I am working on a basic Classification task with Keras and I seem to have stumbled upon a problem where I need some assistance.
I have 200 samples for training and a 100 for validation, I intend to use a ImageDataGenerator to increase the number of training samples for my task. I want to make sure of the total number of training images that are passed to the fit_generator().
I know that the steps_per_epoch defines the total number of batches we get from a generator and ideally it should be number of samples divided by the batch size.
However, this is where things do not add up for me. Here is a snippet of my code:
num_samples = 200
batch_size = 10
gen = ImageDataGenerator(horizontal_flip = True,
vertical_flip = True,
width_shift_range = 0.1,
height_shift_range = 0.1,
zoom_range = 0.1,
rotation_range = 10
)
x,y = shuffle(img_data,img_label, random_state=2)
x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.333, random_state=2)
generator = gen.flow(x_train, y_train, save_to_dir='check_images/sample_run')
new_network.fit_generator(generator, steps_per_epoch=len(x_train)/batch_size, validation_data=(x_test, y_test), epochs=1, verbose=2)
I am saving the augmented images to see how the images turn out from the ImageDataGenerator and also to ascertain the number of images that are generated from it.
After running this code for a single epoch, I get 600 images in my directory, a number which I cannot arrive at, or maybe I am making a mistake.
Any assistance in making me understand the calculation in this code would be deeply appreciated. Has anyone come across similar problems ?
TIA
gen.flow() creates a NumpyArrayIterator internally and that in turn uses Iterator to calculate the steps_per_epoch. Ideally if steps_per_epoch is None, then the calculation is done as steps_per_epoch = (x.shape[0] + batch_size - 1) // batch_size which is approximately same as your calculation.
Not sure why you see more number of samples. Could you compute the x.shape[0] and double check if your code is as per what you explained?
Related
I'm using a pre trained InceptionV3 on Keras to retrain the model to make a binary image classification (data labeled with 0's and 1's).
I'm reaching about 65% of accuracy on my k-fold validation with never seen data, but the problem is the model is overfitting to soon. I need to improve this average accuracy, and I guess there is something related to this overfitting problem.
Here are the loss values on epochs:
Here is the code. The dataset and label variables are Numpy Arrays.
dataset = joblib.load(path_to_dataset)
labels = joblib.load(path_to_labels)
le = LabelEncoder()
labels = le.fit_transform(labels)
labels = to_categorical(labels, 2)
X_train, X_test, y_train, y_test = sk.train_test_split(dataset, labels, test_size=0.2)
X_train, X_val, y_train, y_val = sk.train_test_split(X_train, y_train, test_size=0.25) # 0.25 x 0.8 = 0.2
X_train = np.array(X_train)
y_train = np.array(y_train)
X_val = np.array(X_val)
y_val = np.array(y_val)
X_test = np.array(X_test)
y_test = np.array(y_test)
aug = ImageDataGenerator(
rotation_range=20,
zoom_range=0.15,
horizontal_flip=True,
fill_mode="nearest")
pre_trained_model = InceptionV3(input_shape = (299, 299, 3),
include_top = False,
weights = 'imagenet')
for layer in pre_trained_model.layers:
layer.trainable = False
x = layers.Flatten()(pre_trained_model.output)
x = layers.Dense(1024, activation = 'relu')(x)
x = layers.Dropout(0.2)(x)
x = layers.Dense(2, activation = 'softmax')(x) #already tried with sigmoid activation, same behavior
model = Model(pre_trained_model.input, x)
model.compile(optimizer = RMSprop(lr = 0.0001),
loss = 'binary_crossentropy',
metrics = ['accuracy']) #Already tried with Adam optimizer, same behavior
es = EarlyStopping(monitor='val_loss', mode='min', verbose=1, patience=100)
mc = ModelCheckpoint('best_model_inception_rmsprop.h5', monitor='val_accuracy', mode='max', verbose=1, save_best_only=True)
history = model.fit(x=aug.flow(X_train, y_train, batch_size=32),
validation_data = (X_val, y_val),
epochs = 100,
callbacks=[es, mc])
The training dataset has 2181 images and validation has 727 images.
Something is wrong, but I can't tell what...
Any thoughts of what can be done to improve it?
One way to avoid overfitting is to use a lot of data. The main reason overfitting happens is because you have a small dataset and you try to learn from it. The algorithm will have greater control over this small dataset and it will make sure it satisfies all the datapoints exactly. But if you have a large number of datapoints, then the algorithm is forced to generalize and come up with a good model that suits most of the points.
Suggestions:
Use a lot of data.
Use less deep network if you have a small number of data samples.
If 2nd satisfies then don't use huge number of epochs - Using many epochs leads is kinda forcing your model to learn that and your model will learn it well but can not generalize.
From your loss graph , i see that the model is generalized at early epoch ( where there is intersection of both the train & val score) so plz try to use the model saved at that epoch ( and not the later epochs which seems to overfit)
Second option what you have is use lot of training samples..
If you have less no. of training samples then use data augmentations
Have you tried following?
Using a higher dropout value
Lower Learning Rate (lr=0.00001 or lr=0.000001 ...)
More data augmentation you can use.
It seems to me your data amount is low. You may use a lower ratio for test and validation (10%, 10%).
I have inertial measurement unit (IMU) data for which I am building an anomaly detection autoencoder neural net. I have about 5k training samples of which I am using 10% for validation. I also have about 50 (though I can make more) samples to test anomaly detection. My dataset has 12 IMU features. I train for about 10,000 epochs and I attain mean squared errors for reconstruction (MSE) of about 0.004 during training. After training, I perform an MSE calculation on the test data and I get values very similar to those in the train data (0.003) and I do not know why!
I am making my test set by slicing 50 samples from the overall data (not part of X_train) and changing one of the features to all zeros. I have also tried adding noise to one of the features as well as making multiple features zero.
np.random.seed(404)
np.random.shuffle(all_imu_data)
norm_imu_data = all_imu_data[:len_slice]
anom_imu_data = all_imu_data[len_slice:]
anom_imu_data[:,6] = 0
scaler = MinMaxScaler()
norm_data = scaler.fit_transform(norm_imu_data)
anom_data = scaler.transform(anom_imu_data)
X_train = pd.DataFrame(norm_data)
X_test = pd.DataFrame(anom_data)
I have tried many different network sizes by ranging number of hidden layers and number of hidden nodes/layer. As an example, I show a topology like [12-7-4-7-12]:
input_dim = num_features
input_layer = Input(shape=(input_dim, ))
encoder = Dense(int(7), activation="tanh", activity_regularizer=regularizers.l1(10e-5))(input_layer)
encoder = Dense(int(4), activation="tanh")(encoder)
decoder = Dense(int(7), activation="tanh")(encoder)
decoder = Dense(int(input_dim), activation="tanh")(decoder)
autoencoder = Model(inputs=input_layer, outputs=decoder)
autoencoder.compile(optimizer='adam', loss='mse', metrics=['mse'])
history = autoencoder.fit(X_train, X_train,
epochs=nb_epoch,
batch_size=batch_size,
shuffle=True,
validation_split=0.1,
verbose=1,
callbacks=[checkpointer, tensorboard]).history
pred_train = autoencoder.predict(X_train)
pred_test = autoencoder.predict(X_test)
mse_train = np.mean(np.power(X_train - pred_train, 2), axis=1)
mse_test = np.mean(np.power(X_test - pred_test, 2), axis=1)
print('MSE mean() - X_train:', np.mean(mse_train))
print('MSE mean() - X_test:', np.mean(mse_test))
After doing this, I get MSE mean numbers of 0.004 for Train and 0.003 for Test. Therefore, I cannot select a good threshold for anomalous data, as there are a lot of normal points that have larger MSE scores than the 'anomalous' data.
Any thoughts as to why this network is unable to detect these anomalies?
It is completely normal. You train your autoencoder on a sub sample of your whole data. Therefore, there are also anomalies contaminating your training data. The purpose of the autoencoder is to find a perfect reconstruction of your original data which it does including the anomalies. It is a very powerful tool, so if you show it anomalies in the training data, it will reconstruct them easily.
You need to remove 5% of your anomalous data with another anomaly detection algorithm (for example isolation forest) and do the subsampling on that part of the data (without outliers).
After that, you can find your outliers easily.
I'm trying to learn an embedding for Paris6k images combining VGG and Adrian Ung triplet loss. The problem is that after a small amount of iterations, in the first epoch, the loss becomes nan, and then the accuracy and validation accuracy grow to 1.
I've already tried lowering the learning rate, increasing the batch size (only to 16 beacuse of memory), changing optimizer (Adam and RMSprop), checking if there are None values on my dataset, changing data format from 'float32' to 'float64', adding a little bias to them and simplify the model.
Here is my code:
base_model = VGG16(include_top = False, input_shape = (512, 384, 3))
input_images = base_model.input
input_labels = Input(shape=(1,), name='input_label')
embeddings = Flatten()(base_model.output)
labels_plus_embeddings = concatenate([input_labels, embeddings])
model = Model(inputs=[input_images, input_labels], outputs=labels_plus_embeddings)
batch_size = 16
epochs = 2
embedding_size = 64
opt = Adam(lr=0.0001)
model.compile(loss=tl.triplet_loss_adapted_from_tf, optimizer=opt, metrics=['accuracy'])
label_list = np.vstack(label_list)
x_train = image_list[:2500]
x_val = image_list[2500:]
y_train = label_list[:2500]
y_val = label_list[2500:]
dummy_gt_train = np.zeros((len(x_train), embedding_size + 1))
dummy_gt_val = np.zeros((len(x_val), embedding_size + 1))
H = model.fit(
x=[x_train,y_train],
y=dummy_gt_train,
batch_size=batch_size,
epochs=epochs,
validation_data=([x_val, y_val], dummy_gt_val),callbacks=callbacks_list)
The images are 3366 with values scaled in range [0, 1].
The network takes dummy values because it tries to learn embeddings from images in a way that images of the same class should have small distance, while images of different classes should have high distances and than the real class is part of the training.
I've noticed that I was previously making an incorrect class division (and keeping images that should be discarded), and I didn't have the nan loss problem.
What should I try to do?
Thanks in advance and sorry for my english.
In some case, the random NaN loss can be caused by your data, because if there are no positive pairs in your batch, you will get a NaN loss.
As you can see in Adrian Ung's notebook (or in tensorflow addons triplet loss; it's the same code) :
semi_hard_triplet_loss_distance = math_ops.truediv(
math_ops.reduce_sum(
math_ops.maximum(
math_ops.multiply(loss_mat, mask_positives), 0.0)),
num_positives,
name='triplet_semihard_loss')
There is a division by the number of positives pairs (num_positives), which can lead to NaN.
I suggest you try to inspect your data pipeline in order to ensure there is at least one positive pair in each of your batches. (You can for example adapt some of the code in the triplet_loss_adapted_from_tf to get the num_positives of your batch, and check if it is greater than 0).
Try increasing your batch size. It happened to me also. As mentioned in the previous answer, network is unable to find any num_positives. I had 250 classes and was getting nan loss initially. I increased it to 128/256 and then there was no issue.
I saw that Paris6k has 15 classes or 12 classes. Increase your batch size 32 and if the GPU memory occurs you can try with model with less parameters. You can work on Efficient B0 model for starting. It has 5.3M compared to VGG16 which has 138M parameters.
I have implemented a package for triplet generation so that every batch is guaranteed to include postive pairs. It is compatible with TF/Keras only.
https://github.com/ma7555/kerasgen (Disclaimer: I am the owner)
I would like to compare the classification performance (accuracy) of different classifiers (e.g. CNN, SVM.....), depending on the size of the training data set.
Given is a dataset of images (e.g., MNIST), from which 80% of the images are randomly determined but in compliance with class balance. Subsequently, 80% of the images for the next smaller subset are to be determined from this subset in the same way again. This is repeated until finally a small training amout of about 1000 images is reached.
Each of the classifiers should now be trained with each these subsets.
The aim is to be able to make a statement like for example that from a training size of 5000 images the classifier A is significantly better than classifier B.
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=0, test_size= 0.2, stratify=y)
X_train_2, X_test_2, y_train_2, y_test_2 = train_test_split(X_train, y_train, random_state=0, test_size= 0.2, stratify=y_train)
X_train_3, X_test_3, y_train_3, y_test_3 = train_test_split(X_train_2, y_train_2, random_state=0, test_size= 0.8, stratify=y_train_2)
.....
.....
.....
My problem is that I am not sure if this is really random sampling when I use the above code. Would it better to get the subsets, e.g. using numpy.random.randint?
For any help, I would be very grateful.
I have a simple model for demonstration:
input_layer = Input(shape=(100,))
encoded = Dense(2, activation='relu')(input_layer)
X = np.ones((1000, 100))
Y = np.ones((1000, 2))
print(X.shape)
model = Model(input_layer, encoded)
model.compile(loss='categorical_crossentropy', optimizer='adam')
model.fit(x=X, y=Y, batch_size = 2)
Output is:
2.2.4
(1000, 100)
Epoch 1/1
1000/1000 [==============================] - 3s 3ms/step - loss: 1.3864
Why there are 1000 iterations in one epoch(as shown in the output).
I tried changing this but does not changes the output. I guess it should have been 1000/2 = 500. Please explain what is wrong with my understanding and how can i set the batch size appropriately.
Thanks
In model.fit the numbers in the left part of the progress bar count samples, so it is always the current samples / total number of samples.
Maybe you are confused because it works different in model.fit_generator. There you actually see iterations or batches being counted.
It changes the batch size, the bar progresses faster although you do not explicitly see it as a step. I had the same question in my mind some time ago.
If you want to explicitly see each step, you can use steps_per_epoch and validation_steps.
An example is listed below.
model.fit_generator(training_generator,
steps_per_epoch=steps_per_epoch,
epochs=epochs,
validation_data=validation_generator,
validation_steps=validation_steps)
In this case, steps_per_epoch = number_of_training_samples / batch_size, while validation_steps = number_of_training_samples / batch_size.
During the training, you will see 500 steps instead of 1000 (provided that you have 1000 training samples and your batch_size is 2).