Notebook

How to benchmark models with Transformers¶

With ever-larger language models, it is no longer enough to just compare models on their performance on a specific task. One should always be aware of the computational cost that is attached to a specific model. For a given computation environment (e.g. type of GPU), the computational cost of training a model or deploying it in inference usually depends only on the required memory and the required time.

Being able to accurately benchmark language models on both speed and required memory is therefore very important.

HuggingFace's Transformer library allows users to benchmark models for both TensorFlow 2 and PyTorch using the PyTorchBenchmark and TensorFlowBenchmark classes.

The currently available features for PyTorchBenchmark are summarized in the following table.

		CPU	CPU + torchscript	GPU	GPU + torchscript	GPU + FP16
Speed - Inference	✔	✔	✔	✔	✔	✔
Memory - Inference	✔	✔	✔	✔	✔	✘
Speed - Train	✔	✘	✔	✘	✔	✔
Memory - Train	✔	✘	✔	✘	✔	✘

FP16 stands for mixed-precision meaning that computations within the model are done using a mixture of 16-bit and 32-bit floating-point operations, see here for more detail.
torchscript corresponds to PyTorch's torchscript format, see here.

The currently available features for TensorFlowBenchmark are summarized in the following table.

		CPU	CPU + eager execution	GPU	GPU + eager execution	GPU + XLA	GPU + FP16
Speed - Inference	✔	✔	✔	✔	✔	✘	✔
Memory - Inference	✔	✔	✔	✔	✔	✘	✘
Speed - Train	✔	✘	✔	✘	✘	✘	✔
Memory - Train	✔	✘	✔	✘	✘	✘	✘

eager execution means that the function is run in the eager execution environment of TensorFlow 2, see here.
XLA stands for TensorFlow's Accelerated Linear Algebra (XLA) compiler, see here
FP16 stands for TensorFlow's mixed-precision package and is analogous to PyTorch's FP16 feature, see here.

*Note*: Benchmark training in TensorFlow is not included in v3.0.2, but available in master.

This notebook will show the user how to use PyTorchBenchmark and TensorFlowBenchmark for two different scenarios:

Inference - Pre-trained Model Comparison - A user wants to implement a pre-trained model in production for inference. She wants to compare different models on speed and required memory.
Training - Configuration Comparison - A user wants to train a specific model and searches that for himself most effective model configuration.

Inference - Pre-trained Model Comparison¶

Let's say we want to employ a question-answering model in production. The questions are expected to be of the same format as in SQuAD v2, so that the model to choose should have been fine-tuned on this dataset.

HuggingFace's new dataset webpage lets the user see all relevant information about a dataset and even links the models that have been fine-tuned on this specific dataset. Let's check out the dataset webpage of SQuAD v2 here.

Nice, we can see that there are 7 available models.

Texte alternatif…

Let's assume that we have decided to restrict our pipeline to "encoder-only" models so that we are left with:

a-ware/roberta-large-squad-classification
a-ware/xlmroberta-squadv2
aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2
deepset/roberta-base-squad2
mrm8488/longformer-base-4096-finetuned-squadv2

Great! In this notebook, we will now benchmark these models on both peak memory consumption and inference time to decide which model should be employed in production.

*Note*: None of the models has been tested on performance so that we will just assume that all models perform more or less equally well. The purpose of this notebook is not to find the best model for SQuAD v2, but to showcase how Transformers benchmarking tools can be leveraged.

First, we assume to be limited by the available GPU on this google colab, which in this copy amounts to 16 GB of RAM.

In a first step, we will check which models are the most memory-efficient ones. Let's make sure 100% of the GPU is available to us in this notebook.

In [ ]:

#@title Check available memory of GPU
# Check that we are using 100% of GPU
# memory footprint support libraries/code
!ln -sf /opt/bin/nvidia-smi /usr/bin/nvidia-smi
!pip -q install gputil
!pip -q install psutil
!pip -q install humanize
import psutil
import humanize
import os
import GPUtil as GPU
GPUs = GPU.getGPUs()
# XXX: only one GPU on Colab and isn’t guaranteed
gpu = GPUs[0]
def printm():
 process = psutil.Process(os.getpid())
 print("Gen RAM Free: " + humanize.naturalsize( psutil.virtual_memory().available ), " | Proc size: " + humanize.naturalsize( process.memory_info().rss))
 print("GPU RAM Free: {0:.0f}MB | Used: {1:.0f}MB | Util {2:3.0f}% | Total {3:.0f}MB".format(gpu.memoryFree, gpu.memoryUsed, gpu.memoryUtil*100, gpu.memoryTotal))
printm()

  Building wheel for gputil (setup.py) ... done
Gen RAM Free: 12.8 GB  | Proc size: 160.0 MB
GPU RAM Free: 16280MB | Used: 0MB | Util   0% | Total 16280MB

In [ ]:

# If GPU RAM Util > 0% => crash notebook on purpose
# !kill -9 -1

Looks good! Now we import transformers and download the scripts run_benchmark.py, run_benchmark_tf.py, and plot_csv_file.py which can be found under transformers/examples/benchmarking.

run_benchmark_tf.py and run_benchmark.py are very simple scripts leveraging the PyTorchBenchmark and TensorFlowBenchmark classes, respectively.

In [ ]:

# install transformes
!pip uninstall -y transformers
!pip install -q git+https://github.com/huggingface/transformers.git

# install py3nvml to track GPU memory usage
!pip install -q py3nvml

!rm -f run_benchmark.py
!rm -f run_benchmark_tf.py
!rm -f plot_csv_file.py
!wget https://raw.githubusercontent.com/huggingface/transformers/master/examples/benchmarking/run_benchmark.py -qq
!wget https://raw.githubusercontent.com/huggingface/transformers/master/examples/benchmarking/run_benchmark_tf.py -qq
!wget https://raw.githubusercontent.com/huggingface/transformers/master/examples/benchmarking/plot_csv_file.py -qq

# import pandas to pretty print csv files
import pandas as pd

Information about the input arguments to the run_benchmark scripts can be accessed by running !python run_benchmark.py --help for PyTorch and !python run_benchmark_tf.py --help for TensorFlow.

In [ ]:

!python run_benchmark.py --help

2020-06-26 11:51:47.129203: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1
usage: run_benchmark.py [-h] [--models MODELS [MODELS ...]]
                        [--batch_sizes BATCH_SIZES [BATCH_SIZES ...]]
                        [--sequence_lengths SEQUENCE_LENGTHS [SEQUENCE_LENGTHS ...]]
                        [--no_inference] [--no_cuda] [--no_tpu] [--fp16]
                        [--training] [--verbose] [--no_speed] [--no_memory]
                        [--trace_memory_line_by_line] [--save_to_csv]
                        [--log_print] [--no_env_print] [--no_multi_process]
                        [--with_lm_head]
                        [--inference_time_csv_file INFERENCE_TIME_CSV_FILE]
                        [--inference_memory_csv_file INFERENCE_MEMORY_CSV_FILE]
                        [--train_time_csv_file TRAIN_TIME_CSV_FILE]
                        [--train_memory_csv_file TRAIN_MEMORY_CSV_FILE]
                        [--env_info_csv_file ENV_INFO_CSV_FILE]
                        [--log_filename LOG_FILENAME] [--repeat REPEAT]
                        [--only_pretrain_model] [--torchscript]
                        [--torch_xla_tpu_print_metrics]
                        [--fp16_opt_level FP16_OPT_LEVEL]

optional arguments:
  -h, --help            show this help message and exit
  --models MODELS [MODELS ...]
                        Model checkpoints to be provided to the AutoModel
                        classes. Leave blank to benchmark the base version of
                        all available models
  --batch_sizes BATCH_SIZES [BATCH_SIZES ...]
                        List of batch sizes for which memory and time
                        performance will be evaluated
  --sequence_lengths SEQUENCE_LENGTHS [SEQUENCE_LENGTHS ...]
                        List of sequence lengths for which memory and time
                        performance will be evaluated
  --no_inference        Don't benchmark inference of model
  --no_cuda             Whether to run on available cuda devices
  --no_tpu              Whether to run on available tpu devices
  --fp16                Use FP16 to accelerate inference.
  --training            Benchmark training of model
  --verbose             Verbose memory tracing
  --no_speed            Don't perform speed measurments
  --no_memory           Don't perform memory measurments
  --trace_memory_line_by_line
                        Trace memory line by line
  --save_to_csv         Save result to a CSV file
  --log_print           Save all print statements in a log file
  --no_env_print        Don't print environment information
  --no_multi_process    Don't use multiprocessing for memory and speed
                        measurement. It is highly recommended to use
                        multiprocessing for accurate CPU and GPU memory
                        measurements. This option should only be used for
                        debugging / testing and on TPU.
  --with_lm_head        Use model with its language model head
                        (MODEL_WITH_LM_HEAD_MAPPING instead of MODEL_MAPPING)
  --inference_time_csv_file INFERENCE_TIME_CSV_FILE
                        CSV filename used if saving time results to csv.
  --inference_memory_csv_file INFERENCE_MEMORY_CSV_FILE
                        CSV filename used if saving memory results to csv.
  --train_time_csv_file TRAIN_TIME_CSV_FILE
                        CSV filename used if saving time results to csv for
                        training.
  --train_memory_csv_file TRAIN_MEMORY_CSV_FILE
                        CSV filename used if saving memory results to csv for
                        training.
  --env_info_csv_file ENV_INFO_CSV_FILE
                        CSV filename used if saving environment information.
  --log_filename LOG_FILENAME
                        Log filename used if print statements are saved in
                        log.
  --repeat REPEAT       Times an experiment will be run.
  --only_pretrain_model
                        Instead of loading the model as defined in
                        `config.architectures` if exists, just load the
                        pretrain model weights.
  --torchscript         Trace the models using torchscript
  --torch_xla_tpu_print_metrics
                        Print Xla/PyTorch tpu metrics
  --fp16_opt_level FP16_OPT_LEVEL
                        For fp16: Apex AMP optimization level selected in
                        ['O0', 'O1', 'O2', and 'O3'].See details at
                        https://nvidia.github.io/apex/amp.html

Great, we are ready to run our first memory benchmark. By default, both the required memory and time for inference is enabled. To disable benchmarking on time, we add --no_speed.

The only required parameter is --models which expects a list of model identifiers as defined on the model hub. Here we add the five model identifiers listed above.

Next, we define the sequence_lengths and batch_sizes for which the peak memory is calculated.

Finally, because the results should be stored in a CSV file, the option --save_to_csv is added and the path to save the results is added via the --inference_memory_csv_file argument. Whenever a benchmark is run, the environment information, e.g. GPU type, library versions, ... can be saved using the --env_info_csv_file argument.

In [ ]:

# create plots folder in content
!mkdir -p plots_pt

In [ ]:

# run benchmark
!python run_benchmark.py --no_speed --save_to_csv \
                                --models a-ware/roberta-large-squad-classification \
                                  a-ware/xlmroberta-squadv2 \
                                  aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 \
                                  deepset/roberta-base-squad2 \
                                  mrm8488/longformer-base-4096-finetuned-squadv2 \
                                --sequence_lengths 32 128 512 1024 \
                                --batch_sizes 32 \
                                --inference_memory_csv_file plots_pt/required_memory.csv \
                                --env_info_csv_file plots_pt/env.csv >/dev/null 2>&1  # redirect all prints

Under plots_pt, two files are now created: required_memory.csv and env.csv. Let's check out required_memory.csv first.

In [ ]:

df = pd.read_csv('plots_pt/required_memory.csv')
df

Out[ ]:

	model	batch_size	sequence_length	result
0	a-ware/roberta-large-squad-classification	32	32	2219.0
1	a-ware/roberta-large-squad-classification	32	128	2455.0
2	a-ware/roberta-large-squad-classification	32	512	3641.0
3	a-ware/roberta-large-squad-classification	32	1024	NaN
4	a-ware/xlmroberta-squadv2	32	32	2999.0
5	a-ware/xlmroberta-squadv2	32	128	3235.0
6	a-ware/xlmroberta-squadv2	32	512	4421.0
7	a-ware/xlmroberta-squadv2	32	1024	NaN
8	aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200...	32	32	1025.0
9	aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200...	32	128	1143.0
10	aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200...	32	512	1719.0
11	aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200...	32	1024	NaN
12	deepset/roberta-base-squad2	32	32	1373.0
13	deepset/roberta-base-squad2	32	128	1533.0
14	deepset/roberta-base-squad2	32	512	2433.0
15	deepset/roberta-base-squad2	32	1024	NaN
16	mrm8488/longformer-base-4096-finetuned-squadv2	32	32	3783.0
17	mrm8488/longformer-base-4096-finetuned-squadv2	32	128	3783.0
18	mrm8488/longformer-base-4096-finetuned-squadv2	32	512	3783.0
19	mrm8488/longformer-base-4096-finetuned-squadv2	32	1024	6427.0

Each row in the csv file lists one data point showing the peak memory usage for a given model, batch_size and sequence_length. As can be seen, some values have a NaN result meaning that an Out-of-Memory Error occurred. To better visualize the results, one can make use of the plot_csv_file.py script.

Before, let's take a look at the information about our computation environment.

In [ ]:

df = pd.read_csv('plots_pt/env.csv')
df

Out[ ]:

	transformers_version	2.11.0
0	framework	PyTorch
1	use_torchscript	False
2	framework_version	1.5.1+cu101
3	python_version	3.6.9
4	system	Linux
5	cpu	x86_64
6	architecture	64bit
7	date	2020-06-26
8	time	11:56:37.277009
9	fp16	False
10	use_multiprocessing	True
11	only_pretrain_model	False
12	cpu_ram_mb	13021
13	use_gpu	True
14	num_gpus	1
15	gpu	Tesla P100-PCIE-16GB
16	gpu_ram_mb	16280
17	gpu_power_watts	250.0
18	gpu_performance_state	0
19	use_tpu	False

We can see all relevant information here: the PyTorch version, the Python version, the system, the type of GPU, and available RAM on the GPU, etc...

Note: A different GPU is likely assigned to a copy of this notebook, so that all of the following results may be different. It is very important to always include the environment information when benchmarking your models for both reproducibility and transparency to other users.

Alright, let's plot the results.

In [ ]:

# plot graph and save as image
!python plot_csv_file.py --csv_file plots_pt/required_memory.csv --figure_png_file=plots_pt/required_memory_plot.png --no_log_scale --short_model_names a-ware-roberta a-aware-xlm aodiniz-bert deepset-roberta mrm8488-long

# show image
from IPython.display import Image
Image('plots_pt/required_memory_plot.png')

2020-06-26 11:56:39.671579: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1

Out[ ]:

At this point, it is important to understand how the peak memory is measured. The benchmarking tools measure the peak memory usage the same way the command nvidia-smi does - see here for more information. In short, all memory that is allocated for a given model identifier, batch size and sequence length is measured in a separate process. This way it can be ensured that there is no previously unreleased memory falsely included in the measurement. One should also note that the measured memory even includes the memory allocated by the CUDA driver to load PyTorch and TensorFlow and is, therefore, higher than library-specific memory measurement function, e.g. this one for PyTorch.

Alright, let's analyze the results. It can be noted that the models aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 and deepset/roberta-base-squad2 require significantly less memory than the other three models. Besides mrm8488/longformer-base-4096-finetuned-squadv2 all models more or less follow the same memory consumption pattern with aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 seemingly being able to better scale to larger sequence lengths. mrm8488/longformer-base-4096-finetuned-squadv2 is a Longformer model, which makes use of LocalAttention (check this blog post to learn more about local attention) so that the model scales much better to longer input sequences.

For the sake of this notebook, we assume that the longest required input will be less than 512 tokens so that we settle on the models aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 and deepset/roberta-base-squad2.

To better understand how many API requests of our question-answering pipeline can be run in parallel, we are interested in finding out how many batches the two models run out of memory.

In [ ]:

!python run_benchmark.py --no_speed --save_to_csv \
                                --inference_memory_csv_file plots_pt/required_memory_2.csv \
                                --env_info_csv_file plots_pt/env.csv \
                                --models aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 \
                                  deepset/roberta-base-squad2 \
                                --sequence_lengths 512 \
                                --batch_sizes 64 128 256 512\
                                --no_env_print

2020-06-26 11:56:44.781155: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1
1 / 2
2 / 2
Doesn't fit on GPU. CUDA out of memory. Tried to allocate 6.00 GiB (GPU 0; 15.90 GiB total capacity; 9.47 GiB already allocated; 5.60 GiB free; 9.52 GiB reserved in total by PyTorch)

====================      INFERENCE - MEMORY - RESULT       ====================
--------------------------------------------------------------------------------
          Model Name             Batch Size     Seq Length    Memory in MB 
--------------------------------------------------------------------------------
aodiniz/bert_uncased_L-10_H-51       64             512             2455     
aodiniz/bert_uncased_L-10_H-51      128             512             3929     
aodiniz/bert_uncased_L-10_H-51      256             512             6875     
aodiniz/bert_uncased_L-10_H-51      512             512            12783     
 deepset/roberta-base-squad2         64             512             3539     
 deepset/roberta-base-squad2        128             512             5747     
 deepset/roberta-base-squad2        256             512            10167     
 deepset/roberta-base-squad2        512             512             N/A      
--------------------------------------------------------------------------------
Saving results to csv.

Let's plot the results again, this time changing the x-axis to batch_size however.

In [ ]:

# plot graph and save as image
!python plot_csv_file.py --csv_file plots_pt/required_memory_2.csv \
                          --figure_png_file=plots_pt/required_memory_plot_2.png \
                          --no_log_scale \
                          --short_model_names aodiniz-bert deepset-roberta \
                          --plot_along_batch

# show image
from IPython.display import Image
Image('plots_pt/required_memory_plot_2.png')

2020-06-26 11:57:51.876810: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1

Out[ ]:

Interesting! aodiniz/bert_uncased_L-10_H-51 clearly scales better for higher batch sizes and does not even run out of memory for 512 tokens.

For comparison, let's run the same benchmarking on TensorFlow.

In [ ]:

# create plots folder in content
!mkdir -p plots_tf

!TF_CPP_MIN_LOG_LEVEL=3 python run_benchmark_tf.py --no_speed --save_to_csv \
                                --inference_memory_csv_file plots_tf/required_memory_2.csv \
                                --env_info_csv_file plots_tf/env.csv \
                                --models aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 \
                                         deepset/roberta-base-squad2 \
                                --sequence_lengths 512 \
                                --batch_sizes 64 128 256 512 \
                                --no_env_print \

1 / 2
Doesn't fit on GPU.  OOM when allocating tensor with shape[512,8,512,512] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
	 [[node tf_bert_model/bert/encoder/layer_._0/attention/self/Softmax (defined at /usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_bert.py:267) ]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.
 [Op:__inference_run_in_graph_mode_4243]

Errors may have originated from an input operation.
Input Source operations connected to node tf_bert_model/bert/encoder/layer_._0/attention/self/Softmax:
 tf_bert_model/bert/encoder/layer_._0/attention/self/add (defined at /usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_bert.py:264)

Function call stack:
run_in_graph_mode

2 / 2
Doesn't fit on GPU.  OOM when allocating tensor with shape[512,12,512,512] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
	 [[node tf_roberta_model/roberta/encoder/layer_._0/attention/self/Softmax (defined at /usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_bert.py:267) ]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.
 [Op:__inference_run_in_graph_mode_5047]

Errors may have originated from an input operation.
Input Source operations connected to node tf_roberta_model/roberta/encoder/layer_._0/attention/self/Softmax:
 tf_roberta_model/roberta/encoder/layer_._0/attention/self/add (defined at /usr/local/lib/python3.6/dist-packages/transformers/modeling_tf_bert.py:264)

Function call stack:
run_in_graph_mode


====================      INFERENCE - MEMORY - RESULT       ====================
--------------------------------------------------------------------------------
          Model Name             Batch Size     Seq Length    Memory in MB 
--------------------------------------------------------------------------------
aodiniz/bert_uncased_L-10_H-51       64             512             2885     
aodiniz/bert_uncased_L-10_H-51      128             512             4933     
aodiniz/bert_uncased_L-10_H-51      256             512             9029     
aodiniz/bert_uncased_L-10_H-51      512             512             N/A      
 deepset/roberta-base-squad2         64             512             4933     
 deepset/roberta-base-squad2        128             512             9029     
 deepset/roberta-base-squad2        256             512            15391     
 deepset/roberta-base-squad2        512             512             N/A      
--------------------------------------------------------------------------------
Saving results to csv.

Let's see the same plot for TensorFlow.

In [ ]:

# plot graph and save as image
!python plot_csv_file.py --csv_file plots_tf/required_memory_2.csv --figure_png_file=plots_tf/required_memory_plot_2.png --no_log_scale --short_model_names aodiniz-bert deepset-roberta --plot_along_batch

# show image
from IPython.display import Image
Image('plots_tf/required_memory_plot_2.png')

2020-06-26 11:59:28.790462: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1

Out[ ]:

The model implemented in TensorFlow requires more memory than the one implemented in PyTorch. Let's say for whatever reason we have decided to use TensorFlow instead of PyTorch.

The next step is to measure the inference time of these two models. Instead of disabling time measurement with --no_speed, we will now disable memory measurement with --no_memory.

In [ ]:

!TF_CPP_MIN_LOG_LEVEL=3 python run_benchmark_tf.py --no_memory --save_to_csv \
                                --inference_time_csv_file plots_tf/time_2.csv \
                                --env_info_csv_file plots_tf/env.csv \
                                --models aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 \
                                         deepset/roberta-base-squad2 \
                                --sequence_lengths 8 32 128 512 \
                                --batch_sizes 256 \
                                --no_env_print \

1 / 2
2 / 2

====================       INFERENCE - SPEED - RESULT       ====================
--------------------------------------------------------------------------------
          Model Name             Batch Size     Seq Length     Time in s   
--------------------------------------------------------------------------------
aodiniz/bert_uncased_L-10_H-51      256              8             0.033     
aodiniz/bert_uncased_L-10_H-51      256              32            0.119     
aodiniz/bert_uncased_L-10_H-51      256             128            0.457     
aodiniz/bert_uncased_L-10_H-51      256             512             2.21     
 deepset/roberta-base-squad2        256              8             0.064     
 deepset/roberta-base-squad2        256              32             0.25     
 deepset/roberta-base-squad2        256             128             1.01     
 deepset/roberta-base-squad2        256             512             4.65     
--------------------------------------------------------------------------------
Saving results to csv.

In [ ]:

# plot graph and save as image
!python plot_csv_file.py --csv_file plots_tf/time_2.csv --figure_png_file=plots_tf/time_plot_2.png --no_log_scale --short_model_names aodiniz-bert deepset-roberta --is_time

# show image
from IPython.display import Image
Image('plots_tf/time_plot_2.png')

2020-06-26 12:04:58.002654: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1

Out[ ]:

Ok, this took some time... time measurements take much longer than memory measurements because the forward pass is called multiple times for stable results. Timing measurements leverage Python's timeit module and run 10 times the value given to the --repeat argument (defaults to 3), so in our case 30 times.

Let's focus on the resulting plot. It becomes obvious that aodiniz/bert_uncased_L-10_H-51 is around twice as fast as deepset/roberta-base-squad2. Given that the model is also more memory efficient and assuming that the model performs reasonably well, for the sake of this notebook we will settle on aodiniz/bert_uncased_L-10_H-51. Our model should be able to process input sequences of up to 512 tokens. Latency time of around 2 seconds might be too long though, so let's compare the time for different batch sizes and using TensorFlows XLA package for more speed.

In [ ]:

!TF_CPP_MIN_LOG_LEVEL=3 python run_benchmark_tf.py --no_memory --save_to_csv \
                                --inference_time_csv_file plots_tf/time_xla_1.csv \
                                --env_info_csv_file plots_tf/env.csv \
                                --models aodiniz/bert_uncased_L-10_H-512_A-8_cord19-200616_squad2 \
                                --sequence_lengths 512 \
                                --batch_sizes 8 64 256 \
                                --no_env_print \
                                --use_xla

1 / 1

====================       INFERENCE - SPEED - RESULT       ====================
--------------------------------------------------------------------------------
          Model Name             Batch Size     Seq Length     Time in s   
--------------------------------------------------------------------------------
aodiniz/bert_uncased_L-10_H-51       8              512            0.056     
aodiniz/bert_uncased_L-10_H-51       64             512            0.402     
aodiniz/bert_uncased_L-10_H-51      256             512            1.591     
--------------------------------------------------------------------------------
Saving results to csv.

First of all, it can be noted that XLA reduces latency time by a factor of ca. 1.3 (which is more than observed for other models by TensorFlow here). A batch size of 64 looks like a good choice. More or less half a second for the forward pass is good enough.

Cool, now it should be straightforward to benchmark your favorite models. All the inference time measurements can also be done using the run_benchmark.py script for PyTorch.

Training - Configuration Comparison¶

Next, we will look at how a model can be benchmarked on different configurations. This is especially helpful when one wants to decide how to most efficiently choose the model's configuration parameters for training. In the following different configurations of a Bart MNLI model will be compared to each other using PyTorchBenchmark.

Training in PyTorchBenchmark is defined by running one forward pass to compute the loss: loss = model(input_ids, labels=labels)[0] and one backward pass to compute the gradients loss.backward().

Let's see how to most efficiently train a Bart MNLI model from scratch.

In [ ]:

# Imports
from transformers import BartConfig, PyTorchBenchmark, PyTorchBenchmarkArguments

For the sake of the notebook, we assume that we are looking for a more efficient version of Facebook's bart-large-mnli model. Let's load its configuration and check out the important parameters.

In [ ]:

BartConfig.from_pretrained("facebook/bart-large-mnli").to_diff_dict()

HBox(children=(FloatProgress(value=0.0, description='Downloading', max=908.0, style=ProgressStyle(description_…

Out[ ]:

{'_num_labels': 3,
 'activation_dropout': 0.0,
 'activation_function': 'gelu',
 'add_bias_logits': False,
 'add_final_layer_norm': False,
 'attention_dropout': 0.0,
 'bos_token_id': 0,
 'classif_dropout': 0.0,
 'd_model': 1024,
 'decoder_attention_heads': 16,
 'decoder_ffn_dim': 4096,
 'decoder_layerdrop': 0.0,
 'decoder_layers': 12,
 'dropout': 0.1,
 'encoder_attention_heads': 16,
 'encoder_ffn_dim': 4096,
 'encoder_layerdrop': 0.0,
 'encoder_layers': 12,
 'eos_token_id': 2,
 'extra_pos_embeddings': 2,
 'id2label': {0: 'contradiction', 1: 'neutral', 2: 'entailment'},
 'init_std': 0.02,
 'is_encoder_decoder': True,
 'label2id': {'contradiction': 0, 'entailment': 2, 'neutral': 1},
 'max_position_embeddings': 1024,
 'model_type': 'bart',
 'normalize_before': False,
 'normalize_embedding': True,
 'num_hidden_layers': 12,
 'output_past': False,
 'pad_token_id': 1,
 'scale_embedding': False,
 'static_position_embeddings': False,
 'vocab_size': 50265}

Alright! The important configuration parameters are usually the number of layers config.encoder_num_layers and config.decoder_num_layers, the model's hidden size: config.d_model, the number of attention heads config.encoder_attention_heads and config.decoder_attention_heads and the vocabulary size config.vocab_size.

Let's create 4 configurations different from the baseline and see how they compare in terms of peak memory consumption.

In [ ]:

config_baseline = BartConfig.from_pretrained("facebook/bart-large-mnli")
config_768_hidden = BartConfig.from_pretrained("facebook/bart-large-mnli", d_model=768)
config_8_heads = BartConfig.from_pretrained("facebook/bart-large-mnli", decoder_attention_heads=8, encoder_attention_heads=8)
config_10000_vocab = BartConfig.from_pretrained("facebook/bart-large-mnli", vocab_size=10000)
config_8_layers = BartConfig.from_pretrained("facebook/bart-large-mnli", encoder_layers=8, decoder_layers=8)

Cool, now we can benchmark these configs against the baseline config. This time, instead of using the benchmarking script we will directly use the PyTorchBenchmark class. The class expects the argument args which has to be of type PyTorchBenchmarkArguments and optionally a list of configs.

First, we define the args and give the different configurations appropriate model names. The model names must be in the same order as the configs that are directly passed to PyTorchBenchMark.

If no configs are provided to PyTorchBenchmark, it is assumed that the model names ["bart-base", "bart-768-hid", "bart-8-head", "bart-10000-voc", "bart-8-lay"] correspond to official model identifiers and their corresponding configs are loaded as was shown in the previous section.

It is assumed that the model will be trained on half-precision, so we add the option fp16=True for the following benchmarks.

In [ ]:

# define args
args = PyTorchBenchmarkArguments(models=["bart-base", "bart-768-hid", "bart-8-head", "bart-10000-voc", "bart-8-lay"], 
                                 no_speed=True,
                                 no_inference=True,
                                 training=True, 
                                 train_memory_csv_file="plots_pt/training_mem_fp16.csv", 
                                 save_to_csv=True, 
                                 env_info_csv_file="plots_pt/env.csv",
                                 sequence_lengths=[64, 128, 256, 512],
                                 batch_sizes=[8],
                                 no_env_print=True,
                                 fp16=True)  # let's train on fp16

# create benchmark
benchmark = PyTorchBenchmark(configs=[config_baseline, config_768_hidden, config_8_heads, config_10000_vocab, config_8_layers], args=args)

# run benchmark
result = benchmark.run()

1 / 5
2 / 5
3 / 5
4 / 5
5 / 5

====================        TRAIN - MEMORY - RESULTS        ====================
--------------------------------------------------------------------------------
          Model Name             Batch Size     Seq Length    Memory in MB 
--------------------------------------------------------------------------------
          bart-base                  8               64             2905     
          bart-base                  8              128             3199     
          bart-base                  8              256             5401     
          bart-base                  8              512            11929     
         bart-768-hid                8               64             2441     
         bart-768-hid                8              128             2891     
         bart-768-hid                8              256             4963     
         bart-768-hid                8              512            10865     
         bart-8-head                 8               64             2869     
         bart-8-head                 8              128             3059     
         bart-8-head                 8              256             4825     
         bart-8-head                 8              512             9625     
        bart-10000-voc               8               64             2607     
        bart-10000-voc               8              128             2801     
        bart-10000-voc               8              256             4687     
        bart-10000-voc               8              512            10575     
          bart-8-lay                 8               64             2445     
          bart-8-lay                 8              128             2591     
          bart-8-lay                 8              256             4187     
          bart-8-lay                 8              512             8813     
--------------------------------------------------------------------------------
Saving results to csv.

Nice, let's plot the results again.

In [ ]:

# plot graph and save as image
!python plot_csv_file.py --csv_file plots_pt/training_mem_fp16.csv --figure_png_file=plots_pt/training_mem_fp16.png --no_log_scale

# show image
from IPython.display import Image
Image('plots_pt/training_mem_fp16.png')

2020-06-26 12:11:47.558303: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1

Out[ ]:

As expected the model of the baseline config requires the most memory.

It is interesting to see that the "bart-8-head" model initially requires more memory than bart-10000-voc, but then clearly outperforms bart-10000-voc at an input length of 512. Less surprising is that the "bart-8-lay" is by far the most memory-efficient model when reminding oneself that during the forward pass every layer has to store its activations for the backward pass.

Alright, given the data above, let's say we narrow our candidates down to only the "bart-8-head" and "bart-8-lay" models.

Let's compare these models again on training time.

In [ ]:

# define args
args = PyTorchBenchmarkArguments(models=["bart-8-head", "bart-8-lay"], 
                                 no_inference=True,
                                 training=True,
                                 no_memory=True,
                                 train_time_csv_file="plots_pt/training_speed_fp16.csv", 
                                 save_to_csv=True, 
                                 env_info_csv_file="plots_pt/env.csv",
                                 sequence_lengths=[32, 128, 512],
                                 batch_sizes=[8],
                                 no_env_print=True,
                                 repeat=1, # to make speed measurement faster but less accurate
                                 no_multi_process=True,  # google colab has problems with multi processing
                                 fp16=True
                                 )

# create benchmark
benchmark = PyTorchBenchmark(configs=[config_8_heads, config_8_layers], args=args)

# run benchmark
result = benchmark.run()

1 / 2
2 / 2

====================        TRAIN - SPEED - RESULTS         ====================
--------------------------------------------------------------------------------
          Model Name             Batch Size     Seq Length     Time in s   
--------------------------------------------------------------------------------
         bart-8-head                 8               32            0.127     
         bart-8-head                 8              128            0.398     
         bart-8-head                 8              512            1.567     
          bart-8-lay                 8               32            0.088     
          bart-8-lay                 8              128            0.284     
          bart-8-lay                 8              512            1.153     
--------------------------------------------------------------------------------
Saving results to csv.

The option no_multi_process disabled multi-processing here. This option should in general only be used for testing or debugging. Enabling multi-processing is crucial to ensure accurate memory consumption measurement, but is less important when only measuring speed. The main reason it is disabled here is that google colab sometimes raises "CUDA initialization" due to the notebook's environment. This problem does not arise when running benchmarks outside of a notebook.

Alright, let's plot the last speed results as well.

In [ ]:

# plot graph and save as image
!python plot_csv_file.py --csv_file plots_pt/training_speed_fp16.csv --figure_png_file=plots_pt/training_speed_fp16.png --no_log_scale --is_time

# show image
from IPython.display import Image
Image('plots_pt/training_speed_fp16.png')

2020-06-26 12:13:17.849561: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library libcudart.so.10.1

Out[ ]:

Unsurprisingly, "bart-8-lay" is faster than "bart-8-head" by a factor of ca. 1.3. It might very well be that reducing the layers by a factor of 2 leads to much more performance degradation than reducing the number of heads by a factor of 2. For more information on computational efficient Bart models, check out the new distilbart model here

Alright, that's it! Now you should be able to benchmark your favorite models on your favorite configurations.

Feel free to share your results with the community here or by tweeting us https://twitter.com/HuggingFace 🤗.