def conv2d(x: xb.ArrayList, params: dict = None) -> xb.ArrayList:
(inputs,) = x
w = params["weights"]
conv = jax.lax.conv_general_dilated(
inputs, w, self.stride, self.padding, feature_group_count=groups
)
return xb.pack(conv)
def train_dropout(x: ArrayList, params=None) -> ArrayList:
array = x[0]
mask = xr.bernoulli(array.shape)
# It's possible the input array can be smaller than previous ones, such as the last batch
# in an epoch. So I'll add a bit in there so the mask will be the same shape as the array
output = pack(
array * mask / max(1 - self.drop_p, 0.00001)
) # The max prevents dividing by zero
return output
def batchnorm2d(inputs: ArrayList, params: dict) -> ArrayList:
(x, ) = inputs
weights, bias = params["weights"], params["bias"]
num_features = x.shape[1]
# Reshaping for broadcasting ease
x_mean = jnp.mean(x, axis=(0, 2, 3)).reshape(1, num_features, 1, 1)
x_var = jnp.mean((x - x_mean)**2,
axis=(0, 2, 3)).reshape(1, num_features, 1, 1)
x_norm = (x - x_mean) / jnp.sqrt(x_var + epsilon)
y = weights * x_norm + bias
return pack(y)
def linear(x: xb.ArrayList, params=None) -> xb.ArrayList:
(inputs,) = x
w = params["weights"]
return xb.pack(jnp.matmul(inputs, w))
def flatten(x: ArrayList, params: dict) -> ArrayList:
return pack(func(x[0]))
def avg_pool2d(inputs: ArrayList, params):
sums = jax.lax.reduce_window(
inputs[0], 0.0, jax.lax.add, window_shape, strides, padding
)
return pack(sums / window_size)