def __init__(self, config, constants):
self.none_action = config["num_actions"]
self.landmark_names = get_all_landmark_names()
self.text_module = TextPointerModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
self.final_module = SegmentationFinalModule(
text_module=self.text_module,
text_emb_size=4 * constants["lstm_emb_dim"])
if torch.cuda.is_available():
self.text_module.cuda()
self.final_module.cuda()
def __init__(self, config, constants):
AbstractModel.__init__(self, config, constants)
self.none_action = config["num_actions"]
self.image_module = ImagePositionResnetModule(
image_emb_size=constants["image_emb_dim"],
input_num_channels=3 * constants["max_num_images"],
image_height=config["image_height"],
image_width=config["image_width"])
if config["use_pointer_model"]:
self.text_module = TextPointerModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
else:
self.text_module = TextSimpleModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
# total_emb_size = (constants["image_emb_dim"]
# + constants["lstm_emb_dim"])
total_emb_size = constants["image_emb_dim"]
final_module = MultimodalSimplePositionModule(
image_module=self.image_module,
text_module=self.text_module,
total_emb_size=total_emb_size,
num_grid_x=8,
num_grid_y=8,
num_grid_pose=24)
self.final_module = final_module
if torch.cuda.is_available():
self.image_module.cuda()
self.text_module.cuda()
self.final_module.cuda()
def __init__(self, config, constants):
AbstractModel.__init__(self, config, constants)
self.none_action = config["num_actions"]
self.image_module = ImageTextKernelResnetModule(
image_emb_size=constants["image_emb_dim"],
input_num_channels=3,
image_height=config["image_height"],
image_width=config["image_width"],
text_emb_size=constants["lstm_emb_dim"],
using_recurrence=True)
self.image_recurrence_module = RecurrenceSimpleModule(
input_emb_dim=constants["image_emb_dim"],
output_emb_dim=constants["image_emb_dim"])
if config["use_pointer_model"]:
self.text_module = TextPointerModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
else:
self.text_module = TextSimpleModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
self.action_module = ActionSimpleModule(
num_actions=config["num_actions"],
action_emb_size=constants["action_emb_dim"])
if config["use_pointer_model"]:
total_emb_size = (constants["image_emb_dim"] +
4 * constants["lstm_emb_dim"] +
constants["action_emb_dim"])
else:
total_emb_size = (constants["image_emb_dim"] +
constants["lstm_emb_dim"] +
constants["action_emb_dim"])
final_module = MultimodalTextKernelRecurrentSimpleModule(
image_module=self.image_module,
image_recurrence_module=self.image_recurrence_module,
text_module=self.text_module,
action_module=self.action_module,
total_emb_size=total_emb_size,
num_actions=config["num_actions"])
self.final_module = final_module
if torch.cuda.is_available():
self.image_module.cuda()
self.image_recurrence_module.cuda()
self.text_module.cuda()
self.action_module.cuda()
self.final_module.cuda()
def __init__(self, config, constants):
AbstractModel.__init__(self, config, constants)
self.none_action = config["num_actions"]
landmark_names = get_all_landmark_names()
self.radius_module = RadiusModule(15)
self.angle_module = AngleModule(48)
self.landmark_module = LandmarkModule(63)
self.image_module = SymbolicImageModule(
landmark_names=landmark_names,
radius_module=self.radius_module,
angle_module=self.angle_module,
landmark_module=self.landmark_module)
if config["use_pointer_model"]:
self.text_module = TextPointerModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
else:
self.text_module = TextSimpleModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
self.action_module = ActionSimpleModule(
num_actions=config["num_actions"],
action_emb_size=constants["action_emb_dim"])
total_emb_size = (32 * 3 * 63
+ constants["lstm_emb_dim"]
+ constants["action_emb_dim"])
final_module = MultimodalSimpleModule(
image_module=self.image_module,
text_module=self.text_module,
action_module=self.action_module,
total_emb_size=total_emb_size,
num_actions=config["num_actions"])
self.final_module = final_module
if torch.cuda.is_available():
self.image_module.cuda()
self.text_module.cuda()
self.action_module.cuda()
self.final_module.cuda()
self.radius_module.cuda()
self.angle_module.cuda()
self.landmark_module.cuda()
def __init__(self, config, constants):
AbstractIncrementalModel.__init__(self, config, constants)
self.none_action = config["num_actions"]
self.image_module = ImageResnetModule(
image_emb_size=constants["image_emb_dim"],
input_num_channels=3,
image_height=config["image_height"],
image_width=config["image_width"],
using_recurrence=True)
self.num_cameras = 1
self.image_recurrence_module = IncrementalRecurrenceSimpleModule(
input_emb_dim=(constants["image_emb_dim"] * self.num_cameras + constants["action_emb_dim"]),
output_emb_dim=constants["image_emb_dim"])
if config["use_pointer_model"]:
self.text_module = TextPointerModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
else:
self.text_module = TextBiLSTMModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
self.action_module = ActionSimpleModule(
num_actions=config["num_actions"],
action_emb_size=constants["action_emb_dim"])
if config["use_pointer_model"]:
total_emb_size = (constants["image_emb_dim"]
+ 4 * constants["lstm_emb_dim"]
+ constants["action_emb_dim"])
else:
total_emb_size = ((self.num_cameras + 1) * constants["image_emb_dim"]
+ 2 * constants["lstm_emb_dim"]
+ constants["action_emb_dim"])
if config["do_action_prediction"]:
self.action_prediction_module = ActionPredictionModule(
2 * self.num_cameras * constants["image_emb_dim"], constants["image_emb_dim"], config["num_actions"])
else:
self.action_prediction_module = None
if config["do_temporal_autoencoding"]:
self.temporal_autoencoder_module = TemporalAutoencoderModule(
self.action_module, self.num_cameras * constants["image_emb_dim"],
constants["action_emb_dim"], constants["image_emb_dim"])
else:
self.temporal_autoencoder_module = None
if config["do_object_detection"]:
self.landmark_names = get_all_landmark_names()
self.object_detection_module = ObjectDetectionModule(
image_module=self.image_module, image_emb_size=self.num_cameras * constants["image_emb_dim"], num_objects=67)
else:
self.object_detection_module = None
if config["do_symbolic_language_prediction"]:
self.symbolic_language_prediction_module = SymbolicLanguagePredictionModule(
total_emb_size=2 * constants["lstm_emb_dim"])
else:
self.symbolic_language_prediction_module = None
if config["do_goal_prediction"]:
self.goal_prediction_module = GoalPredictionModule(
total_emb_size=32)
else:
self.goal_prediction_module = None
final_module = TmpIncrementalMultimodalDenseValtsRecurrentSimpleModule(
image_module=self.image_module,
image_recurrence_module=self.image_recurrence_module,
text_module=self.text_module,
action_module=self.action_module,
total_emb_size=total_emb_size,
num_actions=config["num_actions"])
self.final_module = final_module
if torch.cuda.is_available():
self.image_module.cuda()
self.image_recurrence_module.cuda()
self.text_module.cuda()
self.action_module.cuda()
self.final_module.cuda()
if self.action_prediction_module is not None:
self.action_prediction_module.cuda()
if self.temporal_autoencoder_module is not None:
self.temporal_autoencoder_module.cuda()
if self.object_detection_module is not None:
self.object_detection_module.cuda()
if self.symbolic_language_prediction_module is not None:
self.symbolic_language_prediction_module.cuda()
if self.goal_prediction_module is not None:
self.goal_prediction_module.cuda()
class TextSegmentationModel(object):
def __init__(self, config, constants):
self.none_action = config["num_actions"]
self.landmark_names = get_all_landmark_names()
self.text_module = TextPointerModule(
emb_dim=constants["word_emb_dim"],
hidden_dim=constants["lstm_emb_dim"],
vocab_size=config["vocab_size"])
self.final_module = SegmentationFinalModule(
text_module=self.text_module,
text_emb_size=4 * constants["lstm_emb_dim"])
if torch.cuda.is_available():
self.text_module.cuda()
self.final_module.cuda()
def get_segmentation_probs(self, agent_observed_state_list):
for aos in agent_observed_state_list:
assert isinstance(aos, AgentObservedState)
# print "batch size:", len(agent_observed_state_list)
# sort list by instruction length
agent_observed_state_list = sorted(
agent_observed_state_list,
key=lambda aos_: len(aos_.get_instruction()),
reverse=True)
instructions = [
aos.get_instruction() for aos in agent_observed_state_list
]
read_pointers = [
aos.get_read_pointers() for aos in agent_observed_state_list
]
instructions_batch = (instructions, read_pointers)
probs_batch = self.final_module(instructions_batch)
return probs_batch
def load_saved_model(self, load_dir):
if torch.cuda.is_available():
torch_load = torch.load
else:
torch_load = lambda f_: torch.load(f_,
map_location=lambda s_, l_: s_)
text_module_path = os.path.join(load_dir, "text_module_state.bin")
self.text_module.load_state_dict(torch_load(text_module_path))
final_module_path = os.path.join(load_dir, "final_module_state.bin")
self.final_module.load_state_dict(torch_load(final_module_path))
def save_model(self, save_dir):
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# save state file for image nn
text_module_path = os.path.join(save_dir, "text_module_state.bin")
torch.save(self.text_module.state_dict(), text_module_path)
# save state file for final nn
final_module_path = os.path.join(save_dir, "final_module_state.bin")
torch.save(self.final_module.state_dict(), final_module_path)
def get_parameters(self):
parameters = list(self.text_module.parameters())
parameters += list(self.final_module.parameters())
return parameters
def get_named_parameters(self):
named_parameters = list(self.text_module.named_parameters())
named_parameters += list(self.final_module.named_parameters())
return named_parameters