def eval_epoch(val_loader, model, val_meter, cur_epoch, cfg, writer=None):
"""
Evaluate the model on the val set.
Args:
val_loader (loader): data loader to provide validation data.
model (model): model to evaluate the performance.
val_meter (ValMeter): meter instance to record and calculate the metrics.
cur_epoch (int): number of the current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
writer (TensorboardWriter, optional): TensorboardWriter object
to writer Tensorboard log.
"""
# Evaluation mode enabled. The running stats would not be updated.
model.eval()
val_meter.iter_tic()
for cur_iter, (inputs, labels, _, meta) in enumerate(val_loader):
if cfg.NUM_GPUS:
# Transferthe data to the current GPU device.
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list, )):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
val_meter.data_toc()
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
ori_boxes = meta["ori_boxes"]
metadata = meta["metadata"]
if cfg.NUM_GPUS:
preds = preds.cpu()
ori_boxes = ori_boxes.cpu()
metadata = metadata.cpu()
if cfg.NUM_GPUS > 1:
preds = torch.cat(du.all_gather_unaligned(preds), dim=0)
ori_boxes = torch.cat(du.all_gather_unaligned(ori_boxes),
dim=0)
metadata = torch.cat(du.all_gather_unaligned(metadata), dim=0)
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(preds, ori_boxes, metadata)
else:
preds = model(inputs)
if cfg.DATA.MULTI_LABEL:
if cfg.NUM_GPUS > 1:
preds, labels = du.all_gather([preds, labels])
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(
preds, labels, (1, 5))
# Combine the errors across the GPUs.
top1_err, top5_err = [(1.0 - x / preds.size(0)) * 100.0
for x in num_topks_correct]
if cfg.NUM_GPUS > 1:
top1_err, top5_err = du.all_reduce([top1_err, top5_err])
# Copy the errors from GPU to CPU (sync point).
top1_err, top5_err = top1_err.item(), top5_err.item()
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(
top1_err,
top5_err,
inputs[0].size(0) * max(
cfg.NUM_GPUS, 1
), # If running on CPU (cfg.NUM_GPUS == 1), use 1 to represent 1 CPU.
)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Val/Top1_err": top1_err,
"Val/Top5_err": top5_err
},
global_step=len(val_loader) * cur_epoch + cur_iter,
)
val_meter.update_predictions(preds, labels)
val_meter.log_iter_stats(cur_epoch, cur_iter)
val_meter.iter_tic()
# Log epoch stats.
val_meter.log_epoch_stats(cur_epoch)
# write to tensorboard format if available.
if writer is not None:
if cfg.DETECTION.ENABLE:
writer.add_scalars({"Val/mAP": val_meter.full_map},
global_step=cur_epoch)
else:
all_preds = [pred.clone().detach() for pred in val_meter.all_preds]
all_labels = [
label.clone().detach() for label in val_meter.all_labels
]
if cfg.NUM_GPUS:
all_preds = [pred.cpu() for pred in all_preds]
all_labels = [label.cpu() for label in all_labels]
writer.plot_eval(preds=all_preds,
labels=all_labels,
global_step=cur_epoch)
val_meter.reset()
def train_epoch(train_loader, model, optimizer, train_meter, cur_epoch, cfg):
"""
Perform the video training for one epoch.
Args:
train_loader (loader): video training loader.
model (model): the video model to train.
optimizer (optim): the optimizer to perform optimization on the model's
parameters.
train_meter (TrainMeter): training meters to log the training performance.
cur_epoch (int): current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Enable train mode.
model.train()
train_meter.iter_tic()
data_size = len(train_loader)
for cur_iter, (inputs, labels, _, meta) in enumerate(train_loader):
# Transfer the data to the current GPU device.
if isinstance(inputs, (list,)):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list,)):
for i in range(len(val)):
if torch.is_tensor(val[i]):
val[i] = val[i].cuda(non_blocking=True)
else:
val[i] = val[i].to('cuda')
else:
meta[key] = val.cuda(non_blocking=True)
# Update the learning rate.
lr = optim.get_epoch_lr(cur_epoch + float(cur_iter) / data_size, cfg)
optim.set_lr(optimizer, lr)
if cfg.FCOS.ENABLE:
loss_dict = model(inputs, meta["boxes"])
print('centerness: ', loss_dict['loss_centerness'].item(), ' class: ', loss_dict['loss_cls'].item(), ' box reg: ', loss_dict['loss_reg'].item())
loss = sum(loss for loss in loss_dict.values())
# # reduce losses over all GPUs for logging purposes
# loss_dict_reduced = reduce_loss_dict(loss_dict)
# losses_reduced = sum(loss for loss in loss_dict_reduced.values())
else:
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
else:
# Perform the forward pass.
preds = model(inputs)
# Explicitly declare reduction to mean.
loss_fun = losses.get_loss_func(cfg.MODEL.LOSS_FUNC)(reduction="mean")
# Compute the loss.
loss = loss_fun(preds, labels)
# check Nan Loss.
misc.check_nan_losses(loss)
# Perform the backward pass.
optimizer.zero_grad()
loss.backward()
# Update the parameters.
optimizer.step()
if cfg.FCOS.ENABLE:
if cfg.NUM_GPUS > 1:
loss = du.all_reduce([loss])[0]
loss = loss.item()
print(f'step: {cur_iter} lr: ', lr, ' ', loss)
# train_meter.iter_toc()
else:
if cfg.DETECTION.ENABLE:
if cfg.NUM_GPUS > 1:
loss = du.all_reduce([loss])[0]
loss = loss.item()
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(None, None, None, loss, lr)
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(preds, labels, (1, 5))
top1_err, top5_err = [
(1.0 - x / preds.size(0)) * 100.0 for x in num_topks_correct
]
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss, top1_err, top5_err = du.all_reduce(
[loss, top1_err, top5_err]
)
# Copy the stats from GPU to CPU (sync point).
loss, top1_err, top5_err = (
loss.item(),
top1_err.item(),
top5_err.item(),
)
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(
top1_err, top5_err, loss, lr, inputs[0].size(0) * cfg.NUM_GPUS
)
train_meter.log_iter_stats(cur_epoch, cur_iter)
train_meter.iter_tic()
# Log epoch stats.
train_meter.log_epoch_stats(cur_epoch)
train_meter.reset()
def train_epoch(train_loader,
model,
optimizer,
train_meter,
cur_epoch,
cfg,
writer=None):
"""
Perform the video training for one epoch.
Args:
train_loader (loader): video training loader.
model (model): the video model to train.
optimizer (optim): the optimizer to perform optimization on the model's
parameters.
train_meter (TrainMeter): training meters to log the training performance.
cur_epoch (int): current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
writer (TensorboardWriter, optional): TensorboardWriter object
to writer Tensorboard log.
"""
# Enable train mode.
model.train()
train_meter.iter_tic()
data_size = len(train_loader)
for cur_iter, (inputs, labels, _, meta) in enumerate(train_loader):
# Transfer the data to the current GPU device.
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list, )):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
# Update the learning rate.
lr = optim.get_epoch_lr(cur_epoch + float(cur_iter) / data_size, cfg)
optim.set_lr(optimizer, lr)
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
else:
# Perform the forward pass.
preds = model(inputs)
# Explicitly declare reduction to mean.
loss_fun = losses.get_loss_func(cfg.MODEL.LOSS_FUNC)(reduction="mean")
# Compute the loss.
loss = loss_fun(preds, labels)
# check Nan Loss.
misc.check_nan_losses(loss)
# Perform the backward pass.
optimizer.zero_grad()
loss.backward()
# Update the parameters.
optimizer.step()
if cfg.DETECTION.ENABLE:
if cfg.NUM_GPUS > 1:
loss = du.all_reduce([loss])[0]
loss = loss.item()
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(None, None, None, loss, lr)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Train/loss": loss,
"Train/lr": lr
},
global_step=data_size * cur_epoch + cur_iter,
)
else:
top1_err, top5_err = None, None
if cfg.DATA.MULTI_LABEL:
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
[loss] = du.all_reduce([loss])
loss = loss.item()
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(
preds, labels, (1, 5))
top1_err, top5_err = [(1.0 - x / preds.size(0)) * 100.0
for x in num_topks_correct]
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss, top1_err, top5_err = du.all_reduce(
[loss, top1_err, top5_err])
# Copy the stats from GPU to CPU (sync point).
loss, top1_err, top5_err = (
loss.item(),
top1_err.item(),
top5_err.item(),
)
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(top1_err, top5_err, loss, lr,
inputs[0].size(0) * cfg.NUM_GPUS)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Train/loss": loss,
"Train/lr": lr,
"Train/Top1_err": top1_err,
"Train/Top5_err": top5_err,
},
global_step=data_size * cur_epoch + cur_iter,
)
train_meter.log_iter_stats(cur_epoch, cur_iter)
train_meter.iter_tic()
# Log epoch stats.
train_meter.log_epoch_stats(cur_epoch)
train_meter.reset()
def eval_epoch(val_loader, model, val_meter, cur_epoch, cfg):
"""
Evaluate the model on the val set.
Args:
val_loader (loader): data loader to provide validation data.
model (model): model to evaluate the performance.
val_meter (ValMeter): meter instance to record and calculate the metrics.
cur_epoch (int): number of the current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Evaluation mode enabled. The running stats would not be updated.
model.eval()
val_meter.iter_tic()
for cur_iter, (inputs, labels, _, meta) in enumerate(val_loader):
# Transferthe data to the current GPU device.
if isinstance(inputs, (list,)):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list,)):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
preds = preds.cpu()
ori_boxes = meta["ori_boxes"].cpu()
metadata = meta["metadata"].cpu()
if cfg.NUM_GPUS > 1:
preds = torch.cat(du.all_gather_unaligned(preds), dim=0)
ori_boxes = torch.cat(du.all_gather_unaligned(ori_boxes), dim=0)
metadata = torch.cat(du.all_gather_unaligned(metadata), dim=0)
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(preds.cpu(), ori_boxes.cpu(), metadata.cpu())
else:
preds = model(inputs)
# Compute the errors.
num_topks_correct = metrics.topks_correct(preds, labels, (1, 5))
# Combine the errors across the GPUs.
top1_err, top5_err = [
(1.0 - x / preds.size(0)) * 100.0 for x in num_topks_correct
]
if cfg.NUM_GPUS > 1:
top1_err, top5_err = du.all_reduce([top1_err, top5_err])
# Copy the errors from GPU to CPU (sync point).
top1_err, top5_err = top1_err.item(), top5_err.item()
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(
top1_err, top5_err, inputs[0].size(0) * cfg.NUM_GPUS
)
val_meter.log_iter_stats(cur_epoch, cur_iter)
val_meter.iter_tic()
# Log epoch stats.
val_meter.log_epoch_stats(cur_epoch)
val_meter.reset()
def eval_epoch(val_loader, model, val_meter, cur_epoch, cfg, writer=None):
"""
Evaluate the model on the val set.
Args:
val_loader (loader): data loader to provide validation data.
model (model): model to evaluate the performance.
val_meter (ValMeter): meter instance to record and calculate the metrics.
cur_epoch (int): number of the current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
writer (TensorboardWriter, optional): TensorboardWriter object
to writer Tensorboard log.
"""
# Evaluation mode enabled. The running stats would not be updated.
model.eval()
val_meter.iter_tic()
for cur_iter, (inputs, labels, _, meta) in enumerate(val_loader):
# Transferthe data to the current GPU device.
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list, )):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
preds = preds.cpu()
ori_boxes = meta["ori_boxes"].cpu()
metadata = meta["metadata"].cpu()
if cfg.NUM_GPUS > 1:
preds = torch.cat(du.all_gather_unaligned(preds), dim=0)
ori_boxes = torch.cat(du.all_gather_unaligned(ori_boxes),
dim=0)
metadata = torch.cat(du.all_gather_unaligned(metadata), dim=0)
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(preds.cpu(), ori_boxes.cpu(),
metadata.cpu())
else:
preds = model(inputs)
if cfg.DATA.MULTI_LABEL:
if cfg.NUM_GPUS > 1:
preds, labels = du.all_gather([preds, labels])
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(
preds, labels, (1, 1))
# Combine the errors across the GPUs.
top1_err, top5_err = [(1.0 - x / preds.size(0)) * 100.0
for x in num_topks_correct]
if cfg.NUM_GPUS > 1:
top1_err, top5_err = du.all_reduce([top1_err, top5_err])
# Copy the errors from GPU to CPU (sync point).
top1_err, top5_err = top1_err.item(), top5_err.item()
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(top1_err, top5_err,
inputs[0].size(0) * cfg.NUM_GPUS)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Val/Top1_err": top1_err,
"Val/Top5_err": top5_err
},
global_step=len(val_loader) * cur_epoch + cur_iter,
)
val_meter.update_predictions(preds, labels)
val_meter.log_iter_stats(cur_epoch, cur_iter)
val_meter.iter_tic()
logger.info('COMPUTING MCC')
# Log epoch stats.
val_meter.log_epoch_stats(cur_epoch)
all_preds_cpu = [
pred.clone().detach().cpu() for pred in val_meter.all_preds
]
all_labels_cpu = [
label.clone().detach().cpu() for label in val_meter.all_labels
]
logger.info('PREPROC FOR MCC')
preds = torch.cat(all_preds_cpu)
ypreds = torch.argmax(preds, dim=1)
ytrue = torch.cat(all_labels_cpu)
logger.info('COMPUTE CM')
cm = plmetrics.ConfusionMatrix()(ypreds.to('cuda'), ytrue.to('cuda'))
logger.info('CM COMPUTED')
tp, tn, fn, fp = cm[1, 1], cm[0, 0], cm[0, 1], cm[1, 0]
denom = (tp + fp) * (tp + fn) * (tn + fp) * (tn + fn)
mcc = (tp * tn - fp * fn) / torch.sqrt(denom)
logger.info('COMPUTED MCC')
# write to tensorboard format if available.
if writer is not None:
if cfg.DETECTION.ENABLE:
writer.add_scalars({"Val/mAP": val_meter.full_map},
global_step=cur_epoch)
writer.plot_eval(preds=all_preds_cpu,
labels=all_labels_cpu,
global_step=cur_epoch)
val_meter.reset()
return mcc.item()
def train_epoch(train_loader, model, optimizer, train_meter, cur_epoch, cfg):
"""
Perform the video training for one epoch.
Args:
train_loader (loader): video training loader.
model (model): the video model to train.
optimizer (optim): the optimizer to perform optimization on the model's
parameters.
train_meter (TrainMeter): training meters to log the training performance.
cur_epoch (int): current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
"""
# Enable train mode.
model.train()
if cfg.BN.FREEZE:
model.module.freeze_fn('bn_statistics')
train_meter.iter_tic()
data_size = len(train_loader)
for cur_iter, (inputs, labels, _, meta) in enumerate(train_loader):
# Transfer the data to the current GPU device.
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
if isinstance(labels, (dict, )):
labels = {k: v.cuda() for k, v in labels.items()}
else:
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list, )):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
# Update the learning rate.
lr = optim.get_epoch_lr(cur_epoch + float(cur_iter) / data_size, cfg)
optim.set_lr(optimizer, lr)
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
else:
# Perform the forward pass.
preds = model(inputs)
if isinstance(labels, (dict, )):
# Explicitly declare reduction to mean.
loss_fun = losses.get_loss_func(
cfg.MODEL.LOSS_FUNC)(reduction="mean")
# Compute the loss.
loss_verb = loss_fun(preds[0], labels['verb'])
loss_noun = loss_fun(preds[1], labels['noun'])
loss = 0.5 * (loss_verb + loss_noun)
# check Nan Loss.
misc.check_nan_losses(loss)
else:
# Explicitly declare reduction to mean.
loss_fun = losses.get_loss_func(
cfg.MODEL.LOSS_FUNC)(reduction="mean")
# Compute the loss.
loss = loss_fun(preds, labels)
# check Nan Loss.
misc.check_nan_losses(loss)
# Perform the backward pass.
optimizer.zero_grad()
loss.backward()
# Update the parameters.
optimizer.step()
if cfg.DETECTION.ENABLE:
if cfg.NUM_GPUS > 1:
loss = du.all_reduce([loss])[0]
loss = loss.item()
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(None, None, None, loss, lr)
else:
if isinstance(labels, (dict, )):
# Compute the verb accuracies.
verb_top1_acc, verb_top5_acc = metrics.topk_accuracies(
preds[0], labels['verb'], (1, 5))
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss_verb, verb_top1_acc, verb_top5_acc = du.all_reduce(
[loss_verb, verb_top1_acc, verb_top5_acc])
# Copy the stats from GPU to CPU (sync point).
loss_verb, verb_top1_acc, verb_top5_acc = (
loss_verb.item(),
verb_top1_acc.item(),
verb_top5_acc.item(),
)
# Compute the noun accuracies.
noun_top1_acc, noun_top5_acc = metrics.topk_accuracies(
preds[1], labels['noun'], (1, 5))
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss_noun, noun_top1_acc, noun_top5_acc = du.all_reduce(
[loss_noun, noun_top1_acc, noun_top5_acc])
# Copy the stats from GPU to CPU (sync point).
loss_noun, noun_top1_acc, noun_top5_acc = (
loss_noun.item(),
noun_top1_acc.item(),
noun_top5_acc.item(),
)
# Compute the action accuracies.
action_top1_acc, action_top5_acc = metrics.multitask_topk_accuracies(
(preds[0], preds[1]), (labels['verb'], labels['noun']),
(1, 5))
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss, action_top1_acc, action_top5_acc = du.all_reduce(
[loss, action_top1_acc, action_top5_acc])
# Copy the stats from GPU to CPU (sync point).
loss, action_top1_acc, action_top5_acc = (
loss.item(),
action_top1_acc.item(),
action_top5_acc.item(),
)
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(
(verb_top1_acc, noun_top1_acc, action_top1_acc),
(verb_top5_acc, noun_top5_acc, action_top5_acc),
(loss_verb, loss_noun, loss), lr,
inputs[0].size(0) * cfg.NUM_GPUS)
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(
preds, labels, (1, 5))
top1_err, top5_err = [(1.0 - x / preds.size(0)) * 100.0
for x in num_topks_correct]
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss, top1_err, top5_err = du.all_reduce(
[loss, top1_err, top5_err])
# Copy the stats from GPU to CPU (sync point).
loss, top1_err, top5_err = (
loss.item(),
top1_err.item(),
top5_err.item(),
)
train_meter.iter_toc()
# Update and log stats.
train_meter.update_stats(top1_err, top5_err, loss, lr,
inputs[0].size(0) * cfg.NUM_GPUS)
train_meter.log_iter_stats(cur_epoch, cur_iter)
train_meter.iter_tic()
# Log epoch stats.
train_meter.log_epoch_stats(cur_epoch)
train_meter.reset()
def train_epoch(
train_loader,
model,
optimizer,
scaler,
train_meter,
cur_epoch,
cfg,
writer=None,
):
"""
Perform the video training for one epoch.
Args:
train_loader (loader): video training loader.
model (model): the video model to train.
optimizer (optim): the optimizer to perform optimization on the model's
parameters.
train_meter (TrainMeter): training meters to log the training performance.
cur_epoch (int): current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
writer (TensorboardWriter, optional): TensorboardWriter object
to writer Tensorboard log.
"""
# Enable train mode.
model.train()
train_meter.iter_tic()
data_size = len(train_loader)
if cfg.MIXUP.ENABLE:
mixup_fn = MixUp(
mixup_alpha=cfg.MIXUP.ALPHA,
cutmix_alpha=cfg.MIXUP.CUTMIX_ALPHA,
mix_prob=cfg.MIXUP.PROB,
switch_prob=cfg.MIXUP.SWITCH_PROB,
label_smoothing=cfg.MIXUP.LABEL_SMOOTH_VALUE,
num_classes=cfg.MODEL.NUM_CLASSES,
)
iters_noupdate = 0
if (cfg.MODEL.MODEL_NAME == "ContrastiveModel"
and cfg.CONTRASTIVE.TYPE == "moco"):
assert (cfg.CONTRASTIVE.QUEUE_LEN %
(cfg.TRAIN.BATCH_SIZE * cfg.NUM_SHARDS) == 0)
iters_noupdate = (cfg.CONTRASTIVE.QUEUE_LEN // cfg.TRAIN.BATCH_SIZE //
cfg.NUM_SHARDS)
if cfg.MODEL.FROZEN_BN:
misc.frozen_bn_stats(model)
# Explicitly declare reduction to mean.
loss_fun = losses.get_loss_func(cfg.MODEL.LOSS_FUNC)(reduction="mean")
for cur_iter, (inputs, labels, index, time,
meta) in enumerate(train_loader):
# Transfer the data to the current GPU device.
if cfg.NUM_GPUS:
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
if isinstance(inputs[i], (list, )):
for j in range(len(inputs[i])):
inputs[i][j] = inputs[i][j].cuda(non_blocking=True)
else:
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list, )):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
index = index.cuda()
time = time.cuda()
batch_size = (inputs[0][0].size(0)
if isinstance(inputs[0], list) else inputs[0].size(0))
# Update the learning rate.
epoch_exact = cur_epoch + float(cur_iter) / data_size
lr = optim.get_epoch_lr(epoch_exact, cfg)
optim.set_lr(optimizer, lr)
train_meter.data_toc()
if cfg.MIXUP.ENABLE:
samples, labels = mixup_fn(inputs[0], labels)
inputs[0] = samples
with torch.cuda.amp.autocast(enabled=cfg.TRAIN.MIXED_PRECISION):
# Explicitly declare reduction to mean.
perform_backward = True
optimizer.zero_grad()
if cfg.MODEL.MODEL_NAME == "ContrastiveModel":
(
model,
preds,
partial_loss,
perform_backward,
) = contrastive_forward(model, cfg, inputs, index, time,
epoch_exact, scaler)
elif cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
else:
preds = model(inputs)
if cfg.TASK == "ssl" and cfg.MODEL.MODEL_NAME == "ContrastiveModel":
labels = torch.zeros(preds.size(0),
dtype=labels.dtype,
device=labels.device)
if cfg.MODEL.MODEL_NAME == "ContrastiveModel" and partial_loss:
loss = partial_loss
else:
# Compute the loss.
loss = loss_fun(preds, labels)
# check Nan Loss.
misc.check_nan_losses(loss)
if perform_backward:
scaler.scale(loss).backward()
# Unscales the gradients of optimizer's assigned params in-place
scaler.unscale_(optimizer)
# Clip gradients if necessary
if cfg.SOLVER.CLIP_GRAD_VAL:
torch.nn.utils.clip_grad_value_(model.parameters(),
cfg.SOLVER.CLIP_GRAD_VAL)
elif cfg.SOLVER.CLIP_GRAD_L2NORM:
torch.nn.utils.clip_grad_norm_(model.parameters(),
cfg.SOLVER.CLIP_GRAD_L2NORM)
model = cancel_swav_gradients(model, cfg, epoch_exact)
if cur_iter < iters_noupdate and cur_epoch == 0: # for e.g. MoCo
logger.info("Not updating parameters {}/{}".format(
cur_iter, iters_noupdate))
else:
# Update the parameters.
scaler.step(optimizer)
scaler.update()
if cfg.MIXUP.ENABLE:
_top_max_k_vals, top_max_k_inds = torch.topk(labels,
2,
dim=1,
largest=True,
sorted=True)
idx_top1 = torch.arange(labels.shape[0]), top_max_k_inds[:, 0]
idx_top2 = torch.arange(labels.shape[0]), top_max_k_inds[:, 1]
preds = preds.detach()
preds[idx_top1] += preds[idx_top2]
preds[idx_top2] = 0.0
labels = top_max_k_inds[:, 0]
if cfg.DETECTION.ENABLE:
if cfg.NUM_GPUS > 1:
loss = du.all_reduce([loss])[0]
loss = loss.item()
# Update and log stats.
train_meter.update_stats(None, None, None, loss, lr)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Train/loss": loss,
"Train/lr": lr
},
global_step=data_size * cur_epoch + cur_iter,
)
else:
top1_err, top5_err = None, None
if cfg.DATA.MULTI_LABEL:
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
[loss] = du.all_reduce([loss])
loss = loss.item()
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(
preds, labels, (1, 5))
top1_err, top5_err = [(1.0 - x / preds.size(0)) * 100.0
for x in num_topks_correct]
# Gather all the predictions across all the devices.
if cfg.NUM_GPUS > 1:
loss, top1_err, top5_err = du.all_reduce(
[loss.detach(), top1_err, top5_err])
# Copy the stats from GPU to CPU (sync point).
loss, top1_err, top5_err = (
loss.item(),
top1_err.item(),
top5_err.item(),
)
# Update and log stats.
train_meter.update_stats(
top1_err,
top5_err,
loss,
lr,
batch_size * max(
cfg.NUM_GPUS, 1
), # If running on CPU (cfg.NUM_GPUS == 1), use 1 to represent 1 CPU.
)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Train/loss": loss,
"Train/lr": lr,
"Train/Top1_err": top1_err,
"Train/Top5_err": top5_err,
},
global_step=data_size * cur_epoch + cur_iter,
)
torch.cuda.synchronize()
train_meter.iter_toc() # do measure allreduce for this meter
train_meter.log_iter_stats(cur_epoch, cur_iter)
torch.cuda.synchronize()
train_meter.iter_tic()
del inputs
# Log epoch stats.
train_meter.log_epoch_stats(cur_epoch)
train_meter.reset()
def eval_epoch(val_loader, model, val_meter, cur_epoch, cfg, train_loader,
writer):
"""
Evaluate the model on the val set.
Args:
val_loader (loader): data loader to provide validation data.
model (model): model to evaluate the performance.
val_meter (ValMeter): meter instance to record and calculate the metrics.
cur_epoch (int): number of the current epoch of training.
cfg (CfgNode): configs. Details can be found in
slowfast/config/defaults.py
writer (TensorboardWriter, optional): TensorboardWriter object
to writer Tensorboard log.
"""
# Evaluation mode enabled. The running stats would not be updated.
model.eval()
val_meter.iter_tic()
for cur_iter, (inputs, labels, index, time, meta) in enumerate(val_loader):
if cfg.NUM_GPUS:
# Transferthe data to the current GPU device.
if isinstance(inputs, (list, )):
for i in range(len(inputs)):
inputs[i] = inputs[i].cuda(non_blocking=True)
else:
inputs = inputs.cuda(non_blocking=True)
labels = labels.cuda()
for key, val in meta.items():
if isinstance(val, (list, )):
for i in range(len(val)):
val[i] = val[i].cuda(non_blocking=True)
else:
meta[key] = val.cuda(non_blocking=True)
index = index.cuda()
time = time.cuda()
batch_size = (inputs[0][0].size(0)
if isinstance(inputs[0], list) else inputs[0].size(0))
val_meter.data_toc()
if cfg.DETECTION.ENABLE:
# Compute the predictions.
preds = model(inputs, meta["boxes"])
ori_boxes = meta["ori_boxes"]
metadata = meta["metadata"]
if cfg.NUM_GPUS:
preds = preds.cpu()
ori_boxes = ori_boxes.cpu()
metadata = metadata.cpu()
if cfg.NUM_GPUS > 1:
preds = torch.cat(du.all_gather_unaligned(preds), dim=0)
ori_boxes = torch.cat(du.all_gather_unaligned(ori_boxes),
dim=0)
metadata = torch.cat(du.all_gather_unaligned(metadata), dim=0)
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(preds, ori_boxes, metadata)
else:
if cfg.TASK == "ssl" and cfg.MODEL.MODEL_NAME == "ContrastiveModel":
if not cfg.CONTRASTIVE.KNN_ON:
return
train_labels = (model.module.train_labels if hasattr(
model, "module") else model.train_labels)
yd, yi = model(inputs, index, time)
K = yi.shape[1]
C = (cfg.CONTRASTIVE.NUM_CLASSES_DOWNSTREAM
) # eg 400 for Kinetics400
candidates = train_labels.view(1, -1).expand(batch_size, -1)
retrieval = torch.gather(candidates, 1, yi)
retrieval_one_hot = torch.zeros((batch_size * K, C)).cuda()
retrieval_one_hot.scatter_(1, retrieval.view(-1, 1), 1)
yd_transform = yd.clone().div_(cfg.CONTRASTIVE.T).exp_()
probs = torch.mul(
retrieval_one_hot.view(batch_size, -1, C),
yd_transform.view(batch_size, -1, 1),
)
preds = torch.sum(probs, 1)
else:
preds = model(inputs)
if cfg.DATA.MULTI_LABEL:
if cfg.NUM_GPUS > 1:
preds, labels = du.all_gather([preds, labels])
else:
# Compute the errors.
num_topks_correct = metrics.topks_correct(
preds, labels, (1, 5))
# Combine the errors across the GPUs.
top1_err, top5_err = [(1.0 - x / preds.size(0)) * 100.0
for x in num_topks_correct]
if cfg.NUM_GPUS > 1:
top1_err, top5_err = du.all_reduce([top1_err, top5_err])
# Copy the errors from GPU to CPU (sync point).
top1_err, top5_err = top1_err.item(), top5_err.item()
val_meter.iter_toc()
# Update and log stats.
val_meter.update_stats(
top1_err,
top5_err,
batch_size * max(
cfg.NUM_GPUS, 1
), # If running on CPU (cfg.NUM_GPUS == 1), use 1 to represent 1 CPU.
)
# write to tensorboard format if available.
if writer is not None:
writer.add_scalars(
{
"Val/Top1_err": top1_err,
"Val/Top5_err": top5_err
},
global_step=len(val_loader) * cur_epoch + cur_iter,
)
val_meter.update_predictions(preds, labels)
val_meter.log_iter_stats(cur_epoch, cur_iter)
val_meter.iter_tic()
# Log epoch stats.
val_meter.log_epoch_stats(cur_epoch)
# write to tensorboard format if available.
if writer is not None:
if cfg.DETECTION.ENABLE:
writer.add_scalars({"Val/mAP": val_meter.full_map},
global_step=cur_epoch)
else:
all_preds = [pred.clone().detach() for pred in val_meter.all_preds]
all_labels = [
label.clone().detach() for label in val_meter.all_labels
]
if cfg.NUM_GPUS:
all_preds = [pred.cpu() for pred in all_preds]
all_labels = [label.cpu() for label in all_labels]
writer.plot_eval(preds=all_preds,
labels=all_labels,
global_step=cur_epoch)
val_meter.reset()