i was doing some tests using a VGG3 CNN model and i zeroed out all weights and their gradients after every backprop call. i basically allowed only the bias of the model to be trained / updated. what i got was interesting results that i cant per say explain. the accuracy of the model kept on increasing. started off at about 27.12% until after only 10 epochs it reached around 71% accuracy.
this kept me wondering how the accuracy is increasing so significantly without using any weights at all during training. i would love to get some theories and inputs regarding this.
can you actually train a model(not necessarily the best model) without using any weights at all?
If you zero the weights and only use bias then output of the model should be independent of the input. In fact the model only depends on the final layer's bias value. If you implemented it correctly then you should have something akin to a maximum a-priori estimator. I.e. your model predicts the most common class in your training data. That means one of the classes in your training/testing data is overrepresented and your model is always predicting that class.
For PyTorch's tutorial on performing transfer learning for computer vision (https://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html), we can see that there is a higher validation accuracy than training accuracy. Applying the same steps to my own dataset, I see similar results. Why is this the case? Does it have something to do with ResNet 18's architecture?
Assuming there aren't bugs in your code and the train and validation data are in the same domain, then there are a couple reasons why this may occur.
Training loss/acc is computed as the average across an entire training epoch. The network begins the epoch with one set of weights and ends the epoch with a different (hopefully better!) set of weights. During validation you're evaluating everything using only the most recent weights. This means that the comparison between validation and train accuracy is misleading since training accuracy/loss was computed with samples from potentially much worse states of your model. This is usually most noticeable at the start of training or right after the learning rate is adjusted since the network often starts the epoch in a much worse state than it ends. It's also often noticeable when the training data is relatively small (as is the case in your example).
Another difference is the data augmentations used during training that aren't used during validation. During training you randomly crop and flip the training images. While these random augmentations are useful for increasing the ability of your network to generalize they aren't performed during validation because they would diminish performance.
If you were really motivated and didn't mind spending the extra computational power you could get a more meaningful comparison by running the training data back through your network at the end of each epoch using the same data transforms used for validation.
The short answer is that train and validation data are from different distributions, and it's "easier" for model to predict target in validation data then it is for training.
The likely reason for this particular case, as indicated by this answer, is data augmentation during training. This is a way to regularize your model by increasing variability in the training data.
Other architectures can use Dropout (or its modifications), which are deliberately "hurting" training performance, reducing the potential of overfitting.
Notice, that you're using pretrained model, which already contains some information about how to solve classification problem. If your domain is not that different from the data it was trained on, you can expect good performance off-the-shelf.
For a class project, I designed a neural network to approximate sin(x), but ended up with a NN that just memorized my function over the data points I gave it. My NN took in x-values with a batch size of 200. Each x-value was multiplied by 200 different weights, mapping to 200 different neurons in my first layer. My first hidden layer contained 200 neurons, each one a linear combination of the x-values in the batch. My second hidden layer also contained 200 neurons, and my loss function was computed between the 200 neurons in my second layer and the 200 values of sin(x) that the input mapped to.
The problem is, my NN perfectly "approximated" sin(x) with 0 loss, but I know it wouldn't generalize to other data points.
What did I do wrong in designing this neural network, and how can I avoid memorization and instead design my NN's to "learn" about the patterns in my data?
It is same with any machine learning algorithm. You have a dataset based on which you try to learn "the" function f(x), which actually generated the data. In real life datasets, it is impossible to get the original function from the data, and therefore we approximate it using something g(x).
The main goal of any machine learning algorithm is to predict unseen data as best as possible using the function g(x).
Given a dataset D you can always train a model, which will perfectly classify all the datapoints (you can use a hashmap to get 0 error on the train set), but which is overfitting or memorization.
To avoid such things, you yourself have to make sure that the model does not memorise and learns the function. There are a few things which can be done. I am trying to write them down in an informal way (with links).
Train, Validation, Test
If you have large enough dataset, use Train, Validation, Test splits. Split the dataset in three parts. Typically 60%, 20% and 20% for Training, Validation and Test, respectively. (These numbers can vary based on need, also in case of imbalanced data, check how to get stratified partitions which preserve the class ratios in every split). Next, forget about the Test partition, keep it somewhere safe, don't touch it. Your model, will be trained using the Training partition. Once you have trained the model, evaluate the performance of the model using the Validation set. Then select another set of hyper-parameter configuration for your model (eg. number of hidden layer, learaning algorithm, other parameters etc.) and then train the model again, and evaluate based on Validation set. Keep on doing this for several such models. Then select the model, which got you the best validation score.
The role of validation set here is to check what the model has learned. If the model has overfit, then the validation scores will be very bad, and therefore in the above process you will discard those overfit models. But keep in mind, although you did not use the Validation set to train the model, directly, but the Validation set was used indirectly to select the model.
Once you have selected a final model based on Validation set. Now take out your Test set, as if you just got new dataset from real life, which no one has ever seen. The prediction of the model on this Test set will be an indication how well your model has "learned" as it is now trying to predict datapoints which it has never seen (directly or indirectly).
It is key to not go back and tune your model based on the Test score. This is because once you do this, the Test set will start contributing to your mode.
Crossvalidation and bootstrap sampling
On the other hand, if your dataset is small. You can use bootstrap sampling, or k-fold cross-validation. These ideas are similar. For example, for k-fold cross-validation, if k=5, then you split the dataset in 5 parts (also be carefull about stratified sampling). Let's name the parts a,b,c,d,e. Use the partitions [a,b,c,d] to train and get the prediction scores on [e] only. Next, use the partitions [a,b,c,e] and use the prediction scores on [d] only, and continue 5 times, where each time, you keep one partition alone and train the model with the other 4. After this, take an average of these scores. This is indicative of that your model might perform if it sees new data. It is also a good practice to do this multiple times and perform an average. For example, for smaller datasets, perform a 10 time 10-folds cross-validation, which will give a pretty stable score (depending on the dataset) which will be indicative of the prediction performance.
Bootstrap sampling is similar, but you need to sample the same number of datapoints (depends) with replacement from the dataset and use this sample to train. This set will have some datapoints repeated (as it was a sample with replacement). Then use the missing datapoins from the training dataset to evaluate the model. Perform this multiple times and average the performance.
Others
Other ways are to incorporate regularisation techniques in the classifier cost function itself. For example in Support Vector Machines, the cost function enforces conditions such that the decision boundary maintains a "margin" or a gap between two class regions. In neural networks one can also do similar things (although it is not same as in SVM).
In neural network you can use early stopping to stop the training. What this does, is train on the Train dataset, but at each epoch, it evaluates the performance on the Validation dataset. If the model starts to overfit from a specific epoch, then the error for Training dataset will keep on decreasing, but the error of the Validation dataset will start increasing, indicating that your model is overfitting. Based on this one can stop training.
A large dataset from real world tends not to overfit too much (citation needed). Also, if you have too many parameters in your model (to many hidden units and layers), and if the model is unnecessarily complex, it will tend to overfit. A model with lesser pameter will never overfit (though can underfit, if parameters are too low).
In the case of you sin function task, the neural net has to overfit, as it is ... the sin function. These tests can really help debug and experiment with your code.
Another important note, if you try to do a Train, Validation, Test, or k-fold crossvalidation on the data generated by the sin function dataset, then splitting it in the "usual" way will not work as in this case we are dealing with a time-series, and for those cases, one can use techniques mentioned here
First of all, I think it's a great project to approximate sin(x). It would be great if you could share the snippet or some additional details so that we could pin point the exact problem.
However, I think that the problem is that you are overfitting the data hence you are not able to generalize well to other data points.
Few tricks that might work,
Get more training points
Go for regularization
Add a test set so that you know whether you are overfitting or not.
Keep in mind that 0 loss or 100% accuracy is mostly not good on training set.
Any idea about why our training loss is smooth and our validation loss is that noisy (see the link) across epochs? We are implementing a deep learning model for diabetic retinopathy detection (binary classification) using the data set of fundus photographs provided by this Kaggle competition. We are using Keras 2.0 with Tensorflow backend.
As the data set is too big to fit in memory, we are using fit_generator, with ImageDataGenerator randomly taking images from training and validation folders:
# TRAIN THE MODEL
model.fit_generator(
train_generator,
steps_per_epoch= train_generator.samples // training_batch_size,
epochs=int(config['training']['epochs']),
validation_data=validation_generator,
validation_steps= validation_generator.samples // validation_batch_size,
class_weight=None)
Our CNN architecture is VGG16 with dropout = 0.5 in the last two fully connected layers, batch normalization only before the first fully connected layer, and data augmentation (consisting on flipping the images horizontally and vertically). Our training and validation samples are normalized using the training set mean and standard deviation. Batch size is 32. Our activation is a sigmoid and the loss function is the binary_crossentropy. You can find our implementation in Github
It definitely has nothing to do with overfitting, as we tried with a highly regularized model and the behavior was quite the same. Is it related with the sampling from the validation set? Has any of you had a similar problem before?
Thanks!!
I would look, in that order:
bug in validation_generator implementation (incl. steps - does it go through all pics reserved for validation?)
in validation_generator, do not use augmentation (reason: an augmentation might be bad, not learnable, and at train, it does achieve a good score only by hard-coding relationships which are not generalizable)
change train/val split to 50/50
calculate, via a custom callback, the validation loss at the end of the epoch (use the same function, but calling it with a callback produces different (more accurate, at certain, non-standard models) results)
If nothing of the above gives a more smooth validation loss curve, then my next assumption would be that this is the way it is, and I might need to work on the model architecture
I trained GoogLeNet model from scratch. But it didn't give me the promising results.
As an alternative, I would like to do fine tuning of GoogLeNet model on my dataset. Does anyone know what are the steps should I follow?
Assuming you are trying to do image classification. These should be the steps for finetuning a model:
1. Classification layer
The original classification layer "loss3/classifier" outputs predictions for 1000 classes (it's mum_output is set to 1000). You'll need to replace it with a new layer with appropriate num_output. Replacing the classification layer:
Change layer's name (so that when you read the original weights from caffemodel file there will be no conflict with the weights of this layer).
Change num_output to the right number of output classes you are trying to predict.
Note that you need to change ALL classification layers. Usually there is only one, but GoogLeNet happens to have three: "loss1/classifier", "loss2/classifier" and "loss3/classifier".
2. Data
You need to make a new training dataset with the new labels you want to fine tune to. See, for example, this post on how to make an lmdb dataset.
3. How extensive a finetuning you want?
When finetuning a model, you can train ALL model's weights or choose to fix some weights (usually filters of the lower/deeper layers) and train only the weights of the top-most layers. This choice is up to you and it ususally depends on the amount of training data available (the more examples you have the more weights you can afford to finetune).
Each layer (that holds trainable parameters) has param { lr_mult: XX }. This coefficient determines how susceptible these weights to SGD updates. Setting param { lr_mult: 0 } means you FIX the weights of this layer and they will not be changed during the training process.
Edit your train_val.prototxt accordingly.
4. Run caffe
Run caffe train but supply it with caffemodel weights as an initial weights:
~$ $CAFFE_ROOT/build/tools/caffe train -solver /path/to/solver.ptototxt -weights /path/to/orig_googlenet_weights.caffemodel
Fine-tuning is a very useful trick to achieve a promising accuracy compared to past manual feature. #Shai already posted a good tutorial for fine-tuning the Googlenet using Caffe, so I just want to give some recommends and tricks for fine-tuning for general cases.
In most of time, we face a task classification problem that new dataset (e.g. Oxford 102 flower dataset or Cat&Dog) has following four common situations CS231n:
New dataset is small and similar to original dataset.
New dataset is small but is different to original dataset (Most common cases)
New dataset is large and similar to original dataset.
New dataset is large but is different to original dataset.
In practice, most of time we do not have enough data to train the network from scratch, but may be enough for pre-trained model. Whatever which cases I mentions above only thing we must care about is that do we have enough data to train the CNN?
If yes, we can train the CNN from scratch. However, in practice it is still beneficial to initialize the weight from pre-trained model.
If no, we need to check whether data is very different from original datasets? If it is very similar, we can just fine-tune the fully connected neural network or fine-tune with SVM. However, If it is very different from original dataset, we may need to fine-tune the convolutional neural network to improve the generalization.