File size: 10,683 Bytes
6ef31de |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 |
import functools
from absl import app
from absl import flags
from absl import logging
from clu import metric_writers
from clu import platform
import flax.linen as nn
import jax
from ml_collections import config_flags
import tensorflow as tf
import sys, os
from pathlib import Path
# append current path to sys.path
sys.path.append(str(Path(__file__).parent.parent.parent / "scenic"))
import logging
import flax
from flax import jax_utils
from flax.training import checkpoints
from scenic.projects.vid2seq import models, trainer
from scenic.train_lib_deprecated import train_utils
from scenic import app
import ml_collections
import numpy as np
import jax.numpy as jnp
from clu import metric_writers
from scenic.projects.vid2seq.datasets.dense_video_captioning_tfrecord_dataset import get_datasets
from scenic.projects.vid2seq import dvc_eval
MAX_CAPTION_STR_LEN = 200
MAX_KEY_STR_LEN = 400
class ScenicModel:
def __init__(self, flags):
self.FLAGS = flags
jax.config.config_with_absl()
run = (functools.partial(self._run_main, main=self._init_model))
run(self._init_model)
def _run_main(self, argv, *, main):
"""Runs the `main` method after some initial setup."""
del argv
# Hide any GPUs form TensorFlow. Otherwise, TF might reserve memory and make
# it unavailable to JAX.
tf.config.experimental.set_visible_devices([], 'GPU')
# Enable wrapping of all module calls in a named_call for easier profiling:
nn.enable_named_call()
logging.info('JAX host: %d / %d', jax.process_index(), jax.process_count())
logging.info('JAX devices: %r', jax.devices())
# Add a note so that we can tell which task is which JAX host.
# (task 0 is not guaranteed to be the host 0)
platform.work_unit().set_task_status(
f'host_id: {jax.process_index()}, host_count: {jax.process_count()}')
if jax.process_index() == 0:
platform.work_unit().create_artifact(platform.ArtifactType.DIRECTORY,
self.FLAGS.workdir, 'Workdir')
self.FLAGS.config.dataset_configs.base_dir = self.FLAGS.data_dir
self.FLAGS.config.init_from.checkpoint_path = self.FLAGS.ckpt_dir
rng = jax.random.PRNGKey(self.FLAGS.config.rng_seed)
logging.info('RNG: %s', rng)
writer = metric_writers.create_default_writer(
self.FLAGS.workdir, just_logging=jax.process_index() > 0, asynchronous=True)
return main(rng=rng, config=self.FLAGS.config, workdir=self.FLAGS.workdir, writer=writer)
def _init_model(self, rng: jnp.ndarray, config: ml_collections.ConfigDict, workdir: str,
writer: metric_writers.MetricWriter):
data_rng, rng = jax.random.split(rng)
dataset_dict = get_datasets(config, data_rng=data_rng)
datasets_metadata = {
name: ds.meta_data
for name, ds in dataset_dict.items()
}
all_datasets = []
all_datasets_num_train_examples = []
for name, metadata in datasets_metadata.items():
all_datasets.append(name)
all_datasets_num_train_examples.append(
metadata.get('num_train_examples', 0))
dataset = dataset_dict[all_datasets[0]]
model_cls = models.DenseVideoCaptioningModel
model = model_cls(config, dataset.meta_data)
train_state, start_step = trainer.init_state(model, dataset, config,
workdir, rng)
self.train_state = jax_utils.replicate(train_state)
logging.info('Number of processes is %s', jax.process_count())
del rng
import functools
self.infer_step_pmapped = jax.pmap(
functools.partial(
trainer.infer_step,
model=model,
config=config,
debug=config.debug_eval),
axis_name='batch',
)
self.tokenizer = trainer.get_tokenizer(config)
# dsname = 'validation'
# self.iterator = dataset.valid_iter[dsname]
self.config = config
self.data_rng = data_rng
def __call__(self, data_dir=None):
# self.FLAGS.config.dataset_configs.base_dir = data_dir
dataset_dict = get_datasets(self.config, data_rng=self.data_rng)
self.iterator = dataset_dict["youcook"].valid_iter['validation']
batch = next(self.iterator)
train_state = train_utils.sync_model_state_across_replicas(self.train_state)
eval_packs = {}
keys = []
eval_pack = {
'gts':
dvc_eval.convert_strings_to_uint8_arrays(
batch['caption_strings'], MAX_CAPTION_STR_LEN),
'key':
dvc_eval.convert_strings_to_uint8_arrays(
batch['videoid'], MAX_KEY_STR_LEN),
'batch_mask':
batch['batch_mask'],
'duration':
batch['duration'],
'gts_start':
batch['timestamp_start'],
'gts_end':
batch['timestamp_end'],
'split':
batch['split'] if 'split' in batch else
np.ones_like(batch['timestamp_start']),
}
to_del = ['caption_strings', 'key', 'videoid', 'timestamp_start',
'timestamp_end', 'split'] # 'duration',
for x in to_del:
if x in batch:
del batch[x]
# import pdb
# pdb.set_trace()
_, preds = self.infer_step_pmapped(train_state, batch) #model, config)
# import pdb
# pdb.set_trace()
eval_pack['pred'] = preds
eval_pack = jax.tree_map(
lambda x: x.reshape((np.prod(x.shape[:2]),) + x.shape[2:]), eval_pack)
vocabulary_size = self.config.dataset_configs.vocabulary_size
# pred_text = trainer.decode_tokens(preds, tokenizer, vocabulary_size)
# print(preds, pred_text)
format_outputs = []
for i, valid in enumerate(eval_pack['batch_mask']):
print("===============video[", str(0), "]====================")
if valid:
key = dvc_eval.convert_uint8_array_to_string(eval_pack['key'][i])
if key in eval_packs: # redundant video
continue
keys.append(key)
pred, pred_timestamps = [], []
# get indexes in the predicted seq that delimit the pred segments
indexes = [
j for j in range(len(eval_pack['pred'][i]) - 1)
if eval_pack['pred'][i][j] >= vocabulary_size and
eval_pack['pred'][i][j + 1] >= vocabulary_size
] # pylint: disable=g-complex-comprehension
last_processed = -2
order = self.config.dataset_configs.order
# iterate over predicted segments and decode them
for j in range(len(indexes)):
if indexes[j] == last_processed + 1: # 3 timestamps != 2 events
continue
# get predicted tokens and transform to string
if order == 'ld':
start_idx = indexes[j] + 2
end_idx = indexes[j + 1] if j < len(indexes) - 1 else len(
eval_pack['pred'][i])
else:
start_idx = indexes[j - 1] + 2 if j > 0 else 0
end_idx = indexes[j]
pred_seq = [int(eval_pack['pred'][i][k]) for k in range(start_idx, end_idx)]
pred_text = trainer.decode_tokens(pred_seq, self.tokenizer, vocabulary_size)
# get start and end
num_bins = 100 # from config
max_offset = num_bins - 1
pred_time = [
(int(eval_pack['pred'][i][indexes[j]])
- vocabulary_size) *
eval_pack['duration'][i] / max_offset,
(int(eval_pack['pred'][i][indexes[j] + 1]) -
vocabulary_size) *
eval_pack['duration'][i] / max_offset
]
# if pred_time[1] <= pred_time[0]: # remove end < start
# continue
last_processed = indexes[j]
pred.append(pred_text)
pred_timestamps.append(pred_time)
# round to 2 decimal places
format_output = "[{x}s, {y}s] ".format(x=np.around(pred_time[0][0]/1000000, decimals=2), y=np.around(pred_time[1][0]/1000000, decimals=2))
format_output += pred_text
format_outputs.append(format_output)
print(format_outputs)
print("===============================================")
return format_outputs
class ScenicCall:
def __init__(self, main, flags):
self.main = main
self.FLAGS = flags
def __call__(self):
return self.run()
def run(self):
# Provide access to --jax_backend_target and --jax_xla_backend flags.
jax.config.config_with_absl()
run = (functools.partial(self._run_main, main=self.main))
return run(self.main)
def _run_main(self, argv, *, main):
"""Runs the `main` method after some initial setup."""
del argv
# Hide any GPUs form TensorFlow. Otherwise, TF might reserve memory and make
# it unavailable to JAX.
tf.config.experimental.set_visible_devices([], 'GPU')
# Enable wrapping of all module calls in a named_call for easier profiling:
nn.enable_named_call()
logging.info('JAX host: %d / %d', jax.process_index(), jax.process_count())
logging.info('JAX devices: %r', jax.devices())
# Add a note so that we can tell which task is which JAX host.
# (task 0 is not guaranteed to be the host 0)
platform.work_unit().set_task_status(
f'host_id: {jax.process_index()}, host_count: {jax.process_count()}')
if jax.process_index() == 0:
platform.work_unit().create_artifact(platform.ArtifactType.DIRECTORY,
self.FLAGS.workdir, 'Workdir')
self.FLAGS.config.dataset_configs.base_dir = self.FLAGS.data_dir
rng = jax.random.PRNGKey(self.FLAGS.config.rng_seed)
logging.info('RNG: %s', rng)
writer = metric_writers.create_default_writer(
self.FLAGS.workdir, just_logging=jax.process_index() > 0, asynchronous=True)
return main(rng=rng, config=self.FLAGS.config, workdir=self.FLAGS.workdir, writer=writer)
|