kissing-detector/train.py

# adapted from PyTorch tutorials
import copy
import time
from typing import List, Tuple, Optional

import torch
import torch.optim as optim
from torch import nn

from data import AudioVideo, AudioVideo3D
from kissing_detector import KissingDetector, KissingDetector3DConv

ExperimentResults = Tuple[Optional[nn.Module], List[float], List[float]]


def _get_params_to_update(model: nn.Module,
                          feature_extract: bool) -> List[nn.parameter.Parameter]:
    params_to_update = model.parameters()
    if feature_extract:
        print('Params to update')
        params_to_update = []
        for name, param in model.named_parameters():
            if param.requires_grad is True:
                params_to_update.append(param)
                print("*", name)
    else:
        print('Updating ALL params')
    return params_to_update


def train_kd(data_path_base: str,
             conv_model_name: Optional[str],
             num_epochs: int,
             feature_extract: bool,
             batch_size: int,
             use_vggish: bool = True,
             num_workers: int = 4,
             shuffle: bool = True,
             lr: float = 0.001,
             momentum: float = 0.9,
             use_3d: bool = False) -> ExperimentResults:
    num_classes = 2
    try:
        if use_3d:
            kd = KissingDetector3DConv(num_classes, feature_extract, use_vggish)
        else:
            kd = KissingDetector(conv_model_name, num_classes, feature_extract, use_vggish=use_vggish)
    except ValueError:
        # if the combination is not valid
        return None, [-1.0], [-1.0]

    params_to_update = _get_params_to_update(kd, feature_extract)

    av = AudioVideo3D if use_3d else AudioVideo

    datasets = {set_: av(f'{data_path_base}/{set_}') for set_ in ['train', 'val']}
    dataloaders_dict = {x: torch.utils.data.DataLoader(datasets[x],
                                                       batch_size=batch_size,
                                                       shuffle=shuffle, num_workers=num_workers)
                        for x in ['train', 'val']}
    optimizer_ft = optim.SGD(params_to_update, lr=lr, momentum=momentum)

    # Setup the loss fxn
    criterion = nn.CrossEntropyLoss()

    return train_model(kd,
                       dataloaders_dict, criterion, optimizer_ft, num_epochs=num_epochs,
                       is_inception=(conv_model_name == "inception"))


def train_model(model, dataloaders, criterion, optimizer, num_epochs=25, is_inception=False):
    since = time.time()

    val_acc_history = []
    val_f1_history = []

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0
    best_f1 = 0.0

    # Detect if we have a GPU available
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

    for epoch in range(num_epochs):
        print('Epoch {}/{}'.format(epoch, num_epochs - 1))
        print('-' * 10)

        # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # Set model to training mode
            else:
                model.eval()  # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0
            running_tp = 0
            running_fp = 0
            running_fn = 0

            # Iterate over data.
            for a, v, labels in dataloaders[phase]:
                a = a.to(device)
                v = v.to(device)
                labels = labels.to(device)

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    # Get model outputs and calculate loss
                    # Special case for inception because in training it has an auxiliary output. In train
                    #   mode we calculate the loss by summing the final output and the auxiliary output
                    #   but in testing we only consider the final output.
                    if is_inception and phase == 'train':
                        # https://discuss.pytorch.org/t/how-to-optimize-inception-model-with-auxiliary-classifiers/7958
                        outputs, aux_outputs = model(a, v)
                        loss1 = criterion(outputs, labels)
                        loss2 = criterion(aux_outputs, labels)
                        loss = loss1 + 0.4 * loss2
                    else:
                        outputs = model(a, v)
                        loss = criterion(outputs, labels)

                    _, preds = torch.max(outputs, 1)

                    # backward + optimize only if in training phase
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # statistics
                running_loss += loss.item() * a.size(0)
                running_corrects += torch.sum(preds == labels.data)
                running_tp += torch.sum((preds == labels.data)[labels.data == 1])
                running_fp += torch.sum((preds != labels.data)[labels.data == 1])
                running_fn += torch.sum((preds != labels.data)[labels.data == 0])

            epoch_loss = running_loss / len(dataloaders[phase].dataset)
            n = len(dataloaders[phase].dataset)
            epoch_acc = running_corrects.double() / n
            tp = running_tp.double()
            fp = running_fp.double()
            fn = running_fn.double()
            p = tp / (tp + fp)
            r = tp / (tp + fn)
            epoch_f1 = 2 * p * r / (p + r)

            print('{} Loss: {:.4f} F1: {:.4f} Acc: {:.4f}'.format(phase, epoch_loss, epoch_f1, epoch_acc))

            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
            if phase == 'val' and epoch_f1 > best_f1:
                best_f1 = epoch_f1
                best_model_wts = copy.deepcopy(model.state_dict())
            if phase == 'val':
                val_acc_history.append(float(epoch_acc))
                val_f1_history.append(float(epoch_f1))

        print()

    time_elapsed = time.time() - since
    print('Training complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
    print('Best val F1  : {:4f}'.format(best_f1))
    print('Best val Acc : {:4f}'.format(best_acc))

    # load best model weights
    model.load_state_dict(best_model_wts)
    return model, val_acc_history, val_f1_history