'The width of the policy conv layer.')
flags.DEFINE_integer('value_conv_width', 1,
'The width of the value conv layer.')
flags.DEFINE_integer('fc_width', 256 if go.N == 19 else 64,
'The width of the fully connected layer in value head.')
flags.DEFINE_integer('trunk_layers', go.N,
'The number of resnet layers in the shared trunk.')
flags.DEFINE_multi_integer(
'lr_boundaries', [400000, 600000],
'The number of steps at which the learning rate will decay')
flags.DEFINE_multi_float('lr_rates', [0.01, 0.001, 0.0001],
'The different learning rates')
flags.register_multi_flags_validator(
['lr_boundaries', 'lr_rates'],
lambda flags: len(flags['lr_boundaries']) == len(flags['lr_rates']) - 1,
'Number of learning rates must be exactly one greater than the number of boundaries'
)
flags.DEFINE_float('l2_strength', 1e-4,
'The L2 regularization parameter applied to weights.')
flags.DEFINE_float(
'value_cost_weight', 1.0,
'Scalar for value_cost, AGZ paper suggests 1/100 for '
'supervised learning')
'final batch size will be = train_batch_size * num_tpu_cores')
flags.DEFINE_integer('conv_width', 128 if go.N == 19 else 32,
'The width of each conv layer in the shared trunk.')
flags.DEFINE_integer('fc_width', 256 if go.N == 19 else 64,
'The width of the fully connected layer in value head.')
flags.DEFINE_integer('trunk_layers', go.N,
'The number of resnet layers in the shared trunk.')
flags.DEFINE_multi_integer(
'lr_boundaries', [10000, 30000],
'The number of steps at which the learning rate will decay')
flags.DEFINE_multi_float('lr_rates', [2e-2, 1e-3, 1e-4],
'The different learning rates')
flags.DEFINE_float('l2_strength', 1e-4,
'The L2 regularization parameter applied to weights.')
flags.DEFINE_float('sgd_momentum', 0.9,
'Momentum parameter for learning rate.')
flags.DEFINE_string('model_dir', None, 'The working directory of the model')
# See www.moderndescartes.com/essays/shuffle_viz for discussion on sizing
flags.DEFINE_integer('shuffle_buffer_size', 20000,
'Size of buffer used to shuffle train examples.')
flags.DEFINE_bool('use_tpu', False, 'Whether to use TPU for training.')
from absl import app, flags
from mlperf_logging import mllog
import os
flags.DEFINE_multi_float('lr_rates', None, 'lr rates')
flags.DEFINE_multi_float('lr_boundaries', None, 'learning rate boundaries')
flags.DEFINE_float('l2_strength', None, 'weight decay')
flags.DEFINE_integer('conv_width', None, 'conv width')
flags.DEFINE_integer('fc_width', None, 'fc width')
flags.DEFINE_integer('trunk_layers', None, 'trunk layers')
flags.DEFINE_float('value_cost_weight', None, 'value cost weight')
flags.DEFINE_integer('summary_steps', None, 'summary steps')
flags.DEFINE_integer('bool_features', None, 'bool features')
flags.DEFINE_string('input_features', None, 'input features')
flags.DEFINE_string('input_layout', None, 'input layout')
flags.DEFINE_integer('shuffle_buffer_size', None, 'shuffle buffer size')
flags.DEFINE_boolean('shuffle_examples', None, 'shuffle examples')
flags.DEFINE_integer('keep_checkpoint_max', None, 'keep_checkpoint_max')
flags.DEFINE_integer('train_batch_size', None, 'train_batch_size')
FLAGS = flags.FLAGS
def main(argv):
mllogger = mllog.get_mllogger()
mllog.config(filename="train.log")
mllog.config(default_namespace="worker1",
default_stack_offset=1,
default_clear_line=False)
help='use fp16 compression during allreduce')
flags.DEFINE_integer('BATCHES_PER_ALLREDUCE', default=1,
help='number of batches processed locally before '
'executing allreduce across workers; it multiplies '
'total batch size.')
flags.DEFINE_boolean('USE_ADASUM', default=False,
help='use adasum algorithm to do reduction')
flags.DEFINE_integer('LOG_INTERVAL', default=10,
help='how many batches to wait before logging training status')
# Default settings from https://arxiv.org/abs/1706.02677.
flags.DEFINE_string('MODEL_NAME', 'MobileNetV2',
help='The name of the architecture to train.')
flags.DEFINE_string('DATASET_NAME', 'CIFAR100',
help='The name of the dataset to train.')
flags.DEFINE_multi_float('DATA_MEAN', [0.5071, 0.4867, 0.4408],
help='mean value of dataset')
flags.DEFINE_multi_float('DATA_STD', [0.2675, 0.2565, 0.2761],
help='standard deviation value of dataset')
flags.DEFINE_multi_integer('DATA_SHAPE', [3, 32, 32],
help='data dimension of dataset')
flags.DEFINE_list('BLOCK_ARGS', ['wm1.0_rn8_s1',
't1_c16_n1_s1',
't6_c24_n2_s1',
't6_c32_n3_s2',
't6_c64_n4_s2',
't6_c96_n3_s1',
't6_c160_n3_s2',
't6_c320_n1_s1'],
help='argument of blocks in EfficientNet style')
flags.DEFINE_integer('BATCH_SIZE', default=128,
help='input batch size for training')
flags.DEFINE_boolean(
'load_full_weights', False,
'Load full COCO pretrained weights including those of the last detection layers'
)
flags.DEFINE_multi_integer(
'model_size', (608, 608),
'Resolution of DNN input, must be the multiples of 32')
flags.DEFINE_integer('max_out_size', 1,
'maximum detected object amount of one class')
flags.DEFINE_float('iou_threshold', 0.5, 'threshold of non-max suppression')
flags.DEFINE_float('confid_threshold', 0.5, 'threshold of confidence')
flags.DEFINE_float('brightness_delta', 0.3,
'brightness_delta of data augmentation')
flags.DEFINE_multi_float('contrast_range', (0.5, 1.5),
'contrast_range of data augmentation')
flags.DEFINE_float('hue_delta', 0.2,
'hue_delta of data augmentation, only between (0, 0.5)')
flags.DEFINE_float('probability', 0.8, 'percentage of augmented images')
# Anchors of k-means threshold=0.98
_ANCHORS = [(77.0, 91.0), (89.0, 93.0), (83.0, 101.0), (95.0, 102.0),
(92.0, 111.0), (104.0, 108.0), (98.0, 117.0), (110.0, 122.0),
(127.0, 134.0)]
# Anchors of k-means threshold=0.99
# _ANCHORS = [(77.0, 92.0), (89.0, 93.0), (83.0, 101.0), (95.0, 101.0), (93.0, 111.0), (105.0, 109.0), (101.0, 119.0), (114.0, 125.0), (151.0, 152.0)]
# Default anchors of COCO
# _ANCHORS = [(10,13),(16,30),(33,23),(30,61),(62,45),(59,119),(116,90),(156,198),(373,326)]
loss_fns = {
'bce': losses.BCEWithLogits,
'hinge': losses.Hinge,
'was': losses.Wasserstein,
'softplus': losses.Softplus
}
FLAGS = flags.FLAGS
# model and training
flags.DEFINE_enum('dataset', 'cifar10', ['cifar10', 'stl10'], "dataset")
flags.DEFINE_enum('arch', 'cnn32', net_G_models.keys(), "architecture")
flags.DEFINE_integer('total_steps', 100000, "total number of training steps")
flags.DEFINE_integer('batch_size', 128, "batch size")
flags.DEFINE_float('lr_G', 2e-4, "Generator learning rate")
flags.DEFINE_float('lr_D', 2e-4, "Discriminator learning rate")
flags.DEFINE_multi_float('betas', [0.0, 0.9], "for Adam")
flags.DEFINE_integer('n_dis', 5, "update Generator every this steps")
flags.DEFINE_integer('z_dim', 128, "latent space dimension")
flags.DEFINE_float('c', 0.1, "clip value")
flags.DEFINE_enum('loss', 'was', loss_fns.keys(), "loss function")
flags.DEFINE_integer('seed', 0, "random seed")
# logging
flags.DEFINE_integer('eval_step', 5000, "evaluate FID and Inception Score")
flags.DEFINE_integer('sample_step', 500, "sample image every this steps")
flags.DEFINE_integer('sample_size', 64, "sampling size of images")
flags.DEFINE_string('logdir', './logs/WGAN_CIFAR10_CNN', 'logging folder')
flags.DEFINE_bool('record', True, "record inception score and FID")
flags.DEFINE_string('fid_cache', './stats/cifar10.train.npz', 'FID cache')
# generate
flags.DEFINE_bool('generate', False, 'generate images')
flags.DEFINE_string('pretrain', None, 'path to test model')
'Image crop size [height, width] for evaluation.')
flags.DEFINE_integer('eval_interval_secs', 60 * 5,
'How often (in seconds) to run evaluation.')
# For `xception_65`, use atrous_rates = [12, 24, 36] if output_stride = 8, or
# rates = [6, 12, 18] if output_stride = 16. For `mobilenet_v2`, use None. Note
# one could use different atrous_rates/output_stride during training/evaluation.
flags.DEFINE_multi_integer('atrous_rates', None,
'Atrous rates for atrous spatial pyramid pooling.')
flags.DEFINE_integer('output_stride', 16,
'The ratio of input to output spatial resolution.')
# Change to [0.5, 0.75, 1.0, 1.25, 1.5, 1.75] for multi-scale test.
flags.DEFINE_multi_float('eval_scales', [1.0],
'The scales to resize images for evaluation.')
# Change to True for adding flipped images during test.
flags.DEFINE_bool('add_flipped_images', False,
'Add flipped images for evaluation or not.')
flags.DEFINE_integer(
'quantize_delay_step', -1,
'Steps to start quantized training. If < 0, will not quantize model.')
# Dataset settings.
flags.DEFINE_string('dataset', 'pascal_voc_seg',
'Name of the segmentation dataset.')
flags.DEFINE_string('eval_split', 'val',
from absl import app, flags
from time import gmtime, strftime
import json
FLAGS = flags.FLAGS
flags.DEFINE_integer('N', 50, 'number of images')
flags.DEFINE_integer('seed', 42,
'random seed for sampling images from ImageNet')
flags.DEFINE_multi_integer('attack_size', [20, 20], 'size of sticker')
flags.DEFINE_integer('stride', 20, 'stride of sticker')
flags.DEFINE_integer('batchsize', 128, 'batch size')
flags.DEFINE_string('model', 'bagnet33', 'model being evaluated')
flags.DEFINE_string('clip_fn', 'tanh_linear', 'clipping function')
flags.DEFINE_string('param', 'a',
'clip(a*x + b). Which parameter are going to be tested')
flags.DEFINE_multi_float('param_list', None, 'list of parameters to be tested')
flags.DEFINE_float('fixed_param', None, 'the other fixed parameter')
flags.DEFINE_string('data_path', '/mnt/data/imagenet',
'directory where data are stored')
flags.DEFINE_string('output_root', '/mnt/data/clipping_params_searching/',
'directory for storing results')
def main(argv):
"""
FLAGS.output_root/
[NAME]/
[NAME].log
[Name].lst
"""
assert FLAGS.param in ['a', 'b'], 'FLAGS.param must be either a or b'
default='/usr/local/srv/tfrecords/train/*2007*.tfrecords',
help="Dataset glob for train")
flags.DEFINE_string('val_dataset',
default='/usr/local/srv/tfrecords/val/*2007*.tfrecords',
help="Dataset glob for validate")
flags.DEFINE_string('test_dataset',
default='/usr/local/srv/tfrecords/test/*2007*.tfrecords',
help="Dataset glob for test")
flags.DEFINE_string('anchors_path',
default='model_data/yolo_anchors.txt',
help="Anchors path")
flags.DEFINE_string('classes_path',
default='model_data/voc_classes.txt',
help="Classes Path")
flags.DEFINE_multi_float('learning_rate',
default=[1e-3, 1e-4],
lower_bound=0,
help="Learning rate")
flags.DEFINE_enum_class(
'opt',
default=None,
enum_class=OPT,
help="Select optimization, One of {'XLA','DEBUG','MKL'}")
flags.DEFINE_string('tpu_address', default=None, help="TPU address")
flags.DEFINE_bool('freeze', default=False, help="Whether freeze backbone")
flags.DEFINE_bool('prune', default=False, help="Whether prune model")
def parse_tuple(val):
if isinstance(val, str):
return tuple([int(num) for num in val[1:-1].split(',')])
return tuple(val)
'Parent directory in which output TFRecords will be saved.')
flags.DEFINE_string('data_dir', None,
'<data_dir>/<task> contains output of create_splits.py.')
flags.DEFINE_enum('loader_cls', 'Pfam34Loader', ['Pfam34Loader'],
'The loader class to fetch the unpaired sequences with.')
flags.DEFINE_string('task', 'iid_ood_clans',
'Task for which to generate TFRecords.')
flags.DEFINE_string('split', 'train',
'Data split for which to generate TFRecords.')
flags.DEFINE_integer('max_len', 512,
'Maximum sequence length, including any special tokens.')
flags.DEFINE_multi_string(
'index_keys', ['fam_key', 'ci_100'],
'Indexing keys for stratified sampling of sequence pairs.')
flags.DEFINE_multi_float(
'smoothing', [1.0, 1.0],
'Smoothing coefficients for stratified sampling of sequence pairs.')
flags.DEFINE_string(
'branch_key', 'ci_100',
'Branching key for stratified sampling of sequence pairs.')
flags.DEFINE_integer('seed', 0, 'PRNG seed to generate the shard with.')
flags.DEFINE_integer('n_pairs', 102400, 'Number of sequence pairs per shard.')
FLAGS = flags.FLAGS
LOADERS = {
'Pfam34Loader': specs.make_pfam34_loader,
}
def make_serialize_fn():
"""Creates a serialization function for paired examples."""
flags.DEFINE_string('dataset_path', None, 'TFRecord dataset path.')
flags.DEFINE_integer('learner_iterations_per_call', 500,
'Iterations per learner run call.')
flags.DEFINE_integer('policy_save_interval', 10000, 'Policy save interval.')
flags.DEFINE_integer('eval_interval', 10000, 'Evaluation interval.')
flags.DEFINE_integer('summary_interval', 1000, 'Summary interval.')
flags.DEFINE_integer('num_gradient_updates', 1000000,
'Total number of train iterations to perform.')
flags.DEFINE_float(
'reward_shift', 0.0, 'Value to add to reward. Useful for sparse rewards, '
'e.g. set to -0.5 for optimal performance on AntMaze environments which '
'have rewards of 0 (most often) or 1 (when the target position is reached)'
)
flags.DEFINE_multi_float(
'action_clipping', None, 'Optional (min, max) values to clip actions. '
'e.g. set to (-0.995, 0.995) when actions are close to -1 and 1 since'
'tanh_distribution.log_prob(actions) will yield -inf and inf and make '
'actor loss NaN. ')
flags.DEFINE_bool(
'use_trajectories', False,
'Whether dataset samples are stored as trajectories. '
'If False, stored as transitions')
flags.DEFINE_multi_string('gin_file', None, 'Paths to the gin-config files.')
flags.DEFINE_multi_string('gin_param', None, 'Gin binding parameters.')
@gin.configurable
def train_eval(
root_dir,
dataset_path,
env_name,
'each move.')
flags.DEFINE_float(
'mlperf_value_init_penalty', -1.0,
'New children value initialization penalty. '
'Child value = parent\'s value - penalty * color, '
'clamped to [-1, 1]. Penalty should be in [0.0, 2.0]. '
'0 is init-to-parent, 2.0 is init-to-loss [default]. '
'This behaves similiarly to Leela\'s FPU '
'"First Play Urgency".')
flags.DEFINE_float('mlperf_holdout_pct', -1.0,
'Fraction of games to hold out for validation.')
flags.DEFINE_float('mlperf_disable_resign_pct', -1.0,
'Fraction of games to disable resignation for.')
flags.DEFINE_multi_float(
'mlperf_resign_threshold', [0.0, 0.0],
'Each game\'s resign threshold is picked randomly '
'from the range '
'[min_resign_threshold, max_resign_threshold)')
flags.DEFINE_integer(
'mlperf_parallel_games', -1,
'Number of games to play concurrently on each selfplay '
'thread. Inferences from a thread\'s concurrent games are '
'batched up and evaluated together. Increasing '
'concurrent_games_per_thread can help improve GPU or '
'TPU utilization, especially for small models.')
flags.DEFINE_integer('mlperf_virtual_losses', -1,
'Number of virtual losses when running tree search')
flags.DEFINE_float(
'mlperf_gating_win_rate', -1.0,
'Win pct against the target model to define a converged '
'model.')
flags.DEFINE_float('max_v', 2500, 'Maximum value on the y/z-axis.')
flags.DEFINE_float('min_v', 0, 'Minimum value on the y/z-axis.')
# The title for each parameter vector passed above.
flags.DEFINE_string('title1', 'p1', 'Title for p1.')
flags.DEFINE_string('title2', 'p2', 'Title for p2.')
flags.DEFINE_string('title3', 'p3', 'Title for p3.')
# Allows skipping of interpolating just to do plotting.
flags.DEFINE_bool(
'visualize_only', False, 'Will only do visualization. '
'If passed must specify --precomputed_interpolation')
flags.DEFINE_string('precomputed_interpolation', None,
'File path to a precomputed interpolation.')
# Policy and environment details.
flags.DEFINE_string('env', 'Hopper-v1', 'Name of Environment.')
flags.DEFINE_integer('env_seed', 0, 'Seed for the environment.')
flags.DEFINE_integer('global_seed', 1, 'Seed for all the rngs.')
flags.DEFINE_integer('save_every', 10, 'Save results after these many epochs.')
flags.DEFINE_integer('batch_size', 16, 'Number of environments to run.')
flags.DEFINE_integer('n_trajectories', 128, 'Number of trajectories to use.')
flags.DEFINE_integer(
'max_steps_env', 1500, 'Maximum number of steps to run in the environment '
'before termination.')
flags.DEFINE_float('std', None, 'Standard deviations for the policy.')
flags.DEFINE_multi_float('stds', None,
'A list of standard deviations for the policy.')
flags.DEFINE_string('policy_type', 'normal',
'Type of policy. Either `discrete` or `normal`.')
flags.DEFINE_string("baseline_dir", None,
"Path to directory containing the results of the baseline experiments.")
flags.DEFINE_string("filter", "",
"Run only on experiment names that contain this string")
flags.DEFINE_integer("n_bins", 100,
"Number of bins for plot binning")
flags.DEFINE_integer("n_bins_eff_suc", None,
"Number of bins for plot binning")
flags.DEFINE_integer("dpi", 100,
"DPI for saving")
flags.DEFINE_multi_float("x_lim", [-0.01, 0.21],
"X lim for eff_suc plot")
flags.DEFINE_multi_float("y_lim", [0.75, 1.01],
"X lim for eff_suc plot")
flags.DEFINE_float("x_lim_eff", None,
"X lim for eps_eff plot")
flags.DEFINE_float("y_lim_eff", None,
"X lim for eps_eff plot")
flags.DEFINE_float("zoom_factor", None,
"Zoom factor for zoom box")
flags.DEFINE_boolean("y_median", False,
"Aggregate y by median instead of mean")
def avsr_flags():
# Generic flags
flags.DEFINE_integer('buffer_size', 15000, 'Shuffle buffer size')
flags.DEFINE_integer('batch_size', 64, 'Batch Size')
flags.DEFINE_integer('embedding_size', 128, 'Embedding dimension')
flags.DEFINE_integer('beam_width', 10, 'Beam Width')
flags.DEFINE_boolean('enable_function', False, 'Enable Function?')
flags.DEFINE_string('architecture', 'transformer', 'Network Architecture')
flags.DEFINE_string('gpu_id', '0', 'GPU index')
flags.DEFINE_string('input_modality', 'audio',
'Switch between A and V inputs')
flags.DEFINE_integer('noise_level', 0, 'Noise level in range {0, 1, 2, 3}')
flags.DEFINE_boolean('mix_noise', False, 'TBA')
flags.DEFINE_multi_integer('cnn_filters', (16, 32, 48, 64),
'Number of CNN filters per layer')
flags.DEFINE_integer('cnn_dense_units', 256,
'Number of neurons in the CNN output layer')
flags.DEFINE_string('cnn_activation', 'relu',
'Activation function in CNN layers')
flags.DEFINE_string('cnn_final_activation', None,
'Activation function in the final CNN layer')
flags.DEFINE_string('cnn_normalisation', 'layer_norm',
'Normalisation function in CNN blocks')
flags.DEFINE_boolean('cnn_final_clip', True,
'Clip the activation of the final CNN layer')
# flags.DEFINE_float('learning_rate', 0.001, 'Learning rate')
flags.DEFINE_integer(
'lr_warmup_steps', 750,
'Number of steps for the Learning rate linear warmup')
flags.DEFINE_float('max_gradient_norm', 1.0,
'Clip the global gradient norm')
flags.DEFINE_string('optimiser', 'adam', 'Optimiser type')
flags.DEFINE_boolean('amsgrad', False, 'Use AMSGrad ?')
flags.DEFINE_string('logfile', 'default_logfile', 'Logfile name')
flags.DEFINE_boolean('profiling', False,
'Enable profiling for TensorBoard')
flags.DEFINE_boolean('write_halting_history', False,
'Dump segmentations to praat files')
flags.DEFINE_boolean('plot_alignments', False,
'Write alignment summaries in TensorBoard')
flags.DEFINE_boolean('use_tensorboard', False,
'Export TensorBoard summary')
# RNN seq2seq FLAGS
flags.DEFINE_string('encoder', 'RNN', 'Encoder type')
flags.DEFINE_string('recurrent_activation', 'sigmoid',
'Activation function inside LSTM Cell')
flags.DEFINE_string('encoder_input_normalisation', 'layer_norm',
'Normalisation function for the Encoder input')
flags.DEFINE_string('cell_type', 'ln_lstm', 'Recurrent cell type')
flags.DEFINE_multi_integer(
'encoder_units_per_layer', 3 * (256, ),
'Number of encoder cells in each recurrent layer')
flags.DEFINE_multi_integer(
'decoder_units_per_layer', 1 * (256, ),
'Number of decoder cells in each recurrent layer')
flags.DEFINE_multi_float('dropout_probability', 3 * (0.1, ),
'Dropout rate for for RNN cells')
flags.DEFINE_multi_float(
'rnn_l1_l2_weight_regularisation', (0.0, 0.0001),
'Weight regularisation (L1 and L2) for RNN cells')
flags.DEFINE_boolean('recurrent_regularisation', False,
'Use regularisation in the recurrent LSTM kernel')
flags.DEFINE_boolean('regularise_all', True,
'Regularise all model variables ?')
flags.DEFINE_string('recurrent_initialiser', None,
'Recurrent kernel initialiser')
flags.DEFINE_boolean('recurrent_dropout', True,
'Apply dropout on recurrent state')
flags.DEFINE_boolean('enable_attention', True, 'Enable Attention ?')
flags.DEFINE_float('output_sampling', 0.1, 'Output Sampling Rate')
flags.DEFINE_float('lstm_state_dropout', 0.1,
'Dropout applied to the h state of the LSTM')
flags.DEFINE_string('decoder_initialisation', 'final_encoder_state',
'Decoder initialisation scheme')
flags.DEFINE_string('segmental_variant', 'v1', 'Segmental model variant')
# Transformer Model
flags.DEFINE_integer('transformer_hidden_size', 256,
'State size of the Transformer layers')
flags.DEFINE_integer('transformer_num_encoder_layers', 6,
'Number of layers in the encoder stack')
flags.DEFINE_integer('transformer_num_decoder_layers', 6,
'Number of layers in the decoder stack')
flags.DEFINE_integer('transformer_num_heads', 1,
'Number of attention_heads')
flags.DEFINE_integer('transformer_filter_size', 256, 'Filter size')
flags.DEFINE_float('transformer_relu_dropout', 0.1, 'Filter size')
flags.DEFINE_float('transformer_attention_dropout', 0.1, 'Filter size')
flags.DEFINE_float('transformer_layer_postprocess_dropout', 0.1,
'Post-processing layer dropout')
flags.DEFINE_string('transformer_dtype', 'float32', 'Data type')
flags.DEFINE_integer('transformer_extra_decode_length', 0,
'Extra Decode Length')
flags.DEFINE_integer('transformer_beam_size', 10, 'Beam search width')
flags.DEFINE_float('transformer_alpha', 0.6,
'Used for length normalisation in beam search')
flags.DEFINE_boolean('transformer_online_encoder', False,
'Whether or not to use a causal attention bias')
flags.DEFINE_integer('transformer_encoder_lookahead', 11,
'Number of frames for encoder attention lookahead')
flags.DEFINE_integer('transformer_encoder_lookback', 11,
'Number of frames for encoder attention lookback')
flags.DEFINE_boolean('transformer_online_decoder', False,
'Wheter or not to use online decoding')
flags.DEFINE_integer(
'transformer_decoder_lookahead', 5,
'Number of segments for cross-modal attention lookahead')
flags.DEFINE_integer(
'transformer_decoder_lookback', 5,
'Number of segments for cross-modal attention lookback')
flags.DEFINE_float('transformer_l1_regularisation', 0.0,
'Transformer L1 weight regularisation')
flags.DEFINE_float('transformer_l2_regularisation', 0.0,
'Transformer L2 weight regularisation')
## Experimental flags
flags.DEFINE_integer('transformer_num_avalign_layers', 1,
'Number of layers in the AVAlign stack')
flags.DEFINE_boolean('au_loss', False,
'Use the Action Unit Loss for multimodal encoders')
flags.DEFINE_float('au_loss_weight', 10.0,
'Scalar multiplier of the AU loss')
flags.DEFINE_boolean(
'constrain_av_attention', False,
'Limit the search space of the A-V alignment'
' to a window centred on the main diagonal')
flags.DEFINE_integer('constrain_frames', 2, 'Num frames')
flags.DEFINE_boolean('word_loss', False, 'Word counting loss')
flags.DEFINE_float('wordloss_weight', 0.01, 'Num words loss')
flags.DEFINE_string('wb_activation', 'sigmoid',
'Activation for the encoder halting unit')
FLAGS = flags.FLAGS
flags.DEFINE_string('results_patern',
default=None,
help=('File patern for experiments results files'))
flags.DEFINE_string('charts_path',
default=None,
help=('Path where the charts will be saved on'))
flags.DEFINE_list('chart_format_list',
default=['pdf'],
help=('List of file formats to save charts on'))
flags.DEFINE_multi_float('log_prec_at_pr',
default=None,
help=('Log Precision at Recall levels'))
flags.DEFINE_string(
'results_metrics_csv_name',
default=None,
help=('Name to the csv file where metrics will be saved on'))
flags.DEFINE_bool('show_random_guess',
default=False,
help=('Show line for random guess'))
flags.mark_flag_as_required('results_patern')
flags.mark_flag_as_required('charts_path')
ClassifierResults = collections.namedtuple("ClassifierResults", [