def get_model_and_data_loaders(
config: ConfigParser,
logger: logging.Logger,
ckpt_path: Path,
) -> Tuple[torch.nn.Module, module_data.ExpertDataLoader]:
expert_dims, raw_input_dims, text_dim = compute_dims(config)
data_loaders = config.init(
name='data_loader',
module=module_data,
logger=logger,
raw_input_dims=raw_input_dims,
challenge_mode=config.get("challenge_mode", False),
text_dim=text_dim,
text_feat=config["experts"]["text_feat"],
text_agg=config["experts"]["text_agg"],
use_zeros_for_missing=config["experts"].get("use_zeros_for_missing",
False),
task=config.get("task", "retrieval"),
eval_only=True,
distil_params=config.get("distil_params", None),
training_file=config.get("training_file", None),
caption_masks=config.get("caption_masks", None),
ce_shared_dim=config["experts"].get("ce_shared_dim", None),
)
trn_config = compute_trn_config(config)
model = config.init(
name='arch',
module=module_arch,
trn_config=trn_config,
expert_dims=expert_dims,
text_dim=text_dim,
disable_nan_checks=config["disable_nan_checks"],
task=config.get("task", "retrieval"),
ce_shared_dim=config["experts"].get("ce_shared_dim", None),
feat_aggregation=config["data_loader"]["args"]["feat_aggregation"],
trn_cat=config["data_loader"]["args"].get("trn_cat", 0),
)
ckpt_path = config._args.resume
logger.info(f"Loading checkpoint: {ckpt_path} ...")
checkpoint = torch.load(ckpt_path)
state_dict = checkpoint['state_dict']
if config['n_gpu'] > 1:
model = torch.nn.DataParallel(model)
# support backwards compatibility
deprecated = ["ce.moe_fc_bottleneck1", "ce.moe_cg", "ce.moe_fc_proj"]
for mod in deprecated:
for suffix in ("weight", "bias"):
key = f"{mod}.{suffix}"
if key in state_dict:
print(f"WARNING: Removing deprecated key {key} from model")
state_dict.pop(key)
model.load_state_dict(state_dict)
return model, data_loaders
def get_model_and_data_loaders(
config: ConfigParser,
logger: logging.Logger,
ckpt_path: Path,
) -> Tuple[torch.nn.Module, module_data.ExpertDataLoader]:
expert_dims, raw_input_dims = compute_dims(config)
trn_config = compute_trn_config(config)
data_loaders = config.init(
name='data_loader',
module=module_data,
logger=logger,
raw_input_dims=raw_input_dims,
challenge_mode=config.get("challenge_mode", False),
text_feat=config["experts"]["text_feat"],
text_dim=config["experts"]["text_dim"],
text_agg=config["experts"]["text_agg"],
use_zeros_for_missing=config["experts"].get("use_zeros_for_missing", False),
task=config.get("task", "retrieval"),
eval_only=True,
)
model = config.init(
name='arch',
module=module_arch,
trn_config=trn_config,
expert_dims=expert_dims,
text_dim=config["experts"]["text_dim"],
disable_nan_checks=config["disable_nan_checks"],
task=config.get("task", "retrieval"),
ce_shared_dim=config["experts"].get("ce_shared_dim", None),
feat_aggregation=config["data_loader"]["args"]["feat_aggregation"],
trn_cat=config["data_loader"]["args"].get("trn_cat", 0),
)
ckpt_path = config._args.resume
logger.info(f"Loading checkpoint: {ckpt_path} ...")
checkpoint = torch.load(ckpt_path)
state_dict = checkpoint['state_dict']
if config['n_gpu'] > 1:
model = torch.nn.DataParallel(model)
model.load_state_dict(state_dict)
return model, data_loaders
def evaluation(config, logger=None, trainer=None):
if logger is None:
logger = config.get_logger('test')
if getattr(config._args, "eval_from_training_config", False):
eval_conf = copy.deepcopy(config)
merge(eval_conf._config,
config["eval_settings"],
strategy=Strategy.REPLACE)
config = eval_conf
logger.info("Running evaluation with configuration:")
logger.info(config)
expert_dims, raw_input_dims = compute_dims(config)
trn_config = compute_trn_config(config)
# Set the random initial seeds
seed = config["seed"]
logger.info(f"Setting experiment random seed to {seed}")
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
update_src_web_video_dir(config)
visualizer = config.init(
name='visualizer',
module=module_vis,
exp_name=config._exper_name,
web_dir=config._web_log_dir,
)
data_loaders = config.init(
name='data_loader',
module=module_data,
logger=logger,
raw_input_dims=raw_input_dims,
challenge_mode=config.get("challenge_mode", False),
text_feat=config["experts"]["text_feat"],
text_dim=config["experts"]["text_dim"],
text_agg=config["experts"]["text_agg"],
use_zeros_for_missing=config["experts"].get("use_zeros_for_missing",
False),
task=config.get("task", "retrieval"),
eval_only=True,
)
model = config.init(
name='arch',
module=module_arch,
trn_config=trn_config,
expert_dims=expert_dims,
text_dim=config["experts"]["text_dim"],
disable_nan_checks=config["disable_nan_checks"],
task=config.get("task", "retrieval"),
ce_shared_dim=config["experts"].get("ce_shared_dim", None),
feat_aggregation=config["data_loader"]["args"]["feat_aggregation"],
trn_cat=config["data_loader"]["args"].get("trn_cat", 0),
)
logger.info(model)
metrics = [getattr(module_metric, met) for met in config['metrics']]
ckpt_path = config._args.resume
logger.info(f"Loading checkpoint: {ckpt_path} ...")
checkpoint = torch.load(ckpt_path)
state_dict = checkpoint['state_dict']
if config['n_gpu'] > 1:
model = torch.nn.DataParallel(model)
model.load_state_dict(state_dict)
challenge_mode = config.get("challenge_mode", False)
challenge_msg = (
"\n"
"Evaluation ran on challenge features. To obtain a score, upload the similarity"
"matrix for each dataset to the test server after running the "
"`misc/cvpr2020-challenge/prepare_submission.py` script and following the "
"instructions at: "
"https://www.robots.ox.ac.uk/~vgg/challenges/video-pentathlon/"
"\n")
# prepare model for testing. Note that some datasets fail to fit the retrieval
# set on the GPU, so we run them on the CPU
if torch.cuda.is_available() and not config.get("disable_gpu", True):
device = "cuda"
else:
device = "cpu"
logger.info(f"Running evaluation on {device}")
model = model.to(device)
model.eval()
with torch.no_grad():
samples, meta = data_loaders["retrieval"]
# To use the nan-checks safely, we need make temporary copies of the data
disable_nan_checks = config._config["disable_nan_checks"]
with ctxt_mgr(samples, device, disable_nan_checks) as valid:
output = model(**valid)
sims = output["cross_view_conf_matrix"].data.cpu().float().numpy()
dataset = data_loaders.dataset_name
if challenge_mode:
split = data_loaders.dataloaders["dataset"].split_name
prediction_path = config._log_dir / f"{dataset}-{split}-predictions.csv"
compressed_preds = compress_predictions(
query_masks=meta["query_masks"],
sims=sims,
)
np.savetxt(prediction_path,
compressed_preds,
delimiter=',',
fmt="%d")
print(f"Saved similarity matrix predictions to {prediction_path}")
print(challenge_msg)
return
nested_metrics = {}
for metric in metrics:
metric_name = metric.__name__
res = metric(sims, query_masks=meta["query_masks"])
verbose(epoch=0, metrics=res, name=dataset, mode=metric_name)
if trainer is not None:
if not trainer.mini_train:
trainer.writer.set_step(step=0, mode="val")
# avoid tensboard folding by prefixing
metric_name_ = f"test_{metric_name}"
trainer.log_metrics(res, metric_name=metric_name_, mode="val")
nested_metrics[metric_name] = res
if data_loaders.num_test_captions == 1:
visualizer.visualize_ranking(
sims=sims,
meta=meta,
epoch=0,
nested_metrics=nested_metrics,
)
log = {}
for subkey, subval in nested_metrics.items():
for subsubkey, subsubval in subval.items():
log[f"test_{subkey}_{subsubkey}"] = subsubval
for key, value in log.items():
logger.info(" {:15s}: {}".format(str(key), value))
def evaluation(config, logger=None, trainer=None):
if logger is None:
logger = config.get_logger('test')
if getattr(config._args, "eval_from_training_config", False):
eval_conf = copy.deepcopy(config)
merge(eval_conf._config,
config["eval_settings"],
strategy=Strategy.REPLACE)
config = eval_conf
logger.info("Running evaluation with configuration:")
logger.info(config)
expert_dims, raw_input_dims = compute_dims(config)
trn_config = compute_trn_config(config)
# Set the random initial seeds
seed = config["seed"]
logger.info(f"Setting experiment random seed to {seed}")
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
# We use cls defaults for backwards compatibility with the MMIT configs. In the
# long run this should be handled by the json configs themselves
cls_defaults = ["train", "val", "tiny", "challenge"]
data_loaders = config.init(
name='data_loader',
module=module_data,
logger=logger,
raw_input_dims=raw_input_dims,
text_feat=config["experts"]["text_feat"],
text_dim=config["experts"]["text_dim"],
text_agg=config["experts"]["text_agg"],
use_zeros_for_missing=config["experts"].get("use_zeros_for_missing",
False),
task=config.get("task", "retrieval"),
cls_partitions=config.get("cls_partitions", cls_defaults),
)
model = config.init(
name='arch',
module=module_arch,
trn_config=trn_config,
expert_dims=expert_dims,
text_dim=config["experts"]["text_dim"],
disable_nan_checks=config["disable_nan_checks"],
task=config.get("task", "retrieval"),
ce_shared_dim=config["experts"].get("ce_shared_dim", None),
feat_aggregation=config["data_loader"]["args"]["feat_aggregation"],
trn_cat=config["data_loader"]["args"].get("trn_cat", 0),
)
logger.info(model)
metrics = [getattr(module_metric, met) for met in config['metrics']]
visualizer = config.init(
name='visualizer',
module=module_vis,
exp_name=config._exper_name,
web_dir=config._web_log_dir,
)
ckpt_path = config._args.resume
logger.info(f"Loading checkpoint: {ckpt_path} ...")
checkpoint = torch.load(ckpt_path)
state_dict = checkpoint['state_dict']
if config['n_gpu'] > 1:
model = torch.nn.DataParallel(model)
model.load_state_dict(state_dict)
# prepare model for testing. Note that some datasets fail to fit the retrieval
# set on the GPU, so we run them on the CPU
if torch.cuda.is_available() and not config.get("disable_gpu", True):
device = "cuda"
else:
device = "cpu"
logger.info(f"Running evaluation on {device}")
model = model.to(device)
model.eval()
with torch.no_grad():
samples, meta = data_loaders["retrieval"]
# To use the nan-checks safely, we need make temporary copies of the data
disable_nan_checks = config._config["disable_nan_checks"]
with ctxt_mgr(samples, device, disable_nan_checks) as valid:
output = model(**valid)
sims = output["cross_view_conf_matrix"].data.cpu().float().numpy()
dataset = data_loaders.dataset_name
nested_metrics = {}
for metric in metrics:
metric_name = metric.__name__
res = metric(sims, query_masks=meta["query_masks"])
verbose(epoch=0, metrics=res, name=dataset, mode=metric_name)
if trainer is not None:
if not trainer.mini_train:
trainer.writer.set_step(step=0, mode="val")
# avoid tensboard folding by prefixing
metric_name_ = f"test_{metric_name}"
trainer.log_metrics(res, metric_name=metric_name_, mode="val")
nested_metrics[metric_name] = res
if data_loaders.num_test_captions == 1:
visualizer.visualize_ranking(
sims=sims,
meta=meta,
epoch=0,
nested_metrics=nested_metrics,
)
log = {}
for subkey, subval in nested_metrics.items():
for subsubkey, subsubval in subval.items():
log[f"test_{subkey}_{subsubkey}"] = subsubval
for key, value in log.items():
logger.info(" {:15s}: {}".format(str(key), value))
def run_exp(config):
warnings.filterwarnings('ignore')
logger = config.get_logger('train')
leaderboard_path = config._args.leaderboard
Path(leaderboard_path).parent.mkdir(exist_ok=True, parents=True)
with open(leaderboard_path, 'a') as f:
txt_path = f"{config._log_dir}/preds.txt"
print(txt_path, file=f, flush=True)
expert_dims, raw_input_dims = compute_dims(config, logger)
trn_config = compute_trn_config(config)
if config._args.group_seed:
seeds = [int(config._args.group_seed)]
else:
seeds = [int(x) for x in config._args.seeds.split(",")]
# set up local filesystem on the cluster
if socket.gethostname().endswith("cluster"):
os.system(str(Path.home() / "configure_tmp_data.sh"))
for ii, seed in enumerate(seeds):
tic = time.time()
logger.info(f"{ii + 1}/{len(seeds)} Setting experiment random seed to {seed}")
set_seeds(seed)
config["seed"] = seed
# We use cls defaults for backwards compatibility with the MMIT configs. In the
# long run this should be handled by the json configs themselves
cls_defaults = ["train", "val", "tiny", "challenge"]
model = config.init(
name='arch',
module=module_arch,
expert_dims=expert_dims,
text_dim=config["experts"]["text_dim"],
disable_nan_checks=config["disable_nan_checks"],
spatial_feats=config["data_loader"]["args"].get("spatial_feats", False),
task=config.get("task", "retrieval"),
ce_shared_dim=config["experts"].get("ce_shared_dim", None),
feat_aggregation=config["data_loader"]["args"]["feat_aggregation"],
trn_config=trn_config,
trn_cat=config["data_loader"]["args"].get("trn_cat", 0),
)
logger.info(model)
data_loaders = config.init(
name='data_loader',
module=module_data,
logger=logger,
raw_input_dims=raw_input_dims,
text_feat=config["experts"]["text_feat"],
text_dim=config["experts"]["text_dim"],
text_agg=config["experts"]["text_agg"],
use_zeros_for_missing=config["experts"].get("use_zeros_for_missing", False),
task=config.get("task", "retrieval"),
cls_partitions=config.get("cls_partitions", cls_defaults)
)
if config.get("manual_linear_init", False):
logger.info("manually setting init for linear layers")
def init_weights(m):
if isinstance(m, nn.Linear):
torch.nn.init.xavier_uniform(m.weight)
m.bias.data.fill_(0.01)
model.apply(init_weights)
loss = config.init(name="loss", module=module_loss)
metrics = [getattr(module_metric, met) for met in config['metrics']]
trainable_params = filter(lambda p: p.requires_grad, model.parameters())
if config["optimizer"]["type"] == "RAdam":
optimizer = config.init('optimizer', radam, trainable_params)
elif config["optimizer"]["type"] == "Ranger":
optimizer = config.init('optimizer', ranger, trainable_params)
elif config["optimizer"]["type"] == "SWATS":
optimizer = config.init('optimizer', swats, trainable_params)
else:
optimizer = config.init('optimizer', torch.optim, trainable_params)
if config["lr_scheduler"]["type"] == "StepLR":
lr_scheduler = config.init('lr_scheduler', torch.optim.lr_scheduler,
optimizer)
else:
lr_scheduler = config.init('lr_scheduler', cos_restart, optimizer)
visualizer = config.init(
name='visualizer',
module=module_vis,
exp_name=config._exper_name,
web_dir=config._web_log_dir,
)
trainer = Trainer(
model,
loss,
metrics,
optimizer,
config=config,
data_loaders=data_loaders,
lr_scheduler=lr_scheduler,
mini_train=config._args.mini_train,
disable_nan_checks=config["disable_nan_checks"],
visualizer=visualizer,
val_freq=config["trainer"].get("val_freq", 1),
force_cpu_val=config.get("force_cpu_val", False),
skip_first_n_saves=config["trainer"].get("skip_first_n_saves", 0),
include_optim_in_ckpts=config["trainer"].get("include_optim_in_ckpts", 1),
cache_targets=set(config.get("cache_targets", [])),
)
trainer.train()
best_ckpt_path = config.save_dir / "trained_model.pth"
duration = time.strftime('%Hh%Mm%Ss', time.gmtime(time.time() - tic))
logger.info(f"Training took {duration}")
if config._config.get("eval_settings", False):
eval_config = copy.deepcopy(config)
merge(eval_config._config, config["eval_settings"], strategy=Strategy.REPLACE)
eval_config._args.resume = best_ckpt_path
evaluation(eval_config, logger=logger, trainer=trainer)
# If multiple runs were conducted, report relevant statistics
if len(seeds) > 1:
log_summary(
logger=logger,
log_path=config.log_path,
eval_mode=config["eval_mode"],
fixed_num_epochs=config["trainer"]["epochs"],
)
print(f"Log file stored at {config.log_path}")
# Report the location of the "best" checkpoint of the final seeded run (here
# "best" corresponds to the model with the highest geometric mean over the
# R@1, R@5 and R@10 metrics when a validation set is used, or simply the final
# epoch of training for fixed-length schedules).
print(f"The best performing ckpt can be found at {str(best_ckpt_path)}")