the-algorithm/trust_and_safety_models/abusive/abusive_model.py

276 lines
9.0 KiB
Python

import tensorflow as tf
physical_devices = tf.config.list_physical_devices('GPU')
for device in physical_devices:
tf.config.experimental.set_memory_growth(device, True)
from twitter.hmli.nimbus.modeling.model_config import FeatureType, EncodingType, Feature, Model, LogType
from twitter.hmli.nimbus.modeling.feature_loader import BigQueryFeatureLoader
from twitter.cuad.representation.models.text_encoder import TextEncoder
from twitter.cuad.representation.models.optimization import create_optimizer
from twitter.hmli.nimbus.modeling.feature_encoder import FeatureEncoder
import numpy as np
import pandas as pd
import utils
cat_names = [
...
]
category_features = [Feature(name=cat_name, ftype=FeatureType.CONTINUOUS) for cat_name in cat_names]
features = [
Feature(name="tweet_text_with_media_annotations", ftype=FeatureType.STRING, encoding=EncodingType.BERT),
Feature(name="precision_nsfw", ftype=FeatureType.CONTINUOUS),
Feature(name="has_media", ftype=FeatureType.BINARY),
Feature(name="num_media", ftype=FeatureType.DISCRETE)
] + category_features
ptos_prototype = Model(
name='ptos_prototype',
export_path="...",
features=features,
)
print(ptos_prototype)
cq_loader = BigQueryFeatureLoader(gcp_project=COMPUTE_PROJECT)
labels = [
"has_non_punitive_action",
"has_punitive_action",
"has_punitive_action_contains_self_harm",
"has_punitive_action_encourage_self_harm",
"has_punitive_action_episodic",
"has_punitive_action_episodic_hateful_conduct",
"has_punitive_action_other_abuse_policy",
"has_punitive_action_without_self_harm"
]
train_query = f"""
SELECT
{{feature_names}},
{",".join(labels)},
...
"""
val_query = f"""
SELECT
{{feature_names}},
{",".join(labels)},
...
"""
print(train_query)
train = cq_loader.load_features(ptos_prototype, "", "", custom_query=train_query)
val = cq_loader.load_features(ptos_prototype, "", "", custom_query=val_query)
print(train.describe(model=ptos_prototype))
params = {
'max_seq_lengths': 128,
'batch_size': 196,
'lr': 1e-5,
'optimizer_type': 'adamw',
'warmup_steps': 0,
'cls_dropout_rate': 0.1,
'epochs': 30,
'steps_per_epoch': 5000,
'model_type': 'twitter_multilingual_bert_base_cased_mlm',
'mixed_precision': True,
}
params
def parse_labeled_data(row_dict):
label = [row_dict.pop(l) for l in labels]
return row_dict, label
mirrored_strategy = tf.distribute.MirroredStrategy()
BATCH_SIZE = params['batch_size'] * mirrored_strategy.num_replicas_in_sync
train_ds = train.to_tf_dataset().map(parse_labeled_data).shuffle(BATCH_SIZE*100).batch(BATCH_SIZE).repeat()
val_ds = val.to_tf_dataset().map(parse_labeled_data).batch(BATCH_SIZE)
for record in train_ds:
tf.print(record)
break
def get_positive_weights():
"""Computes positive weights used for class imbalance from training data."""
label_weights_df = utils.get_label_weights(
"tos-data-media-full",
project_id="twttr-abusive-interact-prod",
dataset_id="tos_policy"
)
pos_weight_tensor = tf.cast(
label_weights_df.sort_values(by='label').positive_class_weight,
dtype=tf.float32
)
return pos_weight_tensor
pos_weight_tensor = get_positive_weights()
print(pos_weight_tensor)
class TextEncoderPooledOutput(TextEncoder):
def call(self, x):
return super().call([x])["pooled_output"]
def get_config(self):
return super().get_config()
with mirrored_strategy.scope():
text_encoder_pooled_output = TextEncoderPooledOutput(
params['max_seq_lengths'],
model_type=params['model_type'],
trainable=True
)
fe = FeatureEncoder(train)
inputs, preprocessing_head = fe.build_model_head(model=ptos_prototype, text_encoder=text_encoder_pooled_output)
cls_dropout = tf.keras.layers.Dropout(params['cls_dropout_rate'], name="cls_dropout")
outputs = cls_dropout(preprocessing_head)
outputs = tf.keras.layers.Dense(8, name="output", dtype="float32")(outputs)
model = tf.keras.Model(
inputs=inputs,
outputs=outputs
)
pr_auc = tf.keras.metrics.AUC(curve="PR", num_thresholds=1000, multi_label=True, from_logits=True)
custom_loss = lambda y_true, y_pred: utils.multilabel_weighted_loss(y_true, y_pred, weights=pos_weight_tensor)
optimizer = create_optimizer(
init_lr=params["lr"],
num_train_steps=(params["epochs"] * params["steps_per_epoch"]),
num_warmup_steps=params["warmup_steps"],
optimizer_type=params["optimizer_type"],
)
if params.get("mixed_precision"):
optimizer = tf.train.experimental.enable_mixed_precision_graph_rewrite(optimizer)
model.compile(
optimizer=optimizer,
loss=custom_loss,
metrics=[pr_auc]
)
model.weights
model.summary()
pr_auc.name
import getpass
import wandb
from wandb.keras import WandbCallback
try:
wandb_key = ...
wandb.login(...)
run = wandb.init(project='ptos_with_media',
group='new-split-trains',
notes='tweet text with only (num_media, precision_nsfw). on full train set, new split.',
entity='absv',
config=params,
name='tweet-text-w-nsfw-1.1',
sync_tensorboard=True)
except FileNotFoundError:
print('Wandb key not found')
run = wandb.init(mode='disabled')
import datetime
import os
start_train_time = datetime.datetime.now()
print(start_train_time.strftime("%m-%d-%Y (%H:%M:%S)"))
checkpoint_path = os.path.join("...")
print("Saving model checkpoints here: ", checkpoint_path)
cp_callback = tf.keras.callbacks.ModelCheckpoint(
filepath=os.path.join(checkpoint_path, "model.{epoch:04d}.tf"),
verbose=1,
monitor=f'val_{pr_auc.name}',
mode='max',
save_freq='epoch',
save_best_only=True
)
early_stopping_callback = tf.keras.callbacks.EarlyStopping(patience=7,
monitor=f"val_{pr_auc.name}",
mode="max")
model.fit(train_ds, epochs=params["epochs"], validation_data=val_ds, callbacks=[cp_callback, early_stopping_callback],
steps_per_epoch=params["steps_per_epoch"],
verbose=2)
import tensorflow_hub as hub
gs_model_path = ...
reloaded_keras_layer = hub.KerasLayer(gs_model_path)
inputs = tf.keras.layers.Input(name="tweet__core__tweet__text", shape=(1,), dtype=tf.string)
output = reloaded_keras_layer(inputs)
v7_model = tf.keras.models.Model(inputs=inputs, outputs=output)
pr_auc = tf.keras.metrics.AUC(curve="PR", name="pr_auc")
roc_auc = tf.keras.metrics.AUC(curve="ROC", name="roc_auc")
v7_model.compile(metrics=[pr_auc, roc_auc])
model.load_weights("...")
candidate_model = model
with mirrored_strategy.scope():
candidate_eval = candidate_model.evaluate(val_ds)
test_query = f"""
SELECT
{",".join(ptos_prototype.feature_names())},
has_media,
precision_nsfw,
{",".join(labels)},
...
"""
test = cq_loader.load_features(ptos_prototype, "", "", custom_query=test_query)
test = test.to_tf_dataset().map(parse_labeled_data)
print(test)
test_only_media = test.filter(lambda x, y: tf.equal(x["has_media"], True))
test_only_nsfw = test.filter(lambda x, y: tf.greater_equal(x["precision_nsfw"], 0.95))
test_no_media = test.filter(lambda x, y: tf.equal(x["has_media"], False))
test_media_not_nsfw = test.filter(lambda x, y: tf.logical_and(tf.equal(x["has_media"], True), tf.less(x["precision_nsfw"], 0.95)))
for d in [test, test_only_media, test_only_nsfw, test_no_media, test_media_not_nsfw]:
print(d.reduce(0, lambda x, _: x + 1).numpy())
from notebook_eval_utils import SparseMultilabelEvaluator, EvalConfig
from dataclasses import asdict
def display_metrics(probs, targets, labels=labels):
eval_config = EvalConfig(prediction_threshold=0.5, precision_k=0.9)
for eval_mode, y_mask in [("implicit", np.ones(targets.shape))]:
print("Evaluation mode", eval_mode)
metrics = SparseMultilabelEvaluator.evaluate(
targets, np.array(probs), y_mask, classes=labels, eval_config=eval_config
)
metrics_df = pd.DataFrame.from_dict(asdict(metrics)["per_topic_metrics"]).transpose()
metrics_df["pos_to_neg"] = metrics_df["num_pos_samples"] / (metrics_df["num_neg_samples"] + 1)
display(metrics_df.median())
display(metrics_df)
return metrics_df
def eval_model(model, df):
with mirrored_strategy.scope():
targets = np.stack(list(df.map(lambda x, y: y).as_numpy_iterator()), axis=0)
df = df.padded_batch(BATCH_SIZE)
preds = model.predict(df)
return display_metrics(preds, targets)
subsets = {"test": test,
"test_only_media": test_only_media,
"test_only_nsfw": test_only_nsfw,
"test_no_media": test_no_media,
"test_media_not_nsfw": test_media_not_nsfw}
metrics = {}
for name, df in subsets.items():
metrics[name] = eval_model(candidate_model, df)
[(name, m.pr_auc) for name, m in metrics.items()]
for name, x in [(name, m.pr_auc.to_string(index=False).strip().split("\n")) for name, m in metrics.items()]:
print(name)
for y in x:
print(y.strip(), end="\t")
print(".")
for d in [test, test_only_media, test_only_nsfw, test_no_media, test_media_not_nsfw]:
print(d.reduce(0, lambda x, _: x + 1).numpy())