I have a dataset of roughly 34000 images divided in 2 sets: train (30000 images) and validation (4000 images) sets. Each image is the result of the difference between two images taken from a video (the time offset between the images in each pair is about 1 second). The videos have a static background so the diff images contains too much black with only one or two small regions with colors. Each diff image has a label (there has been an action or no.. 1 or 0) so this is sort of binary classification. Briefly, I'm using the slim models pretrained on ImageNet to do the finetuning on my dataset. I've launched 5 separated training using 5 different networks: InceptionV4, InceptionResnetV2, Resnet152, NASNet-mobile, NASNet. I got very good results using the first 4 networks InceptionV4, InceptionResnetV2, Resnet152, NASNet-mobile but it was not the case using NASNet. The thing is that the Area Under the ROC curve on the validation set is always = 0.5 and the logits of the validation images are roughly having the same values which is really weird. In fact, I got this kind of results using NASNet-mobile on the first 10000 mini-batch but after that the model did converge. Here are the values of the hyperparameters I have in my script:
batch_size=10
weight_decay = 0.00004
optimizer = rmsprop
rmsprop_momentum = 0.9
rmsprop_decay = 0.9
learning_rate_decay_type = exponential
learning_rate = 0.01
learning_rate_decay_factor = 0.94
num_epochs_per_decay = 2.0 #'Number of epochs after which learning rate
I'm still newbie in tensorflow and I did not find anything related anywhere else. This is a really weird behavior because I'm using the same parameters and same inputs but it seems using NASNet there is a problem somewhere. I'm not only looking for a solution (if possible because I know it is tough to troubleshoot such things without too much details about the model) but insights about where to look and how to troubleshoot would be great. Does anybody had this problem with finetuning NASNet before? something like the model didn't converge for example? Finally, I know it is really hard to got answers on such questions but I hope to get at least some insights so I can move forward with my investigations.
EDIT:
Here are the plots of the cross entropy and regularization losses:
EDIT:
As proposed in the answer, I did set the drop_path_keep_prob params to 1 and now the model converged and I got good accuracy on the validation set. But now the question is: what does this param mean? Is it one of the params that we should adapt to our dataset (like learning rate etc..)?
The simplest sanity check you can do would be to run the finetuning on a single minibatch. Any deep network should be able to overfit to that, if there aren't any big problems. If you see that it can't do that, then there must be some problem with the definition, or the way you're using the definition.
The only guess I have in your case is that it could be something to do with the drop_path implementation. It's disabled in the mobile version, but it is enabled during training on the large model. It could make the model unstable enough that it wouldn't fine tune, so it may be worth trying to train with it disabled.
Related
I had implemented a CNN with 3 Convolutional layers with Maxpooling and dropout after each layer
I had noticed that when I trained the model for the first time it gave me 88% as testing accuracy but after retraining it for the second time successively, with the same training dataset it gave me 92% as testing accuracy.
I could not understand this behavior, is it possible that the model had overfitting in the second training process?
Thank you in advance for any help!
It is quite possible if you have not provided the seed number set.seed( ) in the R language or tf.random.set_seed(any_no.) in python
Well I am no expert when it comes to machine learning but I do know the math behind it. What you are doing when you train a neural network you basicly find the local minima to the loss function. What this means is that the end result will heavily depend on the initial guess of all of the internal varaibles.
Usually the variables are randomized as a initial estimation and you could therefore reach quite different results from running the training process multiple times.
That being said, from when I studied the subject I was told that you usually reach similar regardless of the initial guess of the parameters. However it is hard to say if 0.88 and 0.92 would be considered similar or not.
Hope this gives a somewhat possible answer to your question.
As mentioned in another answer, you could remove the randomization, both in the parameter initialization of the parameters and the randomization of the data used for each epoch of training by introducing a seed. This would insure that when you run it twice, everything will get "randomized" in the exact same order. In tensorflow this is done using for example tf.random.set_seed(1), the number 1 can be changed to any number to get a new seed.
I am using LSTM for time-series prediction using Keras. I am using 3 LSTM layers with dropout=0.3, hence my training loss is higher than validation loss. To monitor convergence, I using plotting training loss and validation loss together. Results looks like the following.
After researching about the topic, I have seen multiple answers for example ([1][2] but I have found several contradictory arguments on various different places on the internet, which makes me a little confused. I am listing some of them below :
1) Article presented by Jason Brownlee suggests that validation and train data should meet for the convergence and if they don't, I might be under-fitting the data.
https://machinelearningmastery.com/diagnose-overfitting-underfitting-lstm-models/
https://machinelearningmastery.com/learning-curves-for-diagnosing-machine-learning-model-performance/
2) However, following answer on here suggest that my model is just converged :
How do we analyse a loss vs epochs graph?
Hence, I am just bit confused about the whole concept in general. Any help will be appreciated.
Convergence implies you have something to converge to. For a learning system to converge, you would need to know the right model beforehand. Then you would train your model until it was the same as the right model. At that point you could say the model converged! ... but the whole point of machine learning is that we don't know the right model to begin with.
So when do you stop training? In practice, you stop when the model works well enough to do what you want it to do. This might be when validation error drops below a certain threshold. It might just be when you can't afford any more computing power. It's really up to you.
Here is a typical plot of train/test losses behaviour as epoch increases.
I'm not an expert but I have read several topics on similar problems. Well, let me explain what I'm doing.
First, I have used implementation given by https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py for resnet18 & resnet50, and by https://github.com/akamaster/pytorch_resnet_cifar10 for resnet32, resnet56. For all these nets I got the same kind of test-loss hieratic behaviour.
Second, my inputs are images 5x64x64, so I have adapted the first Convolutional Layer, and the output of the last Full-connected consist of 180 neurons. I have used either 64, 128, 256 batch sizes for the training, and 128 for the test: the same behaviour persists. I have also both used 300k or 100k images in input training (100k for the test): same behaviour persists too. The images are not of "standard" RGB photos: first, as you probably have already, , remarked there are 5 channels, second the pixel values can be negative (eg. spanning the range (-0.01, 500))
Third, I am aware of the model.train() statement for the training phase, as well as the model.eval() statement (coupled with the with torch.no_grad():) for the testing phase. It is clear that if I do not use model.eval() during the test phase, the test loss is gently decrasing as the traing loss. But, this is not allowed, isn't it?
I have tried several things after reading post concerning Batch Norm behaviour wo any success
I have used SGD, Adam (& SWATS)
I have tryied lr = 0.1 to lr= 1e-5
I have modified the BN momentum (default = 0.1) : 0.5 and 0.01; as well as the eps parameter.
Now, I have managed to get nice results (ie; good training & testing losses) with a classical CNN (ie. wo any Batch normalization, & short-cuts) but I would like to study Resnet behaviour against adversarial attack. So, I would like to get Resnet fit my images :slight_smile:
Any idea ?
Thanks
After making some tests, I have found something: I have used the standard resnet20 (h=1). Then, I have used as test set the same samples (100,000 images) as for the train set. BUT, for the test set 1) I do not use the shuffling, and 2) I do not make any Horizontal/vertical flip or Rot90deg, Rot180deg or Rot270deg. I observe the same kind of fluctuations for the test loss.
Moreover, when I switch OFF complety the transformations of the train set, and uses the same set for train & test, I got the same behviour:
And finaly, if I switch off the suffling and random transforms (flips & Rotations) of train set, and I use the same set for test, then I get:
Seems that the test loss is converging towards a value, but different from the train loss. Why ???
https://github.com/wenxinxu/resnet-in-tensorflow#overall-structure
The link above is the Resnet model for cifar10.
I am modifying above code to do object detection using Resnet and Cifar10 as training/validating dataset. ( I know the dataset is for object classification) I know that it sounds strange, but hear me out. I use Cifar10 for training and validation then during testing I use a sliding window approach, and then I classify each of the windows to one of 10 classes + "background" classes.
for background classes, I used images from ImageNet. I search ImageNet with following keyword: construction, landscape, byway, mountain, sky, ocean, furniture, forest, room, store, carpet, and floor. then I clean bad images out as much as I can including images that contain Cifar10 classes, for example, I delete a few "floor" images that have dogs in it.
I am currently running the result in Floydhub. Total steps that I am running is 60,000 which is where section under "training curve" from the link about suggests that the result starts to consolidate and do not converge further ( I personally run this code myself and I can back up the claim)
My question is:
what is the cause of the sudden step down in training and validation data which occurs at about the same step?
What if(or Is it possible that)training and validation data don't converge in a step-like fashion at about the same step? what I mean is, for example, training steps down at around 40,000 and validation just converge with no step-down? (smoothly converge)
The sudden step down is caused by the learning rate decay happening at 40k steps (you can find this parameter in hyper_parameters.py). The leraning rate suddenly gets divided by 10, which allows you to tune the parameters more precisely, which in this case improves your performance a lot. You still need the first part, with a pretty big learning rate, to get in a "good" area for your parameters, then the part with a 10x smaller learning rate will refine it and find a very good spot in that area for your parameters.
This would be surprising, since there is a clear difference between before and after 40k, that affects training and validation the same way. You could still see different behaviors from that point: for instance you might start overtraining because of a too small LR, and see you train error drop down and validation go up, because the refinements you're doing are too specific to the training data.
I'm somewhat new to caffe, and I'm getting some strange behavior. I'm trying to use fine tuning on the bvlc_reference_caffenet to accomplish an OCR task.
I've taken their pretrained net, changed the last FC layer to the number of output classes that I have, and retrained. After a few thousand iterations I'm getting loss rates of ~.001, and an accuracy over 90 percent when the network tests. That said, when I try to run my network on data by myself, I get awful results, not exceeding 7 or 8 percent.
The code I'm using to run the net is:
[imports]
net = caffe.Classifier('bvlc_reference_caffenet/deploy.prototxt', 'bvlc_reference_caffenet/caffenet_train_iter_28000.caffemodel',
image_dims=(227, 227, 1))
input_image = caffe.io.load_image('/Training_Processed/6/0.png')
prediction = net.predict([input_image]) # predict takes any number of images, and formats them for the Caffe net automatically
cls = prediction[0].argmax()
Any thoughts on why this performance might be so poor?
Thanks!
PS: Some additional information which may or not be of use. When classifying as shown below, the classifier really seems to favor certain classes. Even though I have a 101 class problem, it seems to only assign a max of 15 different classes
PPS: I'm also fairly certain I'm not overfitting. I've been testing this along the way with snapshots and they all exhibit the same poor results.
Your code for testing the model you posted seem to miss some components:
It seems like you did not subtract the image's mean.
You did not swap channels from RGB to BGR.
You did not scale the inputs to [0..255] range.
Looking at similar instances of caffe.Classifier you may see something like:
net = caffe.Classifier('bvlc_reference_caffenet/deploy.prototxt',
'bvlc_reference_caffenet/caffenet_train_iter_28000.caffemodel',
mean = NP.load( 'ilsvrc_2012_mean.npy' ),
input_scale=1.0, raw_scale=255,
channel_swap=(2,1,0),
image_dims=(227, 227, 1))
It is crucial to have the same input transformation in test as in training.