functional.rst

objax.functional package

objax.functional

objax.functional

objax.functional

Due to the large number of APIs in this section, we organized it into the following sub-sections:

Activation

celu elu leaky_relu log_sigmoid log_softmax logsumexp relu selu sigmoid softmax softplus tanh

celu

elu

leaky_relu

log_sigmoid

log_softmax

logsumexp

relu

selu

sigmoid

softmax

softplus

tanh

Pooling

average_pool_2d batch_to_space2d channel_to_space2d max_pool_2d space_to_batch2d space_to_channel2d

average_pool_2d

For a definition of pooling, including examples see Pooling Layer.

batch_to_space2d

channel_to_space2d

max_pool_2d

For a definition of pooling, including examples see Pooling Layer.

space_to_batch2d

space_to_channel2d

Misc

dynamic_slice flatten interpolate one_hot pad scan stop_gradient top_k rsqrt upsample_2d upscale_nn

dynamic_slice

flatten

interpolate

one_hot

pad

scan

stop_gradient

top_k

rsqrt

upsample_2d

upscale_nn

objax.functional.divergence

objax.functional.divergence

kl

kl

$$kl(p,q) = p \cdot \log{\frac{p + \epsilon}{q + \epsilon}}$$

The ϵ term is added to ensure that neither p nor q are zero.

objax.functional.loss

objax.functional.loss

cross_entropy_logits cross_entropy_logits_sparse l2 mean_absolute_error mean_squared_error mean_squared_log_error sigmoid_cross_entropy_logits

cross_entropy_logits

Calculates the cross entropy loss, defined as follows:

$$\begin{aligned} \begin{eqnarray} l(y,\hat{y}) & = & - \sum_{j=1}^{q} y_j \log \frac{e^{o_j}}{\sum_{k=1}^{q} e^{o_k}} \nonumber \\\ & = & \log \sum_{k=1}^{q} e^{o_k} - \sum_{j=1}^{q} y_j o_j \nonumber \end{eqnarray} \end{aligned}$$

where o_k are the logits and y_k are the labels.

cross_entropy_logits_sparse

l2

Calculates the l2 loss, as:

$$l_2 = \frac{\sum_{i} x_{i}^2}{2}$$

mean_absolute_error

mean_squared_error

sigmoid_cross_entropy_logits

objax.functional.parallel

objax.functional.parallel

pmax pmean pmin psum

pmax

pmean

pmin

psum