def main(argv):
num_epochs = int(argv[1])
learning_rate = float(argv[2])
experiment = avsr.AVSR(
unit='character',
unit_file=
'/home/nas/user/yong/yong_Sigmedia-AVSR/datasets/lrs3/configs/character_list',
video_processing='features',
batch_normalisation=False,
video_train_record=
'/home/nas/user/yong/LIPREADING/TFRecord/LRS_con/video/VGG_M_pre_05_av_clean.tfrecord',
video_test_record=
'/home/nas/user/yong/LIPREADING/TFRecord/LRS_con/video/VGG_M_test_05_av_clean.tfrecord',
audio_processing='features',
audio_train_record=
'/home/nas/user/yong/LIPREADING/TFRecord/LRS_con/audio/lmda_90_pre_05_av_clean.tfrecord',
audio_test_record=
'/home/nas/user/yong/LIPREADING/TFRecord/LRS_con/audio/lmda_90_test_05_av_clean.tfrecord',
labels_train_record=
'/home/nas/user/yong/LIPREADING/TFRecord/LRS_con/characters_pre_05_av.tfrecord',
labels_test_record=
'/home/nas/user/yong/LIPREADING/TFRecord/LRS_con/characters_test_05_av.tfrecord',
encoder_type='unidirectional',
architecture='dual_av_align_ga',
clip_gradients=True,
max_gradient_norm=1.0,
recurrent_l2_regularisation=0.0001,
cell_type='lstm',
highway_encoder=False,
sampling_probability_outputs=0.00,
embedding_size=256,
dropout_probability=(0.9, 0.9, 0.9),
decoding_algorithm='beam_search',
encoder_units_per_layer=((256, 256, 256), (256, 256, 256)),
decoder_units_per_layer=(512, ),
attention_type=(('scaled_luong', ) * 1, ('scaled_luong', ) * 1),
beam_width=10,
batch_size=(128, 128),
optimiser='Adam',
learning_rate=learning_rate,
num_gpus=1,
write_attention_alignment=True,
)
# uer = experiment.evaluate(
# checkpoint_path='/home/nas/user/yong/LIPREADING/logging/checkpoints/lrs_con_av_av_align_pre_05/',
# )
# print(uer)
# return
experiment.train(
num_epochs=num_epochs,
logfile='/home/nas/user/yong/LIPREADING/logging/logs/test',
try_restore_latest_checkpoint=True)
def main(argv):
num_epochs = int(argv[1])
learning_rate = float(argv[2])
experiment = avsr.AVSR(
unit='character',
unit_file='./avsr/misc/character_list',
video_processing='resnet_cnn',
cnn_filters=(8, 16, 32, 64),
cnn_dense_units=64,
batch_normalisation=True,
video_train_record=
'/run/media/john_tukey/download/datasets/tcdtimit/tfrecords4/rgb36lips_train_sd.tfrecord',
video_test_record=
'/run/media/john_tukey/download/datasets/tcdtimit/tfrecords4/rgb36lips_test_sd.tfrecord',
audio_processing='features',
audio_train_record=
'/run/media/john_tukey/download/datasets/tcdtimit/tfrecords4/logmel_train_sd_stack_clean.tfrecord',
audio_test_record=
'/run/media/john_tukey/download/datasets/tcdtimit/tfrecords4/logmel_test_sd_stack_clean.tfrecord',
labels_train_record=
'/run/media/john_tukey/download/datasets/tcdtimit/tfrecords4/characters_train_sd.tfrecord',
labels_test_record=
'/run/media/john_tukey/download/datasets/tcdtimit/tfrecords4/characters_test_sd.tfrecord',
encoder_type='unidirectional',
architecture='av_align',
clip_gradients=True,
max_gradient_norm=1.0,
recurrent_l2_regularisation=0.0001,
cell_type='gru',
highway_encoder=False,
sampling_probability_outputs=0.1,
embedding_size=128,
dropout_probability=(0.9, 0.9, 0.9),
decoding_algorithm='beam_search',
encoder_units_per_layer=((128, 128), (128, 128)),
decoder_units_per_layer=(128, ),
attention_type=(('scaled_luong', ) * 1, ('scaled_luong', ) * 1),
beam_width=10,
batch_size=(48, 64),
optimiser='AMSGrad',
learning_rate=learning_rate,
num_gpus=1,
)
# uer = experiment.evaluate(
# checkpoint_path='./checkpoints/tcd_video_to_chars/checkpoint.ckp-400',
# )
# print(uer)
# return
experiment.train(num_epochs=num_epochs,
logfile='./logs/tcd_av_to_chars',
try_restore_latest_checkpoint=True)