def __init__(self, device, NUM_FRAMES_PER_STEP=5, DETECTIONS_PER_FRAME=20):
super(ModelFC, self).__init__()
self.device = device
self.NUM_FRAMES_PER_STEP = NUM_FRAMES_PER_STEP
self.DETECTIONS_PER_FRAME = DETECTIONS_PER_FRAME
self.DETECTIONS_PER_STEP = self.NUM_FRAMES_PER_STEP * self.DETECTIONS_PER_FRAME
self.lxmert_tokenizer = LxmertTokenizer.from_pretrained(
"unc-nlp/lxmert-base-uncased")
self.lxmert_tokenizer.add_special_tokens({
"additional_special_tokens": [self.NULL, self.ENTITY, self.ACTION]
})
self.lxmert_tokenizer.encode([self.NULL, self.ENTITY, self.ACTION])
self.NULL_TOKEN = self.lxmert_tokenizer.convert_tokens_to_ids(
self.NULL)
self.ENTITY_TOKEN = self.lxmert_tokenizer.convert_tokens_to_ids(
self.ENTITY)
self.ACTION_TOKEN = self.lxmert_tokenizer.convert_tokens_to_ids(
self.ACTION)
self.lxmert = LxmertModel.from_pretrained(
"unc-nlp/lxmert-base-uncased")
self.lxmert.to(device)
self.VG = LxmertVGHead(self.lxmert.config, self.DETECTIONS_PER_STEP)
self.VG.to(device)
def __init__(self, device, NUM_FRAMES_PER_STEP=5, MAX_DETECTIONS=20):
super(Model, self).__init__()
self.device = device
self.NUM_FRAMES_PER_STEP = NUM_FRAMES_PER_STEP
self.MAX_DETECTIONS = MAX_DETECTIONS
self.CANDIDATES = self.NUM_FRAMES_PER_STEP * self.MAX_DETECTIONS
self.lxmert_tokenizer = LxmertTokenizer.from_pretrained(
"unc-nlp/lxmert-base-uncased")
self.lxmert = LxmertModel.from_pretrained(
"unc-nlp/lxmert-base-uncased")
self.lxmert = nn.DataParallel(self.lxmert)
self.lxmert.to(device)
def test_inference_no_head_absolute_embedding(self):
model = LxmertModel.from_pretrained(LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST[0])
input_ids = torch.tensor([[101, 345, 232, 328, 740, 140, 1695, 69, 6078, 1588, 102]])
num_visual_features = 10
_, visual_feats = np.random.seed(0), np.random.rand(1, num_visual_features, model.config.visual_feat_dim)
_, visual_pos = np.random.seed(0), np.random.rand(1, num_visual_features, 4)
visual_feats = torch.as_tensor(visual_feats, dtype=torch.float32)
visual_pos = torch.as_tensor(visual_pos, dtype=torch.float32)
output = model(input_ids, visual_feats=visual_feats, visual_pos=visual_pos)[0]
expected_shape = torch.Size([1, 11, 768])
self.assertEqual(expected_shape, output.shape)
expected_slice = torch.tensor(
[[[0.2417, -0.9807, 0.1480], [1.2541, -0.8320, 0.5112], [1.4070, -1.1052, 0.6990]]]
)
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
def __init__(self,
NUM_FRAMES_PER_STEP=5,
MAX_DETECTIONS=20,
max_epochs=100,
lr=1e-4,
batch_size=4):
super().__init__()
self.NUM_FRAMES_PER_STEP = NUM_FRAMES_PER_STEP
self.MAX_DETECTIONS = MAX_DETECTIONS
self.CANDIDATES = self.NUM_FRAMES_PER_STEP * self.MAX_DETECTIONS
self.lxmert_tokenizer = LxmertTokenizer.from_pretrained(
"unc-nlp/lxmert-base-uncased")
self.lxmert = LxmertModel.from_pretrained(
"unc-nlp/lxmert-base-uncased")
self.save_hyperparameters()
def __init__(self, opt):
super(Model, self).__init__()
self.opt = opt
self.model = LxmertModel.from_pretrained('unc-nlp/lxmert-base-uncased',
return_dict=True)
self.model = nn.DataParallel(self.model)
self.candi_ans_num = opt.train_candi_ans_num
self.batchsize = opt.batch_size
self.Linear_layer = nn.Linear(768, 1)
norm = opt.norm
activation = opt.activation
dropC = opt.dropC
self.debias_loss_fn = LearnedMixin(0.36)
self.classifier = SimpleClassifier(in_dim=768,
hid_dim=2 * 768,
out_dim=1,
dropout=dropC,
norm=norm,
act=activation)
def test_model_from_pretrained(self):
for model_name in LXMERT_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
model = LxmertModel.from_pretrained(model_name)
model.to(torch_device)
self.assertIsNotNone(model)
