# Token Classification (Named Entity Recognition)

## Introduction

TokenClassification Model supports Named entity recognition (NER) and other token level classification tasks, as long as the data follows the format specified below. This model card will focus on the NER task.

Named entity recognition (NER), also referred to as entity chunking, identification or extraction, is the task of detecting and classifying key information (entities) in text. In other words, an NER model takes a piece of text as input and for each word in the text, the model identifies a category the word belongs to. For example, in a sentence: Mary lives in Santa Clara and works at NVIDIA, the model should detect that Mary is a person, Santa Clara is a location and NVIDIA is a company.

TAO Toolkit provides a sample notebook to outline the end-to-end workflow on how to train a TokenClassification model using TAO and deploy it in Riva format on NGC resources.

Copy
Copied!

tao token_classification download_specs -r /results/token_classification/default_specs/ \
-o /specs/nlp/token_classification


• -o: Path to where the spec files will be stored

• -r: Output directory to store logs

After running the above, the spec files would be stored under /specs/nlp/token_classification and you can modify them locally if this directory is mounted to your local folder in ~/.tao_mounts.json file.

## Data Input for Token Classification Model

For pre-training or fine-tuning of the model, the data should be split into two files:

• text.txt

• labels.txt

Each line of the text.txt file contains text sequences, where words are separated with spaces, i.e.: [WORD] [SPACE] [WORD] [SPACE] [WORD]. The labels.txt file contains corresponding labels for each word in text.txt; the labels are separated with spaces, i.e.: [LABEL] [SPACE] [LABEL] [SPACE] [LABEL]. Example of a text.txt file:

Jennifer is from New York City . She likes … …

Corresponding labels.txt file:

B-PER O O B-LOC I-LOC I-LOC O O O … …

## Dataset Conversion

To convert an IOB format (short for inside, outside, beginning) data to the format required for training:

Copy
Copied!

# For conversion from IOB format, for example, for CoNLL-2003 dataset:
tao token_classification dataset_convert [-h] \
-e /specs/nlp/token_classification/dataset_convert.yaml \
source_data_dir=/path/to/source_data_dir \
target_data_dir=/path/to/target_data_dir


The source_data_dir structure should look like this (test.txt is optional):

Copy
Copied!

.
|--sourced_data_dir
|-- dev.txt
|-- test.txt
|-- train.txt


Note, the development set (or dev set) will be used to evaluate the performance of the model during model training. The hyper-parameters search and model selection should be based on the dev set, while the final evaluation of the selected model should be performed on the test set.

An example of a spec file for dataset conversion:

Copy
Copied!

# Path to the folder containing the dataset source files
source_data_dir: ???

# Path to the output folder.
target_data_dir: ???

# list of file names inside source_data_dir in IOB format
list_of_file_names:  ['train.txt','dev.txt']

# name of the file with training data inside sourse_data_dir
# train_file is used to generate label to label_id mapping
train_file_name: 'train.txt'

# Max sequence length use -1 to leave the examples's length as is,
# otherwise long examples will be split into multiple examples'
max_length: -1


Output log after running token_classification dataset_convert:

Copy
Copied!

[NeMo I 2021-01-21 09:07:11 dataset_convert:133] Spec file:
source_data_dir: original/
list_of_file_names:
- train.txt
- dev.txt
train_file_name: train.txt
target_data_dir: original/output/
max_length: -1

[NeMo I token_classification_utils:54] Processing original/output/labels_train.txt
[NeMo I token_classification_utils:92] Labels mapping {'O': 0, 'B-LOC': 1, 'B-MISC': 2, 'B-ORG': 3, 'B-PER': 4, 'I-LOC': 5, 'I-MISC': 6, 'I-ORG': 7, 'I-PER': 8} saved to : original/output/label_ids.csv
[NeMo I token_classification_utils:101] Three most popular labels in original/output/labels_train.txt:
[NeMo I data_preprocessing:131] label: 0, 169578 out of 203621 (83.28%).
[NeMo I data_preprocessing:131] label: 1, 7140 out of 203621 (3.51%).
[NeMo I data_preprocessing:131] label: 4, 6600 out of 203621 (3.24%).
[NeMo I token_classification_utils:103] Total labels: 203621. Label frequencies - {0: 169578, 1: 7140, 4: 6600, 3: 6321, 8: 4528, 7: 3704, 2: 3438, 5: 1157, 6: 1155}
[NeMo I dataset_convert:173] Text and labels for train.txt saved to original/output/.
[NeMo I dataset_convert:174] Processing of train.txt is complete.
[NeMo I token_classification_utils:54] Processing original/output/labels_dev.txt
[NeMo I token_classification_utils:75] Using provided labels mapping {'O': 0, 'B-LOC': 1, 'B-MISC': 2, 'B-ORG': 3, 'B-PER': 4, 'I-LOC': 5, 'I-MISC': 6, 'I-ORG': 7, 'I-PER': 8}
[NeMo I token_classification_utils:98] original/output/labels_dev_label_stats.tsv found, skipping stats calculation.
[NeMo I dataset_convert:173] Text and labels for dev.txt saved to original/output/.
[NeMo I dataset_convert:174] Processing of dev.txt is complete.


### Convert Dataset Required Arguments

• -e: The experiment specification file.

• source_data_dir - path to the raw data

• target_data_dir - path to store the processed files

### Convert Dataset Optional Arguments

• -h, --help: Show this help message and exit

• list_of_file_names: List of files in source_data_dir for conversion

Parameter

Datatype

Default

Description

Supported Values

source_data_dir

string

– | Path to the dataset source data directory

target_data_dir

string

– | Path to the dataset target data directory

list_of_file_names

List of strings

[‘train.txt’,’dev.txt’]

List of files for conversion

train_file_name

string

‘train.txt’

Name of the file with training data inside sourse_data_dir train_file is used to generate string label to integer label_id mapping

After the conversion, the target_data_dir should contain the following files:

Copy
Copied!

.
|--target_data_dir
|-- label_ids.csv
|-- labels_dev.txt
|-- labels_test.txt
|-- labels_train.txt
|-- text_dev.txt
|-- text_test.txt
|-- text_train.txt


Note

The target_data_dir contains file label_ids.csv. This file lists all the labels present in the train data. Each label is written on a separate line. Additionally, a special padding token - O - used to mark input that should be tagged as “no label” for the task. For example, for NER task, words that do not belong to any entities would have O label. During training, this label_ids.csv file is going to be used to create mapping from a text label to an integer.

Example of the label_ids.csv file:

Copy
Copied!

O
B-GPE
B-LOC
B-MISC
B-ORG
B-PER
B-TIME
I-GPE
I-LOC
I-MISC
I-ORG
I-PER
I-TIME


During training, the text labels will be mapped to integers. Each label will have an id that corresponds to the line number of the text label. The above file, would be converted like so:

Copy
Copied!

{'O': 0,
'B-GPE': 1,
'B-LOC': 2,
'B-MISC': 3,
'B-ORG': 4,
'B-PER': 5,
'B-TIME': 6,
'I-GPE': 7,
'I-LOC': 8,
'I-MISC': 9,
'I-ORG': 10,
'I-PER': 11,
'I-TIME': 12}


## Training a Token Classification Model

In the Token Classification Model, we are jointly training a classifier on top of a pre-trained language model, such as BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.

Unless the user provides a pre-trained checkpoint for the language model, the language model is initialized with the pre-trained model from HuggingFace Transformers.

Example spec for training:

Copy
Copied!

trainer:
max_epochs: 5

# Specifies parameters for the Token Classification model
model:
tokenizer:
tokenizer_name: ${model.language_model.pretrained_model_name} # or sentencepiece vocab_file: null # path to vocab file tokenizer_model: null # only used if tokenizer is sentencepiece special_tokens: null # Pre-trained language model such as BERT language_model: pretrained_model_name: bert-base-uncased lm_checkpoint: null config_file: null # json file, precedence over config config: null # Specifies parameters of the token classification head that follows a BERT-based language-model head: num_fc_layers: 2 fc_dropout: 0.5 activation: 'relu' use_transformer_init: True # Path to file with label_ids, generated with dataset_convert.py. # Those labels are used by the model as labels (names of target classes, their number). label_ids: ??? # Path to directory containing both finetuning and validation data. data_dir: ??? # Specifies the parameters of the dataset to be used for training. training_ds: text_file: text_train.txt labels_file: labels_train.txt batch_size: 64 num_samples: -1 # number of samples to be considered, -1 means all the dataset # Specifies the parameters of the dataset to be used for validation. validation_ds: text_file: text_dev.txt labels_file: labels_dev.txt batch_size: 64 num_samples: -1 # number of samples to be considered, -1 means all the dataset # The parameters for the training optimizer, including learning rate, lr schedule, etc. optim: name: adam lr: 5e-5 weight_decay: 0.00 # scheduler setup sched: name: WarmupAnnealing # Scheduler params warmup_steps: null warmup_ratio: 0.1 last_epoch: -1 # pytorch lightning args monitor: val_loss reduce_on_plateau: false  The specification can be roughly grouped into three categories: • Parameters that describe the training process • Parameters that describe the datasets, and • Parameters that describe the model. More details about parameters in the spec file are provided below:  Parameter Data Type Default Description data_dir string – Path to the data converted to the specified above format trainer.max_epochs integer 5 Maximum number of epochs to train the model model.label_ids string – Path to the string labels to integer mapping (is generated during the dataset conversion step) model.tokenizer.tokenizer_name string Will be filled automatically based on model.language_model.pretrained_model_name Tokenizer name model.tokenizer.vocab_file string null Path to tokenizer vocabulary model.tokenizer.tokenizer_model string null Path to tokenizer model (only for sentencepiece tokenizer) model.language_model.pretrained_model_name string bert-base-uncased Pre-trained language model name (choose from bert-base-cased and bert-base-uncased) model.language_model.lm_checkpoint string null Path to the pre-trained language model checkpoint model.language_model.config_file string null Path to the pre-trained language model config file model.language_model.config dictionary null Config of the pre-trained language model model.head.num_fc_layers integer 2 Number of fully connected layers model.head.fc_dropout float 0.5 Activation to use between fully connected layers model.head.activation string ‘relu’ Dropout to apply to the input hidden states model.punct_head.use_transrormer_init bool True Whether to initialize the weights of the classifier head with the same approach used in Transformer training_ds.text_file string text_train.txt Name of the text training file located at data_dir training_ds.labels_file string labels_train.txt Name of the labels training file located at data_dir training_ds.shuffle bool True Whether to shuffle the training data training_ds.num_samples integer -1 Number of samples to use from the training dataset; -1 means all training_ds.batch_size integer 64 Training data batch size validation_ds.text_file string text_dev.txt Name of the text file for evaluation, located at data_dir validation_ds.labels_file string labels_dev.txt Name of the labels dev file located at data_dir validation_ds.shuffle bool False Whether to shuffle the dev data validation_ds.num_samples integer -1 Number of samples to use from the dev set; -1 means all validation_ds.batch_size integer 64 Dev set batch size optim.name string adam Optimizer to use for training optim.lr float 5e-5 Learning rate to use for training optim.weight_decay float 0 Weight decay to use for training optim.sched.name string WarmupAnnealing Warmup schedule optim.sched.warmup_ratio float 0.1 Warmup ratio Example of the command for training the model: Copy Copied!  tao token_classification train [-h] \ -e /specs/nlp/token_classification/train.yaml \ -r /results/token_classification/train/ \ -g 1 \ -k$KEY
data_dir=/path/to/data_dir \
model.label_ids=/path/to/label_ids.csv \
trainer.max_epochs=5 \
training_ds.num_samples=-1 \
validation_ds.num_samples=-1


### Required Arguments for Training

• -e: The experiment specification file to set up training.

• -r: Path to the directory to store the results/logs. Note, the trained-model.tlt would be saved in this

specified folder under a subfolder checkpoints; in our case it will be saved here: /results/token_classification/train/checkpoints/trained-model.tlt

• -k: Encryption key

• data_dir: Path to the data_dir with the processed data files.

• model.label_ids: Path to the label_ids.csv file, usually stored at data_dir

### Optional Arguments

• -h, --help: Show this help message and exit

• -g: The number of GPUs to be used in evaluation in a multi-GPU scenario (default: 1).

• Other arguments to override fields in the specification file.

Note

While the arguments are defined in the spec file, if you wish to override these parameter definitions in the spec file and experiment with them, you may do so over command line by simply defining the param. For example, the sample spec file mentioned above has validation_ds.batch_size set to 64. However, if you see that the GPU utilization can be optimized further by using larger a batch size, you may override to the desired value, by adding the field validation_ds.batch_size=128 over command line. You may repeat this with any of the parameters defined in the sample spec file.

Snippets of the output log from executing the token_classification train command:

Copy
Copied!

# complete model's spec file will be shown
[NeMo I train:93] Spec file:
restore_from: ???
exp_manager:
explicit_log_dir: /results/token_classification/train/
exp_dir: null
name: trained-model
version: null
use_datetime_version: true
resume_if_exists: true
resume_past_end: false
resume_ignore_no_checkpoint: true
create_tensorboard_logger: false
summary_writer_kwargs: null
create_wandb_logger: false
wandb_logger_kwargs: null
create_checkpoint_callback: true
checkpoint_callback_params:
filepath: null
monitor: val_loss
verbose: true
save_last: true
save_top_k: 3
save_weights_only: false
mode: auto
period: 1
prefix: null
postfix: .tlt
save_best_model: false
files_to_copy: null
model:
tokenizer:
tokenizer_name: ...
...

[NeMo I exp_manager:186] Experiments will be logged at /results/token_classification/train/

# The dataset will be processed and tokenized
[NeMo I token_classification_model:61] Reusing label_ids file found at data_dir/label_ids.csv.
Using bos_token, but it is not set yet.
Using eos_token, but it is not set yet.
[NeMo I token_classification_model:105] Setting model.dataset.data_dir to data_dir.

[NeMo I 2021-01-21 17:57:14 token_classification_utils:54] Processing data_dir/labels_train.txt
[NeMo I 2021-01-21 17:57:14 token_classification_utils:75] Using provided labels mapping {'O': 0, 'B-GPE': 1, 'B-LOC': 2, 'B-MISC': 3, 'B-ORG': 4, 'B-PER': 5, 'B-TIME': 6, 'I-GPE': 7, 'I-LOC': 8, 'I-MISC': 9, 'I-ORG': 10, 'I-PER': 11, 'I-TIME': 12}
[NeMo I 2021-01-21 17:57:15 token_classification_utils:101] Three most popular labels in data_dir/labels_train.txt:
[NeMo I 2021-01-21 17:57:15 data_preprocessing:131] label: 0, 18417 out of 21717 (84.80%).
[NeMo I 2021-01-21 17:57:15 data_preprocessing:131] label: 2, 829 out of 21717 (3.82%).
[NeMo I 2021-01-21 17:57:15 data_preprocessing:131] label: 6, 433 out of 21717 (1.99%).
[NeMo I 2021-01-21 17:57:15 token_classification_utils:103] Total labels: 21717. Label frequencies - {0: 18417, 2: 829, 6: 433, 4: 357, 11: 352, 5: 349, 1: 338, 10: 281, 8: 181, 12: 142, 3: 21, 9: 12, 7: 5}
[NeMo I 2021-01-21 17:57:15 token_classification_utils:112] Class Weights: {0: 0.09070632901875775, 2: 2.015124802820822, 6: 3.858056493160419, 4: 4.679379444085327, 11: 4.7458479020979025, 5: 4.786643156270664, 1: 4.942421483841602, 10: 5.9449767314535995, 8: 9.229494262643433, 12: 11.764355362946912, 3: 79.54945054945055, 9: 139.21153846153845, 7: 334.10769230769233}
[NeMo I 2021-01-21 17:57:15 token_classification_utils:116] Class weights saved to data_dir/labels_train_weights.p
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:116] Setting Max Seq length to: 64
[NeMo I 2021-01-21 17:57:19 data_preprocessing:295] Some stats of the lengths of the sequences:
[NeMo I 2021-01-21 17:57:19 data_preprocessing:301] Min: 6 |                  Max: 64 |                  Mean: 26.357 |                  Median: 26.0
[NeMo I 2021-01-21 17:57:19 data_preprocessing:303] 75 percentile: 32.00
[NeMo I 2021-01-21 17:57:19 data_preprocessing:304] 99 percentile: 51.00
[NeMo W 2021-01-21 17:57:19 token_classification_dataset:145] 0 are longer than 64
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:148] *** Example ***
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:149] i: 0
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:150] subtokens: [CLS] new zealand ' s cricket team has scored a morale - boost ##ing win over bangladesh in the first of three one - day internationals in new zealand . [SEP]
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:151] loss_mask: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:152] input_mask: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:153] subtokens_mask: 0 1 1 1 0 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 1 1 0 0 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:155] labels: 0 2 8 0 0 0 0 0 0 0 0 0 0 0 0 0 2 0 0 6 12 12 12 12 12 0 0 2 8 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:19 token_classification_dataset:264] features saved to data_dir/cached_text_train.txt_BertTokenizer_128_30522_-1
[NeMo I 2021-01-21 17:57:19 token_classification_utils:54] Processing data_dir/labels_dev.txt
[NeMo I 2021-01-21 17:57:19 token_classification_utils:75] Using provided labels mapping {'O': 0, 'B-GPE': 1, 'B-LOC': 2, 'B-MISC': 3, 'B-ORG': 4, 'B-PER': 5, 'B-TIME': 6, 'I-GPE': 7, 'I-LOC': 8, 'I-MISC': 9, 'I-ORG': 10, 'I-PER': 11, 'I-TIME': 12}
[NeMo I 2021-01-21 17:57:20 token_classification_utils:101] Three most popular labels in data_dir/labels_dev.txt:
[NeMo I 2021-01-21 17:57:20 data_preprocessing:131] label: 0, 18266 out of 21775 (83.89%).
[NeMo I 2021-01-21 17:57:20 data_preprocessing:131] label: 2, 809 out of 21775 (3.72%).
[NeMo I 2021-01-21 17:57:20 data_preprocessing:131] label: 6, 435 out of 21775 (2.00%).
[NeMo I 2021-01-21 17:57:20 token_classification_utils:103] Total labels: 21775. Label frequencies - {0: 18266, 2: 809, 6: 435, 4: 418, 11: 414, 5: 392, 1: 351, 10: 351, 8: 174, 12: 146, 7: 8, 3: 8, 9: 3}
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:116] Setting Max Seq length to: 70
[NeMo I 2021-01-21 17:57:24 data_preprocessing:295] Some stats of the lengths of the sequences:
[NeMo I 2021-01-21 17:57:24 data_preprocessing:301] Min: 7 |                  Max: 70 |                  Mean: 26.437 |                  Median: 26.0
[NeMo I 2021-01-21 17:57:24 data_preprocessing:303] 75 percentile: 33.00
[NeMo I 2021-01-21 17:57:24 data_preprocessing:304] 99 percentile: 50.00
[NeMo W 2021-01-21 17:57:24 token_classification_dataset:145] 0 are longer than 70
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:148] *** Example ***
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:149] i: 0
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:150] subtokens: [CLS] hamas refuses to recognize israel , and has vowed to undermine palestinian leader mahmoud abbas ' s efforts to make peace with the jewish state . [SEP]
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:151] loss_mask: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:152] input_mask: 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:153] subtokens_mask: 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:155] labels: 0 4 0 0 0 2 0 0 0 0 0 0 1 0 5 11 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[NeMo I 2021-01-21 17:57:24 token_classification_dataset:264] features saved to data_dir/cached_text_dev.txt_BertTokenizer_128_30522_-1

[NeMo I 2021-01-21 17:00:09 modelPT:830] Optimizer config = Adam (
Parameter Group 0
betas: (0.9, 0.999)
eps: 1e-08
lr: 5e-05
weight_decay: 0.0
)
[NeMo I 2021-01-21 17:00:09 lr_scheduler:621] Scheduler "<nemo.core.optim.lr_scheduler.WarmupAnnealing object at 0x7f3b6d05f400>"
will be used during training (effective maximum steps = 16) -
Parameters :
(warmup_steps: null
warmup_ratio: 0.1
last_epoch: -1
max_steps: 16
)
initializing ddp: GLOBAL_RANK: 0, MEMBER: 1/1
[NeMo I 2021-01-21 17:00:11 modelPT:704] No optimizer config provided, therefore no optimizer was created

110 M     Trainable params
0         Non-trainable params
110 M     Total params
Validation sanity check:  50%|████████████████████████████▌                            | 1/2 [00:00<00:00,  1.47it/s][NeMo I 2021-01-21 17:00:13 token_classification_model:178]
label                                                precision    recall       f1           support
O (label_id: 0)                                         82.08     100.00      90.16       2300
B-GPE (label_id: 1)                                      0.00       0.00       0.00         41
B-LOC (label_id: 2)                                      0.00       0.00       0.00        119
B-MISC (label_id: 3)                                     0.00       0.00       0.00          2
B-ORG (label_id: 4)                                      0.00       0.00       0.00         71
B-PER (label_id: 5)                                      0.00       0.00       0.00         62
B-TIME (label_id: 6)                                     0.00       0.00       0.00         56
I-GPE (label_id: 7)                                      0.00       0.00       0.00          4
I-LOC (label_id: 8)                                      0.00       0.00       0.00         18
I-MISC (label_id: 9)                                     0.00       0.00       0.00          0
I-ORG (label_id: 10)                                     0.00       0.00       0.00         52
I-PER (label_id: 11)                                     0.00       0.00       0.00         61
I-TIME (label_id: 12)                                    0.00       0.00       0.00         16
-------------------
micro avg                                               82.08      82.08      82.08       2802
macro avg                                                6.84       8.33       7.51       2802
weighted avg                                            67.38      82.08      74.01       2802

Training: 0it [00:00, ?it/s]
[NeMo I 2021-01-21 17:00:38 train:124] Experiment logs saved to 'output'
[NeMo I 2021-01-21 17:00:38 train:127] Trained model saved to 'output/checkpoints/trained-model.tlt'
INFO: Internal process exited


### Important Parameters

Below is the list of parameters that could help improve the model:

• dropout value between layers (model.head.fc_dropout)

• optimizer (model.optim.name, for example, adam)

• learning rate (model.optim.lr, for example, 5e-5)

## Fine-tuning a Model on a Different Dataset

In the previous section, Training a token classification model,

the Token Classification (NER) model was initialized with a pre-trained language model, but the classifiers were trained from scratch. Now that a user has trained the Token Classification model successfully (e.g., called trained-model.tlt), there may be scenarios where users are required to retrain this trained-model.tlt on a new smaller dataset. TAO conversational AI applications provide a separate tool called fine-tune to enable this.

Note

Labels from the dataset that is used for fine-tuning, should be a subset of the labels of the pre-trained .tlt model. If it is not the case, use the tao token_classification train with your data.

Example for spec for fine-tuning of the model:

Copy
Copied!

trainer:
max_epochs: 5

data_dir: ???

# Fine-tuning settings: training dataset.
finetuning_ds:
num_samples: -1 # number of samples to be considered, -1 means all the dataset

# Fine-tuning settings: validation dataset.
validation_ds:
num_samples: -1 # number of samples to be considered, -1 means all the dataset

# Fine-tuning settings: different optimizer.
optim:
lr: 1e-5


 Parameter Data Type Default Description data_dir string – Path to the data converted to the specified above format trainer.max_epochs integer 5 Maximum number of epochs to train the model finetuning_ds.text_file string text_train.txt Name of the text training file located at data_dir finetuning_ds.labels_file string labels_train.txt Name of the labels training file located at data_dir finetuning_ds.shuffle bool True Whether to shuffle the training data finetuning_ds.num_samples integer -1 Number of samples to use from the training dataset; -1 means all finetuning_ds.batch_size integer 64 Training data batch size validation_ds.text_file string text_dev.txt Name of the text file for evaluation, located at data_dir validation_ds.labels_file string labels_dev.txt Name of the labels dev file located at data_dir validation_ds.shuffle bool False Whether to shuffle the dev data validation_ds.num_samples integer -1 Number of samples to use from the dev set; -1 means all validation_ds.batch_size integer 64 Dev set batch size optim.name string adam Optimizer to use for training optim.lr float 1e-5 Learning rate to use for training

Use the following command to fine-tune the model:

Copy
Copied!

tao token_classification finetune [-h] \
-e /specs/nlp/token_classification/finetune.yaml \
-r /results/token_classification/finetune/ \
-m /results/token_classification/train/checkpoints/trained-model.tlt \
-g 1 \
data_dir=PATH_TO_DATA \
trainer.max_epochs=5 \
-k $KEY  ### Required Arguments for Fine-tuning • -h, --help: Show this help message and exit • -e: The experiment specification file to set up fine-tuning. • -r: Path to the directory to store the results/logs. Note, the finetuned-model.tlt would be saved in this specified folder under a subfolder checkpoints; in our case it will be saved here: /results/token_classification/train/checkpoints/trained-model.tlt • -m: Path to the pre-trained model to use for fine-tuning. • data_dir: Path to data directory with the pre-processed data to use for fine-tuning • -k: Encryption key ### Optional Arguments • -g: The number of GPUs to be used in evaluation in a multi-GPU scenario (default: 1). • Other arguments to override fields in the specification file. Output log for the tao token_calssification finetune command: Copy Copied!  Model restored from '/path/to/trained-model.tlt' # The rest of the log is similar to the output log snippet for :code:token_classification train.  ## Evaluating a Trained Model Spec example to evaluate the pre-trained model: Copy Copied!  restore_from: trained-model.tlt data_dir: ??? # Test settings: dataset. test_ds: text_file: text_dev.txt labels_file: labels_dev.txt batch_size: 1 shuffle: false num_samples: -1 # number of samples to be considered, -1 means the whole the dataset  Use the following command to evaluate the model: Copy Copied!  tao token_classification evaluate [-h] \ -e /specs/nlp/token_classification/evaluate.yaml \ -r /results/token_classification/evaluate/ \ -g 1 \ -m /results/token_classification/train/checkpoints/trained-model.tlt \ -k$KEY \
data_dir=/path/to/data_dir


### Required Arguments for Evaluation

• -e: The experiment specification file to set up evaluation.

• -r: Path to the directory to store the results.

• data_dir: Path to data directory with the pre-processed data to use for evaluation

• -m: Path to the pre-trained model checkpoint for evaluation. Should be a .tlt file.

• -k: Encryption key

### Optional Arguments for Evaluation

• -h, --help: Show this help message and exit

 Parameter Data Type Default Description restore_from string trained-model.tlt Path to the pre-trained model data_dir string – Path to the data converted to the specified above format test_ds.text_file string text_dev.txt Name of the text file to run evaluation on located at data_dir test_ds.labels_file string labels_dev.txt Name of the labels dev file located at data_dir test_ds.shuffle bool False Whether to shuffle the dev data test_ds.num_samples integer -1 Number of samples to use from the dev set; -1 means all test_ds.batch_size integer 64 Dev set batch size

token_classification evaluate generates a classification report that includes the following metrics:

• Precision

• Recall

• F1

More details about these metrics can be found here.

Output log for token_classification evaluate (note, the values below are for demonstration purposes only):

Copy
Copied!

label                                                precision    recall       f1           support
O (label_id: 0)                                         83.89     100.00      91.24      18266
B-GPE (label_id: 1)                                      0.00       0.00       0.00        351
B-LOC (label_id: 2)                                      0.00       0.00       0.00        809
B-MISC (label_id: 3)                                     0.00       0.00       0.00          8
B-ORG (label_id: 4)                                      0.00       0.00       0.00        418
B-PER (label_id: 5)                                      0.00       0.00       0.00        392
B-TIME (label_id: 6)                                     0.00       0.00       0.00        435
I-GPE (label_id: 7)                                      0.00       0.00       0.00          8
I-LOC (label_id: 8)                                      0.00       0.00       0.00        174
I-MISC (label_id: 9)                                     0.00       0.00       0.00          3
I-ORG (label_id: 10)                                     0.00       0.00       0.00        351
I-PER (label_id: 11)                                     0.00       0.00       0.00        414
I-TIME (label_id: 12)                                    0.00       0.00       0.00        146
-------------------
micro avg                                               83.89      83.89      83.89      21775
macro avg                                                6.45       7.69       7.02      21775
weighted avg                                            70.37      83.89      76.53      21775

Testing: 100%|████████████████████████████████████████████████████████████████████████████████████████| 1000/1000 [00:39<00:00, 25.59it/s]
--------------------------------------------------------------------------------
{'f1': tensor(7.0182, device='cuda:0'),
'precision': tensor(6.4527, device='cuda:0'),
'recall': tensor(7.6923, device='cuda:0'),
'test_loss': tensor(1.0170, device='cuda:0')}


## Running Inference using a Trained Model

During inference, a batch of input sentences, listed in the spec files, are passed through the trained model to add token classification label.

To run inference on the model, specify the list of examples in the spec, for example:

Copy
Copied!

input_batch:
- 'We bought four shirts from the Nvidia gear store in Santa Clara.'
- 'Nvidia is a company.'


To run inference:

Copy
Copied!

tao token_classification infer [-h] \
-e /specs/nlp/token_classification/infer.yaml \
-r /results/token_classification/infer/ \
-g 1 \
-m /results/token_classification/checkpoints/trained-model.tlt \
-k $KEY  ### Required Arguments for Inference • -e: The experiment specification file to set up inference. This requires the input_batch with the list of examples to run inference on. • -r: Path to the directory to store the results. • -m: Path to the pre-trained model checkpoint from which to infer. Should be a .tlt file. • -k: Encryption key ### Optional Arguments • -h, --help: Show this help message and exit • -g: The number of GPUs to be used for fine-tuning in a multi-GPU scenario (default: 1). • Other arguments to override fields in the specification file. Output log sample: Copy Copied!  Query : we bought four shirts from the nvidia gear store in santa clara. Result: we bought four shirts from the nvidia[B-LOC] gear store in santa[B-LOC] clara[I-LOC]. Nvidia is a company. Result: Nvidia[B-ORG] is a company.  ## Model Export A pre-trained model could be exported to RIVA format (this format contains model checkpoint along with model artifacts required for successful deployment of the trained .tlt models to Riva Services). For more details about Riva, see this. An example of the spec file for model export: Copy Copied!  # Name of the .tlt EFF archive to be loaded/model to be exported. restore_from: trained-model.tlt # Set export format: RIVA export_format: RIVA # Output EFF archive containing model checkpoint and artifacts required for Riva Services export_to: exported-model.riva   Parameter Data Type Default Description restore_from string trained-model.tlt Path to the pre-trained model export_format string RIVA Export format: RIVA export_to string exported-model.riva Path to the exported model To export a pre-trained model for deployment, run: Copy Copied!  ### For export to Riva format tao token_classification export [-h] \ -e /specs/nlp/token_classification/export.yaml \ -r /results/token_classification/export/ \ -m /results/token_classification/checkpoints/trained-model.tlt \ -k$KEY
export_format=RIVA


### Required Arguments for Export

• -e: The experiment specification file to set up inference. This requires the input_batch with the list of examples to run inference on.

• -r: Path to the directory to store the results.

• -m: Path to the pre-trained model checkpoint from which to infer. Should be a .tlt file.

• -k: Encryption key

### Optional Arguments for Export

• -h, --help: Show this help message and exit

• export_to: To change the default name of the exported model

Output log:

Copy
Copied!

Spec file:
restore_from: path/to/trained-model.tlt
export_to: exported-model.riva
export_format: RIVA
exp_manager:
Exported model to '/results/token_classification/export/exported-model.riva'