We Keep Coding

Model doesn't learn

I am newbie to Machine Learning. For some reason, my CNN doesn't learn at all. I tried on different datasets, but result is the same: loss and accuracy are changing just a little (most likely this is just an inaccuracy). Maybe I configured something incorrectly or did not do something at lot.
The task is to create a lip movement generator based on face photos and audio recordings (supervised learning)
[audio.mp3 + face.jpg] -> [1.jpg, 2.jpg, 3.jpg, 4.jpg, 5.jpg, 6.jpg ... ]
I do normalize input data,
I set loss function = mean_squared_error
I set metrics = [accuracy, mean_squared_error]
I set optimizer = Adam,
The dataset is large enough to see at least some results = 2000 videos
The epoch number = 2,
batch_size = 16
Learning_rate = 0.001 (default)
The model structure diagram is below. I chose the number of layers and their types at random.
Input frame = face, picture 50x60, channels = 3. Shape = (50, 60, 3)
Input mfcc = numerical coefficients. Shape = (20, 43)
Output: 24 images 50x60
I'm trying to generate output = lips movement (video = many frames) by audio (mfcc) and face image.
Here you are my logs and model structure

How to split the Cora dataset to train a GCN model only on training part?

I am training a GCN (Graph Convolutional Network) on Cora dataset.
The Cora dataset has the following attributes:
Number of graphs: 1
Number of features: 1433
Number of classes: 7
Number of nodes: 2708
Number of edges: 10556
Number of training nodes: 140
Training node label rate: 0.05
Is undirected: True
Data(edge_index=[2, 10556], test_mask=[2708], train_mask=[2708], val_mask=[2708], x=[2708, 1433], y=[2708])
Since my code is very long, I only put the relevent parts of my code here. Firstly, I split the Cora dataset as follows:
def to_mask(index, size):
mask = torch.zeros(size, dtype=torch.bool)
mask[index] = 1
return mask
def cora_splits(data, num_classes):
indices = []
for i in range(num_classes):
# returns all indices of the elements = i from data.y tensor
index = (data.y == i).nonzero().view(-1)
# returns a random permutation of integers from 0 to index.size(0).
index = index[torch.randperm(index.size(0))]
# indices is a list of tensors and it has a length of 7
indices.append(index)
# select 20 nodes from each class for training
train_index = torch.cat([i[:20] for i in indices], dim=0)
rest_index = torch.cat([i[20:] for i in indices], dim=0)
rest_index = rest_index[torch.randperm(len(rest_index))]
data.train_mask = to_mask(train_index, size=data.num_nodes)
data.val_mask = to_mask(rest_index[:500], size=data.num_nodes)
data.test_mask = to_mask(rest_index[500:], size=data.num_nodes)
return data
The train is as follows (taken from here with few modifications):
def train(model, optimizer, data, epoch):
t = time.time()
model.train()
optimizer.zero_grad()
output = model(data)
loss_train = F.nll_loss(output[data.train_mask], data.y[data.train_mask])
acc_train = accuracy(output[data.train_mask], data.y[data.train_mask])
loss_train.backward()
optimizer.step()
loss_val = F.nll_loss(output[data.val_mask], data.y[data.val_mask])
acc_val = accuracy(output[data.val_mask], data.y[data.val_mask])
def accuracy(output, labels):
preds = output.max(1)[1].type_as(labels)
correct = preds.eq(labels).double()
correct = correct.sum()
return correct / len(labels)
When I ran my code with 200 epochs in 10 runs I gained:
tensor([0.7690, 0.8030, 0.8530, 0.8760, 0.8600, 0.8550, 0.8850, 0.8580, 0.8940, 0.8830])
Val Loss: 0.5974, Test Accuracy: 0.854 ± 0.039
where each value in the tensor belongs to the model accurracy of each run and the mean accuracy of all those 10 runs is 0.854 with std ± 0.039.
As it can be observed, the accuracy from the first run to the 10th one is increasing substantially. Therefore, I think the model is overfitting. One reason of overfitting is that in the code, the test data has been seen by the model in the training time since in the train function, there is a line output = model(data) so the model is trained over the whole data. What I intend to do is to train my model only on a part of the data (something similar to data[data.train_mask]) but the problem is I cannot pass data[data.train_mask], due to the forward function of the GCN model (from this repository):
def forward(self, data):
x, edge_index = data.x, data.edge_index
x = F.relu(self.conv1(x, edge_index))
for conv in self.convs:
x = F.relu(conv(x, edge_index))
x = F.relu(self.lin1(x))
x = F.dropout(x, p=0.5, training=self.training)
x = self.lin2(x)
return F.log_softmax(x, dim=-1)
If I pass data[data.train_mask] to the GCN model, then in the above forward function in line x, edge_index = data.x, data.edge_index, x and edge_index cannot be retrieved from data[data.train_mask]. Therefore, I need to find a way to split the Cora dataset in a way that I can pass a specefic part of it with the nodes, edge-index and other attributes to the model. My question is how to do it?
Also, any suggestion about k-fold cross validation is much appreciated.

I guess you are a little confused by the nature of transductive learning and the question you asked doesn't actually address the problem you are facing.
As it can be observed, the accuracy from the first run to the 10th one
is increasing substantially. Therefore, I think the model is
overfitting.
Not necessarily, increasing test accuracy could be a normal behavior when your model is learning from the training samples. The learning can last for several dozens of epochs due to the complexity and non-convexity of loss function. The best signal to tell overfit is when your training accuracy increase but test accuracy decreases significantly.
One reason of overfitting is that in the code, the test data has been
seen by the model in the training time since in the train function,
there is a line output = model(data) so the model is trained over the
whole data.
The model indeed has seen the entire graph(adjacency matrix) in the training, but it only sees the labels of nodes in the training set and knows nothing about the labels of nodes in the test set. This is exactly what transductive learning does.
In the end, if you are 100% sure you want to avoid the paradigm of transductive learning, then you might need to write your own split algorithm to achieve that. But I would like to remind that in the real-world use case, transduction is perfectly suitable. An example is to predict the potential links between social network users, where we have the whole network structure as input and want to simply run edge prediction--> transduction. Thus it doesn't make a lot of sense to avoid it.
Depending on your task, you could take a look of how Stellargraph's EdgeSplitter class(docs) and scikit-learn’s train_test_split function (docs) achive the split.
Node classification
If your task is a node classification task, this Node classification with Graph Convolutional Network (GCN) is a good example of how to load data and do train-test-split. It took Cora dataset as an example. The most important steps are the following:
dataset = sg.datasets.Cora()
display(HTML(dataset.description))
G, node_subjects = dataset.load()
train_subjects, test_subjects = model_selection.train_test_split(
node_subjects, train_size=140, test_size=None, stratify=node_subjects
)
val_subjects, test_subjects = model_selection.train_test_split(
test_subjects, train_size=500, test_size=None, stratify=test_subjects
)
train_gen = generator.flow(train_subjects.index, train_targets)
val_gen = generator.flow(val_subjects.index, val_targets)
test_gen = generator.flow(test_subjects.index, test_targets)
Basically, it's the same as train-test-split with a normal classification task, except what we split here is nodes.
Edge classification
If your task is edge classification, you could have a look at this Link prediction example: GCN on the Cora citation dataset. The most relevant code for train-test-split is
# Define an edge splitter on the original graph G:
edge_splitter_test = EdgeSplitter(G)
# Randomly sample a fraction p=0.1 of all positive links, and same number of negative links, from G, and obtain the
# reduced graph G_test with the sampled links removed:
G_test, edge_ids_test, edge_labels_test = edge_splitter_test.train_test_split(
p=0.1, method="global", keep_connected=True
)
# Define an edge splitter on the reduced graph G_test:
edge_splitter_train = EdgeSplitter(G_test)
# Randomly sample a fraction p=0.1 of all positive links, and same number of negative links, from G_test, and obtain the
# reduced graph G_train with the sampled links removed:
G_train, edge_ids_train, edge_labels_train = edge_splitter_train.train_test_split(
p=0.1, method="global", keep_connected=True
)
# For training we create a generator on the G_train graph, and make an
# iterator over the training links using the generator’s flow() method:
train_gen = FullBatchLinkGenerator(G_train, method="gcn")
train_flow = train_gen.flow(edge_ids_train, edge_labels_train)
test_gen = FullBatchLinkGenerator(G_test, method="gcn")
test_flow = train_gen.flow(edge_ids_test, edge_labels_test)
Here the splitting algorithm behind EdgeSplitter class(docs) is more complex, it needs to maintain the graph structure while doing the split, such as keeping the graph connectivity for example. For more details, cf source code for EdgeSplitter

How to write an efficient custom Keras data generator

I would like to train a convolutional recurrent neural network for video frame prediction. The individual frames are quite big so it is challenging to fit the entire training data in memory at once. As such, I followed some tutorials online to create a custom data generator. When testing it, it seems to work but it is slower by a factor of at least 100 than using the pre-loaded data directly. Since I can only fit about a batch size of 8 on the GPU I understand that the data needs to be generated really fast, however, this does not seem to be the case.
I train my model on a single P100 and have 32 GB of memory available to be used by up to 16 cores.
class DataGenerator(tf.keras.utils.Sequence):
def __init__(self, images, input_images=5, predict_images=5, batch_size=16, image_size=(200, 200),
channels=1):
self.images = images
self.input_images = input_images
self.predict_images = predict_images
self.batch_size = batch_size
self.image_size = image_size
self.channels = channels
self.nr_images = int(len(self.images)-input_images-predict_images)
def __len__(self):
return int(np.floor(self.nr_images) / self.batch_size)
def __getitem__(self, item):
# Randomly select the beginning image of each batch
batch_indices = random.sample(range(0, self.nr_images), self.batch_size)
# Allocate the output images
x = np.empty((self.batch_size, self.input_images,
*self.image_size, self.channels), dtype='uint8')
y = np.empty((self.batch_size, self.predict_images,
*self.image_size, self.channels), dtype='uint8')
# Get the list of input an prediction images
for i in range(self.batch_size):
list_images_input = range(batch_indices[i], batch_indices[i]+self.input_images)
list_images_predict = range(batch_indices[i]+self.input_images,
batch_indices[i]+self.input_images+self.predict_images)
for j, ID in enumerate(list_images_input):
x[i, ] = np.load(np.reshape(self.images[ID], (*self.imagesize, self.channels))
# Read in the prediction images
for j, ID in enumerate(list_images_predict):
y[i, ] = np.load(np.reshape(self.images[ID], (*self.imagesize, self.channels))
return x, y
# Training the model using fit_generator
params = {'batch_size': 8,
'input_images': 5,
'predict_images': 5,
'image_size': (100, 100),
'channels': 1
}
data_path = "input_frames/"
input_images = sorted(glob.glob(data_path + "*.png"))
training_generator = DataGenerator(input_images, **params)
model.fit_generator(generator=training_generator, epochs=10, workers=6)
I would have expected that Keras will prepare the next data batch while the current batch is being processed on the GPU but it does not seem to catch up. In other words, preparing the data before sending it to the GPU seems to be the bottleneck.
Any idea on how to improve the performance of a data generator like this? Is there something missing that guarantees that the data is being prepared in a timely manner?
Thanks a lot!

When you use fit_generator, there is a workers= setting that can be used to scale up the number of generator workers. However you should ensure that the 'item' parameter in getitem is taken into account in order to ensure that the different workers (which are not synchronised) return different values depending on item index. i.e. instead of random sample, perhaps just return a slice of the data based on the index. You can shuffle the entire dataset before starting in order to make sure the dataset order is randomised.

Can you please try with use_multiprocessing=True? These are the numbers I observe on my GTX 1080Ti based system with the data generator you provided.
model.fit_generator(generator=training_generator, epochs=10, workers=6)
148/148 [==============================] - 9s 60ms/step
model.fit_generator(generator=training_generator, epochs=10, workers=6, use_multiprocessing=True)
148/148 [==============================] - 2s 11ms/step

You can try the prefetching of tf.data.Dataset. The prefetching allows you to compute the next batch(es) using your CPU while your GPU computes the gradient descent in the same time. Be careful: you need to change the numpy array into tf.constant in the data generator. Then try:
import tensoflow as tf
generator = DataGenerator(images)
spec = [tf.TypeSpec(shape=(generator.batch_size, generator.input_images,
*generator.image_size, generator.channels), dtype='uint8'),
tf.TypeSpec(shape=(generator.batch_size, generator.predict_images,
*generator.image_size, generator.channels), dtype='uint8')
dataset = tf.data.Dataset.from_generator(DataGenerator, output_signature=spec)
dataset.batch(batch_size).prefetch(-1) # this order is important
# a custom training loop is better than model.fit() otherwise prefetching can fail
def train_loop():
...
You can change the "-1" in prefetch() to another value like 1, 2 or more to get the maximum speed depending on your machine and the batch size.

this blog helps in setting up input data pipeline with tf.data and it also is much more efficient than using ImageDataGenerators and the code is also explained by using a custom data directory.
It also enhances the performance with prefetch, cache.
Prefetch processes the next batch while the current batch is being used.

How to get Keras network to not output all 1s

I have a bunch of images that look like this of someone playing a videogame (a simple game I created in Tkinter):
The idea of the game is that the user controls the box at the bottom of the screen in order to dodge the falling balls (they can only dodge left and right).
My goal is to have the neural network output the position of the player on the bottom of the screen. If they're totally on the left, the neural network should output a 0, if they're in the middle, a .5, and all the way right, a 1, and all the values in-between.
My images are 300x400 pixels. I stored my data very simply. I recorded each of the images and position of the player as a tuple for each frame in a 50-frame game. Thus my result was a list in the form [(image, player position), ...] with 50 elements. I then pickled that list.
So in my code I try to create an extremely basic feed-forward network that takes in the image and outputs a value between 0 and 1 representing where the box on the bottom of the image is. But my neural network is only outputting 1s.
What should I change in order to get it to train and output values close to what I want?
Of course, here is my code:
# machine learning code mostly from https://machinelearningmastery.com/tutorial-first-neural-network-python-keras/
from keras.models import Sequential
from keras.layers import Dense
import numpy as np
import pickle
def pil_image_to_np_array(image):
'''Takes an image and converts it to a numpy array'''
# from https://stackoverflow.com/a/45208895
# all my images are black and white, so I only need one channel
return np.array(image)[:, :, 0:1]
def data_to_training_set(data):
# split the list in the form [(frame 1 image, frame 1 player position), ...] into [[all images], [all player positions]]
inputs, outputs = [list(val) for val in zip(*data)]
for index, image in enumerate(inputs):
# convert the PIL images into numpy arrays so Keras can process them
inputs[index] = pil_image_to_np_array(image)
return (inputs, outputs)
if __name__ == "__main__":
# fix random seed for reproducibility
np.random.seed(7)
# load data
# data will be in the form [(frame 1 image, frame 1 player position), (frame 2 image, frame 2 player position), ...]
with open("position_data1.pkl", "rb") as pickled_data:
data = pickle.load(pickled_data)
X, Y = data_to_training_set(data)
# get the width of the images
width = X[0].shape[1] # == 400
# convert the player position (a value between 0 and the width of the image) to values between 0 and 1
for index, output in enumerate(Y):
Y[index] = output / width
# flatten the image inputs so they can be passed to a neural network
for index, inpt in enumerate(X):
X[index] = np.ndarray.flatten(inpt)
# keras expects an array (not a list) of image-arrays for input to the neural network
X = np.array(X)
Y = np.array(Y)
# create model
model = Sequential()
# my images are 300 x 400 pixels, so each input will be a flattened array of 120000 gray-scale pixel values
# keep it super simple by not having any deep learning
model.add(Dense(1, input_dim=120000, activation='sigmoid'))
# Compile model
model.compile(loss='mean_squared_error', optimizer='adam')
# Fit the model
model.fit(X, Y, epochs=15, batch_size=10)
# see what the model is doing
predictions = model.predict(X, batch_size=10)
print(predictions) # this prints all 1s! # TODO fix
EDIT: print(Y) gives me:
so it's definitely not all zeroes.

Of course, a deeper model might give you a better accuracy, but considering the fact that your images are simple, a pretty simple (shallow) model with only one hidden layer should give a medium to high accuracy. So here are the modifications you need to make this happen:
Make sure X and Y are of type float32 (currently, X is of type uint8):
X = np.array(X, dtype=np.float32)
Y = np.array(Y, dtype=np.float32)
When training a neural network it would be much better to normalize the training data. Normalization helps the optimization process to go smoothly and speed up the convergence to a solution. It further prevent large values to cause large gradient updates which would be desruptive. Usually, the values of each feature in the input data should fall in a small range, where two common ranges are [-1,1] and [0,1]. Therefore, to make sure that all values fall in the range [-1,1], we subtract from each feature its mean and divide it by its standard deviation:
X_mean = X.mean(axis=0)
X -= X_mean
X_std = X.std(axis=0)
X /= X_std + 1e-8 # add a very small constant to prevent division by zero
Note that we are normalizing each feature (i.e. each pixel in this case) here not each image. When you want to predict on new data, i.e. in inference or testing mode, you need to subtract X_mean from test data and divide it by X_std (you should NEVER EVER subtract from test data its own mean or divide it by its own standard deviation; rather, use the mean and std of training data):
X_test -= X_mean
X_test /= X_std + 1e-8
If you apply the changes in points one and two, you might notice that the network no longer predicts only ones or only zeros. Rather, it shows some faint signs of learning and predicts a mix of zeros and ones. This is not bad but it is far from good and we have high expectations! The predictions should be much better than a mix of only zeros and ones. There, you should take into account the (forgotten!) learning rate. Since the network has relatively large number of parameters considering a relatively simple problem (and there are a few samples of training data), you should choose a smaller learning rate to smooth the gradient updates and the learning process:
from keras import optimizers
model.compile(loss='mean_squared_error', optimizer=optimizers.Adam(lr=0.0001))
You would notice the difference: the loss value reaches to around 0.01 after 10 epochs. And the network no longer predicts a mix of zeros and ones; rather the predictions are much more accurate and close to what they should be (i.e. Y).
Don't forget! We have high (logical!) expectations. So, how can we do better without adding any new layers to the network (obviously, we assume that adding more layers might help!!)?
4.1. Gather more training data.
4.2. Add weight regularization. Common ones are L1 and L2 regularization (I highly recommend the Jupyter notebooks of the the book Deep Learning with Python written by François Chollet the creator of Keras. Specifically, here is the one which discusses regularization.)
You should always evaluate your model in a proper and unbiased way. Evaluating it on the training data (that you have used to train it) does not tell you anything about how well your model would perform on unseen (i.e. new or real world) data points (e.g. consider a model which stores or memorize all the training data. It would perform perfectly on the training data, but it would be a useless model and perform poorly on new data). So we should have test and train datasets: we train model on the training data and evaluate the model on the test (i.e. new) data. However, during the process of coming up with a good model you are performing lots of experiments: for example, you first change the type and number of layers, train the model and then evaluate it on test data to make sure it is good. Then you change another thing say the learning rate, train it again and then evaluate it again on test data... To make it short, these cycles of tuning and evaluations somehow causes an over-fitting on the test data. Therefore, we would need a third dataset called validation data (read more: What is the difference between test set and validation set?):
# first shuffle the data to make sure it isn't in any particular order
indices = np.arange(X.shape[0])
np.random.shuffle(indices)
X = X[indices]
Y = Y[indices]
# you have 200 images
# we select 100 images for training,
# 50 images for validation and 50 images for test data
X_train = X[:100]
X_val = X[100:150]
X_test = X[150:]
Y_train = Y[:100]
Y_val = Y[100:150]
Y_test = Y[150:]
# train and tune the model
# you can attempt train and tune the model multiple times,
# each time with different architecture, hyper-parameters, etc.
model.fit(X_train, Y_train, epochs=15, batch_size=10, validation_data=(X_val, Y_val))
# only and only after completing the tuning of your model
# you should evaluate it on the test data for just one time
model.evaluate(X_test, Y_test)
# after you are satisfied with the model performance
# and want to deploy your model for production use (i.e. real world)
# you can train your model once more on the whole data available
# with the best configurations you have found out in your tunings
model.fit(X, Y, epochs=15, batch_size=10)
(Actually, when we have few training data available it would be wasteful to separate validation and test data from whole available data. In this case, and if the model is not computationally expensive, instead of separating a validation set which is called cross-validation, one can do K-fold cross-validation or iterated K-fold cross-validation in case of having very few data samples.)
It is around 4 AM at the time of writing this answer and I am feeling sleepy, but I would like to mention one more thing which is not directly related to your question: by using the Numpy library and its functionalities and methods you can write more concise and efficient code and also save yourself a lot time. So make sure you practice using it more as it is heavily used in machine learning community and libraries. To demonstrate this, here is the same code you have written but with more use of Numpy (Note that I have not applied all the changes I mentioned above in this code):
# machine learning code mostly from https://machinelearningmastery.com/tutorial-first-neural-network-python-keras/
from keras.models import Sequential
from keras.layers import Dense
import numpy as np
import pickle
def pil_image_to_np_array(image):
'''Takes an image and converts it to a numpy array'''
# from https://stackoverflow.com/a/45208895
# all my images are black and white, so I only need one channel
return np.array(image)[:, :, 0]
def data_to_training_set(data):
# split the list in the form [(frame 1 image, frame 1 player position), ...] into [[all images], [all player positions]]
inputs, outputs = zip(*data)
inputs = [pil_image_to_np_array(image) for image in inputs]
inputs = np.array(inputs, dtype=np.float32)
outputs = np.array(outputs, dtype=np.float32)
return (inputs, outputs)
if __name__ == "__main__":
# fix random seed for reproducibility
np.random.seed(7)
# load data
# data will be in the form [(frame 1 image, frame 1 player position), (frame 2 image, frame 2 player position), ...]
with open("position_data1.pkl", "rb") as pickled_data:
data = pickle.load(pickled_data)
X, Y = data_to_training_set(data)
# get the width of the images
width = X.shape[2] # == 400
# convert the player position (a value between 0 and the width of the image) to values between 0 and 1
Y /= width
# flatten the image inputs so they can be passed to a neural network
X = np.reshape(X, (X.shape[0], -1))
# create model
model = Sequential()
# my images are 300 x 400 pixels, so each input will be a flattened array of 120000 gray-scale pixel values
# keep it super simple by not having any deep learning
model.add(Dense(1, input_dim=120000, activation='sigmoid'))
# Compile model
model.compile(loss='mean_squared_error', optimizer='adam')
# Fit the model
model.fit(X, Y, epochs=15, batch_size=10)
# see what the model is doing
predictions = model.predict(X, batch_size=10)
print(predictions) # this prints all 1s! # TODO fix

Big HDF5 dataset, how to efficienly shuffle after each epoch

I'm currently working with a big image dataset (~60GB) to train a CNN (Keras/Tensorflow) for a simple classification task.
The images are video frames, and thus highly correlated in time, so I shuffled the data already once when generating the huge .hdf5 file...
To feed the data into the CNN without having to load the whole set at once into memory I wrote a simple batch generator (see code below).
Now my question:
Usually it is recommended to shuffle the data after each training epoch right? (for SGD convergence reasons?) But to do so I'd have to load the whole dataset after each epoch and shuffle it, which is exactly what I wanted to avoid using the batch generator...
So: Is it really that important to shuffle the dataset after each epoch and if yes how could I do that as efficiently as possible?
Here is the current code of my batch generator:
def generate_batches_from_hdf5_file(hdf5_file, batch_size, dimensions, num_classes):
"""
Generator that returns batches of images ('xs') and labels ('ys') from a h5 file.
"""
filesize = len(hdf5_file['labels'])
while 1:
# count how many entries we have read
n_entries = 0
# as long as we haven't read all entries from the file: keep reading
while n_entries < (filesize - batch_size):
# start the next batch at index 0
# create numpy arrays of input data (features)
xs = hdf5_file['images'][n_entries: n_entries + batch_size]
xs = np.reshape(xs, dimensions).astype('float32')
# and label info. Contains more than one label in my case, e.g. is_dog, is_cat, fur_color,...
y_values = hdf5_file['labels'][n_entries:n_entries + batch_size]
#ys = keras.utils.to_categorical(y_values, num_classes)
ys = to_categorical(y_values, num_classes)
# we have read one more batch from this file
n_entries += batch_size
yield (xs, ys)

Yeah, shuffling improves performance since running the data in the same order each time may get you stuck in suboptimal areas.
Don't shuffle the entire data. Create a list of indices into the data, and shuffle that instead. Then move sequentially over the index list and use its values to pick data from the data set.