riashat / active-learning-bayesian-convolutional-neural-networks Goto Github PK

The paper describes computing variational inference via monte carlo estimates -- where is the monte carlo estimation performed in the code?

In file Active-Learning-Bayesian-Convolutional-Neural-Networks/ConvNets/active_learning/BCNN_cifar10.py , the architecture of the model is still a CNN, rather than a BCNN. Besides, the training method in this file is SGD + momentum, which is also related to CNN other than BCNN (the training method for BCNN should be BBB for example), so how do you implement BCNN by Keras in your experiment?

Hello,

In the code for the standard deviation acquisition function (Segnet), the standard deviation seems to not be properly calculated. Lines 236-237 in the demo read:

    for d_iter in range(dropout_iterations):
        L = np.append(L, All_Dropout_Scores[t, r+10])

Notice that the All_Dropout_Scores index doesn't use d_iter, so the L array effectively contains only copies of the same value. The calculated STD is therefore of 0. This would explain why the results for this approach in the related paper are similar to the random acquisition.

can you give me some advises?
I want to realize that applying traditional active learning method to cnn model, such as maximal entropy, but I fail.

[network]
`

model = Sequential()
model.add(Convolution2D(32, 3, 3, border_mode='same',
                        input_shape=X_train.shape[1:]))
model.add(Activation('relu'))
model.add(Convolution2D(32, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Convolution2D(64, 3, 3, border_mode='same'))
model.add(Activation('relu'))
model.add(Convolution2D(64, 3, 3))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Dropout(0.25))

model.add(Flatten())
model.add(Dense(512))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(nb_classes))
model.add(Activation('softmax'))

# let's train the model using SGD + momentum (how original).
sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
model.compile(loss='categorical_crossentropy',
              optimizer=sgd,
              metrics=['accuracy'])

`
active sampling function:

`
def getData(proba,data,label,batch_data,batch_label,num,flag):
tmpdata=np.empty((num,3,32,32),dtype='float32')
tmplabel=np.empty((num,10),dtype='uint8')
if num==batch_size:
Class_Log_Probability = np.log2(proba)
Entropy_Each_Cell = - np.multiply(proba ,Class_Log_Probability)
Entropy = np.sum(Entropy_Each_Cell, axis=1)
index=select_sort(Entropy,num,flag)
else:
index=get_index(flag)
print(index)

for i in range(num):
    t=index[i]
    flag[t]=1
    tmpdata[i]=data[t]
    tmplabel[i]=label[t]
batch_data,batch_label=np.vstack([batch_data,tmpdata]),np.vstack([batch_label,tmplabel])
return batch_data,batch_label,data,label,flag

def select_sort(list_proba,num,flag):
list_len=len(list_proba)
index=[]
while len(index)<num:
max_index = -1
max_value=-10
for j in range(0, list_len):
if(list_proba[j]>max_value) and j not in index and flag[j]==0:
max_index=j
max_value=list_proba[j]
index.append(max_index)
return index
`
dataset: cifar10

the result: random methods is better. why?

When reading the paper "Deep Bayesian Active Learning with Image Data", I was interested in the results of Figure 2. Specifically, I wanted to replicate the Bald VS Deterministic Bald part.
I followed the file-naming logic which lead me to the code in Softmax_Bald_Q10_N1000.py. So my first question is whether this code is the one behind the results of Deterministic Bald, since it uses predict() instead of stochastic_predict()?
Assuming I got it right, I wondered how the calculation of the average entropy have been made even though we only have one single instance of the predictions.
When looking at the code, for softmax_iterations = 1, the values of G_X = Entropy_Average_Pi and F_X = Average_Entropy should be equal because there is no averaging operations involved. However, when I run the code, the values in U_X = G_X - F_X where, in fact, not zeroed-out which they should have been.
Eventually, it turned out that the empty arrays created before the loop, namely score_All and All_Entropy_Softmax had the automatic dtype=np.float64 while the softmax_score resulting from model.predict() was of type np.float32. Hence, subtracting these arrays, or any subsequent results would produce a non-zero difference.

To verify this, it's just a matter of prefixing the dtype parameters as:

score_All = np.zeros(shape=(X_Pool_Dropout.shape[0], nb_classes), dtype=np.float32)
All_Entropy_Softmax = np.zeros(shape=X_Pool_Dropout.shape[0], dtype=np.float32)

Or removing the loop all together since it is only running for one iteration anyway.