (int): The index of the test trajectory. AE_Params (object): Object containing hyperparameters for the model. history_test (dict): Dictionary to store the evaluation results. Returns: float: The mean of the total loss computed over all test examples. """ with torch.no_grad(): model.eval() total_loss_test = total_examples = sum_loss = 0 total_loss_test_1 = total_loss_test_2 = 0 sum_loss_1 = sum_loss_2 = 0 for data in test_loader: data = data.to(device) out, z, z_estimation = model(data, params[test_trajectories, :]) loss_test_mse = F.mse_loss(out, data.x, reduction='mean') loss_test_map = F.mse_loss(z_estimation, z, reduction='mean') loss_test = loss_test_mse + AE_Params.lambda_map * loss_test_map sum_loss += loss_test.item() sum_loss_1 += loss_test_mse.item() sum_loss_2 += loss_test_map.item() total_examples += 1 total_loss_test = sum_loss / total_examples total_loss_test_1 = sum_loss_1 / total_examples total_loss_test_2 = sum_loss_2 / total_examples history_test['test'].append(total_loss_test) history_test['l1'].append(total_loss_test_1) history_test['l2'].append(total_loss_test_2) return total_loss_test
5,063
48.647059
478
py
gca-rom
gca-rom-main/gca_rom/testing.py
import torch from tqdm import tqdm import numpy as np def evaluate(VAR, model, loader, params, AE_Params, test): """ This function evaluates the performance of a trained Autoencoder (AE) model. It encodes the input data using both the model's encoder and a mapping function, and decodes the resulting latent representations to obtain predicted solutions. The relative error between the two latent representations is also computed. Inputs: VAR: np.array, ground truth solution model: object, trained AE model loader: object, data loader for the input data params: np.array, model parameters AE_Params: class, model architecture and training parameters Returns: results: np.array, predicted solutions latents_map: np.array, latent representations obtained using the mapping function latents_gca: np.array, latent representations obtained using the AE encoder """ results = torch.zeros(VAR.shape[0], VAR.shape[1], 1) latents_map = torch.zeros(VAR.shape[0], AE_Params.bottleneck_dim) latents_gca = torch.zeros(VAR.shape[0], AE_Params.bottleneck_dim) index = 0 latents_error = list() with torch.no_grad(): for data in tqdm(loader): z_net = model.solo_encoder(data) z_map = model.mapping(params[test[index], :]) latents_map[index, :] = z_map latents_gca[index, :] = z_net lat_err = np.linalg.norm(z_net - z_map)/np.linalg.norm(z_net) latents_error.append(lat_err) results[index, :, :] = model.solo_decoder(z_map, data) index += 1 np.savetxt(AE_Params.net_dir+'latents'+AE_Params.net_run+'.csv', latents_map.detach(), delimiter =',') latents_error = np.array(latents_error) # print("\nMaximum relative error for latent = ", max(latents_error)) # print("Mean relative error for latent = ", sum(latents_error)/len(latents_error)) # print("Minimum relative error for latent = ", min(latents_error)) return results, latents_map, latents_gca
2,071
44.043478
110
py
gca-rom
gca-rom-main/gca_rom/initialization.py
import os import torch import numpy as np import random import warnings def set_device(): """ Returns the device to be used (GPU or CPU) Returns: device (str): The device to be used ('cuda' if GPU is available, 'cpu' otherwise) """ device = 'cuda' if torch.cuda.is_available() else 'cpu' print("Device used = ", device) torch.set_default_dtype(torch.float64) warnings.filterwarnings("ignore") return device def set_reproducibility(AE_Params): """ Sets the seed for reproducibility of results. Args: AE_Params (class): Contains the hyperparameters of the autoencoder """ seed = AE_Params.seed torch.manual_seed(seed) torch.cuda.manual_seed(seed) torch.cuda.manual_seed_all(seed) np.random.seed(seed) random.seed(seed) torch.backends.cudnn.benchmark = False torch.backends.cudnn.deterministic = True def set_path(AE_Params): """ Creates the directory path to store the network results. Args: AE_Params (class): Contains the hyperparameters of the autoencoder """ path = AE_Params.net_dir isExist = os.path.exists(path) if not isExist: os.makedirs(path, exist_ok=False)
1,224
21.685185
89
py
gca-rom
gca-rom-main/gca_rom/loader.py
import sys from torch_geometric.data import Dataset import torch import scipy class LoadDataset(Dataset): """ A custom dataset class which loads data from a .mat file using scipy.io.loadmat. data_mat : scipy.io.loadmat The loaded data in a scipy.io.loadmat object. U : torch.Tensor The tensor representation of the specified variable from the data_mat. xx : torch.Tensor The tensor representation of the 'xx' key from the data_mat. Refers to X coordinates of the domain yy : torch.Tensor The tensor representation of the 'yy' key from the data_mat.Refers to Y coordinates of the domain T : torch.Tensor The tensor representation of the 'T' key from the data_mat, casted to int. Adjacency Matrix E : torch.Tensor The tensor representation of the 'E' key from the data_mat, casted to int. Connection Matrix __init__(self, root_dir, variable) Initializes the LoadDataset object by loading the data from the .mat file at the root_dir location and converting the specified variable to a tensor representation. """ def __init__(self, root_dir, variable): self.data_mat = scipy.io.loadmat(root_dir) self.U = torch.tensor(self.data_mat[variable]) self.xx = torch.tensor(self.data_mat['xx']) self.yy = torch.tensor(self.data_mat['yy']) self.T = torch.tensor(self.data_mat['T'].astype(int)) self.E = torch.tensor(self.data_mat['E'].astype(int))
1,488
41.542857
172
py
gca-rom
gca-rom-main/gca_rom/gca.py
import torch from torch import nn import torch.nn.functional as F from torch_geometric.nn import GMMConv class Encoder(torch.nn.Module): """ Encoder Class The Encoder class is a subclass of torch.nn.Module that implements a deep neural network for encoding graph data. It uses the Gaussian Mixture convolution (GMMConv) module to extract features from the graph structure and node features. The encoding is then passed through a feed-forward neural network with two fully connected layers to produce the final encoding. Arguments: hidden_channels (list): A list of hidden channel sizes for each layer of the GMMConv module. bottleneck (int): Size of the bottleneck layer in the feed-forward neural network. input_size (int): Size of the node features. ffn (int): Size of the intermediate layer in the feed-forward neural network. skip (bool): If True, the input node features will be concatenated with the GMMConv output at each layer. act (function): Activation function used in the GMMConv layers and feed-forward neural network. Defaults to F.elu. Methods: encoder(data): Encodes the graph data using the GMMConv module and feed-forward neural network. reset_parameters(): Resets the parameters of the GMMConv layers and feed-forward neural network. forward(data): A convenience function that calls the encoder method. """ def __init__(self, hidden_channels, bottleneck, input_size, ffn, skip, act=F.elu): super().__init__() self.hidden_channels = hidden_channels self.depth = len(self.hidden_channels) self.act = act self.ffn = ffn self.skip = skip self.bottleneck = bottleneck self.input_size = input_size self.down_convs = torch.nn.ModuleList() for i in range(self.depth-1): self.down_convs.append(GMMConv(self.hidden_channels[i], self.hidden_channels[i+1], dim=1, kernel_size=5)) self.fc_in1 = nn.Linear(self.input_size*self.hidden_channels[-1], self.ffn) self.fc_in2 = nn.Linear(self.ffn, self.bottleneck) self.reset_parameters() def encoder(self, data): edge_weight = data.edge_attr edge_index = data.edge_index x = data.x idx = 0 for layer in self.down_convs: x = self.act(layer(x, edge_index, edge_weight.unsqueeze(1))) if self.skip: x = x + data.x idx += 1 x = x.reshape(data.num_graphs, self.input_size * self.hidden_channels[-1]) x = self.act(self.fc_in1(x)) x = self.fc_in2(x) return x def reset_parameters(self): for conv in self.down_convs: conv.reset_parameters() for name, param in conv.named_parameters(): if 'bias' in name: nn.init.constant_(param, 0) else: nn.init.kaiming_uniform_(param) def forward(self,data): x = self.encoder(data) return x class Decoder(torch.nn.Module): """ Class Decoder A torch.nn.Module subclass for the decoder part of a neural network. Attributes: hidden_channels (list of ints): A list of hidden channel sizes. depth (int): The length of hidden_channels list. act (function): Activation function to use. ffn (int): Size of output after the first linear layer. skip (bool): Whether to add skip connections. bottleneck (int): Size of bottleneck layer. input_size (int): Size of input data. fc_out1 (torch.nn.Linear): Linear layer from bottleneck to ffn. fc_out2 (torch.nn.Linear): Linear layer from ffn to input_size * hidden_channels[-1]. up_convs (torch.nn.ModuleList): A list of GMMConv layers. Methods: decoder(self, x, data): Decodes the input data x and returns the output. reset_parameters(self): Resets the parameters of the up_convs layer. forward(self, x, data): Performs a forward pass on the input data x and returns the output. """ def __init__(self, hidden_channels, bottleneck, input_size, ffn, skip, act=F.elu): super().__init__() self.hidden_channels = hidden_channels self.depth = len(self.hidden_channels) self.act = act self.ffn = ffn self.skip = skip self.bottleneck = bottleneck self.input_size = input_size self.fc_out1 = nn.Linear(self.bottleneck, self.ffn) self.fc_out2 = nn.Linear(self.ffn, self.input_size * self.hidden_channels[-1]) self.up_convs = torch.nn.ModuleList() for i in range(self.depth-1): self.up_convs.append(GMMConv(self.hidden_channels[self.depth-1-i], self.hidden_channels[self.depth-i-2], dim=1, kernel_size=5)) self.reset_parameters() def decoder(self, x, data): edge_weight = data.edge_attr edge_index = data.edge_index x = self.act(self.fc_out1(x)) x = self.act(self.fc_out2(x)) h = x.reshape(data.num_graphs*self.input_size, self.hidden_channels[-1]) x = h idx = 0 for layer in self.up_convs: if (idx == self.depth - 2): x = layer(x, edge_index, edge_weight.unsqueeze(1)) else: x = self.act(layer(x, edge_index, edge_weight.unsqueeze(1))) if self.skip: x = x + h idx += 1 return x def reset_parameters(self): for conv in self.up_convs: conv.reset_parameters() for name, param in conv.named_parameters(): if 'bias' in name: nn.init.constant_(param, 0) else: nn.init.kaiming_uniform_(param) def forward(self, x, data): x = self.decoder(x, data) return x
5,918
36.226415
139
py
gca-rom
gca-rom-main/gca_rom/scaling.py
from sklearn import preprocessing import torch import sys def scaler_functions(k): match k: case 1: sc_name = "minmax" sc_fun = preprocessing.MinMaxScaler() case 2: sc_name = "robust" sc_fun = preprocessing.RobustScaler() case 3: sc_name = "standard" sc_fun = preprocessing.StandardScaler() return sc_fun, sc_name def tensor_scaling(tensor, scaling_type): scaling_fun_1, _ = scaler_functions(int(sys.argv[3])) scaling_fun_2, _ = scaler_functions(int(sys.argv[3])) match scaling_type: case 1: # print("SAMPLE SCALING") scale = scaling_fun_1.fit(tensor) scaled_data = torch.unsqueeze(torch.tensor(scale.transform(tensor)),0).permute(2, 1, 0) case 2: # print("FEATURE SCALING") scale = scaling_fun_1.fit(torch.t(tensor)) scaled_data = torch.unsqueeze(torch.tensor(scale.transform(torch.t(tensor))), 0).permute(1, 2, 0) case 3: # print("FEATURE-SAMPLE SCALING") scaler_f = scaling_fun_1.fit(torch.t(tensor)) temp = torch.tensor(scaler_f.transform(torch.t(tensor))) scaler_s = scaling_fun_2.fit(temp) scaled_data = torch.unsqueeze(torch.tensor(scaler_s.transform(temp)), 0).permute(1, 2, 0) scale = [scaler_f, scaler_s] case 4: # print("SAMPLE-FEATURE SCALING") scaler_s = scaling_fun_1.fit(tensor) temp = torch.t(torch.tensor(scaler_s.transform(tensor))) scaler_f = scaling_fun_2.fit(temp) scaled_data = torch.unsqueeze(torch.t(torch.tensor(scaler_f.transform(temp))), 0).permute(2, 1, 0) scale = [scaler_s, scaler_f] return scale, scaled_data def inverse_scaling(tensor, scale, scaling_type): match scaling_type: case 1: # print("SAMPLE SCALING") rescaled_data = torch.tensor(scale.inverse_transform(torch.t(torch.tensor(tensor[:, :, 0].detach().numpy().squeeze())))) case 2: # print("FEATURE SCALING") rescaled_data = torch.tensor(torch.t(torch.tensor(scale.inverse_transform(tensor[:, :, 0].detach().numpy().squeeze())))) case 3: # print("FEATURE-SAMPLE SCALING") scaler_f = scale[0] scaler_s = scale[1] rescaled_data = torch.t(torch.tensor(scaler_f.inverse_transform(torch.tensor(scaler_s.inverse_transform(tensor[:, :, 0].detach().numpy().squeeze()))))) case 4: # print("SAMPLE-FEATURE SCALING") scaler_s = scale[0] scaler_f = scale[1] rescaled_data = torch.tensor(scaler_s.inverse_transform(torch.t(torch.tensor(scaler_f.inverse_transform(tensor[:, :, 0].detach().numpy().squeeze()))))) return rescaled_data
2,869
41.205882
163
py
CRST
CRST-master/evaluate.py
import argparse import scipy from scipy import ndimage import numpy as np import sys from packaging import version import time import util import torch import torchvision.models as models import torch.nn.functional as F from torch.utils import data, model_zoo from deeplab.model import Res_Deeplab from deeplab.datasets import GTA5TestDataSet from collections import OrderedDict import os from PIL import Image import matplotlib.pyplot as plt import torch.nn as nn # IMG_MEAN = np.array((104.00698793,116.66876762,122.67891434), dtype=np.float32) IMG_MEAN = np.array((0.406, 0.456, 0.485), dtype=np.float32) # BGR IMG_STD = np.array((0.225, 0.224, 0.229), dtype=np.float32) # BGR DATA_DIRECTORY = './dataset/cityscapes' DATA_LIST_PATH = './dataset/list/cityscapes/val.lst' SAVE_PATH = './cityscapes/eval' TEST_IMAGE_SIZE = '1024,2048' TEST_SCALE = 1.0 IGNORE_LABEL = 255 NUM_CLASSES = 19 NUM_STEPS = 500 # Number of images in the validation set. RESTORE_FROM = './src_model/gta5/src_model.pth' DATA_SRC = 'cityscapes' SET = 'val' LOG_FILE = 'log' MODEL = 'DeeplabRes' def get_arguments(): """Parse all the arguments provided from the CLI. Returns: A list of parsed arguments. """ parser = argparse.ArgumentParser(description="DeepLab-ResNet Network") parser.add_argument("--model", type=str, default=MODEL, help="Model Choice (DeeplabMulti/DeeplabVGG).") parser.add_argument("--data-src", type=str, default=DATA_SRC, help="Data name.") parser.add_argument("--data-dir", type=str, default=DATA_DIRECTORY, help="Path to the directory containing the Cityscapes dataset.") parser.add_argument("--data-list", type=str, default=DATA_LIST_PATH, help="Path to the file listing the images in the dataset.") parser.add_argument("--ignore-label", type=int, default=IGNORE_LABEL, help="The index of the label to ignore during the training.") parser.add_argument("--num-classes", type=int, default=NUM_CLASSES, help="Number of classes to predict (including background).") parser.add_argument("--restore-from", type=str, default=RESTORE_FROM, help="Where restore model parameters from.") parser.add_argument("--gpu", type=int, default=0, help="choose gpu device.") parser.add_argument('--test-flipping', dest='test_flipping', help='If average predictions of original and flipped images.', default=False, action='store_true') parser.add_argument("--set", type=str, default=SET, help="choose evaluation set.") parser.add_argument("--save", type=str, default=SAVE_PATH, help="Path to save result.") parser.add_argument("--log-file", type=str, default=LOG_FILE, help="The name of log file.") parser.add_argument('--debug',help='True means logging debug info.', default=False, action='store_true') parser.add_argument('--test-scale', type=str, default=TEST_SCALE, help='The test scales. Multi-scale supported') parser.add_argument('--test-image-size', default=TEST_IMAGE_SIZE, help='The test image size', type=str) return parser.parse_args() args = get_arguments() # palette if args.data_src == 'gta' or args.data_src == 'cityscapes': # gta: palette = [128, 64, 128, 244, 35, 232, 70, 70, 70, 102, 102, 156, 190, 153, 153, 153, 153, 153, 250, 170, 30, 220, 220, 0, 107, 142, 35, 152, 251, 152, 70, 130, 180, 220, 20, 60, 255, 0, 0, 0, 0, 142, 0, 0, 70, 0, 60, 100, 0, 80, 100, 0, 0, 230, 119, 11, 32] elif args.data_src == 'synthia': # synthia: palette = [128,64,128,244,35,232,70,70,70,102,102,156,64,64,128,153,153,153,250,170,30,220,220,0, 107,142,35,70,130,180,220,20,60,255,0,0,0,0,142,0,60,100,0,0,230,119,11,32] zero_pad = 256 * 3 - len(palette) for i in range(zero_pad): palette.append(0) def colorize_mask(mask): # mask: numpy array of the mask new_mask = Image.fromarray(mask.astype(np.uint8)).convert('P') new_mask.putpalette(palette) return new_mask def main(): """Create the model and start the evaluation process.""" device = torch.device("cuda:" + str(args.gpu)) if not os.path.exists(args.save): os.makedirs(args.save) logger = util.set_logger(args.save, args.log_file, args.debug) logger.info('start with arguments %s', args) x_num = 0 with open(args.data_list) as f: for _ in f.readlines(): x_num = x_num + 1 sys.path.insert(0, 'dataset/helpers') if args.data_src == 'gta' or args.data_src == 'cityscapes': from labels import id2label, trainId2label elif args.data_src == 'synthia': from labels_cityscapes_synthia import id2label, trainId2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) valid_labels = sorted(set(id_2_label.ravel())) scorer = ScoreUpdater(valid_labels, args.num_classes, x_num, logger) scorer.reset() if args.model == 'DeeplabRes': model = Res_Deeplab(num_classes=args.num_classes) # elif args.model == 'DeeplabVGG': # model = DeeplabVGG(num_classes=args.num_classes) # if args.restore_from == RESTORE_FROM: # args.restore_from = RESTORE_FROM_VGG if args.restore_from[:4] == 'http' : saved_state_dict = model_zoo.load_url(args.restore_from) new_params = model.state_dict().copy() for i in saved_state_dict: # Scale.layer5.conv2d_list.3.weight i_parts = str(i).split('.') # print i_parts if not i_parts[0] == 'fc': new_params['.'.join(i_parts[0:])] = saved_state_dict[i] else: loc = "cuda:" + str(args.gpu) saved_state_dict = torch.load(args.restore_from,map_location=loc) new_params = saved_state_dict.copy() model.load_state_dict(new_params) #model.train() model.eval() model.to(device) testloader = data.DataLoader(GTA5TestDataSet(args.data_dir, args.data_list, test_scale = 1.0, test_size=(1024, 512), mean=IMG_MEAN, std=IMG_STD, scale=False, mirror=False), batch_size=1, shuffle=False, pin_memory=True) test_scales = [float(_) for _ in str(args.test_scale).split(',')] h, w = map(int, args.test_image_size.split(',')) if version.parse(torch.__version__) >= version.parse('0.4.0'): interp = nn.Upsample(size=(h, w), mode='bilinear', align_corners=True) else: interp = nn.Upsample(size=(h, w), mode='bilinear') test_image_size = (h, w) mean_rgb = IMG_MEAN[::-1].copy() std_rgb = IMG_STD[::-1].copy() with torch.no_grad(): for index, batch in enumerate(testloader): image, label, _, name = batch img = image.clone() num_scales = len(test_scales) # output_dict = {k: [] for k in range(num_scales)} for scale_idx in range(num_scales): if version.parse(torch.__version__) > version.parse('0.4.0'): image = F.interpolate(image, scale_factor=test_scales[scale_idx], mode='bilinear', align_corners=True) else: test_size = ( int(h*test_scales[scale_idx]), int(w*test_scales[scale_idx]) ) interp_tmp = nn.Upsample(size=test_size, mode='bilinear', align_corners=True) image = interp_tmp(img) if args.model == 'DeeplabRes': output2 = model(image.to(device)) coutput = interp(output2).cpu().data[0].numpy() if args.test_flipping: output2 = model(torch.from_numpy(image.numpy()[:,:,:,::-1].copy()).to(device)) coutput = 0.5 * ( coutput + interp(output2).cpu().data[0].numpy()[:,:,::-1] ) if scale_idx == 0: output = coutput.copy() else: output += coutput output = output/num_scales output = output.transpose(1,2,0) output = np.asarray(np.argmax(output, axis=2), dtype=np.uint8) pred_label = output.copy() label = label_2_id[np.asarray(label.numpy(), dtype=np.uint8)] scorer.update(pred_label.flatten(), label.flatten(), index) output_col = colorize_mask(output) output = Image.fromarray(output) name = name[0].split('/')[-1] output.save('%s/%s' % (args.save, name)) output_col.save('%s/%s_color.png' % (args.save, name.split('.')[0])) class ScoreUpdater(object): # only IoU are computed. accu, cls_accu, etc are ignored. def __init__(self, valid_labels, c_num, x_num, logger=None, label=None, info=None): self._valid_labels = valid_labels self._confs = np.zeros((c_num, c_num)) self._per_cls_iou = np.zeros(c_num) self._logger = logger self._label = label self._info = info self._num_class = c_num self._num_sample = x_num @property def info(self): return self._info def reset(self): self._start = time.time() self._computed = np.zeros(self._num_sample) # one-dimension self._confs[:] = 0 def fast_hist(self,label, pred_label, n): k = (label >= 0) & (label < n) return np.bincount(n * label[k].astype(int) + pred_label[k], minlength=n ** 2).reshape(n, n) def per_class_iu(self,hist): return np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist)) def do_updates(self, conf, i, computed=True): if computed: self._computed[i] = 1 self._per_cls_iou = self.per_class_iu(conf) def update(self, pred_label, label, i, computed=True): conf = self.fast_hist(label, pred_label, self._num_class) self._confs += conf self.do_updates(self._confs, i, computed) self.scores(i) def scores(self, i=None, logger=None): x_num = self._num_sample ious = np.nan_to_num( self._per_cls_iou ) logger = self._logger if logger is None else logger if logger is not None: if i is not None: speed = 1. * self._computed.sum() / (time.time() - self._start) logger.info('Done {}/{} with speed: {:.2f}/s'.format(i + 1, x_num, speed)) name = '' if self._label is None else '{}, '.format(self._label) logger.info('{}mean iou: {:.2f}%'. \ format(name, np.mean(ious) * 100)) with util.np_print_options(formatter={'float': '{:5.2f}'.format}): logger.info('\n{}'.format(ious * 100)) return ious if __name__ == '__main__': main()
11,168
39.762774
176
py
CRST
CRST-master/crst_seg.py
import argparse import sys from packaging import version import time import util import os import os.path as osp import timeit from collections import OrderedDict import scipy.io import torch import torchvision.models as models import torch.nn.functional as F from torch.utils import data, model_zoo import torch.backends.cudnn as cudnn import torch.nn as nn import torch.optim as optim from operator import itemgetter import scipy from scipy import ndimage import math from PIL import Image import numpy as np import shutil import random from deeplab.model import Res_Deeplab from deeplab.datasets import GTA5TestDataSet from deeplab.datasets import SrcSTDataSet, GTA5StMineDataSet, SoftSrcSTDataSet, SoftGTA5StMineDataSet ### shared ### IMG_MEAN = np.array((0.406, 0.456, 0.485), dtype=np.float32) # BGR IMG_STD = np.array((0.225, 0.224, 0.229), dtype=np.float32) # BGR # data ### source ## gta DATA_SRC_DIRECTORY = './dataset/gta5' DATA_SRC_LIST_PATH = './dataset/list/gta5/train.lst' DATA_SRC = 'gta' RESTORE_FROM = './src_model/gta5/src_model.pth' NUM_CLASSES = 19 INIT_SRC_PORT = 0.03 # GTA: 0.03 ### target DATA_TGT_DIRECTORY = './dataset/cityscapes' DATA_TGT_TRAIN_LIST_PATH = './dataset/list/cityscapes/train_ClsConfSet.lst' DATA_TGT_TEST_LIST_PATH = './dataset/list/cityscapes/val.lst' IGNORE_LABEL = 255 # train scales for src and tgt TRAIN_SCALE_SRC = '0.5,1.5' TRAIN_SCALE_TGT = '0.5,1.5' # model MODEL = 'DeeplabRes' # gpu GPU = 0 PIN_MEMORY = False # log files LOG_FILE = 'self_training_log' ### train ### BATCH_SIZE = 2 INPUT_SIZE = '512,1024'# 512,1024 for GTA; RANDSEED = 3 # params for optimizor LEARNING_RATE =5e-5 POWER = 0.0 MOMENTUM = 0.9 WEIGHT_DECAY = 0.0005 NUM_ROUNDS = 4 EPR = 2 SRC_SAMPLING_POLICY = 'r' KC_POLICY = 'cb' KC_VALUE = 'conf' INIT_TGT_PORT = 0.2 MAX_TGT_PORT = 0.5 TGT_PORT_STEP = 0.05 # varies but dataset MAX_SRC_PORT = 0.06 #0.06; SRC_PORT_STEP = 0.0025 #0.0025: MRKLD = 0.0 LRENT = 0.0 MRSRC = 0.0 MINE_PORT = 1e-3 RARE_CLS_NUM = 3 MINE_CHANCE = 0.8 ### val ### SAVE_PATH = 'debug' TEST_IMAGE_SIZE = '1024,2048' EVAL_SCALE = 0.9 TEST_SCALE = '0.9,1.0,1.2' DS_RATE = 4 def seed_torch(seed=0): random.seed(seed) os.environ['PYTHONHASHSEED'] = str(seed) np.random.seed(seed) torch.manual_seed(seed) 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.enabled = False #torch.backends.cudnn.deterministic = True def get_arguments(): """Parse all the arguments provided from the CLI. Returns: A list of parsed arguments. """ parser = argparse.ArgumentParser(description="DeepLab-ResNet Network") ### shared by train & val # data parser.add_argument("--data-src", type=str, default=DATA_SRC, help="Name of source dataset.") parser.add_argument("--data-src-dir", type=str, default=DATA_SRC_DIRECTORY, help="Path to the directory containing the source dataset.") parser.add_argument("--data-src-list", type=str, default=DATA_SRC_LIST_PATH, help="Path to the file listing the images&labels in the source dataset.") parser.add_argument("--data-tgt-dir", type=str, default=DATA_TGT_DIRECTORY, help="Path to the directory containing the target dataset.") parser.add_argument("--data-tgt-train-list", type=str, default=DATA_TGT_TRAIN_LIST_PATH, help="Path to the file listing the images*GT labels in the target train dataset.") parser.add_argument("--data-tgt-test-list", type=str, default=DATA_TGT_TEST_LIST_PATH, help="Path to the file listing the images*GT labels in the target test dataset.") parser.add_argument("--num-classes", type=int, default=NUM_CLASSES, help="Number of classes to predict (including background).") parser.add_argument("--ignore-label", type=int, default=IGNORE_LABEL, help="The index of the label to ignore during the training.") # model parser.add_argument("--model", type=str, default=MODEL, help="Model Choice (DeeplabMulti/DeeplabVGG).") parser.add_argument("--restore-from", type=str, default=RESTORE_FROM, help="Where restore model parameters from.") # gpu parser.add_argument("--gpu", type=int, default=GPU, help="choose gpu device.") parser.add_argument("--pin-memory", type=bool, default=PIN_MEMORY, help="Whether to pin memory in train & eval.") # log files parser.add_argument("--log-file", type=str, default=LOG_FILE, help="The name of log file.") parser.add_argument('--debug',help='True means logging debug info.', default=False, action='store_true') ### train ### parser.add_argument("--batch-size", type=int, default=BATCH_SIZE, help="Number of images sent to the network in one step.") parser.add_argument("--input-size", type=str, default=INPUT_SIZE, help="Comma-separated string with height and width of images.") parser.add_argument("--is-training", action="store_true", help="Whether to updates the running means and variances during the training.") parser.add_argument("--eval-training", action="store_true", help="Use the saved means and variances, or running means and variances during the evaluation.") parser.add_argument("--random-mirror", action="store_true", help="Whether to randomly mirror the inputs during the training.") parser.add_argument("--random-scale", action="store_true", help="Whether to randomly scale the inputs during the training.") parser.add_argument("--train-scale-src", type=str, default=TRAIN_SCALE_SRC, help="The scale for multi-scale training in source domain.") parser.add_argument("--train-scale-tgt", type=str, default=TRAIN_SCALE_TGT, help="The scale for multi-scale training in target domain.") # params for optimizor parser.add_argument("--learning-rate", type=float, default=LEARNING_RATE, help="Base learning rate for training with polynomial decay.") parser.add_argument("--power", type=float, default=POWER, help="Decay parameter to compute the learning rate.") parser.add_argument("--momentum", type=float, default=MOMENTUM, help="Momentum component of the optimiser.") parser.add_argument("--weight-decay", type=float, default=WEIGHT_DECAY, help="Regularisation parameter for L2-loss.") ### val parser.add_argument('--test-flipping', dest='test_flipping', help='If average predictions of original and flipped images.', default=False, action='store_true') parser.add_argument("--test-image-size", type=str, default=TEST_IMAGE_SIZE, help="The test image size.") parser.add_argument("--eval-scale", type=float, default=EVAL_SCALE, help="The test image scale.") parser.add_argument("--test-scale", type=str, default=TEST_SCALE, help="The test image scale.") ### self-training params parser.add_argument("--save", type=str, default=SAVE_PATH, help="Path to save result for self-training.") parser.add_argument("--num-rounds", type=int, default=NUM_ROUNDS, help="Number of rounds for self-training.") parser.add_argument("--epr", type=int, default=EPR, help="Number of epochs per round for self-training.") parser.add_argument('--kc-policy', default=KC_POLICY, type=str, dest='kc_policy', help='The policy to determine kc. "cb" for weighted class-balanced threshold') parser.add_argument('--kc-value', default=KC_VALUE, type=str, help='The way to determine kc values, either "conf", or "prob".') parser.add_argument('--ds-rate', default=DS_RATE, type=int, help='The downsampling rate in kc calculation.') parser.add_argument('--init-tgt-port', default=INIT_TGT_PORT, type=float, dest='init_tgt_port', help='The initial portion of target to determine kc') parser.add_argument('--max-tgt-port', default=MAX_TGT_PORT, type=float, dest='max_tgt_port', help='The max portion of target to determine kc') parser.add_argument('--tgt-port-step', default=TGT_PORT_STEP, type=float, dest='tgt_port_step', help='The portion step in target domain in every round of self-paced self-trained neural network') parser.add_argument('--init-src-port', default=INIT_SRC_PORT, type=float, dest='init_src_port', help='The initial portion of source portion for self-trained neural network') parser.add_argument('--max-src-port', default=MAX_SRC_PORT, type=float, dest='max_src_port', help='The max portion of source portion for self-trained neural network') parser.add_argument('--src-port-step', default=SRC_PORT_STEP, type=float, dest='src_port_step', help='The portion step in source domain in every round of self-paced self-trained neural network') parser.add_argument('--randseed', default=RANDSEED, type=int, help='The random seed to sample the source dataset.') parser.add_argument("--src-sampling-policy", type=str, default=SRC_SAMPLING_POLICY, help="The sampling policy on source dataset: 'c' for 'cumulative' and 'r' for replace ") parser.add_argument('--mine-port', default=MINE_PORT, type=float, help='If a class has a predication portion lower than the mine_port, then mine the patches including the class in self-training.') parser.add_argument('--rare-cls-num', default=RARE_CLS_NUM, type=int, help='The number of classes to be mined.') parser.add_argument('--mine-chance', default=MINE_CHANCE, type=float, help='The chance of patch mining.') parser.add_argument('--rm-prob', help='If remove the probability maps generated in every round.', default=False, action='store_true') parser.add_argument('--mr-weight-kld', default=MRKLD, type=float, dest='mr_weight_kld', help='weight of kld model regularization') parser.add_argument('--lr-weight-ent', default=LRENT, type=float, dest='lr_weight_ent', help='weight of negative entropy label regularization') parser.add_argument('--mr-weight-src', default=MRSRC, type=float, dest='mr_weight_src', help='weight of regularization in source domain') return parser.parse_args() args = get_arguments() # palette if args.data_src == 'gta': # gta: palette = [128, 64, 128, 244, 35, 232, 70, 70, 70, 102, 102, 156, 190, 153, 153, 153, 153, 153, 250, 170, 30, 220, 220, 0, 107, 142, 35, 152, 251, 152, 70, 130, 180, 220, 20, 60, 255, 0, 0, 0, 0, 142, 0, 0, 70, 0, 60, 100, 0, 80, 100, 0, 0, 230, 119, 11, 32] zero_pad = 256 * 3 - len(palette) for i in range(zero_pad): palette.append(0) def colorize_mask(mask): # mask: numpy array of the mask new_mask = Image.fromarray(mask.astype(np.uint8)).convert('P') new_mask.putpalette(palette) return new_mask def main(): randseed = args.randseed seed_torch(randseed) device = torch.device("cuda:" + str(args.gpu)) save_path = args.save save_pseudo_label_path = osp.join(save_path, 'pseudo_label') # in 'save_path'. Save labelIDs, not trainIDs. save_stats_path = osp.join(save_path, 'stats') # in 'save_path' save_lst_path = osp.join(save_path, 'list') if not os.path.exists(save_path): os.makedirs(save_path) if not os.path.exists(save_pseudo_label_path): os.makedirs(save_pseudo_label_path) if not os.path.exists(save_stats_path): os.makedirs(save_stats_path) if not os.path.exists(save_lst_path): os.makedirs(save_lst_path) logger = util.set_logger(args.save, args.log_file, args.debug) logger.info('start with arguments %s', args) if args.model == 'DeeplabRes': model = Res_Deeplab(num_classes=args.num_classes) if args.restore_from[:4] == 'http' : saved_state_dict = model_zoo.load_url(args.restore_from) new_params = model.state_dict().copy() for i in saved_state_dict: # Scale.layer5.conv2d_list.3.weight i_parts = str(i).split('.') # print i_parts if not i_parts[0] == 'fc': new_params['.'.join(i_parts[0:])] = saved_state_dict[i] else: loc = "cuda:" + str(args.gpu) saved_state_dict = torch.load(args.restore_from,map_location=loc) new_params = saved_state_dict.copy() model.load_state_dict(new_params) image_src_list, _, label_src_list, src_num = parse_split_list(args.data_src_list) image_tgt_list, image_name_tgt_list, _, tgt_num = parse_split_list(args.data_tgt_train_list) _, _, _, test_num = parse_split_list(args.data_tgt_test_list) ## label mapping sys.path.insert(0, 'dataset/helpers') if args.data_src == 'synthia': from labels_cityscapes_synthia import id2label, trainId2label elif args.data_src == 'gta': from labels import id2label, trainId2label label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) valid_labels = sorted(set(id_2_label.ravel())) # portions tgt_portion = args.init_tgt_port src_portion = args.init_src_port # training crop size h, w = map(int, args.input_size.split(',')) input_size = (h, w) lscale_src, hscale_src = map(float, args.train_scale_src.split(',')) train_scale_src = (lscale_src, hscale_src) lscale_tgt, hscale_tgt = map(float, args.train_scale_tgt.split(',')) train_scale_tgt = (lscale_tgt, hscale_tgt) for round_idx in range(args.num_rounds): save_round_eval_path = osp.join(args.save,str(round_idx)) save_pseudo_label_color_path = osp.join(save_round_eval_path, 'pseudo_label_color') # in every 'save_round_eval_path' if not os.path.exists(save_round_eval_path): os.makedirs(save_round_eval_path) if not os.path.exists(save_pseudo_label_color_path): os.makedirs(save_pseudo_label_color_path) ########## pseudo-label generation if round_idx != args.num_rounds - 1: # evaluation & save confidence vectors conf_dict, pred_cls_num, save_prob_path, save_pred_path = val(model, device, save_round_eval_path, round_idx, tgt_num, label_2_id, valid_labels, args, logger) # class-balanced thresholds cls_thresh = kc_parameters(conf_dict, pred_cls_num, tgt_portion, round_idx, save_stats_path, args, logger) tgt_portion = min(tgt_portion + args.tgt_port_step, args.max_tgt_port) # pseudo-label maps generation label_selection(cls_thresh, tgt_num, image_name_tgt_list, id_2_label, round_idx, save_prob_path, save_pred_path, save_pseudo_label_path, save_pseudo_label_color_path, save_round_eval_path, args, logger) # save training list if args.src_sampling_policy == 'c': randseed = args.randseed elif args.src_sampling_policy == 'r': randseed += 1 src_train_lst, tgt_train_lst, src_num_sel = savelst_SrcTgt(src_portion, image_tgt_list, image_name_tgt_list, image_src_list, label_src_list, save_lst_path, save_pseudo_label_path, src_num, tgt_num, randseed, args) src_portion = min(src_portion + args.src_port_step, args.max_src_port) ########### model retraining # dataset epoch_per_round = args.epr # reg weights if args.mr_weight_kld == 0: reg_weight_tgt = 0.0 else: # currently only one kind of model regularizer is supported reg_weight_tgt = args.mr_weight_kld reg_weight_src = args.mr_weight_src ### patch mining params # no patch mining in src # patch mining in target rare_id = np.load(save_stats_path + '/rare_id_round' + str(round_idx) + '.npy') mine_id = np.load(save_stats_path + '/mine_id_round' + str(round_idx) + '.npy') mine_chance = args.mine_chance # dataloader if args.lr_weight_ent == 0.0: srctrainset = SrcSTDataSet(args.data_src_dir, src_train_lst, max_iters=src_num_sel,reg_weight=reg_weight_src,data_src=args.data_src, crop_size=input_size,scale=args.random_scale, mirror=args.random_mirror, train_scale=train_scale_src, mean=IMG_MEAN, std=IMG_STD) tgttrainset = GTA5StMineDataSet(args.data_tgt_dir, tgt_train_lst, pseudo_root=save_pseudo_label_path, max_iters=tgt_num,reg_weight=reg_weight_tgt,rare_id = rare_id, mine_id=mine_id, mine_chance = mine_chance, crop_size=input_size,scale=args.random_scale,data_src=args.data_src, mirror=args.random_mirror, train_scale=train_scale_tgt, mean=IMG_MEAN, std=IMG_STD) elif args.lr_weight_ent > 0.0: srctrainset = SoftSrcSTDataSet(args.data_src_dir, src_train_lst, max_iters=src_num_sel,reg_weight=reg_weight_src,data_src=args.data_src,num_classes=args.num_classes, crop_size=input_size,scale=args.random_scale, mirror=args.random_mirror, train_scale=train_scale_src, mean=IMG_MEAN, std=IMG_STD) tgttrainset = SoftGTA5StMineDataSet(args.data_tgt_dir, tgt_train_lst, pseudo_root=save_pseudo_label_path, max_iters=tgt_num,reg_weight=reg_weight_tgt,rare_id = rare_id, mine_id=mine_id, mine_chance = mine_chance, crop_size=input_size,scale=args.random_scale,data_src=args.data_src,num_classes=args.num_classes, mirror=args.random_mirror, train_scale=train_scale_tgt, mean=IMG_MEAN, std=IMG_STD) mixtrainset = torch.utils.data.ConcatDataset([srctrainset, tgttrainset]) mix_trainloader = torch.utils.data.DataLoader(mixtrainset, batch_size=args.batch_size, shuffle=True, num_workers=0, pin_memory=args.pin_memory) # optimizer tot_iter = np.ceil(float(src_num_sel + tgt_num) / args.batch_size) optimizer = optim.SGD([{'params': get_1x_lr_params_NOscale(model), 'lr': args.learning_rate}, {'params': get_10x_lr_params(model), 'lr': 10 * args.learning_rate}], lr=args.learning_rate, momentum=args.momentum, weight_decay=args.weight_decay) interp = nn.Upsample(size=input_size, mode='bilinear', align_corners=True) logger.info('###### Start model retraining dataset in round {}! ######'.format(round_idx)) # model if args.is_training: model.train() else: model.eval() start = timeit.default_timer() # cudnn cudnn.enabled = True # enable cudnn cudnn.benchmark = True # enable the inbuilt cudnn auto-tuner to find the best algorithm to use for your hardware. # start training for epoch in range(epoch_per_round): train(mix_trainloader, model, device, interp, optimizer, tot_iter, round_idx, epoch, args, logger) end = timeit.default_timer() logger.info('###### Finish model retraining dataset in round {}! Time cost: {:.2f} seconds. ######'.format(round_idx, end - start)) # test self-trained model in target domain test set tgt_set = 'test' test(model, device, save_round_eval_path, round_idx, tgt_set, test_num, args.data_tgt_test_list, label_2_id, valid_labels, args, logger) elif round_idx == args.num_rounds - 1: shutil.rmtree(save_pseudo_label_path) tgt_set = 'train' test(model, device, save_round_eval_path, round_idx, tgt_set, tgt_num, args.data_tgt_train_list, label_2_id, valid_labels, args, logger) tgt_set = 'test' test(model, device, save_round_eval_path, round_idx, tgt_set, test_num, args.data_tgt_test_list, label_2_id, valid_labels, args, logger) def val(model, device, save_round_eval_path, round_idx, tgt_num, label_2_id, valid_labels, args, logger): """Create the model and start the evaluation process.""" ## scorer scorer = ScoreUpdater(valid_labels, args.num_classes, tgt_num, logger) scorer.reset() h, w = map(int, args.test_image_size.split(',')) test_image_size = (h, w) test_size = ( int(h*args.eval_scale), int(w*args.eval_scale) ) ## test data loader testloader = data.DataLoader(GTA5TestDataSet(args.data_tgt_dir, args.data_tgt_train_list, test_size=test_size, test_scale=args.eval_scale, mean=IMG_MEAN, std=IMG_STD, scale=False, mirror=False), batch_size=1, shuffle=False, pin_memory=args.pin_memory) ## model for evaluation if args.eval_training: model.train() else: model.eval() # model.to(device) ## upsampling layer if version.parse(torch.__version__) >= version.parse('0.4.0'): interp = nn.Upsample(size=test_image_size, mode='bilinear', align_corners=True) else: interp = nn.Upsample(size=test_image_size, mode='bilinear') ## output of deeplab is logits, not probability softmax2d = nn.Softmax2d() ## output folder save_pred_vis_path = osp.join(save_round_eval_path, 'pred_vis') save_prob_path = osp.join(save_round_eval_path, 'prob') save_pred_path = osp.join(save_round_eval_path, 'pred') if not os.path.exists(save_pred_vis_path): os.makedirs(save_pred_vis_path) if not os.path.exists(save_prob_path): os.makedirs(save_prob_path) if not os.path.exists(save_pred_path): os.makedirs(save_pred_path) # saving output data conf_dict = {k: [] for k in range(args.num_classes)} pred_cls_num = np.zeros(args.num_classes) ## evaluation process logger.info('###### Start evaluating target domain train set in round {}! ######'.format(round_idx)) start_eval = time.time() with torch.no_grad(): for index, batch in enumerate(testloader): image, label, _, name = batch if args.model == 'DeeplabRes': output2 = model(image.to(device)) output = softmax2d(interp(output2)).cpu().data[0].numpy() if args.test_flipping: output2 = model(torch.from_numpy(image.numpy()[:,:,:,::-1].copy()).to(device)) output = 0.5 * ( output + softmax2d(interp(output2)).cpu().data[0].numpy()[:,:,::-1] ) output = output.transpose(1,2,0) amax_output = np.asarray(np.argmax(output, axis=2), dtype=np.uint8) conf = np.amax(output,axis=2) # score pred_label = amax_output.copy() label = label_2_id[np.asarray(label.numpy(), dtype=np.uint8)] scorer.update(pred_label.flatten(), label.flatten(), index) # save visualized seg maps & predication prob map amax_output_col = colorize_mask(amax_output) name = name[0].split('/')[-1] image_name = name.split('.')[0] # prob np.save('%s/%s.npy' % (save_prob_path, image_name), output) # trainIDs/vis seg maps amax_output = Image.fromarray(amax_output) amax_output.save('%s/%s.png' % (save_pred_path, image_name)) amax_output_col.save('%s/%s_color.png' % (save_pred_vis_path, image_name)) # save class-wise confidence maps if args.kc_value == 'conf': for idx_cls in range(args.num_classes): idx_temp = pred_label == idx_cls pred_cls_num[idx_cls] = pred_cls_num[idx_cls] + np.sum(idx_temp) if idx_temp.any(): conf_cls_temp = conf[idx_temp].astype(np.float32) len_cls_temp = conf_cls_temp.size # downsampling by ds_rate conf_cls = conf_cls_temp[0:len_cls_temp:args.ds_rate] conf_dict[idx_cls].extend(conf_cls) elif args.kc_value == 'prob': for idx_cls in range(args.num_classes): idx_temp = pred_label == idx_cls pred_cls_num[idx_cls] = pred_cls_num[idx_cls] + np.sum(idx_temp) # prob slice prob_cls_temp = output[:,:,idx_cls].astype(np.float32).ravel() len_cls_temp = prob_cls_temp.size # downsampling by ds_rate prob_cls = prob_cls_temp[0:len_cls_temp:args.ds_rate] conf_dict[idx_cls].extend(prob_cls) # it should be prob_dict; but for unification, use conf_dict logger.info('###### Finish evaluating target domain train set in round {}! Time cost: {:.2f} seconds. ######'.format(round_idx, time.time()-start_eval)) return conf_dict, pred_cls_num, save_prob_path, save_pred_path # return the dictionary containing all the class-wise confidence vectors def train(mix_trainloader, model, device, interp, optimizer, tot_iter, round_idx, epoch_idx, args, logger): """Create the model and start the training.""" for i_iter, batch in enumerate(mix_trainloader): images, labels, _, _, reg_weights = batch images = images.to(device) labels = labels.to(device) optimizer.zero_grad() adjust_learning_rate(optimizer, i_iter, tot_iter) pred = interp(model(images)) if args.lr_weight_ent == 0.0: loss = reg_loss_calc(pred, labels, reg_weights.to(device), args) if args.lr_weight_ent > 0.0: loss = reg_loss_calc_expand(pred, labels, reg_weights.to(device), args) loss.backward() optimizer.step() logger.info('iter = {} of {} completed, loss = {:.4f}'.format(i_iter+1, tot_iter, loss.data.cpu().numpy())) print('taking snapshot ...') torch.save(model.state_dict(), osp.join(args.save, args.data_src + '2city_round' + str(round_idx) + '_epoch' + str(epoch_idx+1) + '.pth')) def test(model, device, save_round_eval_path, round_idx, tgt_set, test_num, test_list, label_2_id, valid_labels, args, logger): """Create the model and start the evaluation process.""" ## scorer scorer = ScoreUpdater(valid_labels, args.num_classes, test_num, logger) scorer.reset() h, w = map(int, args.test_image_size.split(',')) test_image_size = (h, w) test_size = ( h, w ) test_scales = [float(_) for _ in str(args.test_scale).split(',')] num_scales = len(test_scales) ## test data loader testloader = data.DataLoader(GTA5TestDataSet(args.data_tgt_dir, test_list, test_size=test_size, test_scale=1.0, mean=IMG_MEAN, std=IMG_STD, scale=False, mirror=False), batch_size=1, shuffle=False, pin_memory=args.pin_memory) ## model for evaluation if args.eval_training: model.train() else: model.eval() # model.to(device) ## upsampling layer if version.parse(torch.__version__) >= version.parse('0.4.0'): interp = nn.Upsample(size=test_image_size, mode='bilinear', align_corners=True) else: interp = nn.Upsample(size=test_image_size, mode='bilinear') ## output of deeplab is logits, not probability softmax2d = nn.Softmax2d() ## output folder if tgt_set == 'train': save_test_vis_path = osp.join(save_round_eval_path, 'trainSet_vis') elif tgt_set == 'test': save_test_vis_path = osp.join(save_round_eval_path, 'testSet_vis') if not os.path.exists(save_test_vis_path): os.makedirs(save_test_vis_path) ## evaluation process logger.info('###### Start evaluating in target domain {} set in round {}! ######'.format(tgt_set, round_idx)) start_eval = time.time() with torch.no_grad(): for index, batch in enumerate(testloader): image, label, _, name = batch img = image.clone() for scale_idx in range(num_scales): if version.parse(torch.__version__) > version.parse('0.4.0'): image = F.interpolate(img, scale_factor=test_scales[scale_idx], mode='bilinear', align_corners=True) else: test_size = (int(h * test_scales[scale_idx]), int(w * test_scales[scale_idx])) interp_tmp = nn.Upsample(size=test_size, mode='bilinear', align_corners=True) image = interp_tmp(img) if args.model == 'DeeplabRes': output2 = model(image.to(device)) coutput = interp(output2).cpu().data[0].numpy() if args.test_flipping: output2 = model(torch.from_numpy(image.numpy()[:,:,:,::-1].copy()).to(device)) coutput = 0.5 * ( coutput + interp(output2).cpu().data[0].numpy()[:,:,::-1] ) if scale_idx == 0: output = coutput.copy() else: output = output+coutput output = output/num_scales output = output.transpose(1,2,0) amax_output = np.asarray(np.argmax(output, axis=2), dtype=np.uint8) # score pred_label = amax_output.copy() label = label_2_id[np.asarray(label.numpy(), dtype=np.uint8)] scorer.update(pred_label.flatten(), label.flatten(), index) # save visualized seg maps & predication prob map amax_output_col = colorize_mask(amax_output) name = name[0].split('/')[-1] image_name = name.split('.')[0] # vis seg maps amax_output_col.save('%s/%s_color.png' % (save_test_vis_path, image_name)) logger.info('###### Finish evaluating in target domain {} set in round {}! Time cost: {:.2f} seconds. ######'.format(tgt_set, round_idx, time.time()-start_eval)) return def kc_parameters(conf_dict, pred_cls_num, tgt_portion, round_idx, save_stats_path, args, logger): logger.info('###### Start kc generation in round {} ! ######'.format(round_idx)) start_kc = time.time() # threshold for each class conf_tot = 0.0 cls_thresh = np.ones(args.num_classes,dtype = np.float32) cls_sel_size = np.zeros(args.num_classes, dtype=np.float32) cls_size = np.zeros(args.num_classes, dtype=np.float32) if args.kc_policy == 'cb' and args.kc_value == 'conf': for idx_cls in np.arange(0, args.num_classes): cls_size[idx_cls] = pred_cls_num[idx_cls] if conf_dict[idx_cls] != None: conf_dict[idx_cls].sort(reverse=True) # sort in descending order len_cls = len(conf_dict[idx_cls]) cls_sel_size[idx_cls] = int(math.floor(len_cls * tgt_portion)) len_cls_thresh = int(cls_sel_size[idx_cls]) if len_cls_thresh != 0: cls_thresh[idx_cls] = conf_dict[idx_cls][len_cls_thresh-1] conf_dict[idx_cls] = None # threshold for mine_id with priority num_mine_id = len(np.nonzero(cls_size / np.sum(cls_size) < args.mine_port)[0]) # chose the smallest mine_id id_all = np.argsort(cls_size / np.sum(cls_size)) rare_id = id_all[:args.rare_cls_num] mine_id = id_all[:num_mine_id] # sort mine_id in ascending order w.r.t predication portions # save mine ids np.save(save_stats_path + '/rare_id_round' + str(round_idx) + '.npy', rare_id) np.save(save_stats_path + '/mine_id_round' + str(round_idx) + '.npy', mine_id) logger.info('Mining ids : {}! {} rarest ids: {}!'.format(mine_id,args.rare_cls_num,rare_id)) # save thresholds np.save(save_stats_path + '/cls_thresh_round' + str(round_idx) + '.npy', cls_thresh) np.save(save_stats_path + '/cls_sel_size_round' + str(round_idx) + '.npy', cls_sel_size) logger.info('###### Finish kc generation in round {}! Time cost: {:.2f} seconds. ######'.format(round_idx,time.time() - start_kc)) return cls_thresh def label_selection(cls_thresh, tgt_num, image_name_tgt_list, id_2_label, round_idx, save_prob_path, save_pred_path, save_pseudo_label_path, save_pseudo_label_color_path, save_round_eval_path, args, logger): logger.info('###### Start pseudo-label generation in round {} ! ######'.format(round_idx)) start_pl = time.time() for idx in range(tgt_num): sample_name = image_name_tgt_list[idx].split('.')[0] probmap_path = osp.join(save_prob_path, '{}.npy'.format(sample_name)) pred_path = osp.join(save_pred_path, '{}.png'.format(sample_name)) pred_prob = np.load(probmap_path) pred_label_trainIDs = np.asarray(Image.open(pred_path)) pred_label_labelIDs = id_2_label[pred_label_trainIDs] pred_label_trainIDs = pred_label_trainIDs.copy() if args.kc_policy == 'cb' and args.lr_weight_ent == 0.0: save_wpred_vis_path = osp.join(save_round_eval_path, 'weighted_pred_vis') if not os.path.exists(save_wpred_vis_path): os.makedirs(save_wpred_vis_path) weighted_prob = pred_prob/cls_thresh weighted_pred_trainIDs = np.asarray(np.argmax(weighted_prob, axis=2), dtype=np.uint8) # save weighted predication wpred_label_col = weighted_pred_trainIDs.copy() wpred_label_col = colorize_mask(wpred_label_col) wpred_label_col.save('%s/%s_color.png' % (save_wpred_vis_path, sample_name)) weighted_conf = np.amax(weighted_prob, axis=2) pred_label_trainIDs = weighted_pred_trainIDs.copy() pred_label_labelIDs = id_2_label[pred_label_trainIDs] pred_label_labelIDs[weighted_conf < 1] = 0 # '0' in cityscapes indicates 'unlabaled' for labelIDs pred_label_trainIDs[weighted_conf < 1] = 255 # '255' in cityscapes indicates 'unlabaled' for trainIDs elif args.kc_policy == 'cb' and args.lr_weight_ent > 0.0: # check if cb can be combined with kc_value == conf or prob; also check if \alpha can be larger than 1 save_wpred_vis_path = osp.join(save_round_eval_path, 'weighted_pred_vis') if not os.path.exists(save_wpred_vis_path): os.makedirs(save_wpred_vis_path) # soft pseudo-label soft_pseudo_label = np.power(pred_prob/cls_thresh,1.0/args.lr_weight_ent) # weighted softmax with temperature soft_pseudo_label_sum = soft_pseudo_label.sum(2) soft_pseudo_label = soft_pseudo_label.transpose(2,0,1)/soft_pseudo_label_sum soft_pseudo_label = soft_pseudo_label.transpose(1,2,0).astype(np.float32) np.save('%s/%s.npy' % (save_pseudo_label_path, sample_name), soft_pseudo_label) # hard pseudo-label weighted_pred_trainIDs = np.asarray(np.argmax(soft_pseudo_label, axis=2), dtype=np.uint8) reg_score = np.sum( -soft_pseudo_label*np.log(pred_prob+1e-32) + args.lr_weight_ent*soft_pseudo_label*np.log(soft_pseudo_label+1e-32), axis=2) sel_score = np.sum( -soft_pseudo_label*np.log(cls_thresh+1e-32), axis=2) # save weighted predication wpred_label_col = weighted_pred_trainIDs.copy() wpred_label_col = colorize_mask(wpred_label_col) wpred_label_col.save('%s/%s_color.png' % (save_wpred_vis_path, sample_name)) pred_label_trainIDs = weighted_pred_trainIDs.copy() pred_label_labelIDs = id_2_label[pred_label_trainIDs] pred_label_labelIDs[reg_score >= sel_score] = 0 # '0' in cityscapes indicates 'unlabaled' for labelIDs pred_label_trainIDs[reg_score >= sel_score] = 255 # '255' in cityscapes indicates 'unlabaled' for trainIDs # pseudo-labels with labelID pseudo_label_labelIDs = pred_label_labelIDs.copy() pseudo_label_trainIDs = pred_label_trainIDs.copy() # save colored pseudo-label map pseudo_label_col = colorize_mask(pseudo_label_trainIDs) pseudo_label_col.save('%s/%s_color.png' % (save_pseudo_label_color_path, sample_name)) # save pseudo-label map with label IDs pseudo_label_save = Image.fromarray(pseudo_label_labelIDs.astype(np.uint8)) pseudo_label_save.save('%s/%s.png' % (save_pseudo_label_path, sample_name)) # remove probability maps if args.rm_prob: shutil.rmtree(save_prob_path) logger.info('###### Finish pseudo-label generation in round {}! Time cost: {:.2f} seconds. ######'.format(round_idx,time.time() - start_pl)) def parse_split_list(list_name): image_list = [] image_name_list = [] label_list = [] file_num = 0 with open(list_name) as f: for item in f.readlines(): fields = item.strip().split('\t') image_name = fields[0].split('/')[-1] image_list.append(fields[0]) image_name_list.append(image_name) label_list.append(fields[1]) file_num += 1 return image_list, image_name_list, label_list, file_num def savelst_SrcTgt(src_portion, image_tgt_list, image_name_tgt_list, image_src_list, label_src_list, save_lst_path, save_pseudo_label_path, src_num, tgt_num, randseed, args): src_num_sel = int(np.floor(src_num*src_portion)) np.random.seed(randseed) sel_idx = list( np.random.choice(src_num, src_num_sel, replace=False) ) sel_src_img_list = list( itemgetter(*sel_idx)(image_src_list) ) sel_src_label_list = list(itemgetter(*sel_idx)(label_src_list)) src_train_lst = osp.join(save_lst_path,'src_train.lst') tgt_train_lst = osp.join(save_lst_path, 'tgt_train.lst') # generate src train list with open(src_train_lst, 'w') as f: for idx in range(src_num_sel): f.write("%s\t%s\n" % (sel_src_img_list[idx], sel_src_label_list[idx])) # generate tgt train list if args.lr_weight_ent > 0: with open(tgt_train_lst, 'w') as f: for idx in range(tgt_num): softlabel_name = image_name_tgt_list[idx].split('.')[0] + '.npy' soft_label_tgt_path = osp.join(save_pseudo_label_path, softlabel_name) image_tgt_path = osp.join(save_pseudo_label_path,image_name_tgt_list[idx]) f.write("%s\t%s\t%s\n" % (image_tgt_list[idx], image_tgt_path, soft_label_tgt_path)) elif args.lr_weight_ent == 0: with open(tgt_train_lst, 'w') as f: for idx in range(tgt_num): image_tgt_path = osp.join(save_pseudo_label_path,image_name_tgt_list[idx]) f.write("%s\t%s\n" % (image_tgt_list[idx], image_tgt_path)) return src_train_lst, tgt_train_lst, src_num_sel class ScoreUpdater(object): # only IoU are computed. accu, cls_accu, etc are ignored. def __init__(self, valid_labels, c_num, x_num, logger=None, label=None, info=None): self._valid_labels = valid_labels self._confs = np.zeros((c_num, c_num)) self._per_cls_iou = np.zeros(c_num) self._logger = logger self._label = label self._info = info self._num_class = c_num self._num_sample = x_num @property def info(self): return self._info def reset(self): self._start = time.time() self._computed = np.zeros(self._num_sample) # one-dimension self._confs[:] = 0 def fast_hist(self,label, pred_label, n): k = (label >= 0) & (label < n) return np.bincount(n * label[k].astype(int) + pred_label[k], minlength=n ** 2).reshape(n, n) def per_class_iu(self,hist): return np.diag(hist) / (hist.sum(1) + hist.sum(0) - np.diag(hist)) def do_updates(self, conf, i, computed=True): if computed: self._computed[i] = 1 self._per_cls_iou = self.per_class_iu(conf) def update(self, pred_label, label, i, computed=True): conf = self.fast_hist(label, pred_label, self._num_class) self._confs += conf self.do_updates(self._confs, i, computed) self.scores(i) def scores(self, i=None, logger=None): x_num = self._num_sample ious = np.nan_to_num( self._per_cls_iou ) logger = self._logger if logger is None else logger if logger is not None: if i is not None: speed = 1. * self._computed.sum() / (time.time() - self._start) logger.info('Done {}/{} with speed: {:.2f}/s'.format(i + 1, x_num, speed)) name = '' if self._label is None else '{}, '.format(self._label) logger.info('{}mean iou: {:.2f}%'. \ format(name, np.mean(ious) * 100)) with util.np_print_options(formatter={'float': '{:5.2f}'.format}): logger.info('\n{}'.format(ious * 100)) return ious def loss_calc(pred, label): """ This function returns cross entropy loss for semantic segmentation """ # out shape batch_size x channels x h x w -> batch_size x channels x h x w # label shape h x w x 1 x batch_size -> batch_size x 1 x h x w criterion = torch.nn.CrossEntropyLoss(ignore_index=IGNORE_LABEL).cuda() return criterion(pred, label) def reg_loss_calc(pred, label, reg_weights, args): """ This function returns cross entropy loss for semantic segmentation """ # out shape batch_size x channels x h x w -> batch_size x channels x h x w # label shape h x w x 1 x batch_size -> batch_size x 1 x h x w mr_weight_kld = args.mr_weight_kld num_class = float(args.num_classes) valid_num = torch.sum(label != IGNORE_LABEL).float() label_reg = label[reg_weights != 0,:,:] valid_reg_num = torch.sum(label_reg != IGNORE_LABEL).float() softmax = F.softmax(pred, dim=1) # compute the softmax values logsoftmax = F.log_softmax(pred,dim=1) # compute the log of softmax values label_expand = torch.unsqueeze(label, 1).repeat(1,int(num_class),1,1) labels = label_expand.clone() labels[labels != IGNORE_LABEL] = 1.0 labels[labels == IGNORE_LABEL] = 0.0 labels_valid = labels.clone() # labels = torch.unsqueeze(labels, 1).repeat(1,num_class,1,1) labels = torch.cumsum(labels, dim=1) labels[labels != label_expand + 1] = 0.0 del label_expand labels[labels != 0 ] = 1.0 ### check the vectorized labels # check_labels = torch.argmax(labels, dim=1) # label[label == 255] = 0 # print(torch.sum(check_labels.float() - label)) reg_weights = reg_weights.float().view(len(reg_weights),1,1,1) ce = torch.sum( -logsoftmax*labels ) # cross-entropy loss with vector-form softmax softmax_val = softmax*labels_valid logsoftmax_val = logsoftmax*labels_valid kld = torch.sum( -logsoftmax_val/num_class*reg_weights ) if valid_reg_num > 0: reg_ce = ce/valid_num + (mr_weight_kld*kld)/valid_reg_num else: reg_ce = ce/valid_num return reg_ce def reg_loss_calc_expand(pred, label, reg_weights, args): """ This function returns cross entropy loss for semantic segmentation """ # out shape batch_size x channels x h x w -> batch_size x channels x h x w # label shape h x w x 1 x batch_size -> batch_size x 1 x h x w mr_weight_kld = args.mr_weight_kld num_class = float(args.num_classes) # soft labels regard ignored labels as zero soft labels in data loader # C = label.cpu().numpy() label_sum = torch.sum(label,1) # D = label_sum.cpu().numpy() valid_num = torch.sum(label_sum != 0.0).float() label_reg = label_sum[reg_weights != 0,:,:] valid_reg_num = torch.sum(label_reg != 0.0).float() softmax = F.softmax(pred, dim=1) # compute the softmax values logsoftmax = F.log_softmax(pred,dim=1) # compute the log of softmax values label_expand = torch.unsqueeze(label_sum, 1).repeat(1,num_class,1,1) label_valid = label_expand.clone() label_valid[label_valid != 0] = 1.0 label_valid = label_valid.clone() # # check the vectorized labels # check_labels = torch.argmax(labels, dim=1) # label[label == 255] = 0 # print(torch.sum(check_labels.float() - label)) # reg_weights = reg_weights.float().view(len(reg_weights),1,1,1) ce = torch.sum( -logsoftmax*label ) # cross-entropy loss with vector-form softmax softmax_val = softmax*label_valid logsoftmax_val = logsoftmax*label_valid kld = torch.sum( -logsoftmax_val/num_class*reg_weights ) if valid_reg_num > 0: reg_ce = ce/valid_num + (mr_weight_kld*kld)/valid_reg_num else: reg_ce = ce/valid_num return reg_ce def lr_poly(base_lr, iter, max_iter, power): return base_lr * ((1 - float(iter) / max_iter) ** (power)) def get_1x_lr_params_NOscale(model): """ This generator returns all the parameters of the net except for the last classification layer. Note that for each batchnorm layer, requires_grad is set to False in deeplab_resnet.py, therefore this function does not return any batchnorm parameter """ b = [] b.append(model.conv1) b.append(model.bn1) b.append(model.layer1) b.append(model.layer2) b.append(model.layer3) b.append(model.layer4) for i in range(len(b)): for j in b[i].modules(): jj = 0 for k in j.parameters(): jj += 1 if k.requires_grad: yield k def get_10x_lr_params(model): """ This generator returns all the parameters for the last layer of the net, which does the classification of pixel into classes """ b = [] b.append(model.layer5.parameters()) for j in range(len(b)): for i in b[j]: yield i def adjust_learning_rate(optimizer, i_iter, tot_iter): lr = lr_poly(args.learning_rate, i_iter, tot_iter, args.power) optimizer.param_groups[0]['lr'] = lr optimizer.param_groups[1]['lr'] = lr * 10 if __name__ == '__main__': main()
47,013
48.229319
225
py
CRST
CRST-master/train.py
import argparse import torch import torch.nn as nn from torch.utils import data import numpy as np import pickle import cv2 import torch.optim as optim import scipy.misc import torch.backends.cudnn as cudnn import sys import os import os.path as osp from deeplab.model import Res_Deeplab from deeplab.loss import CrossEntropy2d from deeplab.datasets import GTA5DataSet import matplotlib.pyplot as plt import random import timeit import torchvision.transforms as transforms import util start = timeit.default_timer() #IMG_MEAN = np.array((104.00698793,116.66876762,122.67891434), dtype=np.float32) IMG_MEAN = np.array((0.406, 0.456, 0.485), dtype=np.float32) # BGR IMG_STD = np.array((0.225, 0.224, 0.229), dtype=np.float32) # BGR # IMG_MEAN = [0.485, 0.456, 0.406] # IMG_STD = [0.229, 0.224, 0.225] BATCH_SIZE = 4 DATA_DIRECTORY = './datasets/gta5' DATA_LIST_PATH = './dataset/list/gta5/train.lst' NUM_CLASSES = 19 IGNORE_LABEL = 255 INPUT_SIZE = '500,500' TRAIN_SCALE = '0.5,1.5' LEARNING_RATE = 2.5e-4 MOMENTUM = 0.9 NUM_STEPS = 100000 POWER = 0.9 RANDOM_SEED = 1234 RESTORE_FROM = '' SAVE_NUM_IMAGES = 2 SAVE_PRED_EVERY = 5000 SNAPSHOT_DIR = './gta_src_train/' WEIGHT_DECAY = 0.0005 MODEL = 'DeeplabRes101' LOG_FILE = 'log' PIN_MEMORY = True GPU = '0' def get_arguments(): """Parse all the arguments provided from the CLI. Returns: A list of parsed arguments. """ parser = argparse.ArgumentParser(description="DeepLab-ResNet Network") parser.add_argument("--batch-size", type=int, default=BATCH_SIZE, help="Number of images sent to the network in one step.") parser.add_argument("--data-dir", type=str, default=DATA_DIRECTORY, help="Path to the directory containing the PASCAL VOC dataset.") parser.add_argument("--data-list", type=str, default=DATA_LIST_PATH, help="Path to the file listing the images in the dataset.") parser.add_argument("--ignore-label", type=int, default=IGNORE_LABEL, help="The index of the label to ignore during the training.") parser.add_argument("--input-size", type=str, default=INPUT_SIZE, help="Comma-separated string with height and width of images.") parser.add_argument("--learning-rate", type=float, default=LEARNING_RATE, help="Base learning rate for training with polynomial decay.") parser.add_argument("--momentum", type=float, default=MOMENTUM, help="Momentum component of the optimiser.") parser.add_argument("--model", type=str, default=MODEL, help="The base network.") parser.add_argument("--not-restore-last", action="store_true", help="Whether to not restore last (FC) layers.") parser.add_argument("--num-classes", type=int, default=NUM_CLASSES, help="Number of classes to predict (including background).") parser.add_argument("--num-steps", type=int, default=NUM_STEPS, help="Number of training steps.") parser.add_argument("--power", type=float, default=POWER, help="Decay parameter to compute the learning rate.") parser.add_argument("--random-mirror", action="store_true", help="Whether to randomly mirror the inputs during the training.") parser.add_argument("--random-scale", action="store_true", help="Whether to randomly scale the inputs during the training.") parser.add_argument("--train-scale", type=str, default=TRAIN_SCALE, help="The scale for multi-scale training.") parser.add_argument("--random-seed", type=int, default=RANDOM_SEED, help="Random seed to have reproducible results.") parser.add_argument("--restore-from", type=str, default=RESTORE_FROM, help="Where restore model parameters from.") parser.add_argument("--save-num-images", type=int, default=SAVE_NUM_IMAGES, help="How many images to save.") parser.add_argument("--save-pred-every", type=int, default=SAVE_PRED_EVERY, help="Save summaries and checkpoint every often.") parser.add_argument("--snapshot-dir", type=str, default=SNAPSHOT_DIR, help="Where to save snapshots of the model.") parser.add_argument("--weight-decay", type=float, default=WEIGHT_DECAY, help="Regularisation parameter for L2-loss.") parser.add_argument("--gpu", type=str, default=GPU, help="choose gpu device.") parser.add_argument("--pin-memory", type=bool, default=PIN_MEMORY, help="Whether to pin memory in train & eval.") parser.add_argument("--log-file", type=str, default=LOG_FILE, help="The name of log file.") parser.add_argument('--debug',help='True means logging debug info.', default=False, action='store_true') return parser.parse_args() args = get_arguments() def loss_calc(pred, label): """ This function returns cross entropy loss for semantic segmentation """ # out shape batch_size x channels x h x w -> batch_size x channels x h x w # label shape h x w x 1 x batch_size -> batch_size x 1 x h x w criterion = torch.nn.CrossEntropyLoss(ignore_index=IGNORE_LABEL).cuda() return criterion(pred, label) def lr_poly(base_lr, iter, max_iter, power): return base_lr*((1-float(iter)/max_iter)**(power)) def get_1x_lr_params_NOscale(model): """ This generator returns all the parameters of the net except for the last classification layer. Note that for each batchnorm layer, requires_grad is set to False in deeplab_resnet.py, therefore this function does not return any batchnorm parameter """ b = [] b.append(model.conv1) b.append(model.bn1) b.append(model.layer1) b.append(model.layer2) b.append(model.layer3) b.append(model.layer4) for i in range(len(b)): for j in b[i].modules(): jj = 0 for k in j.parameters(): jj+=1 if k.requires_grad: yield k def get_10x_lr_params(model): """ This generator returns all the parameters for the last layer of the net, which does the classification of pixel into classes """ b = [] b.append(model.layer5.parameters()) for j in range(len(b)): for i in b[j]: yield i def adjust_learning_rate(optimizer, i_iter): """Sets the learning rate to the initial LR divided by 5 at 60th, 120th and 160th epochs""" lr = lr_poly(args.learning_rate, i_iter, args.num_steps, args.power) optimizer.param_groups[0]['lr'] = lr optimizer.param_groups[1]['lr'] = lr * 10 def main(): """Create the model and start the training.""" if not os.path.exists(args.snapshot_dir): os.makedirs(args.snapshot_dir) logger = util.set_logger(args.snapshot_dir, args.log_file, args.debug) logger.info('start with arguments %s', args) h, w = map(int, args.input_size.split(',')) input_size = (h, w) lscale, hscale = map(float, args.train_scale.split(',')) train_scale = (lscale, hscale) cudnn.enabled = True # Create network. model = Res_Deeplab(num_classes=args.num_classes) #saved_state_dict = torch.load(args.restore_from) #new_params = model.state_dict().copy() #for i in saved_state_dict: # #Scale.layer5.conv2d_list.3.weight # i_parts = i.split('.') # # print i_parts # if not args.num_classes == 21 or not i_parts[1]=='layer5': # new_params['.'.join(i_parts[1:])] = saved_state_dict[i] model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', } saved_state_dict = torch.utils.model_zoo.load_url(model_urls['resnet101']) # coco pretrained parameters: # saved_state_dict = torch.load(args.restore_from) new_params = model.state_dict().copy() for i in saved_state_dict: #Scale.layer5.conv2d_list.3.weight i_parts = str(i).split('.') # print i_parts if not i_parts[0]=='fc': new_params['.'.join(i_parts[0:])] = saved_state_dict[i] model.load_state_dict(new_params) #model.float() model.eval() # use_global_stats = True #model.train() device = torch.device("cuda:" + str(args.gpu)) model.to(device) cudnn.benchmark = True trainloader = data.DataLoader(GTA5DataSet(args.data_dir, args.data_list, max_iters=args.num_steps*args.batch_size, crop_size=input_size,train_scale=train_scale, scale=args.random_scale, mirror=args.random_mirror, mean=IMG_MEAN, std = IMG_STD), batch_size=args.batch_size, shuffle=True, num_workers=5, pin_memory=args.pin_memory) optimizer = optim.SGD([{'params': get_1x_lr_params_NOscale(model), 'lr': args.learning_rate }, {'params': get_10x_lr_params(model), 'lr': 10*args.learning_rate}], lr=args.learning_rate, momentum=args.momentum,weight_decay=args.weight_decay) optimizer.zero_grad() interp = nn.Upsample(size=input_size, mode='bilinear', align_corners=True) for i_iter, batch in enumerate(trainloader): images, labels, _, _ = batch images = images.to(device) labels = labels.long().to(device) optimizer.zero_grad() adjust_learning_rate(optimizer, i_iter) pred = interp(model(images)) loss = loss_calc(pred, labels) loss.backward() optimizer.step() # print('iter = ', i_iter, 'of', args.num_steps,'completed, loss = ', loss.data.cpu().numpy()) logger.info('iter = {} of {} completed, loss = {:.4f}'.format(i_iter,args.num_steps,loss.data.cpu().numpy())) if i_iter >= args.num_steps-1: print('save model ...') torch.save(model.state_dict(),osp.join(args.snapshot_dir, 'VOC12_scenes_'+str(args.num_steps)+'.pth')) break if i_iter % args.save_pred_every == 0 and i_iter!=0: print('taking snapshot ...') torch.save(model.state_dict(),osp.join(args.snapshot_dir, 'VOC12_scenes_'+str(i_iter)+'.pth')) end = timeit.default_timer() print(end-start,'seconds') if __name__ == '__main__': main()
10,806
39.324627
164
py
CRST
CRST-master/deeplab/loss.py
import torch import torch.nn.functional as F import torch.nn as nn from torch.autograd import Variable class CrossEntropy2d(nn.Module): def __init__(self, size_average=True, ignore_label=255): super(CrossEntropy2d, self).__init__() self.size_average = size_average self.ignore_label = ignore_label def forward(self, predict, target, weight=None): """ Args: predict:(n, c, h, w) target:(n, h, w) weight (Tensor, optional): a manual rescaling weight given to each class. If given, has to be a Tensor of size "nclasses" """ assert not target.requires_grad assert predict.dim() == 4 assert target.dim() == 3 assert predict.size(0) == target.size(0), "{0} vs {1} ".format(predict.size(0), target.size(0)) assert predict.size(2) == target.size(1), "{0} vs {1} ".format(predict.size(2), target.size(1)) assert predict.size(3) == target.size(2), "{0} vs {1} ".format(predict.size(3), target.size(3)) n, c, h, w = predict.size() target_mask = (target >= 0) * (target != self.ignore_label) target = target[target_mask] if not target.data.dim(): return Variable(torch.zeros(1)) predict = predict.transpose(1, 2).transpose(2, 3).contiguous() predict = predict[target_mask.view(n, h, w, 1).repeat(1, 1, 1, c)].view(-1, c) loss = F.cross_entropy(predict, target, weight=weight, size_average=self.size_average) return loss
1,585
44.314286
103
py
CRST
CRST-master/deeplab/model.py
import torch.nn as nn import math import torch.utils.model_zoo as model_zoo import torch import numpy as np affine_par = True def outS(i): i = int(i) i = (i+1)/2 i = int(np.ceil((i+1)/2.0)) i = (i+1)/2 return i def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, affine = affine_par) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, affine = affine_par) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, stride=stride, bias=False) # change self.bn1 = nn.BatchNorm2d(planes,affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False padding = dilation self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, # change padding=padding, bias=False, dilation = dilation) self.bn2 = nn.BatchNorm2d(planes,affine = affine_par) for i in self.bn2.parameters(): i.requires_grad = False self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * 4, affine = affine_par) for i in self.bn3.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Classifier_Module(nn.Module): def __init__(self, dilation_series, padding_series, num_classes): super(Classifier_Module, self).__init__() self.conv2d_list = nn.ModuleList() for dilation, padding in zip(dilation_series, padding_series): self.conv2d_list.append(nn.Conv2d(2048, num_classes, kernel_size=3, stride=1, padding=padding, dilation=dilation, bias = True)) for m in self.conv2d_list: m.weight.data.normal_(0, 0.01) def forward(self, x): out = self.conv2d_list[0](x) for i in range(len(self.conv2d_list)-1): out += self.conv2d_list[i+1](x) return out class Residual_Covolution(nn.Module): def __init__(self, icol, ocol, num_classes): super(Residual_Covolution, self).__init__() self.conv1 = nn.Conv2d(icol, ocol, kernel_size=3, stride=1, padding=12, dilation=12, bias=True) self.conv2 = nn.Conv2d(ocol, num_classes, kernel_size=3, stride=1, padding=12, dilation=12, bias=True) self.conv3 = nn.Conv2d(num_classes, ocol, kernel_size=1, stride=1, padding=0, dilation=1, bias=True) self.conv4 = nn.Conv2d(ocol, icol, kernel_size=1, stride=1, padding=0, dilation=1, bias=True) self.relu = nn.ReLU(inplace=True) def forward(self, x): dow1 = self.conv1(x) dow1 = self.relu(dow1) seg = self.conv2(dow1) inc1 = self.conv3(seg) add1 = dow1 + self.relu(inc1) inc2 = self.conv4(add1) out = x + self.relu(inc2) return out, seg class Residual_Refinement_Module(nn.Module): def __init__(self, num_classes): super(Residual_Refinement_Module, self).__init__() self.RC1 = Residual_Covolution(2048, 512, num_classes) self.RC2 = Residual_Covolution(2048, 512, num_classes) def forward(self, x): x, seg1 = self.RC1(x) _, seg2 = self.RC2(x) return [seg1, seg1+seg2] class ResNet_Refine(nn.Module): def __init__(self, block, layers, num_classes): self.inplanes = 64 super(ResNet_Refine, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=True) # change self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=1, dilation=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation=4) self.layer5 = Residual_Refinement_Module(num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() # for i in m.parameters(): # i.requires_grad = False def _make_layer(self, block, planes, blocks, stride=1, dilation=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion or dilation == 2 or dilation == 4: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion,affine = affine_par)) for i in downsample._modules['1'].parameters(): i.requires_grad = False layers = [] layers.append(block(self.inplanes, planes, stride,dilation=dilation, downsample=downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, dilation=dilation)) return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.layer5(x) return x class ResNet(nn.Module): def __init__(self, block, layers, num_classes): self.inplanes = 64 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=True) # change self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=1, dilation=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation=4) self.layer5 = self._make_pred_layer(Classifier_Module, [6,12,18,24],[6,12,18,24],num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() # for i in m.parameters(): # i.requires_grad = False def _make_layer(self, block, planes, blocks, stride=1, dilation=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion or dilation == 2 or dilation == 4: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion,affine = affine_par)) for i in downsample._modules['1'].parameters(): i.requires_grad = False layers = [] layers.append(block(self.inplanes, planes, stride,dilation=dilation, downsample=downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, dilation=dilation)) return nn.Sequential(*layers) def _make_pred_layer(self,block, dilation_series, padding_series,num_classes): return block(dilation_series,padding_series,num_classes) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.layer5(x) return x class MS_Deeplab(nn.Module): def __init__(self,block,num_classes): super(MS_Deeplab,self).__init__() self.Scale = ResNet(block,[3, 4, 23, 3],num_classes) #changed to fix #4 def forward(self,x): output = self.Scale(x) # for original scale output_size = output.size()[2] input_size = x.size()[2] self.interp1 = nn.Upsample(size=(int(input_size*0.75)+1, int(input_size*0.75)+1), mode='bilinear') self.interp2 = nn.Upsample(size=(int(input_size*0.5)+1, int(input_size*0.5)+1), mode='bilinear') self.interp3 = nn.Upsample(size=(output_size, output_size), mode='bilinear') x75 = self.interp1(x) output75 = self.interp3(self.Scale(x75)) # for 0.75x scale x5 = self.interp2(x) output5 = self.interp3(self.Scale(x5)) # for 0.5x scale out_max = torch.max(torch.max(output, output75), output5) return [output, output75, output5, out_max] def Res_Ms_Deeplab(num_classes=21): model = MS_Deeplab(Bottleneck, num_classes) return model def Res_Deeplab(num_classes=21, is_refine=False): if is_refine: model = ResNet_Refine(Bottleneck,[3, 4, 23, 3], num_classes) else: model = ResNet(Bottleneck,[3, 4, 23, 3], num_classes) return model
11,127
36.217391
139
py
CRST
CRST-master/deeplab/datasets.py
import os import os.path as osp import numpy as np import random import matplotlib.pyplot as plt import collections import torch import torchvision.transforms as transforms import torchvision import cv2 from torch.utils import data import sys from PIL import Image palette = [128, 64, 128, 244, 35, 232, 70, 70, 70, 102, 102, 156, 190, 153, 153, 153, 153, 153, 250, 170, 30, 220, 220, 0, 107, 142, 35, 152, 251, 152, 70, 130, 180, 220, 20, 60, 255, 0, 0, 0, 0, 142, 0, 0, 70, 0, 60, 100, 0, 80, 100, 0, 0, 230, 119, 11, 32] zero_pad = 256 * 3 - len(palette) for i in range(zero_pad): palette.append(0) def colorize_mask(mask): # mask: numpy array of the mask new_mask = Image.fromarray(mask.astype(np.uint8)).convert('P') new_mask.putpalette(palette) return new_mask class VOCDataSet(data.Dataset): def __init__(self, root, list_path, max_iters=None, crop_size=(321, 321), mean=(128, 128, 128), scale=True, mirror=True, ignore_label=255): self.root = root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.scale = scale self.ignore_label = ignore_label self.mean = mean self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [i_id.strip() for i_id in open(list_path)] if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for name in self.img_ids: img_file = osp.join(self.root, "img/%s.jpg" % name) label_file = osp.join(self.root, "gt/%s.png" % name) self.files.append({ "img": img_file, "label": label_file, "name": name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): f_scale = 0.5 + random.randint(0, 11) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) label = cv2.imread(datafiles["label"], cv2.IMREAD_GRAYSCALE) size = image.shape name = datafiles["name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image -= self.mean img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) #image = image[:, :, ::-1] # change to BGR image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), name class GTA5DataSet(data.Dataset): def __init__(self, root, list_path, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') from labels import id2label, trainId2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # valid_labels = sorted(set(id_2_label.ravel())) label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), img_name class SYNTHIADataSet(data.Dataset): def __init__(self, root, list_path, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') from labels_synthia import id2label, trainId2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # valid_labels = sorted(set(id_2_label.ravel())) label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), img_name class SYNTHIASTDataSet(data.Dataset): def __init__(self, root, list_path, reg_weight = 0.0, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] self.reg_weight = reg_weight with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') from labels_synthia import id2label, trainId2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # valid_labels = sorted(set(id_2_label.ravel())) label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), img_name, self.reg_weight class GTA5STDataSet(data.Dataset): def __init__(self, root, list_path, reg_weight = 0.0, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] self.reg_weight = reg_weight with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') from labels import id2label, trainId2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # valid_labels = sorted(set(id_2_label.ravel())) label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), img_name, self.reg_weight class SrcSTDataSet(data.Dataset): def __init__(self, root, list_path, data_src=None, reg_weight = 0.0, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] self.reg_weight = reg_weight self.data_src = data_src with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') if self.data_src == 'gta': from labels import id2label elif self.data_src == 'synthia': from labels_synthia import id2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # valid_labels = sorted(set(id_2_label.ravel())) label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), img_name, self.reg_weight class SoftSrcSTDataSet(data.Dataset): def __init__(self, root, list_path, data_src = None, num_classes = None, reg_weight = 0.0, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] self.reg_weight = reg_weight self.data_src = data_src self.num_classes = num_classes with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') if self.data_src == 'gta': from labels import id2label elif self.data_src == 'synthia': from labels_synthia import id2label # label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # valid_labels = sorted(set(id_2_label.ravel())) label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) # roi = cv2.Rect(w_off, h_off, self.crop_w, self.crop_h); image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] label_expand = np.tile( np.expand_dims(label, axis=2), (1, 1, self.num_classes) ) labels = label_expand.copy() labels[labels != self.ignore_label] = 1.0 labels[labels == self.ignore_label] = 0.0 labels = np.cumsum(labels, axis=2) labels[labels != label_expand + 1] = 0.0 del label_expand labels[labels != 0.0] = 1.0 labels = labels.transpose((2,0,1)) # weighted_pred_trainIDs = np.asarray(np.argmax(labels, axis=0), dtype=np.uint8) # # save weighted predication # wpred_label_col = weighted_pred_trainIDs.copy() # wpred_label_col = colorize_mask(wpred_label_col) # wpred_label_col.save('%s_color.png' % (index)) # # labels_sum = np.sum(labels,axis=0) # # save weighted predication # weighted_pred_trainIDs[labels_sum == 0] = 255 # wpred_label_col = weighted_pred_trainIDs.copy() # wpred_label_col = colorize_mask(wpred_label_col) # wpred_label_col.save('%s_pseudo_color.png' % (index)) return image.copy(), labels.copy(), np.array(size), img_name, self.reg_weight class SoftGTA5StMineDataSet(data.Dataset): def __init__(self, root, list_path, data_src=None, num_classes = None, reg_weight = 0.0, rare_id = None, mine_id = None, mine_chance = None, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] self.softlabel_ids = [] self.reg_weight = reg_weight self.rare_id = rare_id self.mine_id = mine_id self.mine_chance = mine_chance self.data_src = data_src self.num_classes = num_classes with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) self.softlabel_ids.append(fields[2]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.softlabel_ids = self.softlabel_ids * int(np.ceil(float(max_iters) / len(self.softlabel_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] softlabel_name = self.softlabel_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) softlabel_file = osp.join(self.root, softlabel_name) else: label_file = label_name softlabel_file = softlabel_name self.files.append({ "img": img_file, "label": label_file, "softlabel": softlabel_file, "img_name": img_name, "label_name": label_name, "softlabel_name": softlabel_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label, input_softlabel): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) # interpolate the softlabel by 3-channel groups h,w = label.shape num_group = int(np.ceil(self.num_classes/3.0)) softlabel = np.zeros((h,w,self.num_classes), dtype=np.float32) start_idx = 0 for idx in range(num_group): clabel = input_softlabel[:,:,start_idx:start_idx+3] clabel = cv2.resize(clabel, None, fx=f_scale, fy=f_scale, interpolation=cv2.INTER_LINEAR) softlabel[:,:,start_idx:start_idx+3] = clabel.reshape(h,w,-1) start_idx = start_idx + 3 softlabel = softlabel.transpose(2,0,1)/np.sum(softlabel,2) softlabel = softlabel.transpose(1,2,0) return image, label, softlabel def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) softlabel = np.load(datafiles["softlabel"]) # sys.path.insert(0, 'dataset/helpers') if self.data_src == 'gta': from labels import id2label elif self.data_src == 'synthia': from labels_cityscapes_synthia import id2label label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label, softlabel = self.generate_scale_label(image, label, softlabel) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std # np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) # softlabel_pad h_pad, w_pad = label_pad.shape num_group = int(np.ceil(self.num_classes / 3.0)) softlabel_pad = np.zeros((h_pad, w_pad, self.num_classes), dtype=np.float32) start_idx = 0 for idx in range(num_group): clabel_pad = softlabel[:, :, start_idx:start_idx + 3] clabel_pad = cv2.copyMakeBorder(clabel_pad, 0, pad_h, 0,pad_w, cv2.BORDER_CONSTANT,value=(0.0, 0.0, 0.0)) softlabel_pad[:, :, start_idx:start_idx + 3] = clabel_pad.reshape(h_pad,w_pad,-1) start_idx = start_idx + 3 else: img_pad, label_pad, softlabel_pad = image, label, softlabel img_h, img_w = label_pad.shape # mining or not mine_flag = random.uniform(0, 1) < self.mine_chance if mine_flag and len(self.mine_id) > 0: label_unique = np.unique(label_pad) mine_id_temp = np.array([a for a in self.mine_id if a in label_unique]) # if this image has the mine id if mine_id_temp.size != 0: # decide the single id to be mined mine_id_img = mine_id_temp sel_idx = random.randint(0, mine_id_temp.size-1) sel_mine_id = mine_id_img[sel_idx] # seed for the mined id mine_id_loc = np.where(label_pad == sel_mine_id) # tuple mine_id_len = len(mine_id_loc[0]) seed_loc = random.randint(0, mine_id_len-1) hseed = mine_id_loc[0][seed_loc] wseed = mine_id_loc[1][seed_loc] # patch crop half_crop_h = self.crop_h/2 half_crop_w = self.crop_w/2 # center crop at the seed left_idx = wseed - half_crop_w right_idx = wseed + half_crop_w -1 up_idx = hseed - half_crop_h bottom_idx = hseed + half_crop_h - 1 # shift the left_idx or right_idx if they go beyond the pad margins if left_idx < 0: left_idx = 0 elif right_idx > img_w - 1: left_idx = left_idx - ( ( half_crop_w - 1 ) - (img_w - 1 - wseed) ) # left_idx shifts to the left by the right beyond length if up_idx < 0: up_idx = 0 elif bottom_idx > img_h - 1: up_idx = up_idx - ( ( half_crop_h - 1 ) - (img_h - 1 - hseed) ) # up_idx shifts to the up by the bottom beyond length h_off = up_idx w_off = left_idx else: h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) else: h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) softlabel = np.asarray(softlabel_pad[h_off: h_off + self.crop_h, w_off: w_off + self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) # set ignored label vector to be all zeros label_expand = np.tile( np.expand_dims(label, axis=2), (1, 1, self.num_classes) ) labels_ = label_expand.copy() labels_[labels_ != self.ignore_label] = 1.0 labels_[labels_ == self.ignore_label] = 0.0 labels = labels_*softlabel labels = labels.transpose((2,0,1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] labels = labels[:,:,::flip] # weighted_pred_trainIDs = np.asarray(np.argmax(labels, axis=0), dtype=np.uint8) # # save weighted predication # wpred_label_col = weighted_pred_trainIDs.copy() # wpred_label_col = colorize_mask(wpred_label_col) # wpred_label_col.save('%s_color.png' % (index)) # # labels_sum = np.sum(labels,axis=0) # # save weighted predication # weighted_pred_trainIDs[labels_sum == 0] = 255 # wpred_label_col = weighted_pred_trainIDs.copy() # wpred_label_col = colorize_mask(wpred_label_col) # wpred_label_col.save('%s_pseudo_color.png' % (index)) return image.copy(), labels.copy(), np.array(size), img_name, self.reg_weight class GTA5StMineDataSet(data.Dataset): def __init__(self, root, list_path, data_src=None, reg_weight = 0.0, rare_id = None, mine_id = None, mine_chance = None, pseudo_root = None, max_iters=None, crop_size=(500, 500), train_scale = (0.5, 1.5), mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.pseudo_root = pseudo_root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.crop_h, self.crop_w = crop_size self.lscale, self.hscale = train_scale self.scale = scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] self.reg_weight = reg_weight self.rare_id = rare_id self.mine_id = mine_id self.mine_chance = mine_chance self.data_src = data_src with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) if self.pseudo_root == None: label_file = osp.join(self.root, label_name) else: label_file = label_name self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def generate_scale_label(self, image, label): # f_scale = 0.5 + random.randint(0, 11) / 10.0 f_scale = self.lscale + random.randint(0, int((self.hscale-self.lscale)*10)) / 10.0 image = cv2.resize(image, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_LINEAR) label = cv2.resize(label, None, fx=f_scale, fy=f_scale, interpolation = cv2.INTER_NEAREST) return image, label def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = np.array(Image.open(datafiles["label"])) # sys.path.insert(0, 'dataset/helpers') if self.data_src == 'gta': from labels import id2label elif self.data_src == 'synthia': from labels_cityscapes_synthia import id2label label_2_id = 255 * np.ones((256,)) for l in id2label: if l in (-1, 255): continue label_2_id[l] = id2label[l].trainId label = label_2_id[label] # size = image.shape img_name = datafiles["img_name"] if self.scale: image, label = self.generate_scale_label(image, label) image = np.asarray(image, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std # np.reshape(self.std,(1,1,3)) img_h, img_w = label.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: img_pad = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) label_pad = cv2.copyMakeBorder(label, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(self.ignore_label,)) else: img_pad, label_pad = image, label img_h, img_w = label_pad.shape # mining or not mine_flag = random.uniform(0, 1) < self.mine_chance if mine_flag and len(self.mine_id) > 0: label_unique = np.unique(label_pad) mine_id_temp = np.array([a for a in self.mine_id if a in label_unique]) # if this image has the mine id if mine_id_temp.size != 0: # decide the single id to be mined mine_id_img = mine_id_temp sel_idx = random.randint(0, mine_id_temp.size-1) sel_mine_id = mine_id_img[sel_idx] # seed for the mined id mine_id_loc = np.where(label_pad == sel_mine_id) # tuple mine_id_len = len(mine_id_loc[0]) seed_loc = random.randint(0, mine_id_len-1) hseed = mine_id_loc[0][seed_loc] wseed = mine_id_loc[1][seed_loc] # patch crop half_crop_h = self.crop_h/2 half_crop_w = self.crop_w/2 # center crop at the seed left_idx = wseed - half_crop_w right_idx = wseed + half_crop_w -1 up_idx = hseed - half_crop_h bottom_idx = hseed + half_crop_h - 1 # shift the left_idx or right_idx if they go beyond the pad margins if left_idx < 0: left_idx = 0 elif right_idx > img_w - 1: left_idx = left_idx - ( ( half_crop_w - 1 ) - (img_w - 1 - wseed) ) # left_idx shifts to the left by the right beyond length if up_idx < 0: up_idx = 0 elif bottom_idx > img_h - 1: up_idx = up_idx - ( ( half_crop_h - 1 ) - (img_h - 1 - hseed) ) # up_idx shifts to the up by the bottom beyond length h_off = up_idx w_off = left_idx else: h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) else: h_off = random.randint(0, img_h - self.crop_h) w_off = random.randint(0, img_w - self.crop_w) image = np.asarray(img_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) label = np.asarray(label_pad[h_off : h_off+self.crop_h, w_off : w_off+self.crop_w], np.float32) image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) if self.is_mirror: flip = np.random.choice(2) * 2 - 1 image = image[:, :, ::flip] label = label[:, ::flip] return image.copy(), label.copy(), np.array(size), img_name, self.reg_weight class GTA5TestDataSet(data.Dataset): def __init__(self, root, list_path, max_iters=None, test_size=(1024, 512), test_scale = 1.0, mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.list_path = list_path self.test_h, self.test_w = test_size self.scale = scale self.test_scale = test_scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) label_file = osp.join(self.root, label_name) self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = cv2.imread(datafiles["label"], cv2.IMREAD_GRAYSCALE) # # sys.path.insert(0, 'dataset/helpers') # from labels import id2label, trainId2label # # # label_2_id = 255 * np.ones((256,)) # for l in id2label: # if l in (-1, 255): # continue # label_2_id[l] = id2label[l].trainId # # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # # valid_labels = sorted(set(id_2_label.ravel())) # label = label_2_id[label] # # resize img_name = datafiles["img_name"] # image = cv2.resize(image, (self.test_h, self.test_w), fx=0, fy=0, interpolation = cv2.INTER_LINEAR) # label = cv2.resize(label, (self.test_h, self.test_w), fx=0, fy=0, interpolation = cv2.INTER_NEAREST) image = cv2.resize(image, None, fx=self.test_scale, fy=self.test_scale, interpolation = cv2.INTER_LINEAR) # always keep the resolution of label unchanged # label = cv2.resize(label, None, fx=1, fy=1, interpolation = cv2.INTER_NEAREST) image = np.asarray(image, np.float32) label = np.asarray(label, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) size = image.shape image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) return image.copy(), label.copy(), np.array(size), img_name class GTA5MSTDataSet(data.Dataset): def __init__(self, root, list_path, max_iters=None, test_size=(1024, 512), test_scale = 1.0, mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.list_path = list_path self.test_h, self.test_w = test_size self.scale = scale self.test_scale = test_scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) label_file = osp.join(self.root, label_name) self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = cv2.imread(datafiles["label"], cv2.IMREAD_GRAYSCALE) # # sys.path.insert(0, 'dataset/helpers') # from labels import id2label, trainId2label # # # label_2_id = 255 * np.ones((256,)) # for l in id2label: # if l in (-1, 255): # continue # label_2_id[l] = id2label[l].trainId # # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # # valid_labels = sorted(set(id_2_label.ravel())) # label = label_2_id[label] # # resize img_name = datafiles["img_name"] # image = cv2.resize(image, (self.test_h, self.test_w), fx=0, fy=0, interpolation = cv2.INTER_LINEAR) # label = cv2.resize(label, (self.test_h, self.test_w), fx=0, fy=0, interpolation = cv2.INTER_NEAREST) image = cv2.resize(image, None, fx=self.test_scale, fy=self.test_scale, interpolation = cv2.INTER_LINEAR) # always keep the resolution of label unchanged # label = cv2.resize(label, None, fx=1, fy=1, interpolation = cv2.INTER_NEAREST) image = np.asarray(image, np.float32) label = np.asarray(label, np.float32) # image = image/255.0 # scale to [0,1] # image -= self.mean # BGR # image = image/self.std#np.reshape(self.std,(1,1,3)) size = image.shape image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) return image.copy(), label.copy(), np.array(size), img_name class GTA5TestCRFDataSet(data.Dataset): def __init__(self, root, list_path, max_iters=None, test_size=(1024, 512), test_scale = 1.0, mean=(128, 128, 128), std = (1,1,1), scale=True, mirror=True, ignore_label=255): self.root = root self.list_path = list_path self.test_h, self.test_w = test_size self.scale = scale self.test_scale = test_scale self.ignore_label = ignore_label self.mean = mean self.std = std self.is_mirror = mirror # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [] self.label_ids = [] with open(list_path) as f: for item in f.readlines(): fields = item.strip().split('\t') self.img_ids.append(fields[0]) self.label_ids.append(fields[1]) if not max_iters==None: self.img_ids = self.img_ids * int(np.ceil(float(max_iters) / len(self.img_ids))) self.label_ids = self.label_ids * int(np.ceil(float(max_iters) / len(self.label_ids))) self.files = [] # for split in ["train", "trainval", "val"]: for idx in range(len(self.img_ids)): img_name = self.img_ids[idx] label_name = self.label_ids[idx] img_file = osp.join(self.root, img_name) label_file = osp.join(self.root, label_name) self.files.append({ "img": img_file, "label": label_file, "img_name": img_name, "label_name": label_name }) def __len__(self): return len(self.files) def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) # OpenCV read image as BGR, not RGB label = cv2.imread(datafiles["label"], cv2.IMREAD_GRAYSCALE) # # sys.path.insert(0, 'dataset/helpers') # from labels import id2label, trainId2label # # # label_2_id = 255 * np.ones((256,)) # for l in id2label: # if l in (-1, 255): # continue # label_2_id[l] = id2label[l].trainId # # id_2_label = np.array([trainId2label[_].id for _ in trainId2label if _ not in (-1, 255)]) # # valid_labels = sorted(set(id_2_label.ravel())) # label = label_2_id[label] # # resize img_name = datafiles["img_name"] # image = cv2.resize(image, (self.test_h, self.test_w), fx=0, fy=0, interpolation = cv2.INTER_LINEAR) # label = cv2.resize(label, (self.test_h, self.test_w), fx=0, fy=0, interpolation = cv2.INTER_NEAREST) image_crf = np.asarray(image, np.float32) image_crf = image_crf[:, :, ::-1] # change to RGB image = cv2.resize(image, None, fx=self.test_scale, fy=self.test_scale, interpolation = cv2.INTER_LINEAR) # always keep the resolution of label unchanged # label = cv2.resize(label, None, fx=1, fy=1, interpolation = cv2.INTER_NEAREST) image = np.asarray(image, np.float32) label = np.asarray(label, np.float32) image = image/255.0 # scale to [0,1] image -= self.mean # BGR image = image/self.std#np.reshape(self.std,(1,1,3)) size = image.shape image = image[:, :, ::-1] # change to RGB image = image.transpose((2, 0, 1)) return image.copy(), label.copy(), image_crf.copy(), np.array(size), img_name class VOCDataTestSet(data.Dataset): def __init__(self, root, list_path, crop_size=(505, 505), mean=(128, 128, 128)): self.root = root self.list_path = list_path self.crop_h, self.crop_w = crop_size self.mean = mean # self.mean_bgr = np.array([104.00698793, 116.66876762, 122.67891434]) self.img_ids = [i_id.strip() for i_id in open(list_path)] self.files = [] # for split in ["train", "trainval", "val"]: for name in self.img_ids: img_file = osp.join(self.root, "img/%s.jpg" % name) self.files.append({ "img": img_file }) def __len__(self): return len(self.files) def __getitem__(self, index): datafiles = self.files[index] image = cv2.imread(datafiles["img"], cv2.IMREAD_COLOR) size = image.shape name = osp.splitext(osp.basename(datafiles["img"]))[0] image = np.asarray(image, np.float32) image -= self.mean img_h, img_w, _ = image.shape pad_h = max(self.crop_h - img_h, 0) pad_w = max(self.crop_w - img_w, 0) if pad_h > 0 or pad_w > 0: image = cv2.copyMakeBorder(image, 0, pad_h, 0, pad_w, cv2.BORDER_CONSTANT, value=(0.0, 0.0, 0.0)) image = image.transpose((2, 0, 1)) return image, name, size if __name__ == '__main__': dst = VOCDataSet("./data", is_transform=True) trainloader = data.DataLoader(dst, batch_size=4) for i, data in enumerate(trainloader): imgs, labels = data if i == 0: img = torchvision.utils.make_grid(imgs).numpy() img = np.transpose(img, (1, 2, 0)) img = img[:, :, ::-1] plt.imshow(img) plt.show()
62,075
43.276748
309
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/linear_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbedding(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchExpansion(nn.Module): def __init__(self, dim_scale, channel_dim, img_channels, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, dim_scale**2* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): """ x: B, H, W, C """ x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int, f_hidden: int, ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), nn.Linear(num_patches, num_patches), Rearrange("b c w h -> b c h w"), nn.Linear(num_patches, num_patches), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels*f_hidden) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Linear_Mixer(nn.Module): def __init__( self, img_size: int = 256, img_channels: int = 3, patch_size: int = 4, embed_dim: int = 140, num_layers: int = 19, f_hidden: int = 4, ): super().__init__() self.patch_embed = PatchEmbedding(patch_size, embed_dim, img_channels) layers = [ Mixer(img_size//patch_size, embed_dim, f_hidden) for _ in range(num_layers)] self.mixer_layers = nn.Sequential(*layers) self.patch_expand = PatchExpansion(patch_size, embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.patch_embed(x) x = self.mixer_layers(x) x = self.patch_expand(x) return x
3,644
26.613636
127
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/original_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbeddings(nn.Module): def __init__( self, patch_size: int, hidden_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=hidden_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b (h w) c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchEmbeddings_transpose(nn.Module): def __init__( self, patch_size: int, hidden_dim: int, channels: int, d: int ): super().__init__() self.proj_transpose = nn.Sequential( Rearrange("b (h w) c -> b c h w", h=d), nn.ConvTranspose2d( in_channels=hidden_dim, out_channels=channels, kernel_size=patch_size, stride=patch_size ) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj_transpose(x) class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class MixerBlock(nn.Module): def __init__( self, num_patches: int, num_channels: int, tokens_hidden_dim: int, channels_hidden_dim: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b p c -> b c p"), MLPBlock(num_patches, tokens_hidden_dim), Rearrange("b c p -> b p c") ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, channels_hidden_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Original_Mixer(nn.Module): def __init__( self, image_size: int = 256, channels: int = 3, patch_size: int = 4, num_layers: int = 8, hidden_dim: int = 128, tokens_hidden_dim: int = 96, channels_hidden_dim: int = 256 ): super().__init__() num_patches = (image_size // patch_size) ** 2 d=(image_size-patch_size)//patch_size + 1 self.embed = PatchEmbeddings(patch_size, hidden_dim, channels) layers = [ MixerBlock( num_patches=num_patches, num_channels=hidden_dim, tokens_hidden_dim=tokens_hidden_dim, channels_hidden_dim=channels_hidden_dim ) for _ in range(num_layers) ] self.layers = nn.Sequential(*layers) self.norm = nn.LayerNorm(hidden_dim) self.embed_transpose = PatchEmbeddings_transpose(patch_size, hidden_dim, channels, d) def forward(self, x: torch.Tensor) -> torch.Tensor: b, c, h, w = x.shape x = self.embed(x) x = self.layers(x) x = self.norm(x) x = self.embed_transpose(x) return x
3,674
26.840909
93
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/img2img_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbedding(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchExpansion(nn.Module): def __init__(self, dim_scale, channel_dim, img_channels, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, dim_scale**2* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int, f_hidden: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c w h -> b c h w"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels*f_hidden) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Img2Img_Mixer(nn.Module): def __init__( self, img_size: int = 256, img_channels: int = 3, patch_size: int = 4, embed_dim: int = 128, num_layers: int = 16, f_hidden: int = 4, ): super().__init__() self.patch_embed = PatchEmbedding(patch_size, embed_dim, img_channels) layers = [ Mixer(img_size//patch_size, embed_dim, f_hidden) for _ in range(num_layers)] self.mixer_layers = nn.Sequential(*layers) self.patch_expand = PatchExpansion(patch_size, embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.patch_embed(x) x = self.mixer_layers(x) x = self.patch_expand(x) return x
3,618
27.054264
127
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/vit.py
''' This code is modified from https://github.com/facebookresearch/convit. To adapt the vit/convit to image reconstruction, variable input sizes, and patch sizes for both spatial dimensions. ''' import torch import torch.nn as nn from functools import partial import torch.nn.functional as F from timm.models.helpers import load_pretrained from timm.models.layers import DropPath, to_2tuple, trunc_normal_ class Mlp(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.fc2 = nn.Linear(hidden_features, out_features) self.drop = nn.Dropout(drop) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) x = self.drop(x) return x class GPSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., locality_strength=1., use_local_init=True, grid_size=None): super().__init__() self.num_heads = num_heads self.dim = dim head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.q = nn.Linear(dim, dim, bias=qkv_bias) self.k = nn.Linear(dim, dim, bias=qkv_bias) self.v = nn.Linear(dim, dim, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.pos_proj = nn.Linear(3, num_heads) self.proj_drop = nn.Dropout(proj_drop) self.locality_strength = locality_strength self.gating_param = nn.Parameter(1*torch.ones(self.num_heads)) self.apply(self._init_weights) if use_local_init: self.local_init(locality_strength=locality_strength) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention(self, x): B, N, C = x.shape k = self.k(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) pos_score = self.pos_proj(self.rel_indices).expand(B, -1, -1,-1).permute(0,3,1,2) patch_score = (q @ k.transpose(-2, -1)) * self.scale patch_score = patch_score.softmax(dim=-1) pos_score = pos_score.softmax(dim=-1) gating = self.gating_param.view(1,-1,1,1) attn = (1.-torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score attn = attn / attn.sum(dim=-1).unsqueeze(-1) attn = self.attn_drop(attn) return attn def get_attention_map(self, x, return_map = False): attn_map = self.get_attention(x).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def local_init(self, locality_strength=1.): self.v.weight.data.copy_(torch.eye(self.dim)) locality_distance = 1 #max(1,1/locality_strength**.5) kernel_size = int(self.num_heads**.5) center = (kernel_size-1)/2 if kernel_size%2==0 else kernel_size//2 for h1 in range(kernel_size): for h2 in range(kernel_size): position = h1+kernel_size*h2 self.pos_proj.weight.data[position,2] = -1 self.pos_proj.weight.data[position,1] = 2*(h1-center)*locality_distance self.pos_proj.weight.data[position,0] = 2*(h2-center)*locality_distance self.pos_proj.weight.data *= locality_strength def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.v.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape if not hasattr(self, 'rel_indices') or self.rel_indices.size(1)!=N: self.get_rel_indices() attn = self.get_attention(x) v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class MHSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., grid_size=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) self.apply(self._init_weights) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention_map(self, x, return_map = False): self.get_rel_indices() B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn_map = (q @ k.transpose(-2, -1)) * self.scale attn_map = attn_map.softmax(dim=-1).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.qkv.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Block(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs): super().__init__() self.norm1 = norm_layer(dim) self.use_gpsa = use_gpsa if self.use_gpsa: self.attn = GPSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) else: self.attn = MHSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) def forward(self, x, grid_size): self.attn.current_grid_size = grid_size x = x + self.drop_path(self.attn(self.norm1(x))) x = x + self.drop_path(self.mlp(self.norm2(x))) return x class PatchEmbed(nn.Module): """ Image to Patch Embedding, from timm """ def __init__(self, patch_size, in_chans, embed_dim): super().__init__() self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) self.apply(self._init_weights) def forward(self, x): x = self.proj(x) return x def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) class VisionTransformer(nn.Module): """ Vision Transformer """ def __init__(self, avrg_img_size=320, patch_size=16, in_chans=1, embed_dim=64, depth=8, num_heads=9, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, global_pool=None, gpsa_interval=[-1, -1], locality_strength=1., use_pos_embed=True): super().__init__() self.depth = depth embed_dim *= num_heads self.num_features = embed_dim # num_features for consistency with other models self.locality_strength = locality_strength self.use_pos_embed = use_pos_embed if isinstance(avrg_img_size, int): img_size = to_2tuple(avrg_img_size) if isinstance(patch_size, int): self.patch_size = to_2tuple(patch_size) else: self.patch_size = patch_size self.in_chans = in_chans self.patch_embed = PatchEmbed( patch_size=self.patch_size, in_chans=in_chans, embed_dim=embed_dim) self.pos_drop = nn.Dropout(p=drop_rate) if self.use_pos_embed: self.pos_embed = nn.Parameter( torch.zeros(1, embed_dim, img_size[0] // self.patch_size[0], img_size[1] // self.patch_size[1]) ) trunc_normal_(self.pos_embed, std=.02) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule self.blocks = nn.ModuleList([ Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=True, locality_strength=locality_strength) if i>=gpsa_interval[0]-1 and i<gpsa_interval[1] else Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=False,) for i in range(depth)]) self.norm = norm_layer(embed_dim) # head self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')] self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def seq2img(self, x, img_size): """ Transforms sequence back into image space, input dims: [batch_size, num_patches, channels] output dims: [batch_size, channels, H, W] """ x = x.view(x.shape[0], x.shape[1], self.in_chans, self.patch_size[0], self.patch_size[1]) x = x.chunk(x.shape[1], dim=1) x = torch.cat(x, dim=4).permute(0,1,2,4,3) x = x.chunk(img_size[0]//self.patch_size[0], dim=3) x = torch.cat(x, dim=4).permute(0,1,2,4,3).squeeze(1) return x self.head.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) @torch.jit.ignore def no_weight_decay(self,): return {'pos_embed'} def get_head(self,): return self.head def reset_head(self,): self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def forward_features(self, x, k=None): x = self.patch_embed(x) _, _, H, W = x.shape if self.use_pos_embed: pos_embed = F.interpolate(self.pos_embed, size=[H, W], mode='bilinear', align_corners = False) x = x + pos_embed x = x.flatten(2).transpose(1, 2) x = self.pos_drop(x) for u, blk in enumerate(self.blocks): x = blk(x, (H, W)) if k is not None and u == k: self.attention_map = blk.attn.get_attention_map(x, return_map = True) x = self.norm(x) return x def forward(self, x, k=None): _, _, H, W = x.shape x = self.forward_features(x, k) x = self.head(x) x = self.seq2img(x, (H, W)) return x
15,082
39.007958
186
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/recon_net.py
import torch.nn as nn import torch.nn.functional as F from math import ceil, floor class ReconNet(nn.Module): def __init__(self, net): super().__init__() self.net = net def pad(self, x): _, _, h, w = x.shape hp, wp = self.net.patch_size f1 = ( (wp - w % wp) % wp ) / 2 f2 = ( (hp - h % hp) % hp ) / 2 wpad = [floor(f1), ceil(f1)] hpad = [floor(f2), ceil(f2)] x = F.pad(x, wpad+hpad) return x, wpad, hpad def unpad(self, x, wpad, hpad): return x[..., hpad[0] : x.shape[-2]-hpad[1], wpad[0] : x.shape[-1]-wpad[1]] def forward(self, x, k=None): x, wpad, hpad = self.pad(x) x = self.net(x, k) x = self.unpad(x, wpad, hpad) return x
810
26.033333
90
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/unet.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from torch import nn from torch.nn import functional as F class Unet(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int = 3, out_chans: int = 3, chans: int = 21, num_pool_layers: int = 4, drop_prob: float = 0.0, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append( nn.Sequential( ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1), ) ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
5,981
32.79661
88
py
imaging_MLPs
imaging_MLPs-master/ImageNet/networks/u_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbeddings(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchMerge(nn.Module): r""" Patch Merging Layer. Args: input_resolution (int): Resolution of input feature. dim (int): Number of input channels. norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm """ def __init__(self, channel_dim, norm_layer=nn.LayerNorm): super().__init__() self.dim = channel_dim self.reduction = nn.Linear(4 * channel_dim, 2 * channel_dim, bias=False) self.norm = norm_layer(4 * channel_dim) def forward(self, x): """ x: B, H, W, C """ B, H, W, C = x.shape x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C x = x.view(B, H//2, W//2, 4 * C) # B H/2*W/2 4*C x = self.norm(x) x = self.reduction(x) return x class PatchExpand(nn.Module): def __init__(self, channel_dim, dim_scale=2, norm_layer=nn.LayerNorm): super().__init__() self.dim = channel_dim self.expand = nn.Linear(channel_dim, 2* channel_dim, bias=False) if dim_scale==2 else nn.Identity() self.norm = norm_layer(channel_dim // dim_scale) def forward(self, x): """ x: B, H, W, C """ x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=2, p2=2, c=C//4) x = x.view(B, H*2, W*2 ,C//4) x= self.norm(x) return x class FinalPatchExpand(nn.Module): def __init__(self, channel_dim, img_channels, dim_scale=4, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, 16* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): """ x: B, H, W, C """ x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), MLPBlock(num_patches, num_patches), Rearrange("b c w h -> b c h w"), MLPBlock(num_patches, num_patches), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class U_Mixer(nn.Module): def __init__( self, img_size: int = 256, img_channels: int = 3, embed_dim: int = 96, ): super().__init__() #mixer blocks self.mixer1= Mixer( img_size//4, embed_dim) self.mixer2= Mixer( img_size//8, embed_dim*2) self.mixer3= Mixer( img_size//16, embed_dim*4) self.mixer4= Mixer( img_size//32, embed_dim*8) self.mixer11= Mixer( img_size//4, embed_dim) self.mixer22= Mixer( img_size//8, embed_dim*2) self.mixer33= Mixer( img_size//16, embed_dim*4) self.mixer44= Mixer( img_size//32, embed_dim*8) #encode self.patch_embed = PatchEmbeddings(4, embed_dim, img_channels) self.patch_merge1 = PatchMerge(embed_dim) self.patch_merge2= PatchMerge(embed_dim*2) self.patch_merge3= PatchMerge(embed_dim*4) #decode self.patch_expand1 = PatchExpand(embed_dim*8) self.patch_expand2 = PatchExpand(embed_dim*4) self.patch_expand3 = PatchExpand(embed_dim*2) self.final_expand = FinalPatchExpand(embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: y=[] x = self.patch_embed(x) x = self.mixer1(x) x = self.mixer1(x) y.append(x) x = self.patch_merge1(x) x = self.mixer2(x) x = self.mixer2(x) y.append(x) x = self.patch_merge2(x) x = self.mixer3(x) x = self.mixer3(x) y.append(x) x = self.patch_merge3(x) x = self.mixer4(x) x = self.mixer4(x) x = self.mixer44(x) x = self.mixer44(x) x = self.patch_expand1(x) x = self.mixer33(x) x = self.mixer33(x) + y[2] x = self.patch_expand2(x) x = self.mixer22(x) x = self.mixer22(x) + y[1] x = self.patch_expand3(x) x = self.mixer11(x) x = self.mixer11(x) + y[0] x = self.final_expand(x) return x
6,516
28.488688
127
py
imaging_MLPs
imaging_MLPs-master/SIDD/networks/img2img_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbedding(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchExpansion(nn.Module): def __init__(self, dim_scale, channel_dim, img_channels, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, dim_scale**2* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int, f_hidden: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c w h -> b c h w"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels*f_hidden) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Img2Img_Mixer(nn.Module): def __init__( self, img_size: int = 256, img_channels: int = 3, patch_size: int = 4, embed_dim: int = 128, num_layers: int = 16, f_hidden: int = 4, ): super().__init__() self.patch_embed = PatchEmbedding(patch_size, embed_dim, img_channels) layers = [ Mixer(img_size//patch_size, embed_dim, f_hidden) for _ in range(num_layers)] self.mixer_layers = nn.Sequential(*layers) self.patch_expand = PatchExpansion(patch_size, embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.patch_embed(x) x = self.mixer_layers(x) x = self.patch_expand(x) return x
3,618
27.054264
127
py
imaging_MLPs
imaging_MLPs-master/SIDD/networks/vit.py
''' This code is modified from https://github.com/facebookresearch/convit. To adapt the vit/convit to image reconstruction, variable input sizes, and patch sizes for both spatial dimensions. ''' import torch import torch.nn as nn from functools import partial import torch.nn.functional as F from timm.models.helpers import load_pretrained from timm.models.layers import DropPath, to_2tuple, trunc_normal_ class Mlp(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.fc2 = nn.Linear(hidden_features, out_features) self.drop = nn.Dropout(drop) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) x = self.drop(x) return x class GPSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., locality_strength=1., use_local_init=True, grid_size=None): super().__init__() self.num_heads = num_heads self.dim = dim head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.q = nn.Linear(dim, dim, bias=qkv_bias) self.k = nn.Linear(dim, dim, bias=qkv_bias) self.v = nn.Linear(dim, dim, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.pos_proj = nn.Linear(3, num_heads) self.proj_drop = nn.Dropout(proj_drop) self.locality_strength = locality_strength self.gating_param = nn.Parameter(1*torch.ones(self.num_heads)) self.apply(self._init_weights) if use_local_init: self.local_init(locality_strength=locality_strength) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention(self, x): B, N, C = x.shape k = self.k(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) pos_score = self.pos_proj(self.rel_indices).expand(B, -1, -1,-1).permute(0,3,1,2) patch_score = (q @ k.transpose(-2, -1)) * self.scale patch_score = patch_score.softmax(dim=-1) pos_score = pos_score.softmax(dim=-1) gating = self.gating_param.view(1,-1,1,1) attn = (1.-torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score attn = attn / attn.sum(dim=-1).unsqueeze(-1) attn = self.attn_drop(attn) return attn def get_attention_map(self, x, return_map = False): attn_map = self.get_attention(x).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def local_init(self, locality_strength=1.): self.v.weight.data.copy_(torch.eye(self.dim)) locality_distance = 1 #max(1,1/locality_strength**.5) kernel_size = int(self.num_heads**.5) center = (kernel_size-1)/2 if kernel_size%2==0 else kernel_size//2 for h1 in range(kernel_size): for h2 in range(kernel_size): position = h1+kernel_size*h2 self.pos_proj.weight.data[position,2] = -1 self.pos_proj.weight.data[position,1] = 2*(h1-center)*locality_distance self.pos_proj.weight.data[position,0] = 2*(h2-center)*locality_distance self.pos_proj.weight.data *= locality_strength def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.v.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape if not hasattr(self, 'rel_indices') or self.rel_indices.size(1)!=N: self.get_rel_indices() attn = self.get_attention(x) v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class MHSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., grid_size=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) self.apply(self._init_weights) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention_map(self, x, return_map = False): self.get_rel_indices() B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn_map = (q @ k.transpose(-2, -1)) * self.scale attn_map = attn_map.softmax(dim=-1).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.qkv.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Block(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs): super().__init__() self.norm1 = norm_layer(dim) self.use_gpsa = use_gpsa if self.use_gpsa: self.attn = GPSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) else: self.attn = MHSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) def forward(self, x, grid_size): self.attn.current_grid_size = grid_size x = x + self.drop_path(self.attn(self.norm1(x))) x = x + self.drop_path(self.mlp(self.norm2(x))) return x class PatchEmbed(nn.Module): """ Image to Patch Embedding, from timm """ def __init__(self, patch_size, in_chans, embed_dim): super().__init__() self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) self.apply(self._init_weights) def forward(self, x): x = self.proj(x) return x def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) class VisionTransformer(nn.Module): """ Vision Transformer """ def __init__(self, avrg_img_size=320, patch_size=16, in_chans=1, embed_dim=64, depth=8, num_heads=9, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, global_pool=None, gpsa_interval=[-1, -1], locality_strength=1., use_pos_embed=True): super().__init__() self.depth = depth embed_dim *= num_heads self.num_features = embed_dim # num_features for consistency with other models self.locality_strength = locality_strength self.use_pos_embed = use_pos_embed if isinstance(avrg_img_size, int): img_size = to_2tuple(avrg_img_size) if isinstance(patch_size, int): self.patch_size = to_2tuple(patch_size) else: self.patch_size = patch_size self.in_chans = in_chans self.patch_embed = PatchEmbed( patch_size=self.patch_size, in_chans=in_chans, embed_dim=embed_dim) self.pos_drop = nn.Dropout(p=drop_rate) if self.use_pos_embed: self.pos_embed = nn.Parameter( torch.zeros(1, embed_dim, img_size[0] // self.patch_size[0], img_size[1] // self.patch_size[1]) ) trunc_normal_(self.pos_embed, std=.02) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule self.blocks = nn.ModuleList([ Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=True, locality_strength=locality_strength) if i>=gpsa_interval[0]-1 and i<gpsa_interval[1] else Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=False,) for i in range(depth)]) self.norm = norm_layer(embed_dim) # head self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')] self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def seq2img(self, x, img_size): """ Transforms sequence back into image space, input dims: [batch_size, num_patches, channels] output dims: [batch_size, channels, H, W] """ x = x.view(x.shape[0], x.shape[1], self.in_chans, self.patch_size[0], self.patch_size[1]) x = x.chunk(x.shape[1], dim=1) x = torch.cat(x, dim=4).permute(0,1,2,4,3) x = x.chunk(img_size[0]//self.patch_size[0], dim=3) x = torch.cat(x, dim=4).permute(0,1,2,4,3).squeeze(1) return x self.head.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) @torch.jit.ignore def no_weight_decay(self,): return {'pos_embed'} def get_head(self,): return self.head def reset_head(self,): self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def forward_features(self, x, k=None): x = self.patch_embed(x) _, _, H, W = x.shape if self.use_pos_embed: pos_embed = F.interpolate(self.pos_embed, size=[H, W], mode='bilinear', align_corners = False) x = x + pos_embed x = x.flatten(2).transpose(1, 2) x = self.pos_drop(x) for u, blk in enumerate(self.blocks): x = blk(x, (H, W)) if k is not None and u == k: self.attention_map = blk.attn.get_attention_map(x, return_map = True) x = self.norm(x) return x def forward(self, x, k=None): _, _, H, W = x.shape x = self.forward_features(x, k) x = self.head(x) x = self.seq2img(x, (H, W)) return x
15,082
39.007958
186
py
imaging_MLPs
imaging_MLPs-master/SIDD/networks/recon_net.py
import torch.nn as nn import torch.nn.functional as F from math import ceil, floor class ReconNet(nn.Module): def __init__(self, net): super().__init__() self.net = net def pad(self, x): _, _, h, w = x.shape hp, wp = self.net.patch_size f1 = ( (wp - w % wp) % wp ) / 2 f2 = ( (hp - h % hp) % hp ) / 2 wpad = [floor(f1), ceil(f1)] hpad = [floor(f2), ceil(f2)] x = F.pad(x, wpad+hpad) return x, wpad, hpad def unpad(self, x, wpad, hpad): return x[..., hpad[0] : x.shape[-2]-hpad[1], wpad[0] : x.shape[-1]-wpad[1]] def forward(self, x, k=None): x, wpad, hpad = self.pad(x) x = self.net(x, k) x = self.unpad(x, wpad, hpad) return x
810
26.033333
90
py
imaging_MLPs
imaging_MLPs-master/SIDD/networks/unet.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from torch import nn from torch.nn import functional as F class Unet(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int = 3, out_chans: int = 3, chans: int = 21, num_pool_layers: int = 4, drop_prob: float = 0.0, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append( nn.Sequential( ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1), ) ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
5,981
32.79661
88
py
imaging_MLPs
imaging_MLPs-master/SIDD/networks/u_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbeddings(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchMerge(nn.Module): r""" Patch Merging Layer. Args: input_resolution (int): Resolution of input feature. dim (int): Number of input channels. norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm """ def __init__(self, channel_dim, norm_layer=nn.LayerNorm): super().__init__() self.dim = channel_dim self.reduction = nn.Linear(4 * channel_dim, 2 * channel_dim, bias=False) self.norm = norm_layer(4 * channel_dim) def forward(self, x): """ x: B, H, W, C """ B, H, W, C = x.shape x0 = x[:, 0::2, 0::2, :] # B H/2 W/2 C x1 = x[:, 1::2, 0::2, :] # B H/2 W/2 C x2 = x[:, 0::2, 1::2, :] # B H/2 W/2 C x3 = x[:, 1::2, 1::2, :] # B H/2 W/2 C x = torch.cat([x0, x1, x2, x3], -1) # B H/2 W/2 4*C x = x.view(B, H//2, W//2, 4 * C) # B H/2*W/2 4*C x = self.norm(x) x = self.reduction(x) return x class PatchExpand(nn.Module): def __init__(self, channel_dim, dim_scale=2, norm_layer=nn.LayerNorm): super().__init__() self.dim = channel_dim self.expand = nn.Linear(channel_dim, 2* channel_dim, bias=False) if dim_scale==2 else nn.Identity() self.norm = norm_layer(channel_dim // dim_scale) def forward(self, x): """ x: B, H, W, C """ x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=2, p2=2, c=C//4) x = x.view(B, H*2, W*2 ,C//4) x= self.norm(x) return x class FinalPatchExpand(nn.Module): def __init__(self, channel_dim, img_channels, dim_scale=4, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, 16* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): """ x: B, H, W, C """ x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), MLPBlock(num_patches, num_patches), Rearrange("b c w h -> b c h w"), MLPBlock(num_patches, num_patches), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class U_Mixer(nn.Module): def __init__( self, img_size: int = 256, img_channels: int = 3, embed_dim: int = 96, ): super().__init__() #mixer blocks self.mixer1= Mixer( img_size//4, embed_dim) self.mixer2= Mixer( img_size//8, embed_dim*2) self.mixer3= Mixer( img_size//16, embed_dim*4) self.mixer4= Mixer( img_size//32, embed_dim*8) self.mixer11= Mixer( img_size//4, embed_dim) self.mixer22= Mixer( img_size//8, embed_dim*2) self.mixer33= Mixer( img_size//16, embed_dim*4) self.mixer44= Mixer( img_size//32, embed_dim*8) #encode self.patch_embed = PatchEmbeddings(4, embed_dim, img_channels) self.patch_merge1 = PatchMerge(embed_dim) self.patch_merge2= PatchMerge(embed_dim*2) self.patch_merge3= PatchMerge(embed_dim*4) #decode self.patch_expand1 = PatchExpand(embed_dim*8) self.patch_expand2 = PatchExpand(embed_dim*4) self.patch_expand3 = PatchExpand(embed_dim*2) self.final_expand = FinalPatchExpand(embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: y=[] x = self.patch_embed(x) x = self.mixer1(x) x = self.mixer1(x) y.append(x) x = self.patch_merge1(x) x = self.mixer2(x) x = self.mixer2(x) y.append(x) x = self.patch_merge2(x) x = self.mixer3(x) x = self.mixer3(x) y.append(x) x = self.patch_merge3(x) x = self.mixer4(x) x = self.mixer4(x) x = self.mixer44(x) x = self.mixer44(x) x = self.patch_expand1(x) x = self.mixer33(x) x = self.mixer33(x) + y[2] x = self.patch_expand2(x) x = self.mixer22(x) x = self.mixer22(x) + y[1] x = self.patch_expand3(x) x = self.mixer11(x) x = self.mixer11(x) + y[0] x = self.final_expand(x) return x
6,516
28.488688
127
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/losses.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch import torch.nn as nn import torch.nn.functional as F class SSIMLoss(nn.Module): """ SSIM loss module. """ def __init__(self, win_size: int = 7, k1: float = 0.01, k2: float = 0.03): """ Args: win_size: Window size for SSIM calculation. k1: k1 parameter for SSIM calculation. k2: k2 parameter for SSIM calculation. """ super().__init__() self.win_size = win_size self.k1, self.k2 = k1, k2 self.register_buffer("w", torch.ones(1, 1, win_size, win_size) / win_size ** 2) NP = win_size ** 2 self.cov_norm = NP / (NP - 1) def forward(self, X: torch.Tensor, Y: torch.Tensor, data_range: torch.Tensor): assert isinstance(self.w, torch.Tensor) data_range = data_range[:, None, None, None] C1 = (self.k1 * data_range) ** 2 C2 = (self.k2 * data_range) ** 2 ux = F.conv2d(X, self.w) # typing: ignore uy = F.conv2d(Y, self.w) # uxx = F.conv2d(X * X, self.w) uyy = F.conv2d(Y * Y, self.w) uxy = F.conv2d(X * Y, self.w) vx = self.cov_norm * (uxx - ux * ux) vy = self.cov_norm * (uyy - uy * uy) vxy = self.cov_norm * (uxy - ux * uy) A1, A2, B1, B2 = ( 2 * ux * uy + C1, 2 * vxy + C2, ux ** 2 + uy ** 2 + C1, vx + vy + C2, ) D = B1 * B2 S = (A1 * A2) / D return 1 - S.mean()
1,671
28.857143
87
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/coil_combine.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch import fastmri def rss(data: torch.Tensor, dim: int = 0) -> torch.Tensor: """ Compute the Root Sum of Squares (RSS). RSS is computed assuming that dim is the coil dimension. Args: data: The input tensor dim: The dimensions along which to apply the RSS transform Returns: The RSS value. """ return torch.sqrt((data ** 2).sum(dim)) def rss_complex(data: torch.Tensor, dim: int = 0) -> torch.Tensor: """ Compute the Root Sum of Squares (RSS) for complex inputs. RSS is computed assuming that dim is the coil dimension. Args: data: The input tensor dim: The dimensions along which to apply the RSS transform Returns: The RSS value. """ return torch.sqrt(fastmri.complex_abs_sq(data).sum(dim))
996
22.186047
66
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/math.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import numpy as np import torch def complex_mul(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor: """ Complex multiplication. This multiplies two complex tensors assuming that they are both stored as real arrays with the last dimension being the complex dimension. Args: x: A PyTorch tensor with the last dimension of size 2. y: A PyTorch tensor with the last dimension of size 2. Returns: A PyTorch tensor with the last dimension of size 2. """ if not x.shape[-1] == y.shape[-1] == 2: raise ValueError("Tensors do not have separate complex dim.") re = x[..., 0] * y[..., 0] - x[..., 1] * y[..., 1] im = x[..., 0] * y[..., 1] + x[..., 1] * y[..., 0] return torch.stack((re, im), dim=-1) def complex_conj(x: torch.Tensor) -> torch.Tensor: """ Complex conjugate. This applies the complex conjugate assuming that the input array has the last dimension as the complex dimension. Args: x: A PyTorch tensor with the last dimension of size 2. y: A PyTorch tensor with the last dimension of size 2. Returns: A PyTorch tensor with the last dimension of size 2. """ if not x.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return torch.stack((x[..., 0], -x[..., 1]), dim=-1) def complex_abs(data: torch.Tensor) -> torch.Tensor: """ Compute the absolute value of a complex valued input tensor. Args: data: A complex valued tensor, where the size of the final dimension should be 2. Returns: Absolute value of data. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return (data ** 2).sum(dim=-1).sqrt() def complex_abs_sq(data: torch.Tensor) -> torch.Tensor: """ Compute the squared absolute value of a complex tensor. Args: data: A complex valued tensor, where the size of the final dimension should be 2. Returns: Squared absolute value of data. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return (data ** 2).sum(dim=-1) def tensor_to_complex_np(data: torch.Tensor) -> np.ndarray: """ Converts a complex torch tensor to numpy array. Args: data: Input data to be converted to numpy. Returns: Complex numpy version of data. """ data = data.numpy() return data[..., 0] + 1j * data[..., 1]
2,728
25.754902
77
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/__init__.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from packaging import version from .coil_combine import rss, rss_complex from .fftc import fftshift, ifftshift, roll from .losses import SSIMLoss from .math import ( complex_abs, complex_abs_sq, complex_conj, complex_mul, tensor_to_complex_np, ) from .utils import convert_fnames_to_v2, save_reconstructions if version.parse(torch.__version__) >= version.parse("1.7.0"): from .fftc import fft2c_new as fft2c from .fftc import ifft2c_new as ifft2c else: from .fftc import fft2c_old as fft2c from .fftc import ifft2c_old as ifft2c
758
26.107143
63
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/fftc.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ from typing import List, Optional import torch from packaging import version if version.parse(torch.__version__) >= version.parse("1.7.0"): import torch.fft # type: ignore def fft2c_old(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2 dimensional Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The FFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.fft(data, 2, normalized=True) data = fftshift(data, dim=[-3, -2]) return data def ifft2c_old(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2-dimensional Inverse Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The IFFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.ifft(data, 2, normalized=True) data = fftshift(data, dim=[-3, -2]) return data def fft2c_new(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2 dimensional Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The FFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.view_as_real( torch.fft.fftn( # type: ignore torch.view_as_complex(data), dim=(-2, -1), norm="ortho" ) ) data = fftshift(data, dim=[-3, -2]) return data def ifft2c_new(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2-dimensional Inverse Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The IFFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.view_as_real( torch.fft.ifftn( # type: ignore torch.view_as_complex(data), dim=(-2, -1), norm="ortho" ) ) data = fftshift(data, dim=[-3, -2]) return data # Helper functions def roll_one_dim(x: torch.Tensor, shift: int, dim: int) -> torch.Tensor: """ Similar to roll but for only one dim. Args: x: A PyTorch tensor. shift: Amount to roll. dim: Which dimension to roll. Returns: Rolled version of x. """ shift = shift % x.size(dim) if shift == 0: return x left = x.narrow(dim, 0, x.size(dim) - shift) right = x.narrow(dim, x.size(dim) - shift, shift) return torch.cat((right, left), dim=dim) def roll( x: torch.Tensor, shift: List[int], dim: List[int], ) -> torch.Tensor: """ Similar to np.roll but applies to PyTorch Tensors. Args: x: A PyTorch tensor. shift: Amount to roll. dim: Which dimension to roll. Returns: Rolled version of x. """ if len(shift) != len(dim): raise ValueError("len(shift) must match len(dim)") for (s, d) in zip(shift, dim): x = roll_one_dim(x, s, d) return x def fftshift(x: torch.Tensor, dim: Optional[List[int]] = None) -> torch.Tensor: """ Similar to np.fft.fftshift but applies to PyTorch Tensors Args: x: A PyTorch tensor. dim: Which dimension to fftshift. Returns: fftshifted version of x. """ if dim is None: # this weird code is necessary for toch.jit.script typing dim = [0] * (x.dim()) for i in range(1, x.dim()): dim[i] = i # also necessary for torch.jit.script shift = [0] * len(dim) for i, dim_num in enumerate(dim): shift[i] = x.shape[dim_num] // 2 return roll(x, shift, dim) def ifftshift(x: torch.Tensor, dim: Optional[List[int]] = None) -> torch.Tensor: """ Similar to np.fft.ifftshift but applies to PyTorch Tensors Args: x: A PyTorch tensor. dim: Which dimension to ifftshift. Returns: ifftshifted version of x. """ if dim is None: # this weird code is necessary for toch.jit.script typing dim = [0] * (x.dim()) for i in range(1, x.dim()): dim[i] = i # also necessary for torch.jit.script shift = [0] * len(dim) for i, dim_num in enumerate(dim): shift[i] = (x.shape[dim_num] + 1) // 2 return roll(x, shift, dim)
5,535
25.236967
80
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/data/volume_sampler.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ from typing import List, Optional, Union import torch import torch.distributed as dist from fastmri.data.mri_data import CombinedSliceDataset, SliceDataset from torch.utils.data import Sampler class VolumeSampler(Sampler): """ Sampler for volumetric MRI data. Based on pytorch DistributedSampler, the difference is that all instances from the same MRI volume need to go to the same node for distributed training. Dataset example is a list of tuples (fname, instance), where fname is essentially the volume name (actually a filename). """ def __init__( self, dataset: Union[CombinedSliceDataset, SliceDataset], num_replicas: Optional[int] = None, rank: Optional[int] = None, shuffle: bool = True, seed: int = 0, ): """ Args: dataset: An MRI dataset (e.g., SliceData). num_replicas: Number of processes participating in distributed training. By default, :attr:`rank` is retrieved from the current distributed group. rank: Rank of the current process within :attr:`num_replicas`. By default, :attr:`rank` is retrieved from the current distributed group. shuffle: If ``True`` (default), sampler will shuffle the indices. seed: random seed used to shuffle the sampler if :attr:`shuffle=True`. This number should be identical across all processes in the distributed group. """ if num_replicas is None: if not dist.is_available(): raise RuntimeError("Requires distributed package to be available") num_replicas = dist.get_world_size() if rank is None: if not dist.is_available(): raise RuntimeError("Requires distributed package to be available") rank = dist.get_rank() self.dataset = dataset self.num_replicas = num_replicas self.rank = rank self.epoch = 0 self.shuffle = shuffle self.seed = seed # get all file names and split them based on number of processes self.all_volume_names = sorted( set(str(example[0]) for example in self.dataset.examples) ) self.all_volumes_split: List[List[str]] = [] for rank_num in range(self.num_replicas): self.all_volumes_split.append( [ self.all_volume_names[i] for i in range( rank_num, len(self.all_volume_names), self.num_replicas ) ] ) # get slice indices for each file name rank_indices: List[List[int]] = [[] for _ in range(self.num_replicas)] for i, example in enumerate(self.dataset.examples): vname = str(example[0]) for rank_num in range(self.num_replicas): if vname in self.all_volumes_split[rank_num]: rank_indices[rank_num].append(i) break # need to send equal number of samples to each process - take the max self.num_samples = max(len(indices) for indices in rank_indices) self.total_size = self.num_samples * self.num_replicas self.indices = rank_indices[self.rank] def __iter__(self): if self.shuffle: # deterministically shuffle based on epoch and seed g = torch.Generator() g.manual_seed(self.seed + self.epoch) ordering = torch.randperm(len(self.indices), generator=g).tolist() indices = [self.indices[i] for i in ordering] else: indices = self.indices # add extra samples to match num_samples repeat_times = self.num_samples // len(indices) indices = indices * repeat_times indices = indices + indices[: self.num_samples - len(indices)] assert len(indices) == self.num_samples return iter(indices) def __len__(self): return self.num_samples def set_epoch(self, epoch): self.epoch = epoch
4,332
36.678261
82
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/data/mri_data.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import logging import os import pickle import random import xml.etree.ElementTree as etree from pathlib import Path from typing import Callable, Dict, List, Optional, Sequence, Tuple, Union from warnings import warn import h5py import numpy as np import torch import yaml def et_query( root: etree.Element, qlist: Sequence[str], namespace: str = "http://www.ismrm.org/ISMRMRD", ) -> str: """ ElementTree query function. This can be used to query an xml document via ElementTree. It uses qlist for nested queries. Args: root: Root of the xml to search through. qlist: A list of strings for nested searches, e.g. ["Encoding", "matrixSize"] namespace: Optional; xml namespace to prepend query. Returns: The retrieved data as a string. """ s = "." prefix = "ismrmrd_namespace" ns = {prefix: namespace} for el in qlist: s = s + f"//{prefix}:{el}" value = root.find(s, ns) if value is None: raise RuntimeError("Element not found") return str(value.text) def fetch_dir( key: str, data_config_file: Union[str, Path, os.PathLike] = "fastmri_dirs.yaml" ) -> Path: """ Data directory fetcher. This is a brute-force simple way to configure data directories for a project. Simply overwrite the variables for `knee_path` and `brain_path` and this function will retrieve the requested subsplit of the data for use. Args: key: key to retrieve path from data_config_file. Expected to be in ("knee_path", "brain_path", "log_path"). data_config_file: Optional; Default path config file to fetch path from. Returns: The path to the specified directory. """ data_config_file = Path(data_config_file) if not data_config_file.is_file(): default_config = { "knee_path": "/path/to/knee", "brain_path": "/path/to/brain", "log_path": ".", } with open(data_config_file, "w") as f: yaml.dump(default_config, f) data_dir = default_config[key] warn( f"Path config at {data_config_file.resolve()} does not exist. " "A template has been created for you. " "Please enter the directory paths for your system to have defaults." ) else: with open(data_config_file, "r") as f: data_dir = yaml.safe_load(f)[key] return Path(data_dir) class CombinedSliceDataset(torch.utils.data.Dataset): """ A container for combining slice datasets. """ def __init__( self, roots: Sequence[Path], challenges: Sequence[str], transforms: Optional[Sequence[Optional[Callable]]] = None, sample_rates: Optional[Sequence[Optional[float]]] = None, volume_sample_rates: Optional[Sequence[Optional[float]]] = None, use_dataset_cache: bool = False, dataset_cache_file: Union[str, Path, os.PathLike] = "dataset_cache.pkl", num_cols: Optional[Tuple[int]] = None, ): """ Args: roots: Paths to the datasets. challenges: "singlecoil" or "multicoil" depending on which challenge to use. transforms: Optional; A sequence of callable objects that preprocesses the raw data into appropriate form. The transform function should take 'kspace', 'target', 'attributes', 'filename', and 'slice' as inputs. 'target' may be null for test data. sample_rates: Optional; A sequence of floats between 0 and 1. This controls what fraction of the slices should be loaded. When creating subsampled datasets either set sample_rates (sample by slices) or volume_sample_rates (sample by volumes) but not both. volume_sample_rates: Optional; A sequence of floats between 0 and 1. This controls what fraction of the volumes should be loaded. When creating subsampled datasets either set sample_rates (sample by slices) or volume_sample_rates (sample by volumes) but not both. use_dataset_cache: Whether to cache dataset metadata. This is very useful for large datasets like the brain data. dataset_cache_file: Optional; A file in which to cache dataset information for faster load times. num_cols: Optional; If provided, only slices with the desired number of columns will be considered. """ if sample_rates is not None and volume_sample_rates is not None: raise ValueError( "either set sample_rates (sample by slices) or volume_sample_rates (sample by volumes) but not both" ) if transforms is None: transforms = [None] * len(roots) if sample_rates is None: sample_rates = [None] * len(roots) if volume_sample_rates is None: volume_sample_rates = [None] * len(roots) if not ( len(roots) == len(transforms) == len(challenges) == len(sample_rates) == len(volume_sample_rates) ): raise ValueError( "Lengths of roots, transforms, challenges, sample_rates do not match" ) self.datasets = [] self.examples: List[Tuple[Path, int, Dict[str, object]]] = [] for i in range(len(roots)): self.datasets.append( SliceDataset( root=roots[i], transform=transforms[i], challenge=challenges[i], sample_rate=sample_rates[i], volume_sample_rate=volume_sample_rates[i], use_dataset_cache=use_dataset_cache, dataset_cache_file=dataset_cache_file, num_cols=num_cols, ) ) self.examples = self.examples + self.datasets[-1].examples def __len__(self): return sum(len(dataset) for dataset in self.datasets) def __getitem__(self, i): for dataset in self.datasets: if i < len(dataset): return dataset[i] else: i = i - len(dataset) class SliceDataset(torch.utils.data.Dataset): """ A PyTorch Dataset that provides access to MR image slices. """ def __init__( self, root: Union[str, Path, os.PathLike], challenge: str, transform: Optional[Callable] = None, use_dataset_cache: bool = False, sample_rate: Optional[float] = None, volume_sample_rate: Optional[float] = None, dataset_cache_file: Union[str, Path, os.PathLike] = "dataset_cache.pkl", num_cols: Optional[Tuple[int]] = None, ): """ Args: root: Path to the dataset. challenge: "singlecoil" or "multicoil" depending on which challenge to use. transform: Optional; A callable object that pre-processes the raw data into appropriate form. The transform function should take 'kspace', 'target', 'attributes', 'filename', and 'slice' as inputs. 'target' may be null for test data. use_dataset_cache: Whether to cache dataset metadata. This is very useful for large datasets like the brain data. sample_rate: Optional; A float between 0 and 1. This controls what fraction of the slices should be loaded. Defaults to 1 if no value is given. When creating a sampled dataset either set sample_rate (sample by slices) or volume_sample_rate (sample by volumes) but not both. volume_sample_rate: Optional; A float between 0 and 1. This controls what fraction of the volumes should be loaded. Defaults to 1 if no value is given. When creating a sampled dataset either set sample_rate (sample by slices) or volume_sample_rate (sample by volumes) but not both. dataset_cache_file: Optional; A file in which to cache dataset information for faster load times. num_cols: Optional; If provided, only slices with the desired number of columns will be considered. """ if challenge not in ("singlecoil", "multicoil"): raise ValueError('challenge should be either "singlecoil" or "multicoil"') if sample_rate is not None and volume_sample_rate is not None: raise ValueError( "either set sample_rate (sample by slices) or volume_sample_rate (sample by volumes) but not both" ) self.dataset_cache_file = Path(dataset_cache_file) self.transform = transform self.recons_key = ( "reconstruction_esc" if challenge == "singlecoil" else "reconstruction_rss" ) self.examples = [] # set default sampling mode if none given if sample_rate is None: sample_rate = 1.0 if volume_sample_rate is None: volume_sample_rate = 1.0 # load dataset cache if we have and user wants to use it if self.dataset_cache_file.exists() and use_dataset_cache: with open(self.dataset_cache_file, "rb") as f: dataset_cache = pickle.load(f) else: dataset_cache = {} # check if our dataset is in the cache # if there, use that metadata, if not, then regenerate the metadata if dataset_cache.get(root) is None or not use_dataset_cache: files = list(Path(root).iterdir()) for fname in sorted(files): metadata, num_slices = self._retrieve_metadata(fname) self.examples += [ (fname, slice_ind, metadata) for slice_ind in range(num_slices) ] if dataset_cache.get(root) is None and use_dataset_cache: dataset_cache[root] = self.examples logging.info(f"Saving dataset cache to {self.dataset_cache_file}.") with open(self.dataset_cache_file, "wb") as f: pickle.dump(dataset_cache, f) else: logging.info(f"Using dataset cache from {self.dataset_cache_file}.") self.examples = dataset_cache[root] # subsample if desired if sample_rate < 1.0: # sample by slice random.shuffle(self.examples) num_examples = round(len(self.examples) * sample_rate) self.examples = self.examples[:num_examples] elif volume_sample_rate < 1.0: # sample by volume vol_names = sorted(list(set([f[0].stem for f in self.examples]))) random.shuffle(vol_names) num_volumes = round(len(vol_names) * volume_sample_rate) sampled_vols = vol_names[:num_volumes] self.examples = [ example for example in self.examples if example[0].stem in sampled_vols ] if num_cols: self.examples = [ ex for ex in self.examples if ex[2]["encoding_size"][1] in num_cols # type: ignore ] def _retrieve_metadata(self, fname): with h5py.File(fname, "r") as hf: et_root = etree.fromstring(hf["ismrmrd_header"][()]) enc = ["encoding", "encodedSpace", "matrixSize"] enc_size = ( int(et_query(et_root, enc + ["x"])), int(et_query(et_root, enc + ["y"])), int(et_query(et_root, enc + ["z"])), ) rec = ["encoding", "reconSpace", "matrixSize"] recon_size = ( int(et_query(et_root, rec + ["x"])), int(et_query(et_root, rec + ["y"])), int(et_query(et_root, rec + ["z"])), ) lims = ["encoding", "encodingLimits", "kspace_encoding_step_1"] enc_limits_center = int(et_query(et_root, lims + ["center"])) enc_limits_max = int(et_query(et_root, lims + ["maximum"])) + 1 padding_left = enc_size[1] // 2 - enc_limits_center padding_right = padding_left + enc_limits_max num_slices = hf["kspace"].shape[0] metadata = { "padding_left": padding_left, "padding_right": padding_right, "encoding_size": enc_size, "recon_size": recon_size, } return metadata, num_slices def __len__(self): return len(self.examples) def __getitem__(self, i: int): fname, dataslice, metadata = self.examples[i] with h5py.File(fname, "r") as hf: kspace = hf["kspace"][dataslice] mask = np.asarray(hf["mask"]) if "mask" in hf else None target = hf[self.recons_key][dataslice] if self.recons_key in hf else None attrs = dict(hf.attrs) attrs.update(metadata) if self.transform is None: sample = (kspace, mask, target, attrs, fname.name, dataslice) else: sample = self.transform(kspace, mask, target, attrs, fname.name, dataslice) return sample
13,630
36.759003
116
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/data/subsample.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import contextlib from typing import Optional, Sequence, Tuple, Union import numpy as np import torch @contextlib.contextmanager def temp_seed(rng: np.random, seed: Optional[Union[int, Tuple[int, ...]]]): if seed is None: try: yield finally: pass else: state = rng.get_state() rng.seed(seed) try: yield finally: rng.set_state(state) class MaskFunc: """ An object for GRAPPA-style sampling masks. This crates a sampling mask that densely samples the center while subsampling outer k-space regions based on the undersampling factor. """ def __init__(self, center_fractions: Sequence[float], accelerations: Sequence[int]): """ Args: center_fractions: Fraction of low-frequency columns to be retained. If multiple values are provided, then one of these numbers is chosen uniformly each time. accelerations: Amount of under-sampling. This should have the same length as center_fractions. If multiple values are provided, then one of these is chosen uniformly each time. """ if not len(center_fractions) == len(accelerations): raise ValueError( "Number of center fractions should match number of accelerations" ) self.center_fractions = center_fractions self.accelerations = accelerations self.rng = np.random.RandomState() # pylint: disable=no-member def __call__( self, shape: Sequence[int], seed: Optional[Union[int, Tuple[int, ...]]] = None ) -> torch.Tensor: raise NotImplementedError def choose_acceleration(self): """Choose acceleration based on class parameters.""" choice = self.rng.randint(0, len(self.accelerations)) center_fraction = self.center_fractions[choice] acceleration = self.accelerations[choice] return center_fraction, acceleration class RandomMaskFunc(MaskFunc): """ RandomMaskFunc creates a sub-sampling mask of a given shape. The mask selects a subset of columns from the input k-space data. If the k-space data has N columns, the mask picks out: 1. N_low_freqs = (N * center_fraction) columns in the center corresponding to low-frequencies. 2. The other columns are selected uniformly at random with a probability equal to: prob = (N / acceleration - N_low_freqs) / (N - N_low_freqs). This ensures that the expected number of columns selected is equal to (N / acceleration). It is possible to use multiple center_fractions and accelerations, in which case one possible (center_fraction, acceleration) is chosen uniformly at random each time the RandomMaskFunc object is called. For example, if accelerations = [4, 8] and center_fractions = [0.08, 0.04], then there is a 50% probability that 4-fold acceleration with 8% center fraction is selected and a 50% probability that 8-fold acceleration with 4% center fraction is selected. """ def __call__( self, shape: Sequence[int], seed: Optional[Union[int, Tuple[int, ...]]] = None ) -> torch.Tensor: """ Create the mask. Args: shape: The shape of the mask to be created. The shape should have at least 3 dimensions. Samples are drawn along the second last dimension. seed: Seed for the random number generator. Setting the seed ensures the same mask is generated each time for the same shape. The random state is reset afterwards. Returns: A mask of the specified shape. """ if len(shape) < 3: raise ValueError("Shape should have 3 or more dimensions") with temp_seed(self.rng, seed): num_cols = shape[-2] center_fraction, acceleration = self.choose_acceleration() # create the mask num_low_freqs = int(round(num_cols * center_fraction)) prob = (num_cols / acceleration - num_low_freqs) / ( num_cols - num_low_freqs ) mask = self.rng.uniform(size=num_cols) < prob pad = (num_cols - num_low_freqs + 1) // 2 mask[pad : pad + num_low_freqs] = True # reshape the mask mask_shape = [1 for _ in shape] mask_shape[-2] = num_cols mask = torch.from_numpy(mask.reshape(*mask_shape).astype(np.float32)) return mask class EquispacedMaskFunc(MaskFunc): """ EquispacedMaskFunc creates a sub-sampling mask of a given shape. The mask selects a subset of columns from the input k-space data. If the k-space data has N columns, the mask picks out: 1. N_low_freqs = (N * center_fraction) columns in the center corresponding tovlow-frequencies. 2. The other columns are selected with equal spacing at a proportion that reaches the desired acceleration rate taking into consideration the number of low frequencies. This ensures that the expected number of columns selected is equal to (N / acceleration) It is possible to use multiple center_fractions and accelerations, in which case one possible (center_fraction, acceleration) is chosen uniformly at random each time the EquispacedMaskFunc object is called. Note that this function may not give equispaced samples (documented in https://github.com/facebookresearch/fastMRI/issues/54), which will require modifications to standard GRAPPA approaches. Nonetheless, this aspect of the function has been preserved to match the public multicoil data. """ def __call__( self, shape: Sequence[int], seed: Optional[Union[int, Tuple[int, ...]]] = None ) -> torch.Tensor: """ Args: shape: The shape of the mask to be created. The shape should have at least 3 dimensions. Samples are drawn along the second last dimension. seed: Seed for the random number generator. Setting the seed ensures the same mask is generated each time for the same shape. The random state is reset afterwards. Returns: A mask of the specified shape. """ if len(shape) < 3: raise ValueError("Shape should have 3 or more dimensions") with temp_seed(self.rng, seed): center_fraction, acceleration = self.choose_acceleration() num_cols = shape[-2] num_low_freqs = int(round(num_cols * center_fraction)) # create the mask mask = np.zeros(num_cols, dtype=np.float32) pad = (num_cols - num_low_freqs + 1) // 2 mask[pad : pad + num_low_freqs] = True # determine acceleration rate by adjusting for the number of low frequencies adjusted_accel = (acceleration * (num_low_freqs - num_cols)) / ( num_low_freqs * acceleration - num_cols ) offset = self.rng.randint(0, round(adjusted_accel)) accel_samples = np.arange(offset, num_cols - 1, adjusted_accel) accel_samples = np.around(accel_samples).astype(np.uint) mask[accel_samples] = True # reshape the mask mask_shape = [1 for _ in shape] mask_shape[-2] = num_cols mask = torch.from_numpy(mask.reshape(*mask_shape).astype(np.float32)) return mask def create_mask_for_mask_type( mask_type_str: str, center_fractions: Sequence[float], accelerations: Sequence[int], ) -> MaskFunc: """ Creates a mask of the specified type. Args: center_fractions: What fraction of the center of k-space to include. accelerations: What accelerations to apply. """ if mask_type_str == "random": return RandomMaskFunc(center_fractions, accelerations) elif mask_type_str == "equispaced": return EquispacedMaskFunc(center_fractions, accelerations) else: raise Exception(f"{mask_type_str} not supported")
8,448
36.887892
88
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/fastmri/data/transforms.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ from typing import Dict, Optional, Sequence, Tuple, Union import fastmri import numpy as np import torch from .subsample import MaskFunc def to_tensor(data: np.ndarray) -> torch.Tensor: """ Convert numpy array to PyTorch tensor. For complex arrays, the real and imaginary parts are stacked along the last dimension. Args: data: Input numpy array. Returns: PyTorch version of data. """ if np.iscomplexobj(data): data = np.stack((data.real, data.imag), axis=-1) return torch.from_numpy(data) def tensor_to_complex_np(data: torch.Tensor) -> np.ndarray: """ Converts a complex torch tensor to numpy array. Args: data: Input data to be converted to numpy. Returns: Complex numpy version of data. """ data = data.numpy() return data[..., 0] + 1j * data[..., 1] def apply_mask( data: torch.Tensor, mask_func: MaskFunc, seed: Optional[Union[int, Tuple[int, ...]]] = None, padding: Optional[Sequence[int]] = None, ) -> Tuple[torch.Tensor, torch.Tensor]: """ Subsample given k-space by multiplying with a mask. Args: data: The input k-space data. This should have at least 3 dimensions, where dimensions -3 and -2 are the spatial dimensions, and the final dimension has size 2 (for complex values). mask_func: A function that takes a shape (tuple of ints) and a random number seed and returns a mask. seed: Seed for the random number generator. padding: Padding value to apply for mask. Returns: tuple containing: masked data: Subsampled k-space data mask: The generated mask """ shape = np.array(data.shape) shape[:-3] = 1 mask = mask_func(shape, seed) if padding is not None: mask[:, :, : padding[0]] = 0 mask[:, :, padding[1] :] = 0 # padding value inclusive on right of zeros masked_data = data * mask + 0.0 # the + 0.0 removes the sign of the zeros return masked_data, mask def mask_center(x: torch.Tensor, mask_from: int, mask_to: int) -> torch.Tensor: """ Initializes a mask with the center filled in. Args: mask_from: Part of center to start filling. mask_to: Part of center to end filling. Returns: A mask with the center filled. """ mask = torch.zeros_like(x) mask[:, :, :, mask_from:mask_to] = x[:, :, :, mask_from:mask_to] return mask def center_crop(data: torch.Tensor, shape: Tuple[int, int]) -> torch.Tensor: """ Apply a center crop to the input real image or batch of real images. Args: data: The input tensor to be center cropped. It should have at least 2 dimensions and the cropping is applied along the last two dimensions. shape: The output shape. The shape should be smaller than the corresponding dimensions of data. Returns: The center cropped image. """ if not (0 < shape[0] <= data.shape[-2] and 0 < shape[1] <= data.shape[-1]): raise ValueError("Invalid shapes.") w_from = (data.shape[-2] - shape[0]) // 2 h_from = (data.shape[-1] - shape[1]) // 2 w_to = w_from + shape[0] h_to = h_from + shape[1] return data[..., w_from:w_to, h_from:h_to] def complex_center_crop(data: torch.Tensor, shape: Tuple[int, int]) -> torch.Tensor: """ Apply a center crop to the input image or batch of complex images. Args: data: The complex input tensor to be center cropped. It should have at least 3 dimensions and the cropping is applied along dimensions -3 and -2 and the last dimensions should have a size of 2. shape: The output shape. The shape should be smaller than the corresponding dimensions of data. Returns: The center cropped image """ if not (0 < shape[0] <= data.shape[-3] and 0 < shape[1] <= data.shape[-2]): raise ValueError("Invalid shapes.") w_from = (data.shape[-3] - shape[0]) // 2 h_from = (data.shape[-2] - shape[1]) // 2 w_to = w_from + shape[0] h_to = h_from + shape[1] return data[..., w_from:w_to, h_from:h_to, :] def center_crop_to_smallest( x: torch.Tensor, y: torch.Tensor ) -> Tuple[torch.Tensor, torch.Tensor]: """ Apply a center crop on the larger image to the size of the smaller. The minimum is taken over dim=-1 and dim=-2. If x is smaller than y at dim=-1 and y is smaller than x at dim=-2, then the returned dimension will be a mixture of the two. Args: x: The first image. y: The second image. Returns: tuple of tensors x and y, each cropped to the minimim size. """ smallest_width = min(x.shape[-1], y.shape[-1]) smallest_height = min(x.shape[-2], y.shape[-2]) x = center_crop(x, (smallest_height, smallest_width)) y = center_crop(y, (smallest_height, smallest_width)) return x, y def normalize( data: torch.Tensor, mean: Union[float, torch.Tensor], stddev: Union[float, torch.Tensor], eps: Union[float, torch.Tensor] = 0.0, ) -> torch.Tensor: """ Normalize the given tensor. Applies the formula (data - mean) / (stddev + eps). Args: data: Input data to be normalized. mean: Mean value. stddev: Standard deviation. eps: Added to stddev to prevent dividing by zero. Returns: Normalized tensor. """ return (data - mean) / (stddev + eps) def normalize_instance( data: torch.Tensor, eps: Union[float, torch.Tensor] = 0.0 ) -> Tuple[torch.Tensor, Union[torch.Tensor], Union[torch.Tensor]]: """ Normalize the given tensor with instance norm/ Applies the formula (data - mean) / (stddev + eps), where mean and stddev are computed from the data itself. Args: data: Input data to be normalized eps: Added to stddev to prevent dividing by zero. Returns: torch.Tensor: Normalized tensor """ mean = data.mean() std = data.std() return normalize(data, mean, std, eps), mean, std class UnetDataTransform: """ Data Transformer for training U-Net models. """ def __init__( self, which_challenge: str, mask_func: Optional[MaskFunc] = None, use_seed: bool = True, ): """ Args: which_challenge: Challenge from ("singlecoil", "multicoil"). mask_func: Optional; A function that can create a mask of appropriate shape. use_seed: If true, this class computes a pseudo random number generator seed from the filename. This ensures that the same mask is used for all the slices of a given volume every time. """ if which_challenge not in ("singlecoil", "multicoil"): raise ValueError("Challenge should either be 'singlecoil' or 'multicoil'") self.mask_func = mask_func self.which_challenge = which_challenge self.use_seed = use_seed def __call__( self, kspace: np.ndarray, mask: np.ndarray, target: np.ndarray, attrs: Dict, fname: str, slice_num: int, ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, str, int, float]: """ Args: kspace: Input k-space of shape (num_coils, rows, cols) for multi-coil data or (rows, cols) for single coil data. mask: Mask from the test dataset. target: Target image. attrs: Acquisition related information stored in the HDF5 object. fname: File name. slice_num: Serial number of the slice. Returns: tuple containing: image: Zero-filled input image. target: Target image converted to a torch.Tensor. mean: Mean value used for normalization. std: Standard deviation value used for normalization. fname: File name. slice_num: Serial number of the slice. """ kspace = to_tensor(kspace) # check for max value max_value = attrs["max"] if "max" in attrs.keys() else 0.0 # apply mask if self.mask_func: seed = None if not self.use_seed else tuple(map(ord, fname)) masked_kspace, mask = apply_mask(kspace, self.mask_func, seed) else: masked_kspace = kspace # inverse Fourier transform to get zero filled solution image = fastmri.ifft2c(masked_kspace) # crop input to correct size if target is not None: crop_size = (target.shape[-2], target.shape[-1]) else: crop_size = (attrs["recon_size"][0], attrs["recon_size"][1]) # check for FLAIR 203 if image.shape[-2] < crop_size[1]: crop_size = (image.shape[-2], image.shape[-2]) image = complex_center_crop(image, crop_size) # absolute value image = fastmri.complex_abs(image) # apply Root-Sum-of-Squares if multicoil data if self.which_challenge == "multicoil": image = fastmri.rss(image) # normalize input image, mean, std = normalize_instance(image, eps=1e-11) image = image.clamp(-6, 6) # normalize target if target is not None: target = to_tensor(target) target = center_crop(target, crop_size) target = normalize(target, mean, std, eps=1e-11) target = target.clamp(-6, 6) else: target = torch.Tensor([0]) return image, target, mean, std, fname, slice_num, max_value class VarNetDataTransform: """ Data Transformer for training VarNet models. """ def __init__(self, mask_func: Optional[MaskFunc] = None, use_seed: bool = True): """ Args: mask_func: Optional; A function that can create a mask of appropriate shape. Defaults to None. use_seed: If True, this class computes a pseudo random number generator seed from the filename. This ensures that the same mask is used for all the slices of a given volume every time. """ self.mask_func = mask_func self.use_seed = use_seed def __call__( self, kspace: np.ndarray, mask: np.ndarray, target: np.ndarray, attrs: Dict, fname: str, slice_num: int, ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, str, int, float, torch.Tensor]: """ Args: kspace: Input k-space of shape (num_coils, rows, cols) for multi-coil data. mask: Mask from the test dataset. target: Target image. attrs: Acquisition related information stored in the HDF5 object. fname: File name. slice_num: Serial number of the slice. Returns: tuple containing: masked_kspace: k-space after applying sampling mask. mask: The applied sampling mask target: The target image (if applicable). fname: File name. slice_num: The slice index. max_value: Maximum image value. crop_size: The size to crop the final image. """ if target is not None: target = to_tensor(target) max_value = attrs["max"] else: target = torch.tensor(0) max_value = 0.0 kspace = to_tensor(kspace) seed = None if not self.use_seed else tuple(map(ord, fname)) acq_start = attrs["padding_left"] acq_end = attrs["padding_right"] crop_size = torch.tensor([attrs["recon_size"][0], attrs["recon_size"][1]]) if self.mask_func: masked_kspace, mask = apply_mask( kspace, self.mask_func, seed, (acq_start, acq_end) ) else: masked_kspace = kspace shape = np.array(kspace.shape) num_cols = shape[-2] shape[:-3] = 1 mask_shape = [1] * len(shape) mask_shape[-2] = num_cols mask = torch.from_numpy(mask.reshape(*mask_shape).astype(np.float32)) mask = mask.reshape(*mask_shape) mask[:, :, :acq_start] = 0 mask[:, :, acq_end:] = 0 return ( masked_kspace, mask.byte(), target, fname, slice_num, max_value, crop_size, )
12,887
30.205811
88
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/networks/img2img_mixer.py
import torch import torch.nn as nn import torch.nn.functional as F from torch.nn import init import torch.nn.init as init import numpy as np import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbedding(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchExpansion(nn.Module): def __init__(self, dim_scale, channel_dim, img_channels, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, dim_scale**2* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): """ x: B, H, W, C """ x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int, f_hidden: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c w h -> b c h w"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels*f_hidden) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Img2Img_Mixer(nn.Module): def __init__( self, img_size: int = 320, img_channels: int = 1, patch_size: int = 4, embed_dim: int = 128, num_layers: int = 16, f_hidden: int = 8, ): super().__init__() self.patch_embed = PatchEmbedding(patch_size, embed_dim, img_channels) layers = [ Mixer(img_size//patch_size, embed_dim, f_hidden) for _ in range(num_layers)] self.mixer_layers = nn.Sequential(*layers) self.patch_expand = PatchExpansion(patch_size, embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.patch_embed(x) x = self.mixer_layers(x) x = self.patch_expand(x) return x
3,714
26.932331
127
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/networks/vision_transformer.py
''' This code is modified from https://github.com/facebookresearch/convit. To adapt the vit/convit to image reconstruction, variable input sizes, and patch sizes for both spatial dimensions. ''' import torch import torch.nn as nn from functools import partial import torch.nn.functional as F from timm.models.helpers import load_pretrained from timm.models.layers import DropPath, to_2tuple, trunc_normal_ class Mlp(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.fc2 = nn.Linear(hidden_features, out_features) self.drop = nn.Dropout(drop) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) x = self.drop(x) return x class GPSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., locality_strength=1., use_local_init=True, grid_size=None): super().__init__() self.num_heads = num_heads self.dim = dim head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.q = nn.Linear(dim, dim, bias=qkv_bias) self.k = nn.Linear(dim, dim, bias=qkv_bias) self.v = nn.Linear(dim, dim, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.pos_proj = nn.Linear(3, num_heads) self.proj_drop = nn.Dropout(proj_drop) self.locality_strength = locality_strength self.gating_param = nn.Parameter(1*torch.ones(self.num_heads)) self.apply(self._init_weights) if use_local_init: self.local_init(locality_strength=locality_strength) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention(self, x): B, N, C = x.shape k = self.k(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) pos_score = self.pos_proj(self.rel_indices).expand(B, -1, -1,-1).permute(0,3,1,2) patch_score = (q @ k.transpose(-2, -1)) * self.scale patch_score = patch_score.softmax(dim=-1) pos_score = pos_score.softmax(dim=-1) gating = self.gating_param.view(1,-1,1,1) attn = (1.-torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score attn = attn / attn.sum(dim=-1).unsqueeze(-1) attn = self.attn_drop(attn) return attn def get_attention_map(self, x, return_map = False): attn_map = self.get_attention(x).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def local_init(self, locality_strength=1.): self.v.weight.data.copy_(torch.eye(self.dim)) locality_distance = 1 #max(1,1/locality_strength**.5) kernel_size = int(self.num_heads**.5) center = (kernel_size-1)/2 if kernel_size%2==0 else kernel_size//2 for h1 in range(kernel_size): for h2 in range(kernel_size): position = h1+kernel_size*h2 self.pos_proj.weight.data[position,2] = -1 self.pos_proj.weight.data[position,1] = 2*(h1-center)*locality_distance self.pos_proj.weight.data[position,0] = 2*(h2-center)*locality_distance self.pos_proj.weight.data *= locality_strength def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.v.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape if not hasattr(self, 'rel_indices') or self.rel_indices.size(1)!=N: self.get_rel_indices() attn = self.get_attention(x) v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class MHSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., grid_size=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) self.apply(self._init_weights) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention_map(self, x, return_map = False): self.get_rel_indices() B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn_map = (q @ k.transpose(-2, -1)) * self.scale attn_map = attn_map.softmax(dim=-1).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.qkv.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Block(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs): super().__init__() self.norm1 = norm_layer(dim) self.use_gpsa = use_gpsa if self.use_gpsa: self.attn = GPSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) else: self.attn = MHSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) def forward(self, x, grid_size): self.attn.current_grid_size = grid_size x = x + self.drop_path(self.attn(self.norm1(x))) x = x + self.drop_path(self.mlp(self.norm2(x))) return x class PatchEmbed(nn.Module): """ Image to Patch Embedding, from timm """ def __init__(self, patch_size, in_chans, embed_dim): super().__init__() self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) self.apply(self._init_weights) def forward(self, x): x = self.proj(x) return x def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) class VisionTransformer(nn.Module): """ Vision Transformer """ def __init__(self, avrg_img_size=320, patch_size=16, in_chans=1, embed_dim=64, depth=8, num_heads=9, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, global_pool=None, gpsa_interval=[-1, -1], locality_strength=1., use_pos_embed=True): super().__init__() self.depth = depth embed_dim *= num_heads self.num_features = embed_dim # num_features for consistency with other models self.locality_strength = locality_strength self.use_pos_embed = use_pos_embed if isinstance(avrg_img_size, int): img_size = to_2tuple(avrg_img_size) if isinstance(patch_size, int): self.patch_size = to_2tuple(patch_size) else: self.patch_size = patch_size self.in_chans = in_chans self.patch_embed = PatchEmbed( patch_size=self.patch_size, in_chans=in_chans, embed_dim=embed_dim) self.pos_drop = nn.Dropout(p=drop_rate) if self.use_pos_embed: self.pos_embed = nn.Parameter( torch.zeros(1, embed_dim, img_size[0] // self.patch_size[0], img_size[1] // self.patch_size[1]) ) trunc_normal_(self.pos_embed, std=.02) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule self.blocks = nn.ModuleList([ Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=True, locality_strength=locality_strength) if i>=gpsa_interval[0]-1 and i<gpsa_interval[1] else Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=False,) for i in range(depth)]) self.norm = norm_layer(embed_dim) # head self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')] self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def seq2img(self, x, img_size): """ Transforms sequence back into image space, input dims: [batch_size, num_patches, channels] output dims: [batch_size, channels, H, W] """ x = x.view(x.shape[0], x.shape[1], self.in_chans, self.patch_size[0], self.patch_size[1]) x = x.chunk(x.shape[1], dim=1) x = torch.cat(x, dim=4).permute(0,1,2,4,3) x = x.chunk(img_size[0]//self.patch_size[0], dim=3) x = torch.cat(x, dim=4).permute(0,1,2,4,3).squeeze(1) return x self.head.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) @torch.jit.ignore def no_weight_decay(self,): return {'pos_embed'} def get_head(self,): return self.head def reset_head(self,): self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def forward_features(self, x, k=None): x = self.patch_embed(x) _, _, H, W = x.shape if self.use_pos_embed: pos_embed = F.interpolate(self.pos_embed, size=[H, W], mode='bilinear', align_corners = False) x = x + pos_embed x = x.flatten(2).transpose(1, 2) x = self.pos_drop(x) for u, blk in enumerate(self.blocks): x = blk(x, (H, W)) if k is not None and u == k: self.attention_map = blk.attn.get_attention_map(x, return_map = True) x = self.norm(x) return x def forward(self, x, k=None): _, _, H, W = x.shape x = self.forward_features(x, k) x = self.head(x) x = self.seq2img(x, (H, W)) return x
15,082
39.007958
186
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/networks/recon_net.py
import torch.nn as nn import torch.nn.functional as F from math import ceil, floor from .unet import Unet from .vision_transformer import VisionTransformer class ReconNet(nn.Module): def __init__(self, net): super().__init__() self.net = net def pad(self, x): _, _, h, w = x.shape hp, wp = self.net.patch_size f1 = ( (wp - w % wp) % wp ) / 2 f2 = ( (hp - h % hp) % hp ) / 2 wpad = [floor(f1), ceil(f1)] hpad = [floor(f2), ceil(f2)] x = F.pad(x, wpad+hpad) return x, wpad, hpad def unpad(self, x, wpad, hpad): return x[..., hpad[0] : x.shape[-2]-hpad[1], wpad[0] : x.shape[-1]-wpad[1]] def norm(self, x): mean = x.view(x.shape[0], 1, 1, -1).mean(-1, keepdim=True) std = x.view(x.shape[0], 1, 1, -1,).std(-1, keepdim=True) x = (x-mean)/std return x, mean, std def unnorm(self, x, mean, std): return x * std + mean def vit_forward(self, x, k=None): x, wpad, hpad = self.pad(x) x, mean, std = self.norm(x) x = self.net(x, k) x = self.unnorm(x, mean, std) x = self.unpad(x, wpad, hpad) return x def unet_forward(self, x): x, mean, std = self.norm(x) x = self.net(x) x = self.unnorm(x, mean, std) return x def mixer_forward(self, x): x, mean, std = self.norm(x) x = self.net(x) x = self.unnorm(x, mean, std) return x def forward(self, x, k=None): if isinstance(self.net, Unet): return self.unet_forward(x) elif isinstance(self.net, VisionTransformer): return self.vit_forward(x, k) else: return self.mixer_forward(x)
1,932
25.847222
90
py
imaging_MLPs
imaging_MLPs-master/compressed_sensing/networks/unet.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from torch import nn from torch.nn import functional as F class Unet(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int, out_chans: int, chans: int = 32, num_pool_layers: int = 4, drop_prob: float = 0.0, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append( nn.Sequential( ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1), ) ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
5,979
31.677596
88
py
imaging_MLPs
imaging_MLPs-master/untrained/networks/original_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbeddings(nn.Module): def __init__( self, patch_size: int, hidden_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=hidden_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b (h w) c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchEmbeddings_transpose(nn.Module): def __init__( self, patch_size: int, hidden_dim: int, channels: int, d: int ): super().__init__() self.proj_transpose = nn.Sequential( Rearrange("b (h w) c -> b c h w", h=d), nn.ConvTranspose2d( in_channels=hidden_dim, out_channels=channels, kernel_size=patch_size, stride=patch_size ) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj_transpose(x) class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class MixerBlock(nn.Module): def __init__( self, num_patches: int, num_channels: int, tokens_hidden_dim: int, channels_hidden_dim: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b p c -> b c p"), MLPBlock(num_patches, tokens_hidden_dim), Rearrange("b c p -> b p c") ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, channels_hidden_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Original_Mixer(nn.Module): def __init__( self, image_size: int = 256, channels: int = 1, patch_size: int = 4, num_layers: int = 8, hidden_dim: int = 128, tokens_hidden_dim: int = 96, channels_hidden_dim: int = 256 ): super().__init__() num_patches = (image_size // patch_size) ** 2 d=(image_size-patch_size)//patch_size + 1 self.embed = PatchEmbeddings(patch_size, hidden_dim, channels) layers = [ MixerBlock( num_patches=num_patches, num_channels=hidden_dim, tokens_hidden_dim=tokens_hidden_dim, channels_hidden_dim=channels_hidden_dim ) for _ in range(num_layers) ] self.layers = nn.Sequential(*layers) self.norm = nn.LayerNorm(hidden_dim) self.embed_transpose = PatchEmbeddings_transpose(patch_size, hidden_dim, channels, d) def forward(self, x: torch.Tensor) -> torch.Tensor: b, c, h, w = x.shape x = self.embed(x) x = self.layers(x) x = self.norm(x) x = self.embed_transpose(x) return x
3,674
26.840909
93
py
imaging_MLPs
imaging_MLPs-master/untrained/networks/img2img_mixer.py
import torch import torch.nn as nn from torch.nn import init import torch.nn.init as init import einops from einops.layers.torch import Rearrange from einops import rearrange class PatchEmbedding(nn.Module): def __init__( self, patch_size: int, embed_dim: int, channels: int ): super().__init__() self.proj = nn.Sequential( nn.Conv2d( in_channels=channels, out_channels=embed_dim, kernel_size=patch_size, stride=patch_size ), Rearrange("b c h w -> b h w c") ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.proj(x) class PatchExpansion(nn.Module): def __init__(self, dim_scale, channel_dim, img_channels, norm_layer=nn.LayerNorm): super().__init__() self.dim_scale = dim_scale self.expand = nn.Linear(channel_dim, dim_scale**2* channel_dim, bias=False) self.output_dim = channel_dim self.norm = norm_layer(channel_dim) self.output = nn.Conv2d(in_channels=channel_dim,out_channels=img_channels ,kernel_size=1,bias=False) def forward(self, x): x = self.expand(x) B, H, W, C = x.shape x = rearrange(x, 'b h w (p1 p2 c)-> b (h p1) (w p2) c', p1=self.dim_scale, p2=self.dim_scale, c=C//(self.dim_scale**2)) x = x.view(B,-1,self.output_dim) x= self.norm(x) x = x.view(B,H*self.dim_scale, W*self.dim_scale,-1) x = x.permute(0,3,1,2) x = self.output(x) return x class MLPBlock(nn.Module): def __init__(self, input_dim: int, hidden_dim: int): super().__init__() self.model = nn.Sequential( nn.Linear(input_dim, hidden_dim), nn.GELU(), nn.Linear(hidden_dim, input_dim) ) def forward(self, x: torch.Tensor) -> torch.Tensor: return self.model(x) class Mixer(nn.Module): def __init__( self, num_patches: int, num_channels: int, f_hidden: int ): super().__init__() self.token_mixing = nn.Sequential( nn.LayerNorm(num_channels), Rearrange("b h w c -> b c w h"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c w h -> b c h w"), MLPBlock(num_patches, num_patches*f_hidden), Rearrange("b c h w -> b h w c"), ) self.channel_mixing = nn.Sequential( nn.LayerNorm(num_channels), MLPBlock(num_channels, num_channels*f_hidden) ) def forward(self, x: torch.Tensor) -> torch.Tensor: x = x + self.token_mixing(x) x = x + self.channel_mixing(x) return x class Img2Img_Mixer(nn.Module): def __init__( self, img_size: int = 256, img_channels: int = 1, patch_size: int = 4, embed_dim: int = 128, num_layers: int = 16, f_hidden: int = 8, ): super().__init__() self.patch_embed = PatchEmbedding(patch_size, embed_dim, img_channels) layers = [ Mixer(img_size//patch_size, embed_dim, f_hidden) for _ in range(num_layers)] self.mixer_layers = nn.Sequential(*layers) self.patch_expand = PatchExpansion(patch_size, embed_dim, img_channels) def forward(self, x: torch.Tensor) -> torch.Tensor: x = self.patch_embed(x) x = self.mixer_layers(x) x = self.patch_expand(x) return x
3,618
27.054264
127
py
imaging_MLPs
imaging_MLPs-master/untrained/networks/vit.py
''' This code is modified from https://github.com/facebookresearch/convit. To adapt the vit/convit to image reconstruction, variable input sizes, and patch sizes for both spatial dimensions. ''' import torch import torch.nn as nn from functools import partial import torch.nn.functional as F from timm.models.helpers import load_pretrained from timm.models.layers import DropPath, to_2tuple, trunc_normal_ class Mlp(nn.Module): def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.): super().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features self.fc1 = nn.Linear(in_features, hidden_features) self.act = act_layer() self.fc2 = nn.Linear(hidden_features, out_features) self.drop = nn.Dropout(drop) self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.drop(x) x = self.fc2(x) x = self.drop(x) return x class GPSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., locality_strength=1., use_local_init=True, grid_size=None): super().__init__() self.num_heads = num_heads self.dim = dim head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.q = nn.Linear(dim, dim, bias=qkv_bias) self.k = nn.Linear(dim, dim, bias=qkv_bias) self.v = nn.Linear(dim, dim, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.pos_proj = nn.Linear(3, num_heads) self.proj_drop = nn.Dropout(proj_drop) self.locality_strength = locality_strength self.gating_param = nn.Parameter(1*torch.ones(self.num_heads)) self.apply(self._init_weights) if use_local_init: self.local_init(locality_strength=locality_strength) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention(self, x): B, N, C = x.shape k = self.k(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) q = self.q(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) pos_score = self.pos_proj(self.rel_indices).expand(B, -1, -1,-1).permute(0,3,1,2) patch_score = (q @ k.transpose(-2, -1)) * self.scale patch_score = patch_score.softmax(dim=-1) pos_score = pos_score.softmax(dim=-1) gating = self.gating_param.view(1,-1,1,1) attn = (1.-torch.sigmoid(gating)) * patch_score + torch.sigmoid(gating) * pos_score attn = attn / attn.sum(dim=-1).unsqueeze(-1) attn = self.attn_drop(attn) return attn def get_attention_map(self, x, return_map = False): attn_map = self.get_attention(x).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def local_init(self, locality_strength=1.): self.v.weight.data.copy_(torch.eye(self.dim)) locality_distance = 1 #max(1,1/locality_strength**.5) kernel_size = int(self.num_heads**.5) center = (kernel_size-1)/2 if kernel_size%2==0 else kernel_size//2 for h1 in range(kernel_size): for h2 in range(kernel_size): position = h1+kernel_size*h2 self.pos_proj.weight.data[position,2] = -1 self.pos_proj.weight.data[position,1] = 2*(h1-center)*locality_distance self.pos_proj.weight.data[position,0] = 2*(h2-center)*locality_distance self.pos_proj.weight.data *= locality_strength def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.v.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape if not hasattr(self, 'rel_indices') or self.rel_indices.size(1)!=N: self.get_rel_indices() attn = self.get_attention(x) v = self.v(x).reshape(B, N, self.num_heads, C // self.num_heads).permute(0, 2, 1, 3) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class MHSA(nn.Module): def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., grid_size=None): super().__init__() self.num_heads = num_heads head_dim = dim // num_heads self.scale = qk_scale or head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias) self.attn_drop = nn.Dropout(attn_drop) self.proj = nn.Linear(dim, dim) self.proj_drop = nn.Dropout(proj_drop) self.apply(self._init_weights) self.current_grid_size = grid_size def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) def get_attention_map(self, x, return_map = False): self.get_rel_indices() B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn_map = (q @ k.transpose(-2, -1)) * self.scale attn_map = attn_map.softmax(dim=-1).mean(0) # average over batch distances = self.rel_indices.squeeze()[:,:,-1]**.5 dist = torch.einsum('nm,hnm->h', (distances, attn_map)) dist /= distances.size(0) if return_map: return dist, attn_map else: return dist def get_rel_indices(self, ): H, W = self.current_grid_size N = H*W rel_indices = torch.zeros(1, N, N, 3) indx = torch.arange(W).view(1,-1) - torch.arange(W).view(-1, 1) indx = indx.repeat(H, H) indy = torch.arange(H).view(1,-1) - torch.arange(H).view(-1, 1) indy = indy.repeat_interleave(W, dim=0).repeat_interleave(W, dim=1) indd = indx**2 + indy**2 rel_indices[:,:,:,2] = indd.unsqueeze(0) rel_indices[:,:,:,1] = indy.unsqueeze(0) rel_indices[:,:,:,0] = indx.unsqueeze(0) device = self.qkv.weight.device self.rel_indices = rel_indices.to(device) def forward(self, x): B, N, C = x.shape qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) q, k, v = qkv[0], qkv[1], qkv[2] attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B, N, C) x = self.proj(x) x = self.proj_drop(x) return x class Block(nn.Module): def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0., drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm, use_gpsa=True, **kwargs): super().__init__() self.norm1 = norm_layer(dim) self.use_gpsa = use_gpsa if self.use_gpsa: self.attn = GPSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) else: self.attn = MHSA(dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop, **kwargs) self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() self.norm2 = norm_layer(dim) mlp_hidden_dim = int(dim * mlp_ratio) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop) def forward(self, x, grid_size): self.attn.current_grid_size = grid_size x = x + self.drop_path(self.attn(self.norm1(x))) x = x + self.drop_path(self.mlp(self.norm2(x))) return x class PatchEmbed(nn.Module): """ Image to Patch Embedding, from timm """ def __init__(self, patch_size, in_chans, embed_dim): super().__init__() self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) self.apply(self._init_weights) def forward(self, x): x = self.proj(x) return x def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) class VisionTransformer(nn.Module): """ Vision Transformer """ def __init__(self, avrg_img_size=320, patch_size=16, in_chans=1, embed_dim=64, depth=8, num_heads=9, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0., drop_path_rate=0., norm_layer=nn.LayerNorm, global_pool=None, gpsa_interval=[-1, -1], locality_strength=1., use_pos_embed=True): super().__init__() self.depth = depth embed_dim *= num_heads self.num_features = embed_dim # num_features for consistency with other models self.locality_strength = locality_strength self.use_pos_embed = use_pos_embed if isinstance(avrg_img_size, int): img_size = to_2tuple(avrg_img_size) if isinstance(patch_size, int): self.patch_size = to_2tuple(patch_size) else: self.patch_size = patch_size self.in_chans = in_chans self.patch_embed = PatchEmbed( patch_size=self.patch_size, in_chans=in_chans, embed_dim=embed_dim) self.pos_drop = nn.Dropout(p=drop_rate) if self.use_pos_embed: self.pos_embed = nn.Parameter( torch.zeros(1, embed_dim, img_size[0] // self.patch_size[0], img_size[1] // self.patch_size[1]) ) trunc_normal_(self.pos_embed, std=.02) dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule self.blocks = nn.ModuleList([ Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=True, locality_strength=locality_strength) if i>=gpsa_interval[0]-1 and i<gpsa_interval[1] else Block( dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale, drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer, use_gpsa=False,) for i in range(depth)]) self.norm = norm_layer(embed_dim) # head self.feature_info = [dict(num_chs=embed_dim, reduction=0, module='head')] self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def seq2img(self, x, img_size): """ Transforms sequence back into image space, input dims: [batch_size, num_patches, channels] output dims: [batch_size, channels, H, W] """ x = x.view(x.shape[0], x.shape[1], self.in_chans, self.patch_size[0], self.patch_size[1]) x = x.chunk(x.shape[1], dim=1) x = torch.cat(x, dim=4).permute(0,1,2,4,3) x = x.chunk(img_size[0]//self.patch_size[0], dim=3) x = torch.cat(x, dim=4).permute(0,1,2,4,3).squeeze(1) return x self.head.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): trunc_normal_(m.weight, std=.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) @torch.jit.ignore def no_weight_decay(self,): return {'pos_embed'} def get_head(self,): return self.head def reset_head(self,): self.head = nn.Linear(self.num_features, in_chans*self.patch_size[0]*self.patch_size[1]) def forward_features(self, x, k=None): x = self.patch_embed(x) _, _, H, W = x.shape if self.use_pos_embed: pos_embed = F.interpolate(self.pos_embed, size=[H, W], mode='bilinear', align_corners = False) x = x + pos_embed x = x.flatten(2).transpose(1, 2) x = self.pos_drop(x) for u, blk in enumerate(self.blocks): x = blk(x, (H, W)) if k is not None and u == k: self.attention_map = blk.attn.get_attention_map(x, return_map = True) x = self.norm(x) return x def forward(self, x, k=None): _, _, H, W = x.shape x = self.forward_features(x, k) x = self.head(x) x = self.seq2img(x, (H, W)) return x
15,082
39.007958
186
py
imaging_MLPs
imaging_MLPs-master/untrained/networks/recon_net.py
import torch.nn as nn import torch.nn.functional as F from math import ceil, floor class ReconNet(nn.Module): def __init__(self, net): super().__init__() self.net = net def pad(self, x): _, _, h, w = x.shape hp, wp = self.net.patch_size f1 = ( (wp - w % wp) % wp ) / 2 f2 = ( (hp - h % hp) % hp ) / 2 wpad = [floor(f1), ceil(f1)] hpad = [floor(f2), ceil(f2)] x = F.pad(x, wpad+hpad) return x, wpad, hpad def unpad(self, x, wpad, hpad): return x[..., hpad[0] : x.shape[-2]-hpad[1], wpad[0] : x.shape[-1]-wpad[1]] def forward(self, x, k=None): x, wpad, hpad = self.pad(x) x = self.net(x, k) x = self.unpad(x, wpad, hpad) return x
810
26.033333
90
py
imaging_MLPs
imaging_MLPs-master/untrained/networks/unet.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from torch import nn from torch.nn import functional as F class Unet(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int = 1, out_chans: int = 1, chans: int = 32, num_pool_layers: int = 4, drop_prob: float = 0.0, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append( nn.Sequential( ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1), ) ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
5,981
32.79661
88
py
DeepIR
DeepIR-main/demo.py
#!/usr/bin/env python import os import sys from pprint import pprint # Pytorch requires blocking launch for proper working if sys.platform == 'win32': os.environ['CUDA_LAUNCH_BLOCKING'] = '1' import numpy as np from scipy import io import torch import torch.nn torch.backends.cudnn.enabled = True torch.backends.cudnn.benchmark =True sys.path.append('modules') import utils import motion import dataset import thermal if __name__ == '__main__': imname = 'test1' # Name of the test file name camera = 'sim' # 'sim', 'boson' or 'lepton' scale_sr = 8 # 1 for denoising/NUC, 2, 3, .. for SR nimg = 20 # Number of input images method = 'dip' # 'cvx' for Hardie et al., 'dip' for DeepIR # Load config file -- config = dataset.load_config('configs/%s_%s.ini'%(method, camera)) config['batch_size'] = nimg config['num_workers'] = (0 if sys.platform=='win32' else 4) config['lambda_prior'] *= (scale_sr/nimg) # Load data if not config['real']: # This is simulated data im = utils.get_img(imname, 1) minval = 0 maxval = 1 else: # This is real data im, minval, maxval = utils.get_real_im(imname, camera) # Get data for SR -- this will also get an initial estimate for registration im, imstack, ecc_mats = motion.get_SR_data(im, scale_sr, nimg, config) ecc_mats[:, :, 2] *= scale_sr H, W = im.shape # Load LPIPs function config['gt'] = im # Now run denoising if method == 'cvx': im_dip, profile_dip = thermal.interp_convex(imstack.astype(np.float32), ecc_mats.astype(np.float32), (H, W), config) else: im_dip, profile_dip = thermal.interp_DIP(imstack.astype(np.float32), ecc_mats.astype(np.float32), (H, W), config) # Save data mdict = {'gt': im, 'rec': im_dip, 'gain': profile_dip['gain'], 'offset': profile_dip['offset'], 'snr': profile_dip['metrics']['snrval'], 'psnr': profile_dip['metrics']['psnrval'], 'ssim': profile_dip['metrics']['ssimval'], 'minval': minval, 'maxval': maxval} io.savemat('%s_%s_%s_%dx_%d.mat'%(imname, camera, method, scale_sr, nimg), mdict) pprint(profile_dip['metrics'])
2,636
31.555556
80
py
DeepIR
DeepIR-main/modules/losses.py
#!/usr/bin/env python import torch class TVNorm(): def __init__(self, mode='l1'): self.mode = mode def __call__(self, img): grad_x = img[..., 1:, 1:] - img[..., 1:, :-1] grad_y = img[..., 1:, 1:] - img[..., :-1, 1:] if self.mode == 'isotropic': #return torch.sqrt(grad_x.abs().pow(2) + grad_y.abs().pow(2)).mean() return torch.sqrt(grad_x**2 + grad_y**2).mean() elif self.mode == 'l1': return abs(grad_x).mean() + abs(grad_y).mean() else: return (grad_x.pow(2) + grad_y.pow(2)).mean() class HessianNorm(): def __init__(self): pass def __call__(self, img): # Compute Individual derivatives fxx = img[..., 1:-1, :-2] + img[..., 1:-1, 2:] - 2*img[..., 1:-1, 1:-1] fyy = img[..., :-2, 1:-1] + img[..., 2:, 1:-1] - 2*img[..., 1:-1, 1:-1] fxy = img[..., :-1, :-1] + img[..., 1:, 1:] - \ img[..., 1:, :-1] - img[..., :-1, 1:] return torch.sqrt(fxx.abs().pow(2) +\ 2*fxy[..., :-1, :-1].abs().pow(2) +\ fyy.abs().pow(2)).mean() class L1Norm(): def __init__(self): pass def __call__(self, x1, x2): return abs(x1 - x2).mean() class PoissonNorm(): def __init__(self): pass def __call__(self, x1, x2): return (x1 - torch.log(x1 + 1e-12)*x2).mean() class L2Norm(): def __init__(self): pass def __call__(self, x1, x2): return ((x1 - x2).pow(2)).mean()
1,586
30.117647
80
py
DeepIR
DeepIR-main/modules/utils.py
#!/usr/bin/env python ''' Miscellaneous utilities that are extremely helpful but cannot be clubbed into other modules. ''' import torch # Scientific computing import numpy as np import scipy.linalg as lin from scipy import io # Plotting import cv2 import matplotlib.pyplot as plt def nextpow2(x): ''' Return smallest number larger than x and a power of 2. ''' logx = np.ceil(np.log2(x)) return pow(2, logx) def normalize(x, fullnormalize=False): ''' Normalize input to lie between 0, 1. Inputs: x: Input signal fullnormalize: If True, normalize such that minimum is 0 and maximum is 1. Else, normalize such that maximum is 1 alone. Outputs: xnormalized: Normalized x. ''' if x.sum() == 0: return x xmax = x.max() if fullnormalize: xmin = x.min() else: xmin = 0 xnormalized = (x - xmin)/(xmax - xmin) return xnormalized def asnr(x, xhat, compute_psnr=False): ''' Compute affine SNR, which accounts for any scaling and shift between two signals Inputs: x: Ground truth signal(ndarray) xhat: Approximation of x Outputs: asnr_val: 20log10(||x||/||x - (a.xhat + b)||) where a, b are scalars that miminize MSE between x and xhat ''' mxy = (x*xhat).mean() mxx = (xhat*xhat).mean() mx = xhat.mean() my = x.mean() a = (mxy - mx*my)/(mxx - mx*mx) b = my - a*mx if compute_psnr: return psnr(x, a*xhat + b) else: return rsnr(x, a*xhat + b) def rsnr(x, xhat): ''' Compute reconstruction SNR for a given signal and its reconstruction. Inputs: x: Ground truth signal (ndarray) xhat: Approximation of x Outputs: rsnr_val: RSNR = 20log10(||x||/||x-xhat||) ''' xn = lin.norm(x.reshape(-1)) en = lin.norm((x-xhat).reshape(-1)) + 1e-12 rsnr_val = 20*np.log10(xn/en) return rsnr_val def psnr(x, xhat): ''' Compute Peak Signal to Noise Ratio in dB Inputs: x: Ground truth signal xhat: Reconstructed signal Outputs: snrval: PSNR in dB ''' err = x - xhat denom = np.mean(pow(err, 2)) + 1e-12 snrval = 10*np.log10(np.max(x)/denom) return snrval def embed(im, embedsize): ''' Embed a small image centrally into a larger window. Inputs: im: Image to embed embedsize: 2-tuple of window size Outputs: imembed: Embedded image ''' Hi, Wi = im.shape He, We = embedsize dH = (He - Hi)//2 dW = (We - Wi)//2 imembed = np.zeros((He, We), dtype=im.dtype) imembed[dH:Hi+dH, dW:Wi+dW] = im return imembed def measure(x, noise_snr=40, tau=100): ''' Realistic sensor measurement with readout and photon noise Inputs: noise_snr: Readout noise in electron count tau: Integration time. Poisson noise is created for x*tau. (Default is 100) Outputs: x_meas: x with added noise ''' x_meas = np.copy(x) noise = pow(10, -noise_snr/20)*np.random.randn(x_meas.size).reshape(x_meas.shape) # First add photon noise, provided it is not infinity if tau != float('Inf'): x_meas = x_meas*tau x_meas[x > 0] = np.random.poisson(x_meas[x > 0]) x_meas[x <= 0] = -np.random.poisson(-x_meas[x <= 0]) x_meas = (x_meas + noise)/tau else: x_meas = x_meas + noise return x_meas def grid_plot(imdata): ''' Plot 3D set of images into a 2D grid using subplots. Inputs: imdata: N x H x W image stack Outputs: None ''' N, H, W = imdata.shape nrows = int(np.sqrt(N)) ncols = int(np.ceil(N/nrows)) for idx in range(N): plt.subplot(nrows, ncols, idx+1) plt.imshow(imdata[idx, :, :], cmap='gray') plt.xticks([], []) plt.yticks([], []) def build_montage(images): ''' Build a montage out of images ''' nimg, H, W = images.shape nrows = int(np.ceil(np.sqrt(nimg))) ncols = int(np.ceil(nimg/nrows)) montage_im = np.zeros((H*nrows, W*ncols), dtype=np.float32) cnt = 0 for r in range(nrows): for c in range(ncols): h1 = r*H h2 = (r+1)*H w1 = c*W w2 = (c+1)*W if cnt == nimg: break montage_im[h1:h2, w1:w2] = images[cnt, ...] cnt += 1 return montage_im def ims2rgb(im1, im2): ''' Concatenate images into RGB Inputs: im1, im2: Two images to compare ''' H, W = im1.shape imrgb = np.zeros((H, W, 3)) imrgb[..., 0] = im1 imrgb[..., 2] = im2 return imrgb def textfunc(im, txt): return cv2.putText(im, txt, (30, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (1, 1, 1), 2, cv2.LINE_AA) def get_img(imname, scaling): # Read image im = cv2.resize(plt.imread('data/%s.png'%imname), None, fx=scaling, fy=scaling) if im.ndim == 2: im = im[:, :, np.newaxis] im = im[:, :, [0, 0, 0]] im = np.copy(im, order='C') H, W, _ = im.shape return np.copy(im[..., 1], order='C').astype(np.float32) def get_real_im(imname, camera): im = io.loadmat('data/%s/%s.mat'%(camera, imname))['imstack'] minval = im.min() maxval = im.max() if camera == 'rgb': im = normalize(im[:, ::2, ::2], True) else: im = normalize(im, True).astype(np.float32) return im, minval, maxval def boxify(im, topleft, boxsize, color=[1, 1, 1], width=2): ''' Generate a box around a region. ''' h, w = topleft dh, dw = boxsize im[h:h+dh+1, w:w+width, :] = color im[h:h+width, w:w+dh+width, :] = color im[h:h+dh+1, w+dw:w+dw+width, :] = color im[h+dh:h+dh+width, w:w+dh+width, :] = color return im def get_inp(tensize, const=10.0): ''' Wrapper to get a variable on graph ''' inp = torch.rand(tensize).cuda()/const inp = torch.autograd.Variable(inp, requires_grad=True).cuda() inp = torch.nn.Parameter(inp) return inp
6,530
21.996479
85
py
DeepIR
DeepIR-main/modules/dataset.py
#!/usr/bin/env python import os import sys import tqdm import pdb import math import configparser import numpy as np import torch from torch import nn import torch.nn.functional as F from torch.utils.data import DataLoader, Dataset from PIL import Image from torchvision.transforms import Resize, Compose, ToTensor, Normalize import skimage import matplotlib.pyplot as plt from matplotlib.lines import Line2D import cv2 def get_mgrid(sidelen, dim=2): '''Generates a flattened grid of (x,y,...) coordinates in a range of -1 to 1. sidelen: int dim: int''' tensors = tuple(dim * [torch.linspace(-1, 1, steps=sidelen)]) mgrid = torch.stack(torch.meshgrid(*tensors), dim=-1) mgrid = mgrid.reshape(-1, dim) return mgrid def xy_mgrid(H, W): ''' Generate a flattened meshgrid for heterogenous sizes Inputs: H, W: Input dimensions Outputs: mgrid: H*W x 2 meshgrid ''' Y, X = torch.meshgrid(torch.linspace(-1, 1, H), torch.linspace(-1, 1, W)) mgrid = torch.stack((X, Y), dim=-1).reshape(-1, 2) return mgrid class ImageDataset(Dataset): def __init__(self, img): super().__init__() H, W, nchan = img.shape img = torch.tensor(img)[..., None] self.pixels = img.view(-1, nchan) self.coords = xy_mgrid(H, W) def __len__(self): return 1 def __getitem__(self, idx): if idx > 0: raise IndexError return self.coords, self.pixels class Image3x3Dataset(Dataset): def __init__(self, img): super().__init__() H, W, nchan = img.shape img = torch.tensor(img)[..., None] self.pixels = img.view(-1, nchan) self.coords = xy_mgrid(H, W) # Stack coordinates in the 3x3 neighborhood coords_stack = [] for xshift in [0, 1]: for yshift in [0, 1]: shift_array = np.array([xshift/W, yshift/H]).reshape(1, 2) coords_stack.append(self.coords + shift_array) self.coords = np.hstack(coords_stack).astype(np.float32) def __len__(self): return 1 def __getitem__(self, idx): if idx > 0: raise IndexError return self.coords, self.pixels class ImageFlowDataset(Dataset): def __init__(self, img1, img2): super().__init__() H, W = img1.shape img1 = torch.tensor(img1)[..., None] img2 = torch.tensor(img2)[..., None] self.pixels1 = img1.view(-1, 1) self.pixels2 = img2.view(-1, 1) self.coords = xy_mgrid(H, W) def __len__(self): return 1 def __getitem__(self, idx): if idx > 0: raise IndexError return self.coords, self.pixels1, self.pixels2 class ImageRegDataset(Dataset): def __init__(self, imstack): super().__init__() self.imstack = imstack self.nimg, H, W = imstack.shape def __len__(self): return self.nimg def __getitem__(self, idx): img = torch.tensor(self.imstack[idx, ...])[None, ...] return img, idx class ImageStackDataset(Dataset): def __init__(self, imstack): super().__init__() self.imstack = imstack self.nimg, H, W = imstack.shape self.coords = xy_mgrid(H, W) def __len__(self): return self.nimg def __getitem__(self, idx): img = torch.tensor(self.imstack[idx, ...]) pixels = img[None, ...].permute(1, 2, 0).view(-1, 1) return self.coords, pixels class ImageSRDataset(Dataset): def __init__(self, imstack, Xstack=None, Ystack=None, masks=None, jitter=False, xjitter=None, yjitter=None, get_indices=False): super().__init__() self.imstack = imstack self.Xstack = Xstack self.Ystack = Ystack self.masks = masks self.jitter = jitter self.get_indices = get_indices self.nimg, self.H, self.W = imstack.shape if xjitter is None: self.xjitter = 1/self.W self.yjitter = 1/self.H else: self.xjitter = xjitter self.yjitter = yjitter def __len__(self): return self.nimg def __getitem__(self, idx): img = torch.tensor(self.imstack[idx, ...]) # If Jitter is enabled, return stratified sampled coordinates pixels = img[None, ...].permute(1, 2, 0).view(-1, 1) if self.masks is not None: mask = torch.tensor(self.masks[idx, ...]) mask = mask[None, ...].permute(1, 2, 0).view(-1, 1) else: mask = torch.zeros(1) if self.Xstack is not None: coords = torch.stack((torch.tensor(self.Xstack[idx, ...]), torch.tensor(self.Ystack[idx, ...])), dim=-1).reshape(-1, 2) else: coords = torch.zeros(1) if self.get_indices: return coords, pixels, mask, idx else: return coords, pixels, mask class ImageChunkDataset(Dataset): def __init__(self, imstack, patchsize): super().__init__() self.imstack = imstack self.nimg, self.H, self.W = imstack.shape self.patchsize = patchsize self.patch_coords = xy_mgrid(patchsize[0], patchsize[1]) self.nH = int(np.ceil(self.H/patchsize[0])) self.nW = int(np.ceil(self.W/patchsize[1])) def __len__(self): return (self.nH * self.nW) def __getitem__(self, idx): w_idx = int(idx%self.nH) h_idx = int((idx - w_idx)//self.nH) h1 = h_idx*self.patchsize[0] h2 = h_idx*self.patchsize[0] + self.patchsize[0] w1 = w_idx*self.patchsize[1] w2 = w_idx*self.patchsize[1] + self.patchsize[1] if h2 > self.H: h1 = self.H - self.patchsize[0] h2 = self.H if w2 > self.W: w1 = self.W - self.patchsize[1] w2 = self.W img = torch.tensor(self.imstack[:, h1:h2, w1:w2]) pixels = img.reshape(-1, 1) coords = torch.clone(self.patch_coords) coords[:, 0] = coords[:, 0] + w1 coords[:, 1] = coords[:, 1] + h1 coords = torch.repeat_interleave(coords, self.nimg, 0) return coords, pixels def load_config(configpath): ''' Load configuration file ''' parser = configparser.ConfigParser() parser.read(configpath) params_dict = dict() for section in parser.keys(): for key in parser[section].keys(): token = parser[section][key] if token == 'False': params_dict[key] = False elif token == 'True': params_dict[key] = True elif '.' in token: params_dict[key] = float(token) else: try: params_dict[key] = int(token) except: params_dict[key] = token return params_dict
7,382
27.287356
81
py
DeepIR
DeepIR-main/modules/thermal.py
#!/usr/bin/env python ''' Routines for dealing with thermal images ''' import tqdm import copy import cv2 import numpy as np from skimage.metrics import structural_similarity as ssim_func import torch import kornia import torch.nn.functional as F import utils import losses import motion import deep_prior def get_metrics(gt, estim, pad=True): ''' Compute SNR, PSNR, SSIM, and LPIP between two images. Inputs: gt: Ground truth image estim: Estimated image lpip_func: CUDA function for computing lpip value pad: if True, remove boundaries when computing metrics Outputs: metrics: dictionary with following fields: snrval: SNR of reconstruction psnrval: Peak SNR ssimval: SSIM lpipval: VGG perceptual metrics ''' if min(gt.shape) < 50: pad = False if pad: gt = gt[20:-20, 20:-20] estim = estim[20:-20, 20:-20] snrval = utils.asnr(gt, estim) psnrval = utils.asnr(gt, estim, compute_psnr=True) ssimval = ssim_func(gt, estim) metrics = {'snrval': snrval, 'psnrval': psnrval, 'ssimval': ssimval} return metrics def create_fpn(imsize, vmin=0.9, vmax=1, method='col', rank=1): ''' Generate fixed pattern noise for microbolometer-type sensors Inputs: imsize: (H, W) tuple vmin, vmax: Minimum and maximum value of gain method: 'col' -- generate column only noise 'both' -- generate rank-k noise 'corr_col' -- correlated columns 'corr_both' -- correlated rows and columns rank: if method is 'both' generate noise with this rank. Outputs: fpn: (H, W)-sized fixed pattern noise ''' H, W = imsize if method == 'col': fpn = np.ones((H, 1)).dot(vmin + (vmax-vmin)*np.random.rand(1, W)) elif method == 'both': fpn = 0 for idx in range(rank): col = vmin + (vmax - vmin)*np.random.rand(H, 1) row = vmin + (vmax - vmin)*np.random.rand(1, W) fpn += col.dot(row) fpn /= rank elif method == 'corr_col': row = vmin + (vmax-vmin)*np.random.rand(W) row = np.convolve(row, np.ones(5)/5, mode='same') fpn = np.ones((H, 1)).dot(row.reshape(1, W)) elif method == 'corr_both': row = vmin + (vmax-vmin)*np.random.rand(W) row = np.convolve(row, np.ones(5)/5, mode='same') col = vmin + (vmax-vmin)*np.random.rand(H) col = np.convolve(col, np.ones(5)/5, mode='same') fpn = col.reshape(H, 1).dot(row.reshape(1, W)) return fpn def reg_avg_denoise(imstack, ecc_mats=None): ''' Denoise a thermal stack by registering and averaging. Inputs: Outputs: im_denoised: Denoised image ''' nimg, H, W = imstack.shape # if ecc_mats is none, register the stack if ecc_mats is None: ecc_mats = motion.register_stack(imstack, (H, W))[:, :2, :] # Now warp image back to reference and average ecc_inv = motion.invert_regstack(ecc_mats) imten = torch.tensor(imstack.astype(np.float32))[:, None, ...] ecc_ten = torch.tensor(ecc_inv.astype(np.float32)) imwarped = kornia.geometry.warp_affine(imten, ecc_ten, (H, W), flags='bilinear') im_denoised = imwarped.mean(0)[0, ...].numpy() weights = (imwarped > 0).type(torch.float32).mean(0)[0, ...].numpy() weights[weights == 0] = 1 im_denoised /= weights return im_denoised def interp_DIP(imstack, reg_stack, hr_size, params_dict): ''' Super resolve from a stack of images using deep image prior Inputs: imstack: (nimg, Hl, Wl) stack of low resolution images reg_stack: (nimg, 2, 3) stack of affine matrices hr_size: High resolution image size params_dict: Dictionary containing parameters for optimization kernel_type: Type of downsampling input_type: Type of input input_depth: Depth of input data (number of channels) skip_n33d: Parameter for the neural network skip_n33u: Parameter for the neural network skip_n11: Parameter for the neural network num_scales: Parameter for the neural network upsample_mode: Parameter for the neural network niters: Number of DIP iterations batch_size: Batch size of data num_workers: Workers for data loading learning_rate: Learning rate for optimization prior_type: tv, or hessian lambda_prior: Prior weight optimize_reg: If True, optimize registration parameters visualize: If True, visualize reconstructions at each iteration gt: If visualize is true, gt is the ground truth image reg_final: If True, register the final result to gt lpip_func: If gt is true, evaluate perceptual similarity with this function Returns: im_hr: High resolution image profile: Dictionary containing the following: loss_array: Array with loss at each iteration trained_model: State dictionary for best model metrics: if gt is provided, this is a dictionary with: snrval: SNR of reconstruction psnrval: Peak SNR ssimval: SSIM lpipval: VGG perceptual metrics ''' nimg, Hl, Wl = imstack.shape H, W = hr_size scale_sr = 0.5*(H/Hl + W/Wl) # Internal constant img_every = 2 if params_dict['mul_gain']: lambda_offset = 10 else: lambda_offset = 0 # Create loss functions criterion_fidelity = losses.L1Norm() criterion_offset = losses.TVNorm(mode='l2') if params_dict['prior_type'] == 'tv': criterion_prior = losses.TVNorm() elif params_dict['prior_type'] == 'hessian': criterion_prior = losses.HessianNorm() else: raise ValueError('Prior not implemented') # Create input model_input = deep_prior.get_noise(params_dict['input_depth'], params_dict['input_type'], (H, W)).cuda().detach() # Create the network if params_dict['predmode'] == 'combined': nchan = 3 else: nchan = 1 model = deep_prior.get_net(params_dict['input_depth'], 'skip', 'reflection', n_channels=nchan, skip_n33d=params_dict['skip_n33d'], skip_n33u=params_dict['skip_n33u'], skip_n11=params_dict['skip_n11'], num_scales=params_dict['num_scales'], upsample_mode=params_dict['upsample_mode'] ).cuda() # Set it to training model.train() if params_dict['integrator'] == 'learnable': kernel_size = (int(scale_sr), int(scale_sr)) integrator = torch.nn.Conv2d(1, 1, kernel_size=kernel_size, stride=int(scale_sr), bias=False).cuda() with torch.no_grad(): integrator.weight.fill_(1.0/(scale_sr*scale_sr)) # Create parameters from affine matrices affine_mat = torch.tensor(reg_stack).cuda() affine_var = torch.autograd.Variable(affine_mat, requires_grad=True).cuda() affine_param = torch.nn.Parameter(affine_var) # Create gain parameter vmin = params_dict['fpn_vmin'] params = list(model.parameters()) if params_dict['predmode'] != 'combined': gain = torch.ones(1, 1, Hl, Wl).cuda() gain_var = torch.autograd.Variable(gain, requires_grad=True).cuda() gain_param = torch.nn.Parameter(gain_var) offset = torch.ones(1, 1, Hl, Wl).cuda()*1e-1 offset_var = torch.autograd.Variable(offset, requires_grad=True).cuda() offset_param = torch.nn.Parameter(offset_var) params += [gain_param] + [offset_param] if params_dict['integrator'] == 'learnable': params += integrator.parameters() # Create an ADAM optimizer optimizer = torch.optim.Adam(lr=params_dict['learning_rate'], params=params) # Affine transform requires a separate optimizer reg_optimizer = torch.optim.Adam(lr=params_dict['affine_learning_rate'], params=[affine_param]) loss_array = np.zeros(params_dict['niters']) best_loss = float('inf') best_state_dict = None # We will just use all data gt = torch.tensor(imstack).cuda()[:, None, ...] for epoch in tqdm.tqdm(range(params_dict['niters'])): train_loss = 0 img_and_gain = model(model_input) img_hr = img_and_gain[:, [0], ...] if params_dict['predmode'] == 'combined': gain_param = img_and_gain[:, [1], ...] offset_param = img_and_gain[:, [2], ...] if scale_sr > 1: gain_param = F.interpolate(gain_param, (Hl, Wl)) offset_param = F.interpolate(offset_param, (Hl, Wl)) # Generate low resolution images img_hr_cat = torch.repeat_interleave(img_hr, nimg, 0) if params_dict['integrator'] == 'area': img_hr_affine = kornia.geometry.warp_affine(img_hr_cat, affine_param, (H, W), align_corners=True) img_lr = F.interpolate(img_hr_affine, (Hl, Wl), mode='area') elif params_dict['integrator'] == 'learnable': img_hr_affine = kornia.geometry.warp_affine(img_hr_cat, affine_param, (H, W), align_corners=True) img_lr = integrator(img_hr_affine) else: img_lr = kornia.geometry.warp_affine(img_hr_cat, affine_param/scale_sr, (Hl, Wl), align_corners=False) # Multiply with the gain term mask = img_lr > 0 if params_dict['add_offset']: img_lr = img_lr + offset_param if params_dict['mul_gain']: img_lr = gain_param * img_lr mse_loss = criterion_fidelity(img_lr*mask, gt*mask) prior_loss = params_dict['lambda_prior']*criterion_prior(img_hr) loss = mse_loss + prior_loss if params_dict['add_offset']: offset_loss = lambda_offset*criterion_offset(offset_param) loss = loss + offset_loss optimizer.zero_grad() if params_dict['optimize_reg']: reg_optimizer.zero_grad() loss.backward() optimizer.step() if params_dict['optimize_reg']: reg_optimizer.step() train_loss = loss.item() # Find if we have the best mode if train_loss < best_loss: best_loss = train_loss best_state_dict = copy.deepcopy(model.state_dict()) loss_array[epoch] = train_loss if params_dict['visualize']: if epoch%img_every == 0: with torch.no_grad(): img_hr_cpu = img_hr.cpu().detach().numpy().reshape(H, W) v_idx = np.random.randint(nimg) img_lr_cpu = img_lr[v_idx, ...].cpu().detach().reshape(Hl, Wl) snrval = utils.asnr(params_dict['gt'], img_hr_cpu, compute_psnr=True) ssimval = ssim_func(params_dict['gt'], img_hr_cpu) txt = 'PSNR: %.1f | SSIM: %.2f'%(snrval, ssimval) gain = gain_param.cpu().detach().numpy().reshape(Hl, Wl) offset = offset_param.cpu().detach().numpy().reshape(Hl, Wl) img_hr_ann = utils.textfunc(img_hr_cpu/img_hr_cpu.max(), txt) imtop = np.hstack((imstack[v_idx, ...], img_lr_cpu.numpy())) imbot = np.hstack((gain/gain.max(), offset/offset.max())) imcat = np.vstack((imtop, imbot)) imcat_full = np.hstack((params_dict['gt'], img_hr_ann)) cv2.imshow('Recon LR', np.clip(imcat, 0, 1)) cv2.imshow('Recon HR', np.clip(imcat_full, 0, 1)) cv2.waitKey(1) # We are done, obtain the best model model.eval() with torch.no_grad(): model.load_state_dict(best_state_dict) img_and_gain = model(model_input) img_hr = img_and_gain[[0], [0], ...].reshape(1, 1, H, W) img_hr = kornia.geometry.warp_affine(img_hr, affine_param[[0], ...], (H, W)) img_hr = img_hr.cpu().detach().numpy().reshape(H, W) if params_dict['predmode'] == 'combined': gain_param = img_and_gain[0, 1, ...] offset_param = img_and_gain[0, 2, ...] # In case there's a shift in reconstruction if params_dict['reg_final'] and 'gt' in params_dict: try: img_hr = motion.ecc_flow(params_dict['gt'], img_hr)[1] except: pass # If ground truth is provided, return metrics if 'gt' in params_dict: metrics = get_metrics(params_dict['gt'], img_hr) gain = gain_param.detach().cpu().numpy().reshape(Hl, Wl) offset = offset_param.detach().cpu().numpy().reshape(Hl, Wl) profile = {'loss_array': loss_array, 'trained_model': best_state_dict, 'metrics': metrics, 'ecc_mats': affine_param.detach().cpu().numpy(), 'gain': gain, 'offset': offset} return img_hr, profile def interp_convex(imstack, reg_stack, hr_size, params_dict): ''' Super resolve from a stack of images using convex optimization Inputs: imstack: (nimg, Hl, Wl) stack of low resolution images reg_stack: (nimg, 2, 3) stack of affine matrices hr_size: High resolution image size params_dict: Dictionary containing parameters for optimization niters: Number of SIREN iterations batch_size: Batch size of data num_workers: Workers for data loading learning_rate: Learning rate for optimization prior_type: tv, or hessian lambda_prior: Prior weight optimize_reg: If True, optimize registration parameters visualize: If True, visualize reconstructions at each iteration gt: If visualize is true, gt is the ground truth image reg_final: If True, register the final result to gt lpip_func: If gt is true, evaluate perceptual similarity with this function Returns: im_hr: High resolution image profile: Dictionary containing the following: loss_array: Array with loss at each iteration trained_model: State dictionary for best model metrics: if gt is provided, this is a dictionary with: snrval: SNR of reconstruction psnrval: Peak SNR ssimval: SSIM lpipval: VGG perceptual metrics ''' nimg, Hl, Wl = imstack.shape H, W = hr_size scale_sr = 0.5*(H/Hl + W/Wl) # Internal constant img_every = 10 lambda_offset = 10 # Create loss functions criterion_fidelity = losses.L2Norm() if params_dict['prior_type'] == 'tv': criterion_prior = losses.TVNorm() elif params_dict['prior_type'] == 'hessian': criterion_prior = losses.HessianNorm() elif params_dict['prior_type'] == 'l2': criterion_prior = losses.L2Norm() else: raise ValueError('Prior not implemented') # Initialize solution with linear interpolation #im_init = torch.tensor(interp_SR(imstack, reg_stack, hr_size)) im_init = torch.rand(H, W) gain_init = torch.rand(Hl, Wl) offset_init = torch.ones(Hl, Wl)*1e-2 # Create the variable img_hr_param = torch.autograd.Variable(im_init[None, None, ...], requires_grad=True).cuda() img_hr_param = torch.nn.Parameter(img_hr_param) # Create gain parameter gain_param = torch.autograd.Variable(gain_init[None, None, ...], requires_grad=True).cuda() gain_param = torch.nn.Parameter(gain_param) # Create offset parameter offset_param = torch.autograd.Variable(offset_init[None, None, ...], requires_grad=True).cuda() offset_param = torch.nn.Parameter(offset_param) # Create parameters from affine matrices affine_mat = torch.tensor(reg_stack).cuda() affine_var = torch.autograd.Variable(affine_mat, requires_grad=True).cuda() affine_param = torch.nn.Parameter(affine_var) params = [img_hr_param] + [gain_param] + [offset_param] if params_dict['optimize_reg']: params += [affine_param] #params += [angles_param] + [translations_param] # Create an ADAM optimizer optimizer = torch.optim.Adam(lr=params_dict['learning_rate'], params=params) loss_array = np.zeros(params_dict['niters']) gt = torch.tensor(imstack).cuda()[:, None, ...] for epoch in tqdm.tqdm(range(params_dict['niters'])): # Generate low resolution images img_hr_cat = torch.repeat_interleave(img_hr_param, gt.shape[0], 0) if params_dict['integrator'] == 'area': img_hr_affine = kornia.geometry.warp_affine(img_hr_cat, affine_param, (H, W), align_corners=False) img_lr = F.interpolate(img_hr_affine, (Hl, Wl), mode='area') else: img_lr = kornia.geometry.warp_affine(img_hr_cat, affine_param/scale_sr, (Hl, Wl), align_corners=False) gain_cat = torch.repeat_interleave(gain_param, gt.shape[0], 0) #offset_cat = torch.repeat_interleave(offset_param, gt.shape[0], 0) masks = img_lr > 0 #if epoch > params_dict['niters']: # img_lr = img_lr*(gain_cat + offset_cat) #else: img_lr = img_lr + gain_cat mse_loss = criterion_fidelity(img_lr, gt) prior_loss = criterion_prior(img_hr_param) #offset_loss = criterion_prior(offset_param) loss = mse_loss + params_dict['lambda_prior']*prior_loss # lambda_offset*offset_loss optimizer.zero_grad() loss.backward() optimizer.step() train_loss = loss.item() loss_array[epoch] = train_loss if params_dict['visualize']: if epoch%img_every == 0: img_hr_cpu = img_hr_param.cpu().detach().numpy().reshape(H, W) v_idx = np.random.randint(nimg) img_lr_cpu = img_lr[v_idx, ...] img_lr_cpu = img_lr_cpu.cpu().detach().numpy().reshape(Hl, Wl) gain = gain_param.cpu().detach().numpy().reshape(Hl, Wl) offset = offset_param.cpu().detach().numpy().reshape(Hl, Wl) snrval = utils.psnr(params_dict['gt'], img_hr_cpu) ssimval = ssim_func(params_dict['gt'], img_hr_cpu) txt = 'PSNR: %.1f | SSIM: %.2f'%(snrval, ssimval) img_hr_ann = utils.textfunc(img_hr_cpu, txt) imcat = np.hstack((imstack[v_idx, ...], img_lr_cpu, gain, offset/offset.max())) imcat_full = np.hstack((params_dict['gt'], img_hr_ann)) cv2.imshow('Recon LR', np.clip(imcat, 0, 1)) cv2.imshow('Recon HR', np.clip(imcat_full, 0, 1)) cv2.waitKey(1) # We are done, obtain the best model with torch.no_grad(): img_hr = kornia.geometry.warp_affine(img_hr_param, affine_param[[0], ...], (H, W)) img_hr = img_hr_param.cpu().detach().numpy().reshape(H, W) gain = gain_param.cpu().detach().numpy().reshape(Hl, Wl) offset = offset_param.cpu().detach().numpy().reshape(Hl, Wl) # In case there's a shift in reconstruction if params_dict['reg_final'] and 'gt' in params_dict: try: img_hr = motion.ecc_flow(params_dict['gt'], img_hr)[1] except: pass # If ground truth is provided, return metrics if 'gt' in params_dict: metrics = get_metrics(params_dict['gt'], img_hr) profile = {'loss_array': loss_array, 'metrics': metrics, 'gain': gain, 'offset': offset} return img_hr, profile
21,820
37.485009
84
py
DeepIR
DeepIR-main/modules/motion.py
#!/usr/bin/env python ''' Subroutines for estimating motion between images ''' import os import sys import tqdm import pdb import math import numpy as np from scipy import linalg from scipy import interpolate import torch from torch import nn import torch.nn.functional as F from torch.utils.data import DataLoader, Dataset from PIL import Image from torchvision.transforms import Resize, Compose, ToTensor, Normalize import kornia from pystackreg import StackReg import skimage import matplotlib.pyplot as plt from matplotlib.lines import Line2D import cv2 import utils import thermal def xy_mgrid(H, W): ''' Generate a flattened meshgrid for heterogenous sizes Inputs: H, W: Input dimensions Outputs: mgrid: H*W x 2 meshgrid ''' Y, X = torch.meshgrid(torch.linspace(-1, 1, H), torch.linspace(-1, 1, W)) mgrid = torch.stack((X, Y), dim=-1).reshape(-1, 2) return mgrid def getEuclidianMatrix(theta, shift): ''' Compute 2x3 euclidean matrix ''' mat = np.array([[np.cos(theta), -np.sin(theta), shift[0]], [np.sin(theta), np.cos(theta), shift[1]]]) return mat def fb_flow(frame1, frame2): H, W = frame1.shape Y, X = np.mgrid[:H, :W] hsv = np.zeros((H, W, 3), dtype=np.uint8) hsv[...,1] = 255 flow = cv2.calcOpticalFlowFarneback(frame1, frame2, None, 0.5, 3, 15, 3, 5, 1.2, 0) mag, ang = cv2.cartToPolar(flow[...,0], flow[...,1]) hsv[...,0] = ang*180/np.pi/2 hsv[...,2] = cv2.normalize(mag,None,0,255,cv2.NORM_MINMAX) rgb = cv2.cvtColor(hsv,cv2.COLOR_HSV2BGR) frame2_warped = cv2.remap(frame2.astype(np.float32), (X + flow[..., 0]).astype(np.float32), (Y + flow[..., 1]).astype(np.float32), cv2.INTER_LINEAR) rgb_comp = np.zeros((H, W, 3)) rgb_comp[..., 0] = frame1 rgb_comp[..., 2] = frame2_warped return frame2_warped, flow, rgb, rgb_comp def ecc_flow(im1, im2, warp_mode=cv2.MOTION_HOMOGRAPHY, niters=1000, eps=1e-8): ''' Register images using Opencv intensity based image alignment approach. Inputs: im1, im2: Images to register. im2 will be registered to im1. method: One of cv2.MOTION_*** . Default is MOTION_HOMOGRAPRHY niters: Number of ECC iterations eps: Stopping tolerance Outputs: warp_matrix: Warping matrix im2_aligned: Second image warped to first image's coordinates flow: Flow coordinates to go from im2 to im1 https://learnopencv.com/image-alignment-ecc-in-opencv-c-python/ ''' # Find size of image1 sz = im1.shape # Define 2x3 or 3x3 matrices and initialize the matrix to identity if warp_mode == cv2.MOTION_HOMOGRAPHY : warp_matrix = np.eye(3, 3, dtype=np.float32) else : warp_matrix = np.eye(2, 3, dtype=np.float32) # Define termination criteria criteria = (cv2.TERM_CRITERIA_EPS | cv2.TERM_CRITERIA_COUNT, niters, eps) # Run the ECC algorithm. The results are stored in warp_matrix. (cc, warp_matrix) = cv2.findTransformECC(im1, im2, warp_matrix, warp_mode, criteria=criteria, inputMask=None, gaussFiltSize=5) if warp_mode == cv2.MOTION_HOMOGRAPHY : # Use warpPerspective for Homography flags = cv2.INTER_LINEAR + cv2.WARP_INVERSE_MAP im2_aligned = cv2.warpPerspective(im2, warp_matrix, (sz[1],sz[0]), flags=flags) else : # Use warpAffine for Translation, Euclidean and Affine flags = cv2.INTER_LINEAR + cv2.WARP_INVERSE_MAP im2_aligned = cv2.warpAffine(im2, warp_matrix, (sz[1],sz[0]), flags=flags) # Create flow coordinates Y, X = np.mgrid[:sz[0], :sz[1]] coords = np.ones((3, sz[0]*sz[1])) coords[0, :] = X.reshape(1, -1) coords[1, :] = Y.reshape(1, -1) coords_new = warp_matrix.dot(coords) if warp_mode == cv2.MOTION_HOMOGRAPHY: coords_new = coords_new[:2, :]/coords_new[2, :] flow = np.zeros((sz[0], sz[1], 2), dtype=np.float32) flow[..., 0] = (coords_new[0, :] - coords[0, :]).reshape(sz)*(2/sz[1]) flow[..., 1] = (coords_new[1, :] - coords[1, :]).reshape(sz)*(2/sz[0]) return warp_matrix, im2_aligned, flow def get_SR_data(im, scale, nimg=10, config=None): ''' Wrapper function to get real or simulation data Inputs: im: Image or image stack scale: Scale for resolution nimg: Number of images config: Dictionary containing the following files: simulation: If True, im will be converted to an image stack, else the input will be treated as imstack get_gt: If True, and simulation is also True, return groundtruth registration matrices shift_max, theta_max: See get_imstack downsample: If True, the imstack will be a downsampled version of the data. Only applicable if simulation is false tau, noise_snr: For simulated data, these represente max. photon count, and readout noise add_fpn: If True, add fixed pattern noise to data fpn_vmin: Minimum value of fpn fpn_method: 'col' or 'both' fpn_rank: If 'both', how many patterns to add Outputs: im: Ground truth high resolution image. Only useful if simulation is true, or simulation is false, and downsample is true. Else it is just a nearest neighbor upsampling imstack: (nimg, Hl, Wl) stack of low resolution images ecc_mats: (nimg, 2, 3) affine matrices ''' # Extract configuration values if config is None: simulation = True get_gt = False shift_max = 10 theta_max = np.pi/12 downsample = False add_noise = False tau = None noise_snr = None add_fpn = False fpn_vmin = 0.9 fpn_method = 'col' fpn_rank = 1 else: simulation = not config['real'] get_gt = config['get_gt'] shift_max = config['shift_max']*scale theta_max = config['theta_max']*np.pi/180 downsample = config['downsample'] add_noise = config['add_noise'] tau = config['tau'] noise_snr = config['noise_snr'] add_fpn = config['add_fpn'] fpn_vmin = config['fpn_vmin'] fpn_method = config['fpn_method'] fpn_rank = config['fpn_rank'] if simulation is True: H, W = im.shape imstack, _, _, mats = get_imstack(im, scale, shift_max, theta_max, nimg) imstack /= im.max() im /= im.max() if add_noise: imstack = utils.measure(imstack, noise_snr, tau) if add_fpn: fpn = thermal.create_fpn(imstack.shape[1:], vmin=fpn_vmin, method=fpn_method, rank=fpn_rank) imstack = imstack*fpn[np.newaxis, ...] _, Hl, Wl = imstack.shape if get_gt: ecc_mats = invert_regstack(mats) else: ecc_mats = register_stack(imstack, (Hl, Wl))[:, :2, :] else: _, H, W = im.shape imstack = np.copy(im[:nimg, ...], order='C') interp = cv2.INTER_AREA if downsample: imstack_lr = np.zeros((nimg, H//scale, W//scale)) Hl, Wl = H//scale, W//scale for idx in range(nimg): imstack_lr[idx, ...] = cv2.resize(imstack[idx, ...], (W//scale, H//scale), interpolation=interp) im = imstack[0, ...] imstack = imstack_lr.astype(np.float32) else: im = cv2.resize(imstack[0, ...], (W*scale, H*scale)) Hl, Wl = H, W H, W = Hl*scale, Wl*scale ecc_mats = register_stack(imstack, (Hl, Wl))[:, :2, :] imstack /= im.max() im /= im.max() return im, imstack, ecc_mats def get_random_affine(nimg, shift_max=10, theta_max=np.pi/12, perspective=False): ''' Get affine matrices with random shifts and thetas ''' shifts = np.random.randint(-shift_max, shift_max, size=[nimg, 2]) thetas = (2*np.random.rand(nimg)-1)*theta_max shifts[0, ...] = 0 thetas[0] = 0 affine_mats = np.zeros((nimg, 3, 3)) affine_mats[:, 2, 2] = 1.0 for idx in range(nimg): affine_mats[idx, :2, :] = getEuclidianMatrix(thetas[idx], shifts[idx, ...]) # Set first matrix to identity affine_mats[0, :, :] = 0 affine_mats[0, 0, 0] = 1 affine_mats[0, 1, 1] = 1 affine_mats[0, 2, 2] = 1 if perspective is False: affine_mats = affine_mats[:, :2, :] return affine_mats def get_imstack(im, scale, shift_max=10, theta_max=np.pi/12, nshifts=5): ''' Obtain synthetically generated, low resolution images of im, with random shifts. Inputs: im: Input high resolution image scale: Downsampling factor (> 1) theta_max: Maximum angle of rotation nshifts: Number of shifted images to obtain perturb_coords: If True, perturb the coordinates to study the effect of erroneous registration Outputs: imstack: Stack of images coordstack: Stack of (x ,y) coordinates for each image ''' H, W = im.shape #shifts = np.random.randint(-shift_max, shift_max, size=[nshifts, 2]) shifts = -shift_max + 2*shift_max*np.random.rand(nshifts, 2) thetas = (2*np.random.rand(nshifts)-1)*theta_max Y, X = np.mgrid[:H, :W] tmp = cv2.resize(im, None, fx=1/scale, fy=1/scale) Hl, Wl = tmp.shape imstack = np.zeros((nshifts, Hl, Wl), dtype=np.float32) Xstack = np.zeros_like(imstack) Ystack = np.zeros_like(imstack) mats = np.zeros((nshifts, 2, 3)) # Ensure first shift and theta are zero shifts[0, :] = 0 thetas[0] = 0 coords = np.hstack((X.reshape(-1, 1), Y.reshape(-1, 1), np.ones((H*W, 1)))) for idx in range(nshifts): shift = shifts[idx, :] theta = thetas[idx] mat = getEuclidianMatrix(theta, shift) mats[idx, ...] = mat coords_new = mat.dot(coords.T).T Xnew = coords_new[:, 0].reshape(H, W) Ynew = coords_new[:, 1].reshape(H, W) Xnew = cv2.resize(Xnew, (Wl, Hl), interpolation=cv2.INTER_LINEAR) Ynew = cv2.resize(Ynew, (Wl, Hl), interpolation=cv2.INTER_LINEAR) imstack[idx, ...] = cv2.remap(im, Xnew.astype(np.float32), Ynew.astype(np.float32), cv2.INTER_LINEAR) Xstack[idx, ...] = 2*Xnew/W - 1 Ystack[idx, ...] = 2*Ynew/H - 1 return imstack, Xstack, Ystack, mats def get_downsampled_shifted_images(im, scale, shift_max=10, theta_max=np.pi/12, nshifts=5, perturb_coords=False): ''' Obtain synthetically generated, low resolution images of im, with random shifts. Inputs: im: Input high resolution image scale: Downsampling factor (> 1) theta_max: Maximum angle of rotation nshifts: Number of shifted images to obtain perturb_coords: If True, perturb the coordinates to study the effect of erroneous registration Outputs: imstack: Stack of images coordstack: Stack of (x ,y) coordinates for each image ''' H, W = im.shape shifts = np.random.randint(-shift_max, shift_max, size=[nshifts, 2]) thetas = (2*np.random.rand(nshifts)-1)*theta_max Y, X = np.mgrid[:H, :W] tmp = cv2.resize(im, None, fx=1/scale, fy=1/scale) Hl, Wl = tmp.shape imstack = np.zeros((nshifts, Hl, Wl), dtype=np.float32) Xstack = np.zeros_like(imstack) Ystack = np.zeros_like(imstack) # Ensure first shift and theta are zero shifts[0, :] = 0 thetas[0] = 0 for idx in range(nshifts): shift = shifts[idx, :] theta = thetas[idx] # Shift Xshifted = X - shift[1] Yshifted = Y - shift[0] # Rotate Xrot = (Xshifted-W/2)*np.cos(theta) - (Yshifted-H/2)*np.sin(theta) + W/2 Yrot = (Xshifted-W/2)*np.sin(theta) + (Yshifted-H/2)*np.cos(theta) + H/2 Xnew = cv2.resize(Xrot, (Wl, Hl), interpolation=cv2.INTER_AREA) Ynew = cv2.resize(Yrot, (Wl, Hl), interpolation=cv2.INTER_AREA) imstack[idx, ...] = cv2.remap(im, Xnew.astype(np.float32), Ynew.astype(np.float32), cv2.INTER_AREA) if perturb_coords: # Now ... let's generate noisy estimates Xshifted = X - (1 + np.random.randn(1)*1e-2)*shift[1] Yshifted = Y - (1 + np.random.randn(1)*1e-2)*shift[0] theta = (1 + np.random.randn(1)*1e-2)*theta Xrot = (Xshifted-W/2)*np.cos(theta) -\ (Yshifted-H/2)*np.sin(theta) + W/2 Yrot = (Xshifted-W/2)*np.sin(theta) +\ (Yshifted-H/2)*np.cos(theta) + H/2 Xnew = cv2.resize(Xrot, (Wl, Hl), interpolation=cv2.INTER_AREA) Ynew = cv2.resize(Yrot, (Wl, Hl), interpolation=cv2.INTER_AREA) Xstack[idx, ...] = 2*Xnew/W - 1 Ystack[idx, ...] = 2*Ynew/H - 1 return imstack, Xstack, Ystack, shifts, thetas def register_stack(imstack, full_res, method=StackReg.RIGID_BODY): ''' Register a stack of images and get coordinates Inputs: imstack: nimg x H x W stack of images full_res: Resolution at which images will be super resolved method: Method to use for registration. Default is StackReg.RIGID_BODY Outputs: reg_mats: (nimg, 2, 3) dimensional registration matrices ''' nimg, H, W = imstack.shape Hr, Wr = full_res imstack_full = np.zeros((nimg, Hr, Wr)) # Upsample the images for idx in range(nimg): imstack_full[idx, ...] = cv2.resize(imstack[idx, ...], (Wr, Hr), interpolation=cv2.INTER_AREA) # Now register the stack reg = StackReg(method) reg_mats = reg.register_stack(imstack_full, reference='first', verbose=True) return reg_mats def invert_regstack(regstack): ''' Invert affine matrices ''' nimg = regstack.shape[0] regstack_inv = np.zeros_like(regstack) last_row = np.zeros((1, 3)) last_row[0, 2] = 1 for idx in range(nimg): mat = linalg.inv(np.vstack((regstack[idx, ...], last_row)))[:2, :] regstack_inv[idx, ...] = mat return regstack_inv def mat2coords(reg_stack, full_res, low_res): ''' Computed 2D coordinates from affine matrices Inputs: reg_stack: (nimg, 2, 3) registration stack res: Resolution of images ''' nimg, _, _ = reg_stack.shape H, W = full_res Y, X = np.mgrid[:H, :W] Hl, Wl = low_res coords = np.hstack((X.reshape(-1, 1), Y.reshape(-1, 1), np.ones((H*W, 1)))) Xstack = np.zeros((nimg, Hl, Wl), dtype=np.float32) Ystack = np.zeros_like(Xstack) last_row = np.zeros((1, 3)) last_row[0, 2] = 1 for idx in range(nimg): mat = linalg.inv(np.vstack((reg_stack[idx, ...], last_row))) coords_new = mat.dot(coords.T).T Xstack[idx, ...] = cv2.resize(2*coords_new[:, 0].reshape(H, W)/W - 1, (Wl, Hl), interpolation=cv2.INTER_AREA) Ystack[idx, ...] = cv2.resize(2*coords_new[:, 1].reshape(H, W)/H - 1, (Wl, Hl), interpolation=cv2.INTER_AREA) return Xstack, Ystack def param2theta(params, w, h): ''' Convert affine matrix to parameter that torch can use Inputs: params: nimg x 2 x 3 affine matrices w, h: Width and height of the image Outputs: theta: Matrix to use with grid_sample (for example) Reference: https://discuss.pytorch.org/t/how-to-convert-an-affine-transform-matrix-into-theta-to-use-torch-nn-functional-affine-grid/24315/4 ''' last_row = np.zeros((1, 3), dtype=np.float32) last_row[0, 2] = 1 theta = np.zeros_like(params) for idx in range(params.shape[0]): param = np.vstack((params[idx, ...], last_row)) param = np.linalg.inv(param) theta[idx,0,0] = param[0,0] theta[idx,0,1] = param[0,1]*h/w theta[idx,0,2] = param[0,2]*2/w + theta[idx,0,0] + theta[idx,0,1] - 1 #theta[idx, 0, 2] = param[0, 2]*2/w + param[0, 0] + param[0, 1] - 1 theta[idx,1,0] = param[1,0]*w/h theta[idx,1,1] = param[1,1] theta[idx,1,2] = param[1,2]*2/h + theta[idx,1,0] + theta[idx,1,1] - 1 #theta[idx, 1, 2] = param[1, 2]*2/h + param[1, 0] + param[1, 1] - 1 return theta def affine2rigid(mats): ''' Compute rigid body transformations from affine matrices Inputs: mats: (nmats, 2, 3) affine matrices Outputs: translations: (nmats, 2) translation array angles: (nmats) angles array ''' # Compute average angle to reduce numerical errors if False: angles = (np.arccos(mats[:, 0, 0]) - np.arcsin(mats[:, 0, 1]) + np.arcsin(mats[:, 1, 0]) + np.arccos(mats[:, 1, 1]))/4.0 angles = np.arccos(mats[:, 0, 0]) translations = mats[:, :, 2] return angles, translations def get_transformed_coords(theta, imsize): ''' Compute transformed coordinates for given affine matrices ''' B = theta.shape[0] H, W = imsize return F.affine_grid(theta, (B, 1, H, W)).reshape(-1, H*W, 2) def interp_lr(imref, coords, renderer): ''' Compute affine transformed images from coordinates at high resolution Inputs: imref: (1, 1, H, W) low resolution image, upsampled coords: (B, H, W, 2) high resolution coordinates renderer: Function to downsample the images Outputs: im_lr: (B, 1, Hl, Wl) low resolution transformed images ''' B = coords.shape[0] im_hr = F.grid_sample(torch.repeat_interleave(imref, B, 0), coords, mode='bilinear', align_corners=False) im_lr = renderer.integrator(im_hr) return im_lr def register_stack_ecc(imstack, full_res, method=cv2.MOTION_EUCLIDEAN): ''' Register a stack of images and get coordinates Inputs: imstack: nimg x H x W stack of images full_res: Resolution at which images will be super resolved method: Method to use for ECC registration Outputs: Xstack: X Coordinates for registration Ystack: Y Coordinates for registration mask: (nimg, ) dimensional mask for images that were successfully registered alignment_err: (nimg, ) dimensional array of alignment error ''' nimg, H, W = imstack.shape Hr, Wr = full_res mask = np.zeros(nimg) alignment_err = np.zeros(nimg) Xstack = np.zeros((nimg, H, W), dtype=np.float32) Ystack = np.zeros((nimg, H, W), dtype=np.float32) imref = cv2.resize(imstack[0, ...], (Wr, Hr), interpolation=cv2.INTER_LINEAR) mask[0] = 1 Y, X = np.mgrid[:Hr, :Wr] X = 2*X/Wr - 1 Y = 2*Y/Hr - 1 Xstack[0, ...] = cv2.resize(X, (W, H), interpolation=cv2.INTER_LINEAR) Ystack[0, ...] = cv2.resize(Y, (W, H), interpolation=cv2.INTER_LINEAR) if method == cv2.MOTION_HOMOGRAPHY: ecc_mats = np.zeros((nimg, 3, 3)) ecc_mats[0, 2, 2] = 1 else: ecc_mats = np.zeros((nimg, 2, 3)) # First image is registered ... to itself ecc_mats[0, 0, 0] = 1 ecc_mats[0, 1, 1] = 1 for idx in tqdm.tqdm(range(1, nimg)): im2 = cv2.resize(imstack[idx, ...], (Wr, Hr), interpolation=cv2.INTER_LINEAR) try: mat, im2_aligned, flow = ecc_flow(imref, im2, warp_mode=method) mask[idx] = 1 ecc_mats[idx, :] = mat Xstack[idx, ...] = cv2.resize(X - flow[..., 0], (W, H), interpolation=cv2.INTER_LINEAR) Ystack[idx, ...] = cv2.resize(Y - flow[..., 1], (W, H), interpolation=cv2.INTER_LINEAR) spatial_mask = (im2_aligned != 0) alignment_err[idx] = abs((imref - im2_aligned)*spatial_mask).mean() except: mask[idx] = 0 continue # Now return the coordinates return Xstack, Ystack, mask, ecc_mats, alignment_err def prune_stack(imstack, ecc_mats, full_res, thres=None): ''' Prune a stack of images which are not well registered. Inputs: imstack: nimg x H x W stack of images ecc_mats: nimg x 2 x 3 stack of transformation matrices full_res: Full resolution size thres: Threshold of registration error to consider when rejecting images. If None, 2*median(error_array) is used Outputs: imstack: nimg_good x H x W stack of good images ecc_mats: nimg_good x 2 x 3 stack of good transformation matrices ''' nimg, Hl, Wl = imstack.shape H, W = full_res if thres is None: thres = 1 imref = cv2.resize(imstack[0, ...], (W, H), interpolation=cv2.INTER_AREA) imten = torch.tensor(imref).cuda()[None, None, ...] imstack_ten = torch.tensor(imstack).cuda()[:, None, ...] imten = torch.repeat_interleave(imten, int(nimg), 0) mat = torch.tensor(ecc_mats.astype(np.float32)).cuda() imtrans = kornia.warp_affine(imten, mat, (Hl, Wl)) imdiff = abs(imtrans - imstack_ten).cpu()[:, 0, ...] diff_array = (imdiff/(imstack + 1e-2*imstack.max())).mean(-1).mean(-1) mask = diff_array < thres imstack = np.copy(imstack[mask == 1, ...], order='C') ecc_mats = np.copy(ecc_mats[mask == 1, ...], order='C') imdiff = imdiff[mask == 1, ...] return imstack, ecc_mats, mask, imdiff def flow2rgb(flow): ''' Convert flow to an RGB image to visualize. ''' H, W, _ = flow.shape hsv = np.zeros((H, W, 3), dtype=np.uint8) hsv[..., 1] = 255 mag, ang = cv2.cartToPolar(flow[..., 0], flow[..., 1]) hsv[..., 0] = ang*180/np.pi/2 hsv[..., 2] = cv2.normalize(mag, None, 0, 255, cv2.NORM_MINMAX) return cv2.cvtColor(hsv, cv2.COLOR_HSV2RGB)
23,853
33.772595
137
py
DeepIR
DeepIR-main/modules/deep_prior.py
#!/usr/bin/env ''' One single file for all things Deep Image Prior ''' import os import sys import tqdm import pdb import numpy as np import torch from torch import nn import torchvision import cv2 from dmodels.skip import skip from dmodels.texture_nets import get_texture_nets from dmodels.resnet import ResNet from dmodels.unet import UNet class Downsampler(nn.Module): ''' http://www.realitypixels.com/turk/computergraphics/ResamplingFilters.pdf ''' def __init__(self, n_planes, factor, kernel_type, phase=0, kernel_width=None, support=None, sigma=None, preserve_size=False): super(Downsampler, self).__init__() assert phase in [0, 0.5], 'phase should be 0 or 0.5' if kernel_type == 'lanczos2': support = 2 kernel_width = 4 * factor + 1 kernel_type_ = 'lanczos' elif kernel_type == 'lanczos3': support = 3 kernel_width = 6 * factor + 1 kernel_type_ = 'lanczos' elif kernel_type == 'gauss12': kernel_width = 7 sigma = 1/2 kernel_type_ = 'gauss' elif kernel_type == 'gauss1sq2': kernel_width = 9 sigma = 1./np.sqrt(2) kernel_type_ = 'gauss' elif kernel_type in ['lanczos', 'gauss', 'box']: kernel_type_ = kernel_type else: assert False, 'wrong name kernel' # note that `kernel width` will be different to actual size for phase = 1/2 self.kernel = get_kernel(factor, kernel_type_, phase, kernel_width, support=support, sigma=sigma) downsampler = nn.Conv2d(n_planes, n_planes, kernel_size=self.kernel.shape, stride=factor, padding=0) downsampler.weight.data[:] = 0 downsampler.bias.data[:] = 0 kernel_torch = torch.from_numpy(self.kernel) for i in range(n_planes): downsampler.weight.data[i, i] = kernel_torch self.downsampler_ = downsampler if preserve_size: if self.kernel.shape[0] % 2 == 1: pad = int((self.kernel.shape[0] - 1) / 2.) else: pad = int((self.kernel.shape[0] - factor) / 2.) self.padding = nn.ReplicationPad2d(pad) self.preserve_size = preserve_size def forward(self, input): if self.preserve_size: x = self.padding(input) else: x= input self.x = x return self.downsampler_(x) def get_kernel(factor, kernel_type, phase, kernel_width, support=None, sigma=None): assert kernel_type in ['lanczos', 'gauss', 'box'] # factor = float(factor) if phase == 0.5 and kernel_type != 'box': kernel = np.zeros([kernel_width - 1, kernel_width - 1]) else: kernel = np.zeros([kernel_width, kernel_width]) if kernel_type == 'box': assert phase == 0.5, 'Box filter is always half-phased' kernel[:] = 1./(kernel_width * kernel_width) elif kernel_type == 'gauss': assert sigma, 'sigma is not specified' assert phase != 0.5, 'phase 1/2 for gauss not implemented' center = (kernel_width + 1.)/2. print(center, kernel_width) sigma_sq = sigma * sigma for i in range(1, kernel.shape[0] + 1): for j in range(1, kernel.shape[1] + 1): di = (i - center)/2. dj = (j - center)/2. kernel[i - 1][j - 1] = np.exp(-(di * di + dj * dj)/(2 * sigma_sq)) kernel[i - 1][j - 1] = kernel[i - 1][j - 1]/(2. * np.pi * sigma_sq) elif kernel_type == 'lanczos': assert support, 'support is not specified' center = (kernel_width + 1) / 2. for i in range(1, kernel.shape[0] + 1): for j in range(1, kernel.shape[1] + 1): if phase == 0.5: di = abs(i + 0.5 - center) / factor dj = abs(j + 0.5 - center) / factor else: di = abs(i - center) / factor dj = abs(j - center) / factor pi_sq = np.pi * np.pi val = 1 if di != 0: val = val * support * np.sin(np.pi * di) * np.sin(np.pi * di / support) val = val / (np.pi * np.pi * di * di) if dj != 0: val = val * support * np.sin(np.pi * dj) * np.sin(np.pi * dj / support) val = val / (np.pi * np.pi * dj * dj) kernel[i - 1][j - 1] = val else: assert False, 'wrong method name' kernel /= kernel.sum() return kernel def get_noise(input_depth, method, spatial_size, noise_type='u', var=1./10): """Returns a pytorch.Tensor of size (1 x `input_depth` x `spatial_size[0]` x `spatial_size[1]`) initialized in a specific way. Args: input_depth: number of channels in the tensor method: `noise` for fillting tensor with noise; `meshgrid` for np.meshgrid spatial_size: spatial size of the tensor to initialize noise_type: 'u' for uniform; 'n' for normal var: a factor, a noise will be multiplicated by. Basically it is standard deviation scaler. """ if isinstance(spatial_size, int): spatial_size = (spatial_size, spatial_size) if method == 'noise': shape = [1, input_depth, spatial_size[0], spatial_size[1]] net_input = torch.zeros(shape) fill_noise(net_input, noise_type) net_input *= var elif method == 'meshgrid': assert input_depth == 2 X, Y = np.meshgrid(np.arange(0, spatial_size[1])/float(spatial_size[1]-1), np.arange(0, spatial_size[0])/float(spatial_size[0]-1)) meshgrid = np.concatenate([X[None,:], Y[None,:]]) net_input= np_to_torch(meshgrid) else: assert False return net_input def np_to_torch(img_np): '''Converts image in numpy.array to torch.Tensor. From C x W x H [0..1] to C x W x H [0..1] ''' return torch.from_numpy(img_np)[None, :] def torch_to_np(img_var): '''Converts an image in torch.Tensor format to np.array. From 1 x C x W x H [0..1] to C x W x H [0..1] ''' return img_var.detach().cpu().numpy()[0] def fill_noise(x, noise_type): """Fills tensor `x` with noise of type `noise_type`.""" if noise_type == 'u': x.uniform_() elif noise_type == 'n': x.normal_() else: assert False def get_image_grid(images_np, nrow=8): '''Creates a grid from a list of images by concatenating them.''' images_torch = [torch.from_numpy(x) for x in images_np] torch_grid = torchvision.utils.make_grid(images_torch, nrow) return torch_grid.numpy() def get_net(input_depth, NET_TYPE, pad, upsample_mode, n_channels=3, act_fun='LeakyReLU', skip_n33d=128, skip_n33u=128, skip_n11=4, num_scales=5, downsample_mode='stride'): if NET_TYPE == 'ResNet': # TODO net = ResNet(input_depth, 3, 10, 16, 1, nn.BatchNorm2d, False) elif NET_TYPE == 'skip': net = skip(input_depth, n_channels, num_channels_down = [skip_n33d]*num_scales if isinstance(skip_n33d, int) else skip_n33d, num_channels_up = [skip_n33u]*num_scales if isinstance(skip_n33u, int) else skip_n33u, num_channels_skip = [skip_n11]*num_scales if isinstance(skip_n11, int) else skip_n11, upsample_mode=upsample_mode, downsample_mode=downsample_mode, need_sigmoid=True, need_bias=True, pad=pad, act_fun=act_fun) elif NET_TYPE == 'texture_nets': net = get_texture_nets(inp=input_depth, ratios = [32, 16, 8, 4, 2, 1], fill_noise=False,pad=pad) elif NET_TYPE =='UNet': net = UNet(num_input_channels=input_depth, num_output_channels=3, feature_scale=4, more_layers=0, concat_x=False, upsample_mode=upsample_mode, pad=pad, norm_layer=nn.BatchNorm2d, need_sigmoid=True, need_bias=True) elif NET_TYPE == 'identity': assert input_depth == 3 net = nn.Sequential() else: assert False return net
8,713
33.995984
138
py
DeepIR
DeepIR-main/modules/dmodels/skip.py
import torch import torch.nn as nn from .common import * def skip( num_input_channels=2, num_output_channels=3, num_channels_down=[16, 32, 64, 128, 128], num_channels_up=[16, 32, 64, 128, 128], num_channels_skip=[4, 4, 4, 4, 4], filter_size_down=3, filter_size_up=3, filter_skip_size=1, need_sigmoid=True, need_bias=True, pad='zero', upsample_mode='nearest', downsample_mode='stride', act_fun='LeakyReLU', need1x1_up=True): """Assembles encoder-decoder with skip connections. Arguments: act_fun: Either string 'LeakyReLU|Swish|ELU|none' or module (e.g. nn.ReLU) pad (string): zero|reflection (default: 'zero') upsample_mode (string): 'nearest|bilinear' (default: 'nearest') downsample_mode (string): 'stride|avg|max|lanczos2' (default: 'stride') """ assert len(num_channels_down) == len(num_channels_up) == len(num_channels_skip) n_scales = len(num_channels_down) if not (isinstance(upsample_mode, list) or isinstance(upsample_mode, tuple)) : upsample_mode = [upsample_mode]*n_scales if not (isinstance(downsample_mode, list)or isinstance(downsample_mode, tuple)): downsample_mode = [downsample_mode]*n_scales if not (isinstance(filter_size_down, list) or isinstance(filter_size_down, tuple)) : filter_size_down = [filter_size_down]*n_scales if not (isinstance(filter_size_up, list) or isinstance(filter_size_up, tuple)) : filter_size_up = [filter_size_up]*n_scales last_scale = n_scales - 1 cur_depth = None model = nn.Sequential() model_tmp = model input_depth = num_input_channels for i in range(len(num_channels_down)): deeper = nn.Sequential() skip = nn.Sequential() if num_channels_skip[i] != 0: model_tmp.add(Concat(1, skip, deeper)) else: model_tmp.add(deeper) model_tmp.add(bn(num_channels_skip[i] + (num_channels_up[i + 1] if i < last_scale else num_channels_down[i]))) if num_channels_skip[i] != 0: skip.add(conv(input_depth, num_channels_skip[i], filter_skip_size, bias=need_bias, pad=pad)) skip.add(bn(num_channels_skip[i])) skip.add(act(act_fun)) # skip.add(Concat(2, GenNoise(nums_noise[i]), skip_part)) deeper.add(conv(input_depth, num_channels_down[i], filter_size_down[i], 2, bias=need_bias, pad=pad, downsample_mode=downsample_mode[i])) deeper.add(bn(num_channels_down[i])) deeper.add(act(act_fun)) deeper.add(conv(num_channels_down[i], num_channels_down[i], filter_size_down[i], bias=need_bias, pad=pad)) deeper.add(bn(num_channels_down[i])) deeper.add(act(act_fun)) deeper_main = nn.Sequential() if i == len(num_channels_down) - 1: # The deepest k = num_channels_down[i] else: deeper.add(deeper_main) k = num_channels_up[i + 1] deeper.add(nn.Upsample(scale_factor=2, mode=upsample_mode[i])) model_tmp.add(conv(num_channels_skip[i] + k, num_channels_up[i], filter_size_up[i], 1, bias=need_bias, pad=pad)) model_tmp.add(bn(num_channels_up[i])) model_tmp.add(act(act_fun)) if need1x1_up: model_tmp.add(conv(num_channels_up[i], num_channels_up[i], 1, bias=need_bias, pad=pad)) model_tmp.add(bn(num_channels_up[i])) model_tmp.add(act(act_fun)) input_depth = num_channels_down[i] model_tmp = deeper_main model.add(conv(num_channels_up[0], num_output_channels, 1, bias=need_bias, pad=pad)) if need_sigmoid: model.add(nn.Sigmoid()) return model
3,744
36.079208
144
py
DeepIR
DeepIR-main/modules/dmodels/resnet.py
import torch import torch.nn as nn from numpy.random import normal from numpy.linalg import svd from math import sqrt import torch.nn.init from .common import * class ResidualSequential(nn.Sequential): def __init__(self, *args): super(ResidualSequential, self).__init__(*args) def forward(self, x): out = super(ResidualSequential, self).forward(x) # print(x.size(), out.size()) x_ = None if out.size(2) != x.size(2) or out.size(3) != x.size(3): diff2 = x.size(2) - out.size(2) diff3 = x.size(3) - out.size(3) # print(1) x_ = x[:, :, diff2 /2:out.size(2) + diff2 / 2, diff3 / 2:out.size(3) + diff3 / 2] else: x_ = x return out + x_ def eval(self): print(2) for m in self.modules(): m.eval() exit() def get_block(num_channels, norm_layer, act_fun): layers = [ nn.Conv2d(num_channels, num_channels, 3, 1, 1, bias=False), norm_layer(num_channels, affine=True), act(act_fun), nn.Conv2d(num_channels, num_channels, 3, 1, 1, bias=False), norm_layer(num_channels, affine=True), ] return layers class ResNet(nn.Module): def __init__(self, num_input_channels, num_output_channels, num_blocks, num_channels, need_residual=True, act_fun='LeakyReLU', need_sigmoid=True, norm_layer=nn.BatchNorm2d, pad='reflection'): ''' pad = 'start|zero|replication' ''' super(ResNet, self).__init__() if need_residual: s = ResidualSequential else: s = nn.Sequential stride = 1 # First layers layers = [ # nn.ReplicationPad2d(num_blocks * 2 * stride + 3), conv(num_input_channels, num_channels, 3, stride=1, bias=True, pad=pad), act(act_fun) ] # Residual blocks # layers_residual = [] for i in range(num_blocks): layers += [s(*get_block(num_channels, norm_layer, act_fun))] layers += [ nn.Conv2d(num_channels, num_channels, 3, 1, 1), norm_layer(num_channels, affine=True) ] # if need_residual: # layers += [ResidualSequential(*layers_residual)] # else: # layers += [Sequential(*layers_residual)] # if factor >= 2: # # Do upsampling if needed # layers += [ # nn.Conv2d(num_channels, num_channels * # factor ** 2, 3, 1), # nn.PixelShuffle(factor), # act(act_fun) # ] layers += [ conv(num_channels, num_output_channels, 3, 1, bias=True, pad=pad), nn.Sigmoid() ] self.model = nn.Sequential(*layers) def forward(self, input): return self.model(input) def eval(self): self.model.eval()
2,943
29.350515
195
py
DeepIR
DeepIR-main/modules/dmodels/downsampler.py
import numpy as np import torch import torch.nn as nn class Downsampler(nn.Module): ''' http://www.realitypixels.com/turk/computergraphics/ResamplingFilters.pdf ''' def __init__(self, n_planes, factor, kernel_type, phase=0, kernel_width=None, support=None, sigma=None, preserve_size=False): super(Downsampler, self).__init__() assert phase in [0, 0.5], 'phase should be 0 or 0.5' if kernel_type == 'lanczos2': support = 2 kernel_width = 4 * factor + 1 kernel_type_ = 'lanczos' elif kernel_type == 'lanczos3': support = 3 kernel_width = 6 * factor + 1 kernel_type_ = 'lanczos' elif kernel_type == 'gauss12': kernel_width = 7 sigma = 1/2 kernel_type_ = 'gauss' elif kernel_type == 'gauss1sq2': kernel_width = 9 sigma = 1./np.sqrt(2) kernel_type_ = 'gauss' elif kernel_type in ['lanczos', 'gauss', 'box']: kernel_type_ = kernel_type else: assert False, 'wrong name kernel' # note that `kernel width` will be different to actual size for phase = 1/2 self.kernel = get_kernel(factor, kernel_type_, phase, kernel_width, support=support, sigma=sigma) downsampler = nn.Conv2d(n_planes, n_planes, kernel_size=self.kernel.shape, stride=factor, padding=0) downsampler.weight.data[:] = 0 downsampler.bias.data[:] = 0 kernel_torch = torch.from_numpy(self.kernel) for i in range(n_planes): downsampler.weight.data[i, i] = kernel_torch self.downsampler_ = downsampler if preserve_size: if self.kernel.shape[0] % 2 == 1: pad = int((self.kernel.shape[0] - 1) / 2.) else: pad = int((self.kernel.shape[0] - factor) / 2.) self.padding = nn.ReplicationPad2d(pad) self.preserve_size = preserve_size def forward(self, input): if self.preserve_size: x = self.padding(input) else: x= input self.x = x return self.downsampler_(x) def get_kernel(factor, kernel_type, phase, kernel_width, support=None, sigma=None): assert kernel_type in ['lanczos', 'gauss', 'box'] # factor = float(factor) if phase == 0.5 and kernel_type != 'box': kernel = np.zeros([kernel_width - 1, kernel_width - 1]) else: kernel = np.zeros([kernel_width, kernel_width]) if kernel_type == 'box': assert phase == 0.5, 'Box filter is always half-phased' kernel[:] = 1./(kernel_width * kernel_width) elif kernel_type == 'gauss': assert sigma, 'sigma is not specified' assert phase != 0.5, 'phase 1/2 for gauss not implemented' center = (kernel_width + 1.)/2. print(center, kernel_width) sigma_sq = sigma * sigma for i in range(1, kernel.shape[0] + 1): for j in range(1, kernel.shape[1] + 1): di = (i - center)/2. dj = (j - center)/2. kernel[i - 1][j - 1] = np.exp(-(di * di + dj * dj)/(2 * sigma_sq)) kernel[i - 1][j - 1] = kernel[i - 1][j - 1]/(2. * np.pi * sigma_sq) elif kernel_type == 'lanczos': assert support, 'support is not specified' center = (kernel_width + 1) / 2. for i in range(1, kernel.shape[0] + 1): for j in range(1, kernel.shape[1] + 1): if phase == 0.5: di = abs(i + 0.5 - center) / factor dj = abs(j + 0.5 - center) / factor else: di = abs(i - center) / factor dj = abs(j - center) / factor pi_sq = np.pi * np.pi val = 1 if di != 0: val = val * support * np.sin(np.pi * di) * np.sin(np.pi * di / support) val = val / (np.pi * np.pi * di * di) if dj != 0: val = val * support * np.sin(np.pi * dj) * np.sin(np.pi * dj / support) val = val / (np.pi * np.pi * dj * dj) kernel[i - 1][j - 1] = val else: assert False, 'wrong method name' kernel /= kernel.sum() return kernel #a = Downsampler(n_planes=3, factor=2, kernel_type='lanczos2', phase='1', preserve_size=True) ################# # Learnable downsampler # KS = 32 # dow = nn.Sequential(nn.ReplicationPad2d(int((KS - factor) / 2.)), nn.Conv2d(1,1,KS,factor)) # class Apply(nn.Module): # def __init__(self, what, dim, *args): # super(Apply, self).__init__() # self.dim = dim # self.what = what # def forward(self, input): # inputs = [] # for i in range(input.size(self.dim)): # inputs.append(self.what(input.narrow(self.dim, i, 1))) # return torch.cat(inputs, dim=self.dim) # def __len__(self): # return len(self._modules) # downs = Apply(dow, 1) # downs.type(dtype)(net_input.type(dtype)).size()
5,379
30.83432
129
py
DeepIR
DeepIR-main/modules/dmodels/dcgan.py
import torch import torch.nn as nn def dcgan(inp=2, ndf=32, num_ups=4, need_sigmoid=True, need_bias=True, pad='zero', upsample_mode='nearest', need_convT = True): layers= [nn.ConvTranspose2d(inp, ndf, kernel_size=3, stride=1, padding=0, bias=False), nn.BatchNorm2d(ndf), nn.LeakyReLU(True)] for i in range(num_ups-3): if need_convT: layers += [ nn.ConvTranspose2d(ndf, ndf, kernel_size=4, stride=2, padding=1, bias=False), nn.BatchNorm2d(ndf), nn.LeakyReLU(True)] else: layers += [ nn.Upsample(scale_factor=2, mode=upsample_mode), nn.Conv2d(ndf, ndf, kernel_size=3, stride=1, padding=1, bias=False), nn.BatchNorm2d(ndf), nn.LeakyReLU(True)] if need_convT: layers += [nn.ConvTranspose2d(ndf, 3, 4, 2, 1, bias=False),] else: layers += [nn.Upsample(scale_factor=2, mode='bilinear'), nn.Conv2d(ndf, 3, kernel_size=3, stride=1, padding=1, bias=False)] if need_sigmoid: layers += [nn.Sigmoid()] model =nn.Sequential(*layers) return model
1,244
35.617647
112
py
DeepIR
DeepIR-main/modules/dmodels/texture_nets.py
import torch import torch.nn as nn from .common import * normalization = nn.BatchNorm2d def conv(in_f, out_f, kernel_size, stride=1, bias=True, pad='zero'): if pad == 'zero': return nn.Conv2d(in_f, out_f, kernel_size, stride, padding=(kernel_size - 1) / 2, bias=bias) elif pad == 'reflection': layers = [nn.ReflectionPad2d((kernel_size - 1) / 2), nn.Conv2d(in_f, out_f, kernel_size, stride, padding=0, bias=bias)] return nn.Sequential(*layers) def get_texture_nets(inp=3, ratios = [32, 16, 8, 4, 2, 1], fill_noise=False, pad='zero', need_sigmoid=False, conv_num=8, upsample_mode='nearest'): for i in range(len(ratios)): j = i + 1 seq = nn.Sequential() tmp = nn.AvgPool2d(ratios[i], ratios[i]) seq.add(tmp) if fill_noise: seq.add(GenNoise(inp)) seq.add(conv(inp, conv_num, 3, pad=pad)) seq.add(normalization(conv_num)) seq.add(act()) seq.add(conv(conv_num, conv_num, 3, pad=pad)) seq.add(normalization(conv_num)) seq.add(act()) seq.add(conv(conv_num, conv_num, 1, pad=pad)) seq.add(normalization(conv_num)) seq.add(act()) if i == 0: seq.add(nn.Upsample(scale_factor=2, mode=upsample_mode)) cur = seq else: cur_temp = cur cur = nn.Sequential() # Batch norm before merging seq.add(normalization(conv_num)) cur_temp.add(normalization(conv_num * (j - 1))) cur.add(Concat(1, cur_temp, seq)) cur.add(conv(conv_num * j, conv_num * j, 3, pad=pad)) cur.add(normalization(conv_num * j)) cur.add(act()) cur.add(conv(conv_num * j, conv_num * j, 3, pad=pad)) cur.add(normalization(conv_num * j)) cur.add(act()) cur.add(conv(conv_num * j, conv_num * j, 1, pad=pad)) cur.add(normalization(conv_num * j)) cur.add(act()) if i == len(ratios) - 1: cur.add(conv(conv_num * j, 3, 1, pad=pad)) else: cur.add(nn.Upsample(scale_factor=2, mode=upsample_mode)) model = cur if need_sigmoid: model.add(nn.Sigmoid()) return model
2,315
27.95
146
py
DeepIR
DeepIR-main/modules/dmodels/common.py
import torch import torch.nn as nn import numpy as np from .downsampler import Downsampler def add_module(self, module): self.add_module(str(len(self) + 1), module) torch.nn.Module.add = add_module class Concat(nn.Module): def __init__(self, dim, *args): super(Concat, self).__init__() self.dim = dim for idx, module in enumerate(args): self.add_module(str(idx), module) def forward(self, input): inputs = [] for module in self._modules.values(): inputs.append(module(input)) inputs_shapes2 = [x.shape[2] for x in inputs] inputs_shapes3 = [x.shape[3] for x in inputs] if np.all(np.array(inputs_shapes2) == min(inputs_shapes2)) and np.all(np.array(inputs_shapes3) == min(inputs_shapes3)): inputs_ = inputs else: target_shape2 = min(inputs_shapes2) target_shape3 = min(inputs_shapes3) inputs_ = [] for inp in inputs: diff2 = (inp.size(2) - target_shape2) // 2 diff3 = (inp.size(3) - target_shape3) // 2 inputs_.append(inp[:, :, diff2: diff2 + target_shape2, diff3:diff3 + target_shape3]) return torch.cat(inputs_, dim=self.dim) def __len__(self): return len(self._modules) class GenNoise(nn.Module): def __init__(self, dim2): super(GenNoise, self).__init__() self.dim2 = dim2 def forward(self, input): a = list(input.size()) a[1] = self.dim2 # print (input.data.type()) b = torch.zeros(a).type_as(input.data) b.normal_() x = torch.autograd.Variable(b) return x class Swish(nn.Module): """ https://arxiv.org/abs/1710.05941 The hype was so huge that I could not help but try it """ def __init__(self): super(Swish, self).__init__() self.s = nn.Sigmoid() def forward(self, x): return x * self.s(x) def act(act_fun = 'LeakyReLU'): ''' Either string defining an activation function or module (e.g. nn.ReLU) ''' if isinstance(act_fun, str): if act_fun == 'LeakyReLU': return nn.LeakyReLU(0.2, inplace=True) elif act_fun == 'Swish': return Swish() elif act_fun == 'ELU': return nn.ELU() elif act_fun == 'none': return nn.Sequential() else: assert False else: return act_fun() def bn(num_features): return nn.BatchNorm2d(num_features) def conv(in_f, out_f, kernel_size, stride=1, bias=True, pad='zero', downsample_mode='stride'): downsampler = None if stride != 1 and downsample_mode != 'stride': if downsample_mode == 'avg': downsampler = nn.AvgPool2d(stride, stride) elif downsample_mode == 'max': downsampler = nn.MaxPool2d(stride, stride) elif downsample_mode in ['lanczos2', 'lanczos3']: downsampler = Downsampler(n_planes=out_f, factor=stride, kernel_type=downsample_mode, phase=0.5, preserve_size=True) else: assert False stride = 1 padder = None to_pad = int((kernel_size - 1) / 2) if pad == 'reflection': padder = nn.ReflectionPad2d(to_pad) to_pad = 0 convolver = nn.Conv2d(in_f, out_f, kernel_size, stride, padding=to_pad, bias=bias) layers = filter(lambda x: x is not None, [padder, convolver, downsampler]) return nn.Sequential(*layers)
3,531
27.483871
128
py
DeepIR
DeepIR-main/modules/dmodels/unet.py
import torch.nn as nn import torch import torch.nn as nn import torch.nn.functional as F from .common import * class ListModule(nn.Module): def __init__(self, *args): super(ListModule, self).__init__() idx = 0 for module in args: self.add_module(str(idx), module) idx += 1 def __getitem__(self, idx): if idx >= len(self._modules): raise IndexError('index {} is out of range'.format(idx)) if idx < 0: idx = len(self) + idx it = iter(self._modules.values()) for i in range(idx): next(it) return next(it) def __iter__(self): return iter(self._modules.values()) def __len__(self): return len(self._modules) class UNet(nn.Module): ''' upsample_mode in ['deconv', 'nearest', 'bilinear'] pad in ['zero', 'replication', 'none'] ''' def __init__(self, num_input_channels=3, num_output_channels=3, feature_scale=4, more_layers=0, concat_x=False, upsample_mode='deconv', pad='zero', norm_layer=nn.InstanceNorm2d, need_sigmoid=True, need_bias=True): super(UNet, self).__init__() self.feature_scale = feature_scale self.more_layers = more_layers self.concat_x = concat_x filters = [64, 128, 256, 512, 1024] filters = [x // self.feature_scale for x in filters] self.start = unetConv2(num_input_channels, filters[0] if not concat_x else filters[0] - num_input_channels, norm_layer, need_bias, pad) self.down1 = unetDown(filters[0], filters[1] if not concat_x else filters[1] - num_input_channels, norm_layer, need_bias, pad) self.down2 = unetDown(filters[1], filters[2] if not concat_x else filters[2] - num_input_channels, norm_layer, need_bias, pad) self.down3 = unetDown(filters[2], filters[3] if not concat_x else filters[3] - num_input_channels, norm_layer, need_bias, pad) self.down4 = unetDown(filters[3], filters[4] if not concat_x else filters[4] - num_input_channels, norm_layer, need_bias, pad) # more downsampling layers if self.more_layers > 0: self.more_downs = [ unetDown(filters[4], filters[4] if not concat_x else filters[4] - num_input_channels , norm_layer, need_bias, pad) for i in range(self.more_layers)] self.more_ups = [unetUp(filters[4], upsample_mode, need_bias, pad, same_num_filt =True) for i in range(self.more_layers)] self.more_downs = ListModule(*self.more_downs) self.more_ups = ListModule(*self.more_ups) self.up4 = unetUp(filters[3], upsample_mode, need_bias, pad) self.up3 = unetUp(filters[2], upsample_mode, need_bias, pad) self.up2 = unetUp(filters[1], upsample_mode, need_bias, pad) self.up1 = unetUp(filters[0], upsample_mode, need_bias, pad) self.final = conv(filters[0], num_output_channels, 1, bias=need_bias, pad=pad) if need_sigmoid: self.final = nn.Sequential(self.final, nn.Sigmoid()) def forward(self, inputs): # Downsample downs = [inputs] down = nn.AvgPool2d(2, 2) for i in range(4 + self.more_layers): downs.append(down(downs[-1])) in64 = self.start(inputs) if self.concat_x: in64 = torch.cat([in64, downs[0]], 1) down1 = self.down1(in64) if self.concat_x: down1 = torch.cat([down1, downs[1]], 1) down2 = self.down2(down1) if self.concat_x: down2 = torch.cat([down2, downs[2]], 1) down3 = self.down3(down2) if self.concat_x: down3 = torch.cat([down3, downs[3]], 1) down4 = self.down4(down3) if self.concat_x: down4 = torch.cat([down4, downs[4]], 1) if self.more_layers > 0: prevs = [down4] for kk, d in enumerate(self.more_downs): # print(prevs[-1].size()) out = d(prevs[-1]) if self.concat_x: out = torch.cat([out, downs[kk + 5]], 1) prevs.append(out) up_ = self.more_ups[-1](prevs[-1], prevs[-2]) for idx in range(self.more_layers - 1): l = self.more_ups[self.more - idx - 2] up_= l(up_, prevs[self.more - idx - 2]) else: up_= down4 up4= self.up4(up_, down3) up3= self.up3(up4, down2) up2= self.up2(up3, down1) up1= self.up1(up2, in64) return self.final(up1) class unetConv2(nn.Module): def __init__(self, in_size, out_size, norm_layer, need_bias, pad): super(unetConv2, self).__init__() print(pad) if norm_layer is not None: self.conv1= nn.Sequential(conv(in_size, out_size, 3, bias=need_bias, pad=pad), norm_layer(out_size), nn.ReLU(),) self.conv2= nn.Sequential(conv(out_size, out_size, 3, bias=need_bias, pad=pad), norm_layer(out_size), nn.ReLU(),) else: self.conv1= nn.Sequential(conv(in_size, out_size, 3, bias=need_bias, pad=pad), nn.ReLU(),) self.conv2= nn.Sequential(conv(out_size, out_size, 3, bias=need_bias, pad=pad), nn.ReLU(),) def forward(self, inputs): outputs= self.conv1(inputs) outputs= self.conv2(outputs) return outputs class unetDown(nn.Module): def __init__(self, in_size, out_size, norm_layer, need_bias, pad): super(unetDown, self).__init__() self.conv= unetConv2(in_size, out_size, norm_layer, need_bias, pad) self.down= nn.MaxPool2d(2, 2) def forward(self, inputs): outputs= self.down(inputs) outputs= self.conv(outputs) return outputs class unetUp(nn.Module): def __init__(self, out_size, upsample_mode, need_bias, pad, same_num_filt=False): super(unetUp, self).__init__() num_filt = out_size if same_num_filt else out_size * 2 if upsample_mode == 'deconv': self.up= nn.ConvTranspose2d(num_filt, out_size, 4, stride=2, padding=1) self.conv= unetConv2(out_size * 2, out_size, None, need_bias, pad) elif upsample_mode=='bilinear' or upsample_mode=='nearest': self.up = nn.Sequential(nn.Upsample(scale_factor=2, mode=upsample_mode), conv(num_filt, out_size, 3, bias=need_bias, pad=pad)) self.conv= unetConv2(out_size * 2, out_size, None, need_bias, pad) else: assert False def forward(self, inputs1, inputs2): in1_up= self.up(inputs1) if (inputs2.size(2) != in1_up.size(2)) or (inputs2.size(3) != in1_up.size(3)): diff2 = (inputs2.size(2) - in1_up.size(2)) // 2 diff3 = (inputs2.size(3) - in1_up.size(3)) // 2 inputs2_ = inputs2[:, :, diff2 : diff2 + in1_up.size(2), diff3 : diff3 + in1_up.size(3)] else: inputs2_ = inputs2 output= self.conv(torch.cat([in1_up, inputs2_], 1)) return output
7,324
36.953368
164
py
DeepIR
DeepIR-main/modules/dmodels/__init__.py
from .skip import skip from .texture_nets import get_texture_nets from .resnet import ResNet from .unet import UNet import torch.nn as nn def get_net(input_depth, NET_TYPE, pad, upsample_mode, n_channels=3, act_fun='LeakyReLU', skip_n33d=128, skip_n33u=128, skip_n11=4, num_scales=5, downsample_mode='stride'): if NET_TYPE == 'ResNet': # TODO net = ResNet(input_depth, 3, 10, 16, 1, nn.BatchNorm2d, False) elif NET_TYPE == 'skip': net = skip(input_depth, n_channels, num_channels_down = [skip_n33d]*num_scales if isinstance(skip_n33d, int) else skip_n33d, num_channels_up = [skip_n33u]*num_scales if isinstance(skip_n33u, int) else skip_n33u, num_channels_skip = [skip_n11]*num_scales if isinstance(skip_n11, int) else skip_n11, upsample_mode=upsample_mode, downsample_mode=downsample_mode, need_sigmoid=True, need_bias=True, pad=pad, act_fun=act_fun) elif NET_TYPE == 'texture_nets': net = get_texture_nets(inp=input_depth, ratios = [32, 16, 8, 4, 2, 1], fill_noise=False,pad=pad) elif NET_TYPE =='UNet': net = UNet(num_input_channels=input_depth, num_output_channels=3, feature_scale=4, more_layers=0, concat_x=False, upsample_mode=upsample_mode, pad=pad, norm_layer=nn.BatchNorm2d, need_sigmoid=True, need_bias=True) elif NET_TYPE == 'identity': assert input_depth == 3 net = nn.Sequential() else: assert False return net
1,639
50.25
172
py
AdaptiveGCL
AdaptiveGCL-main/DataHandler.py
import pickle import numpy as np from scipy.sparse import csr_matrix, coo_matrix, dok_matrix from Params import args import scipy.sparse as sp from Utils.TimeLogger import log import torch as t import torch.utils.data as data import torch.utils.data as dataloader class DataHandler: def __init__(self): if args.data == 'yelp': predir = './Datasets/sparse_yelp/' elif args.data == 'lastfm': predir = './Datasets/lastFM/' elif args.data == 'beer': predir = './Datasets/beerAdvocate/' self.predir = predir self.trnfile = predir + 'trnMat.pkl' self.tstfile = predir + 'tstMat.pkl' def loadOneFile(self, filename): with open(filename, 'rb') as fs: ret = (pickle.load(fs) != 0).astype(np.float32) if type(ret) != coo_matrix: ret = sp.coo_matrix(ret) return ret def normalizeAdj(self, mat): degree = np.array(mat.sum(axis=-1)) dInvSqrt = np.reshape(np.power(degree, -0.5), [-1]) dInvSqrt[np.isinf(dInvSqrt)] = 0.0 dInvSqrtMat = sp.diags(dInvSqrt) return mat.dot(dInvSqrtMat).transpose().dot(dInvSqrtMat).tocoo() def makeTorchAdj(self, mat): # make ui adj a = sp.csr_matrix((args.user, args.user)) b = sp.csr_matrix((args.item, args.item)) mat = sp.vstack([sp.hstack([a, mat]), sp.hstack([mat.transpose(), b])]) mat = (mat != 0) * 1.0 mat = (mat + sp.eye(mat.shape[0])) * 1.0 mat = self.normalizeAdj(mat) # make cuda tensor idxs = t.from_numpy(np.vstack([mat.row, mat.col]).astype(np.int64)) vals = t.from_numpy(mat.data.astype(np.float32)) shape = t.Size(mat.shape) return t.sparse.FloatTensor(idxs, vals, shape).cuda() def LoadData(self): trnMat = self.loadOneFile(self.trnfile) tstMat = self.loadOneFile(self.tstfile) self.trnMat = trnMat args.user, args.item = trnMat.shape self.torchBiAdj = self.makeTorchAdj(trnMat) trnData = TrnData(trnMat) self.trnLoader = dataloader.DataLoader(trnData, batch_size=args.batch, shuffle=True, num_workers=0) tstData = TstData(tstMat, trnMat) self.tstLoader = dataloader.DataLoader(tstData, batch_size=args.tstBat, shuffle=False, num_workers=0) class TrnData(data.Dataset): def __init__(self, coomat): self.rows = coomat.row self.cols = coomat.col self.dokmat = coomat.todok() self.negs = np.zeros(len(self.rows)).astype(np.int32) def negSampling(self): for i in range(len(self.rows)): u = self.rows[i] while True: iNeg = np.random.randint(args.item) if (u, iNeg) not in self.dokmat: break self.negs[i] = iNeg def __len__(self): return len(self.rows) def __getitem__(self, idx): return self.rows[idx], self.cols[idx], self.negs[idx] class TstData(data.Dataset): def __init__(self, coomat, trnMat): self.csrmat = (trnMat.tocsr() != 0) * 1.0 tstLocs = [None] * coomat.shape[0] tstUsrs = set() for i in range(len(coomat.data)): row = coomat.row[i] col = coomat.col[i] if tstLocs[row] is None: tstLocs[row] = list() tstLocs[row].append(col) tstUsrs.add(row) tstUsrs = np.array(list(tstUsrs)) self.tstUsrs = tstUsrs self.tstLocs = tstLocs def __len__(self): return len(self.tstUsrs) def __getitem__(self, idx): return self.tstUsrs[idx], np.reshape(self.csrmat[self.tstUsrs[idx]].toarray(), [-1])
3,205
29.245283
103
py
AdaptiveGCL
AdaptiveGCL-main/Main.py
import torch import Utils.TimeLogger as logger from Utils.TimeLogger import log from Params import args from Model import Model, vgae_encoder, vgae_decoder, vgae, DenoisingNet from DataHandler import DataHandler import numpy as np from Utils.Utils import calcRegLoss, pairPredict import os from copy import deepcopy import scipy.sparse as sp import random class Coach: def __init__(self, handler): self.handler = handler print('USER', args.user, 'ITEM', args.item) print('NUM OF INTERACTIONS', self.handler.trnLoader.dataset.__len__()) self.metrics = dict() mets = ['Loss', 'preLoss', 'Recall', 'NDCG'] for met in mets: self.metrics['Train' + met] = list() self.metrics['Test' + met] = list() def makePrint(self, name, ep, reses, save): ret = 'Epoch %d/%d, %s: ' % (ep, args.epoch, name) for metric in reses: val = reses[metric] ret += '%s = %.4f, ' % (metric, val) tem = name + metric if save and tem in self.metrics: self.metrics[tem].append(val) ret = ret[:-2] + ' ' return ret def run(self): self.prepareModel() log('Model Prepared') recallMax = 0 ndcgMax = 0 bestEpoch = 0 stloc = 0 log('Model Initialized') for ep in range(stloc, args.epoch): temperature = max(0.05, args.init_temperature * pow(args.temperature_decay, ep)) tstFlag = (ep % args.tstEpoch == 0) reses = self.trainEpoch(temperature) log(self.makePrint('Train', ep, reses, tstFlag)) if tstFlag: reses = self.testEpoch() if (reses['Recall'] > recallMax): recallMax = reses['Recall'] ndcgMax = reses['NDCG'] bestEpoch = ep log(self.makePrint('Test', ep, reses, tstFlag)) print() print('Best epoch : ', bestEpoch, ' , Recall : ', recallMax, ' , NDCG : ', ndcgMax) def prepareModel(self): self.model = Model().cuda() encoder = vgae_encoder().cuda() decoder = vgae_decoder().cuda() self.generator_1 = vgae(encoder, decoder).cuda() self.generator_2 = DenoisingNet(self.model.getGCN(), self.model.getEmbeds()).cuda() self.generator_2.set_fea_adj(args.user+args.item, deepcopy(self.handler.torchBiAdj).cuda()) self.opt = torch.optim.Adam(self.model.parameters(), lr=args.lr, weight_decay=0) self.opt_gen_1 = torch.optim.Adam(self.generator_1.parameters(), lr=args.lr, weight_decay=0) self.opt_gen_2 = torch.optim.Adam(filter(lambda p: p.requires_grad, self.generator_2.parameters()), lr=args.lr, weight_decay=0, eps=args.eps) def trainEpoch(self, temperature): trnLoader = self.handler.trnLoader trnLoader.dataset.negSampling() generate_loss_1, generate_loss_2, bpr_loss, im_loss, ib_loss, reg_loss = 0, 0, 0, 0, 0, 0 steps = trnLoader.dataset.__len__() // args.batch for i, tem in enumerate(trnLoader): data = deepcopy(self.handler.torchBiAdj).cuda() data1 = self.generator_generate(self.generator_1) self.opt.zero_grad() self.opt_gen_1.zero_grad() self.opt_gen_2.zero_grad() ancs, poss, negs = tem ancs = ancs.long().cuda() poss = poss.long().cuda() negs = negs.long().cuda() out1 = self.model.forward_graphcl(data1) out2 = self.model.forward_graphcl_(self.generator_2) loss = self.model.loss_graphcl(out1, out2, ancs, poss).mean() * args.ssl_reg im_loss += float(loss) loss.backward() # info bottleneck _out1 = self.model.forward_graphcl(data1) _out2 = self.model.forward_graphcl_(self.generator_2) loss_ib = self.model.loss_graphcl(_out1, out1.detach(), ancs, poss) + self.model.loss_graphcl(_out2, out2.detach(), ancs, poss) loss= loss_ib.mean() * args.ib_reg ib_loss += float(loss) loss.backward() # BPR usrEmbeds, itmEmbeds = self.model.forward_gcn(data) ancEmbeds = usrEmbeds[ancs] posEmbeds = itmEmbeds[poss] negEmbeds = itmEmbeds[negs] scoreDiff = pairPredict(ancEmbeds, posEmbeds, negEmbeds) bprLoss = - (scoreDiff).sigmoid().log().sum() / args.batch regLoss = calcRegLoss(self.model) * args.reg loss = bprLoss + regLoss bpr_loss += float(bprLoss) reg_loss += float(regLoss) loss.backward() loss_1 = self.generator_1(deepcopy(self.handler.torchBiAdj).cuda(), ancs, poss, negs) loss_2 = self.generator_2(ancs, poss, negs, temperature) loss = loss_1 + loss_2 generate_loss_1 += float(loss_1) generate_loss_2 += float(loss_2) loss.backward() self.opt.step() self.opt_gen_1.step() self.opt_gen_2.step() log('Step %d/%d: gen 1 : %.3f ; gen 2 : %.3f ; bpr : %.3f ; im : %.3f ; ib : %.3f ; reg : %.3f ' % ( i, steps, generate_loss_1, generate_loss_2, bpr_loss, im_loss, ib_loss, reg_loss, ), save=False, oneline=True) ret = dict() ret['Gen_1 Loss'] = generate_loss_1 / steps ret['Gen_2 Loss'] = generate_loss_2 / steps ret['BPR Loss'] = bpr_loss / steps ret['IM Loss'] = im_loss / steps ret['IB Loss'] = ib_loss / steps ret['Reg Loss'] = reg_loss / steps return ret def testEpoch(self): tstLoader = self.handler.tstLoader epRecall, epNdcg = [0] * 2 i = 0 num = tstLoader.dataset.__len__() steps = num // args.tstBat for usr, trnMask in tstLoader: i += 1 usr = usr.long().cuda() trnMask = trnMask.cuda() usrEmbeds, itmEmbeds = self.model.forward_gcn(self.handler.torchBiAdj) allPreds = torch.mm(usrEmbeds[usr], torch.transpose(itmEmbeds, 1, 0)) * (1 - trnMask) - trnMask * 1e8 _, topLocs = torch.topk(allPreds, args.topk) recall, ndcg = self.calcRes(topLocs.cpu().numpy(), self.handler.tstLoader.dataset.tstLocs, usr) epRecall += recall epNdcg += ndcg log('Steps %d/%d: recall = %.2f, ndcg = %.2f ' % (i, steps, recall, ndcg), save=False, oneline=True) ret = dict() ret['Recall'] = epRecall / num ret['NDCG'] = epNdcg / num return ret def calcRes(self, topLocs, tstLocs, batIds): assert topLocs.shape[0] == len(batIds) allRecall = allNdcg = 0 for i in range(len(batIds)): temTopLocs = list(topLocs[i]) temTstLocs = tstLocs[batIds[i]] tstNum = len(temTstLocs) maxDcg = np.sum([np.reciprocal(np.log2(loc + 2)) for loc in range(min(tstNum, args.topk))]) recall = dcg = 0 for val in temTstLocs: if val in temTopLocs: recall += 1 dcg += np.reciprocal(np.log2(temTopLocs.index(val) + 2)) recall = recall / tstNum ndcg = dcg / maxDcg allRecall += recall allNdcg += ndcg return allRecall, allNdcg def generator_generate(self, generator): edge_index = [] edge_index.append([]) edge_index.append([]) adj = deepcopy(self.handler.torchBiAdj) idxs = adj._indices() with torch.no_grad(): view = generator.generate(self.handler.torchBiAdj, idxs, adj) return view if __name__ == '__main__': with torch.cuda.device(args.gpu): logger.saveDefault = True log('Start') handler = DataHandler() handler.LoadData() log('Load Data') coach = Coach(handler) coach.run()
6,846
29.9819
143
py
AdaptiveGCL
AdaptiveGCL-main/Model.py
from torch import nn import torch.nn.functional as F import torch from Params import args from copy import deepcopy import numpy as np import math import scipy.sparse as sp from Utils.Utils import contrastLoss, calcRegLoss, pairPredict import time import torch_sparse init = nn.init.xavier_uniform_ class Model(nn.Module): def __init__(self): super(Model, self).__init__() self.uEmbeds = nn.Parameter(init(torch.empty(args.user, args.latdim))) self.iEmbeds = nn.Parameter(init(torch.empty(args.item, args.latdim))) self.gcnLayers = nn.Sequential(*[GCNLayer() for i in range(args.gnn_layer)]) def forward_gcn(self, adj): iniEmbeds = torch.concat([self.uEmbeds, self.iEmbeds], axis=0) embedsLst = [iniEmbeds] for gcn in self.gcnLayers: embeds = gcn(adj, embedsLst[-1]) embedsLst.append(embeds) mainEmbeds = sum(embedsLst) return mainEmbeds[:args.user], mainEmbeds[args.user:] def forward_graphcl(self, adj): iniEmbeds = torch.concat([self.uEmbeds, self.iEmbeds], axis=0) embedsLst = [iniEmbeds] for gcn in self.gcnLayers: embeds = gcn(adj, embedsLst[-1]) embedsLst.append(embeds) mainEmbeds = sum(embedsLst) return mainEmbeds def forward_graphcl_(self, generator): iniEmbeds = torch.concat([self.uEmbeds, self.iEmbeds], axis=0) embedsLst = [iniEmbeds] count = 0 for gcn in self.gcnLayers: with torch.no_grad(): adj = generator.generate(x=embedsLst[-1], layer=count) embeds = gcn(adj, embedsLst[-1]) embedsLst.append(embeds) count += 1 mainEmbeds = sum(embedsLst) return mainEmbeds def loss_graphcl(self, x1, x2, users, items): T = args.temp user_embeddings1, item_embeddings1 = torch.split(x1, [args.user, args.item], dim=0) user_embeddings2, item_embeddings2 = torch.split(x2, [args.user, args.item], dim=0) user_embeddings1 = F.normalize(user_embeddings1, dim=1) item_embeddings1 = F.normalize(item_embeddings1, dim=1) user_embeddings2 = F.normalize(user_embeddings2, dim=1) item_embeddings2 = F.normalize(item_embeddings2, dim=1) user_embs1 = F.embedding(users, user_embeddings1) item_embs1 = F.embedding(items, item_embeddings1) user_embs2 = F.embedding(users, user_embeddings2) item_embs2 = F.embedding(items, item_embeddings2) all_embs1 = torch.cat([user_embs1, item_embs1], dim=0) all_embs2 = torch.cat([user_embs2, item_embs2], dim=0) all_embs1_abs = all_embs1.norm(dim=1) all_embs2_abs = all_embs2.norm(dim=1) sim_matrix = torch.einsum('ik,jk->ij', all_embs1, all_embs2) / torch.einsum('i,j->ij', all_embs1_abs, all_embs2_abs) sim_matrix = torch.exp(sim_matrix / T) pos_sim = sim_matrix[np.arange(all_embs1.shape[0]), np.arange(all_embs1.shape[0])] loss = pos_sim / (sim_matrix.sum(dim=1) - pos_sim) loss = - torch.log(loss) return loss def getEmbeds(self): self.unfreeze(self.gcnLayers) return torch.concat([self.uEmbeds, self.iEmbeds], axis=0) def unfreeze(self, layer): for child in layer.children(): for param in child.parameters(): param.requires_grad = True def getGCN(self): return self.gcnLayers class GCNLayer(nn.Module): def __init__(self): super(GCNLayer, self).__init__() def forward(self, adj, embeds, flag=True): if (flag): return torch.spmm(adj, embeds) else: return torch_sparse.spmm(adj.indices(), adj.values(), adj.shape[0], adj.shape[1], embeds) class vgae_encoder(Model): def __init__(self): super(vgae_encoder, self).__init__() hidden = args.latdim self.encoder_mean = nn.Sequential(nn.Linear(hidden, hidden), nn.ReLU(inplace=True), nn.Linear(hidden, hidden)) self.encoder_std = nn.Sequential(nn.Linear(hidden, hidden), nn.ReLU(inplace=True), nn.Linear(hidden, hidden), nn.Softplus()) def forward(self, adj): x = self.forward_graphcl(adj) x_mean = self.encoder_mean(x) x_std = self.encoder_std(x) gaussian_noise = torch.randn(x_mean.shape).cuda() x = gaussian_noise * x_std + x_mean return x, x_mean, x_std class vgae_decoder(nn.Module): def __init__(self, hidden=args.latdim): super(vgae_decoder, self).__init__() self.decoder = nn.Sequential(nn.ReLU(inplace=True), nn.Linear(hidden, hidden), nn.ReLU(inplace=True), nn.Linear(hidden, 1)) self.sigmoid = nn.Sigmoid() self.bceloss = nn.BCELoss(reduction='none') def forward(self, x, x_mean, x_std, users, items, neg_items, encoder): x_user, x_item = torch.split(x, [args.user, args.item], dim=0) edge_pos_pred = self.sigmoid(self.decoder(x_user[users] * x_item[items])) edge_neg_pred = self.sigmoid(self.decoder(x_user[users] * x_item[neg_items])) loss_edge_pos = self.bceloss( edge_pos_pred, torch.ones(edge_pos_pred.shape).cuda() ) loss_edge_neg = self.bceloss( edge_neg_pred, torch.zeros(edge_neg_pred.shape).cuda() ) loss_rec = loss_edge_pos + loss_edge_neg kl_divergence = - 0.5 * (1 + 2 * torch.log(x_std) - x_mean**2 - x_std**2).sum(dim=1) ancEmbeds = x_user[users] posEmbeds = x_item[items] negEmbeds = x_item[neg_items] scoreDiff = pairPredict(ancEmbeds, posEmbeds, negEmbeds) bprLoss = - (scoreDiff).sigmoid().log().sum() / args.batch regLoss = calcRegLoss(encoder) * args.reg beta = 0.1 loss = (loss_rec + beta * kl_divergence.mean() + bprLoss + regLoss).mean() return loss class vgae(nn.Module): def __init__(self, encoder, decoder): super(vgae, self).__init__() self.encoder = encoder self.decoder = decoder def forward(self, data, users, items, neg_items): x, x_mean, x_std = self.encoder(data) loss = self.decoder(x, x_mean, x_std, users, items, neg_items, self.encoder) return loss def generate(self, data, edge_index, adj): x, _, _ = self.encoder(data) edge_pred = self.decoder.sigmoid(self.decoder.decoder(x[edge_index[0]] * x[edge_index[1]])) vals = adj._values() idxs = adj._indices() edgeNum = vals.size() edge_pred = edge_pred[:, 0] mask = ((edge_pred + 0.5).floor()).type(torch.bool) newVals = vals[mask] newVals = newVals / (newVals.shape[0] / edgeNum[0]) newIdxs = idxs[:, mask] return torch.sparse.FloatTensor(newIdxs, newVals, adj.shape) class DenoisingNet(nn.Module): def __init__(self, gcnLayers, features): super(DenoisingNet, self).__init__() self.features = features self.gcnLayers = gcnLayers self.edge_weights = [] self.nblayers = [] self.selflayers = [] self.attentions = [] self.attentions.append([]) self.attentions.append([]) hidden = args.latdim self.nblayers_0 = nn.Sequential(nn.Linear(hidden, hidden), nn.ReLU(inplace=True)) self.nblayers_1 = nn.Sequential(nn.Linear(hidden, hidden), nn.ReLU(inplace=True)) self.selflayers_0 = nn.Sequential(nn.Linear(hidden, hidden), nn.ReLU(inplace=True)) self.selflayers_1 = nn.Sequential(nn.Linear(hidden, hidden), nn.ReLU(inplace=True)) self.attentions_0 = nn.Sequential(nn.Linear( 2 * hidden, 1)) self.attentions_1 = nn.Sequential(nn.Linear( 2 * hidden, 1)) def freeze(self, layer): for child in layer.children(): for param in child.parameters(): param.requires_grad = False def get_attention(self, input1, input2, layer=0): if layer == 0: nb_layer = self.nblayers_0 selflayer = self.selflayers_0 if layer == 1: nb_layer = self.nblayers_1 selflayer = self.selflayers_1 input1 = nb_layer(input1) input2 = selflayer(input2) input10 = torch.concat([input1, input2], axis=1) if layer == 0: weight10 = self.attentions_0(input10) if layer == 1: weight10 = self.attentions_1(input10) return weight10 def hard_concrete_sample(self, log_alpha, beta=1.0, training=True): gamma = args.gamma zeta = args.zeta if training: debug_var = 1e-7 bias = 0.0 np_random = np.random.uniform(low=debug_var, high=1.0-debug_var, size=np.shape(log_alpha.cpu().detach().numpy())) random_noise = bias + torch.tensor(np_random) gate_inputs = torch.log(random_noise) - torch.log(1.0 - random_noise) gate_inputs = (gate_inputs.cuda() + log_alpha) / beta gate_inputs = torch.sigmoid(gate_inputs) else: gate_inputs = torch.sigmoid(log_alpha) stretched_values = gate_inputs * (zeta-gamma) +gamma cliped = torch.clamp(stretched_values, 0.0, 1.0) return cliped.float() def generate(self, x, layer=0): f1_features = x[self.row, :] f2_features = x[self.col, :] weight = self.get_attention(f1_features, f2_features, layer) mask = self.hard_concrete_sample(weight, training=False) mask = torch.squeeze(mask) adj = torch.sparse.FloatTensor(self.adj_mat._indices(), mask, self.adj_mat.shape) ind = deepcopy(adj._indices()) row = ind[0, :] col = ind[1, :] rowsum = torch.sparse.sum(adj, dim=-1).to_dense() d_inv_sqrt = torch.reshape(torch.pow(rowsum, -0.5), [-1]) d_inv_sqrt = torch.clamp(d_inv_sqrt, 0.0, 10.0) row_inv_sqrt = d_inv_sqrt[row] col_inv_sqrt = d_inv_sqrt[col] values = torch.mul(adj._values(), row_inv_sqrt) values = torch.mul(values, col_inv_sqrt) support = torch.sparse.FloatTensor(adj._indices(), values, adj.shape) return support def l0_norm(self, log_alpha, beta): gamma = args.gamma zeta = args.zeta gamma = torch.tensor(gamma) zeta = torch.tensor(zeta) reg_per_weight = torch.sigmoid(log_alpha - beta * torch.log(-gamma/zeta)) return torch.mean(reg_per_weight) def set_fea_adj(self, nodes, adj): self.node_size = nodes self.adj_mat = adj ind = deepcopy(adj._indices()) self.row = ind[0, :] self.col = ind[1, :] def call(self, inputs, training=None): if training: temperature = inputs else: temperature = 1.0 self.maskes = [] x = self.features.detach() layer_index = 0 embedsLst = [self.features.detach()] for layer in self.gcnLayers: xs = [] f1_features = x[self.row, :] f2_features = x[self.col, :] weight = self.get_attention(f1_features, f2_features, layer=layer_index) mask = self.hard_concrete_sample(weight, temperature, training) self.edge_weights.append(weight) self.maskes.append(mask) mask = torch.squeeze(mask) adj = torch.sparse.FloatTensor(self.adj_mat._indices(), mask, self.adj_mat.shape).coalesce() ind = deepcopy(adj._indices()) row = ind[0, :] col = ind[1, :] rowsum = torch.sparse.sum(adj, dim=-1).to_dense() + 1e-6 d_inv_sqrt = torch.reshape(torch.pow(rowsum, -0.5), [-1]) d_inv_sqrt = torch.clamp(d_inv_sqrt, 0.0, 10.0) row_inv_sqrt = d_inv_sqrt[row] col_inv_sqrt = d_inv_sqrt[col] values = torch.mul(adj.values(), row_inv_sqrt) values = torch.mul(values, col_inv_sqrt) support = torch.sparse.FloatTensor(adj._indices(), values, adj.shape).coalesce() nextx = layer(support, x, False) xs.append(nextx) x = xs[0] embedsLst.append(x) layer_index += 1 return sum(embedsLst) def lossl0(self, temperature): l0_loss = torch.zeros([]).cuda() for weight in self.edge_weights: l0_loss += self.l0_norm(weight, temperature) self.edge_weights = [] return l0_loss def forward(self, users, items, neg_items, temperature): self.freeze(self.gcnLayers) x = self.call(temperature, True) x_user, x_item = torch.split(x, [args.user, args.item], dim=0) ancEmbeds = x_user[users] posEmbeds = x_item[items] negEmbeds = x_item[neg_items] scoreDiff = pairPredict(ancEmbeds, posEmbeds, negEmbeds) bprLoss = - (scoreDiff).sigmoid().log().sum() / args.batch regLoss = calcRegLoss(self) * args.reg lossl0 = self.lossl0(temperature) * args.lambda0 return bprLoss + regLoss + lossl0
11,377
29.180371
126
py
AdaptiveGCL
AdaptiveGCL-main/Utils/Utils.py
import torch as t import torch.nn.functional as F def innerProduct(usrEmbeds, itmEmbeds): return t.sum(usrEmbeds * itmEmbeds, dim=-1) def pairPredict(ancEmbeds, posEmbeds, negEmbeds): return innerProduct(ancEmbeds, posEmbeds) - innerProduct(ancEmbeds, negEmbeds) def calcRegLoss(model): ret = 0 for W in model.parameters(): ret += W.norm(2).square() return ret def contrastLoss(embeds1, embeds2, nodes, temp): embeds1 = F.normalize(embeds1, p=2) embeds2 = F.normalize(embeds2, p=2) pckEmbeds1 = embeds1[nodes] pckEmbeds2 = embeds2[nodes] nume = t.exp(t.sum(pckEmbeds1 * pckEmbeds2, dim=-1) / temp) deno = t.exp(pckEmbeds1 @ embeds2.T / temp).sum(-1) return -t.log(nume / deno)
694
29.217391
79
py
RG
RG-master/Image Classification/main.py
from __future__ import print_function import torch import torch.nn as nn import torch.optim as optim import torch.nn.functional as F import torch.backends.cudnn as cudnn from torch.autograd import Variable import torchvision import torchvision.transforms as transforms import os import argparse import random from resnet import * from utils import progress_bar parser = argparse.ArgumentParser(description='PyTorch CIFAR10 Training') parser.add_argument('--lr', default=0.1, type=float, help='learning rate') args = parser.parse_args() device = 'cuda' if torch.cuda.is_available() else 'cpu' best_acc = 0 start_epoch = 0 # Data print('==> Preparing data..') transform_train = 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)), ]) transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) trainset = torchvision.datasets.CIFAR10(root='./data', train=True, download=False, transform=transform_train) trainloader = torch.utils.data.DataLoader(trainset, batch_size=128, shuffle=True, num_workers=0) testset = torchvision.datasets.CIFAR10(root='./data', train=False, download=False, transform=transform_test) testloader = torch.utils.data.DataLoader(testset, batch_size=128, shuffle=False, num_workers=0) def poly(base_lr, epoch,max_iter=100,power=0.9): return base_lr*((1-float(epoch+1)/max_iter)**(power)) # Model print('==> Building model..') net = ResNet34() if device == 'cuda': net.cuda() cudnn.benchmark = True criterion = nn.CrossEntropyLoss().cuda() optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=0.5, weight_decay=0.0001) M_loss = 0 # Training def train(epoch): global M_loss print('\nEpoch: %d' % epoch) net.train() train_loss = 0 correct = 0 total = 0 for batch_idx, (inputs, targets) in enumerate(trainloader): inputs, targets = inputs.cuda(), targets.cuda() inputs, targets = Variable(inputs), Variable(targets) optimizer.zero_grad() outputs = net(inputs) loss = criterion(outputs, targets) loss_ = loss * random.random() # Here represents the Random Gradient loss_.backward() optimizer.step() train_loss += loss_.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() progress_bar(batch_idx, len(trainloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % (train_loss/(batch_idx+1), 100.*correct/total, correct, total)) def test(epoch): global best_acc net.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for batch_idx, (inputs, targets) in enumerate(testloader): inputs, targets = inputs.cuda(), targets.cuda() inputs, targets = Variable(inputs), Variable(targets) outputs = net(inputs) loss = criterion(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() progress_bar(batch_idx, len(testloader), 'Loss: %.3f | Acc: %.3f%% (%d/%d)' % (test_loss/(batch_idx+1), 100.*correct/total, correct, total)) acc = 100.*correct/total if acc > best_acc: print('Saving..') open('./random_gradient_accuracy.txt', 'a').write(str(epoch) + '_' + str(acc) + ',') best_acc = acc print('best_acc:', best_acc) def adjust_learning_rate(optimizer, epoch, net): lr = poly(0.1, epoch) # This is normal way to reduce the LR, you can replace it with CLR print('current lr: ', lr) for param_group in optimizer.param_groups: param_group['lr'] = lr if __name__ == '__main__': for epoch in range(0, 100): train(epoch) test(epoch) adjust_learning_rate(optimizer,epoch,net)
4,105
30.343511
109
py
RG
RG-master/Image Classification/resnet.py
'''ResNet in PyTorch. For Pre-activation ResNet, see 'preact_resnet.py'. Reference: [1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun Deep Residual Learning for Image Recognition. arXiv:1512.03385 ''' import torch import torch.nn as nn import torch.nn.functional as F from torch.autograd import Variable class BasicBlock(nn.Module): expansion = 1 def __init__(self, in_planes, planes, stride=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.shortcut = nn.Sequential() if stride != 1 or in_planes != self.expansion*planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(self.expansion*planes) ) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) out += self.shortcut(x) out = F.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, in_planes, planes, stride=1): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes) self.conv3 = nn.Conv2d(planes, self.expansion*planes, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(self.expansion*planes) self.shortcut = nn.Sequential() if stride != 1 or in_planes != self.expansion*planes: self.shortcut = nn.Sequential( nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(self.expansion*planes) ) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = F.relu(self.bn2(self.conv2(out))) out = self.bn3(self.conv3(out)) out += self.shortcut(x) out = F.relu(out) return out class ResNet(nn.Module): def __init__(self, block, num_blocks, num_classes=10): super(ResNet, self).__init__() self.in_planes = 64 self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False) self.bn1 = nn.BatchNorm2d(64) self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1) self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2) self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2) self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2) self.linear = nn.Linear(512*block.expansion, num_classes) def _make_layer(self, block, planes, num_blocks, stride): strides = [stride] + [1]*(num_blocks-1) layers = [] for stride in strides: layers.append(block(self.in_planes, planes, stride)) self.in_planes = planes * block.expansion return nn.Sequential(*layers) def forward(self, x): out = F.relu(self.bn1(self.conv1(x))) out = self.layer1(out) out = self.layer2(out) out = self.layer3(out) out = self.layer4(out) out = F.avg_pool2d(out, 4) out = out.view(out.size(0), -1) out = self.linear(out) return out def ResNet18(): return ResNet(BasicBlock, [2,2,2,2]) def ResNet34(): return ResNet(BasicBlock, [3,4,6,3]) def ResNet50(): return ResNet(Bottleneck, [3,4,6,3]) def ResNet101(): return ResNet(Bottleneck, [3,4,23,3]) def ResNet152(): return ResNet(Bottleneck, [3,8,36,3])
3,941
32.982759
102
py
RG
RG-master/pix2pix/pix2pix.py
import argparse import os import numpy as np import math import itertools import time import datetime import sys import random import torchvision.transforms as transforms from torchvision.utils import save_image from torch.utils.data import DataLoader from torchvision import datasets from torch.autograd import Variable from models import * from datasets import * import torch.nn as nn import torch.nn.functional as F import torch parser = argparse.ArgumentParser() parser.add_argument('--epoch', type=int, default=0, help='epoch to start training from') parser.add_argument('--n_epochs', type=int, default=200, help='number of epochs of training') parser.add_argument('--dataset_name', type=str, default="facades", help='name of the dataset') parser.add_argument('--batch_size', type=int, default=1, help='size of the batches') parser.add_argument('--lr', type=float, default=0.0002, help='adam: learning rate') parser.add_argument('--b1', type=float, default=0.5, help='adam: decay of first order momentum of gradient') parser.add_argument('--b2', type=float, default=0.999, help='adam: decay of first order momentum of gradient') parser.add_argument('--decay_epoch', type=int, default=100, help='epoch from which to start lr decay') parser.add_argument('--n_cpu', type=int, default=8, help='number of cpu threads to use during batch generation') parser.add_argument('--img_height', type=int, default=256, help='size of image height') parser.add_argument('--img_width', type=int, default=256, help='size of image width') parser.add_argument('--channels', type=int, default=3, help='number of image channels') parser.add_argument('--sample_interval', type=int, default=500, help='interval between sampling of images from generators') parser.add_argument('--checkpoint_interval', type=int, default=-1, help='interval between model checkpoints') opt = parser.parse_args() print(opt) os.makedirs('images/%s' % opt.dataset_name, exist_ok=True) os.makedirs('saved_models/%s' % opt.dataset_name, exist_ok=True) cuda = True if torch.cuda.is_available() else False # Loss functions criterion_GAN = torch.nn.MSELoss() criterion_pixelwise = torch.nn.L1Loss() # Loss weight of L1 pixel-wise loss between translated image and real image lambda_pixel = 100 # Calculate output of image discriminator (PatchGAN) patch = (1, opt.img_height//2**4, opt.img_width//2**4) # Initialize generator and discriminator generator = GeneratorUNet() discriminator = Discriminator() if cuda: generator = generator.cuda() discriminator = discriminator.cuda() criterion_GAN.cuda() criterion_pixelwise.cuda() if opt.epoch != 0: # Load pretrained models generator.load_state_dict(torch.load('saved_models/%s/generator_%d.pth' % (opt.dataset_name, opt.epoch))) discriminator.load_state_dict(torch.load('saved_models/%s/discriminator_%d.pth' % (opt.dataset_name, opt.epoch))) else: # Initialize weights generator.apply(weights_init_normal) discriminator.apply(weights_init_normal) # Optimizers optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2)) optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2)) # Configure dataloaders transforms_ = [ transforms.Resize((opt.img_height, opt.img_width), Image.BICUBIC), transforms.ToTensor(), transforms.Normalize((0.5,0.5,0.5), (0.5,0.5,0.5)) ] dataloader = DataLoader(ImageDataset("../../data/%s" % opt.dataset_name, transforms_=transforms_), batch_size=opt.batch_size, shuffle=True, num_workers=opt.n_cpu) val_dataloader = DataLoader(ImageDataset("../../data/%s" % opt.dataset_name, transforms_=transforms_, mode='val'), batch_size=10, shuffle=True, num_workers=1) # Tensor type Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor def sample_images(batches_done): """Saves a generated sample from the validation set""" imgs = next(iter(val_dataloader)) real_A = Variable(imgs['B'].type(Tensor)) real_B = Variable(imgs['A'].type(Tensor)) fake_B = generator(real_A) img_sample = torch.cat((real_A.data, fake_B.data, real_B.data), -2) save_image(img_sample, 'images/%s/%s.png' % (opt.dataset_name, batches_done), nrow=5, normalize=True) # ---------- # Training # ---------- prev_time = time.time() for epoch in range(opt.epoch, opt.n_epochs): for i, batch in enumerate(dataloader): # Model inputs real_A = Variable(batch['B'].type(Tensor)) real_B = Variable(batch['A'].type(Tensor)) # Adversarial ground truths valid = Variable(Tensor(np.ones((real_A.size(0), *patch))), requires_grad=False) fake = Variable(Tensor(np.zeros((real_A.size(0), *patch))), requires_grad=False) # ------------------ # Train Generators # ------------------ optimizer_G.zero_grad() # GAN loss fake_B = generator(real_A) pred_fake = discriminator(fake_B, real_A) loss_GAN = criterion_GAN(pred_fake, valid) # Pixel-wise loss loss_pixel = criterion_pixelwise(fake_B, real_B) # Total loss loss_G = loss_GAN + lambda_pixel * loss_pixel # Random Gradient loss_G = loss_G * random.random() loss_G.backward() optimizer_G.step() # --------------------- # Train Discriminator # --------------------- optimizer_D.zero_grad() # Real loss pred_real = discriminator(real_B, real_A) loss_real = criterion_GAN(pred_real, valid) # Fake loss pred_fake = discriminator(fake_B.detach(), real_A) loss_fake = criterion_GAN(pred_fake, fake) # Total loss loss_D = 0.5 * (loss_real + loss_fake) loss_D.backward() optimizer_D.step() # -------------- # Log Progress # -------------- # Determine approximate time left batches_done = epoch * len(dataloader) + i batches_left = opt.n_epochs * len(dataloader) - batches_done time_left = datetime.timedelta(seconds=batches_left * (time.time() - prev_time)) prev_time = time.time() # Print log sys.stdout.write("\r[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f, pixel: %f, adv: %f] ETA: %s" % (epoch, opt.n_epochs, i, len(dataloader), loss_D.item(), loss_G.item(), loss_pixel.item(), loss_GAN.item(), time_left)) # If at sample interval save image if batches_done % opt.sample_interval == 0: sample_images(batches_done) if opt.checkpoint_interval != -1 and epoch % opt.checkpoint_interval == 0: # Save model checkpoints torch.save(generator.state_dict(), 'saved_models/%s/generator_%d.pth' % (opt.dataset_name, epoch)) torch.save(discriminator.state_dict(), 'saved_models/%s/discriminator_%d.pth' % (opt.dataset_name, epoch))
7,224
36.827225
123
py
RG
RG-master/pix2pix/datasets.py
import glob import random import os import numpy as np from torch.utils.data import Dataset from PIL import Image import torchvision.transforms as transforms class ImageDataset(Dataset): def __init__(self, root, transforms_=None, mode='train'): self.transform = transforms.Compose(transforms_) self.files = sorted(glob.glob(os.path.join(root, mode) + '/*.*')) if mode == 'train': self.files.extend(sorted(glob.glob(os.path.join(root, 'test') + '/*.*'))) def __getitem__(self, index): img = Image.open(self.files[index % len(self.files)]) w, h = img.size img_A = img.crop((0, 0, w/2, h)) img_B = img.crop((w/2, 0, w, h)) if np.random.random() < 0.5: img_A = Image.fromarray(np.array(img_A)[:, ::-1, :], 'RGB') img_B = Image.fromarray(np.array(img_B)[:, ::-1, :], 'RGB') img_A = self.transform(img_A) img_B = self.transform(img_B) return {'A': img_A, 'B': img_B} def __len__(self): return len(self.files)
1,056
28.361111
85
py
RG
RG-master/pix2pix/models.py
import torch.nn as nn import torch.nn.functional as F import torch def weights_init_normal(m): classname = m.__class__.__name__ if classname.find('Conv') != -1: torch.nn.init.normal_(m.weight.data, 0.0, 0.02) elif classname.find('BatchNorm2d') != -1: torch.nn.init.normal_(m.weight.data, 1.0, 0.02) torch.nn.init.constant_(m.bias.data, 0.0) ############################## # U-NET ############################## class UNetDown(nn.Module): def __init__(self, in_size, out_size, normalize=True, dropout=0.0): super(UNetDown, self).__init__() layers = [nn.Conv2d(in_size, out_size, 4, 2, 1, bias=False)] if normalize: layers.append(nn.InstanceNorm2d(out_size)) layers.append(nn.LeakyReLU(0.2)) if dropout: layers.append(nn.Dropout(dropout)) self.model = nn.Sequential(*layers) def forward(self, x): return self.model(x) class UNetUp(nn.Module): def __init__(self, in_size, out_size, dropout=0.0): super(UNetUp, self).__init__() layers = [ nn.ConvTranspose2d(in_size, out_size, 4, 2, 1, bias=False), nn.InstanceNorm2d(out_size), nn.ReLU(inplace=True)] if dropout: layers.append(nn.Dropout(dropout)) self.model = nn.Sequential(*layers) def forward(self, x, skip_input): x = self.model(x) x = torch.cat((x, skip_input), 1) return x class GeneratorUNet(nn.Module): def __init__(self, in_channels=3, out_channels=3): super(GeneratorUNet, self).__init__() self.down1 = UNetDown(in_channels, 64, normalize=False) self.down2 = UNetDown(64, 128) self.down3 = UNetDown(128, 256) self.down4 = UNetDown(256, 512, dropout=0.5) self.down5 = UNetDown(512, 512, dropout=0.5) self.down6 = UNetDown(512, 512, dropout=0.5) self.down7 = UNetDown(512, 512, dropout=0.5) self.down8 = UNetDown(512, 512, normalize=False, dropout=0.5) self.up1 = UNetUp(512, 512, dropout=0.5) self.up2 = UNetUp(1024, 512, dropout=0.5) self.up3 = UNetUp(1024, 512, dropout=0.5) self.up4 = UNetUp(1024, 512, dropout=0.5) self.up5 = UNetUp(1024, 256) self.up6 = UNetUp(512, 128) self.up7 = UNetUp(256, 64) self.final = nn.Sequential( nn.Upsample(scale_factor=2), nn.ZeroPad2d((1, 0, 1, 0)), nn.Conv2d(128, out_channels, 4, padding=1), nn.Tanh() ) def forward(self, x): # U-Net generator with skip connections from encoder to decoder d1 = self.down1(x) d2 = self.down2(d1) d3 = self.down3(d2) d4 = self.down4(d3) d5 = self.down5(d4) d6 = self.down6(d5) d7 = self.down7(d6) d8 = self.down8(d7) u1 = self.up1(d8, d7) u2 = self.up2(u1, d6) u3 = self.up3(u2, d5) u4 = self.up4(u3, d4) u5 = self.up5(u4, d3) u6 = self.up6(u5, d2) u7 = self.up7(u6, d1) return self.final(u7) ############################## # Discriminator ############################## class Discriminator(nn.Module): def __init__(self, in_channels=3): super(Discriminator, self).__init__() def discriminator_block(in_filters, out_filters, normalization=True): """Returns downsampling layers of each discriminator block""" layers = [nn.Conv2d(in_filters, out_filters, 4, stride=2, padding=1)] if normalization: layers.append(nn.InstanceNorm2d(out_filters)) layers.append(nn.LeakyReLU(0.2, inplace=True)) return layers self.model = nn.Sequential( *discriminator_block(in_channels*2, 64, normalization=False), *discriminator_block(64, 128), *discriminator_block(128, 256), *discriminator_block(256, 512), nn.ZeroPad2d((1, 0, 1, 0)), nn.Conv2d(512, 1, 4, padding=1, bias=False) ) def forward(self, img_A, img_B): # Concatenate image and condition image by channels to produce input img_input = torch.cat((img_A, img_B), 1) return self.model(img_input)
4,289
32.515625
81
py
RG
RG-master/Semantic Segmentation/model.py
import torch.nn as nn import math import torch.utils.model_zoo as model_zoo import torch import numpy as np affine_par = True import torch.nn.functional as F def outS(i): i = int(i) i = (i+1)/2 i = int(np.ceil((i+1)/2.0)) i = (i+1)/2 return i def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, affine = affine_par) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, affine = affine_par) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, stride=stride, bias=False) # change self.bn1 = nn.BatchNorm2d(planes,affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False padding = dilation self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, # change padding=padding, bias=False, dilation = dilation) self.bn2 = nn.BatchNorm2d(planes,affine = affine_par) for i in self.bn2.parameters(): i.requires_grad = False self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * 4, affine = affine_par) for i in self.bn3.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Classifier_Module(nn.Module): def __init__(self, dilation_series, padding_series, num_classes): super(Classifier_Module, self).__init__() self.conv2d_list = nn.ModuleList() for dilation, padding in zip(dilation_series, padding_series): self.conv2d_list.append(nn.Conv2d(2048, num_classes, kernel_size=3, stride=1, padding=padding, dilation=dilation, bias = True)) for m in self.conv2d_list: m.weight.data.normal_(0, 0.01) def forward(self, x): out = self.conv2d_list[0](x) for i in range(len(self.conv2d_list)-1): out += self.conv2d_list[i+1](x) return out class Residual_Covolution(nn.Module): def __init__(self, icol, ocol, num_classes): super(Residual_Covolution, self).__init__() self.conv1 = nn.Conv2d(icol, ocol, kernel_size=3, stride=1, padding=12, dilation=12, bias=True) self.conv2 = nn.Conv2d(ocol, num_classes, kernel_size=3, stride=1, padding=12, dilation=12, bias=True) self.conv3 = nn.Conv2d(num_classes, ocol, kernel_size=1, stride=1, padding=0, dilation=1, bias=True) self.conv4 = nn.Conv2d(ocol, icol, kernel_size=1, stride=1, padding=0, dilation=1, bias=True) self.relu = nn.ReLU(inplace=True) def forward(self, x): dow1 = self.conv1(x) dow1 = self.relu(dow1) seg = self.conv2(dow1) inc1 = self.conv3(seg) add1 = dow1 + self.relu(inc1) inc2 = self.conv4(add1) out = x + self.relu(inc2) return out, seg class Residual_Refinement_Module(nn.Module): def __init__(self, num_classes): super(Residual_Refinement_Module, self).__init__() self.RC1 = Residual_Covolution(2048, 512, num_classes) self.RC2 = Residual_Covolution(2048, 512, num_classes) def forward(self, x): x, seg1 = self.RC1(x) _, seg2 = self.RC2(x) return [seg1, seg1+seg2] class ResNet_Refine(nn.Module): def __init__(self, block, layers, num_classes): self.inplanes = 64 super(ResNet_Refine, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=True) # change self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=1, dilation=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation=4) self.layer5 = Residual_Refinement_Module(num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() # for i in m.parameters(): # i.requires_grad = False def _make_layer(self, block, planes, blocks, stride=1, dilation=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion or dilation == 2 or dilation == 4: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion,affine = affine_par)) for i in downsample._modules['1'].parameters(): i.requires_grad = False layers = [] layers.append(block(self.inplanes, planes, stride,dilation=dilation, downsample=downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, dilation=dilation)) return nn.Sequential(*layers) def forward(self, x): x_size = x.size() x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.layer5(x) interp = F.upsample(x, x_size[2:], mode='bilinear') return interp class ResNet(nn.Module): def __init__(self, block, layers, num_classes): self.inplanes = 64 super(ResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, affine = affine_par) for i in self.bn1.parameters(): i.requires_grad = False self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, ceil_mode=True) # change self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=1, dilation=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation=4) self.layer5 = self._make_pred_layer(Classifier_Module, [6,12,18,24],[6,12,18,24],num_classes) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, 0.01) elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() # for i in m.parameters(): # i.requires_grad = False def _make_layer(self, block, planes, blocks, stride=1, dilation=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion or dilation == 2 or dilation == 4: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion,affine = affine_par)) for i in downsample._modules['1'].parameters(): i.requires_grad = False layers = [] layers.append(block(self.inplanes, planes, stride,dilation=dilation, downsample=downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, dilation=dilation)) return nn.Sequential(*layers) def _make_pred_layer(self,block, dilation_series, padding_series,num_classes): return block(dilation_series,padding_series,num_classes) def forward(self, x): x_size = x.size() x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.layer5(x) interp = F.upsample(x, x_size[2:], mode='bilinear') return interp class MS_Deeplab(nn.Module): def __init__(self,block,num_classes): super(MS_Deeplab,self).__init__() self.Scale = ResNet(block,[3, 4, 23, 3],num_classes) #changed to fix #4 def forward(self,x): output = self.Scale(x) # for original scale output_size = output.size()[2] input_size = x.size()[2] self.interp1 = nn.Upsample(size=(int(input_size*0.75)+1, int(input_size*0.75)+1), mode='bilinear') self.interp2 = nn.Upsample(size=(int(input_size*0.5)+1, int(input_size*0.5)+1), mode='bilinear') self.interp3 = nn.Upsample(size=(output_size, output_size), mode='bilinear') x75 = self.interp1(x) output75 = self.interp3(self.Scale(x75)) # for 0.75x scale x5 = self.interp2(x) output5 = self.interp3(self.Scale(x5)) # for 0.5x scale out_max = torch.max(torch.max(output, output75), output5) return [output, output75, output5, out_max] def Res_Ms_Deeplab(num_classes=21): model = MS_Deeplab(Bottleneck, num_classes) return model def Res_Deeplab(num_classes=21, is_refine=False): if is_refine: model = ResNet_Refine(Bottleneck,[3, 4, 23, 3], num_classes) else: model = ResNet(Bottleneck,[3, 4, 23, 3], num_classes) return model
11,339
36.549669
139
py
RG
RG-master/Semantic Segmentation/train.py
import datetime import os import random import time from math import sqrt import torchvision.transforms as standard_transforms import torchvision.utils as vutils # from tensorboard import SummaryWriter from torch import optim from torch.autograd import Variable from torch.backends import cudnn from torch.optim.lr_scheduler import ReduceLROnPlateau from torch.utils.data import DataLoader import utils.joint_transforms as joint_transforms import utils.transforms as extended_transforms from datasets import voc from models import * from utils import check_mkdir, evaluate, AverageMeter, CrossEntropyLoss2d from tqdm import tqdm as tqdm cudnn.benchmark = True from torchvision.transforms import * from torchvision.transforms import ToTensor, ToPILImage ckpt_path = './ckpt' exp_name = 'RSPPNET' args = { 'epoch_num': 200, 'lr': 0.0001, 'weight_decay': 0.0005, 'momentum': 0.9, 'lr_patience': 100, # large patience denotes fixed lr 'snapshot': '', # empty string denotes learning from scratch 'print_freq': 1, 'val_save_to_img_file': False, 'val_img_sample_rate': 0.1 # randomly sample some validation results to display } def lr_poly(base_lr, iter,max_iter=200,power=0.9): return base_lr*((1-float(iter)/max_iter)**(power)) def adjust_learning_rate(optimizer, i_iter, net, train_args): """Sets the learning rate to the initial LR divided by 5 at 60th, 120th and 160th epochs""" lr = lr_poly(0.0001, i_iter) print('current lr:', lr) # optimizer.step() # optimizer = optim.RMSprop(net.parameters(), lr=lr, alpha=0.99, eps=1e-08, weight_decay=train_args['weight_decay'], momentum=0.9, centered=False) # optimizer = optim.Adam(net.parameters(), lr=lr, betas=(0.9, 0.999), eps=1e-08, weight_decay=train_args['weight_decay']) optimizer = optim.SGD(net.parameters(),lr=lr, momentum=train_args['momentum'],weight_decay=train_args['weight_decay']) # optimizer = optim.SGD([{'params': get_1x_lr_params_NOscale(net), 'lr': lr }, {'params': get_10x_lr_params(net), 'lr': 10*lr} ], lr = lr, momentum = train_args['momentum'],weight_decay = train_args['weight_decay']) # optimizer.zero_grad() # optimizer.param_groups[1]['lr'] = lr * 10 max_label = 20 def get_iou(pred,gt): if pred.shape!= gt.shape: print('pred shape',pred.shape, 'gt shape', gt.shape) assert(pred.shape == gt.shape) gt = gt.astype(np.float32) pred = pred.astype(np.float32) count = np.zeros((max_label+1,)) for j in range(max_label+1): x = np.where(pred==j) p_idx_j = set(zip(x[0].tolist(),x[1].tolist())) x = np.where(gt==j) GT_idx_j = set(zip(x[0].tolist(),x[1].tolist())) #pdb.set_trace() n_jj = set.intersection(p_idx_j,GT_idx_j) u_jj = set.union(p_idx_j,GT_idx_j) if len(GT_idx_j)!=0: count[j] = float(len(n_jj))/float(len(u_jj)) result_class = count Aiou = np.sum(result_class[:])/float(len(np.unique(gt))) return Aiou def main(train_args): net = PSPNet(num_classes=voc.num_classes).cuda() if len(train_args['snapshot']) == 0: curr_epoch = 1 train_args['best_record'] = {'epoch': 0, 'val_loss': 1e10, 'acc': 0, 'acc_cls': 0, 'mean_iu': 0, 'fwavacc': 0} else: print('training resumes from ' + train_args['snapshot']) net.load_state_dict(torch.load(os.path.join(ckpt_path, exp_name, train_args['snapshot']))) split_snapshot = train_args['snapshot'].split('_') curr_epoch = int(split_snapshot[1]) + 1 train_args['best_record'] = {'epoch': int(split_snapshot[1]), 'val_loss': float(split_snapshot[3]), 'acc': float(split_snapshot[5]), 'acc_cls': float(split_snapshot[7]), 'mean_iu': float(split_snapshot[9]), 'fwavacc': float(split_snapshot[11])} mean_std = ([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) input_transform = standard_transforms.Compose([ ToTensor(), Normalize([.485, .456, .406], [.229, .224, .225]), ]) joint_transform = joint_transforms.Compose([ joint_transforms.CenterCrop(224), # joint_transforms.Scale(2), joint_transforms.RandomHorizontallyFlip(), ]) target_transform = standard_transforms.Compose([ extended_transforms.MaskToTensor(), ]) restore_transform = standard_transforms.Compose([ extended_transforms.DeNormalize(*mean_std), standard_transforms.ToPILImage(), ]) visualize = standard_transforms.Compose([ standard_transforms.Scale(400), standard_transforms.CenterCrop(400), standard_transforms.ToTensor() ]) val_input_transform = standard_transforms.Compose([ CenterCrop(224), ToTensor(), Normalize([.485, .456, .406], [.229, .224, .225]), ]) val_target_transform = standard_transforms.Compose([ CenterCrop(224), extended_transforms.MaskToTensor(), ]) train_set = voc.VOC('train', transform=input_transform, target_transform=target_transform, joint_transform=joint_transform) train_loader = DataLoader(train_set, batch_size=4, num_workers=4, shuffle=True) val_set = voc.VOC('val', transform=val_input_transform, target_transform=val_target_transform) val_loader = DataLoader(val_set, batch_size=4, num_workers=4, shuffle=False) # criterion = CrossEntropyLoss2d(size_average=True, ignore_index=voc.ignore_label).cuda() criterion = torch.nn.CrossEntropyLoss(ignore_index=voc.ignore_label).cuda() optimizer = optim.SGD(net.parameters(),lr=train_args['lr'], momentum=train_args['momentum'],weight_decay=train_args['weight_decay']) check_mkdir(ckpt_path) check_mkdir(os.path.join(ckpt_path, exp_name)) # open(os.path.join(ckpt_path, exp_name, 'loss_001_aux_SGD_momentum_95_random_lr_001.txt'), 'w').write(str(train_args) + '\n\n') for epoch in range(curr_epoch, train_args['epoch_num'] + 1): # adjust_learning_rate(optimizer,epoch,net,train_args) train(train_loader, net, criterion, optimizer, epoch, train_args) validate(val_loader, net, criterion, optimizer, epoch, train_args, restore_transform, visualize) adjust_learning_rate(optimizer,epoch,net,train_args) # scheduler.step(val_loss) def train(train_loader, net, criterion, optimizer, epoch, train_args): # interp = nn.Upsample(size=256, mode='bilinear') net.train() train_loss = AverageMeter() curr_iter = (epoch - 1) * len(train_loader) for i, data in enumerate(train_loader): inputs, labels = data assert inputs.size()[2:] == labels.size()[1:] N = inputs.size(0) inputs = Variable(inputs).cuda() labels = Variable(labels).cuda() random_number = random.random() if random_number > 0.5: optimizer.zero_grad() outputs,aux_logits = net(inputs) assert outputs.size()[2:] == labels.size()[1:] assert outputs.size()[1] == voc.num_classes loss_1 = criterion(outputs, labels) loss_2 = criterion(aux_logits, labels) loss = (loss_1 + 0.4*loss_2)*random.random() loss.backward() optimizer.step() train_loss.update(loss.data[0], N) else: optimizer.zero_grad() outputs,aux_logits = net(inputs) assert outputs.size()[2:] == labels.size()[1:] assert outputs.size()[1] == voc.num_classes loss_1 = criterion(outputs, labels) loss_2 = criterion(aux_logits, labels) loss = loss_1 + 0.4*loss_2 loss.backward() optimizer.step() train_loss.update(loss.data[0], N) curr_iter += 1 # writer.add_scalar('train_loss', train_loss.avg, curr_iter) if i % train_args['print_freq'] == 0: print('[epoch %d], [iter %d / %d], [train loss %.5f],[N: %d]' % ( epoch, i + 1, len(train_loader), train_loss.avg, N # , loss_1.data[0], loss_2.data[0],[loss %.3f],[loss2 %.3f] )) def validate(val_loader, net, criterion, optimizer, epoch, train_args, restore, visualize): net.eval() global best_acc val_loss = AverageMeter() inputs_all, gts_all, predictions_all = [], [], [] for vi, data in tqdm(enumerate(val_loader)): inputs, gts = data N = inputs.size(0) inputs = Variable(inputs, volatile=True).cuda() gts = Variable(gts, volatile=True).cuda() outputs = net(inputs) # interp = nn.Upsample(size=256, mode='bilinear') # outputs = interp(net(inputs)) predictions = outputs.data.max(1)[1].squeeze_(1).squeeze_(0).cpu().numpy() val_loss.update(criterion(outputs, gts).data[0], N) # if random.random() > train_args['val_img_sample_rate']: # inputs_all.append(None) # else: # inputs_all.append(inputs.data.squeeze_(0).cpu()) gts_all.append(gts.data.squeeze_(0).cpu().numpy()) predictions_all.append(predictions) # IOU.append(get_iou(outputs,gts)) acc, acc_cls, mean_iu, fwavacc = evaluate(predictions_all, gts_all, voc.num_classes) if mean_iu > train_args['best_record']['mean_iu']: train_args['best_record']['val_loss'] = val_loss.avg train_args['best_record']['epoch'] = epoch train_args['best_record']['acc'] = acc train_args['best_record']['acc_cls'] = acc_cls train_args['best_record']['mean_iu'] = mean_iu train_args['best_record']['fwavacc'] = fwavacc snapshot_name = 'epoch_%d_loss_%.5f_acc_%.5f_acc-cls_%.5f_mean-iu_%.5f_fwavacc_%.5f' % ( epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc ) open(os.path.join(ckpt_path, exp_name, 'loss_0001_dilation_aux_SGD_momentum_090_PSPNet_L3.txt'), 'a').write(str(epoch) + '_' + str(mean_iu) + ',') # torch.save(net.state_dict(), os.path.join(ckpt_path, exp_name, snapshot_name + '.pth')) print('--------------------------------------------------------------------') print('[epoch %d], [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f]' % ( epoch, val_loss.avg, acc, acc_cls, mean_iu, fwavacc)) print('best record: [val loss %.5f], [acc %.5f], [acc_cls %.5f], [mean_iu %.5f], [fwavacc %.5f], [epoch %d]' % ( train_args['best_record']['val_loss'], train_args['best_record']['acc'], train_args['best_record']['acc_cls'], train_args['best_record']['mean_iu'], train_args['best_record']['fwavacc'], train_args['best_record']['epoch'])) print('--------------------------------------------------------------------') if __name__ == '__main__': main(args)
9,955
38.19685
219
py
CD-Flow
CD-Flow-main/main.py
import torch from trainnet import trainNet import pandas as pd import argparse def parse_config(): parser = argparse.ArgumentParser() parser.add_argument("--seed", type=int, default=100) parser.add_argument("--resume_path", type=str, default=None) parser.add_argument("--learning_rate", type=float, default=1e-5) parser.add_argument("--scheduler_step", type=int, default=5) parser.add_argument("--scheduler_gamma", type=float, default=0.5) parser.add_argument("--batch_size_train", type=int, default=4) parser.add_argument("--batch_size_test", type=int, default=4) parser.add_argument("--n_epochs", type=int, default=50) parser.add_argument("--training_datadir", type=str, default='') parser.add_argument("--colorspace", type=str, default='rgb') parser.add_argument("--trainpath1", type=str, default='trainnet.py') parser.add_argument("--trainpath2", type=str, default='main.py') parser.add_argument("--trainpath3", type=str, default='model.py') parser.add_argument("--trainpath4", type=str, default='DataLoader.py') parser.add_argument("--work_path", type=str, default='work_dir') parser.add_argument("--datapath", type=str, default='data') parser.add_argument("--trainset", type=str, default='train.csv') parser.add_argument("--valset", type=str, default='val.csv') parser.add_argument("--testset", type=str, default='test.csv') parser.add_argument("--test_aligned_path", type=str, default=None) parser.add_argument("--test_notaligned_path", type=str, default=None) return parser.parse_args() if __name__ == '__main__': config = parse_config() path = config.datapath modelprediction = pd.DataFrame(columns=['no']) modelprediction_aligned = pd.DataFrame(columns=['no']) modelprediction_notaligned = pd.DataFrame(columns=['no']) work_path = config.work_path trainpath = config.trainset valpath = config.valset testpath = config.testset performance = pd.DataFrame(columns=['stress', 'plcc', 'srcc', 'stress_aligned', 'plcc_aligned', 'srcc_aligned', 'stress_notaligned', 'plcc_notaligned', 'srcc_notaligned']) torch.cuda.empty_cache() i = 0 config.datapath = path+'/{}.csv'.format(i+1) config.work_path = work_path+'/{}'.format(i+1) config.trainset = path+'/{}/'.format(i+1)+trainpath config.valset = path+'/{}/'.format(i+1)+valpath config.testset = path+'/{}/'.format(i+1)+testpath config.test_aligned_path = path+'/{}/test_aligned.csv'.format(i+1) config.test_notaligned_path = path+'/{}/test_notaligned.csv'.format(i+1) dist1, y_true1, stress1, cc_v1, srocc_v1, dist2, y_true2, stress2, cc_v2, srocc_v2,\ dist3, y_true3, stress3, cc_v3, srocc_v3 = trainNet(config, i) performance.loc['{}'.format(i), 'stress'] = stress1 performance.loc['{}'.format(i), 'plcc'] = cc_v1 performance.loc['{}'.format(i), 'srcc'] = srocc_v1 performance.loc['{}'.format(i), 'stress_aligned'] = stress2 performance.loc['{}'.format(i), 'plcc_aligned'] = cc_v2 performance.loc['{}'.format(i), 'srcc_aligned'] = srocc_v2 performance.loc['{}'.format(i), 'stress_notaligned'] = stress3 performance.loc['{}'.format(i), 'plcc_notaligned'] = cc_v3 performance.loc['{}'.format(i), 'srcc_notaligned'] = srocc_v3 performance.to_csv(config.work_path + '/modelperformance.csv', index=None)
3,377
48.676471
175
py
CD-Flow
CD-Flow-main/test.py
import time from EMA import EMA import torch from torch.utils.data import DataLoader from model import CDFlow from DataLoader import CD_128 from coeff_func import * import os from loss import createLossAndOptimizer from torch.autograd import Variable import torchvision import torch.autograd as autograd from function import setup_seed, copy_codes import argparse def parse_config(): parser = argparse.ArgumentParser() parser.add_argument("--batch_size_test", type=int, default=4) parser.add_argument("--work_path", type=str, default='work_dir') parser.add_argument("--datapath", type=str, default='data') parser.add_argument("--dataset", type=str, default='') parser.add_argument("--testset", type=str, default='test.csv') parser.add_argument("--test_aligned_path", type=str, default=None) parser.add_argument("--test_notaligned_path", type=str, default=None) return parser.parse_args() def test(data_val_loader, net): dist = [] y_true = [] for i, data in enumerate(data_val_loader, 0): with torch.no_grad(): x, y, gts = data y_val = gts.numpy() x, y, gts = \ Variable(x).cuda(), \ Variable(y).cuda(), \ Variable(gts).cuda() score, _, _, _, _, _, _, _, _, _ = net(x, y) pred = (torch.squeeze(score)).cpu().detach().numpy().tolist() if isinstance(pred, list): dist.extend(pred) y_true.extend(y_val.tolist()) else: dist.append(np.array(pred)) y_true.append(y_val) dist_np = np.array(dist) y_true_np = np.array(y_true).squeeze() stress = compute_stress(dist_np, y_true_np) _, cc_v, srocc_v, krocc_v, rmse_v = coeff_fit(dist_np, y_true_np) return srocc_v, cc_v, stress, dist, y_true config = parse_config() path = config.datapath work_path = config.work_path testpath = config.testset workspace = work_path + '/{}'.format(1) testset = path + '/{}/'.format(1) + testpath test_aligned_path = path + '/{}/test_aligned.csv'.format(1) test_notaligned_path = path + '/{}/test_notaligned.csv'.format(1) datadir = config.dataset batch_size_test = config.batch_size_test test_pairs = np.genfromtxt(open(testset, encoding='UTF-8-sig'), delimiter=',', dtype=str) test_aligned_pairs = np.genfromtxt(open(test_aligned_path), delimiter=',', dtype=str) test_notaligned_pairs = np.genfromtxt(open(test_notaligned_path), delimiter=',', dtype=str) data_test = CD_128(test_pairs[:], root_dir=datadir, test=True) test_aligned = CD_128(test_aligned_pairs[:], root_dir=datadir, test=True) test_notaligned = CD_128(test_notaligned_pairs[:], root_dir=datadir, test=True) data_test_loader = DataLoader(data_test, batch_size=batch_size_test, shuffle=False, pin_memory=True, num_workers=4) data_test_aligned_loader = DataLoader(test_aligned, batch_size=batch_size_test, shuffle=False, pin_memory=True, num_workers=4) data_test_notaligned_loader = DataLoader(test_notaligned, batch_size=batch_size_test, shuffle=False, pin_memory=True, num_workers=4) print('#############################################################################') print("Testing...") print('#############################################################################') device = torch.device("cuda") pt = os.path.join(workspace, 'checkpoint_best', 'ModelParams_Best_val.pt') checkpoint = torch.load(pt) net = CDFlow().cuda() net = torch.nn.DataParallel(net).cuda() net.load_state_dict(checkpoint['state_dict']) net.eval() srocc_v1, cc_v1, stress1, dist1, y_true1 = test(data_test_loader, net) print('All: plcc{}; srcc{}; stress{}'.format(cc_v1, srocc_v1, stress1)) srocc_v2, cc_v2, stress2, dist2, y_true2 = test(data_test_aligned_loader, net) print('Pixel-wise aligned: plcc{}; srcc{}; stress{}'.format(cc_v2, srocc_v2, stress2)) srocc_v3, cc_v3, stress3, dist3, y_true3 = test(data_test_notaligned_loader, net) print('Non-Pixel-wise aligned: plcc{}; srcc{}; stress{}'.format(cc_v3, srocc_v3, stress3))
4,109
41.8125
117
py
CD-Flow
CD-Flow-main/flow.py
import torch from torch import nn from torch.nn import functional as F from math import log, pi, exp import numpy as np from scipy import linalg as la logabs = lambda x: torch.log(torch.abs(x)) class ActNorm(nn.Module): def __init__(self, in_channel, logdet=True): super().__init__() self.loc = nn.Parameter(torch.zeros(1, in_channel, 1, 1)) self.scale = nn.Parameter(torch.ones(1, in_channel, 1, 1)) self.register_buffer("initialized", torch.tensor(0, dtype=torch.uint8)) self.logdet = logdet def initialize(self, input): with torch.no_grad(): flatten = input.permute(1, 0, 2, 3).contiguous().view(input.shape[1], -1) mean = ( flatten.mean(1) .unsqueeze(1) .unsqueeze(2) .unsqueeze(3) .permute(1, 0, 2, 3) ) std = ( flatten.std(1) .unsqueeze(1) .unsqueeze(2) .unsqueeze(3) .permute(1, 0, 2, 3) ) self.loc.data.copy_(-mean) self.scale.data.copy_(1 / (std + 1e-6)) def forward(self, input): _, _, height, width = input.shape if self.initialized.item() == 0: self.initialize(input) self.initialized.fill_(1) log_abs = logabs(self.scale) logdet = height * width * torch.sum(log_abs) if self.logdet: return self.scale * (input + self.loc), logdet else: return self.scale * (input + self.loc) def reverse(self, output): return output / self.scale - self.loc class InvConv2d(nn.Module): def __init__(self, in_channel): super().__init__() weight = torch.randn(in_channel, in_channel) q, _ = torch.qr(weight) weight = q.unsqueeze(2).unsqueeze(3) self.weight = nn.Parameter(weight) def forward(self, input): _, _, height, width = input.shape out = F.conv2d(input, self.weight) logdet = ( height * width * torch.slogdet(self.weight.squeeze().double())[1].float() ) return out, logdet def reverse(self, output): return F.conv2d( output, self.weight.squeeze().inverse().unsqueeze(2).unsqueeze(3) ) class InvConv2dLU(nn.Module): def __init__(self, in_channel): super().__init__() weight = np.random.randn(in_channel, in_channel) q, _ = la.qr(weight) w_p, w_l, w_u = la.lu(q.astype(np.float32)) w_s = np.diag(w_u) w_u = np.triu(w_u, 1) u_mask = np.triu(np.ones_like(w_u), 1) l_mask = u_mask.T w_p = torch.from_numpy(w_p) w_l = torch.from_numpy(w_l) w_s = torch.from_numpy(w_s) w_u = torch.from_numpy(w_u) self.register_buffer("w_p", w_p) self.register_buffer("u_mask", torch.from_numpy(u_mask)) self.register_buffer("l_mask", torch.from_numpy(l_mask)) self.register_buffer("s_sign", torch.sign(w_s)) self.register_buffer("l_eye", torch.eye(l_mask.shape[0])) self.w_l = nn.Parameter(w_l) self.w_s = nn.Parameter(logabs(w_s)) self.w_u = nn.Parameter(w_u) def forward(self, input): _, _, height, width = input.shape weight = self.calc_weight() out = F.conv2d(input, weight) logdet = height * width * torch.sum(self.w_s) return out, logdet def calc_weight(self): weight = ( self.w_p @ (self.w_l * self.l_mask + self.l_eye) @ ((self.w_u * self.u_mask) + torch.diag(self.s_sign * torch.exp(self.w_s))) ) return weight.unsqueeze(2).unsqueeze(3) def reverse(self, output): weight = self.calc_weight() return F.conv2d(output, weight.squeeze().inverse().unsqueeze(2).unsqueeze(3)) class ZeroConv2d(nn.Module): def __init__(self, in_channel, out_channel, padding=1): super().__init__() self.conv = nn.Conv2d(in_channel, out_channel, 3, padding=0) self.conv.weight.data.zero_() self.conv.bias.data.zero_() self.scale = nn.Parameter(torch.zeros(1, out_channel, 1, 1)) def forward(self, input): out = F.pad(input, [1, 1, 1, 1], value=1) out = self.conv(out) out = out * torch.exp(self.scale * 3) return out class AffineCoupling(nn.Module): def __init__(self, in_channel, filter_size=512, affine=True): super().__init__() self.affine = affine self.net = nn.Sequential( nn.Conv2d(in_channel // 2, filter_size, 3, padding=1), nn.ReLU(inplace=True), nn.Conv2d(filter_size, filter_size, 1), nn.ReLU(inplace=True), ZeroConv2d(filter_size, in_channel if self.affine else in_channel // 2), ) self.net[0].weight.data.normal_(0, 0.05) self.net[0].bias.data.zero_() self.net[2].weight.data.normal_(0, 0.05) self.net[2].bias.data.zero_() def forward(self, input): in_a, in_b = input.chunk(2, 1) if self.affine: log_s, t = self.net(in_a).chunk(2, 1) s = F.sigmoid(log_s + 2) out_b = (in_b + t) * s logdet = torch.sum(torch.log(s).view(input.shape[0], -1), 1) else: net_out = self.net(in_a) out_b = in_b + net_out logdet = None return torch.cat([in_a, out_b], 1), logdet def reverse(self, output): out_a, out_b = output.chunk(2, 1) if self.affine: log_s, t = self.net(out_a).chunk(2, 1) # s = torch.exp(log_s) s = F.sigmoid(log_s + 2) # in_a = (out_a - t) / s in_b = out_b / s - t else: net_out = self.net(out_a) in_b = out_b - net_out return torch.cat([out_a, in_b], 1) class Flow(nn.Module): def __init__(self, in_channel, affine=True, conv_lu=True): super().__init__() self.actnorm = ActNorm(in_channel) if conv_lu: self.invconv = InvConv2dLU(in_channel) else: self.invconv = InvConv2d(in_channel) self.coupling = AffineCoupling(in_channel, affine=affine) def forward(self, input): out, logdet = self.actnorm(input) out, det1 = self.invconv(out) out, det2 = self.coupling(out) logdet = logdet + det1 if det2 is not None: logdet = logdet + det2 return out, logdet def reverse(self, output): input = self.coupling.reverse(output) input = self.invconv.reverse(input) input = self.actnorm.reverse(input) return input def gaussian_log_p(x, mean, log_sd): return -0.5 * log(2 * pi) - log_sd - 0.5 * (x - mean) ** 2 / torch.exp(2 * log_sd) def gaussian_sample(eps, mean, log_sd): return mean + torch.exp(log_sd) * eps class Block(nn.Module): def __init__(self, in_channel, n_flow, split=True, affine=True, conv_lu=True): super().__init__() squeeze_dim = in_channel * 4 self.flows = nn.ModuleList() for i in range(n_flow): self.flows.append(Flow(squeeze_dim, affine=affine, conv_lu=conv_lu)) self.split = split if split: self.prior = ZeroConv2d(in_channel * 2, in_channel * 4) else: self.prior = ZeroConv2d(in_channel * 4, in_channel * 8) def forward(self, input): b_size, n_channel, height, width = input.shape squeezed = input.view(b_size, n_channel, height // 2, 2, width // 2, 2) squeezed = squeezed.permute(0, 1, 3, 5, 2, 4) out = squeezed.contiguous().view(b_size, n_channel * 4, height // 2, width // 2) logdet = 0 for flow in self.flows: out, det = flow(out) logdet = logdet + det if self.split: out, z_new = out.chunk(2, 1) mean, log_sd = self.prior(out).chunk(2, 1) log_p = gaussian_log_p(z_new, mean, log_sd) log_p = log_p.view(b_size, -1).sum(1) else: one = torch.ones_like(out) mean, log_sd = self.prior(one).chunk(2, 1) log_p = gaussian_log_p(out, mean, log_sd) log_p = log_p.view(b_size, -1).sum(1) z_new = out #self.log_sd = log_sd return out, logdet, log_p, z_new def reverse(self, output, eps=None, reconstruct=False): input = output if reconstruct: if self.split: input = torch.cat([output, eps], 1) ## channel-wise concat else: input = eps else: if self.split: mean, log_sd = self.prior(input).chunk(2, 1) z = gaussian_sample(eps, mean, log_sd) input = torch.cat([output, z], 1) else: one = torch.ones_like(input) mean, log_sd = self.prior(one).chunk(2, 1) z = gaussian_sample(eps, mean, log_sd) input = z for flow in self.flows[::-1]: input = flow.reverse(input) b_size, n_channel, height, width = input.shape unsqueezed = input.view(b_size, n_channel // 4, 2, 2, height, width) unsqueezed = unsqueezed.permute(0, 1, 4, 2, 5, 3) unsqueezed = unsqueezed.contiguous().view( b_size, n_channel // 4, height * 2, width * 2 ) return unsqueezed class Glow(nn.Module): def __init__( self, in_channel, n_flow, n_block, affine=True, conv_lu=True ): super().__init__() self.blocks = nn.ModuleList() n_channel = in_channel for i in range(n_block - 1): self.blocks.append(Block(n_channel, n_flow, affine=affine, conv_lu=conv_lu)) n_channel *= 2 self.blocks.append(Block(n_channel, n_flow, split=False, affine=affine)) def forward(self, input): log_p_sum = 0 logdet = 0 out = input z_outs = [] for i, block in enumerate(self.blocks): out, det, log_p, z_new = block(out) z_outs.append(z_new) logdet = logdet + det if log_p is not None: log_p_sum = log_p_sum + log_p return log_p_sum, logdet, z_outs def reverse(self, z_list, reconstruct=True, cd_map=False): for i, block in enumerate(self.blocks[::-1]): if i == 0: input = block.reverse(z_list[-1], z_list[-1], reconstruct=reconstruct) else: input = block.reverse(input, z_list[-(i + 1)], reconstruct=reconstruct) return input
10,847
28.720548
88
py
CD-Flow
CD-Flow-main/DataLoader.py
import os import torch import random import numpy as np from torch.utils.data import Dataset from PIL import Image from torchvision import transforms import torchvision class CD_128(Dataset): def __init__(self, jnd_info, root_dir, test=False): self.ref_name = jnd_info[:, 0] self.test_name = jnd_info[:, 1] self.root_dir = str(root_dir) self.gt = jnd_info[:, 2] self.test = test if test == False: self.trans_org = transforms.Compose([ transforms.Resize(1024), transforms.RandomRotation(3), transforms.RandomCrop(1000), transforms.Resize(768), transforms.ToTensor(), ]) else: self.trans_org = transforms.Compose([ transforms.Resize(1024), transforms.CenterCrop(1024), transforms.ToTensor(), ]) def __len__(self): return len(self.gt) def __getitem__(self, idx): gt = float(self.gt[idx]) full_address = os.path.join(self.root_dir, str(self.ref_name[idx])) ref = Image.open(full_address).convert("RGB") ref1 = self.trans_org(ref) full_address_test = os.path.join(self.root_dir, str(self.test_name[idx])) test = Image.open(full_address_test).convert("RGB") test1 = self.trans_org(test) return ref1, test1, gt
1,425
30
81
py
CD-Flow
CD-Flow-main/loss.py
import torch import numpy as np import torch.optim as optim import torch.nn as nn import torch.nn.functional as F def createLossAndOptimizer(net, learning_rate, scheduler_step, scheduler_gamma): loss = LossFunc() # optimizer = optim.Adam([{'params': net.parameters(), 'lr':learning_rate}], lr = learning_rate, weight_decay=5e-4) optimizer = optim.Adam([{'params': net.parameters(), 'lr': learning_rate}], lr=learning_rate, eps=1e-7) scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=scheduler_step, gamma=scheduler_gamma) return loss, optimizer, scheduler class LossFunc(torch.nn.Module): def __init__(self): super(LossFunc, self).__init__() def mse_loss(self, score, label): score = torch.squeeze(score) return torch.mean((score - label) ** 2) def forward(self, score, label): mse = self.mse_loss(score, label) return mse
909
34
119
py
CD-Flow
CD-Flow-main/model.py
import math import time import torch import torch.nn as nn from flow import * import os class CDFlow(nn.Module): def __init__(self): super(CDFlow, self).__init__() self.glow = Glow(3, 8, 6, affine=True, conv_lu=True) def coordinate_transform(self, x_hat, rev=False): if not rev: log_p, logdet, x_hat = self.glow(x_hat) return log_p, logdet, x_hat else: x_hat = self.glow.reverse(x_hat) return x_hat def forward(self, x, y): log_p_x, logdet_x, x_hat = self.coordinate_transform(x, rev=False) log_p_y, logdet_y, y_hat = self.coordinate_transform(y, rev=False) x_hat_1, y_hat_1 = x_hat[0].view(x_hat[0].shape[0], -1), y_hat[0].view(x_hat[0].shape[0], -1) x_hat_2, y_hat_2 = x_hat[1].view(x_hat[1].shape[0], -1), y_hat[1].view(x_hat[1].shape[0], -1) x_hat_3, y_hat_3 = x_hat[2].view(x_hat[2].shape[0], -1), y_hat[2].view(x_hat[2].shape[0], -1) x_hat_4, y_hat_4 = x_hat[3].view(x_hat[3].shape[0], -1), y_hat[3].view(x_hat[3].shape[0], -1) x_hat_5, y_hat_5 = x_hat[4].view(x_hat[4].shape[0], -1), y_hat[4].view(x_hat[4].shape[0], -1) x_hat_6, y_hat_6 = x_hat[5].view(x_hat[5].shape[0], -1), y_hat[5].view(x_hat[5].shape[0], -1) x_cat_65 = torch.cat((x_hat_6, x_hat_5), dim=1) y_cat_65 = torch.cat((y_hat_6, y_hat_5), dim=1) x_cat_654 = torch.cat((x_hat_6, x_hat_5, x_hat_4), dim=1) y_cat_654 = torch.cat((y_hat_6, y_hat_5, y_hat_4), dim=1) x_cat_6543 = torch.cat((x_hat_6, x_hat_5, x_hat_4, x_hat_3), dim=1) y_cat_6543 = torch.cat((y_hat_6, y_hat_5, y_hat_4, y_hat_3), dim=1) x_cat_65432 = torch.cat((x_hat_6, x_hat_5, x_hat_4, x_hat_3, x_hat_2), dim=1) y_cat_65432 = torch.cat((y_hat_6, y_hat_5, y_hat_4, y_hat_3, y_hat_2), dim=1) x_cat_654321 = torch.cat((x_hat_6, x_hat_5, x_hat_4, x_hat_3, x_hat_2, x_hat_1), dim=1) y_cat_654321 = torch.cat((y_hat_6, y_hat_5, y_hat_4, y_hat_3, y_hat_2, y_hat_1), dim=1) mse6 = (x_hat_6 - y_hat_6).view(x_hat_6.shape[0], -1) mse6 = mse6.unsqueeze(1) mse6 = torch.sqrt(1e-8 + torch.matmul(mse6, mse6.transpose(dim0=-2, dim1=-1))/mse6.shape[2]) mse6 = mse6.squeeze(2) mse65 = (x_cat_65 - y_cat_65).view(x_cat_65.shape[0], -1) mse65 = mse65.unsqueeze(1) mse65 = torch.sqrt(1e-8 + torch.matmul(mse65, mse65.transpose(dim0=-2, dim1=-1))/mse65.shape[2]) mse65 = mse65.squeeze(2) mse654 = (x_cat_654 - y_cat_654).view(x_cat_654.shape[0], -1) mse654 = mse654.unsqueeze(1) mse654 = torch.sqrt(1e-8 + torch.matmul(mse654, mse654.transpose(dim0=-2, dim1=-1))/mse654.shape[2]) mse654 = mse654.squeeze(2) mse6543 = (x_cat_6543 - y_cat_6543).view(x_cat_6543.shape[0], -1) mse6543 = mse6543.unsqueeze(1) mse6543 = torch.sqrt(1e-8 + torch.matmul(mse6543, mse6543.transpose(dim0=-2, dim1=-1))/mse6543.shape[2]) mse6543 = mse6543.squeeze(2) mse65432 = (x_cat_65432 - y_cat_65432).view(x_cat_65432.shape[0], -1) mse65432 = mse65432.unsqueeze(1) mse65432 = torch.sqrt(1e-8 + torch.matmul(mse65432, mse65432.transpose(dim0=-2, dim1=-1)) / mse65432.shape[2]) mse65432 = mse65432.squeeze(2) mse654321 = (x_cat_654321 - y_cat_654321).view(x_cat_654321.shape[0], -1) mse654321 = mse654321.unsqueeze(1) mse654321 = torch.sqrt(1e-8 + torch.matmul(mse654321, mse654321.transpose(dim0=-2, dim1=-1)) / mse654321.shape[2]) mse654321 = mse654321.squeeze(2) return mse654321, mse65432, mse6543, mse654, mse65, mse6, log_p_x, logdet_x, log_p_y, logdet_y
3,702
47.090909
122
py
CD-Flow
CD-Flow-main/function.py
import shutil import random import torch import numpy as np def setup_seed(seed): torch.manual_seed(seed) torch.cuda.manual_seed_all(seed) np.random.seed(seed) random.seed(seed) torch.backends.cudnn.deterministic = True def copy_codes(trainpath1,trainpath2,trainpath3,trainpath4, path1,path2,path3,path4): shutil.copyfile(trainpath1, path1) shutil.copyfile(trainpath2, path2) shutil.copyfile(trainpath3, path3) shutil.copyfile(trainpath4, path4)
484
25.944444
85
py
CD-Flow
CD-Flow-main/trainnet.py
import time from EMA import EMA import torch from torch.utils.data import DataLoader from model import CDFlow from DataLoader import CD_128 from coeff_func import * import os from loss import createLossAndOptimizer from torch.autograd import Variable import torch.autograd as autograd from function import setup_seed, copy_codes from math import log def trainNet(config, times): resume_path = config.resume_path learning_rate = config.learning_rate scheduler_step = config.scheduler_step scheduler_gamma = config.scheduler_gamma batch_size_train = config.batch_size_train batch_size_test = config.batch_size_test n_epochs = config.n_epochs training_datadir = config.training_datadir colorspace = config.colorspace trainpath1 = config.trainpath1 trainpath2 = config.trainpath2 trainpath3 = config.trainpath3 trainpath4 = config.trainpath4 workspace = config.work_path device = torch.device("cuda") # set random seed setup_seed(config.seed) if not os.path.exists(workspace): os.mkdir(workspace) if not os.path.exists(os.path.join(workspace, 'codes')): os.mkdir(os.path.join(workspace, 'codes')) if not os.path.exists(os.path.join(workspace, 'checkpoint')): os.mkdir(os.path.join(workspace, 'checkpoint')) if not os.path.exists(os.path.join(workspace, 'checkpoint_best')): os.mkdir(os.path.join(workspace, 'checkpoint_best')) copy_codes(trainpath1=trainpath1, trainpath2=trainpath2, trainpath3=trainpath3, trainpath4=trainpath4, path1=os.path.join(workspace, 'codes/trainNet.py'), path2=os.path.join(workspace, 'codes/main.py'), path3=os.path.join(workspace, 'codes/net.py'), path4=os.path.join(workspace, 'codes/DataLoader.py')) print("============ HYPERPARAMETERS ==========") print("batch_size_train and test=", batch_size_train, batch_size_test) print("epochs=", n_epochs) print('learning rate=', learning_rate) print('scheduler_step=', scheduler_step) print('scheduler_gamma=', scheduler_gamma) print('training dir=', training_datadir) print('colorspace=', colorspace) print(config.trainset) print(config.valset) print(config.testset) print(config.test_aligned_path) print(config.test_notaligned_path) train_pairs = np.genfromtxt(open(config.trainset, encoding='UTF-8-sig'), delimiter=',', dtype=str) val_pairs = np.genfromtxt(open(config.valset, encoding='UTF-8-sig'), delimiter=',', dtype=str) test_pairs = np.genfromtxt(open(config.testset, encoding='UTF-8-sig'), delimiter=',', dtype=str) test_aligned_pairs = np.genfromtxt(open(config.test_aligned_path), delimiter=',', dtype=str) test_notaligned_pairs = np.genfromtxt(open(config.test_notaligned_path), delimiter=',', dtype=str) data_train = CD_128(train_pairs[:], root_dir=training_datadir, test=False) data_val = CD_128(val_pairs[:], root_dir=training_datadir, test=True) data_test = CD_128(test_pairs[:], root_dir=training_datadir, test=True) test_aligned = CD_128(test_aligned_pairs[:], root_dir=training_datadir, test=True) test_notaligned = CD_128(test_notaligned_pairs[:], root_dir=training_datadir, test=True) net = CDFlow().to(device) net = torch.nn.DataParallel(net) net = net.to(device) loss, optimizer, scheduler = createLossAndOptimizer(net, learning_rate, scheduler_step, scheduler_gamma) data_train_loader = DataLoader(data_train, batch_size=batch_size_train, shuffle=True, pin_memory=True, num_workers=4) data_val_loader = DataLoader(data_val, batch_size=batch_size_test, shuffle=True, pin_memory=True, num_workers=4) data_test_loader = DataLoader(data_test, batch_size=batch_size_test, shuffle=False, pin_memory=True, num_workers=4) data_test_aligned_loader = DataLoader(test_aligned, batch_size=batch_size_test, shuffle=False, pin_memory=True, num_workers=4) data_test_notaligned_loader = DataLoader(test_notaligned, batch_size=batch_size_test, shuffle=False, pin_memory=True, num_workers=4) if resume_path is not None: checkpoint = torch.load(resume_path) start_epoch = checkpoint['epoch'] + 1 net.load_state_dict(checkpoint['state_dict']) optimizer.load_state_dict(checkpoint['optimizer']) print('continue to train: shuffle{} epoch{} '.format(times + 1, start_epoch)) else: start_epoch = 0 training_start_time = time.time() rows, columns = train_pairs.shape n_batches = rows // batch_size_train valsrcc = 0 ema = EMA(net, 0.999) ema.register() autograd.set_detect_anomaly(True) for epoch in range(start_epoch, n_epochs): # initiate parameters for statistic recordings. dist = [] y_true = [] running_loss = 0.0 total_train_loss = 0 start_time = time.time() print_every = 20 train_counter = 0 net.train() print("---------------------train mode-------epoch{}--------------------------".format(epoch)) for i, data in enumerate(data_train_loader, 0): train_counter = train_counter + 1 x, y, gts = data y_val = gts.numpy() x, y, gts = \ Variable(x).to(device), \ Variable(y).to(device), \ Variable(gts).to(device) optimizer.zero_grad() score, score65432, score6543, score654, score65, score6, log_p_x, logdet_x, log_p_y, logdet_y = net(x, y) logdet_x = logdet_x.mean() logdet_y = logdet_y.mean() loss_x, log_p_x, log_det_x = calc_loss(log_p_x, logdet_x, 768, 2.0 ** 5) loss_y, log_p_y, log_det_y = calc_loss(log_p_y, logdet_y, 768, 2.0 ** 5) score_loss = 10 * loss(score, gts) + loss(score65432, gts) + loss(score6543, gts) + loss(score654, gts) + loss(score65, gts) + loss(score6, gts) loss_size = 10 * score_loss + loss_x + loss_y loss_size.backward() optimizer.step() ema.update() running_loss += loss_size.item() total_train_loss += loss_size.item() pred = (torch.squeeze(score)).cpu().detach().numpy().tolist() if isinstance(pred, list): dist.extend(pred) y_true.extend(y_val.tolist()) else: dist.append(np.array(pred)) y_true.append(y_val) if (i + 1) % (print_every + 1) == 0: print("Epoch {}, {:d}% \t train_loss: {:.6f} took: {:.2f}s".format( epoch + 1, int(100 * (i + 1) / n_batches), running_loss / print_every, time.time() - start_time)) running_loss = 0.0 start_time = time.time() torch.save( {"state_dict": net.state_dict(), 'epoch': epoch, 'optimizer': optimizer.state_dict(), 'times': times}, \ os.path.join(workspace, 'checkpoint', 'ModelParams_checkpoint.pt')) # Calculate correlation coefficients between the predicted values and ground truth values on training set. dist = np.array(dist).squeeze() y_true = np.array(y_true).squeeze() _, cc_v, srocc_v, krocc_v, rmse_v = coeff_fit(dist, y_true) print("Training set: PCC{:.4}, SROCC{:.4}, KROCC{:.4}, RMSE{:.4}".format(cc_v, srocc_v, krocc_v, rmse_v)) # validation # EMA ema.apply_shadow() # EMA net.eval() print("----------------------------validation mode---------------------------------") srocc_v, total_val_loss, val_counter, cc_v, krocc_v, rmse_v, stress, dist, y_true, score_val = test( data_val_loader, net, loss) # srocc_a, total_val_loss_a, val_counter_a, cc_a, krocc_a, rmse_a, stress_a, dist_a, y_true_a, score_a = test( # data_test_aligned_loader, net, loss) # srocc_na, total_val_loss_na, val_counter_na, cc_na, krocc_na, rmse_na, stress_na, dist_na, y_true_na, score_na = test( # data_test_notaligned_loader, net, loss) if srocc_v > valsrcc: valsrcc = srocc_v torch.save({"state_dict": net.state_dict()}, os.path.join(workspace, 'checkpoint_best', 'ModelParams_Best_val.pt')) print('update best model...') print("VALIDATION: PCC{:.4}, SROCC{:.4}, STRESS{:.4}, RMSE{:.4}".format(cc_v, srocc_v, stress, rmse_v)) print("loss = {:.6}".format(total_val_loss / val_counter)) # EMA ema.restore() # EMA scheduler.step() print('#############################################################################') print("Training finished, took {:.2f}s".format(time.time() - training_start_time)) pt = os.path.join(workspace, 'checkpoint_best', 'ModelParams_Best_val.pt') checkpoint = torch.load(pt) net = CDFlow().to(device) net = torch.nn.DataParallel(net).to(device) net.load_state_dict(checkpoint['state_dict']) net.eval() srocc_v1, total_val_loss, val_counter, cc_v1, krocc_v, rmse_v, stress1, dist1, y_true1, score_val = test( data_test_loader, net, loss) print('best performance: plcc{} srcc{}'.format(cc_v1, srocc_v1)) srocc_v2, total_val_loss, val_counter, cc_v2, krocc_v, rmse_v, stress2, dist2, y_true2, score_val = test( data_test_aligned_loader, net, loss) print('best performance in Pixel-wise aligned: plcc{} srcc{}'.format(cc_v2, srocc_v2)) srocc_v3, total_val_loss, val_counter, cc_v3, krocc_v, rmse_v, stress3, dist3, y_true3, score_val = test( data_test_notaligned_loader, net, loss) print('best performance in non-Pixel-wise aligned: plcc{} srcc{}'.format(cc_v3, srocc_v3)) return dist1, y_true1, stress1, cc_v1, srocc_v1, dist2, y_true2, stress2, cc_v2, srocc_v2, dist3, y_true3, stress3, cc_v3, srocc_v3 def test(data_val_loader, net, loss): total_val_loss = 0 val_counter = 0 score_val = 0 dist = [] y_true = [] device = torch.device("cuda") for i, data in enumerate(data_val_loader, 0): with torch.no_grad(): x, y, gts = data y_val = gts.numpy() x, y, gts = \ Variable(x).to(device), \ Variable(y).to(device), \ Variable(gts).to(device) score, _, _, _, _, _, _, _, _, _ = net(x, y) score_loss = loss(score, gts) loss_size = score_loss total_val_loss += loss_size.cpu().numpy() score_val = score_val + score_loss.item() val_counter += 1 pred = (torch.squeeze(score)).cpu().detach().numpy().tolist() if isinstance(pred, list): dist.extend(pred) y_true.extend(y_val.tolist()) else: dist.append(np.array(pred)) y_true.append(y_val) # Calculate correlation coefficients between the predicted values and ground truth values on validation set. dist_np = np.array(dist).squeeze() y_true_np = np.array(y_true).squeeze() stress = compute_stress(dist_np, y_true_np) _, cc_v, srocc_v, krocc_v, rmse_v = coeff_fit(dist_np, y_true_np) return srocc_v, total_val_loss, val_counter, cc_v, krocc_v, rmse_v, stress, dist, y_true, score_val def calc_loss(log_p, logdet, image_size, n_bins): n_pixel = image_size * image_size * 3 loss = -log(n_bins) * n_pixel loss = loss + logdet + log_p return (-loss / (log(2) * n_pixel)).mean(), (log_p / (log(2) * n_pixel)).mean(), ( logdet / (log(2) * n_pixel)).mean()
11,783
44.85214
156
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/dqn_algos/demo.py
import numpy as np from arguments import get_args from models import net import torch from rl_utils.env_wrapper.atari_wrapper import make_atari, wrap_deepmind def get_tensors(obs): obs = np.transpose(obs, (2, 0, 1)) obs = np.expand_dims(obs, 0) obs = torch.tensor(obs, dtype=torch.float32) return obs if __name__ == '__main__': args = get_args() # create the environment env = make_atari(args.env_name) env = wrap_deepmind(env, frame_stack=True) # create the network net = net(env.action_space.n, args.use_dueling) # model path model_path = args.save_dir + args.env_name + '/model.pt' # load the models net.load_state_dict(torch.load(model_path, map_location=lambda storage, loc: storage)) # start to test the demo obs = env.reset() for _ in range(2000): env.render() with torch.no_grad(): obs_tensor = get_tensors(obs) action_value = net(obs_tensor) action = torch.argmax(action_value.squeeze()).item() obs, reward, done, _ = env.step(action) if done: obs = env.reset() env.close()
1,134
30.527778
90
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/dqn_algos/models.py
import torch import torch.nn as nn import torch.nn.functional as F # the convolution layer of deepmind class deepmind(nn.Module): def __init__(self): super(deepmind, self).__init__() self.conv1 = nn.Conv2d(4, 32, 8, stride=4) self.conv2 = nn.Conv2d(32, 64, 4, stride=2) self.conv3 = nn.Conv2d(64, 32, 3, stride=1) # start to do the init... nn.init.orthogonal_(self.conv1.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.conv2.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.conv3.weight.data, gain=nn.init.calculate_gain('relu')) # init the bias... nn.init.constant_(self.conv1.bias.data, 0) nn.init.constant_(self.conv2.bias.data, 0) nn.init.constant_(self.conv3.bias.data, 0) def forward(self, x): x = F.relu(self.conv1(x)) x = F.relu(self.conv2(x)) x = F.relu(self.conv3(x)) x = x.view(-1, 32 * 7 * 7) return x # in the initial, just the nature CNN class net(nn.Module): def __init__(self, num_actions, use_dueling=False): super(net, self).__init__() # if use the dueling network self.use_dueling = use_dueling # define the network self.cnn_layer = deepmind() # if not use dueling if not self.use_dueling: self.fc1 = nn.Linear(32 * 7 * 7, 256) self.action_value = nn.Linear(256, num_actions) else: # the layer for dueling network architecture self.action_fc = nn.Linear(32 * 7 * 7, 256) self.state_value_fc = nn.Linear(32 * 7 * 7, 256) self.action_value = nn.Linear(256, num_actions) self.state_value = nn.Linear(256, 1) def forward(self, inputs): x = self.cnn_layer(inputs / 255.0) if not self.use_dueling: x = F.relu(self.fc1(x)) action_value_out = self.action_value(x) else: # get the action value action_fc = F.relu(self.action_fc(x)) action_value = self.action_value(action_fc) # get the state value state_value_fc = F.relu(self.state_value_fc(x)) state_value = self.state_value(state_value_fc) # action value mean action_value_mean = torch.mean(action_value, dim=1, keepdim=True) action_value_center = action_value - action_value_mean # Q = V + A action_value_out = state_value + action_value_center return action_value_out
2,596
37.761194
88
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/dqn_algos/dqn_agent.py
import sys import numpy as np from models import net from utils import linear_schedule, select_actions, reward_recorder from rl_utils.experience_replay.experience_replay import replay_buffer import torch from datetime import datetime import os import copy # define the dqn agent class dqn_agent: def __init__(self, env, args): # define some important self.env = env self.args = args # define the network self.net = net(self.env.action_space.n, self.args.use_dueling) # copy the self.net as the self.target_net = copy.deepcopy(self.net) # make sure the target net has the same weights as the network self.target_net.load_state_dict(self.net.state_dict()) if self.args.cuda: self.net.cuda() self.target_net.cuda() # define the optimizer self.optimizer = torch.optim.Adam(self.net.parameters(), lr=self.args.lr) # define the replay memory self.buffer = replay_buffer(self.args.buffer_size) # define the linear schedule of the exploration self.exploration_schedule = linear_schedule(int(self.args.total_timesteps * self.args.exploration_fraction), \ self.args.final_ratio, self.args.init_ratio) # create the folder to save the models if not os.path.exists(self.args.save_dir): os.mkdir(self.args.save_dir) # set the environment folder self.model_path = os.path.join(self.args.save_dir, self.args.env_name) if not os.path.exists(self.model_path): os.mkdir(self.model_path) # start to do the training def learn(self): # the episode reward episode_reward = reward_recorder() obs = np.array(self.env.reset()) td_loss = 0 for timestep in range(self.args.total_timesteps): explore_eps = self.exploration_schedule.get_value(timestep) with torch.no_grad(): obs_tensor = self._get_tensors(obs) action_value = self.net(obs_tensor) # select actions action = select_actions(action_value, explore_eps) # excute actions obs_, reward, done, _ = self.env.step(action) obs_ = np.array(obs_) # tryint to append the samples self.buffer.add(obs, action, reward, obs_, float(done)) obs = obs_ # add the rewards episode_reward.add_rewards(reward) if done: obs = np.array(self.env.reset()) # start new episode to store rewards episode_reward.start_new_episode() if timestep > self.args.learning_starts and timestep % self.args.train_freq == 0: # start to sample the samples from the replay buffer batch_samples = self.buffer.sample(self.args.batch_size) td_loss = self._update_network(batch_samples) if timestep > self.args.learning_starts and timestep % self.args.target_network_update_freq == 0: # update the target network self.target_net.load_state_dict(self.net.state_dict()) if done and episode_reward.num_episodes % self.args.display_interval == 0: print('[{}] Frames: {}, Episode: {}, Mean: {:.3f}, Loss: {:.3f}'.format(datetime.now(), timestep, episode_reward.num_episodes, \ episode_reward.mean, td_loss)) torch.save(self.net.state_dict(), self.model_path + '/model.pt') # update the network def _update_network(self, samples): obses, actions, rewards, obses_next, dones = samples # convert the data to tensor obses = self._get_tensors(obses) actions = torch.tensor(actions, dtype=torch.int64).unsqueeze(-1) rewards = torch.tensor(rewards, dtype=torch.float32).unsqueeze(-1) obses_next = self._get_tensors(obses_next) dones = torch.tensor(1 - dones, dtype=torch.float32).unsqueeze(-1) # convert into gpu if self.args.cuda: actions = actions.cuda() rewards = rewards.cuda() dones = dones.cuda() # calculate the target value with torch.no_grad(): # if use the double network architecture if self.args.use_double_net: q_value_ = self.net(obses_next) action_max_idx = torch.argmax(q_value_, dim=1, keepdim=True) target_action_value = self.target_net(obses_next) target_action_max_value = target_action_value.gather(1, action_max_idx) else: target_action_value = self.target_net(obses_next) target_action_max_value, _ = torch.max(target_action_value, dim=1, keepdim=True) # target expected_value = rewards + self.args.gamma * target_action_max_value * dones # get the real q value action_value = self.net(obses) real_value = action_value.gather(1, actions) loss = (expected_value - real_value).pow(2).mean() # start to update self.optimizer.zero_grad() loss.backward() self.optimizer.step() return loss.item() # get tensors def _get_tensors(self, obs): if obs.ndim == 3: obs = np.transpose(obs, (2, 0, 1)) obs = np.expand_dims(obs, 0) elif obs.ndim == 4: obs = np.transpose(obs, (0, 3, 1, 2)) obs = torch.tensor(obs, dtype=torch.float32) if self.args.cuda: obs = obs.cuda() return obs
5,646
43.81746
144
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/trpo/trpo_agent.py
import torch import numpy as np import os from models import network from rl_utils.running_filter.running_filter import ZFilter from utils import select_actions, eval_actions, conjugated_gradient, line_search, set_flat_params_to from datetime import datetime class trpo_agent: def __init__(self, env, args): self.env = env self.args = args # define the network self.net = network(self.env.observation_space.shape[0], self.env.action_space.shape[0]) self.old_net = network(self.env.observation_space.shape[0], self.env.action_space.shape[0]) # make sure the net and old net have the same parameters self.old_net.load_state_dict(self.net.state_dict()) # define the optimizer self.optimizer = torch.optim.Adam(self.net.critic.parameters(), lr=self.args.lr) # define the running mean filter self.running_state = ZFilter((self.env.observation_space.shape[0],), clip=5) if not os.path.exists(self.args.save_dir): os.mkdir(self.args.save_dir) self.model_path = self.args.save_dir + self.args.env_name + '/' if not os.path.exists(self.model_path): os.mkdir(self.model_path) def learn(self): num_updates = self.args.total_timesteps // self.args.nsteps obs = self.running_state(self.env.reset()) final_reward = 0 episode_reward = 0 self.dones = False for update in range(num_updates): mb_obs, mb_rewards, mb_actions, mb_dones, mb_values = [], [], [], [], [] for step in range(self.args.nsteps): with torch.no_grad(): obs_tensor = self._get_tensors(obs) value, pi = self.net(obs_tensor) # select actions actions = select_actions(pi) # store informations mb_obs.append(np.copy(obs)) mb_actions.append(actions) mb_dones.append(self.dones) mb_values.append(value.detach().numpy().squeeze()) # start to execute actions in the environment obs_, reward, done, _ = self.env.step(actions) self.dones = done mb_rewards.append(reward) if done: obs_ = self.env.reset() obs = self.running_state(obs_) episode_reward += reward mask = 0.0 if done else 1.0 final_reward *= mask final_reward += (1 - mask) * episode_reward episode_reward *= mask # to process the rollouts mb_obs = np.asarray(mb_obs, dtype=np.float32) mb_rewards = np.asarray(mb_rewards, dtype=np.float32) mb_actions = np.asarray(mb_actions, dtype=np.float32) mb_dones = np.asarray(mb_dones, dtype=np.bool) mb_values = np.asarray(mb_values, dtype=np.float32) # compute the last state value with torch.no_grad(): obs_tensor = self._get_tensors(obs) last_value, _ = self.net(obs_tensor) last_value = last_value.detach().numpy().squeeze() # compute the advantages mb_returns = np.zeros_like(mb_rewards) mb_advs = np.zeros_like(mb_rewards) lastgaelam = 0 for t in reversed(range(self.args.nsteps)): if t == self.args.nsteps - 1: nextnonterminal = 1.0 - self.dones nextvalues = last_value else: nextnonterminal = 1.0 - mb_dones[t + 1] nextvalues = mb_values[t + 1] delta = mb_rewards[t] + self.args.gamma * nextvalues * nextnonterminal - mb_values[t] mb_advs[t] = lastgaelam = delta + self.args.gamma * self.args.tau * nextnonterminal * lastgaelam mb_returns = mb_advs + mb_values # normalize the advantages mb_advs = (mb_advs - mb_advs.mean()) / (mb_advs.std() + 1e-5) # before the update, make the old network has the parameter of the current network self.old_net.load_state_dict(self.net.state_dict()) # start to update the network policy_loss, value_loss = self._update_network(mb_obs, mb_actions, mb_returns, mb_advs) torch.save([self.net.state_dict(), self.running_state], self.model_path + 'model.pt') print('[{}] Update: {} / {}, Frames: {}, Reward: {:.3f}, VL: {:.3f}, PL: {:.3f}'.format(datetime.now(), update, \ num_updates, (update + 1)*self.args.nsteps, final_reward, value_loss, policy_loss)) # start to update network def _update_network(self, mb_obs, mb_actions, mb_returns, mb_advs): mb_obs_tensor = torch.tensor(mb_obs, dtype=torch.float32) mb_actions_tensor = torch.tensor(mb_actions, dtype=torch.float32) mb_returns_tensor = torch.tensor(mb_returns, dtype=torch.float32).unsqueeze(1) mb_advs_tensor = torch.tensor(mb_advs, dtype=torch.float32).unsqueeze(1) # try to get the old policy and current policy values, _ = self.net(mb_obs_tensor) with torch.no_grad(): _, pi_old = self.old_net(mb_obs_tensor) # get the surr loss surr_loss = self._get_surrogate_loss(mb_obs_tensor, mb_advs_tensor, mb_actions_tensor, pi_old) # comupte the surrogate gardient -> g, Ax = g, where A is the fisher information matrix surr_grad = torch.autograd.grad(surr_loss, self.net.actor.parameters()) flat_surr_grad = torch.cat([grad.view(-1) for grad in surr_grad]).data # use the conjugated gradient to calculate the scaled direction vector (natural gradient) nature_grad = conjugated_gradient(self._fisher_vector_product, -flat_surr_grad, 10, mb_obs_tensor, pi_old) # calculate the scaleing ratio non_scale_kl = 0.5 * (nature_grad * self._fisher_vector_product(nature_grad, mb_obs_tensor, pi_old)).sum(0, keepdim=True) scale_ratio = torch.sqrt(non_scale_kl / self.args.max_kl) final_nature_grad = nature_grad / scale_ratio[0] # calculate the expected improvement rate... expected_improve = (-flat_surr_grad * nature_grad).sum(0, keepdim=True) / scale_ratio[0] # get the flat param ... prev_params = torch.cat([param.data.view(-1) for param in self.net.actor.parameters()]) # start to do the line search success, new_params = line_search(self.net.actor, self._get_surrogate_loss, prev_params, final_nature_grad, \ expected_improve, mb_obs_tensor, mb_advs_tensor, mb_actions_tensor, pi_old) set_flat_params_to(self.net.actor, new_params) # then trying to update the critic network inds = np.arange(mb_obs.shape[0]) for _ in range(self.args.vf_itrs): np.random.shuffle(inds) for start in range(0, mb_obs.shape[0], self.args.batch_size): end = start + self.args.batch_size mbinds = inds[start:end] mini_obs = mb_obs[mbinds] mini_returns = mb_returns[mbinds] # put things in the tensor mini_obs = torch.tensor(mini_obs, dtype=torch.float32) mini_returns = torch.tensor(mini_returns, dtype=torch.float32).unsqueeze(1) values, _ = self.net(mini_obs) v_loss = (mini_returns - values).pow(2).mean() self.optimizer.zero_grad() v_loss.backward() self.optimizer.step() return surr_loss.item(), v_loss.item() # get the surrogate loss def _get_surrogate_loss(self, obs, adv, actions, pi_old): _, pi = self.net(obs) log_prob = eval_actions(pi, actions) old_log_prob = eval_actions(pi_old, actions).detach() surr_loss = -torch.exp(log_prob - old_log_prob) * adv return surr_loss.mean() # the product of the fisher informaiton matrix and the nature gradient -> Ax def _fisher_vector_product(self, v, obs, pi_old): kl = self._get_kl(obs, pi_old) kl = kl.mean() # start to calculate the second order gradient of the KL kl_grads = torch.autograd.grad(kl, self.net.actor.parameters(), create_graph=True) flat_kl_grads = torch.cat([grad.view(-1) for grad in kl_grads]) kl_v = (flat_kl_grads * torch.autograd.Variable(v)).sum() kl_second_grads = torch.autograd.grad(kl_v, self.net.actor.parameters()) flat_kl_second_grads = torch.cat([grad.contiguous().view(-1) for grad in kl_second_grads]).data flat_kl_second_grads = flat_kl_second_grads + self.args.damping * v return flat_kl_second_grads # get the kl divergence between two distributions def _get_kl(self, obs, pi_old): mean_old, std_old = pi_old _, pi = self.net(obs) mean, std = pi # start to calculate the kl-divergence kl = -torch.log(std / std_old) + (std.pow(2) + (mean - mean_old).pow(2)) / (2 * std_old.pow(2)) - 0.5 return kl.sum(1, keepdim=True) # get the tensors def _get_tensors(self, obs): return torch.tensor(obs, dtype=torch.float32).unsqueeze(0)
9,299
52.142857
129
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/trpo/utils.py
import numpy as np import torch from torch.distributions.normal import Normal # select actions def select_actions(pi): mean, std = pi normal_dist = Normal(mean, std) return normal_dist.sample().detach().numpy().squeeze() # evaluate the actions def eval_actions(pi, actions): mean, std = pi normal_dist = Normal(mean, std) return normal_dist.log_prob(actions).sum(dim=1, keepdim=True) # conjugated gradient def conjugated_gradient(fvp, b, update_steps, obs, pi_old, residual_tol=1e-10): # the initial solution is zero x = torch.zeros(b.size(), dtype=torch.float32) r = b.clone() p = b.clone() rdotr = torch.dot(r, r) for i in range(update_steps): fv_product = fvp(p, obs, pi_old) alpha = rdotr / torch.dot(p, fv_product) x = x + alpha * p r = r - alpha * fv_product new_rdotr = torch.dot(r, r) beta = new_rdotr / rdotr p = r + beta * p rdotr = new_rdotr # if less than residual tot.. break if rdotr < residual_tol: break return x # line search def line_search(model, loss_fn, x, full_step, expected_rate, obs, adv, actions, pi_old, max_backtracks=10, accept_ratio=0.1): fval = loss_fn(obs, adv, actions, pi_old).data for (_n_backtracks, stepfrac) in enumerate(0.5**np.arange(max_backtracks)): xnew = x + stepfrac * full_step set_flat_params_to(model, xnew) new_fval = loss_fn(obs, adv, actions, pi_old).data actual_improve = fval - new_fval expected_improve = expected_rate * stepfrac ratio = actual_improve / expected_improve if ratio.item() > accept_ratio and actual_improve.item() > 0: return True, xnew return False, x def set_flat_params_to(model, flat_params): prev_indx = 0 for param in model.parameters(): flat_size = int(np.prod(list(param.size()))) param.data.copy_(flat_params[prev_indx:prev_indx + flat_size].view(param.size())) prev_indx += flat_size
2,026
33.355932
125
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/trpo/demo.py
import numpy as np import torch import gym from arguments import get_args from models import network def denormalize(x, mean, std, clip=10): x -= mean x /= (std + 1e-8) return np.clip(x, -clip, clip) def get_tensors(x): return torch.tensor(x, dtype=torch.float32).unsqueeze(0) if __name__ == '__main__': args = get_args() # create the environment env = gym.make(args.env_name) # build up the network net = network(env.observation_space.shape[0], env.action_space.shape[0]) # load the saved model model_path = args.save_dir + args.env_name + '/model.pt' network_model, filters = torch.load(model_path, map_location=lambda storage, loc: storage) net.load_state_dict(network_model) net.eval() for _ in range(10): obs = denormalize(env.reset(), filters.rs.mean, filters.rs.std) reward_total = 0 for _ in range(10000): env.render() obs_tensor = get_tensors(obs) with torch.no_grad(): _, (mean, _) = net(obs_tensor) action = mean.numpy().squeeze() obs, reward, done, _ = env.step(action) reward_total += reward obs = denormalize(obs, filters.rs.mean, filters.rs.std) if done: break print('the reward of this episode is: {}'.format(reward_total)) env.close()
1,383
31.952381
94
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/trpo/models.py
import torch from torch import nn from torch.nn import functional as F class network(nn.Module): def __init__(self, num_states, num_actions): super(network, self).__init__() # define the critic self.critic = critic(num_states) self.actor = actor(num_states, num_actions) def forward(self, x): state_value = self.critic(x) pi = self.actor(x) return state_value, pi class critic(nn.Module): def __init__(self, num_states): super(critic, self).__init__() self.fc1 = nn.Linear(num_states, 64) self.fc2 = nn.Linear(64, 64) self.value = nn.Linear(64, 1) def forward(self, x): x = F.tanh(self.fc1(x)) x = F.tanh(self.fc2(x)) value = self.value(x) return value class actor(nn.Module): def __init__(self, num_states, num_actions): super(actor, self).__init__() self.fc1 = nn.Linear(num_states, 64) self.fc2 = nn.Linear(64, 64) self.action_mean = nn.Linear(64, num_actions) self.sigma_log = nn.Parameter(torch.zeros(1, num_actions)) def forward(self, x): x = F.tanh(self.fc1(x)) x = F.tanh(self.fc2(x)) mean = self.action_mean(x) sigma_log = self.sigma_log.expand_as(mean) sigma = torch.exp(sigma_log) pi = (mean, sigma) return pi
1,376
28.297872
66
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/a2c/a2c_agent.py
import numpy as np import torch from models import net from datetime import datetime from utils import select_actions, evaluate_actions, discount_with_dones import os class a2c_agent: def __init__(self, envs, args): self.envs = envs self.args = args # define the network self.net = net(self.envs.action_space.n) if self.args.cuda: self.net.cuda() # define the optimizer self.optimizer = torch.optim.RMSprop(self.net.parameters(), lr=self.args.lr, eps=self.args.eps, alpha=self.args.alpha) if not os.path.exists(self.args.save_dir): os.mkdir(self.args.save_dir) # check the saved path for envs.. self.model_path = self.args.save_dir + self.args.env_name + '/' if not os.path.exists(self.model_path): os.mkdir(self.model_path) # get the obs.. self.batch_ob_shape = (self.args.num_workers * self.args.nsteps,) + self.envs.observation_space.shape self.obs = np.zeros((self.args.num_workers,) + self.envs.observation_space.shape, dtype=self.envs.observation_space.dtype.name) self.obs[:] = self.envs.reset() self.dones = [False for _ in range(self.args.num_workers)] # train the network.. def learn(self): num_updates = self.args.total_frames // (self.args.num_workers * self.args.nsteps) # get the reward to calculate other information episode_rewards = np.zeros((self.args.num_workers, ), dtype=np.float32) final_rewards = np.zeros((self.args.num_workers, ), dtype=np.float32) # start to update for update in range(num_updates): mb_obs, mb_rewards, mb_actions, mb_dones = [],[],[],[] for step in range(self.args.nsteps): with torch.no_grad(): input_tensor = self._get_tensors(self.obs) _, pi = self.net(input_tensor) # select actions actions = select_actions(pi) cpu_actions = actions.squeeze(1).cpu().numpy() # start to store the information mb_obs.append(np.copy(self.obs)) mb_actions.append(cpu_actions) mb_dones.append(self.dones) # step obs, rewards, dones, _ = self.envs.step(cpu_actions) # start to store the rewards self.dones = dones mb_rewards.append(rewards) for n, done in enumerate(dones): if done: self.obs[n] = self.obs[n]*0 self.obs = obs episode_rewards += rewards # get the masks masks = np.array([0.0 if done else 1.0 for done in dones], dtype=np.float32) final_rewards *= masks final_rewards += (1 - masks) * episode_rewards episode_rewards *= masks # update the obs mb_dones.append(self.dones) # process the rollouts mb_obs = np.asarray(mb_obs, dtype=np.uint8).swapaxes(1, 0).reshape(self.batch_ob_shape) mb_rewards = np.asarray(mb_rewards, dtype=np.float32).swapaxes(1, 0) mb_actions = np.asarray(mb_actions, dtype=np.int32).swapaxes(1, 0) mb_dones = np.asarray(mb_dones, dtype=np.bool).swapaxes(1, 0) mb_masks = mb_dones[:, :-1] mb_dones = mb_dones[:, 1:] # calculate the last value with torch.no_grad(): input_tensor = self._get_tensors(self.obs) last_values, _ = self.net(input_tensor) # compute returns for n, (rewards, dones, value) in enumerate(zip(mb_rewards, mb_dones, last_values.detach().cpu().numpy().squeeze())): rewards = rewards.tolist() dones = dones.tolist() if dones[-1] == 0: rewards = discount_with_dones(rewards+[value], dones+[0], self.args.gamma)[:-1] else: rewards = discount_with_dones(rewards, dones, self.args.gamma) mb_rewards[n] = rewards mb_rewards = mb_rewards.flatten() mb_actions = mb_actions.flatten() # start to update network vl, al, ent = self._update_network(mb_obs, mb_rewards, mb_actions) if update % self.args.log_interval == 0: print('[{}] Update: {}/{}, Frames: {}, Rewards: {:.1f}, VL: {:.3f}, PL: {:.3f}, Ent: {:.2f}, Min: {}, Max:{}'.format(\ datetime.now(), update, num_updates, (update+1)*(self.args.num_workers * self.args.nsteps),\ final_rewards.mean(), vl, al, ent, final_rewards.min(), final_rewards.max())) torch.save(self.net.state_dict(), self.model_path + 'model.pt') # update_network def _update_network(self, obs, returns, actions): # evaluate the actions input_tensor = self._get_tensors(obs) values, pi = self.net(input_tensor) # define the tensor of actions, returns returns = torch.tensor(returns, dtype=torch.float32).unsqueeze(1) actions = torch.tensor(actions, dtype=torch.int64).unsqueeze(1) if self.args.cuda: returns = returns.cuda() actions = actions.cuda() # evaluate actions action_log_probs, dist_entropy = evaluate_actions(pi, actions) # calculate advantages... advantages = returns - values # get the value loss value_loss = advantages.pow(2).mean() # get the action loss action_loss = -(advantages.detach() * action_log_probs).mean() # total loss total_loss = action_loss + self.args.value_loss_coef * value_loss - self.args.entropy_coef * dist_entropy # start to update self.optimizer.zero_grad() total_loss.backward() torch.nn.utils.clip_grad_norm_(self.net.parameters(), self.args.max_grad_norm) self.optimizer.step() return value_loss.item(), action_loss.item(), dist_entropy.item() # get the tensors... def _get_tensors(self, obs): input_tensor = torch.tensor(np.transpose(obs, (0, 3, 1, 2)), dtype=torch.float32) if self.args.cuda: input_tensor = input_tensor.cuda() return input_tensor
6,370
47.633588
135
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/a2c/utils.py
import torch import numpy as np from torch.distributions.categorical import Categorical # select - actions def select_actions(pi, deterministic=False): cate_dist = Categorical(pi) if deterministic: return torch.argmax(pi, dim=1).item() else: return cate_dist.sample().unsqueeze(-1) # get the action log prob and entropy... def evaluate_actions(pi, actions): cate_dist = Categorical(pi) return cate_dist.log_prob(actions.squeeze(-1)).unsqueeze(-1), cate_dist.entropy().mean() def discount_with_dones(rewards, dones, gamma): discounted = [] r = 0 for reward, done in zip(rewards[::-1], dones[::-1]): r = reward + gamma * r * (1.-done) discounted.append(r) return discounted[::-1]
749
29
92
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/a2c/demo.py
from arguments import get_args from models import net import torch from utils import select_actions import cv2 import numpy as np from rl_utils.env_wrapper.frame_stack import VecFrameStack from rl_utils.env_wrapper.atari_wrapper import make_atari, wrap_deepmind # update the current observation def get_tensors(obs): input_tensor = torch.tensor(np.transpose(obs, (2, 0, 1)), dtype=torch.float32).unsqueeze(0) return input_tensor if __name__ == "__main__": args = get_args() # create environment #env = VecFrameStack(wrap_deepmind(make_atari(args.env_name)), 4) env = make_atari(args.env_name) env = wrap_deepmind(env, frame_stack=True) # get the model path model_path = args.save_dir + args.env_name + '/model.pt' network = net(env.action_space.n) network.load_state_dict(torch.load(model_path, map_location=lambda storage, loc: storage)) obs = env.reset() while True: env.render() # get the obs with torch.no_grad(): input_tensor = get_tensors(obs) _, pi = network(input_tensor) actions = select_actions(pi, True) obs, reward, done, _ = env.step([actions]) env.close()
1,193
33.114286
95
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/a2c/models.py
import torch import torch.nn as nn import torch.nn.functional as F # the convolution layer of deepmind class deepmind(nn.Module): def __init__(self): super(deepmind, self).__init__() self.conv1 = nn.Conv2d(4, 32, 8, stride=4) self.conv2 = nn.Conv2d(32, 64, 4, stride=2) self.conv3 = nn.Conv2d(64, 32, 3, stride=1) self.fc1 = nn.Linear(32 * 7 * 7, 512) # start to do the init... nn.init.orthogonal_(self.conv1.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.conv2.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.conv3.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.fc1.weight.data, gain=nn.init.calculate_gain('relu')) # init the bias... nn.init.constant_(self.conv1.bias.data, 0) nn.init.constant_(self.conv2.bias.data, 0) nn.init.constant_(self.conv3.bias.data, 0) nn.init.constant_(self.fc1.bias.data, 0) def forward(self, x): x = F.relu(self.conv1(x)) x = F.relu(self.conv2(x)) x = F.relu(self.conv3(x)) x = x.view(-1, 32 * 7 * 7) x = F.relu(self.fc1(x)) return x # in the initial, just the nature CNN class net(nn.Module): def __init__(self, num_actions): super(net, self).__init__() self.cnn_layer = deepmind() self.critic = nn.Linear(512, 1) self.actor = nn.Linear(512, num_actions) # init the linear layer.. nn.init.orthogonal_(self.critic.weight.data) nn.init.constant_(self.critic.bias.data, 0) # init the policy layer... nn.init.orthogonal_(self.actor.weight.data, gain=0.01) nn.init.constant_(self.actor.bias.data, 0) def forward(self, inputs): x = self.cnn_layer(inputs / 255.0) value = self.critic(x) pi = F.softmax(self.actor(x), dim=1) return value, pi
1,959
37.431373
88
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ddpg/utils.py
import numpy as np import torch # add ounoise here class ounoise(): def __init__(self, std, action_dim, mean=0, theta=0.15, dt=1e-2, x0=None): self.std = std self.mean = mean self.action_dim = action_dim self.theta = theta self.dt = dt self.x0 = x0 # reset the noise def reset(self): self.x_prev = self.x0 if self.x0 is not None else np.zeros(self.action_dim) # generate noise def noise(self): x = self.x_prev + self.theta * (self.mean - self.x_prev) * self.dt + \ self.std * np.sqrt(self.dt) * np.random.normal(size=self.action_dim) self.x_prev = x return x
686
27.625
84
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ddpg/demo.py
from arguments import get_args import gym from models import actor import torch import numpy as np def normalize(obs, mean, std, clip): return np.clip((obs - mean) / std, -clip, clip) if __name__ == '__main__': args = get_args() env = gym.make(args.env_name) # get environment infos obs_dims = env.observation_space.shape[0] action_dims = env.action_space.shape[0] action_max = env.action_space.high[0] # define the network actor_net = actor(obs_dims, action_dims) # load models model_path = args.save_dir + args.env_name + '/model.pt' model, mean, std = torch.load(model_path, map_location=lambda storage, loc: storage) # load models into the network actor_net.load_state_dict(model) for ep in range(10): obs = env.reset() reward_sum = 0 while True: env.render() with torch.no_grad(): norm_obs = normalize(obs, mean, std, args.clip_range) norm_obs_tensor = torch.tensor(norm_obs, dtype=torch.float32).unsqueeze(0) actions = actor_net(norm_obs_tensor) actions = actions.detach().numpy().squeeze() if action_dims == 1: actions = np.array([actions]) obs_, reward, done, _ = env.step(action_max * actions) reward_sum += reward if done: break obs = obs_ print('the episode is: {}, the reward is: {}'.format(ep, reward_sum)) env.close()
1,518
34.325581
90
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ddpg/ddpg_agent.py
import numpy as np from models import actor, critic import torch import os from datetime import datetime from mpi4py import MPI from rl_utils.mpi_utils.normalizer import normalizer from rl_utils.mpi_utils.utils import sync_networks, sync_grads from rl_utils.experience_replay.experience_replay import replay_buffer from utils import ounoise import copy import gym """ ddpg algorithms - revised baseline version support MPI training """ class ddpg_agent: def __init__(self, env, args): self.env = env self.args = args # get the dims and action max of the environment obs_dims = self.env.observation_space.shape[0] self.action_dims = self.env.action_space.shape[0] self.action_max = self.env.action_space.high[0] # define the network self.actor_net = actor(obs_dims, self.action_dims) self.critic_net = critic(obs_dims, self.action_dims) # sync the weights across the mpi sync_networks(self.actor_net) sync_networks(self.critic_net) # build the target newtork self.actor_target_net = copy.deepcopy(self.actor_net) self.critic_target_net = copy.deepcopy(self.critic_net) # create the optimizer self.actor_optim = torch.optim.Adam(self.actor_net.parameters(), self.args.lr_actor) self.critic_optim = torch.optim.Adam(self.critic_net.parameters(), self.args.lr_critic, weight_decay=self.args.critic_l2_reg) # create the replay buffer self.replay_buffer = replay_buffer(self.args.replay_size) # create the normalizer self.o_norm = normalizer(obs_dims, default_clip_range=self.args.clip_range) # create the noise generator self.noise_generator = ounoise(std=0.2, action_dim=self.action_dims) # create the dir to save models if MPI.COMM_WORLD.Get_rank() == 0: if not os.path.exists(self.args.save_dir): os.mkdir(self.args.save_dir) self.model_path = os.path.join(self.args.save_dir, self.args.env_name) if not os.path.exists(self.model_path): os.mkdir(self.model_path) # create a eval environemnt self.eval_env = gym.make(self.args.env_name) # set seeds self.eval_env.seed(self.args.seed * 2 + MPI.COMM_WORLD.Get_rank()) def learn(self): """ the learning part """ self.actor_net.train() # reset the environmenr firstly obs = self.env.reset() self.noise_generator.reset() # get the number of epochs nb_epochs = self.args.total_frames // (self.args.nb_rollout_steps * self.args.nb_cycles) for epoch in range(nb_epochs): for _ in range(self.args.nb_cycles): # used to update the normalizer ep_obs = [] for _ in range(self.args.nb_rollout_steps): with torch.no_grad(): inputs_tensor = self._preproc_inputs(obs) pi = self.actor_net(inputs_tensor) action = self._select_actions(pi) # feed actions into the environment obs_, reward, done, _ = self.env.step(self.action_max * action) # append the rollout information into the memory self.replay_buffer.add(obs, action, reward, obs_, float(done)) ep_obs.append(obs.copy()) obs = obs_ # if done, reset the environment if done: obs = self.env.reset() self.noise_generator.reset() # then start to do the update of the normalizer ep_obs = np.array(ep_obs) self.o_norm.update(ep_obs) self.o_norm.recompute_stats() # then start to update the network for _ in range(self.args.nb_train): a_loss, c_loss = self._update_network() # update the target network self._soft_update_target_network(self.actor_target_net, self.actor_net) self._soft_update_target_network(self.critic_target_net, self.critic_net) # start to do the evaluation success_rate = self._eval_agent() # convert back to normal self.actor_net.train() if epoch % self.args.display_interval == 0: if MPI.COMM_WORLD.Get_rank() == 0: print('[{}] Epoch: {} / {}, Frames: {}, Rewards: {:.3f}, Actor loss: {:.3f}, Critic Loss: {:.3f}'.format(datetime.now(), \ epoch, nb_epochs, (epoch+1) * self.args.nb_rollout_steps * self.args.nb_cycles, success_rate, a_loss, c_loss)) torch.save([self.actor_net.state_dict(), self.o_norm.mean, self.o_norm.std], self.model_path + '/model.pt') # functions to preprocess the image def _preproc_inputs(self, obs): obs_norm = self.o_norm.normalize(obs) inputs_tensor = torch.tensor(obs_norm, dtype=torch.float32).unsqueeze(0) return inputs_tensor # this function will choose action for the agent and do the exploration def _select_actions(self, pi): action = pi.cpu().numpy().squeeze() # TODO: Noise type now - only support ounoise # add the gaussian noise #action = action + np.random.normal(0, 0.1, self.action_dims) # add ou noise action = action + self.noise_generator.noise() action = np.clip(action, -1, 1) return action # update the network def _update_network(self): # sample the samples from the replay buffer samples = self.replay_buffer.sample(self.args.batch_size) obses, actions, rewards, obses_next, dones = samples # try to do the normalization of obses norm_obses = self.o_norm.normalize(obses) norm_obses_next = self.o_norm.normalize(obses_next) # transfer them into tensors norm_obses_tensor = torch.tensor(norm_obses, dtype=torch.float32) norm_obses_next_tensor = torch.tensor(norm_obses_next, dtype=torch.float32) actions_tensor = torch.tensor(actions, dtype=torch.float32) rewards_tensor = torch.tensor(rewards, dtype=torch.float32).unsqueeze(1) dones_tensor = torch.tensor(dones, dtype=torch.float32).unsqueeze(1) with torch.no_grad(): actions_next = self.actor_target_net(norm_obses_next_tensor) q_next_value = self.critic_target_net(norm_obses_next_tensor, actions_next) target_q_value = rewards_tensor + (1 - dones_tensor) * self.args.gamma * q_next_value # the real q value real_q_value = self.critic_net(norm_obses_tensor, actions_tensor) critic_loss = (real_q_value - target_q_value).pow(2).mean() # the actor loss actions_real = self.actor_net(norm_obses_tensor) actor_loss = -self.critic_net(norm_obses_tensor, actions_real).mean() # start to update the network self.actor_optim.zero_grad() actor_loss.backward() sync_grads(self.actor_net) self.actor_optim.step() # update the critic network self.critic_optim.zero_grad() critic_loss.backward() sync_grads(self.critic_net) self.critic_optim.step() return actor_loss.item(), critic_loss.item() # soft update the target network... def _soft_update_target_network(self, target, source): for target_param, param in zip(target.parameters(), source.parameters()): target_param.data.copy_((1 - self.args.polyak) * param.data + self.args.polyak * target_param.data) # do the evaluation def _eval_agent(self): self.actor_net.eval() total_success_rate = [] for _ in range(self.args.nb_test_rollouts): per_success_rate = [] obs = self.eval_env.reset() while True: with torch.no_grad(): inputs_tensor = self._preproc_inputs(obs) pi = self.actor_net(inputs_tensor) actions = pi.detach().cpu().numpy().squeeze() if self.action_dims == 1: actions = np.array([actions]) obs_, reward, done, _ = self.eval_env.step(actions * self.action_max) per_success_rate.append(reward) obs = obs_ if done: break total_success_rate.append(np.sum(per_success_rate)) local_success_rate = np.mean(total_success_rate) global_success_rate = MPI.COMM_WORLD.allreduce(local_success_rate, op=MPI.SUM) return global_success_rate / MPI.COMM_WORLD.Get_size()
8,833
45.494737
142
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ddpg/models.py
import torch import torch.nn as nn import torch.nn.functional as F # define the actor network class actor(nn.Module): def __init__(self, obs_dims, action_dims): super(actor, self).__init__() self.fc1 = nn.Linear(obs_dims, 400) self.fc2 = nn.Linear(400, 300) self.action_out = nn.Linear(300, action_dims) def forward(self, x): x = F.relu(self.fc1(x)) x = F.relu(self.fc2(x)) actions = torch.tanh(self.action_out(x)) return actions class critic(nn.Module): def __init__(self, obs_dims, action_dims): super(critic, self).__init__() self.fc1 = nn.Linear(obs_dims, 400) self.fc2 = nn.Linear(400 + action_dims, 300) self.q_out = nn.Linear(300, 1) def forward(self, x, actions): x = F.relu(self.fc1(x)) x = torch.cat([x, actions], dim=1) x = F.relu(self.fc2(x)) q_value = self.q_out(x) return q_value
950
28.71875
53
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ppo/utils.py
import numpy as np import torch from torch.distributions.normal import Normal from torch.distributions.beta import Beta from torch.distributions.categorical import Categorical import random def select_actions(pi, dist_type, env_type): if env_type == 'atari': actions = Categorical(pi).sample() else: if dist_type == 'gauss': mean, std = pi actions = Normal(mean, std).sample() elif dist_type == 'beta': alpha, beta = pi actions = Beta(alpha.detach().cpu(), beta.detach().cpu()).sample() # return actions return actions.detach().cpu().numpy().squeeze() def evaluate_actions(pi, actions, dist_type, env_type): if env_type == 'atari': cate_dist = Categorical(pi) log_prob = cate_dist.log_prob(actions).unsqueeze(-1) entropy = cate_dist.entropy().mean() else: if dist_type == 'gauss': mean, std = pi normal_dist = Normal(mean, std) log_prob = normal_dist.log_prob(actions).sum(dim=1, keepdim=True) entropy = normal_dist.entropy().mean() elif dist_type == 'beta': alpha, beta = pi beta_dist = Beta(alpha, beta) log_prob = beta_dist.log_prob(actions).sum(dim=1, keepdim=True) entropy = beta_dist.entropy().mean() return log_prob, entropy
1,370
35.078947
78
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ppo/demo.py
from arguments import get_args from models import cnn_net, mlp_net import torch import cv2 import numpy as np import gym from rl_utils.env_wrapper.frame_stack import VecFrameStack from rl_utils.env_wrapper.atari_wrapper import make_atari, wrap_deepmind # denormalize def normalize(x, mean, std, clip=10): x -= mean x /= (std + 1e-8) return np.clip(x, -clip, clip) # get tensors for the agent def get_tensors(obs, env_type, filters=None): if env_type == 'atari': tensor = torch.tensor(np.transpose(obs, (2, 0, 1)), dtype=torch.float32).unsqueeze(0) elif env_type == 'mujoco': tensor = torch.tensor(normalize(obs, filters.rs.mean, filters.rs.std), dtype=torch.float32).unsqueeze(0) return tensor if __name__ == '__main__': # get the arguments args = get_args() # create the environment if args.env_type == 'atari': env = make_atari(args.env_name) env = wrap_deepmind(env, frame_stack=True) elif args.env_type == 'mujoco': env = gym.make(args.env_name) # get the model path model_path = args.save_dir + args.env_name + '/model.pt' # create the network if args.env_type == 'atari': network = cnn_net(env.action_space.n) network.load_state_dict(torch.load(model_path, map_location=lambda storage, loc: storage)) filters = None elif args.env_type == 'mujoco': network = mlp_net(env.observation_space.shape[0], env.action_space.shape[0], args.dist) net_models, filters = torch.load(model_path, map_location=lambda storage, loc: storage) # load models network.load_state_dict(net_models) # start to play the demo obs = env.reset() reward_total = 0 # just one episode while True: env.render() with torch.no_grad(): obs_tensor = get_tensors(obs, args.env_type, filters) _, pi = network(obs_tensor) # get actions if args.env_type == 'atari': actions = torch.argmax(pi, dim=1).item() elif args.env_type == 'mujoco': if args.dist == 'gauss': mean, _ = pi actions = mean.numpy().squeeze() elif args.dist == 'beta': alpha, beta = pi actions = (alpha - 1) / (alpha + beta - 2) actions = actions.numpy().squeeze() actions = -1 + 2 * actions obs_, reward, done, _ = env.step(actions) reward_total += reward if done: break obs = obs_ print('the rewrads is: {}'.format(reward_total))
2,641
35.694444
112
py
reinforcement-learning-algorithms
reinforcement-learning-algorithms-master/rl_algorithms/ppo/models.py
import torch from torch import nn from torch.nn import functional as F """ this network also include gaussian distribution and beta distribution """ class mlp_net(nn.Module): def __init__(self, state_size, num_actions, dist_type): super(mlp_net, self).__init__() self.dist_type = dist_type self.fc1_v = nn.Linear(state_size, 64) self.fc2_v = nn.Linear(64, 64) self.fc1_a = nn.Linear(state_size, 64) self.fc2_a = nn.Linear(64, 64) # check the type of distribution if self.dist_type == 'gauss': self.sigma_log = nn.Parameter(torch.zeros(1, num_actions)) self.action_mean = nn.Linear(64, num_actions) self.action_mean.weight.data.mul_(0.1) self.action_mean.bias.data.zero_() elif self.dist_type == 'beta': self.action_alpha = nn.Linear(64, num_actions) self.action_beta = nn.Linear(64, num_actions) # init.. self.action_alpha.weight.data.mul_(0.1) self.action_alpha.bias.data.zero_() self.action_beta.weight.data.mul_(0.1) self.action_beta.bias.data.zero_() # define layers to output state value self.value = nn.Linear(64, 1) self.value.weight.data.mul_(0.1) self.value.bias.data.zero_() def forward(self, x): x_v = torch.tanh(self.fc1_v(x)) x_v = torch.tanh(self.fc2_v(x_v)) state_value = self.value(x_v) # output the policy... x_a = torch.tanh(self.fc1_a(x)) x_a = torch.tanh(self.fc2_a(x_a)) if self.dist_type == 'gauss': mean = self.action_mean(x_a) sigma_log = self.sigma_log.expand_as(mean) sigma = torch.exp(sigma_log) pi = (mean, sigma) elif self.dist_type == 'beta': alpha = F.softplus(self.action_alpha(x_a)) + 1 beta = F.softplus(self.action_beta(x_a)) + 1 pi = (alpha, beta) return state_value, pi # the convolution layer of deepmind class deepmind(nn.Module): def __init__(self): super(deepmind, self).__init__() self.conv1 = nn.Conv2d(4, 32, 8, stride=4) self.conv2 = nn.Conv2d(32, 64, 4, stride=2) self.conv3 = nn.Conv2d(64, 32, 3, stride=1) self.fc1 = nn.Linear(32 * 7 * 7, 512) # start to do the init... nn.init.orthogonal_(self.conv1.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.conv2.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.conv3.weight.data, gain=nn.init.calculate_gain('relu')) nn.init.orthogonal_(self.fc1.weight.data, gain=nn.init.calculate_gain('relu')) # init the bias... nn.init.constant_(self.conv1.bias.data, 0) nn.init.constant_(self.conv2.bias.data, 0) nn.init.constant_(self.conv3.bias.data, 0) nn.init.constant_(self.fc1.bias.data, 0) def forward(self, x): x = F.relu(self.conv1(x)) x = F.relu(self.conv2(x)) x = F.relu(self.conv3(x)) x = x.view(-1, 32 * 7 * 7) x = F.relu(self.fc1(x)) return x # in the initial, just the nature CNN class cnn_net(nn.Module): def __init__(self, num_actions): super(cnn_net, self).__init__() self.cnn_layer = deepmind() self.critic = nn.Linear(512, 1) self.actor = nn.Linear(512, num_actions) # init the linear layer.. nn.init.orthogonal_(self.critic.weight.data) nn.init.constant_(self.critic.bias.data, 0) # init the policy layer... nn.init.orthogonal_(self.actor.weight.data, gain=0.01) nn.init.constant_(self.actor.bias.data, 0) def forward(self, inputs): x = self.cnn_layer(inputs / 255.0) value = self.critic(x) pi = F.softmax(self.actor(x), dim=1) return value, pi
3,913
37
88
py