from jax import random,vmap
from jax import numpy as jnp
import pprint
def f(s,layers,do,dx):
x = jnp.zeros((do,dx))
for i,layer in enumerate(layers):
x=x.at[i].set( layer( s[i] ) )
return x
class net(nn.Module):
dx: int
do: int
def setup(self):
self.layers = [ nn.Dense( self.dx, use_bias=False )
for _ in range(self.do) ]
def __call__(self, s):
x = vmap(f,in_axes=(0,None,None,None))(s,self.layers,self.do,self.dx)
return x
if __name__ == '__main__':
seed = 123
key = random.PRNGKey( seed )
key,subkey = random.split( key )
outer_batches = 4
s_observations = 5 # AKA the inner batch
x_features = 2
s_features = 3
s_shape = (outer_batches,s_observations, s_features)
s = random.uniform( subkey, s_shape )
key,subkey = random.split( key )
model = net(x_features,s_observations)
p = model.init( subkey, s )
x = model.apply( p, s )
params = p['params']
pkernels = jnp.array([params[key]['kernel'] for key in params.keys()])
x_=jnp.zeros((outer_batches,s_observations,x_features))
g = vmap(vmap(lambda a,b: a@b),in_axes=(0,None))
x_=g(s,pkernels)
print('s shape:',s.shape)
print('p shape:',pkernels.shape)
print('x shape:',x.shape)
print('x_ shape:',x_.shape)
print('sum of difference:',jnp.sum(x-x_))
Hi. I need some “batch-specific” parameters in my model. Here, there is an “inner batch” of length do such that there is a flax.linen.Dense instance for each element in that batch. The outer batch just passes multiple data instances into those layers. I accomplish this by creating a list of flax.linen.Dense instances in the setup method. Then in the __call__ method, I iterate over those layers to fill up an array. This iteration is encapsulated in a function f, and that function is wrapped in jax.vmap.
I have also included some equivalent logic written as matrix multiplication (see the function g) to make it explicit what operation I was hoping to capture with this class.
I would like to replace the for-loop in the __call__ method with a call to jax.vmap. I ofc get an error when I pass a list to vmap, and I ofc get an error when I try to put multiple Dense instances in a jax array. Is there an alternative to using a list to contain my multiple Dense instances? A constraint is that I should be able to create an arbitrary number of Dense instances at the time of model initialization.
vmap can be used to map a single function over batches of data. You are attempting to use it to map multiple functions over batches of data, which it cannot do.
Updated answer based on updated question:
Since each layer is identical aside from the parameters fit to the input data, it sounds like what you want is to map a single dense layer against a batch of data. It might look something like this:
keys = vmap(random.fold_in, in_axes=(None, 0))(subkey, jnp.arange(s_observations))
model = nn.Dense(x_features, use_bias=False)
p = vmap(model.init, in_axes=(0, 1))(keys, s)
x = vmap(model.apply, in_axes=(0, 1), out_axes=1)(p, s)
pkernels = p['params']['kernel']
g = vmap(vmap(lambda a,b: a@b),in_axes=(0,None))
x_=g(s,pkernels)
print('sum of difference:',jnp.sum(x-x_))
# sum of difference: 0.0
Previous answer
In general, the fix would be to define a single parameterized layer that you can pass to vmap. In the example you gave, every layer is identical, and so to achieve the result you’re looking for you could write something like this:
def f(s,layer,dx):
return layer(s)
class net(nn.Module):
dx: int
do: int
def setup(self):
self.layer = nn.Dense( self.dx, use_bias=False )
def __call__(self, s):
x = vmap(f,in_axes=(0,None,None))(s,self.layer,self.dx)
return x
If you had different parameterization per layer, then you could achieve this within vmap by passing those parameters to vmap as well.
7