def evaluate(args, model):
""" Train the model """
dev_dataset = SequenceDataset(
TextTokenIdsCache(args.preprocess_dir, f"{args.mode}-query"),
args.max_seq_length)
collate_fn = get_collate_function(args.max_seq_length)
batch_size = args.pergpu_eval_batch_size
if args.n_gpu > 1:
batch_size *= args.n_gpu
dev_dataloader = DataLoader(dev_dataset,
batch_size=batch_size,
collate_fn=collate_fn)
if args.n_gpu > 1:
model = torch.nn.DataParallel(model)
qembedding_memmap = np.memmap(args.qmemmap_path,
dtype="float32",
shape=(len(dev_dataset), 768),
mode="w+")
with torch.no_grad():
for step, (batch, qoffsets) in enumerate(tqdm(dev_dataloader)):
batch = {k: v.to(args.model_device) for k, v in batch.items()}
model.eval()
embeddings = model(input_ids=batch["input_ids"],
attention_mask=batch["attention_mask"],
is_query=True)
embeddings = embeddings.detach().cpu().numpy()
qembedding_memmap[qoffsets] = embeddings
return qembedding_memmap
def query_inference(model, args, embedding_size):
if os.path.exists(args.query_memmap_path):
print(f"{args.query_memmap_path} exists, skip inference")
return
query_collator = single_get_collate_function(args.max_query_length)
query_dataset = SequenceDataset(ids_cache=TextTokenIdsCache(
data_dir=args.preprocess_dir, prefix=f"{args.mode}-query"),
max_seq_length=args.max_query_length)
query_memmap = np.memmap(args.query_memmap_path,
dtype=np.float32,
mode="w+",
shape=(len(query_dataset), embedding_size))
queryids_memmap = np.memmap(args.queryids_memmap_path,
dtype=np.int32,
mode="w+",
shape=(len(query_dataset), ))
try:
prediction(model,
query_collator,
args,
query_dataset,
query_memmap,
queryids_memmap,
is_query=True)
except:
subprocess.check_call(["rm", args.query_memmap_path])
subprocess.check_call(["rm", args.queryids_memmap_path])
raise
def get_loader(mode, override_modality=None):
if override_modality:
path_to_lmdb = join(args.path_to_data, override_modality)
else:
path_to_lmdb = join(args.path_to_data,
args.modality) if args.modality != 'fusion' else [
join(args.path_to_data, m)
for m in ['rgb', 'flow', 'obj']
]
kargs = {
'path_to_lmdb': path_to_lmdb,
'path_to_csv': join('./data', f"{mode}.csv"),
'time_step': args.alpha,
'img_tmpl': args.img_tmpl,
'action_samples':
args.S_ant if args.task == 'early_recognition' else None,
'past_features': args.task == 'anticipation',
'sequence_length': args.S_enc + args.S_ant,
'label_type':
['verb', 'noun', 'action'] if args.mode != 'train' else 'action',
'challenge': 'test' in mode
}
_set = SequenceDataset(**kargs)
return DataLoader(_set,
batch_size=args.batch_size,
num_workers=args.num_workers,
pin_memory=True,
shuffle=mode == 'training')
def query_inference(model, args, embedding_size):
query_collator = single_get_collate_function(args.max_query_length)
query_dataset = SequenceDataset(ids_cache=TextTokenIdsCache(
data_dir=args.preprocess_dir, prefix=f"{args.mode}-query"),
max_seq_length=args.max_query_length)
query_memmap = np.memmap(args.query_memmap_path,
dtype=np.float32,
mode="w+",
shape=(len(query_dataset), embedding_size))
queryids_memmap = np.memmap(args.queryids_memmap_path,
dtype=np.int32,
mode="w+",
shape=(len(query_dataset), ))
prediction(model,
query_collator,
args,
query_dataset,
query_memmap,
queryids_memmap,
is_query=True)
def get_loader(mode, override_modality = None):
if override_modality:
path_to_lmdb = join(args.path_to_data, override_modality)
else:
path_to_lmdb = join(args.path_to_data, args.modality) if args.modality != 'fusion' else [join(args.path_to_lmdb_data, m) for m in args.fusion_list]
kargs = {
'path_to_lmdb': path_to_lmdb,
'path_to_csv': join(args.path_to_data, "{}.csv".format(mode)),
'time_step': args.alpha,
'img_tmpl': args.img_tmpl,
'past_features': args.task == 'anticipation',
'sequence_length': 1,
'label_type': ['verb', 'noun', 'action'],
'challenge': 'test' in mode,
'args': args
}
_set = SequenceDataset(**kargs)
return DataLoader(_set, batch_size=args.batch_size, num_workers=args.num_workers,
pin_memory=True, shuffle=mode == 'training')
def __init__(self, queryids_cache, rel_file, max_query_length):
SequenceDataset.__init__(self, queryids_cache, max_query_length)
self.reldict = load_rel(rel_file, direction="query2doc")
if __name__ == "__main__":
data = pd.read_excel('./test.xlsx')
label_names = list(data.columns[3:])
infos = data.to_dict(orient='records')
with open('./emb_matrix.pickle', 'rb') as f:
emb_matrix = pickle.load(f)
with open('./w2id_dict.pickle', 'rb') as f:
w2id_dict = pickle.load(f)
test_dataset = SequenceDataset(
label_names=label_names,
max_len=30,
w2id_dict=w2id_dict,
data_info=infos,
is_train=False
)
test_data_loader = DataLoader(
dataset=test_dataset,
batch_size=10,
shuffle=False,
num_workers=2,
)
model = torch.load("./best.model")
model.to(device)
print(model)
data = pd.read_excel('./data.xlsx')
label_names = list(data.columns[2:])
data = data[(data[label_names] != 0).any(1)] # 砍掉通0
print("label names\n", label_names)
infos = data.to_dict(orient='records')
with open('./emb_matrix.pickle', 'rb') as f:
emb_matrix = pickle.load(f)
with open('./w2id_dict.pickle', 'rb') as f:
w2id_dict = pickle.load(f)
train_dataset = SequenceDataset(label_names=label_names,
max_len=30,
w2id_dict=w2id_dict,
data_info=infos)
train_data_loader = DataLoader(
dataset=train_dataset,
batch_size=32,
shuffle=True,
num_workers=2,
)
print("Data loader size: ", len(train_data_loader))
if args.pretrained == 'T':
print("Pretrained...")
model = torch.load("../pretrain/best.model")
model.output_layer = nn.Linear(in_features=128 * 2,