def parse_args():
# Parse commandline
parser = ArgParseDefault()
parser.add_argument('--tpu_ip', required=False, help='IP-address of the TPU')
parser.add_argument('--bucket_name', required=True, help='Bucket name')
parser.add_argument('--tpu_name', required=False, help='Name of the TPU')
parser.add_argument('--tpu_name_project', required=False, help='Name of the TPU project')
parser.add_argument('--pretrain_data', required=True, type=str, help='Folder which contains pretrain data. Should be located under gs://{bucket_name}/{project_name}/pretrain/pretrain_data/')
parser.add_argument('--run_prefix', help='Prefix to be added to all runs. Useful to group runs')
parser.add_argument('--model_class', default='bert_large_uncased_wwm', choices=PRETRAINED_MODELS.keys(), help='Model class to use')
parser.add_argument('--project_name', default='covid-bert', help='Name of subfolder in Google bucket')
parser.add_argument('--num_gpus', default=1, type=int, help='Number of GPUs to use')
parser.add_argument('--eval_steps', default=1000, type=int, help='Number eval steps to run (only active when --do_eval flag is provided)')
parser.add_argument('--init_checkpoint', default=None, help='Run name to initialize checkpoint from. Example: "run2/ctl_step_8000.ckpt-8". or "run2/pretrained/bert_model_8000.ckpt-8". The first contains the mlm/nsp layers. \
By default using a pretrained model from gs://{bucket_name}/pretrained_models/')
parser.add_argument('--load_mlm_nsp_weights', default=None, help="If set to True it will load the mlm/nsp-layers. The init_checkpoint should then be set to a model containing these. Usually in base run-directory named 'ctl_step*'.")
parser.add_argument('--set_trainstep', default=None, help="If set this will set the trainstep. This is only needed when restarting from an old checkpoint and you would like to get the scheduler/optimiser to start at the correct point.")
parser.add_argument('--optimizer_type', default='adamw', choices=['adamw', 'lamb'], type=str, help='Optimizer')
parser.add_argument('--train_batch_size', default=32, type=int, help='Training batch size')
parser.add_argument('--eval_batch_size', default=32, type=int, help='Eval batch size')
parser.add_argument('--num_epochs', default=3, type=int, help='Number of epochs')
parser.add_argument('--num_steps_per_epoch', default=1000, type=int, help='Number of steps per epoch')
parser.add_argument('--warmup_steps', default=10000, type=int, help='Warmup steps')
parser.add_argument('--warmup_proportion', default=None, type=float, help='If set overwrites warmup_steps.')
parser.add_argument('--learning_rate', default=2e-5, type=float, help='Learning rate')
parser.add_argument('--end_lr', default=0, type=float, help='Final learning rate')
parser.add_argument('--max_seq_length', default=96, type=int, help='Maximum sequence length. Sequences longer than this will be truncated, and sequences shorter than this will be padded.')
parser.add_argument('--max_predictions_per_seq', default=14, type=int, help='Maximum predictions per sequence_output.')
parser.add_argument('--dtype', default='fp32', choices=['fp32', 'bf16', 'fp16'], type=str, help='Data type')
parser.add_argument('--steps_per_loop', default=10, type=int, help='Steps per loop')
parser.add_argument('--time_history_log_steps', default=1000, type=int, help='Frequency with which to log timing information with TimeHistory.')
add_bool_arg(parser, 'use_tpu', default=True, help='Use TPU')
add_bool_arg(parser, 'do_eval', default=False, help='Run evaluation (make sure eval data is present in tfrecords folder)')
args = parser.parse_args()
return args
def parse_args():
# Parse commandline
parser = argparse.ArgumentParser()
parser.add_argument(
'--path_to_huggingface',
type=str,
required=True,
help=
'Path to folder containing pytorch model/config.json for huggingface')
parser.add_argument(
'--path_to_tf',
type=str,
required=True,
help='Path to folder containing tf checkpoint and bert_config.json')
parser.add_argument('--input_text',
default='this is some example test',
required=False,
help='Test example')
parser.add_argument('--model_class',
default='bert_large_uncased_wwm',
choices=PRETRAINED_MODELS.keys(),
help='Model class to use')
parser.add_argument('--model_class_huggingface',
default='bert-large-uncased',
help='Model class to use for huggingface tokenizer')
args = parser.parse_args()
return args
def parse_args():
# Parse commandline
parser = ArgParseDefault()
parser.add_argument('--bucket_name', required=True, help='Bucket name')
parser.add_argument('--init_checkpoint', help='Path to checkpoint')
parser.add_argument('--model_class', default='bert_large_uncased_wwm', choices=PRETRAINED_MODELS.keys(), help='Model class to use')
parser.add_argument('--project_name', default='covid-bert', help='Name of subfolder in Google bucket')
parser.add_argument('--output', default=['tf_hub', 'huggingface'], choices=['tf_hub', 'huggingface'], nargs='+', help='Generate output for those model types')
args = parser.parse_args()
return args
def parse_args():
# Parse commandline
parser = ArgParseDefault(usage=USAGE_DESCRIPTION)
parser.add_argument('--run_name', required=True, help='Finetune run name. The model will be loaded from gs://{bucket_name}/{project_name}/finetune/runs/{run_name}.')
parser.add_argument('--bucket_name', required=True, help='Bucket name')
parser.add_argument('--project_name', required=False, default='covid-bert', help='Name of subfolder in Google bucket')
parser.add_argument('--input_text', required=False, help='Predict arbitrary input text and print prediction to stdout')
parser.add_argument('--input_txt_files', nargs='+', required=False, help='Predict text from local txt files. One example per line.')
parser.add_argument('--input_tfrecord_files', nargs='+', required=False, help='Predict text from tfrecord files (local or on bucket).')
parser.add_argument('--tpu_ip', required=False, help='IP-address of the TPU')
parser.add_argument('--model_class', default='bert_large_uncased_wwm', choices=PRETRAINED_MODELS.keys(), help='Model class to use')
parser.add_argument('--num_gpus', default=1, type=int, help='Number of GPUs to use')
parser.add_argument('--eval_batch_size', default=32, type=int, help='Eval batch size')
parser.add_argument('--label_name', default='label', type=str, help='Name of label to predicted')
add_bool_arg(parser, 'interactive_mode', default=False, help='Interactive mode')
add_bool_arg(parser, 'use_tpu', default=False, help='Use TPU (only works when using input_tfrecord_files stored on a Google bucket)')
args = parser.parse_args()
return args
def parse_args():
parser = ArgParseDefault()
parser.add_argument('--input_data', required=True, help='Path to folder with txt files. \
Folder may contain train/dev/test subfolders. Each txt file contains the text of a single tweet per line.')
parser.add_argument('--run_prefix', help='Prefix to be added to all runs. Useful to identify runs')
parser.add_argument('--model_class', default='bert_large_uncased_wwm', choices=PRETRAINED_MODELS.keys(), help='Model class')
parser.add_argument('--username_filler', default='twitteruser', type=str, help='Username filler')
parser.add_argument('--url_filler', default='twitterurl', type=str, help='URL filler (ignored when replace_urls option is false)')
parser.add_argument('--num_logged_samples', default=10, type=int, help='Log first n samples to output')
add_bool_arg(parser, 'run_in_parallel', default=True, help='Run script in parallel')
add_bool_arg(parser, 'replace_usernames', default=True, help='Replace usernames with filler')
add_bool_arg(parser, 'replace_urls', default=True, help='Replace URLs with filler')
add_bool_arg(parser, 'asciify_emojis', default=True, help='Asciifyi emojis')
add_bool_arg(parser, 'replace_multiple_usernames', default=True, help='Replace "@user @user" with "2 <username_filler>"')
add_bool_arg(parser, 'replace_multiple_urls', default=True, help='Replace "http://... http://.." with "2 <url_filler>"')
add_bool_arg(parser, 'remove_unicode_symbols', default=True, help='After preprocessing remove characters which belong to unicode category "So"')
add_bool_arg(parser, 'remove_accented_characters', default=False, help='Remove accents/asciify everything. Probably not recommended.')
args = parser.parse_args()
return args
def parse_args():
# Parse commandline
parser = ArgParseDefault()
parser.add_argument('--finetune_data', required=True, help='Finetune data folder sub path. Path has to be in gs://{bucket_name}/{project_name}/finetune/finetune_data/{finetune_data}.\
This folder includes a meta.json (containing meta info about the dataset), and a file label_mapping.json. \
TFrecord files (train.tfrecords and dev.tfrecords) should be located in a \
subfolder gs://{bucket_name}/{project_name}/finetune/finetune_data/{finetune_data}/tfrecords/')
parser.add_argument('--bucket_name', required=True, help='Bucket name')
parser.add_argument('--tpu_ip', required=False, help='IP-address of the TPU')
parser.add_argument('--preemptible_tpu', default=False, action='store_true', required=False, help='Dynamically create preemptible TPU (this requires you to have glcoud installed with suitable permissions)')
parser.add_argument('--preemptible_tpu_zone', default='us-central1-f', type=str, required=False, help='Preemptible TPU zone (only if --preemptible_tpu flag is provided)')
parser.add_argument('--preemptible_tpu_name', default=None, type=str, required=False, help='Preemptible TPU name (only if --preemptible_tpu flag is provided)')
parser.add_argument('--preemptible_tpu_version', default='nightly', choices=['nightly', '2.1'], type=str, required=False, help='Preemptible TPU version (only if --preemptible_tpu flag is provided)')
parser.add_argument('--run_prefix', help='Prefix to be added to all runs. Useful to group runs')
parser.add_argument('--project_name', default='covid-bert', help='Name of subfolder in Google bucket')
parser.add_argument('--model_class', default='bert_large_uncased_wwm', choices=PRETRAINED_MODELS.keys(), help='Model class to use')
parser.add_argument('--num_gpus', default=1, type=int, help='Number of GPUs to use')
parser.add_argument('--init_checkpoint', default=None, help='Run name to initialize checkpoint from. Example: "run2/ctl_step_8000.ckpt-8". \
By default using a pretrained model from gs://{bucket_name}/pretrained_models/')
parser.add_argument('--init_checkpoint_index', type=int, help='Checkpoint index. This argument is ignored and only added for reporting.')
parser.add_argument('--repeats', default=1, type=int, help='Number of times the script should run. Default is 1')
parser.add_argument('--num_epochs', default=3, type=int, help='Number of epochs')
parser.add_argument('--limit_train_steps', type=int, help='Limit the number of train steps per epoch. Useful for testing.')
parser.add_argument('--limit_eval_steps', type=int, help='Limit the number of eval steps per epoch. Useful for testing.')
parser.add_argument('--train_batch_size', default=32, type=int, help='Training batch size')
parser.add_argument('--eval_batch_size', default=32, type=int, help='Eval batch size')
parser.add_argument('--learning_rate', default=2e-5, type=float, help='Learning rate')
parser.add_argument('--end_lr', default=0, type=float, help='Final learning rate')
parser.add_argument('--warmup_proportion', default=0.1, type=float, help='Learning rate warmup proportion')
parser.add_argument('--max_seq_length', default=96, type=int, help='Maximum sequence length')
parser.add_argument('--early_stopping_epochs', default=-1, type=int, help='Stop when loss hasn\'t decreased during n epochs')
parser.add_argument('--optimizer_type', default='adamw', choices=['adamw', 'lamb'], type=str, help='Optimizer')
parser.add_argument('--dtype', default='fp32', choices=['fp32', 'bf16', 'fp16'], type=str, help='Data type')
parser.add_argument('--steps_per_loop', default=10, type=int, help='Steps per loop (unavailable for Keras fit in TF 2.2, will be added in later version)')
parser.add_argument('--time_history_log_steps', default=10, type=int, help='Frequency with which to log timing information with TimeHistory.')
add_bool_arg(parser, 'use_tpu', default=True, help='Use TPU')
args = parser.parse_args()
return args
def parse_args():
parser = ArgParseDefault()
parser.add_argument('--input_txt_files',
type=str,
nargs='+',
help='Input txt files to process.')
parser.add_argument('--model_class',
default='bert_large_uncased_wwm',
choices=PRETRAINED_MODELS.keys(),
help='Model class')
parser.add_argument('--run_prefix', help='Run prefix')
parser.add_argument('--max_seq_length',
default=96,
type=int,
help='Maximum sequence length')
parser.add_argument('--username_filler',
default='twitteruser',
type=str,
help='Username filler')
parser.add_argument(
'--url_filler',
default='twitterurl',
type=str,
help='URL filler (ignored when replace_urls option is false)')
add_bool_arg(parser,
'replace_usernames',
default=True,
help='Replace usernames with filler')
add_bool_arg(parser,
'replace_urls',
default=True,
help='Replace URLs with filler')
add_bool_arg(parser,
'asciify_emojis',
default=True,
help='Asciifyi emojis')
add_bool_arg(parser,
'replace_multiple_usernames',
default=True,
help='Replace "@user @user" with "2 <username_filler>"')
add_bool_arg(parser,
'replace_multiple_urls',
default=True,
help='Replace "http://... http://.." with "2 <url_filler>"')
add_bool_arg(
parser,
'remove_unicode_symbols',
default=True,
help=
'After preprocessing remove characters which belong to unicode category "So"'
)
add_bool_arg(parser,
'standardize_punctuation',
default=True,
help='Standardize (asciifyi) special punctuation')
add_bool_arg(
parser,
'remove_accented_characters',
default=False,
help='Remove accents/asciify everything. Probably not recommended.')
add_bool_arg(parser,
'write_preprocessed_file',
default=True,
help='Write preprocess output file')
add_bool_arg(parser,
'run_in_parallel',
default=True,
help='Run script in parallel')
return parser.parse_args()
def parse_args():
parser = ArgParseDefault()
parser.add_argument(
'--run_name',
required=True,
help=
'Run name to create tf record files for. Run folder has to be located under \
data/pretrain/{run_name}/preprocessed/ and must contain one or multiple txt files. May also contain train and dev subfolders with txt files.'
)
parser.add_argument('--max_seq_length',
default=96,
type=int,
help='Maximum sequence length')
parser.add_argument('--model_class',
default='bert_large_uncased_wwm',
choices=PRETRAINED_MODELS.keys(),
help='Model class to use')
parser.add_argument(
'--dupe_factor',
default=10,
type=int,
help=
'Number of times to duplicate the input data (with different masks).')
parser.add_argument(
'--short_seq_prob',
default=0.1,
type=float,
help=
'Probability of creating sequences which are shorter than the maximum length.'
)
parser.add_argument(
'--max_predictions_per_seq',
default=14,
type=int,
help='Maximum number of masked LM predictions per sequence.')
parser.add_argument('--random_seed',
default=42,
type=int,
help='Random seed')
parser.add_argument('--masked_lm_prob',
default=0.15,
type=float,
help='Masked LM probabibility')
parser.add_argument('--gzipped',
action='store_true',
default=False,
help='Create gzipped tfrecords files')
parser.add_argument('--num_logged_samples',
default=10,
type=int,
help='Log first n samples to output')
parser.add_argument(
'--max_num_cpus',
default=10,
type=int,
help=
'Adapt this number based on the available memory/size of input files. \
This code was tested on a machine with a lot of memory (250GB). Decrease this number if you run into memory issues.'
)
add_bool_arg(parser,
'run_in_parallel',
default=True,
help='Run script in parallel')
return parser.parse_args()
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('-i', '--input', required=True, type=str, help="Input to tokenize")
parser.add_argument('--model_class', default='covid-twitter-bert', choices=list(PRETRAINED_MODELS.keys()), help="Model class")
args = parser.parse_args()
return args
type='laptop',
folds=FOLDS,
return_val_idxs=True)
VAL_IDX = []
LB, GT = [], []
for idxs in val_idxs:
VAL_IDX.extend(idxs)
for train, val in cv_loader:
for ((rv_raw, lb_raw), x, y) in val:
LB.extend(lb_raw)
GT.extend(rv_raw)
tokenizers = dict([
(model_name,
BertTokenizer.from_pretrained(model_config['path'],
do_lower_case=True))
for model_name, model_config in PRETRAINED_MODELS.items()
])
# print(tokenizers)
cv_loaders = dict([
(model_name,
get_data_loaders_cv(rv_path='../data/TRAIN/Train_laptop_reviews.csv',
lb_path='../data/TRAIN/Train_laptop_labels.csv',
tokenizer=tokenizers[model_name],
batch_size=args.bs,
type='laptop',
folds=FOLDS))
for model_name, model_config in PRETRAINED_MODELS.items()
])
PRED = []
def parse_args():
# Parse commandline
parser = ArgParseDefault()
parser.add_argument('--tpu_ip',
required=True,
help='IP-address of the TPU')
parser.add_argument('--bucket_name', required=True, help='Bucket name')
parser.add_argument('--tpu_name', required=False, help='Name of the TPU')
parser.add_argument('--tpu_name_project',
required=False,
help='Name of the TPU project')
parser.add_argument(
'--pretrain_data',
required=True,
type=str,
help=
'Folder which contains pretrain data. Should be located under gs://{bucket_name}/{project_name}/pretrain/pretrain_data/'
)
parser.add_argument(
'--run_prefix',
help='Prefix to be added to all runs. Useful to group runs')
parser.add_argument('--model_class',
default='bert_large_uncased_wwm',
choices=PRETRAINED_MODELS.keys(),
help='Model class to use')
parser.add_argument('--project_name',
default='covid-bert',
help='Name of subfolder in Google bucket')
parser.add_argument('--num_gpus',
default=1,
type=int,
help='Number of GPUs to use')
parser.add_argument(
'--eval_steps',
default=1000,
type=int,
help=
'Number eval steps to run (only active when --do_eval flag is provided)'
)
parser.add_argument('--optimizer_type',
default='adamw',
choices=['adamw', 'lamb'],
type=str,
help='Optimizer')
parser.add_argument('--train_batch_size',
default=32,
type=int,
help='Training batch size')
parser.add_argument('--eval_batch_size',
default=32,
type=int,
help='Eval batch size')
parser.add_argument('--num_epochs',
default=3,
type=int,
help='Number of epochs')
parser.add_argument('--num_steps_per_epoch',
default=1000,
type=int,
help='Number of steps per epoch')
parser.add_argument('--warmup_steps',
default=10000,
type=int,
help='Warmup steps')
parser.add_argument('--learning_rate',
default=2e-5,
type=float,
help='Learning rate')
parser.add_argument('--end_lr',
default=0,
type=float,
help='Final learning rate')
parser.add_argument(
'--max_seq_length',
default=96,
type=int,
help=
'Maximum sequence length. Sequences longer than this will be truncated, and sequences shorter than this will be padded.'
)
parser.add_argument('--max_predictions_per_seq',
default=14,
type=int,
help='Maximum predictions per sequence_output.')
parser.add_argument('--dtype',
default='fp32',
choices=['fp32', 'bf16', 'fp16'],
type=str,
help='Data type')
parser.add_argument('--steps_per_loop',
default=10,
type=int,
help='Steps per loop')
parser.add_argument(
'--time_history_log_steps',
default=1000,
type=int,
help='Frequency with which to log timing information with TimeHistory.'
)
add_bool_arg(parser, 'use_tpu', default=True, help='Use TPU')
add_bool_arg(
parser,
'do_eval',
default=False,
help=
'Run evaluation (make sure eval data is present in tfrecords folder)')
args = parser.parse_args()
return args
def parse_args():
parser = ArgParseDefault()
parser.add_argument(
'--finetune_datasets',
type=str,
nargs='+',
help=
'Finetune dataset(s) to process. These correspond to folder names in data/finetune. \
Data should be located in data/finetune/originals/{finetune_dataset}/[train.tsv/dev.tsv/test.tsv]. By default runs all datasets.'
)
parser.add_argument('--model_class',
default='bert_large_uncased_wwm',
choices=PRETRAINED_MODELS.keys(),
help='Model class')
parser.add_argument(
'--run_prefix',
help='Prefix to be added to all runs. Useful to identify runs')
parser.add_argument('--max_seq_length',
default=96,
type=int,
help='Maximum sequence length')
parser.add_argument('--username_filler',
default='twitteruser',
type=str,
help='Username filler')
parser.add_argument(
'--url_filler',
default='twitterurl',
type=str,
help='URL filler (ignored when replace_urls option is false)')
add_bool_arg(parser,
'replace_usernames',
default=True,
help='Replace usernames with filler')
add_bool_arg(parser,
'replace_urls',
default=True,
help='Replace URLs with filler')
add_bool_arg(parser,
'asciify_emojis',
default=True,
help='Asciifyi emojis')
add_bool_arg(parser,
'replace_multiple_usernames',
default=True,
help='Replace "@user @user" with "2 <username_filler>"')
add_bool_arg(parser,
'replace_multiple_urls',
default=True,
help='Replace "http://... http://.." with "2 <url_filler>"')
add_bool_arg(parser,
'standardize_punctuation',
default=True,
help='Standardize (asciifyi) special punctuation')
add_bool_arg(
parser,
'remove_unicode_symbols',
default=True,
help=
'After preprocessing remove characters which belong to unicode category "So"'
)
add_bool_arg(
parser,
'remove_accented_characters',
default=False,
help='Remove accents/asciify everything. Probably not recommended.')
return parser.parse_args()
