Distributed keras/tensorflow training on multiple GPUs per host

From #13

import tensorflow as tf
#import keras.backend as k
#https://github.com/microsoft/tensorflow-directml/issues/352

# https://www.tensorflow.org/guide/distributed_training
#
# https://www.tensorflow.org/tutorials/distribute/keras
# https://keras.io/guides/distributed_training/
#strategy = tf.distribute.MirroredStrategy()
#print('Number of devices: {}'.format(strategy.num_replicas_in_sync))

#NUM_GPUS = 2
#strategy = tf.contrib.distribute.MirroredStrategy()#num_gpus=NUM_GPUS)
# not working
strategy = tf.distribute.MirroredStrategy(devices=["/gpu:0", "/gpu:1"])
#WARNING:tensorflow:Some requested devices in `tf.distribute.Strategy` are not visible to TensorFlow: /replica:0/task:0/device:GPU:1,/replica:0/task:0/device:GPU:0
#Number of devices: 2

#central_storage_strategy = tf.distribute.experimental.CentralStorageStrategy()
#strategy = tf.distribute.MultiWorkerMirroredStrategy() # not in tf 1.5
#print("mirrored_strategy: ",mirrored_strategy)
#strategy = tf.distribute.OneDeviceStrategy(device="/gpu:1")
#mirrored_strategy = tf.distribute.MirroredStrategy(devices=["/gpu:0","/gpu:1"],cross_device_ops=tf.contrib.distribute.AllReduceCrossDeviceOps(all_reduce_alg="hierarchical_copy"))
#mirrored_strategy = tf.distribute.MirroredStrategy(devices= ["/gpu:0","/gpu:1"],cross_device_ops=tf.distribute.HierarchicalCopyAllReduce())

#print('Number of devices: {}'.format(strategy.num_replicas_in_sync))

# https://learn.microsoft.com/en-us/windows/ai/directml/gpu-faq
a = tf.constant([1.])
b = tf.constant([2.])
c = tf.add(a, b)

gpu_config = tf.GPUOptions()
gpu_config.visible_device_list = "1"#"0,1"
gpu_config.allow_growth=True

#session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_config))
#print(session.run(c))
#tensorflow.python.framework.errors_impl.AlreadyExistsError: TensorFlow device (DML:0) is being mapped to multiple DML devices (0 now, and 1 previously), which is not supported. This may be the result of providing different GPU configurations (ConfigProto.gpu_options, for example different visible_device_list) when creating multiple Sessions in the same process. This is not  currently supported, see https://github.com/tensorflow/tensorflow/issues/19083
#from keras import backend as K
#K.set_session(session)

cifar = tf.keras.datasets.cifar100
(x_train, y_train), (x_test, y_test) = cifar.load_data()

#with strategy.scope():

# https://www.tensorflow.org/api_docs/python/tf/keras/applications/resnet50/ResNet50
# https://keras.io/api/models/model/
model = tf.keras.applications.ResNet50(
    include_top=True,
    weights=None,
    input_shape=(32, 32, 3),
    classes=100,)
#from tensorflow.python.keras import backend as K
#tf.keras.set_session(session)

loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
# https://keras.io/api/models/model_training_apis/
model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])

model.fit(x_train, y_train, epochs=100, batch_size=7168)

https://saturncloud.io/blog/how-to-do-multigpu-training-with-keras/ around multi_gpu_model

On Dual RTX-4090 - full parallelization

import tensorflow as tf
#import keras
#from keras.utils import multi_gpu_model
#import keras.backend as k
#https://github.com/microsoft/tensorflow-directml/issues/352

# https://www.tensorflow.org/guide/distributed_training
#
# https://www.tensorflow.org/tutorials/distribute/keras
# https://keras.io/guides/distributed_training/
#strategy = tf.distribute.MirroredStrategy()
#print('Number of devices: {}'.format(strategy.num_replicas_in_sync))

#NUM_GPUS = 2
#strategy = tf.contrib.distribute.MirroredStrategy()#num_gpus=NUM_GPUS)
strategy = tf.distribute.MirroredStrategy(devices=["/gpu:0", "/gpu:1"])
# https://learn.microsoft.com/en-us/windows/ai/directml/gpu-faq
cifar = tf.keras.datasets.cifar100
(x_train, y_train), (x_test, y_test) = cifar.load_data()

with strategy.scope():
# https://www.tensorflow.org/api_docs/python/tf/keras/applications/resnet50/ResNet50
# https://keras.io/api/models/model/
  parallel_model = tf.keras.applications.ResNet50(
#model = tf.keras.applications.ResNet50(
    include_top=True,
    weights=None,
    input_shape=(32, 32, 3),
    classes=100,)
# https://saturncloud.io/blog/how-to-do-multigpu-training-with-keras/  
  #parallel_model = multi_gpu_model(model, gpus=2)
  loss_fn = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=False)
# https://keras.io/api/models/model_training_apis/
  parallel_model.compile(optimizer="adam", loss=loss_fn, metrics=["accuracy"])
parallel_model.fit(x_train, y_train, epochs=10, batch_size=256)#5120)#7168)#7168)

Your kernel may have been built without NUMA support.
2023-10-04 02:13:36.322066: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1886] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 21286 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 4090, pci bus id: 0000:01:00.0, compute capability: 8.9
2023-10-04 02:13:36.322333: I tensorflow/compiler/xla/stream_executor/cuda/cuda_gpu_executor.cc:880] could not open file to read NUMA node: /sys/bus/pci/devices/0000:02:00.0/numa_node
Your kernel may have been built without NUMA support.
2023-10-04 02:13:36.322355: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1886] Created device /job:localhost/replica:0/task:0/device:GPU:1 with 21286 MB memory:  -> device: 1, name: NVIDIA GeForce RTX 4090, pci bus id: 0000:02:00.0, compute capability: 8.9
Downloading data from https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz
169001437/169001437 [==============================] - 4s 0us/step
Epoch 1/10
2023-10-04 02:14:01.564898: I tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:442] Loaded cuDNN version 8600
2023-10-04 02:14:01.961456: I tensorflow/compiler/xla/stream_executor/cuda/cuda_dnn.cc:442] Loaded cuDNN version 8600
2023-10-04 02:14:04.719269: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7fa5c69fab80 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
2023-10-04 02:14:04.719297: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): NVIDIA GeForce RTX 4090, Compute Capability 8.9
2023-10-04 02:14:04.719300: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (1): NVIDIA GeForce RTX 4090, Compute Capability 8.9
2023-10-04 02:14:04.722981: I tensorflow/compiler/mlir/tensorflow/utils/dump_mlir_util.cc:269] disabling MLIR crash reproducer, set env var `MLIR_CRASH_REPRODUCER_DIRECTORY` to enable.
2023-10-04 02:14:04.778031: I ./tensorflow/compiler/jit/device_compiler.h:186] Compiled cluster using XLA!  This line is logged at most once for the lifetime of the process.
196/196 [==============================] - 38s 57ms/step - loss: 4.3436 - accuracy: 0.0889
Epoch 2/10
196/196 [==============================] - 10s 52ms/step - loss: 3.6832 - accuracy: 0.1765
Epoch 3/10
196/196 [==============================] - 10s 51ms/step - loss: 3.8586 - accuracy: 0.1447
Epoch 4/10
196/196 [==============================] - 10s 52ms/step - loss: 3.4894 - accuracy: 0.1989
Epoch 5/10
196/196 [==============================] - 10s 52ms/step - loss: 3.2204 - accuracy: 0.2489
Epoch 6/10
196/196 [==============================] - 10s 51ms/step - loss: 3.0356 - accuracy: 0.2802
Epoch 7/10
196/196 [==============================] - 10s 52ms/step - loss: 2.9648 - accuracy: 0.2936
Epoch 8/10
196/196 [==============================] - 10s 52ms/step - loss: 2.6994 - accuracy: 0.3399
Epoch 9/10
196/196 [==============================] - 10s 52ms/step - loss: 2.4836 - accuracy: 0.3792
Epoch 10/10
196/196 [==============================] - 10s 52ms/step - loss: 2.3146 - accuracy: 0.4120

ObrienlabsDev / blog

Distributed keras/tensorflow training on multiple GPUs per host #15