We Keep Coding

How does one fix a `Exception: process 0 terminated with signal SIGSEGV` error and if the single gpu code works fine?

I start 2 processes because I only have 2 gpus but then it gives me a Exception: process 0 terminated with signal SIGSEGV. This code does work with multiple cpus (or at least no error is thrown). Also, it works with a single GPU. Besides that is fails when world_size > 0 and multiple cuda/gpus are present.
My error message this this:
(automl-meta-learning) miranda9~/ML4Coq $ python playground/multiprocessing_playground/ddp_hello_world.py
world_size=2
Traceback (most recent call last):
File "playground/multiprocessing_playground/ddp_hello_world.py", line 49, in <module>
main()
File "playground/multiprocessing_playground/ddp_hello_world.py", line 43, in main
mp.spawn(example,
File "/home/miranda9/miniconda3/envs/automl-meta-learning/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 199, in spawn
return start_processes(fn, args, nprocs, join, daemon, start_method='spawn')
File "/home/miranda9/miniconda3/envs/automl-meta-learning/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 157, in start_processes
while not context.join():
File "/home/miranda9/miniconda3/envs/automl-meta-learning/lib/python3.8/site-packages/torch/multiprocessing/spawn.py", line 105, in join
raise Exception(
Exception: process 0 terminated with signal SIGSEGV
This is the code that gives the error:
import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import torch.nn as nn
import torch.optim as optim
from torch.nn.parallel import DistributedDataParallel as DDP
def example(rank, world_size):
# create default process group
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '8888'
dist.init_process_group("gloo", rank=rank, world_size=world_size)
# create local model
model = nn.Linear(10, 10).to(rank)
# construct DDP model
ddp_model = DDP(model, device_ids=[rank])
# define loss function and optimizer
loss_fn = nn.MSELoss()
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
# forward pass
outputs = ddp_model(torch.randn(20, 10).to(rank))
labels = torch.randn(20, 10).to(rank)
# backward pass
loss_fn(outputs, labels).backward()
# update parameters
optimizer.step()
def main():
# world_size = 2
world_size = torch.cuda.device_count()
mp.spawn(example,
args=(world_size,),
nprocs=world_size,
join=True)
if __name__=="__main__":
main()
print('Done\n\a')
[Optional] Larger self-contained example (gives same error)
Note however, that this slightly more complete example (only missing a distributed dataloader) also gives me the same issue:
"""
Based on: https://pytorch.org/tutorials/intermediate/ddp_tutorial.html
Correctness of code: https://stackoverflow.com/questions/66226135/how-to-parallelize-a-training-loop-ever-samples-of-a-batch-when-cpu-is-only-avai
Note: as opposed to the multiprocessing (torch.multiprocessing) package, processes can use
different communication backends and are not restricted to being executed on the same machine.
"""
import time
from typing import Tuple
import torch
from torch import nn, optim
import torch.distributed as dist
import torch.multiprocessing as mp
from torch.nn.parallel import DistributedDataParallel as DDP
import os
num_epochs = 5
batch_size = 8
Din, Dout = 10, 5
data_x = torch.randn(batch_size, Din)
data_y = torch.randn(batch_size, Dout)
data = [(i*data_x, i*data_y) for i in range(num_epochs)]
class PerDeviceModel(nn.Module):
"""
Toy example for a model ran in parallel but not distributed accross gpus
(only processes with their own gpu or hardware)
"""
def __init__(self):
super().__init__()
self.net1 = nn.Linear(Din, Din)
self.relu = nn.ReLU()
self.net2 = nn.Linear(Din, Dout)
def forward(self, x):
return self.net2(self.relu(self.net1(x)))
def setup_process(rank, world_size, backend='gloo'):
"""
Initialize the distributed environment (for each process).
gloo: is a collective communications library (https://github.com/facebookincubator/gloo). My understanding is that
it's a library/API for process to communicate/coordinate with each other/master. It's a backend library.
"""
# set up the master's ip address so this child process can coordinate
# os.environ['MASTER_ADDR'] = '127.0.0.1'
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '12355'
# - use NCCL if you are using gpus: https://pytorch.org/tutorials/intermediate/dist_tuto.html#communication-backends
if torch.cuda.is_available():
backend = 'nccl'
# Initializes the default distributed process group, and this will also initialize the distributed package.
dist.init_process_group(backend, rank=rank, world_size=world_size)
def cleanup():
""" Destroy a given process group, and deinitialize the distributed package """
dist.destroy_process_group()
def get_batch(batch: Tuple[torch.Tensor, torch.Tensor], rank):
x, y = batch
if torch.cuda.is_available():
x, y = x.to(rank), y.to(rank)
else:
x, y = x.share_memory_(), y.share_memory_()
return x, y
def get_ddp_model(model: nn.Module, rank):
"""
Moves the underlying storage to shared memory.
This is a no-op if the underlying storage is already in shared memory
and for CUDA tensors. Tensors in shared memory cannot be resized.
:return:
TODO: does this have to be done outside or inside the process? my guess is that it doesn't matter because
1) if its on gpu once it's on the right proc it moves it to cpu with id rank via mdl.to(rank)
2) if it's on cpu then mdl.share_memory() or data.share_memory() is a no op if it's already in shared memory o.w.
"""
# if gpu avail do the standard of creating a model and moving the model to the GPU with id rank
if torch.cuda.is_available():
# create model and move it to GPU with id rank
model = model.to(rank)
ddp_model = DDP(model, device_ids=[rank])
else:
# if we want multiple cpu just make sure the model is shared properly accross the cpus with shared_memory()
# note that op is a no op if it's already in shared_memory
model = model.share_memory()
ddp_model = DDP(model) # I think removing the devices ids should be fine...?
return ddp_model
# return OneDeviceModel().to(rank) if torch.cuda.is_available() else OneDeviceModel().share_memory()
def run_parallel_training_loop(rank, world_size):
"""
Distributed function to be implemented later.
This is the function that is actually ran in each distributed process.
Note: as DDP broadcasts model states from rank 0 process to all other processes in the DDP constructor,
you don’t need to worry about different DDP processes start from different model parameter initial values.
"""
setup_process(rank, world_size)
print()
print(f"Start running DDP with model parallel example on rank: {rank}.")
print(f'current process: {mp.current_process()}')
print(f'pid: {os.getpid()}')
# get ddp model
model = PerDeviceModel()
ddp_model = get_ddp_model(model, rank)
# do training
for batch_idx, batch in enumerate(data):
x, y = get_batch(batch, rank)
loss_fn = nn.MSELoss()
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
optimizer.zero_grad()
outputs = ddp_model(x)
# Gradient synchronization communications take place during the backward pass and overlap with the backward computation.
loss_fn(outputs, y).backward() # When the backward() returns, param.grad already contains the synchronized gradient tensor.
optimizer.step() # TODO how does the optimizer know to do the gradient step only once?
print()
print(f"Start running DDP with model parallel example on rank: {rank}.")
print(f'current process: {mp.current_process()}')
print(f'pid: {os.getpid()}')
# Destroy a given process group, and deinitialize the distributed package
cleanup()
def main():
print()
print('running main()')
print(f'current process: {mp.current_process()}')
print(f'pid: {os.getpid()}')
# args
if torch.cuda.is_available():
world_size = torch.cuda.device_count()
else:
world_size = mp.cpu_count()
print(f'world_size={world_size}')
mp.spawn(run_parallel_training_loop, args=(world_size,), nprocs=world_size)
if __name__ == "__main__":
print('starting __main__')
start = time.time()
main()
print(f'execution length = {time.time() - start}')
print('Done!\a\n')
cross posted: https://discuss.pytorch.org/t/why-is-mp-spawn-spawning-4-processes-when-i-only-want-2/112299

I ran your "(minimal) code example" without any change and any error on a server with 4 GPUs (python version: 3.6.9, and pytorch version: 1.5.0+cu101 ).
Does the problem still exist when you run the minimal code example?
If so, and if you are on a linux machine, could you please run the following code instead, and tell me what output you get:
import os
import torch
import torch.distributed as dist
import torch.multiprocessing as mp
import torch.nn as nn
import torch.optim as optim
from torch.nn.parallel import DistributedDataParallel as DDP
def get_visible_gpus():
ns = os.popen('nvidia-smi')
lines_ns = ns.readlines()
# print(lines_ns)
for _i, _line in enumerate(lines_ns):
if _line.find('|=') >= 0:
break
line_gpus = lines_ns[_i:]
for _i, _line in enumerate(line_gpus):
if _line.find('Processes') >= 0:
break
line_gpus = line_gpus[:_i-3]
# print(line_gpus)
idx_gpu_lines = []
for _i, _line in enumerate(line_gpus):
if _line.find('+') >= 0:
idx_gpu_lines.append(_i+1)
idx_gpus = []
for _line_gpu in idx_gpu_lines:
idx_gpus.append(int(line_gpus[_line_gpu].split()[1]))
# print(idx_gpus)
return idx_gpus
def example(rank, world_size):
print('rank:{}'.format(rank))
# create default process group
os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '8888'
dist.init_process_group("gloo", rank=rank, world_size=world_size)
# create local model
model = nn.Linear(10, 10).to(rank)
# construct DDP model
ddp_model = DDP(model, device_ids=[rank])
# define loss function and optimizer
loss_fn = nn.MSELoss()
optimizer = optim.SGD(ddp_model.parameters(), lr=0.001)
# forward pass
outputs = ddp_model(torch.randn(20, 10).to(rank))
labels = torch.randn(20, 10).to(rank)
# backward pass
loss_fn(outputs, labels).backward()
# update parameters
optimizer.step()
def main():
# world_size = 2
world_size = torch.cuda.device_count()
print('world_size:{}'.format(world_size))
print('get_visible_gpus():{}'.format(get_visible_gpus()))
mp.spawn(example,
args=(world_size,),
nprocs=world_size,
join=True)
if __name__ == "__main__":
print(torch.__version__)
main()
print('Done\n\a')
In my case, I simply get:
1.5.0+cu101
world_size:4
get_visible_gpus():[0, 1, 2, 3]
rank:1
rank:3
rank:0
rank:2
Done
get_visible_gpus() is simply text parsing an nvidia-smi shell cmd to get the ids of the gpus that cuda can see.
NB: Please excuse me, I would have commented instead of "answering" -as I am not directly solving your problem, but asking for more details- but my reputation is not good enough T.T

Solution: increase shm-size
docker run -it \
--shm-size=64g
Reason:
If you run on docker container, it's probably because the shm_size of docker is not large enough. By default, Docker containers are allocated 64 MB of shared memory. This shared memory is not a memory limit, but a /dev/shm temporary file storage file system that uses RAM to store files. This is used for IPC.
For check shm-size. After entering the container you can use df to view the shm size.

AttributeError: module 'keras.layers' has no attribute 'regularizers'

I try to run a Python project from Git for the first time. After having installed all needed librariers I started the program and I received following error:
AttributeError: module 'keras.layers' has no attribute 'regularizers'
Source:
from keras import layers, models, optimizers
from keras import backend as K
class Actor:
'''
Actor(policy) Model
'''
def __init__(self, state_size, action_size):
self.state_size = state_size
self.action_size = action_size
self.build_model()
def build_model(self):
states = layers.Input(shape=(self.state_size, ), name='states')
net = layers.Dense(units=16, kernel_regularizer=layers.regularizers.l2(1e-6))(states)
net = layers.BatchNormalization()(net)
net = layers.Activation('relu')(net)
net = layers.Dense(units=32, kernel_regularizer=layers.regularizers.l2(1e-6))(net)
net = layers.BatchNormalization()(net)
net = layers.Activation('relu')(net)
actions = layers.Dense(units=self.action_size, activation='softmax', name='actions')(net)
self.model = models.Model(inputs=states, outputs=actions)
action_gradients = layers.Input(shape=(self.action_size, ))
loss = K.mean(-action_gradients * actions)
optimizer = optimizers.Adam(lr=.0001)
updates_op = optimizer.get_updates(params=self.model.trainable_weights, loss=loss)
self.train_fn = K.function(
inputs = [self.model.input, action_gradients, K.learning_phase()],
outputs = [],
updates = updates_op
)

I think the program was written in TensorFlow version 1 and also keras version 1. You may be solved the problem by writing that way
net = layers.Dense(32, kernel_regularizer=tensorflow.keras.regularizers.L1(1e-6))(states)
but will occur so many errors. As there used some functionality that has been removed in later versions such as
optimizer.get_updates
no longer works. So, better work with later version

Create a temporary keras session?

I have a bunch of models floating around, I clone them, cross-validate them, do hyperparameter selection and what have you. As such, my keras global session can get quite mucked up. The solution per various threads is to call .clear_session(). However, this will throw away any models that I want to keep. One option is to train all of my models in a multiprocessing thread. However, it would be convenient to just instantiate a new session for each model as one might do with Tensorflow:
def score_model(**hyperparameters):
with tf.Graph().as_default()
my_model = build_model(**hyperparameters)
with tf.Session() as sess:
my_model.train(X,y)
score = my_model.score()
# now it's all gone, I have the score, so I don't need the model anymore
# the rest of my_model should get garbage collected, hooray!
return score
Can I do this sort of thing with keras?
UPDATE
The sess.as_default() method is crashing my kernel. My memory does not seem to be running low, and it gives no error whatsoever. In the following loop I can't even make it to i=2 before crashing.
from sklearn.datasets import load_iris
import numpy as np
import sklearn
import keras
import keras.wrappers.scikit_learn
import tensorflow as tf
import keras.models
import os
def sessioned(f):
def sessioned_f(self, *args, **kwargs):
if not hasattr(self, "sess"):
self.sess = tf.Session()
with self.sess.as_default():
return f(self, *args, **kwargs)
return result
return sessioned_f
class LogisticRegression(keras.wrappers.scikit_learn.KerasClassifier):
def __init__(self, n_epochs=100, **kwargs):
self.n_epochs = n_epochs
super().__init__(**kwargs)
#sessioned
def fit(self, X, y,**kwargs):
# get the shape of X and one hot y
self.input_shape = X.shape[-1]
self.label_encoder = sklearn.preprocessing.LabelEncoder()
self.label_encoder.fit(y)
self.output_shape = len(self.label_encoder.classes_)
label_encoded = self.label_encoder.transform(y).reshape((-1,1))
y_onehot = sklearn.preprocessing.OneHotEncoder().fit_transform(label_encoded).toarray()
super().fit(X,y_onehot,epochs=self.n_epochs,verbose=1,**kwargs)
return self
#sessioned
def predict_proba(self, X):
return super().predict_proba(X)
def check_params(self, params):
#fuckit
pass
#sessioned
def __call__(self): # the build_fn thing
# create model
model = keras.models.Sequential()
model.add(keras.layers.Dense(self.output_shape, input_dim=self.input_shape, kernel_initializer="normal", activation="softmax"))
# Compile model
model.compile(loss='categorical_crossentropy', optimizer='adam')
return model
data = load_iris()
i=0
while True:
print(i)
graph = tf.Graph()
with graph.as_default():
model = LogisticRegression()
model.fit(data.data, data.target)
model.sess.close()
del model
i+=1
del graph

You can use Keras exactly as you described, except instead of running Tensorflow code inside the with statements you run the Keras code.
To set the session you would use
with sess.as_default()
Here is a link with with more information:
https://blog.keras.io/keras-as-a-simplified-interface-to-tensorflow-tutorial.html
I have also found it helpful to look at the source code inside keras.backend. If you look at get_session() you can see that Keras first looks to see if there is a tensorflow default session. Otherwise it uses the session set to Keras using set_session(). Finally if no session has been set then it creates one.

TensorRT for upscaling images serves not expected results

I'm fighting with TensorRT (TensorRT 4 for python right now) since several weeks. I passed a lot of problems to get TensorRT running. The example code from NVIDIA works well for me :
TensorRT MNIST example
Now, i created my own network in tensorflow (a very simple one) for upscaling images, let's say (in HWC) 320x240x3 into 640x480x3 .The usual way by creating a frozen-graph and running an inferencer just based on Tensorflow gave me expected results but not by using TensorRT.
I have a strange feeling about that i made something wrong by feeding the images into the GPU-memory (This would be probably an issue about pycuda and/or TensorRT).
The worst case scenario would be that TensorRT destroys my network by the optimization process.
I hope someone has just a little idea for saving my life.
This is my Tensorflow-model (i just wrapped the functions):
net = conv2d(input,
64,
k_size=3,
activation=tf.nn.relu,
name='conv1')
net = deconv2d(net,
3,
k_size=5,
activation=tf.tanh,
stride=self.params.resize_factor,
scale=self.params.resize_factor,
name='deconv')
This is the important snippet of my inferencer:
import tensorrt as trt
import uff
from tensorrt.parsers import uffparser
import pycuda.driver as cuda
import numpy as np
...
def _init_infer(self, uff_model):
g_logger = trt.infer.ConsoleLogger(trt.infer.LogSeverity.ERROR)
parser = uffparser.create_uff_parser()
parser.register_input(self.input_node, (self.channels, self.height, self.width), 0)
parser.register_output(self.output_node)
self.engine = trt.utils.uff_to_trt_engine(g_logger, uff_model, parser, self.max_batch_size,
self.max_workspace_size)
parser.destroy()
self.runtime = trt.infer.create_infer_runtime(g_logger)
self.context = self.engine.create_execution_context()
self.output = np.empty(self.output_size, dtype=self.dtype)
# create CUDA stream
self.stream = cuda.Stream()
# allocate device memory
self.d_input = cuda.mem_alloc(self.channels * self.max_batch_size * self.width *
self.height * self.output.dtype.itemsize)
self.d_output = cuda.mem_alloc(self.output_size * self.output.dtype.itemsize)
self.bindings = [int(self.d_input), int(self.d_output)]
def infer(self, input_batch, batch_size=1):
# transfer input data to device
cuda.memcpy_htod_async(self.d_input, input_batch, self.stream)
# execute model
self.context.enqueue(batch_size, self.bindings, self.stream.handle, None)
# transfer predictions back
cuda.memcpy_dtoh_async(self.output, self.d_output, self.stream)
# synchronize threads
self.stream.synchronize()
return self.output
And the executable snippet:
...
# create trt inferencer
trt_inferencer = TensorRTInferencer(params=params)
img = [misc.imread('./test_images/lion.png')]
img[0] = normalize(img[0])
img = img[0]
# inferencing method
result = trt_inferencer.infer(img)
result = inormalize(result, dtype=np.uint8)
result = result.reshape(1, params.height * 2, params.width * 2, 3)
...
And the weird result by comparison :(
upscaled lion TensorRT, Tensorflow, Original

I got it now, finally. The problem was a wrong dimension and order of the input images and output. And for everyone who run into the same problem, this is the adopted executable snippet, dependent on my initialization:
...
# create trt inferencer
trt_inferencer = TensorRTInferencer(params=params)
img = [misc.imread('./test_images/lion.png')]
img[0] = normalize(img[0])
img = img[0]
img = np.transpose(img, (2, 0, 1))
img = img.ravel()
# inferencing method
result = trt_inferencer.infer(img)
result = inormalize(result, dtype=np.uint8)
result = np.reshape(result, newshape=[3, params.height * 2, params.width * 2])
result = np.transpose(result, (1, 2, 0))
...

How to classify images using Spark and Caffe

I am using Caffe to do image classification, can I am using MAC OS X, Pyhton.
Right now I know how to classify a list of images using Caffe with Spark python, but if I want to make it faster, I want to use Spark.
Therefore, I tried to apply the image classification on each element of an RDD, the RDD created from a list of image_path. However, Spark does not allow me to do so.
Here is my code:
This is the code for image classification:
# display image name, class number, predicted label
def classify_image(image_path, transformer, net):
image = caffe.io.load_image(image_path)
transformed_image = transformer.preprocess('data', image)
net.blobs['data'].data[...] = transformed_image
output = net.forward()
output_prob = output['prob'][0]
pred = output_prob.argmax()
labels_file = caffe_root + 'data/ilsvrc12/synset_words.txt'
labels = np.loadtxt(labels_file, str, delimiter='\t')
lb = labels[pred]
image_name = image_path.split(images_folder_path)[1]
result_str = 'image: '+image_name+' prediction: '+str(pred)+' label: '+lb
return result_str
This this the code generates Caffe parameters and apply the classify_image method on each element of the RDD:
def main():
sys.path.insert(0, caffe_root + 'python')
caffe.set_mode_cpu()
model_def = caffe_root + 'models/bvlc_reference_caffenet/deploy.prototxt'
model_weights = caffe_root + 'models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel'
net = caffe.Net(model_def,
model_weights,
caffe.TEST)
mu = np.load(caffe_root + 'python/caffe/imagenet/ilsvrc_2012_mean.npy')
mu = mu.mean(1).mean(1)
transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})
transformer.set_transpose('data', (2,0,1))
transformer.set_mean('data', mu)
transformer.set_raw_scale('data', 255)
transformer.set_channel_swap('data', (2,1,0))
net.blobs['data'].reshape(50,
3,
227, 227)
image_list= []
for image_path in glob.glob(images_folder_path+'*.jpg'):
image_list.append(image_path)
images_rdd = sc.parallelize(image_list)
transformer_bc = sc.broadcast(transformer)
net_bc = sc.broadcast(net)
image_predictions = images_rdd.map(lambda image_path: classify_image(image_path, transformer_bc, net_bc))
print image_predictions
if __name__ == '__main__':
main()
As you can see, here I tried to broadcast the caffe parameters, transformer_bc = sc.broadcast(transformer), net_bc = sc.broadcast(net)
The error is:
RuntimeError: Pickling of "caffe._caffe.Net" instances is not enabled
Before I am doing the broadcast, the error was :
Driver stacktrace.... Caused by: org.apache.spark.api.python.PythonException: Traceback (most recent call last):....
So, do you know, is there any way I can classify images using Caffe and Spark but also take advantage of Spark?

When you work with complex, non-native objects initialization has to moved directly to the workers for example with singleton module:
net_builder.py:
import cafe
net = None
def build_net(*args, **kwargs):
... # Initialize net here
return net
def get_net(*args, **kwargs):
global net
if net is None:
net = build_net(*args, **kwargs)
return net
main.py:
import net_builder
sc.addPyFile("net_builder.py")
def classify_image(image_path, transformer, *args, **kwargs):
net = net_builder.get_net(*args, **kwargs)
It means you'll have to distribute all required files as well. It can be done either manually or using SparkFiles mechanism.
On a side note you should take a look at the SparkNet package.