The sequence_Loss module's source_code has three parameters that are required they list them as outputs, targets, and weights.
Outputs and targets are self explanatory, but I'm looking to better understand is what is the weight parameter?
The other thing I find confusing is that it states that the targets should be the same length as the outputs, what exactly do they mean by the length of a tensor? Especially if its a 3 dimensional tensor.
Think of the weights as a mask applied to the input tensor. In some NLP applications, we often have different sentence length for each sentence. In order to parallel/batch multiple instance sentences into a minibatch to feed into a neural net, people use a mask matrix to denotes which element in the the input tensor is actually a valid input. For instance, the weight can be a np.ones([batch, max_length]) that means all of the input elements are legit.
We can also use a matrix of the same shape as the labels such as np.asarray([[1,1,1,0],[1,1,0,0],[1,1,1,1]]) (we assume the labels shape is 3x4), then the crossEntropy of the first row last column will be masked out as 0.
You can also use weight to calculate weighted accumulation of cross entropy.
We used this in a class and our professor said we could just pass it ones of the right shape (the comment says "list of 1D batch-sized float-Tensors of the same length as logits"). That doesn't help with what they mean, but maybe it will help you get your code to run. Worked for me.
This code should do the trick: [tf.ones(batch_size, tf.float32) for _ in logits].
Edit: from TF code:
for logit, target, weight in zip(logits, targets, weights):
if softmax_loss_function is None:
# TODO(irving,ebrevdo): This reshape is needed because
# sequence_loss_by_example is called with scalars sometimes, which
# violates our general scalar strictness policy.
target = array_ops.reshape(target, [-1])
crossent = nn_ops.sparse_softmax_cross_entropy_with_logits(
logit, target)
else:
crossent = softmax_loss_function(logit, target)
log_perp_list.append(crossent * weight)
The weights that are passed are multiplied by the loss for that particular logit. So I guess if you want to take a particular prediction extra-seriously you can increase the weight above 1.
Related
I have a model that outputs a Softmax, and I would like to develop a custom loss function. The desired behaviour would be:
1) Softmax to one-hot (normally I do numpy.argmax(softmax_vector) and set that index to 1 in a null vector, but this is not allowed in a loss function).
2) Multiply the resulting one-hot vector by my embedding matrix to get an embedding vector (in my context: the word-vector that is associated to a given word, where words have been tokenized and assigned to indices, or classes for the Softmax output).
3) Compare this vector with the target (this could be a normal Keras loss function).
I know how to write a custom loss function in general, but not to do this. I found this closely related question (unanswered), but my case is a bit different, since I would like to preserve my softmax output.
It is possible to mix tensorflow and keras in you customer loss function. Once you can access to all Tensorflow function, things become very easy. I just give you a example of how this function could be imlement.
import tensorflow as tf
def custom_loss(target, softmax):
max_indices = tf.argmax(softmax, -1)
# Get the embedding matrix. In Tensorflow, this can be directly done
# with tf.nn.embedding_lookup
embedding_vectors = tf.nn.embedding_lookup(you_embedding_matrix, max_indices)
# Do anything you want with normal keras loss function
loss = some_keras_loss_function(target, embedding_vectors)
loss = tf.reduce_mean(loss)
return loss
Fan Luo's answer points in the right direction, but ultimately will not work because it involves non-derivable operations. Note such operations are acceptable for the real value (a loss function takes a real value and a predicted value, non-derivable operations are only fine for the real value).
To be fair, that was what I was asking in the first place. It is not possible to do what I wanted, but we can get a similar and derivable behaviour:
1) Element-wise power of the softmax values. This makes smaller values much smaller. For example, with a power of 4 [0.5, 0.2, 0.7] becomes [0.0625, 0.0016, 0.2400]. Note that 0.2 is comparable to 0.7, but 0.0016 is negligible with respect to 0.24. The higher my_power is, the more similar to a one-hot the final result will be.
soft_extreme = Lambda(lambda x: x ** my_power)(softmax)
2) Importantly, both softmax and one-hot vectors are normalized, but not our "soft_extreme". First, find the sum of the array:
norm = tf.reduce_sum(soft_extreme, 1)
3) Normalize soft_extreme:
almost_one_hot = Lambda(lambda x: x / norm)(soft_extreme)
Note: Setting my_power too high in 1) will result in NaNs. If you need a better softmax to one-hot conversion, then you may do steps 1 to 3 two or more times in a row.
4) Finally we want the vector from the dictionary. Lookup is forbidden, but we can take the average vector using matrix multiplication. Because our soft_normalized is similar to one-hot encoding this average will be similar to the vector associated to the highest argument (original intended behaviour). The higher my_power is in (1), the truer this will be:
target_vectors = tf.tensordot(almost_one_hot, embedding_matrix, axes=[[1], [0]])
Note: This will not work directly using batches! In my case, I reshaped my "one hot" (from [batch, dictionary_length] to [batch, 1, dictionary_length] using tf.reshape. Then tiled my embedding_matrix batch times and finally used:
predicted_vectors = tf.matmul(reshaped_one_hot, tiled_embedding)
There may be more elegant solutions (or less memory-hungry, if tiling the embedding matrix is not an option), so feel free to explore more.
I am using the following code for standard GRU implementation:
def BiRNN_deep_dynamic_FAST_FULL_autolength(x,batch_size,dropout,hidden_dim):
seq_len=length_rnn(x)
with tf.variable_scope('forward'):
lstm_cell_fwd =tf.contrib.rnn.GRUCell(hidden_dim,kernel_initializer=tf.contrib.layers.xavier_initializer(),bias_initializer=tf.contrib.layers.xavier_initializer())
lstm_cell_fwd = tf.contrib.rnn.DropoutWrapper(lstm_cell_fwd, output_keep_prob=dropout)
with tf.variable_scope('backward'):
lstm_cell_back =tf.contrib.rnn.GRUCell(hidden_dim,kernel_initializer=tf.contrib.layers.xavier_initializer(),bias_initializer=tf.contrib.layers.xavier_initializer())
lstm_cell_back = tf.contrib.rnn.DropoutWrapper(lstm_cell_back, output_keep_prob=dropout)
outputs,_= tf.nn.bidirectional_dynamic_rnn(cell_fw=lstm_cell_fwd,cell_bw= lstm_cell_back,inputs=x,sequence_length=seq_len,dtype=tf.float32,time_major=False)
outputs_fwd,outputs_bck=outputs
### fwd matrix is the matrix that keeps all the last [-1] vectors
fwd_matrix=tf.gather_nd(outputs_fwd, tf.stack([tf.range(batch_size), seq_len-1], axis=1)) ### 99,64
outputs_fwd=tf.transpose(outputs_fwd,[1,0,2])
outputs_bck=tf.transpose(outputs_bck,[1,0,2])
return outputs_fwd,outputs_bck,fwd_matrix
Can anyone provide a simple example of how to use the tf.contrib.cudnn_rnn.CudnnGRU Cell in a similar fashion? Just swapping out the cells doesn't work.
First issue is that there is no dropout wrapper for CuDnnGRU cell, which is fine. Second it doesnt seem to work with tf.nn.bidirectional_dynamic_rnn. Any help appreciated.
CudnnGRU is not an RNNCell instance. It's more akin to dynamic_rnn.
The tensor manipulations below are equivalent, where input_tensor is a time-major tensor, i.e. of shape [max_sequence_length, batch_size, embedding_size]. CudnnGRU expects the input tensor to be time-major (as opposed to the more standard batch-major format i.e. of shape [batch_size, max_sequence_length, embedding_size]), and it's a good practice to use time-major tensors with RNN ops anyways since they're somewhat faster.
CudnnGRU:
rnn = tf.contrib.cudnn_rnn.CudnnGRU(
num_rnn_layers, hidden_size, direction='bidirectional')
rnn_output = rnn(input_tensor)
CudnnCompatibleGRUCell:
rnn_output = input_tensor
sequence_length = tf.reduce_sum(
tf.sign(inputs),
reduction_indices=0) # 1 if `input_tensor` is batch-major.
for _ in range(num_rnn_layers):
fw_cell = tf.contrib.cudnn_rnn.CudnnCompatibleGRUCell(hidden_size)
bw_cell = tf.contrib.cudnn_rnn.CudnnCompatibleGRUCell(hidden_size)
rnn_output = tf.nn.bidirectional_dynamic_rnn(
fw_cell, bw_cell, rnn_output, sequence_length=sequence_length,
dtype=tf.float32, time_major=True)[1] # Set `time_major` accordingly
Note the following:
If you were using LSTMs, you need not use CudnnCompatibleLSTMCell; you can use the standard LSTMCell. But with GRUs, the Cudnn implementation has inherently different math operations, and in particular, more weights (see the documentation).
Unlike dynamic_rnn, CudnnGRU doesn't allow you to specify sequence lengths. Still, it is over an order of magnitude faster, but you will have to be careful on how you extract your outputs (e.g. if you're interested in the final hidden state of each sequence that is padded and of varying length, you will need each sequence's length).
rnn_output is probably a tuple with lots of (distinct) stuff in both cases. Refer to the documentation, or just print it out, to inspect what parts of the output you need.
My understanding of noise contrastive estimation is that we sample some vectors from our word embeddings (the negative sample), and then calculate the log-likelihood of each. Then we want to maximize the difference between the probability of the target word and the log-likelihood of each of the negative sample words (So if I'm correct about this, we want to optimize the loss function so that it gets as close to 1 as possible).
My question is this:
What is the purpose of the num_classes parameters to the nce_loss function? My best guess is that the number of classes is passed in so that Tensorflow knows the size of the distribution from which the negative samples our drawn, but this might not make sense, since we could just infer the size of the distribution from the variable itself. Otherwise, I can't think of a reason for why we would need to know the total possible number of classes, especially if the language model is only outputting k + 1 predictions (negative sample size + 1 for the target word).
Your guess is correct. The num_classes argument is used to sample negative labels from the log-uniform (Zipfian) distribution.
Here's the link to the source code:
# Sample the negative labels.
# sampled shape: [num_sampled] tensor
# true_expected_count shape = [batch_size, 1] tensor
# sampled_expected_count shape = [num_sampled] tensor
if sampled_values is None:
sampled_values = candidate_sampling_ops.log_uniform_candidate_sampler(
true_classes=labels,
num_true=num_true,
num_sampled=num_sampled,
unique=True,
range_max=num_classes)
The range_max=num_classes argument basically defines the shape of this distribution and also the range of the sampled values - [0, range_max). Note that this range can't be accurately inferred from the labels, because a particular mini-batch can have only small word ids, which would skew the distribution significantly.
I am aware that there is a similar topic at LSTM Followed by Mean Pooling, but that is about Keras and I work in pure TensorFlow.
I have an LSTM network where the recurrence is handled by:
outputs, final_state = tf.nn.dynamic_rnn(cell,
embed,
sequence_length=seq_lengths,
initial_state=initial_state)
where I pass the correct sequence lengths for each sample (padding by zeros). In any case, outputs contains irrelevant outputs since some samples produce longer outputs than others, based on sequence lengths.
Right now I'm extracting the last relevant output by means of the following method:
def extract_axis_1(data, ind):
"""
Get specified elements along the first axis of tensor.
:param data: Tensorflow tensor that will be subsetted.
:param ind: Indices to take (one for each element along axis 0 of data).
:return: Subsetted tensor.
"""
batch_range = tf.range(tf.shape(data)[0])
indices = tf.stack([batch_range, ind], axis=1)
res = tf.reduce_mean(tf.gather_nd(data, indices), axis=0)
where I pass sequence_length - 1 as indices. In reference to the last topic, I would like to select all relevant outputs followed by average pooling, instead of just the last one.
Now, I tried passing nested lists as indeces to extract_axis_1 but tf.stack does not accept this.
Any solution directions for this?
You can exploit the weight parameter of the tf.contrib.seq2seq.sequence_loss function.
From the documentation:
weights: A Tensor of shape [batch_size, sequence_length] and dtype float. weights constitutes the weighting of each prediction in the sequence. When using weights as masking, set all valid timesteps to 1 and all padded timesteps to 0, e.g. a mask returned by tf.sequence_mask.
You need to compute a binary mask that distinguish between your valid outputs and invalid ones. Then you can just provide this mask to the weights parameter of the loss function (probably, you will want to use a loss like this one); the function will not consider the outputs with a 0 weight in the computation of the loss.
If you can't/don't need to use a sequence loss you can do exactly the same thing manually. You compute a binarymask and then multiply your outputs by this mask and provide these as inputs to your fully connected layer.
How's the following codes different?
with tf.contrib.rnn.DropoutWrapper
enc_cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.DropoutWrapper(tf.contrib.rnn.BasicLSTMCell(rnn_sizes, output_keep_prob=1-keep_prob) for _ in range(num_layers)])
_, encoding_state = tf.nn.dynamic_rnn(enc_cell, rnn_inputs, dtype=tf.float32)
with tf.nn.droupout
enc_cell = tf.contrib.rnn.MultiRNNCell([tf.contrib.rnn.BasicLSTMCell(rnn_size) for _ in range(num_layers)])
_, encoding_state = tf.nn.dynamic_rnn(enc_cell, tf.nn.dropout(rnn_inputs, 1 - keep_prob), dtype=tf.float32)
It seems that there is a difference in the number of states we get from tf.nn.dynamic_rnn. len(encoding state) is greater with tf.nn.dropout.
An explanation will be highly appreciated.
Thank you.
The idea behind both is the same and it is dropout: the network "drops" (i.e does not use) some of its nodes in the prediction. This means reducing during training the capacity of the model to prevent overfitting. Thanks to dropout, the network learns not to rely exclusively on particular nodes for its prediction.
The difference between the two methods is that:
tf.nn.droputis a generic function to perform droput to a given input tensor. Looking at the documentation:
Computes dropout.
With probability keep_prob, outputs the input element scaled up by 1 /
keep_prob, otherwise outputs 0. The scaling is so that the expected
sum is unchanged.
tf.contrib.rnn.DropoutWrapper or tf.nn.rnn_cell.DropoutWrapper is a specific class to define Recurrent Neural Network cells with dropout applied both at the input and the output of the cell. Looking at the documentation:
Operator adding dropout to inputs and outputs of the given cell.
In particular, it uses tf.nn.droput to mask the input to the cell, the state and the output.
The difference between your two pieces of code is that when you are using tf.nn.dropout you are masking the inputs of the first layer only. In the wrapper case, layer per layer, you are masking the outputs of the cells (since you are providing only the output probabilities )
I think dropout can only mask one end, like what you did with rnn_inputs.
DropoutWrapper can mask multi end, like a lstm cell.