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.
Related
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 guess i have made something in folowing simple neural network with PyTorch, because this runs much slower with CUDA then in CPU, can you find the mistake pls. The using function like
def backward(ctx, input):
return backward_sigm(ctx, input)
seems have no real impact on preformance
import torch
import torch.nn as nn
import torch.nn.functional as f
dname = 'cuda:0'
dname = 'cpu'
device = torch.device(dname)
print(torch.version.cuda)
def forward_sigm(ctx, input):
sigm = 1 / (1 + torch.exp(-input))
ctx.save_for_backward(sigm)
return sigm
def forward_step(ctx, input):
return torch.tensor(input > 0.5, dtype = torch.float32, device = device)
def backward_sigm(ctx, grad_output):
sigm, = ctx.saved_tensors
return grad_output * sigm * (1-sigm)
def backward_step(ctx, grad_output):
return grad_output
class StepAF(torch.autograd.Function):
#staticmethod
def forward(ctx, input):
return forward_sigm(ctx, input)
#staticmethod
def backward(ctx, input):
return backward_sigm(ctx, input)
#else return grad_output
class StepNN(torch.nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(StepNN, self).__init__()
self.linear1 = torch.nn.Linear(input_size, hidden_size)
#self.linear1.cuda()
self.linear2 = torch.nn.Linear(hidden_size, output_size)
#self.linear2.cuda()
#self.StepAF = StepAF.apply
def forward(self,x):
h_line_1 = self.linear1(x)
h_thrash_1 = StepAF.apply(h_line_1)
h_line_2 = self.linear2(h_thrash_1)
output = StepAF.apply(h_line_2)
return output
inputs = torch.tensor( [[1,0,1,0],[1,0,0,1],[0,1,0,1],[0,1,1,0],[1,0,0,0],[0,0,0,1],[1,1,0,1],[0,1,0,0],], dtype = torch.float32, device = device)
expected = torch.tensor( [[1,0,0],[1,0,0],[0,1,0],[0,1,0],[1,0,0],[0,0,1],[0,1,0],[0,0,1],], dtype = torch.float32, device = device)
nn = StepNN(4,8,3)
#print(*(x for x in nn.parameters()))
criterion = torch.nn.MSELoss(reduction='sum')
optimizer = torch.optim.SGD(nn.parameters(), lr=1e-3)
steps = 50000
print_steps = steps // 20
good_loss = 1e-5
for t in range(steps):
output = nn(inputs)
loss = criterion(output, expected)
if t % print_steps == 0:
print('step ',t, ', loss :' , loss.item())
if loss < good_loss:
print('step ',t, ', loss :' , loss.item())
break
optimizer.zero_grad()
loss.backward()
optimizer.step()
test = torch.tensor( [[0,1,0,1],[0,1,1,0],[1,0,1,0],[1,1,0,1],], dtype = torch.float32, device=device)
print(nn(test))
Unless you have large enough data, you won't see any performance improvement while using GPU. The problem is that GPUs use parallel processing, so unless you have large amounts of data, the CPU can process the samples almost as fast as the GPU.
As far as I can see in your example, you are using 8 samples of size (4, 1). I would imagine maybe when having over hundreds or thousands of samples, then you would see the performance improvement on a GPU. In your case, the sample size is (4, 1), and the hidden layer size is 8, so the CPU can perform the calculations fairly quickly.
There are lots of example notebooks online of people using MNIST data (it has around 60000 images for training), so you could load one in maybe Google Colab and then try training on the CPU and then on GPU and observe the training times. You could try this link for example. It uses TensorFlow instead of PyTorch but it will give you an idea of the performance improvement of a GPU.
Note : If you haven't used Google Colab before, then you need to change the runtime type (None for CPU and GPU for GPU) in the runtime menu at the top.
Also, I will post the results from this notebook here itself (look at the time mentioned in the brackets, and if you run it, you can see firsthand how fast it runs) :
On CPU :
INFO:tensorflow:loss = 294.3736, step = 1
INFO:tensorflow:loss = 28.285727, step = 101 (23.769 sec)
INFO:tensorflow:loss = 23.518856, step = 201 (24.128 sec)
On GPU :
INFO:tensorflow:loss = 295.08328, step = 0
INFO:tensorflow:loss = 47.37291, step = 100 (4.709 sec)
INFO:tensorflow:loss = 23.31364, step = 200 (4.581 sec)
INFO:tensorflow:loss = 9.980572, step = 300 (4.572 sec)
INFO:tensorflow:loss = 17.769928, step = 400 (4.560 sec)
INFO:tensorflow:loss = 16.345463, step = 500 (4.531 sec)
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.
I have a simple model created with Keras and I need to measure the execution time for prediction per image. Right now I just do this:
start = time.clock()
my_model.predict(images_test)
end = time.clock()
print("Time per image: {} ".format((end-start)/len(images_test)))
But I noticed that the calculated time is bigger when len(images_test) is smaller. For example when len(images_test) = 32 I get: 0.06 and when len(images_test) = 1024 I get: 0.006
Is there a "right" way to do this ?
if use TF it seems no Asynchronous problem
but if use pytorch it has Asynchronous problem.
in TF:
start = time.clock()
result = my_model.predict(images_test)
end = time.clock()
in pytorch:
torch.cuda.synchronize()
start = time.clock()
my_model.predict(images_test)
torch.cuda.synchronize()
end = time.clock()
But i think you can do 10 times Loop model_predict
and print time_list
(computer need load keras model so first time load slower than other times )
in TF:
pred_time_list=[]
for i in range(10):
start = time.clock()
result = my_model.predict(images_test)
end = time.clock()
pred_time_list.append(end-start)
print(pred_time_list)
(print the pred_time_list and you may find out why the times incorrect)
Reference:
[1]
https://discuss.pytorch.org/t/doing-qr-decomposition-on-gpu-is-much-slower-than-on-cpu/21213/6
[2]
https://discuss.pytorch.org/t/is-there-any-code-torch-backends-cudnn-benchmark-torch-cuda-synchronize-similar-in-tensorflow/51484/2
I am currently runnuing training in matlab on a matrix of logspecrum samples I am constantly dealing with underflow problems.I understood that I need to work with log's in order to deal with underflowing.
I am still strugling with uderflow though , when i calculate the mean (mue) bucause it is negetive i cant work with logs so i need the real values that underflow.
These are equasions i am working with:
In MATLAB code i calulate log_tau in oreder avoid underflow but when calulating mue i need exp(log(tau)) which goes to zero.
I am attaching relevent MATLAB code
**in the code i called the variable alpha is tau ...
for i = 1 : 50
log_c = Logsum(log_alpha,1) - log(N);
c = exp(log_c);
mue = DataMat*alpha./(repmat(exp(Logsum(log_alpha,1)),FrameSize,1));
log_abs_mue = log(abs(mue));
log_SigmaSqr = log((DataMat.^2)*alpha) - repmat(Logsum(log_alpha,1),FrameSize,1) - 2*log_abs_mue;
SigmaSqr = exp(log_SigmaSqr);
for j=1:N
rep_DataMat(:,:,j) = repmat(DataMat(:,j),1,M);
log_gamma(j,:) = log_c - 0.5*(FrameSize*log(2*pi)+sum(log_SigmaSqr)) + sum((rep_DataMat(:,:,j) - mue).^2./(2*SigmaSqr));
end
log_alpha = log_gamma - repmat(Logsum(log_gamma,2),1,M);
alpha = exp(log_alpha);
end
c = exp(log_c);
SigmaSqr = exp(log_SigmaSqr);
does any one see how i can avoid this? or what needs to be fixed in code?
What i did was add this line to the MATLAB code:
mue(isnan(mue))=0; %fix 0/0 problem
and this one:
SigmaSqr(SigmaSqr==0)=1;%fix if mue_k = x_k
not sure if this is the best solution but is seems to work...
any have a better idea?