thushv89 / adacnn Goto Github PK

will give a speed up

replace with tensorflow later

1st Epoch add actions randomly
2nd epoch add actions deterministically
3rd epoch remove actions randomly
4th epoch onwards prune network

The weights were being updated wrongly. If w is the weights of actor/critic and w' are the weights of target actor/critic, I was doing
w' <- w' * tau + (1-tau) * w'
instead of
w' <- w * tau + (1-tau) * w'

filters/neurons with mean(abs(weights))<threshold will be removed

All the values for actions are sigmoids
Global action -1 - 0 - +1 decide wheterh to remove that proportion from all layers or add or no adapt

Probably due to the very last part of the script, as it should be withing the epoch scope, but was in the batch_id scope

Previously was using all the actions for a single experience tuple, but seems I should have optimized a single action per single experience tuple.

Be careful how you use dropout. For small networks using large dropout causes masive performance degradations.
Use a small validation set accumulation rate
Restart momentum at every start of a new task (or use pool momentum because the pool gets to a uniform distribution over time)

If you are scaling weights to make the expected sum 1, do this consistently for both weights and bias.
normalizing factor for adaptive nets. Very bad

Also have the weighted reward

top_k_accuracy
image_size
...

Say we got the following binned data distribution
0.2 0.2 0.2 0.2 0.2 0.0 0.0 0.0 0.0 0.0
divide by max
1.0 1.0 1.0 1.0 1.0 0 0 0 0 0
use (1 - this) as the dropout vector to drop cnn filters

After add operation,
for the next m batches, (until next action) optimize only the newly added filters with the new data batches.
Then finetune the whole net

After remove operation
Finetune the whole net

Possible causes:
L2 decay: No
Adaptive dropout: No
Pool Momentum:No
Way params removed:
Data distribution

Currently,
Set use_dropout to False
use_l2_loss to False
pool_momentum = 0.0
rm_indices are a continuous block
** no finetune after remove action ** This seems to be the cause of NaNs