train_p4.py

from __future__ import print_function
import datetime
import os
import time
import sys
import numpy as np
import torch
import torch.utils.data
from torch.utils.data.dataloader import default_collate
from torch import nn
import torch.nn.functional as F
import torchvision
from torchvision import transforms

import utils

from scheduler import WarmupMultiStepLR

from datasets.AS_base import SegDataset
import models.AS_p4_base as Models


def get_labels_start_end_time(frame_wise_labels, bg_class=["background"]):
    labels = []
    starts = []
    ends = []
    last_label = frame_wise_labels[0]
    if frame_wise_labels[0] not in bg_class:
        labels.append(frame_wise_labels[0])
        starts.append(0)
    for i in range(len(frame_wise_labels)):
        if frame_wise_labels[i] != last_label:
            if frame_wise_labels[i] not in bg_class:
                labels.append(frame_wise_labels[i])
                starts.append(i)
            if last_label not in bg_class:
                ends.append(i)
            last_label = frame_wise_labels[i]
    if last_label not in bg_class:
        ends.append(i)
    return labels, starts, ends
 
 
def levenstein(p, y, norm=False):
    m_row = len(p)    
    n_col = len(y)
    D = np.zeros([m_row+1, n_col+1], np.float)
    for i in range(m_row+1):
        D[i, 0] = i
    for i in range(n_col+1):
        D[0, i] = i
 
    for j in range(1, n_col+1):
        for i in range(1, m_row+1):
            if y[j-1] == p[i-1]:
                D[i, j] = D[i-1, j-1]
            else:
                D[i, j] = min(D[i-1, j] + 1,
                              D[i, j-1] + 1,
                              D[i-1, j-1] + 1)
     
    if norm:
        score = (1 - D[-1, -1]/max(m_row, n_col)) * 100
    else:
        score = D[-1, -1]
 
    return score
 
 
def edit_score(recognized, ground_truth, norm=True, bg_class=["background"]):
    P, _, _ = get_labels_start_end_time(recognized, bg_class)
    Y, _, _ = get_labels_start_end_time(ground_truth, bg_class)
    return levenstein(P, Y, norm)
 
 
def f_score(recognized, ground_truth, overlap, bg_class=["background"]):
    p_label, p_start, p_end = get_labels_start_end_time(recognized, bg_class)
    y_label, y_start, y_end = get_labels_start_end_time(ground_truth, bg_class)
 
    tp = 0
    fp = 0
 
    hits = np.zeros(len(y_label))
 
    for j in range(len(p_label)):
        intersection = np.minimum(p_end[j], y_end) - np.maximum(p_start[j], y_start)
        union = np.maximum(p_end[j], y_end) - np.minimum(p_start[j], y_start)
        IoU = (1.0*intersection / union)*([p_label[j] == y_label[x] for x in range(len(y_label))])
        # Get the best scoring segment
        idx = np.array(IoU).argmax()
 
        if IoU[idx] >= overlap and not hits[idx]:
            tp += 1
            hits[idx] = 1
        else:
            fp += 1
    fn = len(y_label) - sum(hits)
    return float(tp), float(fp), float(fn)

def train_one_epoch(model, criterion, optimizer, lr_scheduler, data_loader, device, epoch, print_freq):
    model.train()
    metric_logger = utils.MetricLogger(delimiter="  ")
    metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value}'))
    metric_logger.add_meter('clips/s', utils.SmoothedValue(window_size=10, fmt='{value:.3f}'))

    header = 'Epoch: [{}]'.format(epoch)
    for clip, target in metric_logger.log_every(data_loader, print_freq, header):
        start_time = time.time()
        clip, target = clip.to(device), target.to(device)
        output = model(clip)
        loss = criterion(output.permute(0,2,1), target)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        output = torch.max(output,dim=-1)[1]
        # acc = torch.mean(torch.tensor(output==target,dtype=torch.float))
        output, target = output.cpu().numpy().astype(np.int32), target.cpu().numpy().astype(np.int32)
        acc = np.mean(output == target)

        batch_size = clip.shape[0]
        metric_logger.update(loss=loss.item(), lr=optimizer.param_groups[0]["lr"])
        metric_logger.meters['acc'].update(acc.item(), n=batch_size)
        metric_logger.meters['clips/s'].update(batch_size / (time.time() - start_time))
        lr_scheduler.step()
        sys.stdout.flush()

def evaluate(model, criterion, data_loader, device, len_test):
    model.eval()
    metric_logger = utils.MetricLogger(delimiter="  ")
    header = 'Test:'
    acc_list = []

    total_correct_class = [0] * 19
    total_class = [0] * 19

    with torch.no_grad():
        overlap = [.1, .25, .5]
        tp, fp, fn = np.zeros(3), np.zeros(3), np.zeros(3)
        edit = 0
        length = 0
        for clip, target in metric_logger.log_every(data_loader, 20, header):
            clip = clip.to(device, non_blocking=True)
            target = target.to(device, non_blocking=True)
            output = model(clip)
            loss = criterion(output.permute(0,2,1), target)

            output = torch.max(output,dim=-1)[1]
            output, target = output.cpu().numpy().astype(np.int32), target.cpu().numpy().astype(np.int32)
            acc = np.mean(output == target)
            # acc = torch.mean(torch.tensor(output==target,dtype=torch.float))
            acc_list.append(acc)
            for b in range(output.shape[0]):
                # print(output[b].shape)
                # print(target[b].shape)
                edit += edit_score(output[b], target[b])
            for b in range(output.shape[0]):
                for s in range(len(overlap)):
                    tp1, fp1, fn1 = f_score(output[b], target[b], overlap[s])
                    tp[s] += tp1
                    fp[s] += fp1
                    fn[s] += fn1


            batch_size = clip.shape[0]
            metric_logger.update(loss=loss.item())
            metric_logger.meters['acc'].update(acc.item(), n=batch_size)

    metric_logger.synchronize_between_processes()
    total_acc = np.mean((np.array(acc_list)))
    edit = (1.0 * edit) / len_test
    print('Edit: %.4f' % (edit))
    f1s = np.array([0, 0 ,0], dtype=float)
    for s in range(len(overlap)):
        precision = tp[s] / float(tp[s] + fp[s])
        recall = tp[s] / float(tp[s] + fn[s])
 
        f1 = 2.0 * (precision * recall) / (precision + recall)
 
        f1 = np.nan_to_num(f1) * 100
        print('F1@%0.2f: %.4f' % (overlap[s], f1))
        f1s[s] = f1
    print("total acc:", total_acc)
    return total_acc

def main(args):

    if args.output_dir:
        utils.mkdir(args.output_dir)

    print(args)
    print("torch version: ", torch.__version__)
    print("torchvision version: ", torchvision.__version__)

    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    torch.cuda.manual_seed(args.seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

    device = torch.device('cuda')

    # Data loading code
    print("Loading data")

    st = time.time()

    dataset = SegDataset(root='/datasets/AS_data_base', train=True)

    dataset_test = SegDataset(root='/datasets/AS_data_base', train=False)

    print("Creating data loaders")

    data_loader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True)
    data_loader_test = torch.utils.data.DataLoader(dataset_test, batch_size=args.batch_size, num_workers=args.workers, pin_memory=True)
    
    print("Creating model")
    Model = getattr(Models, args.model)
    model = Model(radius=args.radius, nsamples=args.nsamples, spatial_stride=args.spatial_stride,
                  temporal_kernel_size=args.temporal_kernel_size, temporal_stride=args.temporal_stride,
                  emb_relu=args.emb_relu,
                  dim=args.dim, depth=args.depth, heads=args.heads, dim_head=args.dim_head,
                  mlp_dim=args.mlp_dim, num_classes=19)

    if torch.cuda.device_count() > 1:
        model = nn.DataParallel(model)
    model.to(device)

    if args.resume:
        checkpoint = torch.load(args.resume, map_location='cpu')
        pre_state = checkpoint['model'] 
        # for name in pre_state.keys():
        #     print(name)
        update_dict = {k: v for k, v in pre_state.items() if k.startswith("module.tube_embedding.") or k.startswith("module.transformer1.") or k.startswith("module.pos")}
        # update_dict = {k: v for k, v in pre_state.items() if k.startswith("tube_embedding.") or k.startswith("transformer1.") or k.startswith("transformer2.") or k.startswith("pos")}  
        for name in update_dict.keys():
            print(name)
        net_state_dict = model.state_dict()
        # for name in net_state_dict.keys():
        #     print(name)
        net_state_dict.update(update_dict)
        model.load_state_dict(net_state_dict)
        # print(pre_state['transformer1.layers.1.0.fn.fn.to_qkv.weight'])
        # print(model.state_dict()['transformer1.layers.1.0.fn.fn.to_qkv.weight'])

    criterion = nn.CrossEntropyLoss()

    lr = args.lr
    optimizer = torch.optim.SGD(model.parameters(), lr=lr, momentum=args.momentum, weight_decay=args.weight_decay)

    # convert scheduler to be per iteration, not per epoch, for warmup that lasts
    # between different epochs
    warmup_iters = args.lr_warmup_epochs * len(data_loader)
    lr_milestones = [len(data_loader) * m for m in args.lr_milestones]
    lr_scheduler = WarmupMultiStepLR(optimizer, milestones=lr_milestones, gamma=args.lr_gamma, warmup_iters=warmup_iters, warmup_factor=1e-5)

    # model_without_ddp = model

    print("Start training")
    start_time = time.time()
    acc = 0
    for epoch in range(args.start_epoch, args.epochs):
        train_one_epoch(model, criterion, optimizer, lr_scheduler, data_loader, device, epoch, args.print_freq)

        acc = max(acc, evaluate(model, criterion, data_loader_test, device, len(dataset_test)))

def parse_args():
    import argparse
    parser = argparse.ArgumentParser(description='P4Transformer Model Training')

    parser.add_argument('--seed', default=0, type=int, help='random seed')
    parser.add_argument('--model', default='P4Transformer', type=str, help='model')
    # P4D
    parser.add_argument('--radius', default=0.9, type=float, help='radius for the ball query')
    parser.add_argument('--nsamples', default=32, type=int, help='number of neighbors for the ball query')
    parser.add_argument('--spatial-stride', default=32, type=int, help='spatial subsampling rate')
    parser.add_argument('--temporal-kernel-size', default=3, type=int, help='temporal kernel size')
    parser.add_argument('--temporal-stride', default=1, type=int, help='temporal stride')
    # embedding
    parser.add_argument('--emb-relu', default=False, action='store_true')
    # transformer
    parser.add_argument('--dim', default=2048, type=int, help='transformer dim')
    parser.add_argument('--depth', default=5, type=int, help='transformer depth')
    parser.add_argument('--heads', default=8, type=int, help='transformer head')
    parser.add_argument('--dim-head', default=128, type=int, help='transformer dim for each head')
    parser.add_argument('--mlp-dim', default=1024, type=int, help='transformer mlp dim')
    # training
    parser.add_argument('-b', '--batch-size', default=8, type=int)
    parser.add_argument('--epochs', default=50, type=int, metavar='N', help='number of total epochs to run')
    parser.add_argument('-j', '--workers', default=10, type=int, metavar='N', help='number of data loading workers (default: 16)')
    parser.add_argument('--lr', default=0.05, type=float, help='initial learning rate')
    parser.add_argument('--momentum', default=0.9, type=float, metavar='M', help='momentum')
    parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float, metavar='W', help='weight decay (default: 1e-4)', dest='weight_decay')
    parser.add_argument('--lr-milestones', nargs='+', default=[20, 35], type=int, help='decrease lr on milestones')
    parser.add_argument('--lr-gamma', default=0.5, type=float, help='decrease lr by a factor of lr-gamma')
    parser.add_argument('--lr-warmup-epochs', default=5, type=int, help='number of warmup epochs')
    # output
    parser.add_argument('--print-freq', default=20, type=int, help='print frequency')
    parser.add_argument('--output-dir', default='', type=str, help='path where to save')
    # resume
    parser.add_argument('--resume', default='', help='resume from checkpoint')
    parser.add_argument('--start-epoch', default=0, type=int, metavar='N', help='start epoch')

    args = parser.parse_args()

    return args

if __name__ == "__main__":
    args = parse_args()
    main(args)