# Copyright (c) 2018-2020 Chris Choy (chrischoy@ai.stanford.edu). # # Permission is hereby granted, free of charge, to any person obtaining a copy of # this software and associated documentation files (the "Software"), to deal in # the Software without restriction, including without limitation the rights to # use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies # of the Software, and to permit persons to whom the Software is furnished to do # so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. # # Please cite "4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural # Networks", CVPR'19 (https://arxiv.org/abs/1904.08755) if you use any part # of the code. import os import glob import argparse import sklearn.metrics as metrics import numpy as np import GPUtil import torch import torch.nn as nn import torch.utils.data from torch.utils.data import DataLoader import torch.optim as optim import torch.nn.functional as F import MinkowskiEngine as ME from torch.utils.data import Dataset, DataLoader from examples.common import seed_all parser = argparse.ArgumentParser() parser.add_argument("--voxel_size", type=float, default=0.05) parser.add_argument("--max_steps", type=int, default=200000) parser.add_argument("--val_freq", type=int, default=20) parser.add_argument("--batch_size", default=1, type=int) parser.add_argument("--lr", default=1e-1, type=float) parser.add_argument("--weight_decay", type=float, default=1e-5) parser.add_argument("--num_workers", type=int, default=2) parser.add_argument("--stat_freq", type=int, default=20) parser.add_argument("--weights", type=str, default="modelnet.pth") parser.add_argument("--seed", type=int, default=777) parser.add_argument("--translation", type=float, default=0.2) parser.add_argument("--test_translation", type=float, default=0.0) parser.add_argument( "--network", type=str, choices=["pointnet", "minkpointnet", "minkfcnn"], default="minkfcnn", ) class clusterCNN(ME.MinkowskiNetwork): def __init__( self, in_channel, out_channel, embedding_channel=1024, channels = (64, 64, 128, 128, 256, 256, 256, 512, 512, 512, 512, 512, 512), D=2, ): ME.MinkowskiNetwork.__init__(self, D) self.network_initialization( in_channel, out_channel, channels = channels, embedding_channel=embedding_channel, kernel_size=3, D=D, ) self.weight_initialization() def get_mlp_block(self, in_channel, out_channel): return nn.Sequential( ME.MinkowskiLinear(in_channel, out_channel, bias=False), ME.MinkowskiBatchNorm(out_channel), ME.MinkowskiReLU(), ) def get_conv_block(self, in_channel, out_channel, kernel_size, stride): return nn.Sequential( ME.MinkowskiConvolution( in_channel, out_channel, kernel_size=kernel_size, stride=stride, dimension=self.D, ), ME.MinkowskiBatchNorm(out_channel), ME.MinkowskiReLU(), ) def network_initialization( self, in_channel, out_channel, channels, embedding_channel, kernel_size, D=2, ): self.conv1_1 = self.get_conv_block(in_channel, channels[0], kernel_size=kernel_size, stride=1) self.conv1_2 = self.get_conv_block(channels[0], channels[1], kernel_size=kernel_size, stride=1) self.pool1 = ME.MinkowskiMaxPooling(kernel_size=3, stride=2, dimension=D) self.conv2_1 = self.get_conv_block(channels[1], channels[2], kernel_size=kernel_size, stride=1) self.conv2_2 = self.get_conv_block(channels[2], channels[3], kernel_size=kernel_size, stride=1) self.pool2 = ME.MinkowskiMaxPooling(kernel_size=3, stride=2, dimension=D) self.conv3_1 = self.get_conv_block(channels[3], channels[4], kernel_size=kernel_size, stride=1) self.conv3_2 = self.get_conv_block(channels[4], channels[5], kernel_size=kernel_size, stride=1) self.conv3_3 = self.get_conv_block(channels[5], channels[6], kernel_size=kernel_size, stride=1) self.pool3 = ME.MinkowskiMaxPooling(kernel_size=3, stride=2, dimension=D) self.conv4_1 = self.get_conv_block(channels[6], channels[7], kernel_size=kernel_size, stride=1) self.conv4_2 = self.get_conv_block(channels[7], channels[8], kernel_size=kernel_size, stride=1) self.conv4_3 = self.get_conv_block(channels[8], channels[9], kernel_size=kernel_size, stride=1) self.pool4 = ME.MinkowskiMaxPooling(kernel_size=3, stride=2, dimension=D) self.conv5_1 = self.get_conv_block(channels[9], channels[10], kernel_size=kernel_size, stride=1) self.conv5_2 = self.get_conv_block(channels[10], channels[11], kernel_size=kernel_size, stride=1) self.conv5_3 = self.get_conv_block(channels[11], channels[12], kernel_size=kernel_size, stride=1) self.pool5 = ME.MinkowskiMaxPooling(kernel_size=3, stride=2, dimension=D) self.global_pool = ME.MinkowskiGlobalPooling() self.final = nn.Sequential( self.get_mlp_block(512, 512), ME.MinkowskiDropout(), self.get_mlp_block(512, 512), self.get_mlp_block(512, 2), #ME.MinkowskiFunctional.softmax(), ) def weight_initialization(self): for m in self.modules(): if isinstance(m, ME.MinkowskiConvolution): ME.utils.kaiming_normal_(m.kernel, mode="fan_out", nonlinearity="relu") if isinstance(m, ME.MinkowskiBatchNorm): nn.init.constant_(m.bn.weight, 1) nn.init.constant_(m.bn.bias, 0) def forward(self, x: ME.SparseTensor): x = self.conv1_1(x) x = self.conv1_2(x) x = self.pool1(x) x = self.conv2_1(x) x = self.conv2_2(x) x = self.pool2(x) x = self.conv3_1(x) x = self.conv3_2(x) x = self.conv3_3(x) x = self.pool3(x) x = self.conv4_1(x) x = self.conv4_2(x) x = self.conv4_3(x) x = self.pool4(x) x = self.conv5_1(x) x = self.conv5_2(x) x = self.conv5_3(x) x = self.pool5(x) x = self.global_pool(x) return self.final(x).F class nnbarDataset(Dataset): print(Dataset.__dict__) total_num = np.array([1974, 247,247]) input_size = np.array([[990, 124, 124], [984, 123, 123]]) def __init__( self, phase: str, data_root: str = "/data/yjwa/sparse_torch/MinkowskiEngine/nnbar_overlay/official_npys/dataloader_test", ): Dataset.__init__(self) self.phase = phase self.i_f, self.i_e = self.index_map(iter) self.w_xy, self.w_val, self.u_xy, self.u_val, self.v_xy, self.v_val, self.label = self.load_one_file(data_root, phase, self.i_f) def __len__(self): if self.phase == "train": return self.total_num[0] elif self.phase == "val": return total_num[1] else: return total_num[2] def index_map(self, i): idx_f = 0 if self.phase == "train": idx_e = self.total_num[0] while idx_e > self.input_size[idx_f, 0]: idx_e -= self.input_size[idx_f, 0] idx_f += 1 elif self.phase == "val": idx_e = total_num[1] while idx_e > input_size[idx_f, 1]: idx_e -= input_size[idx_f, 1] idx_f += 1 elif self.phase == "test": idx_e = total_num[2] while idx_e > input_size[idx_f, 2]: idx_e -= input_size[idx_f, 2] idx_f += 1 return idx_f, idx_e def load_one_file(self, data_root, phase, i_f): #print("In load_data") w_xy, w_val, labels = [], [], [] u_xy, u_val = [], [] v_xy, v_val = [], [] assert os.path.exists(data_root), f"{data_root} does not exist" files = glob.glob(os.path.join(data_root, "*_%s_bdt.npy" %phase)) #assert len(files) == 1, "no file or more than 1 file found" print(files) npy_name = files[i_f] #for npy_name in files: a=np.load(npy_name, allow_pickle=True) w_xy.extend(a[:,10:12]) w_val.extend(a[:,12]) u_xy.extend(a[:,13:15]) u_val.extend(a[:,15]) v_xy.extend(a[:,16:18]) v_val.extend(a[:,18]) #print("data extent") labels.extend(a[:,21].astype("int64")) #20 for trad, 21for bdt, to be fixed #print("labels extent") w_xy = np.stack(w_xy, axis=0) u_xy = np.stack(u_xy, axis=0) v_xy = np.stack(v_xy, axis=0) #w_val = np.stack(w_val, axis=0) #print(np.shape(w_xy), np.shape(w_val), np.shape(labels)) labels = np.stack(labels, axis=0) return w_xy, w_val, u_xy, u_val, v_xy, v_val, labels def __getitem__(self): w_xy = self.w_xy[self.i_e] w_val = self.w_val[self.i_e] u_xy = self.u_xy[self.i_e] u_val = self.u_val[self.i_e] v_xy = self.v_xy[self.i_e] v_val = self.v_val[self.i_e] label = self.label[self.i_e] w_val = torch.from_numpy(np.asarray(w_val)) u_val = torch.from_numpy(np.asarray(u_val)) v_val = torch.from_numpy(np.asarray(v_val)) padding = np.asarray(np.zeros((4096,))).astype("int64") w_x = np.array(w_xy[0]) w_x = np.concatenate((w_x,padding)) w_x = w_x[:4096] w_y = np.array(w_xy[1]) w_y = np.concatenate((w_y,padding)) w_y = w_y[:4096] u_x = np.array(u_xy[0]) u_x = np.concatenate((u_x,padding)) u_x = u_x[:4096] u_y = np.array(u_xy[1]) u_y = np.concatenate((u_y,padding)) u_y = u_y[:4096] v_x = np.array(v_xy[0]) v_x = np.concatenate((v_x,padding)) v_x = v_x[:4096] v_y = np.array(v_xy[1]) v_y = np.concatenate((v_y,padding)) v_y = v_y[:4096] x = np.concatenate((w_x, u_x, v_x)) y = np.concatenate((w_y, u_y, v_y)) #print(w_y) w_xy = np.dstack((w_x,w_y)) w_xy = np.reshape(w_xy, (4096,2)) w_xy = torch.from_numpy(np.asarray(w_xy)) w_val = np.concatenate((w_val, padding)) w_val = w_val[:4096] w_val = np.reshape(w_val, (4096,1)) w_val = torch.from_numpy(np.asarray(w_val)) u_xy = np.dstack((u_x,u_y)) u_xy = np.reshape(u_xy, (4096,2)) u_xy = torch.from_numpy(np.asarray(u_xy)) u_val = np.concatenate((u_val, padding)) u_val = u_val[:4096] u_val = np.reshape(u_val, (4096,1)) u_val = torch.from_numpy(np.asarray(u_val)) v_xy = np.dstack((v_x,v_y)) v_xy = np.reshape(v_xy, (4096,2)) v_xy = torch.from_numpy(np.asarray(v_xy)) v_val = np.concatenate((v_val, padding)) v_val = v_val[:4096] v_val = np.reshape(v_val, (4096,1)) v_val = torch.from_numpy(np.asarray(v_val)) xy = np.dstack((x,y)) xy = np.reshape(xy, (4096*3,2)) xy = torch.from_numpy(np.asarray(xy)) val = np.concatenate((w_val, u_val, v_val)) val = torch.from_numpy(np.asarray(val)) label = np.reshape(label, (1)) label = torch.from_numpy(np.asarray(label)) # return { # "coordinates": w_xy.to(torch.float32), # "feats": w_val.to(torch.float32), # "label": label, # } return xy.to(torch.float32), val.to(torch.float32), label # def __len__(self): # return self.w_xy.shape[0] def make_data_loader_custom(phase, config): assert phase in ["train", "val", "test"] is_train = phase == "train" dataset = nnbarDataset( phase=phase, ) #print(dataset) return DataLoader( dataset, num_workers=config.num_workers, shuffle = is_train, batch_size=config.batch_size, collate_fn=ME.utils.batch_sparse_collate, ) def criterion(pred, labels, smoothing=True): """ Calculate cross entropy loss, apply label smoothing if needed. """ labels = labels.contiguous().view(-1) if smoothing: eps = 0.2 n_class = pred.size(1) one_hot = torch.zeros_like(pred).scatter(1, labels.view(-1, 1), 1) one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1) log_prb = F.log_softmax(pred, dim=1) loss = -(one_hot * log_prb).sum(dim=1).mean() else: loss = F.cross_entropy(pred, labels, reduction="mean") return loss def val(net, device, config, phase="val"): is_minknet = isinstance(net, ME.MinkowskiNetwork) data_loader = make_data_loader_custom("val", config=config,) torch.cuda.set_device('cuda:1') net.eval() labels_val, preds_val = [], [] with torch.no_grad(): for batch in data_loader: coords, feats, labels = batch input = ME.SparseTensor(feats.float(), coords, device="cuda:1") logit = net(input) pred = torch.argmax(logit, 1) #print("val_labels", labels) #print("val_logit", logit) #print("val_pred", pred) #labels.append(labels.numpy()) labels_val.append(labels.cpu().numpy()) preds_val.append(pred.cpu().numpy()) torch.cuda.empty_cache() torch.cuda.empty_cache() return metrics.accuracy_score(np.concatenate(labels_val), np.concatenate(preds_val)) def test(net, device, config, phase="test"): torch.cuda.set_device('cuda:1') is_minknet = isinstance(net, ME.MinkowskiNetwork) data_loader = make_data_loader_custom("test", config=config,) net.eval() labels_val, preds_val, sfs_val = [], [], [] with torch.no_grad(): for batch in data_loader: coords, feats, labels = batch input = ME.SparseTensor(feats.float(), coords, device="cuda:1") logit = net(input) sf_val = torch.softmax(logit, dim=1) pred = torch.argmax(logit, 1) labels_val.append(labels.cpu().numpy()) preds_val.append(pred.cpu().numpy()) sfs_val.append(sf_val.cpu().numpy()) torch.cuda.empty_cache() return np.concatenate(labels_val), np.concatenate(preds_val), np.concatenate(sfs_val) #return metrics.accuracy_score(np.concatenate(labels_val), np.concatenate(preds_val)) def train(net, device, config): if torch.cuda.device_count() > 1: print("Let's use", torch.cuda.device_count(), "GPUs!") # dim = 0 [30, xxx] -> [10, ...], [10, ...], [10, ...] on 3 GPUs net = nn.DataParallel(net) net.to(device) is_minknet = isinstance(net, ME.MinkowskiNetwork) optimizer = optim.SGD( net.parameters(), lr=config.lr, momentum=0.9, weight_decay=config.weight_decay, ) scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=config.max_steps,) print(optimizer) print(scheduler) torch.cuda.set_device('cuda:1') GPUtil.showUtilization() train_iter = iter(make_data_loader_custom("train", config)) GPUtil.showUtilization() best_metric = 0 net.train() for i in range(config.max_steps): optimizer.zero_grad() try: data = train_iter.next() except StopIteration: train_iter = iter(make_data_loader_custom("train", config)) data = train_iter.next() coords, feats, labels = data input = ME.SparseTensor(feats.float(), coords, device="cuda:1") logit = net(input) loss = criterion(logit, labels.to(device)) loss.backward() optimizer.step() scheduler.step() torch.cuda.empty_cache() if i % config.stat_freq == 0: print(f"Iter: {i}, Loss: {loss.item():.3e}") if i % config.val_freq == 0 and i > 0: torch.save( { "state_dict": net.state_dict(), "optimizer": optimizer.state_dict(), "scheduler": scheduler.state_dict(), "curr_iter": i, }, config.weights, ) accuracy = val(net, device, config, phase="val") if best_metric < accuracy: best_metric = accuracy print(f"Validation accuracy: {accuracy}. Best accuracy: {best_metric}") net.train() if __name__ == "__main__": config = parser.parse_args() seed_all(config.seed) device = torch.device("cuda:1" if torch.cuda.is_available() else "cpu")# changed "cuda" to "cuda:0" print("===================ModelNet40 Dataset===================") print(f"Training with translation {config.translation}") print(f"Evaluating with translation {config.test_translation}") print("=============================================\n\n") net = clusterCNN( in_channel=1, out_channel=2,).to(device) print("===================Network===================") print(net) print("=============================================\n\n") train(net, device, config) y_label, y_pred, y_sf = test(net, device, config, phase="test") accuracy = metrics.accuracy_score(y_label, y_pred) print(f"Test accuracy: {accuracy}") dim0 = len(y_label) y_label = np.reshape(y_label,(dim0,1)) y_pred = np.reshape(y_pred,(dim0,1)) y_sf = np.reshape(y_sf,(dim0,2)) y_combine = np.hstack((y_label, y_pred, y_sf)) #y_pred_sf = torch.softmax(torch.from_numpy(y_pred), dim=2) np.savetxt('label_pred_sf_all_planes_2021aug.csv',(y_combine), delimiter=' ')