class BUTD(Pythia):
def __init__(self, config):
super().__init__(config)
@classmethod
def config_path(cls):
return "configs/models/butd/defaults.yaml"
def build(self):
self._build_word_embedding()
self._init_feature_encoders("image")
self._init_feature_embeddings("image")
self._init_classifier()
self._init_extras()
def _build_word_embedding(self):
self.text_processor = registry.get(self._datasets[0] +
"_text_processor")
self.vocab = self.text_processor.vocab
self.vocab_size = self.vocab.get_size()
self.word_embedding = self.vocab.get_embedding(
torch.nn.Embedding, embedding_dim=self.config.embedding_dim)
self.text_embeddings_out_dim = self.config.embedding_dim
def _init_classifier(self):
self.classifier = ClassifierLayer(
self.config.classifier.type,
in_dim=self.config.classifier.params.feature_dim,
out_dim=self.vocab_size,
**self.config.classifier.params)
def get_optimizer_parameters(self, config):
params = [
{
"params": self.word_embedding.parameters()
},
{
"params": self.image_feature_embeddings_list.parameters()
},
{
"params": self.classifier.parameters()
},
{
"params": self.image_feature_encoders.parameters(),
"lr": (config.optimizer.params.lr * 0.1),
},
]
return params
def prepare_data(self, sample_list, batch_size):
# turn off teacher forcing during beam search
# (otherwise one cannot run beam search on val set)
self.teacher_forcing = self.config.inference.type != "beam_search" and hasattr(
sample_list, "text")
data = {}
if self.teacher_forcing:
caption_lengths, sort_ind = sample_list.caption_len.sort(
dim=0, descending=True)
data["decode_lengths"] = (caption_lengths - 1).tolist()
sample_list.text = sample_list.text[sort_ind]
sample_list.answers = sample_list.answers[sort_ind]
sample_list.image_feature_0 = sample_list.image_feature_0[sort_ind]
data["texts"] = sample_list.text
timesteps = max(data["decode_lengths"])
sample_list.add_field("targets", sample_list.text[:, 1:])
else:
data["texts"] = sample_list.answers.new_full((batch_size, 1),
self.vocab.SOS_INDEX,
dtype=torch.long)
timesteps = self.text_processor.max_length
sample_list.add_field("targets", sample_list.answers[:, 0, 1:])
return data, sample_list, timesteps
def init_hidden_state(self, features):
h = features.new_zeros(
(features.size(0), self.config.classifier.params.hidden_dim),
dtype=torch.float,
)
c = features.new_zeros(
(features.size(0), self.config.classifier.params.hidden_dim),
dtype=torch.float,
)
return h, c
def get_data_t(self, t, data, batch_size_t, prev_output):
if self.teacher_forcing:
# Modify batch_size for timestep t
batch_size_t = sum([l > t for l in data["decode_lengths"]])
elif prev_output is not None and self.config.inference.type == "greedy":
# Adding t-1 output words to data["text"] for greedy decoding
output_softmax = torch.log_softmax(prev_output, dim=1)
_, indices = torch.max(output_softmax, dim=1, keepdim=True)
data["texts"] = torch.cat(
(data["texts"], indices.view(batch_size_t, 1)), dim=1)
# Slice data based on batch_size at timestep t
data["texts"] = data["texts"][:batch_size_t]
if "state" in data:
h1 = data["state"]["td_hidden"][0][:batch_size_t]
c1 = data["state"]["td_hidden"][1][:batch_size_t]
h2 = data["state"]["lm_hidden"][0][:batch_size_t]
c2 = data["state"]["lm_hidden"][1][:batch_size_t]
else:
h1, c1 = self.init_hidden_state(data["texts"])
h2, c2 = self.init_hidden_state(data["texts"])
data["state"] = {"td_hidden": (h1, c1), "lm_hidden": (h2, c2)}
registry.register("{}_lstm_state".format(h1.device), data["state"])
return data, batch_size_t
def forward(self, sample_list):
# Stores the output probabilites.
scores = sample_list.answers.new_ones(
(
sample_list.answers.size(0),
self.text_processor.max_length,
self.vocab_size,
),
dtype=torch.float,
)
if self.config["inference"]["type"] in [
"beam_search", "nucleus_sampling"
]:
decoder = registry.get_decoder_class(
self.config["inference"]["type"])(self.vocab, self.config)
sample_list = decoder.init_batch(sample_list)
batch_size = sample_list.image_feature_0.size(0)
data, sample_list, timesteps = self.prepare_data(
sample_list, batch_size)
output = None
batch_size_t = batch_size
for t in range(timesteps):
data, batch_size_t = self.get_data_t(t, data, batch_size_t, output)
if self.config.inference.type in [
"beam_search", "nucleus_sampling"
]:
pi_t = data["texts"]
else:
pi_t = data["texts"][:, t].unsqueeze(-1)
embedding = self.word_embedding(pi_t)
attention_feature, _ = self.process_feature_embedding(
"image",
sample_list,
embedding[:, 0, :],
batch_size_t=batch_size_t)
output = self.classifier(attention_feature)
# Compute decoding
if self.config.inference.type in [
"beam_search", "nucleus_sampling"
]:
finish, data, batch_size_t = decoder.decode(t, data, output)
if finish:
break
else:
scores[:batch_size_t, t] = output
model_output = {"scores": scores}
if self.config.inference.type in ["beam_search", "nucleus_sampling"]:
model_output["captions"] = decoder.get_result()
return model_output
class Pythia(BaseModel):
def __init__(self, config):
super().__init__(config)
self.config = config
self._global_config = registry.get("config")
self._datasets = self._global_config.datasets.split(",")
@classmethod
def config_path(cls):
return "configs/models/pythia/defaults.yaml"
@classmethod
def format_state_key(cls, key):
return key.replace("fa_history", "fa_context")
def build(self):
self._build_word_embedding()
self._init_text_embeddings("text")
self._init_feature_encoders("image")
self._init_feature_embeddings("image")
self._init_combine_layer("image", "text")
self._init_classifier(self._get_classifier_input_dim())
self._init_extras()
def _build_word_embedding(self):
assert len(self._datasets) > 0
text_processor = registry.get(self._datasets[0] + "_text_processor")
vocab = text_processor.vocab
self.word_embedding = vocab.get_embedding(torch.nn.Embedding,
embedding_dim=300)
def _init_text_embeddings(self, attr="text"):
if "embeddings" not in attr:
attr += "_embeddings"
text_embeddings = []
text_embeddings_list_config = self.config[attr]
embeddings_out_dim = 0
for text_embedding in text_embeddings_list_config:
embedding_type = text_embedding.type
embedding_kwargs = copy.deepcopy(text_embedding.params)
self._update_text_embedding_args(embedding_kwargs)
embedding = TextEmbedding(embedding_type, **embedding_kwargs)
text_embeddings.append(embedding)
embeddings_out_dim += embedding.text_out_dim
setattr(self, attr + "_out_dim", embeddings_out_dim)
setattr(self, attr, nn.ModuleList(text_embeddings))
def _update_text_embedding_args(self, args):
# Add model_data_dir to kwargs
args.model_data_dir = self.config.model_data_dir
def _init_feature_encoders(self, attr):
feat_encoders = []
feat_encoders_list_config = self.config[attr + "_feature_encodings"]
feature_dim = self.config[attr + "_feature_dim"]
setattr(self, attr + "_feature_dim", feature_dim)
for feat_encoder in feat_encoders_list_config:
feat_encoder_config = copy.deepcopy(feat_encoder)
feat_encoder_config.params.model_data_dir = self.config.model_data_dir
feat_model = build_image_encoder(feat_encoder_config,
direct_features=True)
feat_encoders.append(feat_model)
setattr(self, attr + "_feature_dim", feat_model.out_dim)
setattr(self, attr + "_feature_encoders", nn.ModuleList(feat_encoders))
def _init_feature_embeddings(self, attr):
feature_embeddings_list = []
num_feature_feat = len(
getattr(self.config, f"{attr}_feature_encodings"))
self.feature_embeddings_out_dim = 0
for _ in range(num_feature_feat):
feature_embeddings = []
feature_attn_model_list = self.config[attr + "_feature_embeddings"]
for feature_attn_model_params in feature_attn_model_list:
feature_embedding = ImageFeatureEmbedding(
getattr(self, attr + "_feature_dim"),
self.text_embeddings_out_dim,
**feature_attn_model_params,
)
feature_embeddings.append(feature_embedding)
self.feature_embeddings_out_dim += feature_embedding.out_dim
feature_embeddings = nn.ModuleList(feature_embeddings)
feature_embeddings_list.append(feature_embeddings)
self.feature_embeddings_out_dim *= getattr(self, attr + "_feature_dim")
setattr(self, attr + "_feature_embeddings_out_dim",
self.feature_embeddings_out_dim)
del self.feature_embeddings_out_dim
setattr(
self,
attr + "_feature_embeddings_list",
nn.ModuleList(feature_embeddings_list),
)
def _get_embeddings_attr(self, attr):
embedding_attr1 = attr
if hasattr(self, attr + "_embeddings_out_dim"):
embedding_attr1 = attr + "_embeddings_out_dim"
else:
embedding_attr1 = attr + "_feature_embeddings_out_dim"
return embedding_attr1
def _init_combine_layer(self, attr1, attr2):
config_attr = attr1 + "_" + attr2 + "_modal_combine"
multi_modal_combine_layer = ModalCombineLayer(
self.config[config_attr].type,
getattr(self, self._get_embeddings_attr(attr1)),
getattr(self, self._get_embeddings_attr(attr2)),
**self.config[config_attr].params,
)
setattr(
self,
attr1 + "_" + attr2 + "_multi_modal_combine_layer",
multi_modal_combine_layer,
)
def _init_classifier(self, combined_embedding_dim):
# TODO: Later support multihead
num_choices = registry.get(self._datasets[0] + "_num_final_outputs")
self.classifier = ClassifierLayer(
self.config.classifier.type,
in_dim=combined_embedding_dim,
out_dim=num_choices,
**self.config.classifier.params,
)
def _init_extras(self):
self.inter_model = None
def get_optimizer_parameters(self, config):
combine_layer = self.image_text_multi_modal_combine_layer
params = [
{
"params": self.word_embedding.parameters()
},
{
"params": self.image_feature_embeddings_list.parameters()
},
{
"params": self.text_embeddings.parameters()
},
{
"params": combine_layer.parameters()
},
{
"params": self.classifier.parameters()
},
{
"params": self.image_feature_encoders.parameters(),
"lr": (config.optimizer.params.lr * 0.1),
},
]
return params
def _get_classifier_input_dim(self):
return self.image_text_multi_modal_combine_layer.out_dim
def process_text_embedding(self,
sample_list,
embedding_attr="text_embeddings",
info=None):
text_embeddings = []
# Get "text" attribute in case of "text_embeddings" case
# and "context" attribute in case of "context_embeddings"
texts = getattr(sample_list, embedding_attr.split("_")[0])
# Get embedding models
text_embedding_models = getattr(self, embedding_attr)
for text_embedding_model in text_embedding_models:
# TODO: Move this logic inside
if isinstance(text_embedding_model, PreExtractedEmbedding):
embedding = text_embedding_model(sample_list.question_id)
else:
embedding = text_embedding_model(texts)
text_embeddings.append(embedding)
text_embeddding_total = torch.cat(text_embeddings, dim=1)
return text_embeddding_total
def process_feature_embedding(self,
attr,
sample_list,
text_embedding_total,
extra=None,
batch_size_t=None):
if extra is None:
extra = []
feature_embeddings = []
feature_attentions = []
features = []
batch_size_t = (sample_list.get_batch_size()
if batch_size_t is None else batch_size_t)
# Convert list of keys to the actual values
extra = sample_list.get_fields(extra)
feature_idx = 0
# Get all of the features, which are in the form, "image_feature_0"
# "image_feature_1" ...
while True:
feature = getattr(sample_list, f"{attr}_feature_{feature_idx:d}",
None)
if feature is None:
break
feature_idx += 1
feature = feature[:batch_size_t]
features.append(feature)
feature_encoders = getattr(self, attr + "_feature_encoders")
# Each feature should have a separate image feature encoders
assert len(features) == len(feature_encoders), (
"Number of feature encoders, {} are not equal "
"to number of features, {}.".format(len(feature_encoders),
len(features)))
# Now, iterate to get final attended image features
for i, feature in enumerate(features):
# Get info related to the current feature. info is generally
# in key of format "image_info_0" for 0th feature
feature_info = getattr(sample_list, f"{attr}_info_{i:d}", {})
# For Pythia, we need max_features to mask attention
feature_dim = getattr(feature_info, "max_features", None)
if feature_dim is not None:
feature_dim = feature_dim[:batch_size_t]
# Attribute in which encoders are saved, for "image" it
# will be "image_feature_encoders", other example is
# "context_feature_encoders"
encoders_attr = attr + "_feature_encoders"
feature_encoder = getattr(self, encoders_attr)[i]
# Encode the features
encoded_feature = feature_encoder(feature)
# Get all of the feature embeddings
list_attr = attr + "_feature_embeddings_list"
feature_embedding_models = getattr(self, list_attr)[i]
# Forward through these embeddings one by one
for feature_embedding_model in feature_embedding_models:
inp = (encoded_feature, text_embedding_total, feature_dim,
extra)
embedding, attention = feature_embedding_model(*inp)
feature_embeddings.append(embedding)
feature_attentions.append(attention.squeeze(-1))
# Concatenate all features embeddings and return along with attention
feature_embedding_total = torch.cat(feature_embeddings, dim=1)
return feature_embedding_total, feature_attentions
def combine_embeddings(self, *args):
feature_names = args[0]
feature_embeddings = args[1]
layer = "_".join(feature_names) + "_multi_modal_combine_layer"
return getattr(self, layer)(*feature_embeddings)
def calculate_logits(self, joint_embedding, **kwargs):
return self.classifier(joint_embedding)
def forward(self, sample_list):
sample_list.text = self.word_embedding(sample_list.text)
text_embedding_total = self.process_text_embedding(sample_list)
image_embedding_total, _ = self.process_feature_embedding(
"image", sample_list, text_embedding_total)
if self.inter_model is not None:
image_embedding_total = self.inter_model(image_embedding_total)
joint_embedding = self.combine_embeddings(
["image", "text"], [image_embedding_total, text_embedding_total])
model_output = {"scores": self.calculate_logits(joint_embedding)}
return model_output
class MoVieMcan(BaseModel):
def __init__(self, config):
super().__init__(config)
self.config = config
self._global_config = registry.get("config")
self._datasets = self._global_config.datasets.split(",")
@classmethod
def config_path(cls):
return "configs/models/movie_mcan/defaults.yaml"
def build(self):
self.image_feature_dim = 2048
self._build_word_embedding()
self._init_text_embeddings("text")
self._init_feature_encoders("image")
self._init_feature_embeddings("image")
self._init_combine_layer("image", "text")
self._init_classifier(self._get_classifier_input_dim())
self._init_extras()
def _build_word_embedding(self):
assert len(self._datasets) > 0
text_processor = registry.get(self._datasets[0] + "_text_processor")
vocab = text_processor.vocab
self.word_embedding = vocab.get_embedding(torch.nn.Embedding,
embedding_dim=300)
def _init_text_embeddings(self, attr: str = "text"):
if "embeddings" not in attr:
attr += "_embeddings"
module_config = self.config[attr]
embedding_type = module_config.type
embedding_kwargs = copy.deepcopy(module_config.params)
self._update_text_embedding_args(embedding_kwargs)
embedding = TextEmbedding(embedding_type, **embedding_kwargs)
embeddings_out_dim = embedding.text_out_dim
setattr(self, attr + "_out_dim", embeddings_out_dim)
setattr(self, attr, embedding)
def _update_text_embedding_args(self, args):
# Add model_data_dir to kwargs
args.model_data_dir = self.config.model_data_dir
def _init_feature_encoders(self, attr: str):
feat_encoder = self.config[attr + "_feature_encodings"]
feature_dim = self.config[attr + "_feature_dim"]
setattr(self, attr + "_feature_dim", feature_dim)
feat_encoder_config = copy.deepcopy(feat_encoder)
with omegaconf.open_dict(feat_encoder_config):
feat_encoder_config.params.model_data_dir = self.config.model_data_dir
feat_encoder_config.params.in_dim = feature_dim
feat_model = build_image_encoder(feat_encoder_config,
direct_features=True)
setattr(self, attr + "_feature_dim", feat_model.out_dim)
setattr(self, attr + "_feature_encoders", feat_model)
def _init_feature_embeddings(self, attr: str):
embedding_kwargs = self.config[attr + "_feature_embeddings"]["params"]
setattr(self, attr + "_feature_embeddings_out_dim",
embedding_kwargs["hidden_dim"])
assert (getattr(self, attr + "_feature_embeddings_out_dim") ==
self.text_embeddings_out_dim), "dim1: {}, dim2: {}".format(
getattr(self, attr + "_feature_embeddings_out_dim"),
self.text_embeddings_out_dim,
)
feature_embedding = TwoBranchEmbedding(
getattr(self, attr + "_feature_dim"), **embedding_kwargs)
setattr(self, attr + "_feature_embeddings_list", feature_embedding)
def _get_embeddings_attr(self, attr: str):
embedding_attr1 = attr
if hasattr(self, attr + "_embeddings_out_dim"):
embedding_attr1 = attr + "_embeddings_out_dim"
else:
embedding_attr1 = attr + "_feature_embeddings_out_dim"
return embedding_attr1
def _init_combine_layer(self, attr1: str, attr2: str):
multi_modal_combine_layer = BranchCombineLayer(
getattr(self, self._get_embeddings_attr(attr1)),
getattr(self, self._get_embeddings_attr(attr2)),
)
setattr(
self,
attr1 + "_" + attr2 + "_multi_modal_combine_layer",
multi_modal_combine_layer,
)
def _init_classifier(self, combined_embedding_dim: int):
# TODO: Later support multihead
num_choices = registry.get(self._datasets[0] + "_num_final_outputs")
params = self.config["classifier"].get("params")
if params is None:
params = {}
self.classifier = ClassifierLayer(self.config.classifier.type,
in_dim=combined_embedding_dim,
out_dim=num_choices,
**params)
def _init_extras(self):
self.inter_model = None
def get_optimizer_parameters(self,
config: DictConfig) -> List[Dict[str, Any]]:
combine_layer = self.image_text_multi_modal_combine_layer
params = [
{
"params": filter_grads(self.word_embedding.parameters())
},
{
"params":
filter_grads(
self.image_feature_embeddings_list.sga.parameters())
},
{
"params":
filter_grads(
self.image_feature_embeddings_list.sga_pool.parameters())
},
{
"params":
filter_grads(
self.image_feature_embeddings_list.cbn.parameters()),
"lr": (config.optimizer.params.lr *
config.training.encoder_lr_multiply),
},
{
"params": filter_grads(self.text_embeddings.parameters())
},
{
"params": filter_grads(combine_layer.parameters())
},
{
"params": filter_grads(self.classifier.parameters())
},
{
"params":
filter_grads(self.image_feature_encoders.parameters())
},
]
return params
def get_mapping(self):
mapping = [
"word_embedding",
"image_feature_embeddings_list_sga",
"image_feature_embeddings_list_sga_pool",
"image_feature_embeddings_list_cbn",
"text_embeddings",
"combine_layer",
"classifier",
"image_feature_encoders",
]
return mapping
def _get_classifier_input_dim(self):
return self.image_text_multi_modal_combine_layer.out_dim
def process_text_embedding(
self,
sample_list: Dict[str, Any],
embedding_attr: str = "text_embeddings"
) -> Tuple[torch.Tensor, torch.Tensor]:
# Get "text" attribute in case of "text_embeddings" case
# and "context" attribute in case of "context_embeddings"
texts = getattr(sample_list, embedding_attr.split("_")[0])
# Get embedding models
text_embedding_model = getattr(self, embedding_attr)
# TODO: Move this logic inside
if isinstance(text_embedding_model, PreExtractedEmbedding):
text_embedding_total = text_embedding_model(
sample_list.question_id)
else:
text_embedding_total, text_embedding_vec = text_embedding_model(
texts, sample_list.text_mask)
return text_embedding_total, text_embedding_vec
def process_feature_embedding(
self,
attr: str,
sample_list: Dict[str, Any],
text_embedding_total: torch.Tensor,
text_embedding_vec: torch.Tensor,
extra: list = [],
batch_size_t: Optional[int] = None,
):
batch_size_t = (sample_list.get_batch_size()
if batch_size_t is None else batch_size_t)
# Convert list of keys to the actual values
if hasattr(sample_list, "image"):
feature = sample_list.image
feature_encoder = getattr(self, attr + "_feature_encoders")
encoded_feature = feature_encoder(feature, text_embedding_vec)
else:
feature = sample_list.image_feature_0
feature_encoder = getattr(self, attr + "_feature_encoders")
encoded_feature = feature_encoder(feature)
feature_embedding = getattr(self, attr + "_feature_embeddings_list")
feature_sga, feature_cbn = feature_embedding(
encoded_feature,
text_embedding_total,
text_embedding_vec,
None,
sample_list.text_mask,
)
return feature_sga, feature_cbn
def combine_embeddings(self, *args):
feature_names = args[0]
v1, v2, q = args[1]
layer = "_".join(feature_names) + "_multi_modal_combine_layer"
return getattr(self, layer)(v1, v2, q)
def calculate_logits(self, joint_embedding: torch.Tensor, **kwargs):
return self.classifier(joint_embedding)
def forward(self, sample_list: Dict[str, Any]) -> Dict[str, torch.Tensor]:
sample_list.text_mask = sample_list.text.eq(0)
sample_list.text = self.word_embedding(sample_list.text)
text_embedding_total, text_embedding_vec = self.process_text_embedding(
sample_list)
feature_sga, feature_cbn = self.process_feature_embedding(
"image", sample_list, text_embedding_total, text_embedding_vec[:,
0])
joint_embedding = self.combine_embeddings(
["image", "text"],
[feature_sga, feature_cbn, text_embedding_vec[:, 1]])
model_output = {"scores": self.calculate_logits(joint_embedding)}
return model_output
def build_classifier_layer(config, *args, **kwargs):
from mmf.modules.layers import ClassifierLayer
classifier = ClassifierLayer(config.type, *args, **config.params, **kwargs)
return classifier.module
