ft_classify.py

"""Finetune 3D CNN."""
import os
import argparse
import time
import random

import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from torch.utils.data import DataLoader, random_split
from torchvision import transforms
import torch.optim as optim
from tensorboardX import SummaryWriter

from datasets.ucf101 import UCF101Dataset
from models.c3d import C3D
from models.r3d import R3DNet
from models.r21d import R2Plus1DNet

import ast

def load_pretrained_weights(ckpt_path):
    """load pretrained weights and adjust params name."""
    adjusted_weights = {}
    pretrained_weights = torch.load(ckpt_path)
    for name, params in pretrained_weights.items():
        if 'base_network' in name:
            name = name[name.find('.')+1:]
            adjusted_weights[name] = params
            #print('Pretrained weight name: [{}]'.format(name))
    return adjusted_weights


def diff(x):
    shift_x = torch.roll(x,1,2)
    return ((shift_x -x) + 1)/2


def train(args, model, criterion, optimizer, device, train_dataloader, writer, epoch):
    torch.set_grad_enabled(True)
    model.train()

    running_loss = 0.0
    correct = 0
    for i, data in enumerate(train_dataloader, 1):
        # get inputs
        clips, idxs = data
        inputs = clips.to(device)
        targets = idxs.to(device)
        # zero the parameter gradients
        optimizer.zero_grad()
        # forward and backward
        if args.modality == 'res':
            outputs = model(diff(inputs))
        else:
            outputs = model(inputs)
        #outputs = model(inputs) # return logits here
        loss = criterion(outputs, targets)
        loss.backward()
        optimizer.step()
        # compute loss and acc
        running_loss += loss.item()
        pts = torch.argmax(outputs, dim=1)
        correct += torch.sum(targets == pts).item()
        # print statistics and write summary every N batch
        if i % args.pf == 0:
            avg_loss = running_loss / args.pf
            avg_acc = correct / (args.pf * args.bs)
            print('Train epoch: [{0:3d}/{1:3d}][{2:4d}/{3:4d}]\t'
                'loss: {4:.3f}, acc: {5:.3f}, lr: {lr}'.format(
                epoch, args.epochs, i + 1, len(train_dataloader),
                avg_loss, avg_acc, lr=optimizer.param_groups[0]['lr']), end='\r')
            step = (epoch-1)*len(train_dataloader) + i
            writer.add_scalar('train/CrossEntropyLoss', avg_loss, step)
            writer.add_scalar('train/Accuracy', avg_acc, step)
            running_loss = 0.0
            correct = 0
    print('')
    # summary params and grads per eopch
    for name, param in model.named_parameters():
        writer.add_histogram('params/{}'.format(name), param, epoch)
        writer.add_histogram('grads/{}'.format(name), param.grad, epoch)


def validate(args, model, criterion, device, val_dataloader, writer, epoch):
    torch.set_grad_enabled(False)
    model.eval()
    
    total_loss = 0.0
    correct = 0
    for i, data in enumerate(val_dataloader):
        # get inputs
        clips, idxs = data
        inputs = clips.to(device)
        targets = idxs.to(device)
        # forward
        if args.modality == 'res':
            outputs = model(diff(inputs))
        else:
            outputs = model(inputs)
        #outputs = model(inputs) # return logits here
        loss = criterion(outputs, targets)
        # compute loss and acc
        total_loss += loss.item()
        pts = torch.argmax(outputs, dim=1)
        correct += torch.sum(targets == pts).item()
        # print('correct: {}, {}, {}'.format(correct, targets, pts))
    avg_loss = total_loss / len(val_dataloader)
    avg_acc = correct / len(val_dataloader.dataset)
    writer.add_scalar('val/CrossEntropyLoss', avg_loss, epoch)
    writer.add_scalar('val/Accuracy', avg_acc, epoch)
    print('[VAL] loss: {:.3f}, acc: {:.3f}'.format(avg_loss, avg_acc))
    return avg_loss


def parse_args():
    parser = argparse.ArgumentParser(description='Finetune 3D CNN from VCOP pretrained weights')
    parser.add_argument('--cl', type=int, default=16, help='clip length')
    parser.add_argument('--model', type=str, default='r3d', help='c3d/r3d/r21d')
    parser.add_argument('--dataset', type=str, default='ucf101', help='ucf101/hmdb51')
    parser.add_argument('--split', type=str, default='1', help='dataset split')
    parser.add_argument('--gpu', type=int, default=0, help='GPU id')
    parser.add_argument('--lr', type=float, default=1e-3, help='learning rate')
    parser.add_argument('--ft_lr', type=float, default=1e-3, help='finetune learning rate')
    parser.add_argument('--momentum', type=float, default=9e-1, help='momentum')
    parser.add_argument('--wd', type=float, default=5e-4, help='weight decay')
    parser.add_argument('--log', type=str, default='logs', help='log directory')
    parser.add_argument('--ckpt', type=str, help='checkpoint path')
    parser.add_argument('--desp', type=str, help='additional description')
    parser.add_argument('--epochs', type=int, default=150, help='number of total epochs to run')
    parser.add_argument('--start-epoch', type=int, default=1, help='manual epoch number (useful on restarts)')
    parser.add_argument('--bs', type=int, default=8, help='mini-batch size')
    parser.add_argument('--workers', type=int, default=8, help='number of data loading workers')
    parser.add_argument('--pf', type=int, default=10, help='print frequency every batch')
    parser.add_argument('--seed', type=int, default=632, help='seed for initializing training.')
    parser.add_argument('--modality', default='rgb', type=str, help='currently support [rgb, res]')
    args = parser.parse_args()
    return args


if __name__ == '__main__':
    args = parse_args()
    print(vars(args))

    torch.backends.cudnn.benchmark = True
    # Force the pytorch to create context on the specific device 
    #os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu)
    device = torch.device('cuda:0' if torch.cuda.is_available() else "cpu")

    if args.seed:
        random.seed(args.seed)
        np.random.seed(args.seed)
        torch.manual_seed(args.seed)
        if args.gpu:
            torch.cuda.manual_seed_all(args.seed)

    ########### model ##############
    if args.dataset == 'ucf101':
        class_num = 101
    elif args.dataset == 'hmdb51':
        class_num = 51

    if args.model == 'c3d':
        model = C3D(with_classifier=True, num_classes=class_num).to(device)
    elif args.model == 'r3d':
        model = R3DNet(layer_sizes=(1,1,1,1), with_classifier=True, num_classes=class_num).to(device)
    elif args.model == 'r21d':   
        model = R2Plus1DNet(layer_sizes=(1,1,1,1), with_classifier=True, num_classes=class_num).to(device)
    pretrained_weights = load_pretrained_weights(args.ckpt)
    model.load_state_dict(pretrained_weights, strict=False)

    if args.desp:
        exp_name = '{}_{}_cls_cl{}_{}_{}'.format(args.model, args.modality, args.cl, args.desp, time.strftime('%m%d'))
    else:
        exp_name = '{}_{}_cls_cl{}_{}'.format(args.model, args.modality, args.cl, time.strftime('%m%d'))
    print(exp_name)
    log_dir = os.path.join(args.log, exp_name)
    writer = SummaryWriter(log_dir)

    train_transforms = transforms.Compose([
        transforms.Resize((128, 171)),
        transforms.RandomCrop(112),
        transforms.ToTensor()
    ])

    if args.dataset == 'ucf101':
        train_dataset = UCF101Dataset('data', args.cl, args.split, True, False, train_transforms)
        val_size = 800
        #val_dataset = UCF101Dataset('data', args.cl, args.split, False, True, train_transforms)

    # split val for 800 videos
    train_dataset, val_dataset = random_split(train_dataset, (len(train_dataset)-val_size, val_size))
    print('TRAIN video number: {}, VAL video number: {}.'.format(len(train_dataset), len(val_dataset)))
    train_dataloader = DataLoader(train_dataset, batch_size=args.bs, shuffle=True,
                                num_workers=args.workers, pin_memory=True)
    val_dataloader = DataLoader(val_dataset, batch_size=args.bs, shuffle=False,
                                num_workers=args.workers, pin_memory=True)

    # save graph and clips_order samples
    for data in train_dataloader:
        clips, idxs = data
        writer.add_video('train/clips', clips, 0, fps=8)
        writer.add_text('train/idxs', str(idxs.tolist()), 0)
        clips = clips.to(device)
        writer.add_graph(model, clips)
        break
    # save init params at step 0
    for name, param in model.named_parameters():
        writer.add_histogram('params/{}'.format(name), param, 0)

    ### loss funciton, optimizer and scheduler ###
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD([
        {'params': [param for name, param in model.named_parameters() if 'linear' not in name and 'conv5' not in name and 'conv4' not in name]},
        {'params': [param for name, param in model.named_parameters() if 'linear' in name or 'conv5' in name or 'conv4' in name], 'lr': args.ft_lr}],
        lr=args.lr, momentum=args.momentum, weight_decay=args.wd)
    scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', min_lr=1e-5, patience=30, factor=0.1)

    prev_best_val_loss = float('inf')
    prev_best_model_path = None
    for epoch in range(args.start_epoch, args.epochs):
        time_start = time.time()
        train(args, model, criterion, optimizer, device, train_dataloader, writer, epoch)
        print('Epoch time: {:.2f} s.'.format(time.time() - time_start))
        val_loss = validate(args, model, criterion, device, val_dataloader, writer, epoch)
        scheduler.step(val_loss)
        writer.add_scalar('train/lr', optimizer.param_groups[0]['lr'], epoch)
        writer.add_scalar('train/ft_lr', optimizer.param_groups[1]['lr'], epoch)
        # save model every 20 epoches
        if epoch % 20 == 0:
            torch.save(model.state_dict(), os.path.join(log_dir, 'model_{}.pt'.format(epoch)))
        # save model for the best val
        if val_loss < prev_best_val_loss:
            model_path = os.path.join(log_dir, 'best_model_{}.pt'.format(epoch))
            torch.save(model.state_dict(), model_path)
            prev_best_val_loss = val_loss
            if prev_best_model_path:
                os.remove(prev_best_model_path)
            prev_best_model_path = model_path