Skip to content

Parametric input in single CNN model #163

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
sahilyadav27 wants to merge 15 commits into
base: main
Choose a base branch
from
Open

Parametric input in single CNN model #163

Changes from 1 commit
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
7fa649f
Update single_cnn.py
sahilyadav27 Dec 2, 2022
fbc2a1d
Add files via upload
sahilyadav27 Dec 2, 2022
bd92e9b
Update single_cnn.py
sahilyadav27 Dec 2, 2022
893c2a3
Update single_cnn.py
sahilyadav27 Dec 2, 2022
edd8b23
Update single_cnn.py
sahilyadav27 Dec 2, 2022
fe6d9d6
Update singleCNN_param.yml
sahilyadav27 Dec 2, 2022
cfceeaf
Update single_cnn.py
sahilyadav27 Dec 6, 2022
f8194ca
Update single_cnn.py
sahilyadav27 Dec 10, 2022
60a4fb9
Update single_cnn.py
sahilyadav27 Dec 10, 2022
1bbae3f
Update single_cnn.py
sahilyadav27 Dec 10, 2022
5a6510c
Update cnn_rnn.py
sahilyadav27 Dec 10, 2022
66d4e94
Add files via upload
sahilyadav27 Dec 10, 2022
328ebb4
Delete single_cnn.py
sahilyadav27 Dec 10, 2022
2a7acb7
Delete singleCNN_param.yml
sahilyadav27 Dec 10, 2022
c27937f
Update single_nn.py
sahilyadav27 Dec 10, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
29 changes: 16 additions & 13 deletions ctlearn/default_models/single_cnn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,7 +8,10 @@ def single_cnn_model(data, model_params):

# Load neural network model
network_input_img = tf.keras.Input(shape=data.img_shape, name=f"images")
Copy link
Member

@TjarkMiener TjarkMiener Dec 9, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm not sure if we want to support MLP only without the CNNs here or in a separate file. There is no need for an additional file I guess. @nietootein What are your thoughts about this?
I ran a quick check. It is possible to request only the parameter list without the images from dl1dh. @sahilyadav27 So I guess you can just check here if images are selected or not with data.img_shape != None

Copy link
Author

@sahilyadav27 sahilyadav27 Dec 9, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought having a separate MLP model would make sense because we might want to add the MLP model to other models as well like CNN_RNN and ResNet. So it would be better to have a common MLP model independent of the image model rather than having it inside the Single CNN model?

Copy link
Author

@sahilyadav27 sahilyadav27 Dec 10, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should work now

network_input_param = tf.keras.Input(shape=data.prm_shape, name=f"parameters")
flag_prm = 0
if data.prm_shape != None:
flag_prm = 1
network_input_param = tf.keras.Input(shape=data.prm_shape, name=f"parameters")
backbone_name = model_params.get("name", "CNN") + "_block"
trainable_backbone = model_params.get("trainable_backbone", True)
pretrained_weights = model_params.get("pretrained_weights", None)
Expand All @@ -21,10 +24,12 @@ def single_cnn_model(data, model_params):
else:
sys.path.append(model_params["model_directory"])
engine_cnn_module = importlib.import_module(model_params["engine_cnn"]["module"])
engine_mlp_module = importlib.import_module(model_params["engine_mlp"]["module"])
engine_cnn = getattr(engine_cnn_module, model_params["engine_cnn"]["function"])
engine_mlp = getattr(engine_mlp_module, model_params["engine_mlp"]["function"])

if flag_prm == 1:
engine_mlp_module = importlib.import_module(model_params["engine_mlp"]["module"])
engine_mlp = getattr(engine_mlp_module, model_params["engine_mlp"]["function"])
engine_output_mlp = engine_mlp(network_input_param, params=model_params, name=backbone_name)
output_mlp = tf.keras.layers.Flatten()(engine_output_mlp)
# The original ResNet implementation use this padding, but we pad the images in the ImageMapper.
# x = tf.pad(telescope_data, tf.constant([[3, 3], [3, 3]]), name='conv1_pad')
init_layer = model_params.get("init_layer", False)
Expand All @@ -45,15 +50,13 @@ def single_cnn_model(data, model_params):
)(network_input_img)

engine_output_cnn = engine_cnn(network_input_img, params=model_params, name=backbone_name)
engine_output_mlp = engine_mlp(network_input_param, params=model_params, name=backbone_name)

output_cnn = tf.keras.layers.GlobalAveragePooling2D(
name=backbone_name + "_global_avgpool"
)(engine_output_cnn)
output_mlp = tf.keras.layers.Flatten()(engine_output_mlp)
concat = tf.keras.layers.Concatenate()([output_cnn, output_mlp])

singlecnn_model = tf.keras.Model(inputs=[network_input_img, network_input_param], outputs = [concat], name=backbone_name)


return singlecnn_model, [network_input_img, network_input_param]
if flag_prm == 1:
concat = tf.keras.layers.Concatenate()([output_cnn, output_mlp])
singlecnn_model = tf.keras.Model(inputs=[network_input_img, network_input_param], outputs = [concat], name=backbone_name)
return singlecnn_model, [network_input_img, network_input_param]
else:
singlecnn_model = tf.keras.Model(network_input_img, output_cnn, name=backbone_name)
return singlecnn_model, [network_input_img]