Add candle fe #1603
Add candle fe #1603
Conversation
| dummy = lbann.Dummy(input_, name='dummy') | ||
|
|
||
| # Encoder | ||
| encode1 = lbann.FullyConnected(data, |
There was a problem hiding this comment.
These are repeated layers, try a simple loop/block
| label = lbann.Split(input_, name='label') | ||
|
|
||
| # Encoder | ||
| encode1 = lbann.FullyConnected(finetunedata, |
There was a problem hiding this comment.
Same as above, you can even create a "model class" that separate driver from the actual model, please see JAG or ATOM applications examples
| dummy = lbann.Dummy(input_, name='dummy') | ||
|
|
||
| # Encoder | ||
| encode1 = lbann.FullyConnected(image, |
There was a problem hiding this comment.
Same as above and possibly put this in the vision directory
| gene_dropout1 = lbann.Dropout(gene_relu1, | ||
| name="gene_dropout1", | ||
| data_layout="model_parallel", | ||
| keep_prob=0.95) |
There was a problem hiding this comment.
Please see previous comments.
| keep_prob=0.95) | ||
|
|
||
| # Shared Weights for Drug Tracks | ||
| drug_fc1_w = lbann.Initializer() |
There was a problem hiding this comment.
I dont think this correct, shouldn't it be lbann.Weight(....)?
|
Also, I suggest a comparison of existing prototext with PFE to verify correctness. |
There was a problem hiding this comment.
The models look reasonable to me, aside from some correctness issue with weights. I'd like if all of the CANDLE stuff was moved into a stand-alone directory in the app directory and removed from the model zoo directory. The var_mnist model can go in the vision app.
| encode1 = lbann.FullyConnected(recon_data, | ||
| name="encode1", | ||
| data_layout="model_parallel", | ||
| weights="w1", |
There was a problem hiding this comment.
| weights="w1", | |
| weights=w1, |
|
|
||
| # Encoder | ||
|
|
||
| w1 = lbann.Initializer() |
There was a problem hiding this comment.
| w1 = lbann.Initializer() | |
| w1 = lbann.Weights(initializer=lbann.GlorotUniformInitializer()) |
No description provided.