Related
I am working on a multi-class classification task using my own images.
filenames = [] # a list of filenames
labels = [] # a list of labels corresponding to the filenames
full_ds = tf.data.Dataset.from_tensor_slices((filenames, labels))
This full dataset will be shuffled and split into train, valid and test dataset
full_ds_size = len(filenames)
full_ds = full_ds.shuffle(buffer_size=full_ds_size*2, seed=128) # seed is used for reproducibility
train_ds_size = int(0.64 * full_ds_size)
valid_ds_size = int(0.16 * full_ds_size)
train_ds = full_ds.take(train_ds_size)
remaining = full_ds.skip(train_ds_size)
valid_ds = remaining.take(valid_ds_size)
test_ds = remaining.skip(valid_ds_size)
Now I am struggling to understand how each class is distributed in train_ds, valid_ds and test_ds. An ugly solution is to iterate all the element in the dataset and count the occurrence of each class. Is there any better way to solve it?
My ugly solution:
def get_class_distribution(dataset):
class_distribution = {}
for element in dataset.as_numpy_iterator():
label = element[1]
if label in class_distribution.keys():
class_distribution[label] += 1
else:
class_distribution[label] = 0
# sort dict by key
class_distribution = collections.OrderedDict(sorted(class_distribution.items()))
return class_distribution
train_ds_class_dist = get_class_distribution(train_ds)
valid_ds_class_dist = get_class_distribution(valid_ds)
test_ds_class_dist = get_class_distribution(test_ds)
print(train_ds_class_dist)
print(valid_ds_class_dist)
print(test_ds_class_dist)
The answer below assumes:
there are five classes.
labels are integers from 0 to 4.
It can be modified to suit your needs.
Define a counter function:
def count_class(counts, batch, num_classes=5):
labels = batch['label']
for i in range(num_classes):
cc = tf.cast(labels == i, tf.int32)
counts[i] += tf.reduce_sum(cc)
return counts
Use the reduce operation:
initial_state = dict((i, 0) for i in range(5))
counts = train_ds.reduce(initial_state=initial_state,
reduce_func=count_class)
print([(k, v.numpy()) for k, v in counts.items()])
A solution inspired by user650654 's answer, only using TensorFlow primitives (with tf.unique_with_counts instead of for loop):
In theory, this should have better performance and scale better to large datasets, batches or class count.
num_classes = 5
#tf.function
def count_class(counts, batch):
y, _, c = tf.unique_with_counts(batch[1])
return tf.tensor_scatter_nd_add(counts, tf.expand_dims(y, axis=1), c)
counts = train_ds.reduce(
initial_state=tf.zeros(num_classes, tf.int32),
reduce_func=count_class)
print(counts.numpy())
Similar and simpler version with numpy that actually had better performances for my simple use-case:
count = np.zeros(num_classes, dtype=np.int32)
for _, labels in train_ds:
y, _, c = tf.unique_with_counts(labels)
count[y.numpy()] += c.numpy()
print(count)
I am trying to train a triple loss model using a fit_generator. it requires three input and no output. so i have a function that generates hard triplets. the output from the triplets generator has a shape of (3,5,279) which is 3 inputs(anchor,positive and negative) for 5 batches and a total of 279 features. When i run the fit_generator it throws this error that "the list of Numpy arrays that you are passing to your model is not the size the model expected. Expected to see 3 array(s), but instead got the following list of 1 arrays" meanwhile i have passed a list of three arrays. the code is below. it works when i use the fit, however, i want to always call the generator function to generate my triplets as my batches. thanks in advance..this has taken me three days
def load_data():
path = "arrhythmia_data.txt"
f = open( path, "r")
data = []
#remove line breaker, comma separate and store in array
for line in f:
line = line.replace('\n','').replace('?','0')
line = line.split(",")
data.append(line)
f.close()
data = np.array(data).astype(np.float64)
#print(data.shape)
#create the class labels for input data
Y_train = data[:,-1:]
train = data[:,:-1]
normaliser = preprocessing.MinMaxScaler()
train = normaliser.fit_transform(train)
val = train[320:,:]
train = train[:320,:]
#create one hot encoding of the class labels of the data and separate them into train and test data
lb = LabelBinarizer()
encode = lb.fit_transform(Y_train)
nb_classes = int(len(encode[0]))
#one_hot_labels = keras.utils.to_categorical(labels, num_classes=10) this could also be used for one hot encoding
Y_val_e = encode[320:,:]
Y_train_e = encode[:320,:]
print(Y_train_e[0])
print(np.argmax(Y_train_e[0]))
val_in = []
train_in = []
#grouping and sorting the input data based on label id or name
for n in range(nb_classes):
images_class_n = np.asarray([row for idx,row in enumerate(train) if np.argmax(Y_train_e[idx])==n])
train_in.append(images_class_n)
images_class_n = np.asarray([row for idx,row in enumerate(val) if np.argmax(Y_val_e[idx])==n])
val_in.append(images_class_n)
#print(train_in[0].shape)
return train_in,val_in,Y_train_e,Y_val_e,nb_classes
train_in,val,Y_train,Y_val,nb_classes = load_data()
input_shape = (train_in[0].shape[1],)
def build_network(input_shape , embeddingsize):
'''
Define the neural network to learn image similarity
Input :
input_shape : shape of input images
embeddingsize : vectorsize used to encode our picture
'''
#in_ = Input(train.shape)
net = Sequential()
net.add(Dense(128, activation='relu', input_shape=input_shape))
net.add(Dense(128, activation='relu'))
net.add(Dense(256, activation='relu'))
net.add(Dense(4096, activation='sigmoid'))
net.add(Dense(embeddingsize, activation= None))
#Force the encoding to live on the d-dimentional hypershpere
net.add(Lambda(lambda x: K.l2_normalize(x,axis=-1)))
return net
class TripletLossLayer(Layer):
def __init__(self, alpha, **kwargs):
self.alpha = alpha
super(TripletLossLayer, self).__init__(**kwargs)
def triplet_loss(self, inputs):
anchor, positive, negative = inputs
p_dist = K.sum(K.square(anchor-positive), axis=-1)
n_dist = K.sum(K.square(anchor-negative), axis=-1)
return K.sum(K.maximum(p_dist - n_dist + self.alpha, 0), axis=0)
def call(self, inputs):
loss = self.triplet_loss(inputs)
self.add_loss(loss)
return loss
def build_model(input_shape, network, margin=0.2):
'''
Define the Keras Model for training
Input :
input_shape : shape of input images
network : Neural network to train outputing embeddings
margin : minimal distance between Anchor-Positive and Anchor-Negative for the lossfunction (alpha)
'''
# Define the tensors for the three input images
anchor_input = Input(input_shape, name="anchor_input")
positive_input = Input(input_shape, name="positive_input")
negative_input = Input(input_shape, name="negative_input")
# Generate the encodings (feature vectors) for the three images
encoded_a = network(anchor_input)
encoded_p = network(positive_input)
encoded_n = network(negative_input)
#TripletLoss Layer
loss_layer = TripletLossLayer(alpha=margin,name='triplet_loss_layer')([encoded_a,encoded_p,encoded_n])
# Connect the inputs with the outputs
network_train = Model(inputs=[anchor_input,positive_input,negative_input],outputs=loss_layer)
# return the model
return network_train
def get_batch_random(batch_size,s="train"):
# initialize result
triplets=[np.zeros((batch_size,m)) for i in range(3)]
for i in range(batch_size):
#Pick one random class for anchor
anchor_class = np.random.randint(0, nb_classes)
nb_sample_available_for_class_AP = X[anchor_class].shape[0]
#Pick two different random pics for this class => A and P. You can use same anchor as P if there is one one element for anchor
if nb_sample_available_for_class_AP<=1:
continue
[idx_A,idx_P] = np.random.choice(nb_sample_available_for_class_AP,size=2 ,replace=False)
#Pick another class for N, different from anchor_class
negative_class = (anchor_class + np.random.randint(1,nb_classes)) % nb_classes
nb_sample_available_for_class_N = X[negative_class].shape[0]
#Pick a random pic for this negative class => N
idx_N = np.random.randint(0, nb_sample_available_for_class_N)
triplets[0][i,:] = X[anchor_class][idx_A,:]
triplets[1][i,:] = X[anchor_class][idx_P,:]
triplets[2][i,:] = X[negative_class][idx_N,:]
return np.array(triplets)
def get_batch_hard(draw_batch_size,hard_batchs_size,norm_batchs_size,network,s="train"):
if s == 'train':
X = train_in
else:
X = val
#m, features = X[0].shape
#while True:
#Step 1 : pick a random batch to study
studybatch = get_batch_random(draw_batch_size,X)
#Step 2 : compute the loss with current network : d(A,P)-d(A,N). The alpha parameter here is omited here since we want only to order them
studybatchloss = np.zeros((draw_batch_size))
#Compute embeddings for anchors, positive and negatives
A = network.predict(studybatch[0])
P = network.predict(studybatch[1])
N = network.predict(studybatch[2])
#Compute d(A,P)-d(A,N)
studybatchloss = np.sum(np.square(A-P),axis=1) - np.sum(np.square(A-N),axis=1)
#Sort by distance (high distance first) and take the
selection = np.argsort(studybatchloss)[::-1][:hard_batchs_size]
#Draw other random samples from the batch
selection2 = np.random.choice(np.delete(np.arange(draw_batch_size),selection),norm_batchs_size,replace=False)
selection = np.append(selection,selection2)
triplets = [studybatch[0][selection,:], studybatch[1][selection,:],studybatch[2][selection,:]]
triplets = triplets.reshape(triplets.shape[0],triplets.shape[1],triplets.shape[2])
yield triplets
network = build_network(input_shape,embeddingsize=10)
hard = get_batch_hard(5,4,1,network,s="train")
network_train = build_model(input_shape,network)
optimizer = Adam(lr = 0.00006)
network_train.compile(loss=None,optimizer=optimizer)
#this works
#history = network_train.fit(hard,epochs=100,steps_per_epoch=1, verbose=2)
history = network_train.fit_generator(hard,epochs=10,steps_per_epoch=16, verbose=2)
# error:: the list of Numpy arrays that you are passing to your model is not the size the model
expected. Expected to see 3 array(s), but instead got the following list of 1 arrays:
I think that's beacause in your generator you are yielding the 3 inputs array in one list, you need to yield the 3 arrays independently:
triplet_1 = studybatch[0][selection,:]
triplet_2 = studybatch[1][selection,:]
triplet_3 = studybatch[2][selection,:]
yield [triplet_1, triplet_2, triplet_3]
Is there a way in Keras to retrieve the cell state (i.e., c vector) of a LSTM layer at every timestep of a given input?
It seems the return_state argument returns the last cell state after the computation is done, but I need also the intermediate ones. Also, I don't want to pass these cell states to the next layer, I only want to be able to access them.
Preferably using TensorFlow as backend.
Thanks
I was looking for a solution to this issue and after reading the guidance for creating your own custom RNN Cell in tf.keras (https://www.tensorflow.org/api_docs/python/tf/keras/layers/AbstractRNNCell), I believe the following is the most concise and easy to read way of doing this for Tensorflow 2:
import tensorflow as tf
from tensorflow.keras.layers import LSTMCell
class LSTMCellReturnCellState(LSTMCell):
def call(self, inputs, states, training=None):
real_inputs = inputs[:,:self.units] # decouple [h, c]
outputs, [h,c] = super().call(real_inputs, states, training=training)
return tf.concat([h, c], axis=1), [h,c]
num_units = 512
test_input = tf.random.uniform([5,100,num_units])
rnn = tf.keras.layers.RNN(LSTMCellReturnCellState(num_units),
return_sequences=True, return_state=True)
whole_seq_output, final_memory_state, final_carry_state = rnn(test_input)
print(whole_seq_output.shape)
>>> (5,100,1024)
# Hidden state sequence
h_seq = whole_seq_output[:,:,:num_units] # (5,100,512)
# Cell state sequence
c_seq = whole_seq_output[:,:,num_units:] # (5,100,512)
As mentioned in an above solution, you can see the advantage of this is that it can be easily wrapped into tf.keras.layers.RNN as a drop-in for the normal LSTMCell.
Here is a Colab Notebook with the code running as expected for tensorflow==2.6.0
I know it's pretty late, I hope this can help.
what you are asking, technically, is possible by modifying the LSTM-cell in call method. I modify it and make it return 4 dimension instead of 3 when you give return_sequences=True.
Code
from keras.layers.recurrent import _generate_dropout_mask
class Mod_LSTMCELL(LSTMCell):
def call(self, inputs, states, training=None):
if 0 < self.dropout < 1 and self._dropout_mask is None:
self._dropout_mask = _generate_dropout_mask(
K.ones_like(inputs),
self.dropout,
training=training,
count=4)
if (0 < self.recurrent_dropout < 1 and
self._recurrent_dropout_mask is None):
self._recurrent_dropout_mask = _generate_dropout_mask(
K.ones_like(states[0]),
self.recurrent_dropout,
training=training,
count=4)
# dropout matrices for input units
dp_mask = self._dropout_mask
# dropout matrices for recurrent units
rec_dp_mask = self._recurrent_dropout_mask
h_tm1 = states[0] # previous memory state
c_tm1 = states[1] # previous carry state
if self.implementation == 1:
if 0 < self.dropout < 1.:
inputs_i = inputs * dp_mask[0]
inputs_f = inputs * dp_mask[1]
inputs_c = inputs * dp_mask[2]
inputs_o = inputs * dp_mask[3]
else:
inputs_i = inputs
inputs_f = inputs
inputs_c = inputs
inputs_o = inputs
x_i = K.dot(inputs_i, self.kernel_i)
x_f = K.dot(inputs_f, self.kernel_f)
x_c = K.dot(inputs_c, self.kernel_c)
x_o = K.dot(inputs_o, self.kernel_o)
if self.use_bias:
x_i = K.bias_add(x_i, self.bias_i)
x_f = K.bias_add(x_f, self.bias_f)
x_c = K.bias_add(x_c, self.bias_c)
x_o = K.bias_add(x_o, self.bias_o)
if 0 < self.recurrent_dropout < 1.:
h_tm1_i = h_tm1 * rec_dp_mask[0]
h_tm1_f = h_tm1 * rec_dp_mask[1]
h_tm1_c = h_tm1 * rec_dp_mask[2]
h_tm1_o = h_tm1 * rec_dp_mask[3]
else:
h_tm1_i = h_tm1
h_tm1_f = h_tm1
h_tm1_c = h_tm1
h_tm1_o = h_tm1
i = self.recurrent_activation(x_i + K.dot(h_tm1_i,
self.recurrent_kernel_i))
f = self.recurrent_activation(x_f + K.dot(h_tm1_f,
self.recurrent_kernel_f))
c = f * c_tm1 + i * self.activation(x_c + K.dot(h_tm1_c,
self.recurrent_kernel_c))
o = self.recurrent_activation(x_o + K.dot(h_tm1_o,
self.recurrent_kernel_o))
else:
if 0. < self.dropout < 1.:
inputs *= dp_mask[0]
z = K.dot(inputs, self.kernel)
if 0. < self.recurrent_dropout < 1.:
h_tm1 *= rec_dp_mask[0]
z += K.dot(h_tm1, self.recurrent_kernel)
if self.use_bias:
z = K.bias_add(z, self.bias)
z0 = z[:, :self.units]
z1 = z[:, self.units: 2 * self.units]
z2 = z[:, 2 * self.units: 3 * self.units]
z3 = z[:, 3 * self.units:]
i = self.recurrent_activation(z0)
f = self.recurrent_activation(z1)
c = f * c_tm1 + i * self.activation(z2)
o = self.recurrent_activation(z3)
h = o * self.activation(c)
if 0 < self.dropout + self.recurrent_dropout:
if training is None:
h._uses_learning_phase = True
return tf.expand_dims(tf.concat([h,c],axis=0),0), [h, c]
Sample code
# create a cell
test = Mod_LSTMCELL(100)
# Input timesteps=10, features=7
in1 = Input(shape=(10,7))
out1 = RNN(test, return_sequences=True)(in1)
M = Model(inputs=[in1],outputs=[out1])
M.compile(keras.optimizers.Adam(),loss='mse')
ans = M.predict(np.arange(7*10,dtype=np.float32).reshape(1, 10, 7))
print(ans.shape)
# state_h
print(ans[0,0,0,:])
# state_c
print(ans[0,0,1,:])
First, this is not possible do with the tf.keras.layers.LSTM. You have to use LSTMCell instead or subclass LSTM. Second, there is no need to subclass LSTMCell to get the sequence of cell states. LSTMCell already returns a list of the hidden state (h) and cell state (c) everytime you call it.
For those not familiar with LSTMCell, it takes in the current [h, c] tensors, and the input at the current timestep (it cannot take in a sequence of times) and returns the activations, and the updated [h,c].
Here is an example of showing how to use LSTMCell to process a sequence of timesteps and to return the accumulated cell states.
# example inputs
inputs = tf.convert_to_tensor(np.random.rand(3, 4), dtype='float32') # 3 timesteps, 4 features
h_c = [tf.zeros((1,2)), tf.zeros((1,2))] # must initialize hidden/cell state for lstm cell
h_c = tf.convert_to_tensor(h_c, dtype='float32')
lstm = tf.keras.layers.LSTMCell(2)
# example of how you accumulate cell state over repeated calls to LSTMCell
inputs = tf.unstack(inputs, axis=0)
c_states = []
for cur_inputs in inputs:
out, h_c = lstm(tf.expand_dims(cur_inputs, axis=0), h_c)
h, c = h_c
c_states.append(c)
You can access the states of any RNN by setting return_sequences = True in the initializer. You can find more information about this parameter here.
I'm trying to build a convolutional lstm autoencoder (that also predicts future and past) with Tensorflow, and it works to a certain degree, but the error sometimes jumps back up, so essentially, it never converges.
The model is as follows:
The encoder starts with a 64x64 frame from a 20 frame bouncing mnist video for each time step of the lstm. Every stacking layer of LSTM halfs it and increases the depth via 2x2 convolutions with a stride of 2. (so -->32x32x3 -->...--> 1x1x96)
On the other hand, the lstm performs 3x3 convolutions with a stride of 1 on its state. Both results are concatenated to form the new state. In the same way, the decoder uses transposed convolutions to go back to the original format. Then the squared error is calculated.
The error starts at around 2700 and it takes around 20 hours (geforce1060) to get down to ~1700. At which point the jumping back up (and it sometimes jumps back up to 2300 or even ridiculous values like 440300) happens often enough that I can't really get any lower. Also at that point, it can usually pinpoint where the number should be, but its too fuzzy to actually make out the digit...
I tried different learning rates and optimizers, so if anybody knows why that jumping happens, that'd make me happy :)
Here is a graph of the loss with epochs:
import tensorflow as tf
import numpy as np
import matplotlib
import matplotlib.pyplot as plt
import os
os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
#based on code by loliverhennigh (Github)
class ConvCell(tf.contrib.rnn.RNNCell):
count = 0 #exists only to remove issues with variable scope
def __init__(self, shape, num_features, transpose = False):
self.shape = shape
self.num_features = num_features
self._state_is_tuple = True
self._transpose = transpose
ConvCell.count+=1
self.count = ConvCell.count
#property
def state_size(self):
return (tf.contrib.rnn.LSTMStateTuple(self.shape[0:4],self.shape[0:4]))
#property
def output_size(self):
return tf.TensorShape(self.shape[1:4])
#here comes to the actual conv lstm implementation, if transpose = true, it performs a deconvolution on the input
def __call__(self, inputs, state, scope=None):
with tf.variable_scope(scope or type(self).__name__+str(self.count)):
c, h = state
state_shape = h.shape
input_shape = inputs.shape
#filter variables and convolutions on data coming from the same cell, a time step previous
h_filters = tf.get_variable("h_filters",[3,3,state_shape[3],self.num_features])
h_filters_gates = tf.get_variable("h_filters_gates",[3,3,state_shape[3],3])
h_partial = tf.nn.conv2d(h,h_filters,[1,1,1,1],'SAME')
h_partial_gates = tf.nn.conv2d(h,h_filters_gates,[1,1,1,1],'SAME')
c_filters = tf.get_variable("c_filters",[3,3,state_shape[3],3])
c_partial = tf.nn.conv2d(c,c_filters,[1,1,1,1],'SAME')
#filters and convolutions/deconvolutions on data coming fromthe cell input
if self._transpose:
x_filters = tf.get_variable("x_filters",[2,2,self.num_features,input_shape[3]])
x_filters_gates = tf.get_variable("x_filters_gates",[2,2,3,input_shape[3]])
x_partial = tf.nn.conv2d_transpose(inputs,x_filters,[int(state_shape[0]),int(state_shape[1]),int(state_shape[2]),self.num_features],[1,2,2,1],'VALID')
x_partial_gates = tf.nn.conv2d_transpose(inputs,x_filters_gates,[int(state_shape[0]),int(state_shape[1]),int(state_shape[2]),3],[1,2,2,1],'VALID')
else:
x_filters = tf.get_variable("x_filters",[2,2,input_shape[3],self.num_features])
x_filters_gates = tf.get_variable("x_filters_gates",[2,2,input_shape[3],3])
x_partial = tf.nn.conv2d(inputs,x_filters,[1,2,2,1],'VALID')
x_partial_gates = tf.nn.conv2d(inputs,x_filters_gates,[1,2,2,1],'VALID')
#some more lstm gate business
gate_bias = tf.get_variable("gate_bias",[1,1,1,3])
h_bias = tf.get_variable("h_bias",[1,1,1,self.num_features*2])
gates = h_partial_gates + x_partial_gates + c_partial + gate_bias
i,f,o = tf.split(gates,3,axis=3)
#concatenate the units coming from the spacial and the temporal dimension to build a unified state
concat = tf.concat([h_partial,x_partial],3) + h_bias
new_c = tf.nn.relu(concat)*tf.sigmoid(i)+c*tf.sigmoid(f)
new_h = new_c * tf.sigmoid(o)
new_state = tf.contrib.rnn.LSTMStateTuple(new_c,new_h)
return new_h, new_state #its redundant, but thats how tensorflow likes it, apparently
#global variables
LEARNING_RATE = 0.005
ITERATIONS_PER_EPOCH = 80
BATCH_SIZE = 75
TEST = False #manual switch to go from training to testing
if TEST:
BATCH_SIZE = 1
inputs = tf.placeholder(tf.float32, (20, BATCH_SIZE, 64, 64,1))
shape0 = [BATCH_SIZE,64,64,2]
shape1 = [BATCH_SIZE,32,32,6]
shape2 = [BATCH_SIZE,16,16,12]
shape3 = [BATCH_SIZE,8,8,24]
shape4 = [BATCH_SIZE,4,4,48]
shape5 = [BATCH_SIZE,2,2,96]
shape6 = [BATCH_SIZE,1,1,192]
#apparently tf.multirnncell has very specific requirements for the initial states oO
initial_state1 = (tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape1),tf.zeros(shape1)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape2),tf.zeros(shape2)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape3),tf.zeros(shape3)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape4),tf.zeros(shape4)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape5),tf.zeros(shape5)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape6),tf.zeros(shape6)))
initial_state2 = (tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape5),tf.zeros(shape5)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape4),tf.zeros(shape4)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape3),tf.zeros(shape3)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape2),tf.zeros(shape2)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape1),tf.zeros(shape1)),tf.contrib.rnn.LSTMStateTuple(tf.zeros(shape0),tf.zeros(shape0)))
#encoding part of the autoencoder graph
cell1 = ConvCell(shape1,3)
cell2 = ConvCell(shape2,6)
cell3 = ConvCell(shape3,12)
cell4 = ConvCell(shape4,24)
cell5 = ConvCell(shape5,48)
cell6 = ConvCell(shape6,96)
mcell = tf.contrib.rnn.MultiRNNCell([cell1,cell2,cell3,cell4,cell5,cell6])
rnn_outputs, rnn_states = tf.nn.dynamic_rnn(mcell, inputs[0:20,:,:,:],initial_state=initial_state1,dtype=tf.float32, time_major=True)
#decoding part of the autoencoder graph, forward block and backwards block
cell9a = ConvCell(shape5,48,transpose = True)
cell10a = ConvCell(shape4,24,transpose = True)
cell11a = ConvCell(shape3,12,transpose = True)
cell12a = ConvCell(shape2,6,transpose = True)
cell13a = ConvCell(shape1,3,transpose = True)
cell14a = ConvCell(shape0,1,transpose = True)
mcella = tf.contrib.rnn.MultiRNNCell([cell9a,cell10a,cell11a,cell12a,cell13a,cell14a])
cell9b = ConvCell(shape5,48,transpose = True)
cell10b = ConvCell(shape4,24,transpose = True)
cell11b= ConvCell(shape3,12,transpose = True)
cell12b = ConvCell(shape2,6,transpose = True)
cell13b = ConvCell(shape1,3,transpose = True)
cell14b = ConvCell(shape0,1,transpose = True)
mcellb = tf.contrib.rnn.MultiRNNCell([cell9b,cell10b,cell11b,cell12b,cell13b,cell14b])
def PredictionLayer(rnn_outputs,viewPoint = 11, reverse = False):
predLength = viewPoint-2 if reverse else 20-viewPoint #vision is the input for the decoder
vision = tf.concat([rnn_outputs[viewPoint-1:viewPoint,:,:,:],tf.zeros([predLength,BATCH_SIZE,1,1,192])],0)
if reverse:
rnn_outputs2, rnn_states = tf.nn.dynamic_rnn(mcellb, vision, initial_state = initial_state2, time_major=True)
else:
rnn_outputs2, rnn_states = tf.nn.dynamic_rnn(mcella, vision, initial_state = initial_state2, time_major=True)
mean = tf.reduce_mean(rnn_outputs2,4)
if TEST:
return mean
if reverse:
return tf.reduce_sum(tf.square(mean-inputs[viewPoint-2::-1,:,:,:,0]))
else:
return tf.reduce_sum(tf.square(mean-inputs[viewPoint-1:20,:,:,:,0]))
if TEST:
mean = tf.concat([PredictionLayer(rnn_outputs,11,True)[::-1,:,:,:],createPredictionLayer(rnn_outputs,11)],0)
else: #training part of the graph
error = tf.zeros([1])
for i in range(8,15): #range size of 7 or less works, 9 or more does not, no idea why
error += PredictionLayer(rnn_outputs, i)
error += PredictionLayer(rnn_outputs, i, True)
train_fn = tf.train.RMSPropOptimizer(learning_rate=LEARNING_RATE).minimize(error)
################################################################################
## TRAINING LOOP ##
################################################################################
#code based on siemanko/tf_lstm.py (Github)
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.8)
saver = tf.train.Saver(restore_sequentially=True, allow_empty=True,)
session = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
session.run(tf.global_variables_initializer())
vids = np.load("mnist_test_seq.npy") #20/10000/64/64 , moving mnist dataset from http://www.cs.toronto.edu/~nitish/unsupervised_video/
vids = vids[:,0:6000,:,:] #training set
saver.restore(session,tf.train.latest_checkpoint('./conv_lstm_multiples_v2/'))
#saver.restore(session,'.\conv_lstm_multiples\iteration-74')
for epoch in range(1000):
if TEST:
break
epoch_error = 0
#randomize batches each epoch
vids = np.swapaxes(vids,0,1)
np.random.shuffle(vids)
vids = np.swapaxes(vids,0,1)
for i in range(ITERATIONS_PER_EPOCH):
#running the graph and feeding data
err,_ = session.run([error, train_fn], {inputs: np.expand_dims(vids[:,i*BATCH_SIZE:(i+1)*BATCH_SIZE,:,:],axis=4)})
print(err)
epoch_error += err
#training error each epoch and regular saving
epoch_error /= (ITERATIONS_PER_EPOCH*BATCH_SIZE*4096*20*7)
if (epoch+1) % 5 == 0:
saver.save(session,'.\conv_lstm_multiples_v2\iteration',global_step=epoch)
print("saved")
print("Epoch %d, train error: %f" % (epoch, epoch_error))
#testing
plt.ion()
f, axarr = plt.subplots(2)
vids = np.load("mnist_test_seq.npy")
for i in range(6000,10000):
img = session.run([mean], {inputs: np.expand_dims(vids[:,i:i+1,:,:],axis=4)})
for j in range(20):
axarr[0].imshow(img[0][j,0,:,:])
axarr[1].imshow(vids[j,i,:,:])
plt.show()
plt.pause(0.1)
Usually this happens when gradients' magnitude is very high at some point and causes your network parameters to change a lot. To verify that it is indeed the case, you can produce the same plot of gradient magnitudes and see if they jump right before the loss jump. Assuming this is the case, the classic approach is to use gradient clipping (or go all the way to natural gradient).
The following snippet is from a fairly large piece of code but hopefully I can give all the information necessary:
y2 = tf.matmul(y1,ymask)
dist = tf.norm(ystar-y2,axis=0)
y1 and y2 are 128x30 and ymask is 30x30. ystar is 128x30. dist is 1x30. When ymask is the identity matrix, everything works fine. But when I set it to be all zeros, apart from a single 1 along the diagonal (so as to set all columns but one in y2 to be zero), I get nans for the gradient of dist with respect to y2, using tf.gradients(dist, [y2]). The specific value of dist is [0,0,7.9,0,...], with all the ystar-y2 values being around the range (-1,1) in the third column and zero elsewhere.
I'm pretty confused as to why a numerical issue would occur here, given there are no logs or divisions, is this underflow? Am I missing something in the maths?
For context, I'm doing this to try to train individual dimensions of y, one at a time, using the whole network.
longer version to reproduce:
import tensorflow as tf
import numpy as np
import pandas as pd
batchSize = 128
eta = 0.8
tasks = 30
imageSize = 32**2
groups = 3
tasksPerGroup = 10
trainDatapoints = 10000
w = np.zeros([imageSize, groups * tasksPerGroup])
toyIndex = 0
for toyLoop in range(groups):
m = np.ones([imageSize]) * np.random.randn(imageSize)
for taskLoop in range(tasksPerGroup):
w[:, toyIndex] = m * 0.1 * np.random.randn(1)
toyIndex += 1
xRand = np.random.normal(0, 0.5, (trainDatapoints, imageSize))
taskLabels = np.matmul(xRand, w) + np.random.normal(0,0.5,(trainDatapoints, groups * tasksPerGroup))
DF = np.concatenate((xRand, taskLabels), axis=1)
trainDF = pd.DataFrame(DF[:trainDatapoints, ])
# define graph variables
x = tf.placeholder(tf.float32, [None, imageSize])
W = tf.Variable(tf.zeros([imageSize, tasks]))
b = tf.Variable(tf.zeros([tasks]))
ystar = tf.placeholder(tf.float32, [None, tasks])
ymask = tf.placeholder(tf.float32, [tasks, tasks])
dataLength = tf.cast(tf.shape(ystar)[0],dtype=tf.float32)
y1 = tf.matmul(x, W) + b
y2 = tf.matmul(y1,ymask)
dist = tf.norm(ystar-y2,axis=0)
mse = tf.reciprocal(dataLength) * tf.reduce_mean(tf.square(dist))
grads = tf.gradients(dist, [y2])
trainStep = tf.train.GradientDescentOptimizer(eta).minimize(mse)
# build graph
init = tf.global_variables_initializer()
sess = tf.Session()
sess.run(init)
randTask = np.random.randint(0, 9)
ymaskIn = np.zeros([tasks, tasks])
ymaskIn[randTask, randTask] = 1
batch = trainDF.sample(batchSize)
batch_xs = batch.iloc[:, :imageSize]
batch_ys = np.zeros([batchSize, tasks])
batch_ys[:, randTask] = batch.iloc[:, imageSize + randTask]
gradOut = sess.run(grads, feed_dict={x: batch_xs, ystar: batch_ys, ymask: ymaskIn})
sess.run(trainStep, feed_dict={x: batch_xs, ystar: batch_ys, ymask:ymaskIn})
Here's a very simple reproduction:
import tensorflow as tf
with tf.Graph().as_default():
y = tf.zeros(shape=[1], dtype=tf.float32)
dist = tf.norm(y,axis=0)
(grad,) = tf.gradients(dist, [y])
with tf.Session():
print(grad.eval())
Prints:
[ nan]
The issue is that tf.norm computes sum(x**2)**0.5. The gradient is x / sum(x**2) ** 0.5 (see e.g. https://math.stackexchange.com/a/84333), so when sum(x**2) is zero we're dividing by zero.
There's not much to be done in terms of a special case: the gradient as x approaches all zeros depends on which direction it's approaching from. For example if x is a single-element vector, the limit as x approaches 0 could either be 1 or -1 depending on which side of zero it's approaching from.
So in terms of solutions, you could just add a small epsilon:
import tensorflow as tf
def safe_norm(x, epsilon=1e-12, axis=None):
return tf.sqrt(tf.reduce_sum(x ** 2, axis=axis) + epsilon)
with tf.Graph().as_default():
y = tf.constant([0.])
dist = safe_norm(y,axis=0)
(grad,) = tf.gradients(dist, [y])
with tf.Session():
print(grad.eval())
Prints:
[ 0.]
Note that this is not actually the Euclidean norm. It's a good approximation as long as the input is much larger than epsilon.