File size: 12,358 Bytes
a560c26
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
# coding=utf-8
# Copyright 2023 The Google Research Authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

"""The main model training loop."""

import functools
import os
import time
from typing import Dict, Iterable, Mapping, Optional, Tuple, Type, Union

from absl import logging
from clu import checkpoint
from clu import metric_writers
from clu import metrics
from clu import parameter_overview
from clu import periodic_actions
import flax
from flax import linen as nn

import jax
import jax.numpy as jnp
import ml_collections
import numpy as np
import optax

from scenic.train_lib import lr_schedules
from scenic.train_lib import optimizers

import tensorflow as tf

from invariant_slot_attention.lib import evaluator
from invariant_slot_attention.lib import input_pipeline
from invariant_slot_attention.lib import losses
from invariant_slot_attention.lib import utils

Array = jnp.ndarray
ArrayTree = Union[Array, Iterable["ArrayTree"], Mapping[str, "ArrayTree"]]  # pytype: disable=not-supported-yet
PRNGKey = Array


def train_step(
    model,
    tx,
    rng,
    step,
    state_vars,
    opt_state,
    params,
    batch,
    loss_fn,
    train_metrics_cls,
    predicted_max_num_instances,
    ground_truth_max_num_instances,
    conditioning_key = None,
    ):
  """Perform a single training step.

  Args:
    model: Model used in training step.
    tx: The optimizer to use to minimize loss_fn.
    rng: Random number key
    step: Which training step we are on.
    state_vars: Accessory variables.
    opt_state: The state of the optimizer.
    params: The current parameters to be updated.
    batch: Training inputs for this step.
    loss_fn: Loss function that takes model predictions and a batch of data.
    train_metrics_cls: The metrics collection for computing training metrics.
    predicted_max_num_instances: Maximum number of instances in prediction.
    ground_truth_max_num_instances: Maximum number of instances in ground truth,
      including background (which counts as a separate instance).
    conditioning_key: Optional string. If provided, defines the batch key to be
      used as conditioning signal for the model. Otherwise this is inferred from
      the available keys in the batch.

  Returns:
    Tuple of the updated opt, state_vars, new random number key,
      metrics update, and step + 1. Note that some of this info is stored in
      TrainState, but here it is unpacked.
  """

  # Split PRNGKey and bind to host / device.
  new_rng, rng = jax.random.split(rng)
  rng = jax.random.fold_in(rng, jax.host_id())
  rng = jax.random.fold_in(rng, jax.lax.axis_index("batch"))
  init_rng, dropout_rng = jax.random.split(rng, 2)

  mutable_var_keys = list(state_vars.keys()) + ["intermediates"]

  conditioning = batch[conditioning_key] if conditioning_key else None

  def train_loss_fn(params, state_vars):
    preds, mutable_vars = model.apply(
        {"params": params, **state_vars}, video=batch["video"],
        conditioning=conditioning, mutable=mutable_var_keys,
        rngs={"state_init": init_rng, "dropout": dropout_rng}, train=True,
        padding_mask=batch.get("padding_mask"))
    # Filter intermediates, as we do not want to store them in the TrainState.
    state_vars = utils.filter_key_from_frozen_dict(
        mutable_vars, key="intermediates")
    loss, loss_aux = loss_fn(preds, batch)
    return loss, (state_vars, preds, loss_aux)

  grad_fn = jax.value_and_grad(train_loss_fn, has_aux=True)
  (loss, (state_vars, preds, loss_aux)), grad = grad_fn(params, state_vars)

  # Compute average gradient across multiple workers.
  grad = jax.lax.pmean(grad, axis_name="batch")

  updates, new_opt_state = tx.update(grad, opt_state, params)
  new_params = optax.apply_updates(params, updates)

  # Compute metrics.
  metrics_update = train_metrics_cls.gather_from_model_output(
      loss=loss,
      **loss_aux,
      predicted_segmentations=utils.remove_singleton_dim(
          preds["outputs"].get("segmentations")),  # pytype: disable=attribute-error
      ground_truth_segmentations=batch.get("segmentations"),
      predicted_max_num_instances=predicted_max_num_instances,
      ground_truth_max_num_instances=ground_truth_max_num_instances,
      padding_mask=batch.get("padding_mask"),
      mask=batch.get("mask"))
  return (
      new_opt_state, new_params, state_vars, new_rng, metrics_update, step + 1)


def train_and_evaluate(config,
                       workdir):
  """Runs a training and evaluation loop.

  Args:
    config: Configuration to use.
    workdir: Working directory for checkpoints and TF summaries. If this
      contains checkpoint training will be resumed from the latest checkpoint.
  """
  rng = jax.random.PRNGKey(config.seed)

  tf.io.gfile.makedirs(workdir)

  # Input pipeline.
  rng, data_rng = jax.random.split(rng)
  # Make sure each host uses a different RNG for the training data.
  if config.get("seed_data", True):  # Default to seeding data if not specified.
    data_rng = jax.random.fold_in(data_rng, jax.host_id())
  else:
    data_rng = None
  train_ds, eval_ds = input_pipeline.create_datasets(config, data_rng)
  train_iter = iter(train_ds)  # pytype: disable=wrong-arg-types

  # Initialize model
  model = utils.build_model_from_config(config.model)

  # Construct TrainMetrics and EvalMetrics, metrics collections.
  train_metrics_cls = utils.make_metrics_collection("TrainMetrics",
                                                    config.train_metrics_spec)
  eval_metrics_cls = utils.make_metrics_collection("EvalMetrics",
                                                   config.eval_metrics_spec)

  def init_model(rng):
    rng, init_rng, model_rng, dropout_rng = jax.random.split(rng, num=4)

    init_conditioning = None
    if config.get("conditioning_key"):
      init_conditioning = jnp.ones(
          [1] + list(train_ds.element_spec[config.conditioning_key].shape)[2:],
          jnp.int32)
    init_inputs = jnp.ones(
        [1] + list(train_ds.element_spec["video"].shape)[2:],
        jnp.float32)
    initial_vars = model.init(
        {"params": model_rng, "state_init": init_rng, "dropout": dropout_rng},
        video=init_inputs, conditioning=init_conditioning,
        padding_mask=jnp.ones(init_inputs.shape[:-1], jnp.int32))

    # Split into state variables (e.g. for batchnorm stats) and model params.
    # Note that `pop()` on a FrozenDict performs a deep copy.
    state_vars, initial_params = initial_vars.pop("params")  # pytype: disable=attribute-error

    # Filter out intermediates (we don't want to store these in the TrainState).
    state_vars = utils.filter_key_from_frozen_dict(
        state_vars, key="intermediates")
    return state_vars, initial_params

  state_vars, initial_params = init_model(rng)
  parameter_overview.log_parameter_overview(initial_params)  # pytype: disable=wrong-arg-types

  learning_rate_fn = lr_schedules.get_learning_rate_fn(config)
  tx = optimizers.get_optimizer(
      config.optimizer_configs, learning_rate_fn, params=initial_params)

  opt_state = tx.init(initial_params)

  state = utils.TrainState(
      step=1, opt_state=opt_state, params=initial_params, rng=rng,
      variables=state_vars)

  loss_fn = functools.partial(
      losses.compute_full_loss, loss_config=config.losses)

  checkpoint_dir = os.path.join(workdir, "checkpoints")
  ckpt = checkpoint.MultihostCheckpoint(checkpoint_dir)
  state = ckpt.restore_or_initialize(state)
  initial_step = int(state.step)

  # Replicate our parameters.
  state = flax.jax_utils.replicate(state, devices=jax.local_devices())
  del rng  # rng is stored in the state.

  # Only write metrics on host 0, write to logs on all other hosts.
  writer = metric_writers.create_default_writer(
      workdir, just_logging=jax.host_id() > 0)
  writer.write_hparams(utils.prepare_dict_for_logging(config.to_dict()))

  logging.info("Starting training loop at step %d.", initial_step)
  report_progress = periodic_actions.ReportProgress(
      num_train_steps=config.num_train_steps, writer=writer)
  if jax.process_index() == 0:
    profiler = periodic_actions.Profile(num_profile_steps=5, logdir=workdir)
  p_train_step = jax.pmap(
      train_step,
      axis_name="batch",
      donate_argnums=(2, 3, 4, 5, 6, 7),
      static_broadcasted_argnums=(0, 1, 8, 9, 10, 11, 12))

  train_metrics = None
  with metric_writers.ensure_flushes(writer):
    if config.num_train_steps == 0:
      with report_progress.timed("eval"):
        evaluate(model, state, eval_ds, loss_fn, eval_metrics_cls, config,
                 writer, step=0)
      with report_progress.timed("checkpoint"):
        ckpt.save(flax.jax_utils.unreplicate(state))
      return

    for step in range(initial_step, config.num_train_steps + 1):
      # `step` is a Python integer. `state.step` is JAX integer on GPU/TPU.
      is_last_step = step == config.num_train_steps

      with jax.profiler.StepTraceAnnotation("train", step_num=step):
        batch = jax.tree_map(np.asarray, next(train_iter))
        (opt_state, params, state_vars, rng, metrics_update, p_step
         ) = p_train_step(
             model, tx, state.rng, state.step, state.variables,
             state.opt_state, state.params, batch, loss_fn,
             train_metrics_cls,
             config.num_slots,
             config.max_instances + 1,  # Incl. background.
             config.get("conditioning_key"))

        state = state.replace(  # pytype: disable=attribute-error
            opt_state=opt_state,
            params=params,
            step=p_step,
            variables=state_vars,
            rng=rng,
        )

        metric_update = flax.jax_utils.unreplicate(metrics_update)
        train_metrics = (
            metric_update
            if train_metrics is None else train_metrics.merge(metric_update))

      # Quick indication that training is happening.
      logging.log_first_n(logging.INFO, "Finished training step %d.", 5, step)
      report_progress(step, time.time())

      if jax.process_index() == 0:
        profiler(step)

      if step % config.log_loss_every_steps == 0 or is_last_step:
        metrics_res = train_metrics.compute()
        writer.write_scalars(step, jax.tree_map(np.array, metrics_res))
        train_metrics = None

      if step % config.eval_every_steps == 0 or is_last_step:
        with report_progress.timed("eval"):
          evaluate(model, state, eval_ds, loss_fn, eval_metrics_cls,
                   config, writer, step=step)

      if step % config.checkpoint_every_steps == 0 or is_last_step:
        with report_progress.timed("checkpoint"):
          ckpt.save(flax.jax_utils.unreplicate(state))


def evaluate(model, state, eval_ds, loss_fn_eval, eval_metrics_cls, config,
             writer, step):
  """Evaluate the model."""
  eval_metrics, eval_batch, eval_preds = evaluator.evaluate(
      model,
      state,
      eval_ds,
      loss_fn_eval,
      eval_metrics_cls,
      predicted_max_num_instances=config.num_slots,
      ground_truth_max_num_instances=config.max_instances + 1,  # Incl. bg.
      slice_size=config.get("eval_slice_size"),
      slice_keys=config.get("eval_slice_keys"),
      conditioning_key=config.get("conditioning_key"),
      remove_from_predictions=config.get("remove_from_predictions"),
      metrics_on_cpu=config.get("metrics_on_cpu", False))

  metrics_res = eval_metrics.compute()
  writer.write_scalars(
      step, jax.tree_map(np.array, utils.flatten_named_dicttree(metrics_res)))
  writer.write_images(
      step,
      jax.tree_map(
          np.array,
          utils.prepare_images_for_logging(
              config,
              eval_batch,
              eval_preds,
              n_samples=config.get("n_samples", 5),
              n_frames=config.get("n_frames", 1),
              min_n_colors=config.get("logging_min_n_colors", 1))))