Open geoffwoollard opened 5 years ago
This architecture can get 100% validation roc auc in just several epochs on 0.5 snr (J109, J110).
model = deepconsensus_layers_wrapper(
input_shape=X_val.shape[1::],
num_hidden_layers=4,
conv2d1_k=(15,9,7,3),
conv2d2_k=(15,9,7,3),
conv2d1_n=(8,8,32,64),
conv2d2_n=(8,16,32,64),
mp_k=(7,5,3,4),
mp_strides=(4,2,2,2),
pooling_type=('max','max','max','av'),
dense13_n=512,
dropout13_rate=0.5
)
I can get the same results by changing the dense layer all the way down to 8 neurons, although it takes 5-6 epochs instead of 1-2.
Let's try more noise.
I can vastly simplify the architecture to only have 20k trainable params, and still get perfect performance
model = deepconsensus_layers_wrapper(
input_shape=X_val.shape[1::],
num_hidden_layers=4,
conv2d1_k=(5,5,3,3),
conv2d2_k=(5,5,3,3),
conv2d1_n=(8,8,8,16),
conv2d2_n=(8,8,8,16),
mp_k=(7,5,3,4),
mp_strides=(2,2,2,2),
pooling_type=('max','max','max','av'),
dense13_n=8,
dropout13_rate=0.5
)
However, this is getting close to the limit. For instance this architecture only has 8,8 vs 16.16 channels and it needs 6 epochs
model = deepconsensus_layers_wrapper(
input_shape=X_val.shape[1::],
num_hidden_layers=4,
conv2d1_k=(5,5,3,3),
conv2d2_k=(5,5,3,3),
conv2d1_n=(8,8,8,8),
conv2d2_n=(8,8,8,8),
mp_k=(7,5,3,4),
mp_strides=(2,2,2,2),
pooling_type=('max','max','max','av'),
dense13_n=8,
dropout13_rate=0.5
)
@Davjes15 please make a picture of the smaller network named model-sc-nt0.25-ligand-J179-J180-8dense-ep-9-24kparams-20190403-1832
etc. It's in the model folder in drive.
With snr=0.1 can get mid 90s auc in validation data. This is just for the ligand, which is surprising to me since it's so small.
model = deepconsensus_layers_wrapper(
input_shape=X_val.shape[1::],
num_hidden_layers=4,
conv2d1_k=(9,7,3,3),
conv2d2_k=(9,7,3,3),
conv2d1_n=(8,8,32,64),
conv2d2_n=(8,8,32,64),
mp_k=(7,5,3,4),
mp_strides=(4,2,2,2),
pooling_type=('max','max','max','av'),
dense13_n=512,
dropout13_rate=0.5
)
@geoffwoollard I sent the image and it is also on the google drive sorry for the delay
0.5 snr: J109, J110