def construct_model(params, saver, logger):
if params['model'] != "CaptionGenerator":
raise Exception('Model Not Implemented')
if params['checkpoint'] is not None:
logger.warn('use checkpoint: %s', params['checkpoint'])
state_dict, params_ = saver.load_model(params['checkpoint'])
else:
raise Exception('checkpoint is None!')
for key in params_:
if 'data' in key:
continue
params[key] = params_[key]
model = CaptionGenerator(
params['hidden_dim'], params['rnn_layer'], params['rnn_cell'],
params['rnn_dropout'], params['bidirectional'],
params['attention_type'], params['context_type'],
params['softmask_scale'], params['vocab_size'],
params['sent_embedding_dim'], params['video_feature_dim'],
params['video_use_residual'], params['max_cap_length'])
model.load_state_dict(state_dict)
logger.info('*' * 100)
sys.stdout.flush()
print(model)
sys.stdout.flush()
logger.info('*' * 100)
return model
def construct_model(params, saver, logger):
if params['checkpoint'] is None:
logger.error('checkpoints are required for evaluation')
exit()
logger.warn('use checkpoint: %s', params['checkpoint'])
state_dict_sl, state_dict_cg, params_ = saver.load_model_slcg(
params['checkpoint'])
# params_['anchor_list'] = ANCHOR_LIST
# params['anchor_list'] = params_['anchor_list']
# params['regressor_scale'] = params_['regressor_scale']
# params = params
for key in params_:
if 'data' in key:
continue
params[key] = params_[key]
params['batch_size'] = 8
model_sl = SentenceLocalizer(
params['hidden_dim'], params['rnn_layer'], params['rnn_cell'],
params['rnn_dropout'], params['bidirectional'],
params['attention_type_sl'], params['regressor_scale'],
params['vocab_size'], params['sent_embedding_dim'],
params['video_feature_dim'], params['fc_dropout'],
params['anchor_list'], params['feature_mixer_type'],
params['video_use_residual'], params['sent_use_residual'])
model_cg = CaptionGenerator(
params['hidden_dim'], params['rnn_layer'], params['rnn_cell'],
params['rnn_dropout'], params['bidirectional'],
params['attention_type_cg'], params['context_type'],
params['softmask_scale'], params['vocab_size'],
params['sent_embedding_dim'], params['video_feature_dim'],
params['video_use_residual'], params['max_cap_length'])
logger.info('*' * 100)
sys.stdout.flush()
print('caption generator' + '*' * 90)
print(model_cg)
print('sentence localizer' + '*' * 90)
print(model_sl)
sys.stdout.flush()
logger.info('*' * 100)
model_sl.load_state_dict(state_dict_sl)
model_cg.load_state_dict(state_dict_cg)
return model_sl, model_cg
def construct_model(params, saver, logger):
if params['model'] != "CaptionGenerator":
raise Exception('Model Not Implemented')
# def __init__(self, hidden_dim, rnn_layer, rnn_cell, rnn_dropout, bidirectional, attention_type, context_type, scale,
# sent_vocab_size, sent_embedding_dim, video_feature_dim, video_use_residual, max_cap_length):
model = CaptionGenerator(params['hidden_dim'], params['rnn_layer'], params['rnn_cell'], params['rnn_dropout'],
params['bidirectional'], params['attention_type'], params['context_type'],
params['softmask_scale'], params['vocab_size'], params['sent_embedding_dim'],
params['video_feature_dim'], params['video_use_residual'], params['max_cap_length'])
logger.info('*' * 100)
sys.stdout.flush()
print(model)
sys.stdout.flush()
logger.info('*' * 100)
if params['checkpoint'] is not None:
logger.warn('use checkpoint: %s', params['checkpoint'])
state_dict, params_ = saver.load_model(params['checkpoint'])
param_refine(params, params_)
model.load_state_dict(state_dict)
return model
def construct_model(params, saver, logger):
params['anchor_list'] = ANCHOR_LIST
if params['checkpoint'] is not None:
state_dict_sl, state_dict_cg, params_ = saver.load_model_slcg(
params['checkpoint'])
params['anchor_list'] = params_['anchor_list']
model_sl = SentenceLocalizer(
params['hidden_dim'], params['rnn_layer'], params['rnn_cell'],
params['rnn_dropout'], params['bidirectional'],
params['attention_type_sl'], params['regressor_scale'],
params['vocab_size'], params['sent_embedding_dim'],
params['video_feature_dim'], params['fc_dropout'],
params['anchor_list'], params['feature_mixer_type'],
params['video_use_residual'], params['sent_use_residual'],
params['pe_video'], params['pe_sent'])
model_cg = CaptionGenerator(
params['hidden_dim'], params['rnn_layer'], params['rnn_cell'],
params['rnn_dropout'], params['bidirectional'],
params['attention_type_cg'], params['context_type'],
params['softmask_scale'], params['vocab_size'],
params['sent_embedding_dim'], params['video_feature_dim'],
params['video_use_residual'], params['max_cap_length'])
logger.info('*' * 100)
sys.stdout.flush()
print('caption generator' + '*' * 90)
print(model_cg)
print('sentence localizer' + '*' * 90)
print(model_sl)
sys.stdout.flush()
logger.info('*' * 100)
if params['checkpoint'] is not None:
logger.warn('use checkpoint: %s', params['checkpoint'])
model_sl.load_state_dict(state_dict_sl)
model_cg.load_state_dict(state_dict_cg)
if params['checkpoint_cg'] is not None:
state_dict_cg, _ = saver.load_model(params['checkpoint_cg'])
logger.warn('use checkpoint: %s', params['checkpoint_cg'])
model_cg.load_state_dict(state_dict_cg)
if params['checkpoint_sl'] is not None:
state_dict_sl, _ = saver.load_model(params['checkpoint_sl'])
logger.warn('use checkpoint: %s', params['checkpoint_sl'])
model_sl.load_state_dict(state_dict_sl)
return model_sl, model_cg