-
Notifications
You must be signed in to change notification settings - Fork 8
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
[feat] Add support with TFRA Dynamic Embedding Horovod Training.
Adding script also support test TF Embedding when use Horovod training.
- Loading branch information
Showing
3 changed files
with
301 additions
and
16 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,283 @@ | ||
| # Copyright 2020 The TensorFlow Authors. All Rights Reserved. | ||
| # | ||
| # 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. | ||
| # ============================================================================== | ||
| """unit tests of dynamic embedding optimizer ops | ||
| """ | ||
| from __future__ import absolute_import | ||
| from __future__ import division | ||
| from __future__ import print_function | ||
|
|
||
| import itertools | ||
| import os | ||
| import shutil | ||
|
|
||
| from absl import app, flags | ||
| import tensorflow as tf | ||
|
|
||
| try: | ||
| from tensorflow_recommenders_addons import dynamic_embedding as de | ||
| except: | ||
| pass | ||
|
|
||
| from tensorflow.core.protobuf import config_pb2 | ||
| from tensorflow.python.framework import dtypes | ||
| from tensorflow.python.framework import test_util | ||
| from tensorflow.python.framework.errors_impl import NotFoundError | ||
| from tensorflow.python.ops import math_ops | ||
| from tensorflow.python.platform import test | ||
| from tensorflow.python.training import monitored_session | ||
| from tensorflow.python.training import training_util | ||
| try: | ||
| from tensorflow.keras.optimizers.legacy import Adam | ||
| except: | ||
| from tensorflow.keras.optimizers import Adam | ||
|
|
||
| import deepray | ||
|
|
||
| FLAGS = flags.FLAGS | ||
| FLAGS.use_horovod = True | ||
| FLAGS.keras_use_ctl = True | ||
| FLAGS.use_dynamic_embedding = True | ||
| FLAGS.epochs = 1 | ||
| FLAGS.batch_size = 8 | ||
| FLAGS.steps_per_summary = 2 | ||
| FLAGS.save_checkpoint_steps = 2 | ||
| FLAGS.stop_steps = 10 | ||
|
|
||
| from deepray.core.base_trainer import Trainer | ||
| from deepray.layers.dynamic_embedding import DistributedDynamicEmbedding, DynamicEmbeddingOption | ||
| from deepray.utils.export import export_to_savedmodel | ||
|
|
||
| default_config = config_pb2.ConfigProto(allow_soft_placement=True, gpu_options=config_pb2.GPUOptions(allow_growth=True)) | ||
|
|
||
|
|
||
| def get_emb_sequential_model(emb_t, *args, **kwargs): | ||
| l0 = tf.keras.layers.InputLayer(input_shape=(None,), dtype=dtypes.int64) | ||
| l1 = emb_t(*args, **kwargs) | ||
| l2 = tf.keras.layers.Dense(8, 'relu', kernel_initializer='zeros') | ||
| l3 = tf.keras.layers.Dense(1, 'sigmoid', kernel_initializer='zeros') | ||
| if emb_t == tf.keras.layers.Embedding: | ||
| model = tf.keras.Sequential([l0, l1, l2, l3]) | ||
| elif emb_t == DistributedDynamicEmbedding: | ||
| model = tf.keras.Sequential([l0, l1, l2, l3]) | ||
| else: | ||
| raise TypeError('Unsupported embedding layer {}'.format(emb_t)) | ||
| return model | ||
|
|
||
|
|
||
| class HorovodTest(test.TestCase): | ||
|
|
||
| @test_util.deprecated_graph_mode_only | ||
| def test_adam_minimize_trainable(self): | ||
| tf.keras.backend.clear_session() | ||
| keras_base_opt = Adam(1.0) | ||
| keras_test_opt = Adam(1.0) | ||
| self.common_minimize_trainable_v2(keras_base_opt, keras_test_opt, name="keras_adam") | ||
|
|
||
| @test_util.run_all_in_graph_and_eager_modes | ||
| def test_all_to_all_embedding_trainable(self): | ||
| # TODO: Resolve the conflict arising from the 'save' function incompatibility with TensorFlow 2.11. | ||
| if (tf.__version__ == "2.11.0" or tf.__version__ == "2.11.1"): | ||
| self.skipTest("The save function doesn't work with TF 2.11, skip the test.") | ||
| keras_base_opt = Adam(1.0) | ||
| keras_test_opt = Adam(1.0) | ||
| self.common_all_to_all_embedding_trainable_v2(keras_base_opt, keras_test_opt, name="keras_adam") | ||
|
|
||
| def common_minimize_trainable_v2(self, base_opt, test_opt, name): | ||
| try: | ||
| import horovod.tensorflow as hvd | ||
| except (NotFoundError): | ||
| self.skipTest("Skip the test for horovod because it's not available.") | ||
| try: | ||
| from tensorflow_recommenders_addons import dynamic_embedding as de | ||
| except (NotFoundError): | ||
| self.skipTest("Skip the test for TFRA DE because it's not available.") | ||
|
|
||
| tf.config.set_soft_device_placement(True) | ||
|
|
||
| hvd.init() | ||
|
|
||
| # These cases need 2 GPUs at least if available. | ||
| logical_devices = tf.config.list_logical_devices('GPU') | ||
| _device = "GPU" if len(logical_devices) >= hvd.size() else "CPU" | ||
| _device_id = hvd.local_rank() if _device == "GPU" and len(logical_devices) >= 2 else 0 | ||
|
|
||
| if _device == "GPU": | ||
| os.environ["CUDA_VISIBLE_DEVICES"] = str(_device_id) | ||
|
|
||
| base_opt = de.DynamicEmbeddingOptimizer(base_opt, synchronous=True) | ||
| for dtype, run_step, dim in itertools.product([dtypes.float32], [1], [10]): | ||
| print("device=", "/{}:{}".format(_device, _device_id)) | ||
| with tf.device("/{}:{}".format(_device, _device_id)): | ||
| x = tf.random.uniform(shape=[32, dim]) | ||
| y = tf.zeros([32, 1]) | ||
|
|
||
| with tf.GradientTape() as tape_base: | ||
| base_weight = tf.compat.v1.get_variable(name="base_weights", initializer=tf.ones([10, 1])) | ||
|
|
||
| base_logits = tf.nn.relu(math_ops.matmul(x, base_weight)) | ||
| base_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=base_logits) | ||
|
|
||
| # grad_base = tape_base.gradient(base_loss, base_weight) | ||
| # base_opt. | ||
| base_opt_op = base_opt.minimize(base_loss, var_list=[base_weight], tape=tape_base) | ||
|
|
||
| with tf.GradientTape() as tape_test: | ||
| test_weight = tf.compat.v1.get_variable(name="test_weights", initializer=tf.ones([10, 1])) | ||
|
|
||
| test_logits = tf.nn.relu(math_ops.matmul(x, test_weight)) | ||
| test_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=test_logits) | ||
|
|
||
| grads_and_vars = test_opt._compute_gradients(test_loss, var_list=[test_weight], tape=tape_test) | ||
| var_list = [] | ||
| aggregated_grad = [] | ||
| for grad, var in grads_and_vars: | ||
| var_list.append(var) | ||
| aggregated_grad.append(hvd.allreduce(grad, op=hvd.Sum)) | ||
| aggregated_grads_and_vars = zip(aggregated_grad, var_list) | ||
| test_opt_op = test_opt.apply_gradients(aggregated_grads_and_vars) | ||
|
|
||
| with monitored_session.MonitoredTrainingSession(is_chief=True, config=default_config) as sess: | ||
| for _ in range(run_step): | ||
| sess.run(base_opt_op) | ||
| sess.run(test_opt_op) | ||
|
|
||
| self.assertAllCloseAccordingToType( | ||
| sess.run(base_weight), | ||
| sess.run(test_weight), | ||
| msg="Cond:{},{},{}".format(dtype, run_step, dim), | ||
| ) | ||
|
|
||
| def common_all_to_all_embedding_trainable_v2(self, base_opt, test_opt, name): | ||
| try: | ||
| import horovod.tensorflow as hvd | ||
| except (NotFoundError): | ||
| self.skipTest("Skip the test for horovod because it's not available.") | ||
| try: | ||
| from tensorflow_recommenders_addons import dynamic_embedding as de | ||
| except (NotFoundError): | ||
| self.skipTest("Skip the test for TFRA DE because it's not available.") | ||
|
|
||
| tf.config.set_soft_device_placement(True) | ||
|
|
||
| hvd.init() | ||
|
|
||
| # These cases need 2 GPUs at least if available. | ||
| logical_devices = tf.config.list_logical_devices('GPU') | ||
| _device = "GPU" if len(logical_devices) >= hvd.size() else "CPU" | ||
| _device_id = hvd.local_rank() if _device == "GPU" and len(logical_devices) >= 2 else 0 | ||
|
|
||
| if _device == "GPU": | ||
| os.environ["CUDA_VISIBLE_DEVICES"] = str(_device_id) | ||
| _de_device = 'HBM' | ||
| else: | ||
| _de_device = 'DRAM' | ||
|
|
||
| base_opt = de.DynamicEmbeddingOptimizer(base_opt, synchronous=True) | ||
| test_opt = test_opt | ||
| init = tf.keras.initializers.Zeros() | ||
| batch_size = FLAGS.batch_size | ||
| for dtype, run_step, dim in itertools.product([dtypes.float32], [FLAGS.stop_steps], [10]): | ||
| print("device=", "/{}:{}".format(_device, _device_id)) | ||
| with tf.device("/{}:{}".format(_device, _device_id)): | ||
| total_data_num = batch_size * run_step | ||
| x = math_ops.range(0, total_data_num, dtype=dtypes.int64) | ||
| x = tf.reshape(x, (-1, 1)) | ||
| y = tf.zeros((total_data_num, 1), dtype=dtypes.float32) | ||
| train_input_fn = tf.data.Dataset.from_tensor_slices((x, y)).cache().repeat() | ||
| train_input_fn = train_input_fn.batch(batch_size) | ||
|
|
||
| base_model = get_emb_sequential_model( | ||
| DistributedDynamicEmbedding, | ||
| embedding_dim=dim, | ||
| key_dtype=tf.int64, | ||
| value_dtype=dtype, | ||
| initializer=init, | ||
| de_option=DynamicEmbeddingOption(device=_de_device), | ||
| name='all2all_emb' | ||
| ) | ||
| test_model = get_emb_sequential_model( | ||
| tf.keras.layers.Embedding, | ||
| input_dim=batch_size * run_step, | ||
| output_dim=dim, | ||
| embeddings_initializer=init, | ||
| name='tf_emb' | ||
| ) | ||
| loss_func = tf.keras.losses.MeanSquaredError(reduction=tf.keras.losses.Reduction.SUM) | ||
| base_trainer = Trainer(model=base_model, optimizer=base_opt, loss=loss_func) | ||
| test_trainer = Trainer(model=test_model, optimizer=test_opt, loss=loss_func) | ||
|
|
||
| hvd_model_dir = "/tmp/hvd_save_restore" + str(hvd.size()) + str(run_step) + str( | ||
| dim | ||
| ) # All ranks should share same save directory | ||
| tf_model_dir = "/tmp/tf_save_restore" + str(hvd.size()) + str(run_step) + str( | ||
| dim | ||
| ) # All ranks should share same save directory | ||
|
|
||
| FLAGS.model_dir = tf_model_dir | ||
| test_trainer.fit(train_input=train_input_fn,) | ||
|
|
||
| FLAGS.model_dir = hvd_model_dir | ||
| base_trainer.fit(train_input=train_input_fn,) | ||
|
|
||
| self.assertAllCloseAccordingToType( | ||
| base_model.layers[1].weights[0], | ||
| test_model.layers[1].weights[0], | ||
| msg="Cond:{},{},{}".format(dtype, run_step, dim), | ||
| ) | ||
|
|
||
| self.assertAllCloseAccordingToType( | ||
| base_model.layers[2].weights[0], | ||
| test_model.layers[2].weights[0], | ||
| msg="Cond:{},{},{}".format(dtype, run_step, dim), | ||
| ) | ||
|
|
||
| a2aemb_size = base_model.layers[0].emb.params.size() | ||
|
|
||
| export_to_savedmodel(base_model) | ||
|
|
||
| del base_model | ||
| del base_trainer | ||
| del base_opt | ||
| tf.keras.backend.clear_session() | ||
| tf.compat.v1.reset_default_graph() | ||
| new_base_model = get_emb_sequential_model( | ||
| DistributedDynamicEmbedding, | ||
| embedding_dim=dim, | ||
| key_dtype=tf.int64, | ||
| value_dtype=dtype, | ||
| initializer=init, | ||
| de_option=DynamicEmbeddingOption(device=_de_device,), | ||
| name='all2all_emb' | ||
| ) | ||
| new_base_opt = de.DynamicEmbeddingOptimizer(Adam(1.0), synchronous=True) | ||
| FLAGS.init_checkpoint = [os.path.join(FLAGS.model_dir, 'ckpt_main_model')] | ||
| FLAGS.stop_steps = 0 | ||
| new_base_trainer = Trainer(model=new_base_model, optimizer=new_base_opt, loss=loss_func) | ||
| new_base_trainer.fit(train_input=train_input_fn,) | ||
|
|
||
| new_a2aemb_size = new_base_model.layers[0].emb.params.size() | ||
| self.assertEqual(a2aemb_size, new_a2aemb_size) | ||
|
|
||
| hvd.join() # Sync for avoiding files conflict | ||
| tf.keras.backend.clear_session() | ||
| tf.compat.v1.reset_default_graph() | ||
| new_base_model.load_weights(FLAGS.model_dir + '/export_main/variables/variables') | ||
| new_a2aemb_size = new_base_model.layers[0].emb.params.size() | ||
| self.assertEqual(a2aemb_size, new_a2aemb_size) | ||
| hvd.join() # Sync for avoiding files conflict | ||
|
|
||
|
|
||
| if __name__ == "__main__": | ||
| app.run(test.main()) |