For the main.py of the px2graph project, the part of training and validation is shown as below:
splits = [s for s in ['train', 'valid'] if opt.iters[s] > 0]
start_round = opt.last_round - opt.num_rounds
# Main training loop
for round_idx in range(start_round, opt.last_round):
for split in splits:
print("Round %d: %s" % (round_idx, split))
loader.start_epoch(sess, split, train_flag, opt.iters[split] * opt.batchsize)
flag_val = split == 'train'
for step in tqdm(range(opt.iters[split]), ascii=True):
global_step = step + round_idx * opt.iters[split]
to_run = [sample_idx, summaries[split], loss, accuracy]
if split == 'train': to_run += [optim]
# Do image summaries at the end of each round
do_image_summary = step == opt.iters[split] - 1
if do_image_summary: to_run[1] = image_summaries[split]
# Start with lower learning rate to prevent early divergence
t = 1/(1+np.exp(-(global_step-5000)/1000))
lr_start = opt.learning_rate / 15
lr_end = opt.learning_rate
tmp_lr = (1-t) * lr_start + t * lr_end
# Run computation graph
result = sess.run(to_run, feed_dict={train_flag:flag_val, lr:tmp_lr})
out_loss = result[2]
out_accuracy = result[3]
if sum(out_loss) > 1e5:
print("Loss diverging...exiting before code freezes due to NaN values.")
print("If this continues you may need to try a lower learning rate, a")
print("different optimizer, or a larger batch size.")
return
time_str = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
print("{}: step {}, loss {:g}, acc {:g}".format(time_str, global_step, out_loss, out_accuracy))
# Log data
if split == 'valid' or (split == 'train' and step % 20 == 0) or do_image_summary:
writer.add_summary(result[1], global_step)
writer.flush()
# Save training snapshot
saver.save(sess, 'exp/' + opt.exp_id + '/snapshot')
with open('exp/' + opt.exp_id + '/last_round', 'w') as f:
f.write('%d\n' % round_idx)
It seems that the author only get the result of each batch of the validation set. I am wondering, if I want to observe whether the model is improving or reaching the best performance, should I use the result on the whole validation set?
If the validation set is small enough, we could calculate the loss, accuracy on the whole validation set during training to observe the performance. However, if the validation set is too large, it is better to calculate batch-wise validation results and for multiple steps.
Related
I'm self study of Python and it's my first code.
I'm working for analyze logs from the servers. Usually I need analyze full day logs. I created script (this is example, simple logic) just for check speed. If I use normal coding the duration of analyzing 20mil rows about 12-13 minutes. I need 200mil rows by 5 min.
What I tried:
Use multiprocessing (met issue with share memory, think that fix it). But as the result - 300K rows = 20 sec and no matter how many processes. (PS: Also need control processors count in advance)
Use threading (I found that it's not give any speed, 300K rows = 2 sec. But normal code same, 300K = 2 sec)
Use asyncio (I think that script is slow because need reads many files). Result same as threading - 300K = 2 sec.
Finally I think that all three my script incorrect and didn't work correctly.
PS: I try to avoid use specific python modules (like pandas) because in this case it will be more difficult to execute on different servers. Better to use common lib.
Please help to check 1st - multiprocessing.
import csv
import os
from multiprocessing import Process, Queue, Value, Manager
file = {"hcs.log", "hcs1.log", "hcs2.log", "hcs3.log"}
def argument(m, a, n):
proc_num = os.getpid()
a_temp_m = a["vod_miss"]
a_temp_h = a["vod_hit"]
with open(os.getcwd() + '/' + m, newline='') as hcs_1:
hcs_2 = csv.reader(hcs_1, delimiter=' ')
for j in hcs_2:
if j[3].find('MISS') != -1:
a_temp_m[n] = a_temp_m[n] + 1
elif j[3].find('HIT') != -1:
a_temp_h[n] = a_temp_h[n] + 1
a["vod_miss"][n] = a_temp_m[n]
a["vod_hit"][n] = a_temp_h[n]
if __name__ == '__main__':
procs = []
manager = Manager()
vod_live_cuts = manager.dict()
i = "vod_hit"
ii = "vod_miss"
cpu = 1
n = 1
vod_live_cuts[i] = manager.list([0] * cpu)
vod_live_cuts[ii] = manager.list([0] * cpu)
for m in file:
proc = Process(target=argument, args=(m, vod_live_cuts, (n-1)))
procs.append(proc)
proc.start()
if n >= cpu:
n = 1
proc.join()
else:
n += 1
[proc.join() for proc in procs]
[proc.close() for proc in procs]
I'm expect, each file by def argument will be processed by independent process and finally all results will be saved in dict vod_live_cuts. For each process I added independent list in dict. I think it will help cross operation for use this parameter. But maybe it's wrong way :(
using IPC is costly, so only use "shared objects" for saving the final result, not for intermediate results while parsing the file.
limiting the number of processes is done by using a multiprocessing.Pool, the following code uses it to reach the max hard-disk speed, you only need to post-process the results.
you can only parse data as fast as your HDD can read it (typically 30-80 MB/s), so if you need to improve the performance further you should use SSD or RAID0 for higher disk speed, you cannot get much faster than this without changing your hardware.
import csv
import os
from multiprocessing import Process, Queue, Value, Manager, Pool
file = {"hcs.log", "hcs1.log", "hcs2.log", "hcs3.log"}
def argument(m, a):
proc_num = os.getpid()
a_temp_m_n = 0 # make it local to process
a_temp_h_n = 0 # as shared lists use IPC
with open(os.getcwd() + '/' + m, newline='') as hcs_1:
hcs_2 = csv.reader(hcs_1, delimiter=' ')
for j in hcs_2:
if j[3].find('MISS') != -1:
a_temp_m_n = a_temp_m_n + 1
elif j[3].find('HIT') != -1:
a_temp_h_n = a_temp_h_n + 1
a["vod_miss"].append(a_temp_m_n)
a["vod_hit"].append(a_temp_h_n)
if __name__ == '__main__':
manager = Manager()
vod_live_cuts = manager.dict()
i = "vod_hit"
ii = "vod_miss"
cpu = 1
vod_live_cuts[i] = manager.list()
vod_live_cuts[ii] = manager.list()
with Pool(cpu) as pool:
tasks = []
for m in file:
task = pool.apply_async(argument, args=(m, vod_live_cuts))
tasks.append(task)
for task in tasks:
task.get()
print(list(vod_live_cuts[i]))
print(list(vod_live_cuts[ii]))
I am training a VAE (using federated learning, but that is not so important) and wanted to keep the loss and train functions simple to exchange. The initial approach was to have a tf.function as loss function and a tf.function as train function as follows:
#tf.function
def kl_reconstruction_loss(model, model_input, beta):
x, y = model_input
mean, logvar = model.encode(x, y)
z = model.reparameterize(mean, logvar)
x_logit = model.decode(z, y)
cross_ent = tf.nn.sigmoid_cross_entropy_with_logits(logits=x_logit, labels=x)
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(cross_ent, axis=[1, 2, 3]), axis=0)
kl_loss = tf.reduce_mean(0.5 * tf.reduce_sum(tf.exp(logvar) + tf.square(mean) - 1. - logvar, axis=-1), axis=0)
loss = reconstruction_loss + beta * kl_loss
return loss, kl_loss, reconstruction_loss
#tf.function
def train_fn(model: tf.keras.Model, batch, optimizer, kl_beta):
"""Trains the model on a single batch.
Args:
model: The VAE model.
batch: A batch of inputs [images, labels] for the vae.
optimizer: The optimizer to train the model.
beta: Weighting of KL loss
Returns:
The loss.
"""
def vae_loss():
"""Does the forward pass and computes losses for the generator."""
# N.B. The complete pass must be inside loss() for gradient tracing.
return kl_reconstruction_loss(model, batch, kl_beta)
with tf.GradientTape() as tape:
loss, kl_loss, rc_loss = vae_loss()
grads = tape.gradient(loss, model.trainable_variables)
grads_and_vars = zip(grads, model.trainable_variables)
optimizer.apply_gradients(grads_and_vars)
return loss
For my dataset this results in an epoch duration of approx. 25 seconds. However, since I have to call those functions directly in my code, I would have to enter different ones if I would want to try out different loss/train functions.
So, alternatively, I followed https://github.com/google-research/federated/tree/master/gans and wrapped the loss function in a class and the train function in another function. Now I have:
class VaeKlReconstructionLossFns(AbstractVaeLossFns):
#tf.function
def vae_loss(self, model, model_input, labels, global_round):
# KL Reconstruction loss
mean, logvar = model.encode(model_input, labels)
z = model.reparameterize(mean, logvar)
x_logit = model.decode(z, labels)
cross_ent = tf.nn.sigmoid_cross_entropy_with_logits(logits=x_logit, labels=model_input)
reconstruction_loss = tf.reduce_mean(tf.reduce_sum(cross_ent, axis=[1, 2, 3]), axis=0)
kl_loss = tf.reduce_mean(0.5 * tf.reduce_sum(tf.exp(logvar) + tf.square(mean) - 1. - logvar, axis=-1), axis=0)
loss = reconstruction_loss + self._get_beta(global_round) * kl_loss
if model.losses:
loss += tf.add_n(model.losses)
return loss, kl_loss, reconstruction_loss
def create_train_vae_fn(
vae_loss_fns: vae_losses.AbstractVaeLossFns,
vae_optimizer: tf.keras.optimizers.Optimizer):
"""Create a function that trains VAE, binding loss and optimizer.
Args:
vae_loss_fns: Instance of gan_losses.AbstractVAELossFns interface,
specifying the VAE training loss.
vae_optimizer: Optimizer for training the VAE.
Returns:
Function that executes one step of VAE training.
"""
# We check that the optimizer has not been used previously, which ensures
# that when it is bound the train fn isn't holding onto a different copy of
# the optimizer variables then the copy that is being exchanged b/w server and
# clients.
if vae_optimizer.variables():
raise ValueError(
'Expected vae_optimizer to not have been used previously, but '
'variables were already initialized.')
#tf.function
def train_vae_fn(model: tf.keras.Model,
model_inputs,
labels,
global_round,
new_optimizer_state=None):
"""Trains the model on a single batch.
Args:
model: The VAE model.
model_inputs: A batch of inputs (usually images) for the VAE.
labels: A batch of labels corresponding to the inputs.
global_round: The current glob al FL round for beta calculation
new_optimizer_state: A possible optimizer state to overwrite the current one with.
Returns:
The number of examples trained on.
The loss.
The updated optimizer state.
"""
def vae_loss():
"""Does the forward pass and computes losses for the generator."""
# N.B. The complete pass must be inside loss() for gradient tracing.
return vae_loss_fns.vae_loss(model, model_inputs, labels, global_round)
# Set optimizer vars
optimizer_state = get_optimizer_state(vae_optimizer)
if new_optimizer_state is not None:
# if optimizer is uninitialised, initialise vars
try:
tf.nest.assert_same_structure(optimizer_state, new_optimizer_state)
except ValueError:
initialize_optimizer_vars(vae_optimizer, model)
optimizer_state = get_optimizer_state(vae_optimizer)
tf.nest.assert_same_structure(optimizer_state, new_optimizer_state)
tf.nest.map_structure(lambda a, b: a.assign(b), optimizer_state, new_optimizer_state)
with tf.GradientTape() as tape:
loss, kl_loss, rc_loss = vae_loss()
grads = tape.gradient(loss, model.trainable_variables)
grads_and_vars = zip(grads, model.trainable_variables)
vae_optimizer.apply_gradients(grads_and_vars)
return tf.shape(model_inputs)[0], loss, optimizer_state
return train_vae_fn
This new formulation takes about 86 seconds per epoch.
I am struggling to understand why the second version performs so much worse than the first one. Does anyone have a good explanation for this?
Thanks in advance!
EDIT: My Tensorflow version is 2.5.0
I copied the idea from mnslarcher and wrote the following two functions for my keypoint detector (resnet50 backbone) algorithm.
def build_valid_loader(cfg):
_cfg = cfg.clone()
_cfg.defrost() # make this cfg mutable.
_cfg.DATASETS.TRAIN = cfg.DATASETS.TEST
return build_detection_train_loader(_cfg)
def store_valid_loss(model, data, storage):
training_mode = model.training
with torch.no_grad():
loss_dict = model(data)
losses = sum(loss_dict.values())
assert torch.isfinite(losses).all(), loss_dict
loss_dict_reduced = {k: v.item()
for k, v in comm.reduce_dict(loss_dict).items()}
losses_reduced = sum(loss for loss in loss_dict_reduced.values())
if comm.is_main_process():
storage.put_scalars(val_loss=losses_reduced, **loss_dict_reduced)
model.train(training_mode)
then in plain_train_net.py I am calling them as bellow.
val_data_loader = build_valid_loader(cfg)
logger.info("Starting training from iteration {}".format(start_iter))
with EventStorage(start_iter) as storage:
for data, val_data, iteration in zip(data_loader, val_data_loader, range(start_iter, max_iter)):
iteration = iteration + 1
..
..
#At the end of the for loop.
# Calculate and log validation loss.
store_valid_loss(model, val_data, storage)
after 1k iteration, loss_keypoint is increasing, but total_loss is same compared to without store_valid_loss call. What am I missing? Can anyone please help to understand?
I am using 4 GeForce RTX 2080 Ti.
I am working on a deep learning problem. I am solving it using pytorch. I have two GPU's which are on the same machine (16273MiB,12193MiB). I want to use both the GPU's for my training (video dataset).
I get a warning:
There is an imbalance between your GPUs. You may want to exclude GPU 1 which
has less than 75% of the memory or cores of GPU 0. You can do so by setting
the device_ids argument to DataParallel, or by setting the CUDA_VISIBLE_DEVICES
environment variable.
warnings.warn(imbalance_warn.format(device_ids[min_pos], device_ids[max_pos]))
I also get an error:
raise TypeError('Broadcast function not implemented for CPU tensors')
TypeError: Broadcast function not implemented for CPU tensors
if __name__ == '__main__':
opt.scales = [opt.initial_scale]
for i in range(1, opt.n_scales):
opt.scales.append(opt.scales[-1] * opt.scale_step)
opt.arch = '{}-{}'.format(opt.model, opt.model_depth)
opt.mean = get_mean(opt.norm_value)
opt.std = get_std(opt.norm_value)
print("opt",opt)
with open(os.path.join(opt.result_path, 'opts.json'), 'w') as opt_file:
json.dump(vars(opt), opt_file)
torch.manual_seed(opt.manual_seed)
model, parameters = generate_model(opt)
#print(model)
pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print("Total number of trainable parameters: ", pytorch_total_params)
# Define Class weights
if opt.weighted:
print("Weighted Loss is created")
if opt.n_finetune_classes == 2:
weight = torch.tensor([1.0, 3.0])
else:
weight = torch.ones(opt.n_finetune_classes)
else:
weight = None
criterion = nn.CrossEntropyLoss()
if not opt.no_cuda:
criterion = nn.DataParallel(criterion.cuda())
if opt.no_mean_norm and not opt.std_norm:
norm_method = Normalize([0, 0, 0], [1, 1, 1])
elif not opt.std_norm:
norm_method = Normalize(opt.mean, [1, 1, 1])
else:
norm_method = Normalize(opt.mean, opt.std)
train_loader = torch.utils.data.DataLoader(
training_data,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.n_threads,
pin_memory=True)
train_logger = Logger(
os.path.join(opt.result_path, 'train.log'),
['epoch', 'loss', 'acc', 'precision','recall','lr'])
train_batch_logger = Logger(
os.path.join(opt.result_path, 'train_batch.log'),
['epoch', 'batch', 'iter', 'loss', 'acc', 'precision', 'recall', 'lr'])
if opt.nesterov:
dampening = 0
else:
dampening = opt.dampening
optimizer = optim.SGD(
parameters,
lr=opt.learning_rate,
momentum=opt.momentum,
dampening=dampening,
weight_decay=opt.weight_decay,
nesterov=opt.nesterov)
# scheduler = lr_scheduler.ReduceLROnPlateau(
# optimizer, 'min', patience=opt.lr_patience)
if not opt.no_val:
spatial_transform = Compose([
Scale(opt.sample_size),
CenterCrop(opt.sample_size),
ToTensor(opt.norm_value), norm_method
])
print('run')
for i in range(opt.begin_epoch, opt.n_epochs + 1):
if not opt.no_train:
adjust_learning_rate(optimizer, i, opt.lr_steps)
train_epoch(i, train_loader, model, criterion, optimizer, opt,
train_logger, train_batch_logger)
I have also made changes in my train file:
model = nn.DataParallel(model(),device_ids=[0,1]).cuda()
outputs = model(inputs)
It does not seem to work properly and is giving error. Please advice, I am new to pytorch.
Thanks
As mentioned in this link, you have to do model.cuda() before passing it to nn.DataParallel.
net = nn.DataParallel(model.cuda(), device_ids=[0,1])
https://github.com/pytorch/pytorch/issues/17065
I have to deal with very big data (Point clouds generally more than 30 000 000 points) using Matlab. I can read ascii data using textscan function. After reading, I need to detect invalid data (points with 0,0,0 coordinates) and then I need to do some mathematical operations on each point or each line in the data. In my way, first I read data with textscan and then I assign this data to a matrix. Secondly, I use for loops for detecting invalid points and doing some mathematical operations on each point or line in the data. A sample of my code is shown as below. According to profile tool of Matlab textscan takes 37% and line
transformed_list((i:i),(1:4)) = coordinate_list((i:i),(1:4))*t_matrix;
takes 35% of all computation time.
I tried it with another point cloud (stores around 5 500 000) and profile tool reported same results. Is there a way of avoiding for loops, or is there another way of speeding up this computation?
fileID = fopen('C:\Users\Mustafa\Desktop\ptx_all_data\dede5.ptx');
original_data = textscan(fileID,'%f %f %f %f %f %f %f', 'delimiter',' ');
fclose(fileID);
column = original_data{1}(1);
row = original_data{1}(2);
t_matrix = [original_data{1}(7) original_data{2}(7) original_data{3}(7) original_data{4}(7)
original_data{1}(8) original_data{2}(8) original_data{3}(8) original_data{4}(8)
original_data{1}(9) original_data{2}(9) original_data{3}(9) original_data{4}(9)
original_data{1}(10) original_data{2}(10) original_data{3}(10) original_data{4}(10)];
coordinate_list(:,1) = original_data{1}(11:length(original_data{1}));
coordinate_list(:,2) = original_data{2}(11:length(original_data{2}));
coordinate_list(:,3) = original_data{3}(11:length(original_data{3}));
coordinate_list(:,4) = 0;
coordinate_list(:,5) = original_data{4}(11:length(original_data{4}));
transformed_list = zeros(length(coordinate_list),5);
for i = 1:length(coordinate_list)
if coordinate_list(i,1) == 0 && coordinate_list(i,2) == 0 && coordinate_list(i,3) == 0
transformed_list(i,:) = NaN;
else
%transformed_list(i,:) = coordinate_list(i,:)*t_matrix;
transformed_list((i:i),(1:4)) = coordinate_list((i:i),(1:4))*t_matrix;
transformed_list(i,5) = coordinate_list(i,5);
end
%i
end
Thanks in advance
for loops with conditional statements like those will take ages to run. But what Matlab lacks in loop speed it makes up with vectorization and indexing.
Let's try some logical indexing like this to solve the first step:
coordinate_list(coordinate_list(:,1) == 0 .* ...
coordinate_list(:,2) == 0 .* ...
coordinate_list(:,3) == 0)=nan;
And then vectorize the second statement:
transformed_list(:,(1:4)) = coordinate_list(:,(1:4))*t_matrix;
As EBH mentioned above this might be a bit heavy on your RAM. If it's more than your computer can handle asks yourself if the coordinates really have to be doubles, maybe single precision will do. If that still doesn't do, try slicing the vector and performing the operation in parts.
Small example to give you an idea because I had a 2million element point cloud around here:
In R2015a
transformed_list = zeros(length(coordinate_list),5);
tic
for i = 1:length(coordinate_list)
if coordinate_list(i,1) == 0 && coordinate_list(i,2) == 0 && coordinate_list(i,3) == 0
transformed_list(i,:) = NaN;
else
%transformed_list(i,:) = coordinate_list(i,:)*t_matrix;
transformed_list((i:i),(1:3)) = coordinate_list((i:i),(1:3))*t_matrix;
transformed_list(i,5) = 1;
end
%i
end
toc
Returns Elapsed time is 10.928142 seconds.
transformed_list=coordinate_list;
tic
coordinate_list(coordinate_list(:,1) == 0 .* ...
coordinate_list(:,2) == 0 .* ...
coordinate_list(:,3) == 0)=nan;
transformed_list(:,(1:3)) = coordinate_list(:,(1:3))*t_matrix;
toc
Returns Elapsed time is 0.101696 seconds.
Rather than read the whole file, you'd be better off using a loop with
fscanf(fileID, '%f', 7)
and processing input as you read it.