Python Distributed Training of Neural Networks

Introduction

PyDTNN is a lightweight library developed at Universitat Jaume I (Spain) for distributed and federated deep learning training and inference of convolutional and transformer-based neural networks, intended as an initial starting point for interacting with training and inference processes. PyDTNN prioritizes simplicity over peak performance, offering an approachable user interface that enables a gentle learning curve. To carry out training and inference, PyDTNN exploits inter-process parallelism (via MPI) and intra-process parallelism (via multithreading), leveraging the capabilities of multicore processors and GPUs at the node level. For this purpose, PyDTNN relies on mpi4py/pympi/NCCL for message passing, BLAS routines accessed through NumPy/Cython for multicore processors, and PyCUDA/cuDNN/cuBLAS for GPU acceleration.

Supported layers:

• Fully-connected
• Convolutional 2D
• Max pooling 2D
• Average pooling 2D
• Dropout
• Flatten
• Feed Forward
• Multi-head attention
• Batch normalization
• Encoder & Decoder (for transformer nets, e.g., Bert)
• Addition block (for residual nets, e.g., ResNet)
• Concatenation block (for channel concatenation-based nets, e.g., Inception, GoogleNet, DenseNet, etc.)

Supported datasets:

• MNIST: handwritten digit database. This dataset is included into the repository. Its binary version can be downloaded from: https://github.com/hpca-uji/PyDTNN
• CIFAR10: database of the 80 million tiny images dataset. This dataset is not included into the repository. Its binary version can be downloaded from: https://www.cs.toronto.edu/~kriz/cifar.html
• ImageNet: the most highly-used subset of ImageNet is the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) 2012-2017 image classification and localization dataset. This dataset spans 1000 object classes and contains 1,281, 167 training images, 50,000 validation images and 100,000 test images. This dataset is not included into the repository. It can be downloaded from: https://image-net.org/challenges/LSVRC/2012/2012-downloads.php
• ChestXray: the NIH Chest X-ray dataset consists of 100,000 de-identified images of chest x-rays. The images are in PNG format. It can be downloaded from: https://nihcc.app.box.com/v/ChestXray-NIHCC
• IWSLT: the IWSLT 2017 Multilingual Task addresses text translation, including zero-shot translation, with a single MT system across all directions including English, German, Dutch, Italian and Romanian. As unofficial task, conventional bilingual text translation is offered between English and Arabic, French, Japanese, Chinese, German and Korean. This dataset is included into the repository. Its binary version can be downloaded from: https://github.com/hpca-uji/PyDTNN
• And others via generic data loaders.

Installing PyDTNN from source

Download PyDTNN source code from its GitHub repository and enter the PyDTNN directory:

git clone https://github.com/hpca-uji/PyDTNN.git
cd PyDTNN

Then package itself must be installed:

pip install .

If you plan to modify the PyDTNN code, instead of using the previous line, you can install PyDTNN in editable mode (see CONTRIBUTING.md for more details):

pip install --config-settings editable_mode=compat -e .

Optionally, if you are going to use MPI, you should have installed the corresponding system libraries, and install the required Python packages with:

git submodule update --init ./vendor/net-queue
pip install ./vendor/net-queue

git submodule update --init ./vendor/pympi
pip install ./vendor/pympi

pip install .[mpi]

Optionally, if you are going to use CUDA, you should have installed the corresponding system libraries, and install the required Python packages with:

pip install nvidia-pyindex
pip install .[cuda]

Optionally, if you are going to use FHE, you should have installed the corresponding system libraries, and install the required Python packages with:

pip install .[fhe]

Optionally, if you are going to use MPI/TCP, you should enable the protocol with:

export PYMPI_PROTO=tcp

Optionally, if you are going to use MPI/gRPC, you should enable the protocol with:

export PYMPI_PROTO=grpc

Optionally, if you are going to use MPI/MQTT, you should have installed a MQTT broker server, you should enable the protocol with:

export PYMPI_PROTO=mqtt

Optionally, if you are going to use MPI/SSL, you should enable the transport with:

export PYMPI_SSL=yes
export PYMPI_SSL_KEY=comms/ssl/key.pem    # server private key
export PYMPI_SSL_CERT=comms/ssl/cert.pem  # server certificate

For more information on how to manage external dependencies see vendor/README.md.

Launcher options

The PyDTNN framework comes with a utility launcher called pydtnn-benchmark that supports the following options:

• Model parameters:
  • --model: Neural network model: simplemlp, simplecnn, alexnet, vgg11, vgg16, etc. Default: None.
  • --batch-size: Batch size per MPI rank. Default: None.
  • --global-batch-size: Batch size between all MPI ranks. Default: None.
  • --dtype: Datatype to use: float32, float64. Default: float32.
  • --num-epochs: Number of epochs to perform. Default: 1.
  • --steps-per-epoch: Trims the training data depending on the given number of steps per epoch. Default: 0, i.e., do not trim.
  • --evaluate: Evaluate the model before and after training the model. Default: False.
  • --evaluate-only: Only evaluate the model. Default: False.
  • --weights-and-bias-filename: Load weights and bias from file. Default: None.
  • --history-file: Filename to save training loss and metrics.
  • --shared-storage: If True ranks assume they share the file system. Default: True.
  • --model-sync-freq: Number of batches between model synchronization. The 0 value synchronizes gradients every batch. Positive values synchronizes gradients and weights every N batches. Negative values disables synchronization. Default: 0.
  • --model-sync-alg: Aggregation method used to synchronize models: avg, wavg or invwavg. Default: avg.
  • --model-sync-participation: Rank participation to synchronize models: all or avail2all. Default: all.
  • --model-sync-min-avail: Minimum ranks with data required to synchronize models. Default: 0.
  • --initial-model-sync: Synchronize models on training start. Default: True.
  • --final-model-sync: Synchronize models on training end. Default: True.
  • --tensor-format: Data format to be used: NHWC or NCHW. Optionally, the AUTO value sets NCHW when the option --enable-gpu is set and NHWC otherwise. Default: NHWC.
• Dataset parameters:
  • --dataset: Dataset to train: mnist, cifar10, synthetic, …. Default: None.
  • --dataset-path: Path to dataset folder.
  • --dataset-lang: Dataset language. Default: en.
  • --dataset-lang2: Dataset second language. Default: de.
  • --synthetic-train-samples: Number of synthetic train sample. Default: 1000.
  • --synthetic-test-samples: Number of synthetic train sample. Default: 100.
  • --synthetic-input-shape: Synthetic input shape (coma separated). Default: 3,32,32.
  • --synthetic-output-shape: Synthetic output shape (coma separated). Default: 10.
  • --dataset-percentage: Percentage of dataset that will be used. If it is 0: it is deactivated; if is is a value below 1 (and above 0): it will perform undersampling; and if is is a value above 1: it will perform oversampling. Default: 0.
  • --test-as-validation: Prevent making partitions on training data for training+validation data, use test data for validation. True if specified.
  • --validation-split: Split between training and validation data.
  • --augment-flip: Flip horizontally training images. Default: False.
  • --augment-flip-prob: Probability to flip training images. Default: 0.5.
  • --augment-crop: Crop training images. Default: False.
  • --augment-crop-size: Size to crop training images. Default: 16.
  • --augment-crop-prob: Probability to crop training images. Default: 0.5.
  • --validation-split: Split between training and validation data.
  • --transform-crop: Crop the images. True if specified.
  • --transform-crop-perc: Central crop of the images. Default: 0.875.
  • --transform-resize: Resize the images. True if specified.
  • --transform-resize-size: New size of the images. Default: 300.
  • --normalize: Normalize dataset. Default: False.
  • --normalize-offset: Offset samples by a value. Default: -0.45.
  • --normalize-scale: Scale samples by a value. Default: 3.75.
• Optimization parameters:
  • --enable-best-of: Enable the BestOf auto-tuner.
  • --enable-memory-cache: Enable the memory cache module to use persistent memory.
  • --enable-fused-bn-relu: Fuse BatchNormalization and Relu layers. True if specified.
  • --enable-fused-conv-relu: Fuse Conv2D and Relu layers. True if specified.
  • --enable-fused-conv-bn: Fuse Conv2D and BatchNormalization layers. True if specified.
  • --enable-fused-conv-bn-relu: Fuse Conv2D and BatchNormalization and Relu layers. Default: False.
• Convolution operation parameters:
  • --conv-variant:Select the standard 2D Convolutional module. Options:
    • i2c (default): Use the ConvI2C algorithm.
    • gemm: Use the ConvGemm algorithm.
    • winograd: Use the CondWinograd algorithm.
    • direct: Use the ConvDirect algorithm.
  • --conv-direct-method: Use ConvDirect module to realize convolutions in Conv2D layers. True if specified.
  • --conv-direct-methods-for-best-of: ConvDirect modules to compare in best_of option if specified.
• Optimizer parameters:
  • --optimizer: Optimizers: sgd, rmsprop, adam, nadam. Default: sgd.
  • --learning-rate: Learning rate. Default: 0.01.
  • --learning-rate-scaling: Scale learning rate in data parallelism: new_lr = lr/num_procs. True if specified.
  • --optimizer-momentum: Decay rate for sgd optimizer. Default: 0.9. optimizers. Default: 1e-8.
  • --optimizer-decay: Decay rate for optimizers. Default: 0.0.
  • --optimizer-nesterov: Whether to apply Nesterov momentum. Default: False.
  • --optimizer-beta1: Variable for adam, nadam optimizers. Default: 0.99.
  • --optimizer-beta2: Variable for adam, nadam optimizers. Default: 0.999.
  • --optimizer-epsilon: Variable for rmsprop, adam, nadam. Default: 1e-7.
  • --optimizer-rho: Variable for rmsprop optimizers. Default: 0.99.
  • --loss-func: Loss functions that is evaluated on each trained batch: categorical_cross_entropy, binary_cross_entropy or kl_divergence. Default categorical_cross_entropy.
  • --metrics: List of comma-separated metrics that are evaluated on each trained batch: categorical_accuracy, categorical_hinge, categorical_mse, categorical_mae, regression_mse, regression_mae, binary_confusion_matrix, multiclass_confusion_matrix, precision, recall, f1_score. Default: categorical_accuracy.
• Learning rate schedulers parameters:
  • --schedulers: List of comma-separated LR schedulers: warm_up, early_stopping, reduce_lr_on_plateau, reduce_lr_every_nepochs, model_checkpoint. Default: early_stopping,reduce_lr_on_plateau,model_checkpoint.
  • --warm-up-batches: Number of batches (ramp up) that the LR is scaled up from 0 until LR. Default: 5.
  • --early-stopping-metric: Loss metric monitored by early_stopping scheduler. Default: val_categorical_cross_entropy.
  • --early-stopping-patience: Number of epochs with no improvement after which training will be stopped. Default: 10.
  • --early-stopping-minimize: Whether to minimize the metric. If False, it will maximize. Default: True.
  • --reduce-lr-on-plateau-metric: Loss metric monitored by reduce_lr_on_plateau scheduler. Default: val_categorical_cross_entropy.
  • --reduce-lr-on-plateau-factor: Factor by which the learning rate will be reduced. new_lr = lr *factor. Default: 0.1.
  • --reduce-lr-on-plateau-patience: Number of epochs with no improvement after which LR will be reduced. Default: 5.
  • --reduce-lr-on-plateau-min-lr: Lower bound on the learning rate. Default: 0.
  • --reduce-lr-every-nepochs-factor: Factor by which the learning rate will be reduced. new_lr = lr*factor. Default: 0.1.
  • --reduce-lr-every-nepochs-nepochs: Number of epochs after which LR will be periodically reduced. Default: 5.
  • --reduce-lr-every-nepochs-min-lr: Lower bound on the learning rate. Default: 0.
  • --stop-at-loss-metric: Loss metric monitored by stop_at_loss scheduler. Default: val_accuracy.
  • --stop-at-loss-threshold: Metric threshold monitored by stop_at_loss scheduler. Default: 0.
  • --model-checkpoint-metric: Loss metric monitored by model_checkpoint scheduler. Default: val_categorical_cross_entropy
  • --model-checkpoint-save-freq: Frequency (in epochs) at which the model weights and bias will be saved by the model_checkpoint scheduler. Default: 2.
• Parallelization and other performance-related parameters:
  • --parallel: Data parallelization modes: sequential, data (MPI). Default: sequential.
  • --use-blocking-mpi: Enable blocking MPI primitives. Default: True.
  • --use-mpi-buffers: Enable the use of MPI buffers. Possible values: True (MPI operations by buffer), False (MPI operations by object) or None (auto-select the better option). Default: None.
  • --enable-gpu: Enable GPU, use cuDNN library. Default: False.
  • --enable-gpudirect: Enable GPU pinned memory for gradients when using a CUDA-aware MPI version. Default: False.
  • --enable-nccl: Enable the use of the NCCL library for collective communications on GPUs. This option can only be set with --enable-gpu. Default. False.
  • --enable-cudnn-auto-conv-alg: Let cuDNN to select the best performing convolution algorithm. Default: True.
• Encryption parameters:
  • --encryption: Encryption library: tenseal, openfhe, None. Default None.
  • --encryption-poly-degree: Encryption polynomial degree. 2 ^ value. Default: 13.
  • --encryption-global-scale: Encryption global scale. 2 ^ value. Default: 40.
  • --encryption-security-level: Encryption security level: 0 (Not set), 128, 192, 256. Default: 128.
• Tracing and profiling parameters:
  • --tracing: Obtain Simple/Extrae-based traces. Default: False.
  • --tracer-output: Output file to store the Simple/Extrae-based traces.
  • --tracer-pmlib-server: Address of PMlib tracer server. Default: 127.0.0.1.
  • --tracer-pmlib-port: Port of PMlib tracer server. Default: 6526.
  • --tracer-pmlib-device: Port of PMlib tracer device.
  • --profile: Obtain cProfile profiles. Default: False.

Example: distributed training of a CNN for the MNIST dataset

In this example, we train a simple CNN for the MNIST dataset using data parallelism and 12 MPI ranks each using 4 OpenMP threads:

$ export OMP_NUM_THREADS=4
$ mpirun -np 12 \
    pydtnn-benchmark \
      --model=simplecnn \
      --dataset=mnist \
      --dataset-path=datasets/mnist \
      --dataset-train-path=datasets/mnist \
      --dataset-test-path=datasets/mnist \
      --test-as-validation=False \
      --augment-flip=True \
      --batch-size=64 \
      --validation-split=0.2 \
      --num-epochs=50 \
      --evaluate=True \
      --optimizer=adam \
      --learning-rate=0.01 \
      --loss-func=categorical_cross_entropy \
      --schedulers=warm_up,reduce_lr_every_nepochs \
      --reduce-lr-every-nepochs-factor=0.5 \
      --reduce-lr-every-nepochs-nepochs=30 \
      --reduce-lr-every-nepochs-min-lr=0.001 \
      --early-stopping-metric=val_categorical_cross_entropy \
      --early-stopping-patience=20 \
      --parallel=sequential \
      --tracing=False \
      --profile=False \
      --enable-gpu=True \
      --dtype=float32


**** simplecnn model...
+-------+--------------------------+---------+---------------+-------------------+------------------------+
| Layer |           Type           | #Params | Output shape  |   Weights shape   |       Parameters       |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   0   |          Input           |    0    |  (1, 28, 28)  |                   |                        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   1   |          Conv2D          |   40    |  (4, 28, 28)  |   (4, 1, 3, 3)    |padd=(1,1), stride=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   2   |          Conv2D          |   148   |  (4, 28, 28)  |   (4, 4, 3, 3)    |padd=(1,1), stride=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   3   |        MaxPool2D         |    0    |  (4, 14, 14)  |      (2, 2)       |padd=(0,0), stride=(2,2)|
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   4   |         Flatten          |    0    |    (784,)     |                   |                        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   5   |            FC            | 100480  |    (128,)     |    (784, 128)     |                        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   6   |           Relu           |    0    |    (128,)     |                   |                        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   7   |         Dropout          |    0    |    (128,)     |                   |       rate=0.50        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   8   |            FC            |  1290   |     (10,)     |     (128, 10)     |                        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|   9   |         Softmax          |    0    |     (10,)     |                   |                        |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
|             Total parameters       101958    398.27 KBytes                                              |
+-------+--------------------------+---------+---------------+-------------------+------------------------+
**** Loading mnist dataset...
**** Parameters:
  model                          : simplecnn
  dataset                        : mnist
  dataset_train_path             : datasets/mnist
  dataset_test_path              : datasets/mnist
  test_as_validation             : False
  augment_flip                    : True
  augment_flip_prob               : 0.5
  augment_crop                    : False
  augment_crop_size               : 16
  augment_crop_prob               : 0.5
  batch_size                     : 64
  global_batch_size              : None
  validation_split               : 0.2
  steps_per_epoch                : 0
  num_epochs                     : 50
  evaluate                       : True
  weights_and_bias_filename      : None
  shared_storage                 : True
  history_file                   : None
  optimizer                      : adam
  learning_rate                  : 0.01
  learning_rate_scaling          : True
  momentum                       : 0.9
  decay                          : 0.0
  nesterov                       : False
  beta1                          : 0.99
  beta2                          : 0.999
  epsilon                        : 1e-07
  rho                            : 0.9
  loss_func                      : categorical_cross_entropy
  metrics                        : categorical_accuracy
  schedulers                  : warm_up,reduce_lr_every_nepochs
  warm_up_epochs                 : 5
  early_stopping_metric          : val_categorical_cross_entropy
  early_stopping_patience        : 20
  reduce_lr_on_plateau_metric    : val_categorical_cross_entropy
  reduce_lr_on_plateau_factor    : 0.1
  reduce_lr_on_plateau_patience  : 5
  reduce_lr_on_plateau_min_lr    : 0
  reduce_lr_every_nepochs_factor : 0.5
  reduce_lr_every_nepochs_nepochs: 30
  reduce_lr_every_nepochs_min_lr : 0.001
  stop_at_loss_metric            : val_accuracy
  stop_at_loss_threshold         : 0
  model_checkpoint_metric        : val_categorical_cross_entropy
  model_checkpoint_save_freq     : 2
  mpi_processes                  : 12
  threads_per_process            : 4
  parallel                       : data
  non_blocking_mpi               : False
  tracing                        : False
  profile                        : False
  gpus_per_node                  : 0
  enable_conv_gemm               : False
  enable_gpu                     : False
  enable_gpudirect               : False
  enable_nccl                    : False
  dtype                          : float32
**** Evaluating on test dataset...
Testing: 100%|████████████████████| 10000/10000 [00:00<00:00, 29732.29 samples/s, test_acc: 12.50%, test_cro: 2.3008704]
**** Training...
Epoch  1/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11184.77 samples/s, acc: 71.35%, cro: 1.2238941, val_acc: 88.49%, val_cro: 0.4369879]
Epoch  2/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10691.66 samples/s, acc: 88.87%, cro: 0.4051699, val_acc: 91.10%, val_cro: 0.3070377]
Epoch  3/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10617.42 samples/s, acc: 90.98%, cro: 0.3086980, val_acc: 92.56%, val_cro: 0.2624177]
Epoch  4/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10874.49 samples/s, acc: 92.43%, cro: 0.2576146, val_acc: 93.83%, val_cro: 0.2232232]
Epoch  5/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10451.37 samples/s, acc: 93.48%, cro: 0.2159374, val_acc: 94.76%, val_cro: 0.1868786]
Epoch  6/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10756.92 samples/s, acc: 94.81%, cro: 0.1748247, val_acc: 95.63%, val_cro: 0.1544418]
Epoch  7/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10901.69 samples/s, acc: 95.77%, cro: 0.1417673, val_acc: 96.25%, val_cro: 0.1331401]
Epoch  8/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11046.87 samples/s, acc: 96.55%, cro: 0.1164078, val_acc: 96.80%, val_cro: 0.1134956]
Epoch  9/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10944.06 samples/s, acc: 97.05%, cro: 0.0992564, val_acc: 96.98%, val_cro: 0.1033213]
Epoch 10/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11017.46 samples/s, acc: 97.48%, cro: 0.0866701, val_acc: 97.28%, val_cro: 0.0972526]
Epoch 11/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10871.28 samples/s, acc: 97.67%, cro: 0.0769905, val_acc: 97.58%, val_cro: 0.0862264]
Epoch 12/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10982.79 samples/s, acc: 97.99%, cro: 0.0682642, val_acc: 97.55%, val_cro: 0.0828536]
Epoch 13/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11115.45 samples/s, acc: 98.16%, cro: 0.0616423, val_acc: 97.77%, val_cro: 0.0782390]
Epoch 14/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10783.11 samples/s, acc: 98.30%, cro: 0.0562393, val_acc: 97.91%, val_cro: 0.0716845]
Epoch 15/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10642.71 samples/s, acc: 98.49%, cro: 0.0515601, val_acc: 97.93%, val_cro: 0.0696817]
Epoch 16/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10643.74 samples/s, acc: 98.62%, cro: 0.0468920, val_acc: 97.98%, val_cro: 0.0688842]
Epoch 17/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10726.90 samples/s, acc: 98.70%, cro: 0.0434075, val_acc: 98.10%, val_cro: 0.0675637]
Epoch 18/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10558.22 samples/s, acc: 98.71%, cro: 0.0424472, val_acc: 98.25%, val_cro: 0.0641221]
Epoch 19/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10809.42 samples/s, acc: 98.86%, cro: 0.0382850, val_acc: 98.19%, val_cro: 0.0646157]
Epoch 20/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10712.11 samples/s, acc: 98.95%, cro: 0.0348660, val_acc: 98.25%, val_cro: 0.0617139]
Epoch 21/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11068.46 samples/s, acc: 99.05%, cro: 0.0323043, val_acc: 98.14%, val_cro: 0.0658118]
Epoch 22/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11436.88 samples/s, acc: 99.06%, cro: 0.0306285, val_acc: 98.17%, val_cro: 0.0648578]
Epoch 23/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11093.08 samples/s, acc: 99.17%, cro: 0.0282567, val_acc: 98.22%, val_cro: 0.0661603]
Epoch 24/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11058.23 samples/s, acc: 99.14%, cro: 0.0275220, val_acc: 98.28%, val_cro: 0.0638472]
Epoch 25/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11362.12 samples/s, acc: 99.27%, cro: 0.0242397, val_acc: 98.32%, val_cro: 0.0616558]
Epoch 26/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10929.57 samples/s, acc: 99.33%, cro: 0.0228250, val_acc: 98.41%, val_cro: 0.0614293]
Epoch 27/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10582.72 samples/s, acc: 99.33%, cro: 0.0218627, val_acc: 98.30%, val_cro: 0.0647660]
Epoch 28/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11540.73 samples/s, acc: 99.40%, cro: 0.0202375, val_acc: 98.31%, val_cro: 0.0653990]
Epoch 29/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11089.71 samples/s, acc: 99.47%, cro: 0.0187735, val_acc: 98.33%, val_cro: 0.0642570]
Epoch 30/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11112.27 samples/s, acc: 99.51%, cro: 0.0166023, val_acc: 98.40%, val_cro: 0.0630408]
Epoch 31/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11004.81 samples/s, acc: 99.56%, cro: 0.0154129, val_acc: 98.24%, val_cro: 0.0669048]
LRScheduler ReduceLROnPlateau: metric val_categorical_cross_entropy did not improve for 5 epochs, setting learning rate to 0.01000000
Epoch 32/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11015.29 samples/s, acc: 99.70%, cro: 0.0122010, val_acc: 98.39%, val_cro: 0.0635789]
Epoch 33/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11166.31 samples/s, acc: 99.74%, cro: 0.0111252, val_acc: 98.44%, val_cro: 0.0624000]
Epoch 34/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11112.63 samples/s, acc: 99.74%, cro: 0.0108013, val_acc: 98.44%, val_cro: 0.0627380]
Epoch 35/50: 100%|████████████████| 48000/48000 [00:04<00:00, 10914.84 samples/s, acc: 99.76%, cro: 0.0105415, val_acc: 98.47%, val_cro: 0.0627000]
Epoch 36/50: 100%|████████████████| 48000/48000 [00:04<00:00, 11017.57 samples/s, acc: 99.76%, cro: 0.0103665, val_acc: 98.50%, val_cro: 0.0628462]
LRScheduler EarlyStopping: metric val_categorical_cross_entropy did not improve for 10 epochs, stop training!
LRScheduler ReduceLROnPlateau: metric val_categorical_cross_entropy did not improve for 5 epochs, setting learning rate to 0.00100000
**** Done...
Time: 173.59 s
Throughput: 17282.50 samples/s
**** Evaluating on test dataset...
Testing: 100%|███████████████████| 10000/10000 [00:00<00:00, 28720.12 samples/s, test_acc: 100.00%, test_cro: 0.0000443]

Example: inference of the VGG16 CNN for the CIFAR-10 dataset

In this example, we perform inference with the CNN VGG16 for the CIFAR-10 dataset using 4 OpenMP threads:

$ export OMP_NUM_THREADS=4
$ pydtnn-benchmark \
    --model=vgg16_cifar10 \
    --dataset=cifar10 \
    --dataset-path=datasets/cifar10/cifar-10-binary.tar.gz \
    --evaluate-only=True \
    --batch-size=64 \
    --validation-split=0.2 \
    --weights-and-bias-filename=vgg16-weights-nhwc.npz \
    --tracing=False \
    --profile=False \
    --enable-gpu=True \
    --dtype=float32


**** vgg16_cifar10 model...
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
| Layer |           Type           | #Params | Output shape  |   Weights shape   |             Parameters              |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   0   |         InputCPU         |    0    |  (32, 32, 3)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   1   |        Conv2DCPU         |  1792   | (32, 32, 64)  |   (3, 3, 3, 64)   |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   2   |         ReluCPU          |    0    | (32, 32, 64)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   3   |        Conv2DCPU         |  36928  | (32, 32, 64)  |  (64, 3, 3, 64)   |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   4   |         ReluCPU          |    0    | (32, 32, 64)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   5   |       MaxPool2DCPU       |    0    | (16, 16, 64)  |      (2, 2)       |padd=(0,0), stride=(2,2), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   6   |        Conv2DCPU         |  73856  | (16, 16, 128) |  (64, 3, 3, 128)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   7   |         ReluCPU          |    0    | (16, 16, 128) |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   8   |        Conv2DCPU         | 147584  | (16, 16, 128) | (128, 3, 3, 128)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|   9   |         ReluCPU          |    0    | (16, 16, 128) |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  10   |       MaxPool2DCPU       |    0    |  (8, 8, 128)  |      (2, 2)       |padd=(0,0), stride=(2,2), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  11   |        Conv2DCPU         | 295168  |  (8, 8, 256)  | (128, 3, 3, 256)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  12   |         ReluCPU          |    0    |  (8, 8, 256)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  13   |        Conv2DCPU         | 590080  |  (8, 8, 256)  | (256, 3, 3, 256)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  14   |         ReluCPU          |    0    |  (8, 8, 256)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  15   |        Conv2DCPU         | 590080  |  (8, 8, 256)  | (256, 3, 3, 256)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  16   |         ReluCPU          |    0    |  (8, 8, 256)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  17   |       MaxPool2DCPU       |    0    |  (4, 4, 256)  |      (2, 2)       |padd=(0,0), stride=(2,2), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  18   |        Conv2DCPU         | 1180160 |  (4, 4, 512)  | (256, 3, 3, 512)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  19   |         ReluCPU          |    0    |  (4, 4, 512)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  20   |        Conv2DCPU         | 2359808 |  (4, 4, 512)  | (512, 3, 3, 512)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  21   |         ReluCPU          |    0    |  (4, 4, 512)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  22   |        Conv2DCPU         | 2359808 |  (4, 4, 512)  | (512, 3, 3, 512)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  23   |         ReluCPU          |    0    |  (4, 4, 512)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  24   |       MaxPool2DCPU       |    0    |  (2, 2, 512)  |      (2, 2)       |padd=(0,0), stride=(2,2), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  25   |        Conv2DCPU         | 2359808 |  (2, 2, 512)  | (512, 3, 3, 512)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  26   |         ReluCPU          |    0    |  (2, 2, 512)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  27   |        Conv2DCPU         | 2359808 |  (2, 2, 512)  | (512, 3, 3, 512)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  28   |         ReluCPU          |    0    |  (2, 2, 512)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  29   |        Conv2DCPU         | 2359808 |  (2, 2, 512)  | (512, 3, 3, 512)  |padd=(1,1), stride=(1,1), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  30   |         ReluCPU          |    0    |  (2, 2, 512)  |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  31   |       MaxPool2DCPU       |    0    |  (1, 1, 512)  |      (2, 2)       |padd=(0,0), stride=(2,2), dilat=(1,1)|
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  32   |        FlattenCPU        |    0    |    (512,)     |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  33   |          FCCPU           | 262656  |    (512,)     |    (512, 512)     |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  34   |         ReluCPU          |    0    |    (512,)     |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  35   |        DropoutCPU        |    0    |    (512,)     |                   |              rate=0.50              |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  36   |          FCCPU           | 262656  |    (512,)     |    (512, 512)     |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  37   |         ReluCPU          |    0    |    (512,)     |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  38   |        DropoutCPU        |    0    |    (512,)     |                   |              rate=0.50              |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  39   |          FCCPU           |  5130   |     (10,)     |     (512, 10)     |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|  40   |        SoftmaxCPU        |    0    |     (10,)     |                   |                                     |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
|             Total parameters      15245130   58.16 MBytes                                                            |
+-------+--------------------------+---------+---------------+-------------------+-------------------------------------+
**** Loading cifar10 dataset...
**** Parameters:
  model_name                     : vgg16_cifar10
  batch_size                     : 64
  global_batch_size              : None
  dtype                          : <class 'numpy.float32'>
  num_epochs                     : 400
  steps_per_epoch                : 0
  evaluate_on_train              : True
  evaluate_only                  : True
  weights_and_bias_filename      : vgg16-weights-nhwc.npz
  history_file                   : None
  shared_storage                 : False
  enable_fused_bn_relu           : False
  enable_fused_conv_relu         : False
  enable_fused_conv_bn           : False
  enable_fused_conv_bn_relu      : False
  tensor_format                  : NHWC
  enable_best_of                 : False
  dataset_name                   : cifar10
  use_synthetic_data             : False
  dataset_path                   : datasets/cifar10/cifar-10-binary.tar.gz
  test_as_validation             : True
  augment_flip                    : True
  augment_flip_prob               : 0.5
  augment_crop                    : True
  augment_crop_size               : 16
  augment_crop_prob               : 0.5
  validation_split               : 0.2
  optimizer_name                 : sgd
  learning_rate                  : 0.01
  learning_rate_scaling          : True
  momentum                       : 0.9
  decay                          : 0.0001
  nesterov                       : False
  beta1                          : 0.99
  beta2                          : 0.999
  epsilon                        : 1e-07
  rho                            : 0.9
  loss_func                      : categorical_cross_entropy
  metrics                        : categorical_accuracy
  schedulers_names            : warm_up,reduce_lr_on_plateau,model_checkpoint,early_stopping
  warm_up_epochs                 : 5
  early_stopping_metric          : val_categorical_cross_entropy
  early_stopping_patience        : 20
  reduce_lr_on_plateau_metric    : val_categorical_cross_entropy
  reduce_lr_on_plateau_factor    : 0.1
  reduce_lr_on_plateau_patience  : 15
  reduce_lr_on_plateau_min_lr    : 1e-05
  reduce_lr_every_nepochs_factor : 0.5
  reduce_lr_every_nepochs_nepochs: 50
  reduce_lr_every_nepochs_min_lr : 0.001
  stop_at_loss_metric            : val_categorical_accuracy
  stop_at_loss_threshold         : 70.0
  model_checkpoint_metric        : categorical_accuracy
  model_checkpoint_save_freq     : 2
  enable_conv_gemm               : False
  enable_memory_cache            : True
  enable_conv_winograd           : False
  mpi_processes                  : 1
  threads_per_process            : 4
  parallel                       : sequential
  non_blocking_mpi               : False
  gpus_per_node                  : 2
  enable_gpu                     : False
  enable_gpudirect               : False
  enable_nccl                    : False
  enable_cudnn_auto_conv_alg     : True
  tracing                        : True
  tracer_output                  : prueba.trc
  profile                        : False
**** Evaluating on test dataset...
Testing: 100%|██████████████████████| 10000/10000 [00:13<00:00, 715.46 samples/s, test_cce: 0.4376189, test_acc: 89.24%]

Credits

The main contributors to PyDTNN are:

• Miguel Ángel Prosper Quirós ([email protected])
• Paul Ximo Pluijter Izquierdo ([email protected])
• Manuel Francisco Dolz Zaragozá ([email protected])
• Sergio Barrachina Mir ([email protected])
• Miguel Pardo Navarro ([email protected])
• Andrés Enrique Tomás Domínguez ([email protected])
• Adrián Castelló Gimeno ([email protected])
• Adrián Bartolomé López ([email protected])
• Mar Catalán Carbó ([email protected])
• Jose Ignacio Mestre Miravet ([email protected])
• Enrique Salvador Quintana Ortí ([email protected])

If you have questions or comments about PyDTNN, please contact:

• Manuel Francisco Dolz Zaragozá ([email protected])

Citing PyDTNN

If you use PyDTNN, and you would like to acknowledge the project in your academic publication, we suggest citing the following paper:

• PyDTNN: A user-friendly and extensible framework for distributed deep learning. Sergio Barrachina, Adrián Castelló, Mar Catalán, Manuel F. Dolz, Jose I. Mestre. Journal of Supercomputing 77(9), pp. 9971-9987 (2021) ISSN: 1573-0484. DOI: 10.1007/s11227-021-03673-z.

Other references:

• A Flexible Research-Oriented Framework for Distributed Training of Deep Neural Networks. Sergio Barrachina, Adrián Castelló, Mar Catalán, Manuel F. Dolz and Jose I. Mestre. 2021 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), pp. 730-739 (2021) DOI: 10.1109/IPDPSW52791.2021.00110.

Acknowledgments

The PyDTNN library has been partially supported by:

• Project TIN2017-82972-R "Algorithmic Techniques for Energy-Aware and Error-Resilient High Performance Computing" funded by the Spanish Ministry of Economy and Competitiveness (2018-2020).
• Project RTI2018-098156-B-C51 "Innovative Technologies of Processors, Accelerators and Networks for Data Centers and High Performance Computing" funded by the Spanish Ministry of Science, Innovation and Universities.
• Project CDEIGENT/2017/04 "High Performance Computing for Neural Networks" funded by the Valencian Government.
• Project UJI-A2019-11 "Energy-Aware High Performance Computing for Deep Neural Networks" funded by the Universitat Jaume I.
• Project CIDEXG/2022/13 "AT4SUSDL: Advanced Techniques for Sustainable Deep Learning" funded by the Valencian Government.
• Project RYC2021-033973-I "Dotación ayuda Ramón y Cajal" funded by the Spanish Ministry of Science, Innovation and Universities.
• Project PID2023-146569NB-C22 "Inteligencia sostenible en el Borde-UJI" funded by the Spanish Ministry of Science, Innovation and Universities.
• Project C121/23 Convenio "CIBERseguridad post-Cuántica para el Aprendizaje Federado en procesadores de bajo consumo y aceleradores (CIBER-CAFE)" funded by the Spanish National Cybersecurity Institute.